Control Systems Technology Group - User contributions [en]
https://cstwiki.wtb.tue.nl/wiki/Special:Contributions/S143110
User contributionsenMediaWiki 1.37.6Thu, 01 Jun 2023 08:19:48 GMTPRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39352
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39352<p>S143110: /* Section 2: Willingness to follow the system */</p>
<hr />
<div><table cellspacing=1<br />
style="float:right; margin: 0 0 0 1em; border: 1px solid #a2a9b1;background-color: #f9f9f9; padding: .7em;"><br />
<tr><td colspan="2" style="border-color: #a2a9b1; padding: 0.2em 0.4em; text-align: center;"><b>Authors</b></td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">David Dekker</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0936100</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Haico Dorenbos</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0959563</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Jori van Falier</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0910959</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Tim Muyrers</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0907928</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Thomas Schouten</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0938927</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Sten Wessel</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0941508</td></tr><br />
</table><br />
<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self-learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self-driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self-Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
<div style="clear: left; float:left; margin-bottom:0.5em; padding:0.5em 1.4em 0.8em 0;" class="toclimit-2">__TOC__</div><br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people commute to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these:<br />
<br />
# The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
# The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but is also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that cannot be predicted. Because of all these problems we will be using self-learning software to deal with these problems. This program will try to compensate for all the unknowns.<br />
<br />
== Objectives ==<br />
To make a solution for this problem fit in the time we have for this project, we can only focus on some of these aspects. To account for all of them, this would take a lot more years of research and manpower to create this ground-breaking system. Therefore, we will only focus on a simpler situation with one goal and a finite number of cars leaving for that goal. Hence, we only focus at traffic as a ''snapshot'' of time and only for one goal. We think the results of this can be reasonably generalized to a situation with multiple goals and a continuous flow of traffic. The reason for that is that we can extend the behaviour we create to apply to the full-scale problem, by evaluating the situation repeatedly (from ''snapshot'' to continuous) and doing so sequentially for every goal (from single to multi goal). We will elaborate on how this problem scales in the section about complexity analysis: [[#A Small Complexity Analysis]]. For now, we state the objectives for this problem explicitly. We make a distinction between two separate goals.<br />
<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams. This goal is to be satisfied for individual users.<br />
<br />
=== Maximize overall utility ===<br />
Not only minimized travel time individually is important, but also society interest overall. We capture a variety of aspects of interest for society under the umbrella term ''utility''.<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time overall, for all users collectively.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprise ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== Requirements ==<br />
Algorithms are measured in completeness, optimality, time complexity and space complexity. Furthermore the costs of our algorithm and system is also important. So completeness, optimality, time complexity, space complexity and costs are included in the requirements. Because we have combined the completeness and optimality in our main requirement, we also have included time complexity in some of our requirement (we don’t use big O notation in the requirements, because customers are interested in the running time not the big O notation, however we do explain why we meet this requirement by using the big O notation) and last of all cost and space complexity is also combined in the cost requirement (space complexity is included in this, because if you need more memory then you also need to buy more memory or computers).<br />
<br />
=== Main Requirement ===<br />
''For every situation our algorithm has to return for everyone the optimal driving route, such that the sum of the total driving time of everyone is minimal.''<br />
<br />
Note that with ''minimal'' we mean that the sum of the total driving time cannot become any lower than this. For the algorithm implementation, we focus on fulfilling this requirement, because we can in within project bounds a simulation where the fulfilment of this requirement can be measured. However, we will provide a way to extend the algorithm to take the other objectives into account as well.<br />
<br />
=== Cost Requirement ===<br />
''The variable cost per day added with the fixed cost divided by the deprecation period (measured in days) may not exceed €8,96 times the total reduced travel time (also measured in hours) for everyone per day.''<br />
<br />
We explain why we choose this amount. The minimum wage in the Netherlands is €8,96 per hour for people over 23 years. Suppose our system variable cost per hour added with the fixed cost divided by the deprecation period is equal to €7,00 and only one person travel time is reduced by an hour every day and nobody else's travel time is changed, then we have a societal profit of at least €1,96 because that person can start 1 hour earlier with his work and his wage is at least €8,96 per hour. So at least 8,96 – 7,00 = €1,96 societal profit.<br />
<br />
==== Example ====<br />
We want to test if this requirement is satisfied. So let's take a look at a certain algorithm that has a variable cost of €500 per day (electricity costs, hiring employees to maintain the code, etcetera) and the fixed costs are €100.000 (development cost, buying a lot of computers, etcetera). Suppose that the deprecation period of this algorithm is 10 years (after 10 years a new algorithm is found to solve the problem better). Then the variable cost per day added with the fixed cost divided by the deprecation period is equal to 500 + 100.000 / (10 * 365) = 527,40. Now suppose for 100.000 people the driving time on average is decreased by 15 minutes every day and nothing else is changed in driving time. Then we have a societal revenue of 100.000 * 0,25 * 8,96 = €224.000,00. 527,40 <= 224000,00, thus in this case our system satisfies this requirement and has a societal benefit of 224.000,00 – 527,40 = €223.472,60.<br />
<br />
=== After Running Time Requirement ===<br />
''The after running time may be no more than 5 seconds.''<br />
<br />
The after running time is the running time when a customer wants the best route to his destination. The calculation of the best route might only take 5 seconds, this is because customers get impatient if it takes longer than 5 seconds.<br />
<br />
=== Pre Running Time Requirement ===<br />
''The pre running time may be no more than 7 days.''<br />
<br />
This is the running time that the artificial intelligent program takes to prepare the system for being used by drivers (for example a NEAT algorithm, introduced later, that optimizes a neural network is included in this running time). The goal is to optimize for a specific road network, where the traffic distributions are still variable. Hence, this running time may be high, because program only has to be run when the roadmap changes and once in beginning. So users do not have to wait on this algorithm, because this algorithm can run parallel with the calculation of the best route for the users. This is because changes in a road network are known in advance, so when generating a new neural network for the new road situation, the calculation of best routes for users is still for the old road network.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices currently available. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to the current situation only, instead of trying to avoid certain situations beforehand.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will reroute all traffic over this faster route. Right now, this is not really a problem since not all users will follow that advice. However, in the long term (consider for instance self-driving cars), this approach will only create a continuous shifting of recommended routes as a solely reactive system. A self-learning system however find a stable solution, since it can predict traffic jams that haven't even happened yet, and account for the effects of rerouting traffic. It can reroute the user many kilometres before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
[[File:Idea visualize prof.jpg|300 px|thumb|Visualization of the rerouting process.]]<br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously with developing the algorithm. The group of researchers will be split where four members of the team will work on making the self-learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representative for a final implementation as possible.<br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external parties to use and understand the way the system operates, and what conclusions it presents.<br />
<br />
== User and Society ==<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish and less considerate to other drivers. As long as people are driving their car (instead of autonomous), we can investigate further if and how we want to incorporate this.<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey, with the goal of decreasing overall travel time. When looking at literature we find something called ''risk aversion''<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favourable bet (overall travel time decrease) when it means they can directly lose (more travel time)<ref>YouTube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== Traffic research ==<br />
<br />
=== What is a traffic jam ===<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can identify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: unnecessary braking resulting in so called ''phantom jams''<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans) and probably also when cars become fully autonomous, we do see however that accidents can be massively decreased when we decrease the amount of cars driving at a certain road. The busier it is, the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the risk factor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
[[File:FormulaUsed1PNG.png|380 px|thumb|Relation between traffic volume and road crashes.]]<br />
<br />
Too many drivers on the road also causes direct traffic jams through unnecessary braking. When a car in front brakes, the one behind it brakes too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill. This phenomenon is called a ''phantom jam'' which seem to occur and disappear suddenly on crowded roads.<br />
<br />
Road works cannot really be avoided. But if the system is connected to a database, it should be able to take this into account. We can see and handle planned roadwork the same as a (however temporary) change in the road network.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self-learning system should be able to handle occasional accidents and reroute accordingly.<br />
<br />
In conclusion, to make sure the navigational system works in real life you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focused on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be beneficial. Moving on, we abstract away the details of how unpredictable jams occur and disappear and focus training the system based on experiments where these traffic jam situations just happen.<br />
<br />
=== Traffic data ===<br />
<br />
For our system to work, traffic data has to be acquired for realistic large scale data, to help train the system. As our system is still in an early development stage, the data found will not be used as it requires some parsing to get it in a format where the system can work with. As this is too much time consuming, we will focus on a fictional road network inspired on a rush hour commute situation. However, it will show that this data is obtainable and could be used for our system. The data has been acquired from ''Nationale Databank Wegverkeersgegevens'' (NDW) <ref> Nationale Databank Wegverkeersgegevens (NDW), Retrieved March, 2017 from http://www.ndw.nu/nl/ </ref>. Open historic data has been applied for and obtained. It gives information about the traffic density and speed of cars on 27000 locations in the Netherlands. This information can be used for our program to determine how many people travel these roads and when a traffic jam occurs. The data is delivered in the DATEX II form, which is a European standard for traffic data <ref> T. Delissen, Nederlands Profiel DATEX II, 2015-2a, Retrieved April, 2017 from http://www.ndw.nu/downloaddocument/039ee3e1098f37830e26df848f6202ca/Nederlands%20Profiel%20DATEX-II%202015-2a%20(14-01-2016).pdf </ref>. By doing further research about this data, the data can be read by our program and used to train the system. The data base also has current traffic data, that can be used when our system becomes operational. By knowing about the traffic on these 27000 points, the system can get a good understanding of the current traffic situation, and therefore make better decisions. All of this data has to be implemented in the system when it is further developed and can be beneficial to the functioning of our system.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by ''Kennisinstituut voor Mobiliteitsbeleid'' (KiM) from the Ministry of Infrastructure and the Environment in the Netherlands<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3924 respondents from car drivers of the ''Mobiliteitspanel Nederland'' (MPN). <br />
<br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought specifically for that it provides traffic information. Users get most of the traffic information from the internet or the radio and during travel also information panels above the road (DRIPs). From the questions in the KiM research follows that ease-of-use and quality of information are important reasons for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long-distance travel. The system is also used for navigating to points of interest. The most used points of interest are parking lots (75%), gas stations (49%) and places of entertainment (38%). Furthermore, navigation systems are used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of importance follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route themselves, heard of or have seen a traffic jam, or because of road works. Because our system will work optimal when everyone is using the system, and following the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can prove that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors in choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Forms is used to host the questionnaire, because it has all features (kind of questions and setup) required for this research. The questionnaire can be found [https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform here in English] or [https://docs.google.com/forms/d/e/1FAIpQLScx_kPjyN1QNCiAW3xYT-Q77lZks4vDtjyr6YV5x0momVwqVw/viewform here in Dutch], which are identical in content. Please note that by choosing certain answers it is possible that the questionnaire forwards you past irrelevant questions and thus skipping certain sections. In order to see the full questionnaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone receiving the questionnaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non-frequent drivers. While the system is designed for non-self-driving cars it would also work on autonomous cars. The system will first be tested and deployed in the Netherlands and thus this is also the target group for distributing the questionnaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigation system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCNS is designed to be applicable to society at this moment, it thus is dependent on the willingness of the user to follow its instructions. If too many users stray from the path the system proposes, the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing whatever the system proposes should still be faster than not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistle-blower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this they are taken to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily life. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account, and so on. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom-made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know a reasonable amount of time in advance when users are going to get on the road. This section asks for different occasions (different kinds of destinations) the time the user would be able and willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigation system. In this section the participants are asked about possible approaches the system could take regarding traffic and travel time distribution.<br />
<br />
A participant is free to not answer a question if they do not want to answer, without having to provide a reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representative participants for the target group (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well received with up to 95% of the participants showing interest in the product<br />
* 72.4% of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) versus sending the users on secondary roads)<br />
* Point-based system or reward for driving longer routes is preferred among the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travel ahead of time and inform the navigation system of those plans.<br />
* 88.7% 91.9% and 95.1% of participants indicate to be interested in this goal in the form of this kind of product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until April 7). Seven people responded to the English version and 56 filled in the Dutch version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent.<br />
<br />
==== Study outcome section 1: Introduction questions ====<br />
The introduction questions focused on introducing the product and for us, the researchers, to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answers are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9%, respectively.<br />
<br />
So what does this mean? To start, it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter-device communication. As for environmental nuisances it is hard to say how the device would be optimized at this point since the answers are divided among all four categories. <br />
<br />
==== Study outcome section 2: Willingness to follow the system ====<br />
<br />
The second question category tests whether or not participants are willing to follow the system in four different situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
<br />
==== Study outcome section 3: Fairness ====<br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally only one would be the best. The preferred options are a point-based system that will send you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not further specified how this would take shape) sounds appealing to a large part of the participants in both sections respectively scoring 81.9% and 73.8% on ''very good'' or ''good'' compared to the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made).<br />
<br />
==== Study outcome section 4: Privacy ====<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section about privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
==== Study outcome section 5: System interaction ====<br />
This section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people, this loss in performance will most likely be compensated by the extra information the system is able to receive from more users. The responses show that 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be investigated further, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
==== Study outcome section 6: Finalizing questions ====<br />
In the beginning of the questionnaire participants were asked on their view on the importance of environmental factors that influence the functioning of a navigation system. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take air and noise pollution into account or not, as well as having an option to change this setting, or no option.<br />
64% would like standard incorporation of air pollution into the system, with 72% preferring to have an option to change the setting the system is on default. <br />
For noise pollution these values are slightly lower with 60% wanting a default incorporation of noise into the system, and a slightly higher amount of participants (76%) preferring to have an option to change noise incorporation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency, etc.), but will also need a incorporation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
Making system interaction easier is a strong aim in future planning of travel. For this a section of the questionnaire is designed specifically to understand how drivers would like to do this. The aim is to design one to two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants, which is a mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre-planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car as well an option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regarding daily commutes as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
[[File:Goede Graph user research.JPG|666 px|framed|center|Research results on the time ahead users would inform the system of their travel plans for a variety of purposes.]]<br />
As we see in the graph above, on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interested in the system with all possible goals achieving over 50% of interested participants. The three most popular goals are "The overall travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overall travel time is lower", "The overall travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interested in this goal or product. What this means for the design is mostly that the goal can be customized as long as users do not have an increased travel time themselves. <br />
<br />
===='''Final conclusion'''====<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in-depth interviews, the information collected through this limited resources survey gives a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user-influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user-friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of effects that our system introduces could not be specified. For example, the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. The same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things cannot be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed, more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the whole population, although the target group has been reached. Furthermore, the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point-based reward systems are favoured by the users and have to be further researched, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things cannot be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
We start by introducing a very easy model for our problem. It is not very elaborated and therefore the model does not really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etc. is taken into account, due to time constraints). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is static and for now only looks at a distribution of cars at the beginning of the simulation. Moreover, we assume that everyone uses our clever navigation system and that everyone is obedient to the system and therefore they will always follow their directions. Because we first only focus on this simple road network, we make these assumptions to make sure we can measure in a simple and effective way the effects our navigation system has.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like the graph pictured on the right. [[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation used to run the algorithms on.]]<br />
The double-circled vertex is the destination, which means that the drivers from cities B, C and D have to go to this location (we do not take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each bidirectional road there are two edges, one for each direction (the roads connected directly to A however are unidirectional, because if you are at destination A then you do not have to drive away from it). Every driveway in this graph is indicated with a number and every city is indicated with a capital letter.<br />
<br />
==== Data about roads and cities ====<br />
To model a daily commute situation, we will assume that 30 000 people live in city B who all need to go to A. We will furthermore assume that 20 000 people live in city C, who all need to go to city A. Moreover we will assume that 10 000 people live in city D, who all need to go to city A. This is however only for demonstration purposes, as we want our algorithm to be independent on the starting distribution of cars. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower than it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are high-capacity highways, therefore they can for instance have have the function 0.00025x + 15 (in the case when you need at least 15 minutes to drive over that certain road).</li><br />
<li>Roads 2, 4, 5, 6 and 7 are short secondary roads with a lower capacity (the ''N-wegen'' in Dutch) and these roads can for instance have the function 0.0005x + 5 (you need at least 5 minutes to drive over that certain road).</li><br />
<li>Roads 8 en 9 are also secondary roads, but shorter and with an even lower capacity and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm no cars are present on the roads yet. Suppose we have some driver that lives in city C, which drives over the driveways 6, 5 and 1 (in this order to arrive at his destination A. If this happens, then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10 000 people have driven over road 1 and they have waited 17.5 minutes then the cost value for this road is 175 000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving round” a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This process will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit this process for 10 driving rounds, for example (you can take a different number than 10). If after 10 driving rounds not everyone has arrived at destination A, then it will get a high cost value (and therefore this is automatically a bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was to penalize this behaviour, by increasing the cost value for those solutions proportional to the amount of cars that have not reached their destination. This cost value has to be strictly higher than a cost value of a solution for which it is possible for everyone to arrive within 10 driving rounds to destination A. If we define our cost value this way, then a solution for which in 10 driving rounds not everyone gets to their destination is worse than a solution in which everyone gets to their destination instead.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it does not need neural Nntworks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also it can be used for evaluation at different times. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are quite constraining the real-life situation (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time to evaluate. Especially because this algorithm need to determine distributions for every combination of driveway, destination and different parts of the day. If you look for example at an area the size of Germany and the Netherlands, then there is a huge amount of driveways and a huge amount of destinations. This scales badly.</li><br />
<li> Because of the same reason as before, this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a side note, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
====''Optimizing the distribution matrix''====<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
====''The program''====<br />
<br />
The program, written in Python with TensorFlow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is slow for large networks.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an extended version of the “Brick Algorithm”. This algorithm is an improve of the “Brick Algorithm” because it also works for a larger scale. The idea of this algorithm is based on that traffic jams only occur at a local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven. This however seems to be an unrealistic assumption overall for the real world, but also this assumption can be generalized as we will show below.<br />
<br />
==== Situation Representation ====<br />
We are again see the road network as a graph with nodes and vertices, just as in the “Brick Algorithm”. However, this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm). This graph is pictured on the right.<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb|A more complicated road graph example.]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 cars that need to go to A. If a city does not have subscript A, then 200 cars need to go to A from that city. For subscript G and J the same holds analogously. Furthermore, we assume that every black-coloured road has a single lane and has a minimal travel time of 10 minutes. Every orange road has two lanes and has a minimal travel time of 20 minutes. And every green road has three lanes and has a minimal time of 30 minutes. Then we also have red roads, which have a minimal travel time of 15 minutes and consists of a single lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm, which is described below.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes as an example. Thus, every node that is in ''t'' minutes of a goal is taken into account. We take ''t'' to be equal to 30 minutes in this example and we note that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round.<br />
<br />
==== First iteration round ====<br />
In the first iteration round we will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within a minimum of ''t'' waiting time from its goal. We will discard every node in this example that is further than ''t'' time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example we will use the “Brick Algorithm” for three zones with subscript A, G and J (note that zones do not have to be disjunctive) and we find the optimal solution for these zones. After that (still within the same iteration round) you will take back the discarded nodes and vertices and take a look at cars that travel to a goal and are outside of that zone (so for example someone from a city that does not have subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travellers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as cars arrived at that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account and so on. This iteration is stopped after ''u'' iteration rounds have passed where we will determine ''u'' based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is. However, if two goals are close to each other then it will not take the other goal into account, which might not be desirable behaviour. Therefore we say that when we calculate the local solution for a certain goal that then we also take other goals within a radius of 2*''t'' into account, so we use the “Brick Algorithm” for multiple goals at the same time. Then we discard the result of the other goals and we use only the result of the goal that we were calculating. We call this process the “ILDB+ Algorithm”.<br />
<br />
==== Motivation of assumptions ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is seems logical for some situations, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
However, in situations of a daily commute, this is often not the case. We see traffic jams occurring because of accidents, or merging traffic at interchanges. This can happen alongside any route and is often not only near a destination node. This does not mean that the ILDB algorithm approach cannot be used, it is still applicable with some modifications. The main idea of splitting route finding on a local and global level stays in place, only now extended to zones alongside routes as well. We split the complete road network in multiple overlapping zones, such that every inch of road is covered. For the ILDB algorithm to work, it is important that the zones capture traffic hotspots in full. A way we can accomplish this, is to dynamically make zones active and inactive based on the current traffic situations in such a way that jams are as much covered by a single zone while still having all roads covered by zones.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>The assumption that traffic jams going to a certain location only occur close to that location is a assumption, then this algorithm does not work without modification. However, it is unclear and hard to evaluate the performance of this extended algorithm in terms of completeness and optimality.</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine by ourselves how much iteration rounds are needed and what the value for ''t'' is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius ''t'' of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the "+" part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is run is limited). We also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius ''t'' the total driving time of all traffic does not decrease much if you continue with increasing it and what stands out is that the total driving time increases a bit sometimes for higher radius ''t'', because with a higher radius ''t'' you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm can get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is high and therefore because we only use small graphs in our final implementation we do not use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, that we call the “Auction House Algorithm” because it is similar to an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list.<br />
<br />
==== Definition List ====<br />
;Route<br />
:A set of roads that will take you from your start location to your destination.<br />
;In Advance Best<br />
:A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etc.). m is in this case some constant, which will will determine.<br />
;Capacity<br />
:The maximum amount of cars that can take a certain road each hour.<br />
;Claiming<br />
:If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road.<br />
;Neural<br />
:<code>Neural(arg 1, arg 2, ..., arg n)</code> is the output of a Neural Network with n input arguments.<br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm, again here depicted on the right.<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation as used earlier for the Brick Algorithm.]]<br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 lane: 1750 vehicles per hour </li><br />
<li> 2 lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving or waiting time of a road is easily determined. We will assume that the daily commute rush hour will have a duration of 2 hours. And that when x amount of people in the rush hour need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the driving or waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location and destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<br />
;Shortest path with road (a)<br />
:The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams into account.<br />
;Shortest path without road (b)<br />
:The total driving time that someone needs to go from its location to their destination without taking traffic jams into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.)<br />
;Own total time with road (c)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account.<br />
:Own total time without road (d)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account. Also given that the m in advance best routes might not use this road.<br />
;How much other claims (e)<br />
:How much other start location/destination combinations make a claim on this road.<br />
;Sum shortest path others with road (f)<br />
:The total sum of the shortest path with road for other start location and destination combinations (including other destinations from the same city).<br />
;Sum shortest path others without road (g)<br />
:The total sum of the shortest path without road for other start location and destination combinations.<br />
;Sum total time others with road (h)<br />
:The total sum of the own total time with road for other start location and destination combinations.<br />
;Sum total time others without road (i)<br />
:The total sum of the own total time without road for other start location and destination combinations.<br />
;How much people claim the road (j)<br />
:The total sum of all drivers that claim this road.<br />
;Capacity of a road (k)<br />
:The maximum capacity of a road.<br />
<br />
We define: <code>l = Neural(a, b, c, d, e, f, g, h, i, j, k)</code>. l is probably a value between 0 and 1 (for simplicity, because it is the output of a neural network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location and destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + ... + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unbounded m? ====<br />
A good question is why we bound m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m does not have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evaluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behaviour of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility <!-- NOTE: Capital letter O is intended spelling. -->, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behaviour?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behaviour of a solution, however don’t conclude that one of the solutions is bad because it haves different behaviour. Because the total driving time might be the same for solutions with different behaviour. To illustrate this we will show a graph of a 3D function, depicted on the right.<br />
[[File:Different_Behavior_Same_Result.png|thumb|Graph on a 3D function with an infinite number of absolute minimum values.]]<br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behaviour for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like a tetrahedron or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="Some predetermined graphs"><br />
File:Brick Algorithm - Situation Representation.png|Tetrahedron model.<br />
File:ILDB(+) Algorithm - Situation Representation.png|Complex Rectangle model.<br />
</gallery><br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is shown on the right.<br />
[[File:RandomGraph.png|300px|framed|center|Generated Total Random Graph.]]<br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward). This can be seen in the figure below.<br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get the other figure shown below.<br />
<br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="RR-Grid construction"><br />
File:USE_-_Grid.png|A square neighbor-fully-connected rectangular grid.<br />
File:USE - RRGrid.png|An RR-Grid based on the grid above.<br />
</gallery><br />
<br />
== A Small Complexity Analysis ==<br />
<br />
We have a look at how the training time increases as the amount of cities and routes increases. Unfortunately we can only run this for our very small problem, which may not say very much about training for bigger road networks.<br />
We use the tetrahedron model mentioned before in the situation representation of the Brick Algorithm, where road 2, 4 and 6 are highways with maximum speed 30 and length 280. The other roads used, 1 and 3, secondary roads with maximum speed 17 and length 354. People start from cities B, C and D and all want to arrive at goal A.<br />
<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|The tetrahedron model used for the complexity analysis.]]<br />
<br />
We test three situations, first with only city A and B and one road from B to A, then we add the city D with three routes to the goal. These routes that people can drive are from B to A directly, from B to A via D and from D to A. Similarly we extend to four cities with five routes.<br />
<br />
We train using only one species, a population size of 5 and train for only 5 generations to keep it fast to test.<br />
<br />
We run the training script five times and then take the average running time, so that the time found will not differ more than a few tenths on different runs of the complexity analysis.<br />
<br />
* Two cities (A and B) and one route: 7.3 seconds<br />
* Three cities (A, B and D) and three routes: 10.4 seconds<br />
* Full graph with four cities and nine routes: 12.7 seconds<br />
<br />
From this we could conclude that if we want to use very big road networks, we probably need to devise optimizations to keep the training tractable. But remember this is just a small test with possibly numbers too small to draw any proper conclusions from, so this conclusion is not definite at all.<br />
<br />
== Running the simulation ==<br />
[[File:Pre-2016-q3-group13-simulation.gif|thumb|333px|Simulation ([[:File:PRE16-3-group13-neat-implementation.zip|code]]) of our system compared to a conventional situation where all drivers use the common highway route. Click to run the animation.]]<br />
With the NEAT algorithm, compiling all of the approaches and thoughts into a final implementation, we run a simulation to show the results of our system. This adds to all research done above and is a proof-of-concept of the technical capabilities of this system.<br />
We use a slightly modified version if the tetrahedron model for the simulation, in an attempt to make it somewhat more realistic. It's shape is modeled after the real-life network between the Dutch cities Amsterdam (the goal at the north side), Utrecht (in the middle), The Hague and Eindhoven (south), and Leiden and Amersfoort (west and east of Utrecht, respectively). The roads connecting the large cities (Amsterdam, Utrecht, The Hague and Eindhoven) are 2-lane highways, while the other roads are 1-lane secondary roads with a lower maximum speed. Because this simulates on a low scale, the For this simulation, a random amount of cars between 20 and 25 (for demonstration purposes) leaves a city with destination Amsterdam. However, the cars sent on the left simulation and the right simulation is the same, for a fair comparison.<br />
<br />
On the left in the animation, we simulate our system, where cars are distributed based on the output the evolved neural network provides. On the right the same situation in terms of initial car distribution is shown, but cars always choose the fastest highway route from their city to Amsterdam. As clearly visible in the animation, our system outperforms the situation on the right in terms of total duration before all cars reach the destination. Our system here is based on the NEAT algorithm approach primarily, which is evolved on variable starting distribution of cars on this fixed road network. The generation of this network took less than 15 minutes, while the evaluation once generated was almost instant. For this small example, the time requirements are met, as well as the main requirement as we see improvement of traffic flow in the new situation. The cost requirement is harder to check on this small scale, but since overall travel times are shorter, it seems promising.<br />
<br />
We know that this is still far from a real-life implementation as a navigation system ready to use in consumer cars. For that, we have laid the groundwork on how to expand this simple situation to larger areas with the sections above. With that, the conducted user research will start to show its results in the implementation directly. Also, the scaling of the running time also becomes important, as covered in the previous section. We have shown that it is not easy to make a finished product for use in society, but we hope that with all the research we have done, the best we can, we have shown ways on how it is possible to extend our basic simulation to a more realistic implementation, in terms of scaling and letting the user interests into the design. We are proud to show that we can apply self-learning techniques in an unconventional way, with using the NEAT algorithm approach, into this kind of problem: something we have not seen applied before.<br />
<br />
==References==<br />
''For some reason, references are not showing up when not logged in to the wiki. They are visible when looking at the source of this page.''<br />
<br />
<references/></div>Fri, 14 Apr 2017 10:14:04 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39351
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39351<p>S143110: /* Section 2: Willingness to follow the system */</p>
<hr />
<div><table cellspacing=1<br />
style="float:right; margin: 0 0 0 1em; border: 1px solid #a2a9b1;background-color: #f9f9f9; padding: .7em;"><br />
<tr><td colspan="2" style="border-color: #a2a9b1; padding: 0.2em 0.4em; text-align: center;"><b>Authors</b></td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">David Dekker</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0936100</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Haico Dorenbos</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0959563</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Jori van Falier</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0910959</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Tim Muyrers</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0907928</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Thomas Schouten</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0938927</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Sten Wessel</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0941508</td></tr><br />
</table><br />
<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self-learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self-driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self-Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
<div style="clear: left; float:left; margin-bottom:0.5em; padding:0.5em 1.4em 0.8em 0;" class="toclimit-2">__TOC__</div><br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people commute to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these:<br />
<br />
# The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
# The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but is also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that cannot be predicted. Because of all these problems we will be using self-learning software to deal with these problems. This program will try to compensate for all the unknowns.<br />
<br />
== Objectives ==<br />
To make a solution for this problem fit in the time we have for this project, we can only focus on some of these aspects. To account for all of them, this would take a lot more years of research and manpower to create this ground-breaking system. Therefore, we will only focus on a simpler situation with one goal and a finite number of cars leaving for that goal. Hence, we only focus at traffic as a ''snapshot'' of time and only for one goal. We think the results of this can be reasonably generalized to a situation with multiple goals and a continuous flow of traffic. The reason for that is that we can extend the behaviour we create to apply to the full-scale problem, by evaluating the situation repeatedly (from ''snapshot'' to continuous) and doing so sequentially for every goal (from single to multi goal). We will elaborate on how this problem scales in the section about complexity analysis: [[#A Small Complexity Analysis]]. For now, we state the objectives for this problem explicitly. We make a distinction between two separate goals.<br />
<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams. This goal is to be satisfied for individual users.<br />
<br />
=== Maximize overall utility ===<br />
Not only minimized travel time individually is important, but also society interest overall. We capture a variety of aspects of interest for society under the umbrella term ''utility''.<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time overall, for all users collectively.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprise ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== Requirements ==<br />
Algorithms are measured in completeness, optimality, time complexity and space complexity. Furthermore the costs of our algorithm and system is also important. So completeness, optimality, time complexity, space complexity and costs are included in the requirements. Because we have combined the completeness and optimality in our main requirement, we also have included time complexity in some of our requirement (we don’t use big O notation in the requirements, because customers are interested in the running time not the big O notation, however we do explain why we meet this requirement by using the big O notation) and last of all cost and space complexity is also combined in the cost requirement (space complexity is included in this, because if you need more memory then you also need to buy more memory or computers).<br />
<br />
=== Main Requirement ===<br />
''For every situation our algorithm has to return for everyone the optimal driving route, such that the sum of the total driving time of everyone is minimal.''<br />
<br />
Note that with ''minimal'' we mean that the sum of the total driving time cannot become any lower than this. For the algorithm implementation, we focus on fulfilling this requirement, because we can in within project bounds a simulation where the fulfilment of this requirement can be measured. However, we will provide a way to extend the algorithm to take the other objectives into account as well.<br />
<br />
=== Cost Requirement ===<br />
''The variable cost per day added with the fixed cost divided by the deprecation period (measured in days) may not exceed €8,96 times the total reduced travel time (also measured in hours) for everyone per day.''<br />
<br />
We explain why we choose this amount. The minimum wage in the Netherlands is €8,96 per hour for people over 23 years. Suppose our system variable cost per hour added with the fixed cost divided by the deprecation period is equal to €7,00 and only one person travel time is reduced by an hour every day and nobody else's travel time is changed, then we have a societal profit of at least €1,96 because that person can start 1 hour earlier with his work and his wage is at least €8,96 per hour. So at least 8,96 – 7,00 = €1,96 societal profit.<br />
<br />
==== Example ====<br />
We want to test if this requirement is satisfied. So let's take a look at a certain algorithm that has a variable cost of €500 per day (electricity costs, hiring employees to maintain the code, etcetera) and the fixed costs are €100.000 (development cost, buying a lot of computers, etcetera). Suppose that the deprecation period of this algorithm is 10 years (after 10 years a new algorithm is found to solve the problem better). Then the variable cost per day added with the fixed cost divided by the deprecation period is equal to 500 + 100.000 / (10 * 365) = 527,40. Now suppose for 100.000 people the driving time on average is decreased by 15 minutes every day and nothing else is changed in driving time. Then we have a societal revenue of 100.000 * 0,25 * 8,96 = €224.000,00. 527,40 <= 224000,00, thus in this case our system satisfies this requirement and has a societal benefit of 224.000,00 – 527,40 = €223.472,60.<br />
<br />
=== After Running Time Requirement ===<br />
''The after running time may be no more than 5 seconds.''<br />
<br />
The after running time is the running time when a customer wants the best route to his destination. The calculation of the best route might only take 5 seconds, this is because customers get impatient if it takes longer than 5 seconds.<br />
<br />
=== Pre Running Time Requirement ===<br />
''The pre running time may be no more than 7 days.''<br />
<br />
This is the running time that the artificial intelligent program takes to prepare the system for being used by drivers (for example a NEAT algorithm, introduced later, that optimizes a neural network is included in this running time). The goal is to optimize for a specific road network, where the traffic distributions are still variable. Hence, this running time may be high, because program only has to be run when the roadmap changes and once in beginning. So users do not have to wait on this algorithm, because this algorithm can run parallel with the calculation of the best route for the users. This is because changes in a road network are known in advance, so when generating a new neural network for the new road situation, the calculation of best routes for users is still for the old road network.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices currently available. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to the current situation only, instead of trying to avoid certain situations beforehand.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will reroute all traffic over this faster route. Right now, this is not really a problem since not all users will follow that advice. However, in the long term (consider for instance self-driving cars), this approach will only create a continuous shifting of recommended routes as a solely reactive system. A self-learning system however find a stable solution, since it can predict traffic jams that haven't even happened yet, and account for the effects of rerouting traffic. It can reroute the user many kilometres before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
[[File:Idea visualize prof.jpg|300 px|thumb|Visualization of the rerouting process.]]<br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously with developing the algorithm. The group of researchers will be split where four members of the team will work on making the self-learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representative for a final implementation as possible.<br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external parties to use and understand the way the system operates, and what conclusions it presents.<br />
<br />
== User and Society ==<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish and less considerate to other drivers. As long as people are driving their car (instead of autonomous), we can investigate further if and how we want to incorporate this.<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey, with the goal of decreasing overall travel time. When looking at literature we find something called ''risk aversion''<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favourable bet (overall travel time decrease) when it means they can directly lose (more travel time)<ref>YouTube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== Traffic research ==<br />
<br />
=== What is a traffic jam ===<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can identify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: unnecessary braking resulting in so called ''phantom jams''<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans) and probably also when cars become fully autonomous, we do see however that accidents can be massively decreased when we decrease the amount of cars driving at a certain road. The busier it is, the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the risk factor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
[[File:FormulaUsed1PNG.png|380 px|thumb|Relation between traffic volume and road crashes.]]<br />
<br />
Too many drivers on the road also causes direct traffic jams through unnecessary braking. When a car in front brakes, the one behind it brakes too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill. This phenomenon is called a ''phantom jam'' which seem to occur and disappear suddenly on crowded roads.<br />
<br />
Road works cannot really be avoided. But if the system is connected to a database, it should be able to take this into account. We can see and handle planned roadwork the same as a (however temporary) change in the road network.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self-learning system should be able to handle occasional accidents and reroute accordingly.<br />
<br />
In conclusion, to make sure the navigational system works in real life you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focused on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be beneficial. Moving on, we abstract away the details of how unpredictable jams occur and disappear and focus training the system based on experiments where these traffic jam situations just happen.<br />
<br />
=== Traffic data ===<br />
<br />
For our system to work, traffic data has to be acquired for realistic large scale data, to help train the system. As our system is still in an early development stage, the data found will not be used as it requires some parsing to get it in a format where the system can work with. As this is too much time consuming, we will focus on a fictional road network inspired on a rush hour commute situation. However, it will show that this data is obtainable and could be used for our system. The data has been acquired from ''Nationale Databank Wegverkeersgegevens'' (NDW) <ref> Nationale Databank Wegverkeersgegevens (NDW), Retrieved March, 2017 from http://www.ndw.nu/nl/ </ref>. Open historic data has been applied for and obtained. It gives information about the traffic density and speed of cars on 27000 locations in the Netherlands. This information can be used for our program to determine how many people travel these roads and when a traffic jam occurs. The data is delivered in the DATEX II form, which is a European standard for traffic data <ref> T. Delissen, Nederlands Profiel DATEX II, 2015-2a, Retrieved April, 2017 from http://www.ndw.nu/downloaddocument/039ee3e1098f37830e26df848f6202ca/Nederlands%20Profiel%20DATEX-II%202015-2a%20(14-01-2016).pdf </ref>. By doing further research about this data, the data can be read by our program and used to train the system. The data base also has current traffic data, that can be used when our system becomes operational. By knowing about the traffic on these 27000 points, the system can get a good understanding of the current traffic situation, and therefore make better decisions. All of this data has to be implemented in the system when it is further developed and can be beneficial to the functioning of our system.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by ''Kennisinstituut voor Mobiliteitsbeleid'' (KiM) from the Ministry of Infrastructure and the Environment in the Netherlands<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3924 respondents from car drivers of the ''Mobiliteitspanel Nederland'' (MPN). <br />
<br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought specifically for that it provides traffic information. Users get most of the traffic information from the internet or the radio and during travel also information panels above the road (DRIPs). From the questions in the KiM research follows that ease-of-use and quality of information are important reasons for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long-distance travel. The system is also used for navigating to points of interest. The most used points of interest are parking lots (75%), gas stations (49%) and places of entertainment (38%). Furthermore, navigation systems are used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of importance follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route themselves, heard of or have seen a traffic jam, or because of road works. Because our system will work optimal when everyone is using the system, and following the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can prove that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors in choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Forms is used to host the questionnaire, because it has all features (kind of questions and setup) required for this research. The questionnaire can be found [https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform here in English] or [https://docs.google.com/forms/d/e/1FAIpQLScx_kPjyN1QNCiAW3xYT-Q77lZks4vDtjyr6YV5x0momVwqVw/viewform here in Dutch], which are identical in content. Please note that by choosing certain answers it is possible that the questionnaire forwards you past irrelevant questions and thus skipping certain sections. In order to see the full questionnaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone receiving the questionnaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non-frequent drivers. While the system is designed for non-self-driving cars it would also work on autonomous cars. The system will first be tested and deployed in the Netherlands and thus this is also the target group for distributing the questionnaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigation system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCNS is designed to be applicable to society at this moment, it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing whatever the system proposes should still be faster than not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistle-blower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this they are taken to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily life. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account, and so on. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom-made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know a reasonable amount of time in advance when users are going to get on the road. This section asks for different occasions (different kinds of destinations) the time the user would be able and willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigation system. In this section the participants are asked about possible approaches the system could take regarding traffic and travel time distribution.<br />
<br />
A participant is free to not answer a question if they do not want to answer, without having to provide a reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representative participants for the target group (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well received with up to 95% of the participants showing interest in the product<br />
* 72.4% of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) versus sending the users on secondary roads)<br />
* Point-based system or reward for driving longer routes is preferred among the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travel ahead of time and inform the navigation system of those plans.<br />
* 88.7% 91.9% and 95.1% of participants indicate to be interested in this goal in the form of this kind of product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until April 7). Seven people responded to the English version and 56 filled in the Dutch version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent.<br />
<br />
==== Study outcome section 1: Introduction questions ====<br />
The introduction questions focused on introducing the product and for us, the researchers, to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answers are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9%, respectively.<br />
<br />
So what does this mean? To start, it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter-device communication. As for environmental nuisances it is hard to say how the device would be optimized at this point since the answers are divided among all four categories. <br />
<br />
==== Study outcome section 2: Willingness to follow the system ====<br />
<br />
The second question category tests whether or not participants are willing to follow the system in four different situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
<br />
==== Study outcome section 3: Fairness ====<br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally only one would be the best. The preferred options are a point-based system that will send you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not further specified how this would take shape) sounds appealing to a large part of the participants in both sections respectively scoring 81.9% and 73.8% on ''very good'' or ''good'' compared to the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made).<br />
<br />
==== Study outcome section 4: Privacy ====<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section about privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
==== Study outcome section 5: System interaction ====<br />
This section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people, this loss in performance will most likely be compensated by the extra information the system is able to receive from more users. The responses show that 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be investigated further, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
==== Study outcome section 6: Finalizing questions ====<br />
In the beginning of the questionnaire participants were asked on their view on the importance of environmental factors that influence the functioning of a navigation system. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take air and noise pollution into account or not, as well as having an option to change this setting, or no option.<br />
64% would like standard incorporation of air pollution into the system, with 72% preferring to have an option to change the setting the system is on default. <br />
For noise pollution these values are slightly lower with 60% wanting a default incorporation of noise into the system, and a slightly higher amount of participants (76%) preferring to have an option to change noise incorporation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency, etc.), but will also need a incorporation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
Making system interaction easier is a strong aim in future planning of travel. For this a section of the questionnaire is designed specifically to understand how drivers would like to do this. The aim is to design one to two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants, which is a mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre-planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car as well an option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regarding daily commutes as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
[[File:Goede Graph user research.JPG|666 px|framed|center|Research results on the time ahead users would inform the system of their travel plans for a variety of purposes.]]<br />
As we see in the graph above, on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interested in the system with all possible goals achieving over 50% of interested participants. The three most popular goals are "The overall travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overall travel time is lower", "The overall travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interested in this goal or product. What this means for the design is mostly that the goal can be customized as long as users do not have an increased travel time themselves. <br />
<br />
===='''Final conclusion'''====<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in-depth interviews, the information collected through this limited resources survey gives a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user-influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user-friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of effects that our system introduces could not be specified. For example, the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. The same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things cannot be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed, more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the whole population, although the target group has been reached. Furthermore, the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point-based reward systems are favoured by the users and have to be further researched, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things cannot be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
We start by introducing a very easy model for our problem. It is not very elaborated and therefore the model does not really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etc. is taken into account, due to time constraints). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is static and for now only looks at a distribution of cars at the beginning of the simulation. Moreover, we assume that everyone uses our clever navigation system and that everyone is obedient to the system and therefore they will always follow their directions. Because we first only focus on this simple road network, we make these assumptions to make sure we can measure in a simple and effective way the effects our navigation system has.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like the graph pictured on the right. [[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation used to run the algorithms on.]]<br />
The double-circled vertex is the destination, which means that the drivers from cities B, C and D have to go to this location (we do not take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each bidirectional road there are two edges, one for each direction (the roads connected directly to A however are unidirectional, because if you are at destination A then you do not have to drive away from it). Every driveway in this graph is indicated with a number and every city is indicated with a capital letter.<br />
<br />
==== Data about roads and cities ====<br />
To model a daily commute situation, we will assume that 30 000 people live in city B who all need to go to A. We will furthermore assume that 20 000 people live in city C, who all need to go to city A. Moreover we will assume that 10 000 people live in city D, who all need to go to city A. This is however only for demonstration purposes, as we want our algorithm to be independent on the starting distribution of cars. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower than it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are high-capacity highways, therefore they can for instance have have the function 0.00025x + 15 (in the case when you need at least 15 minutes to drive over that certain road).</li><br />
<li>Roads 2, 4, 5, 6 and 7 are short secondary roads with a lower capacity (the ''N-wegen'' in Dutch) and these roads can for instance have the function 0.0005x + 5 (you need at least 5 minutes to drive over that certain road).</li><br />
<li>Roads 8 en 9 are also secondary roads, but shorter and with an even lower capacity and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm no cars are present on the roads yet. Suppose we have some driver that lives in city C, which drives over the driveways 6, 5 and 1 (in this order to arrive at his destination A. If this happens, then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10 000 people have driven over road 1 and they have waited 17.5 minutes then the cost value for this road is 175 000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving round” a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This process will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit this process for 10 driving rounds, for example (you can take a different number than 10). If after 10 driving rounds not everyone has arrived at destination A, then it will get a high cost value (and therefore this is automatically a bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was to penalize this behaviour, by increasing the cost value for those solutions proportional to the amount of cars that have not reached their destination. This cost value has to be strictly higher than a cost value of a solution for which it is possible for everyone to arrive within 10 driving rounds to destination A. If we define our cost value this way, then a solution for which in 10 driving rounds not everyone gets to their destination is worse than a solution in which everyone gets to their destination instead.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it does not need neural Nntworks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also it can be used for evaluation at different times. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are quite constraining the real-life situation (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time to evaluate. Especially because this algorithm need to determine distributions for every combination of driveway, destination and different parts of the day. If you look for example at an area the size of Germany and the Netherlands, then there is a huge amount of driveways and a huge amount of destinations. This scales badly.</li><br />
<li> Because of the same reason as before, this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a side note, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
====''Optimizing the distribution matrix''====<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
====''The program''====<br />
<br />
The program, written in Python with TensorFlow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is slow for large networks.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an extended version of the “Brick Algorithm”. This algorithm is an improve of the “Brick Algorithm” because it also works for a larger scale. The idea of this algorithm is based on that traffic jams only occur at a local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven. This however seems to be an unrealistic assumption overall for the real world, but also this assumption can be generalized as we will show below.<br />
<br />
==== Situation Representation ====<br />
We are again see the road network as a graph with nodes and vertices, just as in the “Brick Algorithm”. However, this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm). This graph is pictured on the right.<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb|A more complicated road graph example.]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 cars that need to go to A. If a city does not have subscript A, then 200 cars need to go to A from that city. For subscript G and J the same holds analogously. Furthermore, we assume that every black-coloured road has a single lane and has a minimal travel time of 10 minutes. Every orange road has two lanes and has a minimal travel time of 20 minutes. And every green road has three lanes and has a minimal time of 30 minutes. Then we also have red roads, which have a minimal travel time of 15 minutes and consists of a single lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm, which is described below.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes as an example. Thus, every node that is in ''t'' minutes of a goal is taken into account. We take ''t'' to be equal to 30 minutes in this example and we note that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round.<br />
<br />
==== First iteration round ====<br />
In the first iteration round we will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within a minimum of ''t'' waiting time from its goal. We will discard every node in this example that is further than ''t'' time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example we will use the “Brick Algorithm” for three zones with subscript A, G and J (note that zones do not have to be disjunctive) and we find the optimal solution for these zones. After that (still within the same iteration round) you will take back the discarded nodes and vertices and take a look at cars that travel to a goal and are outside of that zone (so for example someone from a city that does not have subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travellers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as cars arrived at that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account and so on. This iteration is stopped after ''u'' iteration rounds have passed where we will determine ''u'' based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is. However, if two goals are close to each other then it will not take the other goal into account, which might not be desirable behaviour. Therefore we say that when we calculate the local solution for a certain goal that then we also take other goals within a radius of 2*''t'' into account, so we use the “Brick Algorithm” for multiple goals at the same time. Then we discard the result of the other goals and we use only the result of the goal that we were calculating. We call this process the “ILDB+ Algorithm”.<br />
<br />
==== Motivation of assumptions ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is seems logical for some situations, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
However, in situations of a daily commute, this is often not the case. We see traffic jams occurring because of accidents, or merging traffic at interchanges. This can happen alongside any route and is often not only near a destination node. This does not mean that the ILDB algorithm approach cannot be used, it is still applicable with some modifications. The main idea of splitting route finding on a local and global level stays in place, only now extended to zones alongside routes as well. We split the complete road network in multiple overlapping zones, such that every inch of road is covered. For the ILDB algorithm to work, it is important that the zones capture traffic hotspots in full. A way we can accomplish this, is to dynamically make zones active and inactive based on the current traffic situations in such a way that jams are as much covered by a single zone while still having all roads covered by zones.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>The assumption that traffic jams going to a certain location only occur close to that location is a assumption, then this algorithm does not work without modification. However, it is unclear and hard to evaluate the performance of this extended algorithm in terms of completeness and optimality.</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine by ourselves how much iteration rounds are needed and what the value for ''t'' is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius ''t'' of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the "+" part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is run is limited). We also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius ''t'' the total driving time of all traffic does not decrease much if you continue with increasing it and what stands out is that the total driving time increases a bit sometimes for higher radius ''t'', because with a higher radius ''t'' you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm can get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is high and therefore because we only use small graphs in our final implementation we do not use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, that we call the “Auction House Algorithm” because it is similar to an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list.<br />
<br />
==== Definition List ====<br />
;Route<br />
:A set of roads that will take you from your start location to your destination.<br />
;In Advance Best<br />
:A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etc.). m is in this case some constant, which will will determine.<br />
;Capacity<br />
:The maximum amount of cars that can take a certain road each hour.<br />
;Claiming<br />
:If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road.<br />
;Neural<br />
:<code>Neural(arg 1, arg 2, ..., arg n)</code> is the output of a Neural Network with n input arguments.<br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm, again here depicted on the right.<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation as used earlier for the Brick Algorithm.]]<br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 lane: 1750 vehicles per hour </li><br />
<li> 2 lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving or waiting time of a road is easily determined. We will assume that the daily commute rush hour will have a duration of 2 hours. And that when x amount of people in the rush hour need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the driving or waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location and destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<br />
;Shortest path with road (a)<br />
:The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams into account.<br />
;Shortest path without road (b)<br />
:The total driving time that someone needs to go from its location to their destination without taking traffic jams into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.)<br />
;Own total time with road (c)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account.<br />
:Own total time without road (d)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account. Also given that the m in advance best routes might not use this road.<br />
;How much other claims (e)<br />
:How much other start location/destination combinations make a claim on this road.<br />
;Sum shortest path others with road (f)<br />
:The total sum of the shortest path with road for other start location and destination combinations (including other destinations from the same city).<br />
;Sum shortest path others without road (g)<br />
:The total sum of the shortest path without road for other start location and destination combinations.<br />
;Sum total time others with road (h)<br />
:The total sum of the own total time with road for other start location and destination combinations.<br />
;Sum total time others without road (i)<br />
:The total sum of the own total time without road for other start location and destination combinations.<br />
;How much people claim the road (j)<br />
:The total sum of all drivers that claim this road.<br />
;Capacity of a road (k)<br />
:The maximum capacity of a road.<br />
<br />
We define: <code>l = Neural(a, b, c, d, e, f, g, h, i, j, k)</code>. l is probably a value between 0 and 1 (for simplicity, because it is the output of a neural network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location and destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + ... + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unbounded m? ====<br />
A good question is why we bound m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m does not have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evaluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behaviour of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility <!-- NOTE: Capital letter O is intended spelling. -->, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behaviour?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behaviour of a solution, however don’t conclude that one of the solutions is bad because it haves different behaviour. Because the total driving time might be the same for solutions with different behaviour. To illustrate this we will show a graph of a 3D function, depicted on the right.<br />
[[File:Different_Behavior_Same_Result.png|thumb|Graph on a 3D function with an infinite number of absolute minimum values.]]<br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behaviour for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like a tetrahedron or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="Some predetermined graphs"><br />
File:Brick Algorithm - Situation Representation.png|Tetrahedron model.<br />
File:ILDB(+) Algorithm - Situation Representation.png|Complex Rectangle model.<br />
</gallery><br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is shown on the right.<br />
[[File:RandomGraph.png|300px|framed|center|Generated Total Random Graph.]]<br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward). This can be seen in the figure below.<br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get the other figure shown below.<br />
<br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="RR-Grid construction"><br />
File:USE_-_Grid.png|A square neighbor-fully-connected rectangular grid.<br />
File:USE - RRGrid.png|An RR-Grid based on the grid above.<br />
</gallery><br />
<br />
== A Small Complexity Analysis ==<br />
<br />
We have a look at how the training time increases as the amount of cities and routes increases. Unfortunately we can only run this for our very small problem, which may not say very much about training for bigger road networks.<br />
We use the tetrahedron model mentioned before in the situation representation of the Brick Algorithm, where road 2, 4 and 6 are highways with maximum speed 30 and length 280. The other roads used, 1 and 3, secondary roads with maximum speed 17 and length 354. People start from cities B, C and D and all want to arrive at goal A.<br />
<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|The tetrahedron model used for the complexity analysis.]]<br />
<br />
We test three situations, first with only city A and B and one road from B to A, then we add the city D with three routes to the goal. These routes that people can drive are from B to A directly, from B to A via D and from D to A. Similarly we extend to four cities with five routes.<br />
<br />
We train using only one species, a population size of 5 and train for only 5 generations to keep it fast to test.<br />
<br />
We run the training script five times and then take the average running time, so that the time found will not differ more than a few tenths on different runs of the complexity analysis.<br />
<br />
* Two cities (A and B) and one route: 7.3 seconds<br />
* Three cities (A, B and D) and three routes: 10.4 seconds<br />
* Full graph with four cities and nine routes: 12.7 seconds<br />
<br />
From this we could conclude that if we want to use very big road networks, we probably need to devise optimizations to keep the training tractable. But remember this is just a small test with possibly numbers too small to draw any proper conclusions from, so this conclusion is not definite at all.<br />
<br />
== Running the simulation ==<br />
[[File:Pre-2016-q3-group13-simulation.gif|thumb|333px|Simulation ([[:File:PRE16-3-group13-neat-implementation.zip|code]]) of our system compared to a conventional situation where all drivers use the common highway route. Click to run the animation.]]<br />
With the NEAT algorithm, compiling all of the approaches and thoughts into a final implementation, we run a simulation to show the results of our system. This adds to all research done above and is a proof-of-concept of the technical capabilities of this system.<br />
We use a slightly modified version if the tetrahedron model for the simulation, in an attempt to make it somewhat more realistic. It's shape is modeled after the real-life network between the Dutch cities Amsterdam (the goal at the north side), Utrecht (in the middle), The Hague and Eindhoven (south), and Leiden and Amersfoort (west and east of Utrecht, respectively). The roads connecting the large cities (Amsterdam, Utrecht, The Hague and Eindhoven) are 2-lane highways, while the other roads are 1-lane secondary roads with a lower maximum speed. Because this simulates on a low scale, the For this simulation, a random amount of cars between 20 and 25 (for demonstration purposes) leaves a city with destination Amsterdam. However, the cars sent on the left simulation and the right simulation is the same, for a fair comparison.<br />
<br />
On the left in the animation, we simulate our system, where cars are distributed based on the output the evolved neural network provides. On the right the same situation in terms of initial car distribution is shown, but cars always choose the fastest highway route from their city to Amsterdam. As clearly visible in the animation, our system outperforms the situation on the right in terms of total duration before all cars reach the destination. Our system here is based on the NEAT algorithm approach primarily, which is evolved on variable starting distribution of cars on this fixed road network. The generation of this network took less than 15 minutes, while the evaluation once generated was almost instant. For this small example, the time requirements are met, as well as the main requirement as we see improvement of traffic flow in the new situation. The cost requirement is harder to check on this small scale, but since overall travel times are shorter, it seems promising.<br />
<br />
We know that this is still far from a real-life implementation as a navigation system ready to use in consumer cars. For that, we have laid the groundwork on how to expand this simple situation to larger areas with the sections above. With that, the conducted user research will start to show its results in the implementation directly. Also, the scaling of the running time also becomes important, as covered in the previous section. We have shown that it is not easy to make a finished product for use in society, but we hope that with all the research we have done, the best we can, we have shown ways on how it is possible to extend our basic simulation to a more realistic implementation, in terms of scaling and letting the user interests into the design. We are proud to show that we can apply self-learning techniques in an unconventional way, with using the NEAT algorithm approach, into this kind of problem: something we have not seen applied before.<br />
<br />
==References==<br />
''For some reason, references are not showing up when not logged in to the wiki. They are visible when looking at the source of this page.''<br />
<br />
<references/></div>Fri, 14 Apr 2017 10:12:23 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39350
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39350<p>S143110: /* Traffic data */</p>
<hr />
<div><table cellspacing=1<br />
style="float:right; margin: 0 0 0 1em; border: 1px solid #a2a9b1;background-color: #f9f9f9; padding: .7em;"><br />
<tr><td colspan="2" style="border-color: #a2a9b1; padding: 0.2em 0.4em; text-align: center;"><b>Authors</b></td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">David Dekker</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0936100</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Haico Dorenbos</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0959563</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Jori van Falier</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0910959</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Tim Muyrers</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0907928</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Thomas Schouten</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0938927</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Sten Wessel</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0941508</td></tr><br />
</table><br />
<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self-learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self-driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self-Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
<div style="clear: left; float:left; margin-bottom:0.5em; padding:0.5em 1.4em 0.8em 0;" class="toclimit-2">__TOC__</div><br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people commute to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these:<br />
<br />
# The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
# The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but is also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that cannot be predicted. Because of all these problems we will be using self-learning software to deal with these problems. This program will try to compensate for all the unknowns.<br />
<br />
== Objectives ==<br />
To make a solution for this problem fit in the time we have for this project, we can only focus on some of these aspects. To account for all of them, this would take a lot more years of research and manpower to create this ground-breaking system. Therefore, we will only focus on a simpler situation with one goal and a finite number of cars leaving for that goal. Hence, we only focus at traffic as a ''snapshot'' of time and only for one goal. We think the results of this can be reasonably generalized to a situation with multiple goals and a continuous flow of traffic. The reason for that is that we can extend the behaviour we create to apply to the full-scale problem, by evaluating the situation repeatedly (from ''snapshot'' to continuous) and doing so sequentially for every goal (from single to multi goal). We will elaborate on how this problem scales in the section about complexity analysis: [[#A Small Complexity Analysis]]. For now, we state the objectives for this problem explicitly. We make a distinction between two separate goals.<br />
<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams. This goal is to be satisfied for individual users.<br />
<br />
=== Maximize overall utility ===<br />
Not only minimized travel time individually is important, but also society interest overall. We capture a variety of aspects of interest for society under the umbrella term ''utility''.<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time overall, for all users collectively.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprise ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== Requirements ==<br />
Algorithms are measured in completeness, optimality, time complexity and space complexity. Furthermore the costs of our algorithm and system is also important. So completeness, optimality, time complexity, space complexity and costs are included in the requirements. Because we have combined the completeness and optimality in our main requirement, we also have included time complexity in some of our requirement (we don’t use big O notation in the requirements, because customers are interested in the running time not the big O notation, however we do explain why we meet this requirement by using the big O notation) and last of all cost and space complexity is also combined in the cost requirement (space complexity is included in this, because if you need more memory then you also need to buy more memory or computers).<br />
<br />
=== Main Requirement ===<br />
''For every situation our algorithm has to return for everyone the optimal driving route, such that the sum of the total driving time of everyone is minimal.''<br />
<br />
Note that with ''minimal'' we mean that the sum of the total driving time cannot become any lower than this. For the algorithm implementation, we focus on fulfilling this requirement, because we can in within project bounds a simulation where the fulfilment of this requirement can be measured. However, we will provide a way to extend the algorithm to take the other objectives into account as well.<br />
<br />
=== Cost Requirement ===<br />
''The variable cost per day added with the fixed cost divided by the deprecation period (measured in days) may not exceed €8,96 times the total reduced travel time (also measured in hours) for everyone per day.''<br />
<br />
We explain why we choose this amount. The minimum wage in the Netherlands is €8,96 per hour for people over 23 years. Suppose our system variable cost per hour added with the fixed cost divided by the deprecation period is equal to €7,00 and only one person travel time is reduced by an hour every day and nobody else's travel time is changed, then we have a societal profit of at least €1,96 because that person can start 1 hour earlier with his work and his wage is at least €8,96 per hour. So at least 8,96 – 7,00 = €1,96 societal profit.<br />
<br />
==== Example ====<br />
We want to test if this requirement is satisfied. So let's take a look at a certain algorithm that has a variable cost of €500 per day (electricity costs, hiring employees to maintain the code, etcetera) and the fixed costs are €100.000 (development cost, buying a lot of computers, etcetera). Suppose that the deprecation period of this algorithm is 10 years (after 10 years a new algorithm is found to solve the problem better). Then the variable cost per day added with the fixed cost divided by the deprecation period is equal to 500 + 100.000 / (10 * 365) = 527,40. Now suppose for 100.000 people the driving time on average is decreased by 15 minutes every day and nothing else is changed in driving time. Then we have a societal revenue of 100.000 * 0,25 * 8,96 = €224.000,00. 527,40 <= 224000,00, thus in this case our system satisfies this requirement and has a societal benefit of 224.000,00 – 527,40 = €223.472,60.<br />
<br />
=== After Running Time Requirement ===<br />
''The after running time may be no more than 5 seconds.''<br />
<br />
The after running time is the running time when a customer wants the best route to his destination. The calculation of the best route might only take 5 seconds, this is because customers get impatient if it takes longer than 5 seconds.<br />
<br />
=== Pre Running Time Requirement ===<br />
''The pre running time may be no more than 7 days.''<br />
<br />
This is the running time that the artificial intelligent program takes to prepare the system for being used by drivers (for example a NEAT algorithm, introduced later, that optimizes a neural network is included in this running time). The goal is to optimize for a specific road network, where the traffic distributions are still variable. Hence, this running time may be high, because program only has to be run when the roadmap changes and once in beginning. So users do not have to wait on this algorithm, because this algorithm can run parallel with the calculation of the best route for the users. This is because changes in a road network are known in advance, so when generating a new neural network for the new road situation, the calculation of best routes for users is still for the old road network.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices currently available. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to the current situation only, instead of trying to avoid certain situations beforehand.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will reroute all traffic over this faster route. Right now, this is not really a problem since not all users will follow that advice. However, in the long term (consider for instance self-driving cars), this approach will only create a continuous shifting of recommended routes as a solely reactive system. A self-learning system however find a stable solution, since it can predict traffic jams that haven't even happened yet, and account for the effects of rerouting traffic. It can reroute the user many kilometres before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
[[File:Idea visualize prof.jpg|300 px|thumb|Visualization of the rerouting process.]]<br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously with developing the algorithm. The group of researchers will be split where four members of the team will work on making the self-learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representative for a final implementation as possible.<br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external parties to use and understand the way the system operates, and what conclusions it presents.<br />
<br />
== User and Society ==<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish and less considerate to other drivers. As long as people are driving their car (instead of autonomous), we can investigate further if and how we want to incorporate this.<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey, with the goal of decreasing overall travel time. When looking at literature we find something called ''risk aversion''<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favourable bet (overall travel time decrease) when it means they can directly lose (more travel time)<ref>YouTube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== Traffic research ==<br />
<br />
=== What is a traffic jam ===<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can identify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: unnecessary braking resulting in so called ''phantom jams''<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans) and probably also when cars become fully autonomous, we do see however that accidents can be massively decreased when we decrease the amount of cars driving at a certain road. The busier it is, the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the risk factor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
[[File:FormulaUsed1PNG.png|380 px|thumb|Relation between traffic volume and road crashes.]]<br />
<br />
Too many drivers on the road also causes direct traffic jams through unnecessary braking. When a car in front brakes, the one behind it brakes too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill. This phenomenon is called a ''phantom jam'' which seem to occur and disappear suddenly on crowded roads.<br />
<br />
Road works cannot really be avoided. But if the system is connected to a database, it should be able to take this into account. We can see and handle planned roadwork the same as a (however temporary) change in the road network.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self-learning system should be able to handle occasional accidents and reroute accordingly.<br />
<br />
In conclusion, to make sure the navigational system works in real life you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focused on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be beneficial. Moving on, we abstract away the details of how unpredictable jams occur and disappear and focus training the system based on experiments where these traffic jam situations just happen.<br />
<br />
=== Traffic data ===<br />
<br />
For our system to work, traffic data has to be acquired for realistic large scale data, to help train the system. As our system is still in an early development stage, the data found will not be used as it requires some parsing to get it in a format where the system can work with. As this is too much time consuming, we will focus on a fictional road network inspired on a rush hour commute situation. However, it will show that this data is obtainable and could be used for our system. The data has been acquired from ''Nationale Databank Wegverkeersgegevens'' (NDW) <ref> Nationale Databank Wegverkeersgegevens (NDW), Retrieved March, 2017 from http://www.ndw.nu/nl/ </ref>. Open historic data has been applied for and obtained. It gives information about the traffic density and speed of cars on 27000 locations in the Netherlands. This information can be used for our program to determine how many people travel these roads and when a traffic jam occurs. The data is delivered in the DATEX II form, which is a European standard for traffic data <ref> T. Delissen, Nederlands Profiel DATEX II, 2015-2a, Retrieved April, 2017 from http://www.ndw.nu/downloaddocument/039ee3e1098f37830e26df848f6202ca/Nederlands%20Profiel%20DATEX-II%202015-2a%20(14-01-2016).pdf </ref>. By doing further research about this data, the data can be read by our program and used to train the system. The data base also has current traffic data, that can be used when our system becomes operational. By knowing about the traffic on these 27000 points, the system can get a good understanding of the current traffic situation, and therefore make better decisions. All of this data has to be implemented in the system when it is further developed and can be beneficial to the functioning of our system.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by ''Kennisinstituut voor Mobiliteitsbeleid'' (KiM) from the Ministry of Infrastructure and the Environment in the Netherlands<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3924 respondents from car drivers of the ''Mobiliteitspanel Nederland'' (MPN). <br />
<br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought specifically for that it provides traffic information. Users get most of the traffic information from the internet or the radio and during travel also information panels above the road (DRIPs). From the questions in the KiM research follows that ease-of-use and quality of information are important reasons for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long-distance travel. The system is also used for navigating to points of interest. The most used points of interest are parking lots (75%), gas stations (49%) and places of entertainment (38%). Furthermore, navigation systems are used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of importance follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route themselves, heard of or have seen a traffic jam, or because of road works. Because our system will work optimal when everyone is using the system, and following the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can prove that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors in choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Forms is used to host the questionnaire, because it has all features (kind of questions and setup) required for this research. The questionnaire can be found [https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform here in English] or [https://docs.google.com/forms/d/e/1FAIpQLScx_kPjyN1QNCiAW3xYT-Q77lZks4vDtjyr6YV5x0momVwqVw/viewform here in Dutch], which are identical in content. Please note that by choosing certain answers it is possible that the questionnaire forwards you past irrelevant questions and thus skipping certain sections. In order to see the full questionnaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone receiving the questionnaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non-frequent drivers. While the system is designed for non-self-driving cars it would also work on autonomous cars. The system will first be tested and deployed in the Netherlands and thus this is also the target group for distributing the questionnaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigation system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCNS is designed to be applicable to society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing whatever the system proposes should still be faster than not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistle-blower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this they are taken to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily life. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account, and so on. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom-made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know a reasonable amount of time in advance when users are going to get on the road. This section asks for different occasions (different kinds of destinations) the time the user would be able and willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigation system. In this section the participants are asked about possible approaches the system could take regarding traffic and travel time distribution.<br />
<br />
A participant is free to not answer a question if they do not want to answer, without having to provide a reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representative participants for the target group (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well received with up to 95% of the participants showing interest in the product<br />
* 72.4% of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) versus sending the users on secondary roads)<br />
* Point-based system or reward for driving longer routes is preferred among the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travel ahead of time and inform the navigation system of those plans.<br />
* 88.7% 91.9% and 95.1% of participants indicate to be interested in this goal in the form of this kind of product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until April 7). Seven people responded to the English version and 56 filled in the Dutch version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent.<br />
<br />
==== Study outcome section 1: Introduction questions ====<br />
The introduction questions focused on introducing the product and for us, the researchers, to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answers are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9%, respectively.<br />
<br />
So what does this mean? To start, it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter-device communication. As for environmental nuisances it is hard to say how the device would be optimized at this point since the answers are divided among all four categories. <br />
<br />
==== Study outcome section 2: Willingness to follow the system ====<br />
<br />
The second question category tests whether or not participants are willing to follow the system in four different situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
<br />
==== Study outcome section 3: Fairness ====<br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally only one would be the best. The preferred options are a point-based system that will send you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not further specified how this would take shape) sounds appealing to a large part of the participants in both sections respectively scoring 81.9% and 73.8% on ''very good'' or ''good'' compared to the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made).<br />
<br />
==== Study outcome section 4: Privacy ====<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section about privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
==== Study outcome section 5: System interaction ====<br />
This section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people, this loss in performance will most likely be compensated by the extra information the system is able to receive from more users. The responses show that 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be investigated further, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
==== Study outcome section 6: Finalizing questions ====<br />
In the beginning of the questionnaire participants were asked on their view on the importance of environmental factors that influence the functioning of a navigation system. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take air and noise pollution into account or not, as well as having an option to change this setting, or no option.<br />
64% would like standard incorporation of air pollution into the system, with 72% preferring to have an option to change the setting the system is on default. <br />
For noise pollution these values are slightly lower with 60% wanting a default incorporation of noise into the system, and a slightly higher amount of participants (76%) preferring to have an option to change noise incorporation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency, etc.), but will also need a incorporation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
Making system interaction easier is a strong aim in future planning of travel. For this a section of the questionnaire is designed specifically to understand how drivers would like to do this. The aim is to design one to two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants, which is a mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre-planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car as well an option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regarding daily commutes as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
[[File:Goede Graph user research.JPG|666 px|framed|center|Research results on the time ahead users would inform the system of their travel plans for a variety of purposes.]]<br />
As we see in the graph above, on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interested in the system with all possible goals achieving over 50% of interested participants. The three most popular goals are "The overall travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overall travel time is lower", "The overall travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interested in this goal or product. What this means for the design is mostly that the goal can be customized as long as users do not have an increased travel time themselves. <br />
<br />
===='''Final conclusion'''====<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in-depth interviews, the information collected through this limited resources survey gives a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user-influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user-friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of effects that our system introduces could not be specified. For example, the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. The same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things cannot be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed, more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the whole population, although the target group has been reached. Furthermore, the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point-based reward systems are favoured by the users and have to be further researched, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things cannot be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
We start by introducing a very easy model for our problem. It is not very elaborated and therefore the model does not really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etc. is taken into account, due to time constraints). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is static and for now only looks at a distribution of cars at the beginning of the simulation. Moreover, we assume that everyone uses our clever navigation system and that everyone is obedient to the system and therefore they will always follow their directions. Because we first only focus on this simple road network, we make these assumptions to make sure we can measure in a simple and effective way the effects our navigation system has.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like the graph pictured on the right. [[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation used to run the algorithms on.]]<br />
The double-circled vertex is the destination, which means that the drivers from cities B, C and D have to go to this location (we do not take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each bidirectional road there are two edges, one for each direction (the roads connected directly to A however are unidirectional, because if you are at destination A then you do not have to drive away from it). Every driveway in this graph is indicated with a number and every city is indicated with a capital letter.<br />
<br />
==== Data about roads and cities ====<br />
To model a daily commute situation, we will assume that 30 000 people live in city B who all need to go to A. We will furthermore assume that 20 000 people live in city C, who all need to go to city A. Moreover we will assume that 10 000 people live in city D, who all need to go to city A. This is however only for demonstration purposes, as we want our algorithm to be independent on the starting distribution of cars. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower than it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are high-capacity highways, therefore they can for instance have have the function 0.00025x + 15 (in the case when you need at least 15 minutes to drive over that certain road).</li><br />
<li>Roads 2, 4, 5, 6 and 7 are short secondary roads with a lower capacity (the ''N-wegen'' in Dutch) and these roads can for instance have the function 0.0005x + 5 (you need at least 5 minutes to drive over that certain road).</li><br />
<li>Roads 8 en 9 are also secondary roads, but shorter and with an even lower capacity and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm no cars are present on the roads yet. Suppose we have some driver that lives in city C, which drives over the driveways 6, 5 and 1 (in this order to arrive at his destination A. If this happens, then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10 000 people have driven over road 1 and they have waited 17.5 minutes then the cost value for this road is 175 000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving round” a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This process will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit this process for 10 driving rounds, for example (you can take a different number than 10). If after 10 driving rounds not everyone has arrived at destination A, then it will get a high cost value (and therefore this is automatically a bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was to penalize this behaviour, by increasing the cost value for those solutions proportional to the amount of cars that have not reached their destination. This cost value has to be strictly higher than a cost value of a solution for which it is possible for everyone to arrive within 10 driving rounds to destination A. If we define our cost value this way, then a solution for which in 10 driving rounds not everyone gets to their destination is worse than a solution in which everyone gets to their destination instead.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it does not need neural Nntworks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also it can be used for evaluation at different times. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are quite constraining the real-life situation (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time to evaluate. Especially because this algorithm need to determine distributions for every combination of driveway, destination and different parts of the day. If you look for example at an area the size of Germany and the Netherlands, then there is a huge amount of driveways and a huge amount of destinations. This scales badly.</li><br />
<li> Because of the same reason as before, this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a side note, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
====''Optimizing the distribution matrix''====<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
====''The program''====<br />
<br />
The program, written in Python with TensorFlow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is slow for large networks.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an extended version of the “Brick Algorithm”. This algorithm is an improve of the “Brick Algorithm” because it also works for a larger scale. The idea of this algorithm is based on that traffic jams only occur at a local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven. This however seems to be an unrealistic assumption overall for the real world, but also this assumption can be generalized as we will show below.<br />
<br />
==== Situation Representation ====<br />
We are again see the road network as a graph with nodes and vertices, just as in the “Brick Algorithm”. However, this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm). This graph is pictured on the right.<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb|A more complicated road graph example.]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 cars that need to go to A. If a city does not have subscript A, then 200 cars need to go to A from that city. For subscript G and J the same holds analogously. Furthermore, we assume that every black-coloured road has a single lane and has a minimal travel time of 10 minutes. Every orange road has two lanes and has a minimal travel time of 20 minutes. And every green road has three lanes and has a minimal time of 30 minutes. Then we also have red roads, which have a minimal travel time of 15 minutes and consists of a single lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm, which is described below.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes as an example. Thus, every node that is in ''t'' minutes of a goal is taken into account. We take ''t'' to be equal to 30 minutes in this example and we note that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round.<br />
<br />
==== First iteration round ====<br />
In the first iteration round we will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within a minimum of ''t'' waiting time from its goal. We will discard every node in this example that is further than ''t'' time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example we will use the “Brick Algorithm” for three zones with subscript A, G and J (note that zones do not have to be disjunctive) and we find the optimal solution for these zones. After that (still within the same iteration round) you will take back the discarded nodes and vertices and take a look at cars that travel to a goal and are outside of that zone (so for example someone from a city that does not have subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travellers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as cars arrived at that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account and so on. This iteration is stopped after ''u'' iteration rounds have passed where we will determine ''u'' based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is. However, if two goals are close to each other then it will not take the other goal into account, which might not be desirable behaviour. Therefore we say that when we calculate the local solution for a certain goal that then we also take other goals within a radius of 2*''t'' into account, so we use the “Brick Algorithm” for multiple goals at the same time. Then we discard the result of the other goals and we use only the result of the goal that we were calculating. We call this process the “ILDB+ Algorithm”.<br />
<br />
==== Motivation of assumptions ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is seems logical for some situations, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
However, in situations of a daily commute, this is often not the case. We see traffic jams occurring because of accidents, or merging traffic at interchanges. This can happen alongside any route and is often not only near a destination node. This does not mean that the ILDB algorithm approach cannot be used, it is still applicable with some modifications. The main idea of splitting route finding on a local and global level stays in place, only now extended to zones alongside routes as well. We split the complete road network in multiple overlapping zones, such that every inch of road is covered. For the ILDB algorithm to work, it is important that the zones capture traffic hotspots in full. A way we can accomplish this, is to dynamically make zones active and inactive based on the current traffic situations in such a way that jams are as much covered by a single zone while still having all roads covered by zones.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>The assumption that traffic jams going to a certain location only occur close to that location is a assumption, then this algorithm does not work without modification. However, it is unclear and hard to evaluate the performance of this extended algorithm in terms of completeness and optimality.</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine by ourselves how much iteration rounds are needed and what the value for ''t'' is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius ''t'' of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the "+" part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is run is limited). We also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius ''t'' the total driving time of all traffic does not decrease much if you continue with increasing it and what stands out is that the total driving time increases a bit sometimes for higher radius ''t'', because with a higher radius ''t'' you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm can get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is high and therefore because we only use small graphs in our final implementation we do not use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, that we call the “Auction House Algorithm” because it is similar to an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list.<br />
<br />
==== Definition List ====<br />
;Route<br />
:A set of roads that will take you from your start location to your destination.<br />
;In Advance Best<br />
:A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etc.). m is in this case some constant, which will will determine.<br />
;Capacity<br />
:The maximum amount of cars that can take a certain road each hour.<br />
;Claiming<br />
:If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road.<br />
;Neural<br />
:<code>Neural(arg 1, arg 2, ..., arg n)</code> is the output of a Neural Network with n input arguments.<br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm, again here depicted on the right.<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation as used earlier for the Brick Algorithm.]]<br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 lane: 1750 vehicles per hour </li><br />
<li> 2 lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving or waiting time of a road is easily determined. We will assume that the daily commute rush hour will have a duration of 2 hours. And that when x amount of people in the rush hour need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the driving or waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location and destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<br />
;Shortest path with road (a)<br />
:The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams into account.<br />
;Shortest path without road (b)<br />
:The total driving time that someone needs to go from its location to their destination without taking traffic jams into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.)<br />
;Own total time with road (c)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account.<br />
:Own total time without road (d)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account. Also given that the m in advance best routes might not use this road.<br />
;How much other claims (e)<br />
:How much other start location/destination combinations make a claim on this road.<br />
;Sum shortest path others with road (f)<br />
:The total sum of the shortest path with road for other start location and destination combinations (including other destinations from the same city).<br />
;Sum shortest path others without road (g)<br />
:The total sum of the shortest path without road for other start location and destination combinations.<br />
;Sum total time others with road (h)<br />
:The total sum of the own total time with road for other start location and destination combinations.<br />
;Sum total time others without road (i)<br />
:The total sum of the own total time without road for other start location and destination combinations.<br />
;How much people claim the road (j)<br />
:The total sum of all drivers that claim this road.<br />
;Capacity of a road (k)<br />
:The maximum capacity of a road.<br />
<br />
We define: <code>l = Neural(a, b, c, d, e, f, g, h, i, j, k)</code>. l is probably a value between 0 and 1 (for simplicity, because it is the output of a neural network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location and destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + ... + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unbounded m? ====<br />
A good question is why we bound m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m does not have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evaluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behaviour of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility <!-- NOTE: Capital letter O is intended spelling. -->, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behaviour?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behaviour of a solution, however don’t conclude that one of the solutions is bad because it haves different behaviour. Because the total driving time might be the same for solutions with different behaviour. To illustrate this we will show a graph of a 3D function, depicted on the right.<br />
[[File:Different_Behavior_Same_Result.png|thumb|Graph on a 3D function with an infinite number of absolute minimum values.]]<br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behaviour for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like a tetrahedron or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="Some predetermined graphs"><br />
File:Brick Algorithm - Situation Representation.png|Tetrahedron model.<br />
File:ILDB(+) Algorithm - Situation Representation.png|Complex Rectangle model.<br />
</gallery><br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is shown on the right.<br />
[[File:RandomGraph.png|300px|framed|center|Generated Total Random Graph.]]<br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward). This can be seen in the figure below.<br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get the other figure shown below.<br />
<br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="RR-Grid construction"><br />
File:USE_-_Grid.png|A square neighbor-fully-connected rectangular grid.<br />
File:USE - RRGrid.png|An RR-Grid based on the grid above.<br />
</gallery><br />
<br />
== A Small Complexity Analysis ==<br />
<br />
We have a look at how the training time increases as the amount of cities and routes increases. Unfortunately we can only run this for our very small problem, which may not say very much about training for bigger road networks.<br />
We use the tetrahedron model mentioned before in the situation representation of the Brick Algorithm, where road 2, 4 and 6 are highways with maximum speed 30 and length 280. The other roads used, 1 and 3, secondary roads with maximum speed 17 and length 354. People start from cities B, C and D and all want to arrive at goal A.<br />
<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|The tetrahedron model used for the complexity analysis.]]<br />
<br />
We test three situations, first with only city A and B and one road from B to A, then we add the city D with three routes to the goal. These routes that people can drive are from B to A directly, from B to A via D and from D to A. Similarly we extend to four cities with five routes.<br />
<br />
We train using only one species, a population size of 5 and train for only 5 generations to keep it fast to test.<br />
<br />
We run the training script five times and then take the average running time, so that the time found will not differ more than a few tenths on different runs of the complexity analysis.<br />
<br />
* Two cities (A and B) and one route: 7.3 seconds<br />
* Three cities (A, B and D) and three routes: 10.4 seconds<br />
* Full graph with four cities and nine routes: 12.7 seconds<br />
<br />
From this we could conclude that if we want to use very big road networks, we probably need to devise optimizations to keep the training tractable. But remember this is just a small test with possibly numbers too small to draw any proper conclusions from, so this conclusion is not definite at all.<br />
<br />
== Running the simulation ==<br />
[[File:Pre-2016-q3-group13-simulation.gif|thumb|333px|Simulation ([[:File:PRE16-3-group13-neat-implementation.zip|code]]) of our system compared to a conventional situation where all drivers use the common highway route. Click to run the animation.]]<br />
With the NEAT algorithm, compiling all of the approaches and thoughts into a final implementation, we run a simulation to show the results of our system. This adds to all research done above and is a proof-of-concept of the technical capabilities of this system.<br />
We use a slightly modified version if the tetrahedron model for the simulation, in an attempt to make it somewhat more realistic. It's shape is modeled after the real-life network between the Dutch cities Amsterdam (the goal at the north side), Utrecht (in the middle), The Hague and Eindhoven (south), and Leiden and Amersfoort (west and east of Utrecht, respectively). The roads connecting the large cities (Amsterdam, Utrecht, The Hague and Eindhoven) are 2-lane highways, while the other roads are 1-lane secondary roads with a lower maximum speed. Because this simulates on a low scale, the For this simulation, a random amount of cars between 20 and 25 (for demonstration purposes) leaves a city with destination Amsterdam. However, the cars sent on the left simulation and the right simulation is the same, for a fair comparison.<br />
<br />
On the left in the animation, we simulate our system, where cars are distributed based on the output the evolved neural network provides. On the right the same situation in terms of initial car distribution is shown, but cars always choose the fastest highway route from their city to Amsterdam. As clearly visible in the animation, our system outperforms the situation on the right in terms of total duration before all cars reach the destination. Our system here is based on the NEAT algorithm approach primarily, which is evolved on variable starting distribution of cars on this fixed road network. The generation of this network took less than 15 minutes, while the evaluation once generated was almost instant. For this small example, the time requirements are met, as well as the main requirement as we see improvement of traffic flow in the new situation. The cost requirement is harder to check on this small scale, but since overall travel times are shorter, it seems promising.<br />
<br />
We know that this is still far from a real-life implementation as a navigation system ready to use in consumer cars. For that, we have laid the groundwork on how to expand this simple situation to larger areas with the sections above. With that, the conducted user research will start to show its results in the implementation directly. Also, the scaling of the running time also becomes important, as covered in the previous section. We have shown that it is not easy to make a finished product for use in society, but we hope that with all the research we have done, the best we can, we have shown ways on how it is possible to extend our basic simulation to a more realistic implementation, in terms of scaling and letting the user interests into the design. We are proud to show that we can apply self-learning techniques in an unconventional way, with using the NEAT algorithm approach, into this kind of problem: something we have not seen applied before.<br />
<br />
==References==<br />
''For some reason, references are not showing up when not logged in to the wiki. They are visible when looking at the source of this page.''<br />
<br />
<references/></div>Fri, 14 Apr 2017 10:08:01 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39349
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39349<p>S143110: /* Objectives */</p>
<hr />
<div><table cellspacing=1<br />
style="float:right; margin: 0 0 0 1em; border: 1px solid #a2a9b1;background-color: #f9f9f9; padding: .7em;"><br />
<tr><td colspan="2" style="border-color: #a2a9b1; padding: 0.2em 0.4em; text-align: center;"><b>Authors</b></td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">David Dekker</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0936100</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Haico Dorenbos</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0959563</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Jori van Falier</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0910959</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Tim Muyrers</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0907928</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Thomas Schouten</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0938927</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Sten Wessel</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0941508</td></tr><br />
</table><br />
<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self-learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self-driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self-Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
<div style="clear: left; float:left; margin-bottom:0.5em; padding:0.5em 1.4em 0.8em 0;" class="toclimit-2">__TOC__</div><br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people commute to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these:<br />
<br />
# The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
# The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but is also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that cannot be predicted. Because of all these problems we will be using self-learning software to deal with these problems. This program will try to compensate for all the unknowns.<br />
<br />
== Objectives ==<br />
To make a solution for this problem fit in the time we have for this project, we can only focus on some of these aspects. To account for all of them, this would take a lot more years of research and manpower to create this ground-breaking system. Therefore, we will only focus on a simpler situation with one goal and a finite number of cars leaving for that goal. Hence, we only focus at traffic as a ''snapshot'' of time and only for one goal. We think the results of this can be reasonably generalized to a situation with multiple goals and a continuous flow of traffic. The reason for that is that we can extend the behaviour we create to apply to the full-scale problem, by evaluating the situation repeatedly (from ''snapshot'' to continuous) and doing so sequentially for every goal (from single to multi goal). We will elaborate on how this problem scales in the section about complexity analysis: [[#A Small Complexity Analysis]]. For now, we state the objectives for this problem explicitly. We make a distinction between two separate goals.<br />
<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams. This goal is to be satisfied for individual users.<br />
<br />
=== Maximize overall utility ===<br />
Not only minimized travel time individually is important, but also society interest overall. We capture a variety of aspects of interest for society under the umbrella term ''utility''.<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time overall, for all users collectively.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprise ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== Requirements ==<br />
Algorithms are measured in completeness, optimality, time complexity and space complexity. Furthermore the costs of our algorithm and system is also important. So completeness, optimality, time complexity, space complexity and costs are included in the requirements. Because we have combined the completeness and optimality in our main requirement, we also have included time complexity in some of our requirement (we don’t use big O notation in the requirements, because customers are interested in the running time not the big O notation, however we do explain why we meet this requirement by using the big O notation) and last of all cost and space complexity is also combined in the cost requirement (space complexity is included in this, because if you need more memory then you also need to buy more memory or computers).<br />
<br />
=== Main Requirement ===<br />
''For every situation our algorithm has to return for everyone the optimal driving route, such that the sum of the total driving time of everyone is minimal.''<br />
<br />
Note that with ''minimal'' we mean that the sum of the total driving time cannot become any lower than this. For the algorithm implementation, we focus on fulfilling this requirement, because we can in within project bounds a simulation where the fulfilment of this requirement can be measured. However, we will provide a way to extend the algorithm to take the other objectives into account as well.<br />
<br />
=== Cost Requirement ===<br />
''The variable cost per day added with the fixed cost divided by the deprecation period (measured in days) may not exceed €8,96 times the total reduced travel time (also measured in hours) for everyone per day.''<br />
<br />
We explain why we choose this amount. The minimum wage in the Netherlands is €8,96 per hour for people over 23 years. Suppose our system variable cost per hour added with the fixed cost divided by the deprecation period is equal to €7,00 and only one person travel time is reduced by an hour every day and nobody else's travel time is changed, then we have a societal profit of at least €1,96 because that person can start 1 hour earlier with his work and his wage is at least €8,96 per hour. So at least 8,96 – 7,00 = €1,96 societal profit.<br />
<br />
==== Example ====<br />
We want to test if this requirement is satisfied. So let's take a look at a certain algorithm that has a variable cost of €500 per day (electricity costs, hiring employees to maintain the code, etcetera) and the fixed costs are €100.000 (development cost, buying a lot of computers, etcetera). Suppose that the deprecation period of this algorithm is 10 years (after 10 years a new algorithm is found to solve the problem better). Then the variable cost per day added with the fixed cost divided by the deprecation period is equal to 500 + 100.000 / (10 * 365) = 527,40. Now suppose for 100.000 people the driving time on average is decreased by 15 minutes every day and nothing else is changed in driving time. Then we have a societal revenue of 100.000 * 0,25 * 8,96 = €224.000,00. 527,40 <= 224000,00, thus in this case our system satisfies this requirement and has a societal benefit of 224.000,00 – 527,40 = €223.472,60.<br />
<br />
=== After Running Time Requirement ===<br />
''The after running time may be no more than 5 seconds.''<br />
<br />
The after running time is the running time when a customer wants the best route to his destination. The calculation of the best route might only take 5 seconds, this is because customers get impatient if it takes longer than 5 seconds.<br />
<br />
=== Pre Running Time Requirement ===<br />
''The pre running time may be no more than 7 days.''<br />
<br />
This is the running time that the artificial intelligent program takes to prepare the system for being used by drivers (for example a NEAT algorithm, introduced later, that optimizes a neural network is included in this running time). The goal is to optimize for a specific road network, where the traffic distributions are still variable. Hence, this running time may be high, because program only has to be run when the roadmap changes and once in beginning. So users do not have to wait on this algorithm, because this algorithm can run parallel with the calculation of the best route for the users. This is because changes in a road network are known in advance, so when generating a new neural network for the new road situation, the calculation of best routes for users is still for the old road network.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices currently available. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to the current situation only, instead of trying to avoid certain situations beforehand.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will reroute all traffic over this faster route. Right now, this is not really a problem since not all users will follow that advice. However, in the long term (consider for instance self-driving cars), this approach will only create a continuous shifting of recommended routes as a solely reactive system. A self-learning system however find a stable solution, since it can predict traffic jams that haven't even happened yet, and account for the effects of rerouting traffic. It can reroute the user many kilometres before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
[[File:Idea visualize prof.jpg|300 px|thumb|Visualization of the rerouting process.]]<br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously with developing the algorithm. The group of researchers will be split where four members of the team will work on making the self-learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representative for a final implementation as possible.<br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external parties to use and understand the way the system operates, and what conclusions it presents.<br />
<br />
== User and Society ==<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish and less considerate to other drivers. As long as people are driving their car (instead of autonomous), we can investigate further if and how we want to incorporate this.<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey, with the goal of decreasing overall travel time. When looking at literature we find something called ''risk aversion''<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favourable bet (overall travel time decrease) when it means they can directly lose (more travel time)<ref>YouTube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== Traffic research ==<br />
<br />
=== What is a traffic jam ===<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can identify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: unnecessary braking resulting in so called ''phantom jams''<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans) and probably also when cars become fully autonomous, we do see however that accidents can be massively decreased when we decrease the amount of cars driving at a certain road. The busier it is, the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the risk factor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
[[File:FormulaUsed1PNG.png|380 px|thumb|Relation between traffic volume and road crashes.]]<br />
<br />
Too many drivers on the road also causes direct traffic jams through unnecessary braking. When a car in front brakes, the one behind it brakes too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill. This phenomenon is called a ''phantom jam'' which seem to occur and disappear suddenly on crowded roads.<br />
<br />
Road works cannot really be avoided. But if the system is connected to a database, it should be able to take this into account. We can see and handle planned roadwork the same as a (however temporary) change in the road network.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self-learning system should be able to handle occasional accidents and reroute accordingly.<br />
<br />
In conclusion, to make sure the navigational system works in real life you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focused on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be beneficial. Moving on, we abstract away the details of how unpredictable jams occur and disappear and focus training the system based on experiments where these traffic jam situations just happen.<br />
<br />
=== Traffic data ===<br />
<br />
For our system to work, traffic data has to be acquired for realistic large scale data, to help train the system. As our system is still in an early development stage, the data found will not be used as it requires some parsing to get it in a format where the system can work with. As this is too much time consuming, we will focus on a fictional road network inspired on a rush hour commute situation. However, it will show that this data is obtainable and could be used for our system. The data has been acquired from ''Nationale Databank Wegverkeersgegevens'' (NDW) <ref> Nationale Databank Wegverkeersgegevens (NDW), Retrieved March, 2017 from http://www.ndw.nu/nl/ </ref>. Open historic data has been applied for and obtained. It gives information about the traffic density and speed of cars on 27000 locations in the Netherlands. This information can be used for our program to determine how many people travel these roads and when a traffic jam occurs. The data is delivered in the DATEX II form, which is a European standard for traffic data <ref> T. Delissen, Nederlands Profiel DATEX II, 2015-2a, Retrieved April, 2017 from http://www.ndw.nu/downloaddocument/039ee3e1098f37830e26df848f6202ca/Nederlands%20Profiel%20DATEX-II%202015-2a%20(14-01-2016).pdf </ref>. By doing further research about this data, the data can be read by our program and used to train the system. The data base also has current traffic data, that can be used when our system becomes operational. By knowing about the traffic on these 25000 points, the system can get a good understanding of the current traffic situation, and therefore make better decisions. All of this data has to be implemented in the system when it is further developed and can be beneficial to the functioning of our system.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by ''Kennisinstituut voor Mobiliteitsbeleid'' (KiM) from the Ministry of Infrastructure and the Environment in the Netherlands<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3924 respondents from car drivers of the ''Mobiliteitspanel Nederland'' (MPN). <br />
<br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought specifically for that it provides traffic information. Users get most of the traffic information from the internet or the radio and during travel also information panels above the road (DRIPs). From the questions in the KiM research follows that ease-of-use and quality of information are important reasons for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long-distance travel. The system is also used for navigating to points of interest. The most used points of interest are parking lots (75%), gas stations (49%) and places of entertainment (38%). Furthermore, navigation systems are used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of importance follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route themselves, heard of or have seen a traffic jam, or because of road works. Because our system will work optimal when everyone is using the system, and following the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can prove that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors in choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Forms is used to host the questionnaire, because it has all features (kind of questions and setup) required for this research. The questionnaire can be found [https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform here in English] or [https://docs.google.com/forms/d/e/1FAIpQLScx_kPjyN1QNCiAW3xYT-Q77lZks4vDtjyr6YV5x0momVwqVw/viewform here in Dutch], which are identical in content. Please note that by choosing certain answers it is possible that the questionnaire forwards you past irrelevant questions and thus skipping certain sections. In order to see the full questionnaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone receiving the questionnaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non-frequent drivers. While the system is designed for non-self-driving cars it would also work on autonomous cars. The system will first be tested and deployed in the Netherlands and thus this is also the target group for distributing the questionnaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigation system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCNS is designed to be applicable to society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing whatever the system proposes should still be faster than not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistle-blower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this they are taken to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily life. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account, and so on. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom-made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know a reasonable amount of time in advance when users are going to get on the road. This section asks for different occasions (different kinds of destinations) the time the user would be able and willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigation system. In this section the participants are asked about possible approaches the system could take regarding traffic and travel time distribution.<br />
<br />
A participant is free to not answer a question if they do not want to answer, without having to provide a reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representative participants for the target group (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well received with up to 95% of the participants showing interest in the product<br />
* 72.4% of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) versus sending the users on secondary roads)<br />
* Point-based system or reward for driving longer routes is preferred among the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travel ahead of time and inform the navigation system of those plans.<br />
* 88.7% 91.9% and 95.1% of participants indicate to be interested in this goal in the form of this kind of product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until April 7). Seven people responded to the English version and 56 filled in the Dutch version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent.<br />
<br />
==== Study outcome section 1: Introduction questions ====<br />
The introduction questions focused on introducing the product and for us, the researchers, to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answers are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9%, respectively.<br />
<br />
So what does this mean? To start, it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter-device communication. As for environmental nuisances it is hard to say how the device would be optimized at this point since the answers are divided among all four categories. <br />
<br />
==== Study outcome section 2: Willingness to follow the system ====<br />
<br />
The second question category tests whether or not participants are willing to follow the system in four different situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
<br />
==== Study outcome section 3: Fairness ====<br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally only one would be the best. The preferred options are a point-based system that will send you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not further specified how this would take shape) sounds appealing to a large part of the participants in both sections respectively scoring 81.9% and 73.8% on ''very good'' or ''good'' compared to the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made).<br />
<br />
==== Study outcome section 4: Privacy ====<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section about privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
==== Study outcome section 5: System interaction ====<br />
This section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people, this loss in performance will most likely be compensated by the extra information the system is able to receive from more users. The responses show that 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be investigated further, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
==== Study outcome section 6: Finalizing questions ====<br />
In the beginning of the questionnaire participants were asked on their view on the importance of environmental factors that influence the functioning of a navigation system. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take air and noise pollution into account or not, as well as having an option to change this setting, or no option.<br />
64% would like standard incorporation of air pollution into the system, with 72% preferring to have an option to change the setting the system is on default. <br />
For noise pollution these values are slightly lower with 60% wanting a default incorporation of noise into the system, and a slightly higher amount of participants (76%) preferring to have an option to change noise incorporation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency, etc.), but will also need a incorporation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
Making system interaction easier is a strong aim in future planning of travel. For this a section of the questionnaire is designed specifically to understand how drivers would like to do this. The aim is to design one to two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants, which is a mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre-planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car as well an option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regarding daily commutes as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
[[File:Goede Graph user research.JPG|666 px|framed|center|Research results on the time ahead users would inform the system of their travel plans for a variety of purposes.]]<br />
As we see in the graph above, on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interested in the system with all possible goals achieving over 50% of interested participants. The three most popular goals are "The overall travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overall travel time is lower", "The overall travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interested in this goal or product. What this means for the design is mostly that the goal can be customized as long as users do not have an increased travel time themselves. <br />
<br />
===='''Final conclusion'''====<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in-depth interviews, the information collected through this limited resources survey gives a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user-influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user-friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of effects that our system introduces could not be specified. For example, the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. The same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things cannot be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed, more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the whole population, although the target group has been reached. Furthermore, the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point-based reward systems are favoured by the users and have to be further researched, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things cannot be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
We start by introducing a very easy model for our problem. It is not very elaborated and therefore the model does not really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etc. is taken into account, due to time constraints). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is static and for now only looks at a distribution of cars at the beginning of the simulation. Moreover, we assume that everyone uses our clever navigation system and that everyone is obedient to the system and therefore they will always follow their directions. Because we first only focus on this simple road network, we make these assumptions to make sure we can measure in a simple and effective way the effects our navigation system has.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like the graph pictured on the right. [[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation used to run the algorithms on.]]<br />
The double-circled vertex is the destination, which means that the drivers from cities B, C and D have to go to this location (we do not take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each bidirectional road there are two edges, one for each direction (the roads connected directly to A however are unidirectional, because if you are at destination A then you do not have to drive away from it). Every driveway in this graph is indicated with a number and every city is indicated with a capital letter.<br />
<br />
==== Data about roads and cities ====<br />
To model a daily commute situation, we will assume that 30 000 people live in city B who all need to go to A. We will furthermore assume that 20 000 people live in city C, who all need to go to city A. Moreover we will assume that 10 000 people live in city D, who all need to go to city A. This is however only for demonstration purposes, as we want our algorithm to be independent on the starting distribution of cars. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower than it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are high-capacity highways, therefore they can for instance have have the function 0.00025x + 15 (in the case when you need at least 15 minutes to drive over that certain road).</li><br />
<li>Roads 2, 4, 5, 6 and 7 are short secondary roads with a lower capacity (the ''N-wegen'' in Dutch) and these roads can for instance have the function 0.0005x + 5 (you need at least 5 minutes to drive over that certain road).</li><br />
<li>Roads 8 en 9 are also secondary roads, but shorter and with an even lower capacity and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm no cars are present on the roads yet. Suppose we have some driver that lives in city C, which drives over the driveways 6, 5 and 1 (in this order to arrive at his destination A. If this happens, then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10 000 people have driven over road 1 and they have waited 17.5 minutes then the cost value for this road is 175 000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving round” a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This process will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit this process for 10 driving rounds, for example (you can take a different number than 10). If after 10 driving rounds not everyone has arrived at destination A, then it will get a high cost value (and therefore this is automatically a bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was to penalize this behaviour, by increasing the cost value for those solutions proportional to the amount of cars that have not reached their destination. This cost value has to be strictly higher than a cost value of a solution for which it is possible for everyone to arrive within 10 driving rounds to destination A. If we define our cost value this way, then a solution for which in 10 driving rounds not everyone gets to their destination is worse than a solution in which everyone gets to their destination instead.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it does not need neural Nntworks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also it can be used for evaluation at different times. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are quite constraining the real-life situation (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time to evaluate. Especially because this algorithm need to determine distributions for every combination of driveway, destination and different parts of the day. If you look for example at an area the size of Germany and the Netherlands, then there is a huge amount of driveways and a huge amount of destinations. This scales badly.</li><br />
<li> Because of the same reason as before, this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a side note, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
====''Optimizing the distribution matrix''====<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
====''The program''====<br />
<br />
The program, written in Python with TensorFlow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is slow for large networks.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an extended version of the “Brick Algorithm”. This algorithm is an improve of the “Brick Algorithm” because it also works for a larger scale. The idea of this algorithm is based on that traffic jams only occur at a local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven. This however seems to be an unrealistic assumption overall for the real world, but also this assumption can be generalized as we will show below.<br />
<br />
==== Situation Representation ====<br />
We are again see the road network as a graph with nodes and vertices, just as in the “Brick Algorithm”. However, this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm). This graph is pictured on the right.<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb|A more complicated road graph example.]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 cars that need to go to A. If a city does not have subscript A, then 200 cars need to go to A from that city. For subscript G and J the same holds analogously. Furthermore, we assume that every black-coloured road has a single lane and has a minimal travel time of 10 minutes. Every orange road has two lanes and has a minimal travel time of 20 minutes. And every green road has three lanes and has a minimal time of 30 minutes. Then we also have red roads, which have a minimal travel time of 15 minutes and consists of a single lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm, which is described below.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes as an example. Thus, every node that is in ''t'' minutes of a goal is taken into account. We take ''t'' to be equal to 30 minutes in this example and we note that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round.<br />
<br />
==== First iteration round ====<br />
In the first iteration round we will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within a minimum of ''t'' waiting time from its goal. We will discard every node in this example that is further than ''t'' time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example we will use the “Brick Algorithm” for three zones with subscript A, G and J (note that zones do not have to be disjunctive) and we find the optimal solution for these zones. After that (still within the same iteration round) you will take back the discarded nodes and vertices and take a look at cars that travel to a goal and are outside of that zone (so for example someone from a city that does not have subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travellers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as cars arrived at that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account and so on. This iteration is stopped after ''u'' iteration rounds have passed where we will determine ''u'' based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is. However, if two goals are close to each other then it will not take the other goal into account, which might not be desirable behaviour. Therefore we say that when we calculate the local solution for a certain goal that then we also take other goals within a radius of 2*''t'' into account, so we use the “Brick Algorithm” for multiple goals at the same time. Then we discard the result of the other goals and we use only the result of the goal that we were calculating. We call this process the “ILDB+ Algorithm”.<br />
<br />
==== Motivation of assumptions ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is seems logical for some situations, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
However, in situations of a daily commute, this is often not the case. We see traffic jams occurring because of accidents, or merging traffic at interchanges. This can happen alongside any route and is often not only near a destination node. This does not mean that the ILDB algorithm approach cannot be used, it is still applicable with some modifications. The main idea of splitting route finding on a local and global level stays in place, only now extended to zones alongside routes as well. We split the complete road network in multiple overlapping zones, such that every inch of road is covered. For the ILDB algorithm to work, it is important that the zones capture traffic hotspots in full. A way we can accomplish this, is to dynamically make zones active and inactive based on the current traffic situations in such a way that jams are as much covered by a single zone while still having all roads covered by zones.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>The assumption that traffic jams going to a certain location only occur close to that location is a assumption, then this algorithm does not work without modification. However, it is unclear and hard to evaluate the performance of this extended algorithm in terms of completeness and optimality.</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine by ourselves how much iteration rounds are needed and what the value for ''t'' is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius ''t'' of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the "+" part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is run is limited). We also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius ''t'' the total driving time of all traffic does not decrease much if you continue with increasing it and what stands out is that the total driving time increases a bit sometimes for higher radius ''t'', because with a higher radius ''t'' you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm can get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is high and therefore because we only use small graphs in our final implementation we do not use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, that we call the “Auction House Algorithm” because it is similar to an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list.<br />
<br />
==== Definition List ====<br />
;Route<br />
:A set of roads that will take you from your start location to your destination.<br />
;In Advance Best<br />
:A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etc.). m is in this case some constant, which will will determine.<br />
;Capacity<br />
:The maximum amount of cars that can take a certain road each hour.<br />
;Claiming<br />
:If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road.<br />
;Neural<br />
:<code>Neural(arg 1, arg 2, ..., arg n)</code> is the output of a Neural Network with n input arguments.<br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm, again here depicted on the right.<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation as used earlier for the Brick Algorithm.]]<br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 lane: 1750 vehicles per hour </li><br />
<li> 2 lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving or waiting time of a road is easily determined. We will assume that the daily commute rush hour will have a duration of 2 hours. And that when x amount of people in the rush hour need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the driving or waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location and destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<br />
;Shortest path with road (a)<br />
:The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams into account.<br />
;Shortest path without road (b)<br />
:The total driving time that someone needs to go from its location to their destination without taking traffic jams into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.)<br />
;Own total time with road (c)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account.<br />
:Own total time without road (d)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account. Also given that the m in advance best routes might not use this road.<br />
;How much other claims (e)<br />
:How much other start location/destination combinations make a claim on this road.<br />
;Sum shortest path others with road (f)<br />
:The total sum of the shortest path with road for other start location and destination combinations (including other destinations from the same city).<br />
;Sum shortest path others without road (g)<br />
:The total sum of the shortest path without road for other start location and destination combinations.<br />
;Sum total time others with road (h)<br />
:The total sum of the own total time with road for other start location and destination combinations.<br />
;Sum total time others without road (i)<br />
:The total sum of the own total time without road for other start location and destination combinations.<br />
;How much people claim the road (j)<br />
:The total sum of all drivers that claim this road.<br />
;Capacity of a road (k)<br />
:The maximum capacity of a road.<br />
<br />
We define: <code>l = Neural(a, b, c, d, e, f, g, h, i, j, k)</code>. l is probably a value between 0 and 1 (for simplicity, because it is the output of a neural network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location and destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + ... + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unbounded m? ====<br />
A good question is why we bound m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m does not have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evaluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behaviour of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility <!-- NOTE: Capital letter O is intended spelling. -->, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behaviour?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behaviour of a solution, however don’t conclude that one of the solutions is bad because it haves different behaviour. Because the total driving time might be the same for solutions with different behaviour. To illustrate this we will show a graph of a 3D function, depicted on the right.<br />
[[File:Different_Behavior_Same_Result.png|thumb|Graph on a 3D function with an infinite number of absolute minimum values.]]<br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behaviour for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like a tetrahedron or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="Some predetermined graphs"><br />
File:Brick Algorithm - Situation Representation.png|Tetrahedron model.<br />
File:ILDB(+) Algorithm - Situation Representation.png|Complex Rectangle model.<br />
</gallery><br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is shown on the right.<br />
[[File:RandomGraph.png|300px|framed|center|Generated Total Random Graph.]]<br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward). This can be seen in the figure below.<br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get the other figure shown below.<br />
<br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="RR-Grid construction"><br />
File:USE_-_Grid.png|A square neighbor-fully-connected rectangular grid.<br />
File:USE - RRGrid.png|An RR-Grid based on the grid above.<br />
</gallery><br />
<br />
== A Small Complexity Analysis ==<br />
<br />
We have a look at how the training time increases as the amount of cities and routes increases. Unfortunately we can only run this for our very small problem, which may not say very much about training for bigger road networks.<br />
We use the tetrahedron model mentioned before in the situation representation of the Brick Algorithm, where road 2, 4 and 6 are highways with maximum speed 30 and length 280. The other roads used, 1 and 3, secondary roads with maximum speed 17 and length 354. People start from cities B, C and D and all want to arrive at goal A.<br />
<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|The tetrahedron model used for the complexity analysis.]]<br />
<br />
We test three situations, first with only city A and B and one road from B to A, then we add the city D with three routes to the goal. These routes that people can drive are from B to A directly, from B to A via D and from D to A. Similarly we extend to four cities with five routes.<br />
<br />
We train using only one species, a population size of 5 and train for only 5 generations to keep it fast to test.<br />
<br />
We run the training script five times and then take the average running time, so that the time found will not differ more than a few tenths on different runs of the complexity analysis.<br />
<br />
* Two cities (A and B) and one route: 7.3 seconds<br />
* Three cities (A, B and D) and three routes: 10.4 seconds<br />
* Full graph with four cities and nine routes: 12.7 seconds<br />
<br />
From this we could conclude that if we want to use very big road networks, we probably need to devise optimizations to keep the training tractable. But remember this is just a small test with possibly numbers too small to draw any proper conclusions from, so this conclusion is not definite at all.<br />
<br />
== Running the simulation ==<br />
[[File:Pre-2016-q3-group13-simulation.gif|thumb|333px|Simulation ([[:File:PRE16-3-group13-neat-implementation.zip|code]]) of our system compared to a conventional situation where all drivers use the common highway route. Click to run the animation.]]<br />
With the NEAT algorithm, compiling all of the approaches and thoughts into a final implementation, we run a simulation to show the results of our system. This adds to all research done above and is a proof-of-concept of the technical capabilities of this system.<br />
We use a slightly modified version if the tetrahedron model for the simulation, in an attempt to make it somewhat more realistic. It's shape is modeled after the real-life network between the Dutch cities Amsterdam (the goal at the north side), Utrecht (in the middle), The Hague and Eindhoven (south), and Leiden and Amersfoort (west and east of Utrecht, respectively). The roads connecting the large cities (Amsterdam, Utrecht, The Hague and Eindhoven) are 2-lane highways, while the other roads are 1-lane secondary roads with a lower maximum speed. Because this simulates on a low scale, the For this simulation, a random amount of cars between 20 and 25 (for demonstration purposes) leaves a city with destination Amsterdam. However, the cars sent on the left simulation and the right simulation is the same, for a fair comparison.<br />
<br />
On the left in the animation, we simulate our system, where cars are distributed based on the output the evolved neural network provides. On the right the same situation in terms of initial car distribution is shown, but cars always choose the fastest highway route from their city to Amsterdam. As clearly visible in the animation, our system outperforms the situation on the right in terms of total duration before all cars reach the destination. Our system here is based on the NEAT algorithm approach primarily, which is evolved on variable starting distribution of cars on this fixed road network. The generation of this network took less than 15 minutes, while the evaluation once generated was almost instant. For this small example, the time requirements are met, as well as the main requirement as we see improvement of traffic flow in the new situation. The cost requirement is harder to check on this small scale, but since overall travel times are shorter, it seems promising.<br />
<br />
We know that this is still far from a real-life implementation as a navigation system ready to use in consumer cars. For that, we have laid the groundwork on how to expand this simple situation to larger areas with the sections above. With that, the conducted user research will start to show its results in the implementation directly. Also, the scaling of the running time also becomes important, as covered in the previous section. We have shown that it is not easy to make a finished product for use in society, but we hope that with all the research we have done, the best we can, we have shown ways on how it is possible to extend our basic simulation to a more realistic implementation, in terms of scaling and letting the user interests into the design. We are proud to show that we can apply self-learning techniques in an unconventional way, with using the NEAT algorithm approach, into this kind of problem: something we have not seen applied before.<br />
<br />
==References==<br />
''For some reason, references are not showing up when not logged in to the wiki. They are visible when looking at the source of this page.''<br />
<br />
<references/></div>Fri, 14 Apr 2017 09:58:49 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39348
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=39348<p>S143110: /* Objectives */</p>
<hr />
<div><table cellspacing=1<br />
style="float:right; margin: 0 0 0 1em; border: 1px solid #a2a9b1;background-color: #f9f9f9; padding: .7em;"><br />
<tr><td colspan="2" style="border-color: #a2a9b1; padding: 0.2em 0.4em; text-align: center;"><b>Authors</b></td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">David Dekker</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0936100</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Haico Dorenbos</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0959563</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Jori van Falier</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0910959</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Tim Muyrers</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0907928</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Thomas Schouten</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0938927</td></tr><br />
<tr><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">Sten Wessel</td><td style="border-color: #a2a9b1; padding: 0.2em 0.4em;">0941508</td></tr><br />
</table><br />
<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self-learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self-driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self-Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
<div style="clear: left; float:left; margin-bottom:0.5em; padding:0.5em 1.4em 0.8em 0;" class="toclimit-2">__TOC__</div><br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people commute to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these:<br />
<br />
# The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
# The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but is also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that cannot be predicted. Because of all these problems we will be using self-learning software to deal with these problems. This program will try to compensate for all the unknowns.<br />
<br />
== Objectives ==<br />
To make a solution for this problem fit in the time we have for this project, we can only focus on some of these aspects. To account for all of them, this would take a lot more years of research and manpower to create this ground-breaking system. Therefore, we will only focus on a simpler situation with one goal and a finite number of cars leaving for that goal. Hence, we only focus at traffic as a ''snapshot'' of time and only for one goal. We think the results of this can be reasonably generalized to a situation with multiple goals and a continuous flow of traffic. The reason for that is that we can extend the behaviour we create to apply to the full-scale problem, by evaluating the situation repeatedly (from ''snapshot'' to continuous) and doing so sequentially for every goal (from single to multi goal). We will elaborate on how this problem scales in the section about complexity analysis: [[A Small Complexity Analysis]]. For now, we state the objectives for this problem explicitly. We make a distinction between two separate goals.<br />
<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams. This goal is to be satisfied for individual users.<br />
<br />
=== Maximize overall utility ===<br />
Not only minimized travel time individually is important, but also society interest overall. We capture a variety of aspects of interest for society under the umbrella term ''utility''.<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time overall, for all users collectively.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprise ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== Requirements ==<br />
Algorithms are measured in completeness, optimality, time complexity and space complexity. Furthermore the costs of our algorithm and system is also important. So completeness, optimality, time complexity, space complexity and costs are included in the requirements. Because we have combined the completeness and optimality in our main requirement, we also have included time complexity in some of our requirement (we don’t use big O notation in the requirements, because customers are interested in the running time not the big O notation, however we do explain why we meet this requirement by using the big O notation) and last of all cost and space complexity is also combined in the cost requirement (space complexity is included in this, because if you need more memory then you also need to buy more memory or computers).<br />
<br />
=== Main Requirement ===<br />
''For every situation our algorithm has to return for everyone the optimal driving route, such that the sum of the total driving time of everyone is minimal.''<br />
<br />
Note that with ''minimal'' we mean that the sum of the total driving time cannot become any lower than this. For the algorithm implementation, we focus on fulfilling this requirement, because we can in within project bounds a simulation where the fulfilment of this requirement can be measured. However, we will provide a way to extend the algorithm to take the other objectives into account as well.<br />
<br />
=== Cost Requirement ===<br />
''The variable cost per day added with the fixed cost divided by the deprecation period (measured in days) may not exceed €8,96 times the total reduced travel time (also measured in hours) for everyone per day.''<br />
<br />
We explain why we choose this amount. The minimum wage in the Netherlands is €8,96 per hour for people over 23 years. Suppose our system variable cost per hour added with the fixed cost divided by the deprecation period is equal to €7,00 and only one person travel time is reduced by an hour every day and nobody else's travel time is changed, then we have a societal profit of at least €1,96 because that person can start 1 hour earlier with his work and his wage is at least €8,96 per hour. So at least 8,96 – 7,00 = €1,96 societal profit.<br />
<br />
==== Example ====<br />
We want to test if this requirement is satisfied. So let's take a look at a certain algorithm that has a variable cost of €500 per day (electricity costs, hiring employees to maintain the code, etcetera) and the fixed costs are €100.000 (development cost, buying a lot of computers, etcetera). Suppose that the deprecation period of this algorithm is 10 years (after 10 years a new algorithm is found to solve the problem better). Then the variable cost per day added with the fixed cost divided by the deprecation period is equal to 500 + 100.000 / (10 * 365) = 527,40. Now suppose for 100.000 people the driving time on average is decreased by 15 minutes every day and nothing else is changed in driving time. Then we have a societal revenue of 100.000 * 0,25 * 8,96 = €224.000,00. 527,40 <= 224000,00, thus in this case our system satisfies this requirement and has a societal benefit of 224.000,00 – 527,40 = €223.472,60.<br />
<br />
=== After Running Time Requirement ===<br />
''The after running time may be no more than 5 seconds.''<br />
<br />
The after running time is the running time when a customer wants the best route to his destination. The calculation of the best route might only take 5 seconds, this is because customers get impatient if it takes longer than 5 seconds.<br />
<br />
=== Pre Running Time Requirement ===<br />
''The pre running time may be no more than 7 days.''<br />
<br />
This is the running time that the artificial intelligent program takes to prepare the system for being used by drivers (for example a NEAT algorithm, introduced later, that optimizes a neural network is included in this running time). The goal is to optimize for a specific road network, where the traffic distributions are still variable. Hence, this running time may be high, because program only has to be run when the roadmap changes and once in beginning. So users do not have to wait on this algorithm, because this algorithm can run parallel with the calculation of the best route for the users. This is because changes in a road network are known in advance, so when generating a new neural network for the new road situation, the calculation of best routes for users is still for the old road network.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices currently available. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to the current situation only, instead of trying to avoid certain situations beforehand.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will reroute all traffic over this faster route. Right now, this is not really a problem since not all users will follow that advice. However, in the long term (consider for instance self-driving cars), this approach will only create a continuous shifting of recommended routes as a solely reactive system. A self-learning system however find a stable solution, since it can predict traffic jams that haven't even happened yet, and account for the effects of rerouting traffic. It can reroute the user many kilometres before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
[[File:Idea visualize prof.jpg|300 px|thumb|Visualization of the rerouting process.]]<br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously with developing the algorithm. The group of researchers will be split where four members of the team will work on making the self-learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representative for a final implementation as possible.<br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external parties to use and understand the way the system operates, and what conclusions it presents.<br />
<br />
== User and Society ==<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish and less considerate to other drivers. As long as people are driving their car (instead of autonomous), we can investigate further if and how we want to incorporate this.<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey, with the goal of decreasing overall travel time. When looking at literature we find something called ''risk aversion''<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favourable bet (overall travel time decrease) when it means they can directly lose (more travel time)<ref>YouTube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== Traffic research ==<br />
<br />
=== What is a traffic jam ===<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can identify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: unnecessary braking resulting in so called ''phantom jams''<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans) and probably also when cars become fully autonomous, we do see however that accidents can be massively decreased when we decrease the amount of cars driving at a certain road. The busier it is, the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the risk factor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
[[File:FormulaUsed1PNG.png|380 px|thumb|Relation between traffic volume and road crashes.]]<br />
<br />
Too many drivers on the road also causes direct traffic jams through unnecessary braking. When a car in front brakes, the one behind it brakes too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill. This phenomenon is called a ''phantom jam'' which seem to occur and disappear suddenly on crowded roads.<br />
<br />
Road works cannot really be avoided. But if the system is connected to a database, it should be able to take this into account. We can see and handle planned roadwork the same as a (however temporary) change in the road network.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self-learning system should be able to handle occasional accidents and reroute accordingly.<br />
<br />
In conclusion, to make sure the navigational system works in real life you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focused on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be beneficial. Moving on, we abstract away the details of how unpredictable jams occur and disappear and focus training the system based on experiments where these traffic jam situations just happen.<br />
<br />
=== Traffic data ===<br />
<br />
For our system to work, traffic data has to be acquired for realistic large scale data, to help train the system. As our system is still in an early development stage, the data found will not be used as it requires some parsing to get it in a format where the system can work with. As this is too much time consuming, we will focus on a fictional road network inspired on a rush hour commute situation. However, it will show that this data is obtainable and could be used for our system. The data has been acquired from ''Nationale Databank Wegverkeersgegevens'' (NDW) <ref> Nationale Databank Wegverkeersgegevens (NDW), Retrieved March, 2017 from http://www.ndw.nu/nl/ </ref>. Open historic data has been applied for and obtained. It gives information about the traffic density and speed of cars on 27000 locations in the Netherlands. This information can be used for our program to determine how many people travel these roads and when a traffic jam occurs. The data is delivered in the DATEX II form, which is a European standard for traffic data <ref> T. Delissen, Nederlands Profiel DATEX II, 2015-2a, Retrieved April, 2017 from http://www.ndw.nu/downloaddocument/039ee3e1098f37830e26df848f6202ca/Nederlands%20Profiel%20DATEX-II%202015-2a%20(14-01-2016).pdf </ref>. By doing further research about this data, the data can be read by our program and used to train the system. The data base also has current traffic data, that can be used when our system becomes operational. By knowing about the traffic on these 25000 points, the system can get a good understanding of the current traffic situation, and therefore make better decisions. All of this data has to be implemented in the system when it is further developed and can be beneficial to the functioning of our system.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by ''Kennisinstituut voor Mobiliteitsbeleid'' (KiM) from the Ministry of Infrastructure and the Environment in the Netherlands<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3924 respondents from car drivers of the ''Mobiliteitspanel Nederland'' (MPN). <br />
<br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought specifically for that it provides traffic information. Users get most of the traffic information from the internet or the radio and during travel also information panels above the road (DRIPs). From the questions in the KiM research follows that ease-of-use and quality of information are important reasons for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long-distance travel. The system is also used for navigating to points of interest. The most used points of interest are parking lots (75%), gas stations (49%) and places of entertainment (38%). Furthermore, navigation systems are used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of importance follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route themselves, heard of or have seen a traffic jam, or because of road works. Because our system will work optimal when everyone is using the system, and following the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can prove that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors in choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Forms is used to host the questionnaire, because it has all features (kind of questions and setup) required for this research. The questionnaire can be found [https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform here in English] or [https://docs.google.com/forms/d/e/1FAIpQLScx_kPjyN1QNCiAW3xYT-Q77lZks4vDtjyr6YV5x0momVwqVw/viewform here in Dutch], which are identical in content. Please note that by choosing certain answers it is possible that the questionnaire forwards you past irrelevant questions and thus skipping certain sections. In order to see the full questionnaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone receiving the questionnaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non-frequent drivers. While the system is designed for non-self-driving cars it would also work on autonomous cars. The system will first be tested and deployed in the Netherlands and thus this is also the target group for distributing the questionnaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigation system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCNS is designed to be applicable to society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing whatever the system proposes should still be faster than not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistle-blower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this they are taken to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily life. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account, and so on. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom-made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know a reasonable amount of time in advance when users are going to get on the road. This section asks for different occasions (different kinds of destinations) the time the user would be able and willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigation system. In this section the participants are asked about possible approaches the system could take regarding traffic and travel time distribution.<br />
<br />
A participant is free to not answer a question if they do not want to answer, without having to provide a reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representative participants for the target group (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well received with up to 95% of the participants showing interest in the product<br />
* 72.4% of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) versus sending the users on secondary roads)<br />
* Point-based system or reward for driving longer routes is preferred among the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travel ahead of time and inform the navigation system of those plans.<br />
* 88.7% 91.9% and 95.1% of participants indicate to be interested in this goal in the form of this kind of product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until April 7). Seven people responded to the English version and 56 filled in the Dutch version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent.<br />
<br />
==== Study outcome section 1: Introduction questions ====<br />
The introduction questions focused on introducing the product and for us, the researchers, to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answers are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9%, respectively.<br />
<br />
So what does this mean? To start, it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter-device communication. As for environmental nuisances it is hard to say how the device would be optimized at this point since the answers are divided among all four categories. <br />
<br />
==== Study outcome section 2: Willingness to follow the system ====<br />
<br />
The second question category tests whether or not participants are willing to follow the system in four different situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
<br />
==== Study outcome section 3: Fairness ====<br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally only one would be the best. The preferred options are a point-based system that will send you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not further specified how this would take shape) sounds appealing to a large part of the participants in both sections respectively scoring 81.9% and 73.8% on ''very good'' or ''good'' compared to the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made).<br />
<br />
==== Study outcome section 4: Privacy ====<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section about privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
==== Study outcome section 5: System interaction ====<br />
This section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people, this loss in performance will most likely be compensated by the extra information the system is able to receive from more users. The responses show that 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be investigated further, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
==== Study outcome section 6: Finalizing questions ====<br />
In the beginning of the questionnaire participants were asked on their view on the importance of environmental factors that influence the functioning of a navigation system. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take air and noise pollution into account or not, as well as having an option to change this setting, or no option.<br />
64% would like standard incorporation of air pollution into the system, with 72% preferring to have an option to change the setting the system is on default. <br />
For noise pollution these values are slightly lower with 60% wanting a default incorporation of noise into the system, and a slightly higher amount of participants (76%) preferring to have an option to change noise incorporation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency, etc.), but will also need a incorporation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
Making system interaction easier is a strong aim in future planning of travel. For this a section of the questionnaire is designed specifically to understand how drivers would like to do this. The aim is to design one to two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants, which is a mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre-planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car as well an option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regarding daily commutes as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
[[File:Goede Graph user research.JPG|666 px|framed|center|Research results on the time ahead users would inform the system of their travel plans for a variety of purposes.]]<br />
As we see in the graph above, on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interested in the system with all possible goals achieving over 50% of interested participants. The three most popular goals are "The overall travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overall travel time is lower", "The overall travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interested in this goal or product. What this means for the design is mostly that the goal can be customized as long as users do not have an increased travel time themselves. <br />
<br />
===='''Final conclusion'''====<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in-depth interviews, the information collected through this limited resources survey gives a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user-influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user-friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of effects that our system introduces could not be specified. For example, the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. The same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things cannot be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed, more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the whole population, although the target group has been reached. Furthermore, the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point-based reward systems are favoured by the users and have to be further researched, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things cannot be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
We start by introducing a very easy model for our problem. It is not very elaborated and therefore the model does not really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etc. is taken into account, due to time constraints). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is static and for now only looks at a distribution of cars at the beginning of the simulation. Moreover, we assume that everyone uses our clever navigation system and that everyone is obedient to the system and therefore they will always follow their directions. Because we first only focus on this simple road network, we make these assumptions to make sure we can measure in a simple and effective way the effects our navigation system has.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like the graph pictured on the right. [[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation used to run the algorithms on.]]<br />
The double-circled vertex is the destination, which means that the drivers from cities B, C and D have to go to this location (we do not take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each bidirectional road there are two edges, one for each direction (the roads connected directly to A however are unidirectional, because if you are at destination A then you do not have to drive away from it). Every driveway in this graph is indicated with a number and every city is indicated with a capital letter.<br />
<br />
==== Data about roads and cities ====<br />
To model a daily commute situation, we will assume that 30 000 people live in city B who all need to go to A. We will furthermore assume that 20 000 people live in city C, who all need to go to city A. Moreover we will assume that 10 000 people live in city D, who all need to go to city A. This is however only for demonstration purposes, as we want our algorithm to be independent on the starting distribution of cars. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower than it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are high-capacity highways, therefore they can for instance have have the function 0.00025x + 15 (in the case when you need at least 15 minutes to drive over that certain road).</li><br />
<li>Roads 2, 4, 5, 6 and 7 are short secondary roads with a lower capacity (the ''N-wegen'' in Dutch) and these roads can for instance have the function 0.0005x + 5 (you need at least 5 minutes to drive over that certain road).</li><br />
<li>Roads 8 en 9 are also secondary roads, but shorter and with an even lower capacity and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm no cars are present on the roads yet. Suppose we have some driver that lives in city C, which drives over the driveways 6, 5 and 1 (in this order to arrive at his destination A. If this happens, then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10 000 people have driven over road 1 and they have waited 17.5 minutes then the cost value for this road is 175 000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving round” a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This process will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit this process for 10 driving rounds, for example (you can take a different number than 10). If after 10 driving rounds not everyone has arrived at destination A, then it will get a high cost value (and therefore this is automatically a bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was to penalize this behaviour, by increasing the cost value for those solutions proportional to the amount of cars that have not reached their destination. This cost value has to be strictly higher than a cost value of a solution for which it is possible for everyone to arrive within 10 driving rounds to destination A. If we define our cost value this way, then a solution for which in 10 driving rounds not everyone gets to their destination is worse than a solution in which everyone gets to their destination instead.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it does not need neural Nntworks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also it can be used for evaluation at different times. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are quite constraining the real-life situation (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time to evaluate. Especially because this algorithm need to determine distributions for every combination of driveway, destination and different parts of the day. If you look for example at an area the size of Germany and the Netherlands, then there is a huge amount of driveways and a huge amount of destinations. This scales badly.</li><br />
<li> Because of the same reason as before, this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a side note, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
====''Optimizing the distribution matrix''====<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
====''The program''====<br />
<br />
The program, written in Python with TensorFlow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is slow for large networks.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an extended version of the “Brick Algorithm”. This algorithm is an improve of the “Brick Algorithm” because it also works for a larger scale. The idea of this algorithm is based on that traffic jams only occur at a local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven. This however seems to be an unrealistic assumption overall for the real world, but also this assumption can be generalized as we will show below.<br />
<br />
==== Situation Representation ====<br />
We are again see the road network as a graph with nodes and vertices, just as in the “Brick Algorithm”. However, this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm). This graph is pictured on the right.<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb|A more complicated road graph example.]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 cars that need to go to A. If a city does not have subscript A, then 200 cars need to go to A from that city. For subscript G and J the same holds analogously. Furthermore, we assume that every black-coloured road has a single lane and has a minimal travel time of 10 minutes. Every orange road has two lanes and has a minimal travel time of 20 minutes. And every green road has three lanes and has a minimal time of 30 minutes. Then we also have red roads, which have a minimal travel time of 15 minutes and consists of a single lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm, which is described below.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes as an example. Thus, every node that is in ''t'' minutes of a goal is taken into account. We take ''t'' to be equal to 30 minutes in this example and we note that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round.<br />
<br />
==== First iteration round ====<br />
In the first iteration round we will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within a minimum of ''t'' waiting time from its goal. We will discard every node in this example that is further than ''t'' time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example we will use the “Brick Algorithm” for three zones with subscript A, G and J (note that zones do not have to be disjunctive) and we find the optimal solution for these zones. After that (still within the same iteration round) you will take back the discarded nodes and vertices and take a look at cars that travel to a goal and are outside of that zone (so for example someone from a city that does not have subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travellers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as cars arrived at that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account and so on. This iteration is stopped after ''u'' iteration rounds have passed where we will determine ''u'' based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is. However, if two goals are close to each other then it will not take the other goal into account, which might not be desirable behaviour. Therefore we say that when we calculate the local solution for a certain goal that then we also take other goals within a radius of 2*''t'' into account, so we use the “Brick Algorithm” for multiple goals at the same time. Then we discard the result of the other goals and we use only the result of the goal that we were calculating. We call this process the “ILDB+ Algorithm”.<br />
<br />
==== Motivation of assumptions ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is seems logical for some situations, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
However, in situations of a daily commute, this is often not the case. We see traffic jams occurring because of accidents, or merging traffic at interchanges. This can happen alongside any route and is often not only near a destination node. This does not mean that the ILDB algorithm approach cannot be used, it is still applicable with some modifications. The main idea of splitting route finding on a local and global level stays in place, only now extended to zones alongside routes as well. We split the complete road network in multiple overlapping zones, such that every inch of road is covered. For the ILDB algorithm to work, it is important that the zones capture traffic hotspots in full. A way we can accomplish this, is to dynamically make zones active and inactive based on the current traffic situations in such a way that jams are as much covered by a single zone while still having all roads covered by zones.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>The assumption that traffic jams going to a certain location only occur close to that location is a assumption, then this algorithm does not work without modification. However, it is unclear and hard to evaluate the performance of this extended algorithm in terms of completeness and optimality.</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine by ourselves how much iteration rounds are needed and what the value for ''t'' is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius ''t'' of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the "+" part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is run is limited). We also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius ''t'' the total driving time of all traffic does not decrease much if you continue with increasing it and what stands out is that the total driving time increases a bit sometimes for higher radius ''t'', because with a higher radius ''t'' you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm can get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is high and therefore because we only use small graphs in our final implementation we do not use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, that we call the “Auction House Algorithm” because it is similar to an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list.<br />
<br />
==== Definition List ====<br />
;Route<br />
:A set of roads that will take you from your start location to your destination.<br />
;In Advance Best<br />
:A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etc.). m is in this case some constant, which will will determine.<br />
;Capacity<br />
:The maximum amount of cars that can take a certain road each hour.<br />
;Claiming<br />
:If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road.<br />
;Neural<br />
:<code>Neural(arg 1, arg 2, ..., arg n)</code> is the output of a Neural Network with n input arguments.<br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm, again here depicted on the right.<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|Situation representation as used earlier for the Brick Algorithm.]]<br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 lane: 1750 vehicles per hour </li><br />
<li> 2 lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving or waiting time of a road is easily determined. We will assume that the daily commute rush hour will have a duration of 2 hours. And that when x amount of people in the rush hour need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the driving or waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location and destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<br />
;Shortest path with road (a)<br />
:The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams into account.<br />
;Shortest path without road (b)<br />
:The total driving time that someone needs to go from its location to their destination without taking traffic jams into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.)<br />
;Own total time with road (c)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account.<br />
:Own total time without road (d)<br />
:The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams into account. Also given that the m in advance best routes might not use this road.<br />
;How much other claims (e)<br />
:How much other start location/destination combinations make a claim on this road.<br />
;Sum shortest path others with road (f)<br />
:The total sum of the shortest path with road for other start location and destination combinations (including other destinations from the same city).<br />
;Sum shortest path others without road (g)<br />
:The total sum of the shortest path without road for other start location and destination combinations.<br />
;Sum total time others with road (h)<br />
:The total sum of the own total time with road for other start location and destination combinations.<br />
;Sum total time others without road (i)<br />
:The total sum of the own total time without road for other start location and destination combinations.<br />
;How much people claim the road (j)<br />
:The total sum of all drivers that claim this road.<br />
;Capacity of a road (k)<br />
:The maximum capacity of a road.<br />
<br />
We define: <code>l = Neural(a, b, c, d, e, f, g, h, i, j, k)</code>. l is probably a value between 0 and 1 (for simplicity, because it is the output of a neural network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location and destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + ... + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unbounded m? ====<br />
A good question is why we bound m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m does not have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evaluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behaviour of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility <!-- NOTE: Capital letter O is intended spelling. -->, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behaviour?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behaviour of a solution, however don’t conclude that one of the solutions is bad because it haves different behaviour. Because the total driving time might be the same for solutions with different behaviour. To illustrate this we will show a graph of a 3D function, depicted on the right.<br />
[[File:Different_Behavior_Same_Result.png|thumb|Graph on a 3D function with an infinite number of absolute minimum values.]]<br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behaviour for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like a tetrahedron or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="Some predetermined graphs"><br />
File:Brick Algorithm - Situation Representation.png|Tetrahedron model.<br />
File:ILDB(+) Algorithm - Situation Representation.png|Complex Rectangle model.<br />
</gallery><br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is shown on the right.<br />
[[File:RandomGraph.png|300px|framed|center|Generated Total Random Graph.]]<br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward). This can be seen in the figure below.<br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get the other figure shown below.<br />
<br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
<gallery style="margin-left: auto; margin-right: auto;" widths=250px heights=250px caption="RR-Grid construction"><br />
File:USE_-_Grid.png|A square neighbor-fully-connected rectangular grid.<br />
File:USE - RRGrid.png|An RR-Grid based on the grid above.<br />
</gallery><br />
<br />
== A Small Complexity Analysis ==<br />
<br />
We have a look at how the training time increases as the amount of cities and routes increases. Unfortunately we can only run this for our very small problem, which may not say very much about training for bigger road networks.<br />
We use the tetrahedron model mentioned before in the situation representation of the Brick Algorithm, where road 2, 4 and 6 are highways with maximum speed 30 and length 280. The other roads used, 1 and 3, secondary roads with maximum speed 17 and length 354. People start from cities B, C and D and all want to arrive at goal A.<br />
<br />
[[File:Brick Algorithm - Situation Representation.png|thumb|The tetrahedron model used for the complexity analysis.]]<br />
<br />
We test three situations, first with only city A and B and one road from B to A, then we add the city D with three routes to the goal. These routes that people can drive are from B to A directly, from B to A via D and from D to A. Similarly we extend to four cities with five routes.<br />
<br />
We train using only one species, a population size of 5 and train for only 5 generations to keep it fast to test.<br />
<br />
We run the training script five times and then take the average running time, so that the time found will not differ more than a few tenths on different runs of the complexity analysis.<br />
<br />
* Two cities (A and B) and one route: 7.3 seconds<br />
* Three cities (A, B and D) and three routes: 10.4 seconds<br />
* Full graph with four cities and nine routes: 12.7 seconds<br />
<br />
From this we could conclude that if we want to use very big road networks, we probably need to devise optimizations to keep the training tractable. But remember this is just a small test with possibly numbers too small to draw any proper conclusions from, so this conclusion is not definite at all.<br />
<br />
== Running the simulation ==<br />
[[File:Pre-2016-q3-group13-simulation.gif|thumb|333px|Simulation ([[:File:PRE16-3-group13-neat-implementation.zip|code]]) of our system compared to a conventional situation where all drivers use the common highway route. Click to run the animation.]]<br />
With the NEAT algorithm, compiling all of the approaches and thoughts into a final implementation, we run a simulation to show the results of our system. This adds to all research done above and is a proof-of-concept of the technical capabilities of this system.<br />
We use a slightly modified version if the tetrahedron model for the simulation, in an attempt to make it somewhat more realistic. It's shape is modeled after the real-life network between the Dutch cities Amsterdam (the goal at the north side), Utrecht (in the middle), The Hague and Eindhoven (south), and Leiden and Amersfoort (west and east of Utrecht, respectively). The roads connecting the large cities (Amsterdam, Utrecht, The Hague and Eindhoven) are 2-lane highways, while the other roads are 1-lane secondary roads with a lower maximum speed. Because this simulates on a low scale, the For this simulation, a random amount of cars between 20 and 25 (for demonstration purposes) leaves a city with destination Amsterdam. However, the cars sent on the left simulation and the right simulation is the same, for a fair comparison.<br />
<br />
On the left in the animation, we simulate our system, where cars are distributed based on the output the evolved neural network provides. On the right the same situation in terms of initial car distribution is shown, but cars always choose the fastest highway route from their city to Amsterdam. As clearly visible in the animation, our system outperforms the situation on the right in terms of total duration before all cars reach the destination. Our system here is based on the NEAT algorithm approach primarily, which is evolved on variable starting distribution of cars on this fixed road network. The generation of this network took less than 15 minutes, while the evaluation once generated was almost instant. For this small example, the time requirements are met, as well as the main requirement as we see improvement of traffic flow in the new situation. The cost requirement is harder to check on this small scale, but since overall travel times are shorter, it seems promising.<br />
<br />
We know that this is still far from a real-life implementation as a navigation system ready to use in consumer cars. For that, we have laid the groundwork on how to expand this simple situation to larger areas with the sections above. With that, the conducted user research will start to show its results in the implementation directly. Also, the scaling of the running time also becomes important, as covered in the previous section. We have shown that it is not easy to make a finished product for use in society, but we hope that with all the research we have done, the best we can, we have shown ways on how it is possible to extend our basic simulation to a more realistic implementation, in terms of scaling and letting the user interests into the design. We are proud to show that we can apply self-learning techniques in an unconventional way, with using the NEAT algorithm approach, into this kind of problem: something we have not seen applied before.<br />
<br />
==References==<br />
''For some reason, references are not showing up when not logged in to the wiki. They are visible when looking at the source of this page.''<br />
<br />
<references/></div>Fri, 14 Apr 2017 09:58:30 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37395
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37395<p>S143110: /* What is a traffic jam */</p>
<hr />
<div>David Dekker - 0936100<br />
<br />
Haico Dorenbos - 0959563<br />
<br />
Jori van Falier - 0910959<br />
<br />
Tim Muyrers - 0907928<br />
<br />
Thomas Schouten - 0938927<br />
<br />
Sten Wessel - 0941508<br />
<br />
<br />
<br />
== Introduction ==<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people go to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these: <br />
1. The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
2. The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that can not be predicted. Because of all these problems we will be using self learning software to deal with these problems. This program will try to compensate for all the unknowns. <br />
<br />
== Objectives ==<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams.<br />
<br />
=== Maximize overall utility ===<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time, this is the main objective as mentioned above.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprises ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices available currently. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to a situation instead trying to avoid certain situations.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will distribute traffic equally over the roads available, which may seem like the optimal situation. A self learning system however should have more roads available, since it can predict traffic jams that haven't even happened yet. It can reroute the user many kilometers before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Idea visualize prof.jpg|300 px]]</div><br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously. The group of researchers will be split where four members of the team will work on making the self learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representable as possible. <br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external personnel to use and understand the way the system operates, and what conclusions it provides.<br />
<br />
== User and Society ==<br />
<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish. As long as people are driving their car (instead of autonomous), we can investigate further if/how we want to incorporate this. <br />
<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey. When looking at literature we find something called: “risk aversion”<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favorable bet (overall travel time decrease) when it means they can directly lose out (more travel time) <ref>Youtube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== Traffic research ==<br />
<br />
=== What is a traffic jam ===<br />
<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can indentify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: Over Breaking<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans). We do see though that accidents can be massively decreased when we decrease the amount of cars driving at a road. The busier it is the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the riskfactor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsed1PNG.png|380 px]]</div><br />
Too many drivers on the road also causes direct traffic jams through over breaking. When a car in front breakers the one after it breaks too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill.<br />
<br />
Road works can’t really be avoided. But if the system is connected to a database, it should be able to take this into account.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self learning system should be able to.<br />
<br />
<br />
<br />
So to make sure the navigational system works in society you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. <br />
More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focussed on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be benicial.<br />
<br />
=== Traffic data ===<br />
<br />
For our system to work, traffic data has to be acquired for realistic large scale data, to help train the system. As our system is still in an early development stage, the data found will not be used, but it will show that this data is obtainable and could be used for our system. The data has been acquired from Nationale Databank Wegverkeersgegevens (NDW) <ref> Nationale Databank Wegverkeersgegevens (NDW), Retrieved March, 2017 from http://www.ndw.nu/nl/ </ref>. Open historic data has been applied for and obtained. It gives information about the traffic density and speed of cars on 27000 locations in the Netherlands. This information can be used for our program to determine how many people travel these roads and when a traffic jam occurs. The data is delivered in the DATEX II form, which is a European standard for traffic data <ref> T. Delissen, Nederlands Profiel DATEX II, 2015-2a, Retrieved April, 2017 from http://www.ndw.nu/downloaddocument/039ee3e1098f37830e26df848f6202ca/Nederlands%20Profiel%20DATEX-II%202015-2a%20(14-01-2016).pdf </ref>. By doing further research about this data, the data can be read by our program and used to train the system. The data base also has current traffic data, that can be used when our system becomes operational. By knowing about the traffic on these 25000 points, the system can get a good understanding of the current traffic situation, and therefore make better decisions. All of this data has to be implemented in the system when it is further developed and can be very beneficial to the functionality of our system.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by “Kennisinstituut voor Mobiliteitsbeleid (KiM)” from the ministry of infrastructure and milieu<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3.924 respondents from car drivers of the “Mobiliteitspanel Nederland (MPN)”. <br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought because of traffic information. Most of the information comes from the internet and the radio and during the travel also information panels above the road (DRIPs). From the questions follows that easy in use and quality of information are important reason for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long distant travels. The system is also used for points of interest. The most used points of interest are parking lot (75%), gas stations (49%) and places of entertainment (38%). Furthermore the system is used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of important follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route, heard or seen a traffic jam, or because of road works. Because our system will work optimal when everyone using the system, follows the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can proof that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors into choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Drive is used to host the questionnaire, because it has all features (sort of questions/ set-up) required for this research. <br />
<br />
The link below will bring you to the questionnaire*:<br />
<br />
https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform<br />
<br />
* Please note that by choosing certain answers it is possible that the questionaire forwards you and thus skipping certain sections. In order to see the full questionaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone recieving the questionaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non frequent drivers. While the system is designed for non selfdriving cars it would also work on autonomous cars. The system will first be tested and deployed in The Netherlands and thus this is also the target group for distributing the questionaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigational system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCS is designed to be applicable into society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing and whatever the system proposes it should still be faster then not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistleblower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this he/she is taking to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily lives. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account etc. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know way in advance when users are going to drive. This section asks of different occasions (different destinations) the time a user would be able/willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigational system. In this section the participants are asked about possible approaches the system could take regarding traffic/travel time distribution.<br />
<br />
A participant is free to not answer a question if he/she does not want to answer, without having to provide an reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representable participants (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well recieved with up to 95% of the participants showing interrest in the product<br />
* 72.4 % of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) vs sending the users on secondairy roads)<br />
* Point based system or reward for driving longer routes is preferred amongst the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travels ahead of time and inform the navigation system of said plans.<br />
* 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until 7th of April). Seven people responded to the English version and fifty-six filled in the English version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent. <br />
<br />
The introduction questions focused on introducing the product and for us the investigators to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answer are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9% respectively.<br />
<br />
So what does this mean. To start it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter device communication. As for environmental nuisances it’s hard to say how the device would be optimized at this point since the answers are divided amongst all four categories. <br />
<br />
The second question category tests whether or not participants are willing to follow the system in four situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally one would be the best. The preferred options are a point based system that will sent you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not defined) sounds appealing to a large part of the participants with both sections respectively scoring 81.9% and 73.8% on Very good or Good in comparison with the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made). <br />
<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
<br />
<br />
The next section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people this fall in performance will most likely be compensated by the extra information the system is able to receive from more users. 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be more investigated, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
<br />
In the beginning of the questionaire participants were asked on their point of view on the important of environmental factors influencing navigation systems functioning. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take polution/noise into account or not aswell as having a option to change this setting or no option.<br />
64% would like standard incorperation of polution into the system, with 72% prefering to have an option to change the setting the system is on standard. <br />
For noise these values are slightly lower with 60% wanting a standard incorperation of noise into the system, and a slightly higher amount of participants (76%) prefering to have a option to change noise incorperation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency etc), but will also need a incorperation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
<br />
Making system interaction easier is a strong aim in future planning of travels. For this a section of the questionnaire is design specifically to understand how participants/drivers would like to do this. The aim is to design one maximally two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants which is the mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car aswell a option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regard work travel as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Goede Graph user research.JPG|666 px]]</div><br />
As you can see in the graph above on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interrested in the system with all possible goals achieving over 50% of interrested participants. The three most popular goals are "The overal travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overal travel time is lower", "The overal travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product. What this means for the design mostly means that the goal can be customized as long as users dont lose out on travel time themselves. <br />
<br />
'''Final conclusion'''<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in depth interviews, the information collected through this limited resources survey give a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of things could not be specified. For example the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. Same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things can not be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the population. Furthermore the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point and reward systems are favoured by the users and have to be researched even more, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things can not be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
==== Introduction ====<br />
We start by introducing a very easy model for our problem. It is not very expanded and therefore the model doesn’t really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etcetera is taken into account ). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is completely static and ironically it doesn’t take time into account. Moreover we assume that everyone uses our clever navigation system and that everyone is obedient to our clever navigation system and therefore they will always follow their directions.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like this: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
The vertex that is two times circled is the destination, which means that the drivers from cities B, C and D have to go to this location (we don’t take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each road there are two edges, because you can take the driveway forward or the driveway backward (for destination A however there are none driveway backwards, because if you are at destination A then you don’t have to drive back). Every driveway in this graph is indicated with a number and every city is indicated with a letter.<br />
<br />
==== Data about roads and cities ====<br />
We will assume that 30.000 people live in city B who all need to go to A. We will furthermore assume that 20.000 people live in city C, who all need to go to city A. Moreover we will assume that 10.000 people live in city D, who all need to go to city A. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower then it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are huge highways (they can handle a lot of drivers), therefore they have the function 0.00025x + 15 (you need at least 15 minutes to drive over that certain road)</li><br />
<li>Roads 2, 4, 5, 6 and 7 are little, short roads (In dutch: N-wegen) (and they can only handle a small amount of drivers) and these roads have the function 0.0005x + 5 (you need at least 5 minutes to driver over that certain road).</li><br />
<li>Roads 8 en 9 are small, very short roads (you are soon at the new node by using this vertex) and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm. Suppose we have some driver that live in city C, which drives over the driveways 6, 5 and 1 (in this order) and finally arrives at his destination A. If this happens then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10.000 people have driven over road 1 and they have waited 17,5 minutes then the cost value for this road is 175.000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving” round a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This script will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit our script till 10 “driving” rounds for example (you can take a different number than 10). If after 10 “driving” rounds not everyone has arrived at destination A then it will get a very high cost value (and therefore this is automatically a very bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was for example to say that this value is equal to (Float_MaxValue) * (amount_drivers_at_start - amount_drivers_reached_destination) / (amount_drivers_at_start). This cost value has to be definitely higher than a cost value of a solution for which it is possible for everyone to arrive within 10 “driving” to destination A. If we define our cost value this way then a solution for which 10 “driving” rounds is too less to for everyone to get to their destination is worse than a solution in which everyone get to their destination in 10 “driving” rounds.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also for different times it can be used. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are not very logical (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time. Especially because this algorithm need to determine driveways * destinations * different_parts_of_the_day amount of constant. If you look for example at an area Germany and the Netherland then there is a huge amount of driveways and a huge amount of destinations. Hence you will have to determine a lot of constants.</li><br />
<li> Because of the reason before this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a sidenote, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
''Optimizing the distribution matrix''<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
''The program''<br />
<br />
The program, written in Python with Tensorflow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is very slow.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an expanded version of the “Brick Algorithm”. Therefore it is really important to have read the “Brick Algorithm” and understand how it works. This algorithm is an improve of the “Brick Algorithm” because it also work for a more larger scale. The idea of this algorithm is based on that traffic jams only occur at local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven.<br />
<br />
==== Situation Representation ====<br />
We are trying again to see the problem as a graph with nodes and vertices, likewise the “Brick Algorithm”. However this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm):<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 humans that need to go to A. If a city doesn’t have subscript A then 200 humans need to go to A from that city. For subscript G and J the same holds. Furthermore we assume that every black road has 1 lane and has a minimal time of 10 minutes. Every orange road has 2 lanes and has a minimal time of 20 minutes. And every green road has 3 lanes and has a minimal time of 30 minutes. Then we also have red road, which has a minimal time of 15 minutes and has 1 lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes for example. So every node that is in t minutes of a goal is taken into account. We take t is 30 minutes in this example and you might notice that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round:<br />
<br />
==== First iteration round ====<br />
In the first iteration round you will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within t minimal waiting time from its goal. You will discard every node in this example that is further than t time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example you will use the “Brick Algorithm” for three zones with subscript: A, G and J (note that zones doesn’t have to be disjunctive) and you will found the optimal solution for these zones. After that (same iteration round) you will take back the discarded nodes and vertices and take a look at humans that will travel to a goal and are outside of that zone (so for example someone from a city that hasn’t got subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travelers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as new inhabitants of that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account etc. This iteration is stopped after u iteration rounds have passed where we will determine u based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is, however if two goals are close to each other then it will not take the other goal into account, which might be bad. Therefore you might say that when you calculate the local solution for a certain goal that you also take other goals within a radius of 2*t into account, so you will use the “Brick Algorithm” for multiple goals at the same time. Then you discard the result of the other goals and you will use only the result of the goal that you were calculating. This algorithm will be called “ILDB+ Algorithm”.<br />
<br />
==== Defense of Assumption ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is very logical, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>If the assumption that traffic jams going to a certain location only occur close to that location is a very wrong assumption, then this algorithm doesn’t work (if it is partially true then this algorithm might still work).</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine ourselves how much iteration rounds are needed and what the value for t is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius t of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the + part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is runned is limited). You can also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius t the total driving time of all traffic doesn’t decrease much if you continue with increasing it and strange enough the total driving time increases a bit sometime for higher radius t, because with a higher radius t you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is very high and therefore because we only use small graphs in our final implementation we don’t use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
Before reading and understanding this algorithm you have to read and understand the Brick algorithm. This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, we called this algorithm “Auction House Algorithm” because it is a little bit comparable with an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list:<br />
<br />
==== Definition List ====<br />
<ul><br />
<li>'''Route:''' A set of roads that will take you from your start location to your destination. </li><br />
<li>'''In Advance Best:''' A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etcetera). m is in this case some constant, which will will determine. </li><br />
<li> '''Capacity:''' The maximal amount of cars that can take a certain road each hour. </li><br />
<li> '''Claiming:''' If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road. </li><br />
<li> '''Neural:''' Neural(arg1, arg2, …, arg n) is the output of a Neural Network with n input arguments. </li><br />
</ul><br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 Lane: 1750 vehicles per hour </li><br />
<li> 2 Lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving/waiting time of a road is easily determined. We will assume that the spits (Dutch word for the moment that everyone needs to go to work etcetera) will have a duration of 2 hours. And that when x amount of people in the spits need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the the driving/waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location, destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<ul><br />
<li> '''Shortest path with road (a):''' The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams, etcetera into account. </li><br />
<li> '''Shortest path without road (b):''' The total driving time that someone needs to go from its location to their destination without taking traffic jams, etcetera into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.) </li><br />
<li> '''Own total time with road (c):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. </li><br />
<li> '''Own total time without road (d):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. Also given that the m in advance best routes might not use this road. </li><br />
<li> '''How much other claims (e):''' How much other start location/destination combinations make a claim on this road. </li><br />
<li> '''Sum shortest path others with road (f):''' The total sum of the shortest path with road for other start location/destination combinations (including other destinations from the same city). </li><br />
<li> '''Sum shortest path others without road (g):''' The total sum of the shortest path without road for other start location/destination combinations. </li><br />
<li> '''Sum total time others with road (h):''' The total sum of the own total time with road for other start location/destination combinations. </li><br />
<li> '''Sum total time others without road (i):''' The total sum of the own total time without road for other start location/destination combinations. </li><br />
<li> '''How much people claim the road (j):''' The total sum of all drivers that claim this road. </li><br />
<li> '''Capacity of a road (k):''' The maximal capacity of a road. </li><br />
</ul><br />
We will define: l = Neural(a, b, c, d, e, f, g, h, i, j, k). l is probably a value between 0 and 1 (because it is the output of a Neural Network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location/destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + … + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unlimited m? ====<br />
A good question is why not unlimit m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m doesn’t have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evoluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behavior of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behavior?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behavior of a solution, however don’t conclude that one of the solutions is bad because it haves different behavior. Because the total driving time might be the same for solutions with different behavior. To illustrate this we will show a graph of a 3D function: <br><br />
[[File:Different_Behavior_Same_Result.png]] <br><br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behavior for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like Tetraëder or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create. <br><br />
[[File:Brick Algorithm - Situation Representation.png|250px]]<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|150px]]<br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is: <br><br />
[[File:RandomGraph.png|300px]] <br><br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward): <br><br />
[[File:USE_-_Grid.png|250px]] <br><br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get: <br><br />
[[File:USE - RRGrid.png|250px]] <br><br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
=='''References'''==<br />
<references/></div>Wed, 05 Apr 2017 09:46:02 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37394
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37394<p>S143110: /* Traffic research */</p>
<hr />
<div>David Dekker - 0936100<br />
<br />
Haico Dorenbos - 0959563<br />
<br />
Jori van Falier - 0910959<br />
<br />
Tim Muyrers - 0907928<br />
<br />
Thomas Schouten - 0938927<br />
<br />
Sten Wessel - 0941508<br />
<br />
<br />
<br />
== Introduction ==<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people go to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these: <br />
1. The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
2. The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that can not be predicted. Because of all these problems we will be using self learning software to deal with these problems. This program will try to compensate for all the unknowns. <br />
<br />
== Objectives ==<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams.<br />
<br />
=== Maximize overall utility ===<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time, this is the main objective as mentioned above.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprises ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices available currently. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to a situation instead trying to avoid certain situations.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will distribute traffic equally over the roads available, which may seem like the optimal situation. A self learning system however should have more roads available, since it can predict traffic jams that haven't even happened yet. It can reroute the user many kilometers before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Idea visualize prof.jpg|300 px]]</div><br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously. The group of researchers will be split where four members of the team will work on making the self learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representable as possible. <br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external personnel to use and understand the way the system operates, and what conclusions it provides.<br />
<br />
== User and Society ==<br />
<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish. As long as people are driving their car (instead of autonomous), we can investigate further if/how we want to incorporate this. <br />
<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey. When looking at literature we find something called: “risk aversion”<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favorable bet (overall travel time decrease) when it means they can directly lose out (more travel time) <ref>Youtube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== Traffic research ==<br />
<br />
== What is a traffic jam ==<br />
<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can indentify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: Over Breaking<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans). We do see though that accidents can be massively decreased when we decrease the amount of cars driving at a road. The busier it is the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the riskfactor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsed1PNG.png|380 px]]</div><br />
Too many drivers on the road also causes direct traffic jams through over breaking. When a car in front breakers the one after it breaks too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill.<br />
<br />
Road works can’t really be avoided. But if the system is connected to a database, it should be able to take this into account.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self learning system should be able to.<br />
<br />
<br />
<br />
So to make sure the navigational system works in society you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. <br />
More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focussed on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be benicial.<br />
<br />
=== Traffic data ===<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by “Kennisinstituut voor Mobiliteitsbeleid (KiM)” from the ministry of infrastructure and milieu<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3.924 respondents from car drivers of the “Mobiliteitspanel Nederland (MPN)”. <br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought because of traffic information. Most of the information comes from the internet and the radio and during the travel also information panels above the road (DRIPs). From the questions follows that easy in use and quality of information are important reason for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long distant travels. The system is also used for points of interest. The most used points of interest are parking lot (75%), gas stations (49%) and places of entertainment (38%). Furthermore the system is used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of important follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route, heard or seen a traffic jam, or because of road works. Because our system will work optimal when everyone using the system, follows the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can proof that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors into choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Drive is used to host the questionnaire, because it has all features (sort of questions/ set-up) required for this research. <br />
<br />
The link below will bring you to the questionnaire*:<br />
<br />
https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform<br />
<br />
* Please note that by choosing certain answers it is possible that the questionaire forwards you and thus skipping certain sections. In order to see the full questionaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone recieving the questionaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non frequent drivers. While the system is designed for non selfdriving cars it would also work on autonomous cars. The system will first be tested and deployed in The Netherlands and thus this is also the target group for distributing the questionaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigational system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCS is designed to be applicable into society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing and whatever the system proposes it should still be faster then not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistleblower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this he/she is taking to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily lives. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account etc. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know way in advance when users are going to drive. This section asks of different occasions (different destinations) the time a user would be able/willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigational system. In this section the participants are asked about possible approaches the system could take regarding traffic/travel time distribution.<br />
<br />
A participant is free to not answer a question if he/she does not want to answer, without having to provide an reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representable participants (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well recieved with up to 95% of the participants showing interrest in the product<br />
* 72.4 % of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) vs sending the users on secondairy roads)<br />
* Point based system or reward for driving longer routes is preferred amongst the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travels ahead of time and inform the navigation system of said plans.<br />
* 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until 7th of April). Seven people responded to the English version and fifty-six filled in the English version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent. <br />
<br />
The introduction questions focused on introducing the product and for us the investigators to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answer are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9% respectively.<br />
<br />
So what does this mean. To start it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter device communication. As for environmental nuisances it’s hard to say how the device would be optimized at this point since the answers are divided amongst all four categories. <br />
<br />
The second question category tests whether or not participants are willing to follow the system in four situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally one would be the best. The preferred options are a point based system that will sent you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not defined) sounds appealing to a large part of the participants with both sections respectively scoring 81.9% and 73.8% on Very good or Good in comparison with the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made). <br />
<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
<br />
<br />
The next section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people this fall in performance will most likely be compensated by the extra information the system is able to receive from more users. 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be more investigated, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
<br />
In the beginning of the questionaire participants were asked on their point of view on the important of environmental factors influencing navigation systems functioning. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take polution/noise into account or not aswell as having a option to change this setting or no option.<br />
64% would like standard incorperation of polution into the system, with 72% prefering to have an option to change the setting the system is on standard. <br />
For noise these values are slightly lower with 60% wanting a standard incorperation of noise into the system, and a slightly higher amount of participants (76%) prefering to have a option to change noise incorperation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency etc), but will also need a incorperation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
<br />
Making system interaction easier is a strong aim in future planning of travels. For this a section of the questionnaire is design specifically to understand how participants/drivers would like to do this. The aim is to design one maximally two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants which is the mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car aswell a option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regard work travel as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Goede Graph user research.JPG|666 px]]</div><br />
As you can see in the graph above on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interrested in the system with all possible goals achieving over 50% of interrested participants. The three most popular goals are "The overal travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overal travel time is lower", "The overal travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product. What this means for the design mostly means that the goal can be customized as long as users dont lose out on travel time themselves. <br />
<br />
'''Final conclusion'''<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in depth interviews, the information collected through this limited resources survey give a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of things could not be specified. For example the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. Same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things can not be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the population. Furthermore the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point and reward systems are favoured by the users and have to be researched even more, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things can not be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
==== Introduction ====<br />
We start by introducing a very easy model for our problem. It is not very expanded and therefore the model doesn’t really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etcetera is taken into account ). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is completely static and ironically it doesn’t take time into account. Moreover we assume that everyone uses our clever navigation system and that everyone is obedient to our clever navigation system and therefore they will always follow their directions.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like this: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
The vertex that is two times circled is the destination, which means that the drivers from cities B, C and D have to go to this location (we don’t take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each road there are two edges, because you can take the driveway forward or the driveway backward (for destination A however there are none driveway backwards, because if you are at destination A then you don’t have to drive back). Every driveway in this graph is indicated with a number and every city is indicated with a letter.<br />
<br />
==== Data about roads and cities ====<br />
We will assume that 30.000 people live in city B who all need to go to A. We will furthermore assume that 20.000 people live in city C, who all need to go to city A. Moreover we will assume that 10.000 people live in city D, who all need to go to city A. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower then it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are huge highways (they can handle a lot of drivers), therefore they have the function 0.00025x + 15 (you need at least 15 minutes to drive over that certain road)</li><br />
<li>Roads 2, 4, 5, 6 and 7 are little, short roads (In dutch: N-wegen) (and they can only handle a small amount of drivers) and these roads have the function 0.0005x + 5 (you need at least 5 minutes to driver over that certain road).</li><br />
<li>Roads 8 en 9 are small, very short roads (you are soon at the new node by using this vertex) and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm. Suppose we have some driver that live in city C, which drives over the driveways 6, 5 and 1 (in this order) and finally arrives at his destination A. If this happens then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10.000 people have driven over road 1 and they have waited 17,5 minutes then the cost value for this road is 175.000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving” round a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This script will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit our script till 10 “driving” rounds for example (you can take a different number than 10). If after 10 “driving” rounds not everyone has arrived at destination A then it will get a very high cost value (and therefore this is automatically a very bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was for example to say that this value is equal to (Float_MaxValue) * (amount_drivers_at_start - amount_drivers_reached_destination) / (amount_drivers_at_start). This cost value has to be definitely higher than a cost value of a solution for which it is possible for everyone to arrive within 10 “driving” to destination A. If we define our cost value this way then a solution for which 10 “driving” rounds is too less to for everyone to get to their destination is worse than a solution in which everyone get to their destination in 10 “driving” rounds.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also for different times it can be used. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are not very logical (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time. Especially because this algorithm need to determine driveways * destinations * different_parts_of_the_day amount of constant. If you look for example at an area Germany and the Netherland then there is a huge amount of driveways and a huge amount of destinations. Hence you will have to determine a lot of constants.</li><br />
<li> Because of the reason before this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a sidenote, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
''Optimizing the distribution matrix''<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
''The program''<br />
<br />
The program, written in Python with Tensorflow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is very slow.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an expanded version of the “Brick Algorithm”. Therefore it is really important to have read the “Brick Algorithm” and understand how it works. This algorithm is an improve of the “Brick Algorithm” because it also work for a more larger scale. The idea of this algorithm is based on that traffic jams only occur at local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven.<br />
<br />
==== Situation Representation ====<br />
We are trying again to see the problem as a graph with nodes and vertices, likewise the “Brick Algorithm”. However this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm):<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 humans that need to go to A. If a city doesn’t have subscript A then 200 humans need to go to A from that city. For subscript G and J the same holds. Furthermore we assume that every black road has 1 lane and has a minimal time of 10 minutes. Every orange road has 2 lanes and has a minimal time of 20 minutes. And every green road has 3 lanes and has a minimal time of 30 minutes. Then we also have red road, which has a minimal time of 15 minutes and has 1 lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes for example. So every node that is in t minutes of a goal is taken into account. We take t is 30 minutes in this example and you might notice that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round:<br />
<br />
==== First iteration round ====<br />
In the first iteration round you will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within t minimal waiting time from its goal. You will discard every node in this example that is further than t time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example you will use the “Brick Algorithm” for three zones with subscript: A, G and J (note that zones doesn’t have to be disjunctive) and you will found the optimal solution for these zones. After that (same iteration round) you will take back the discarded nodes and vertices and take a look at humans that will travel to a goal and are outside of that zone (so for example someone from a city that hasn’t got subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travelers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as new inhabitants of that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account etc. This iteration is stopped after u iteration rounds have passed where we will determine u based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is, however if two goals are close to each other then it will not take the other goal into account, which might be bad. Therefore you might say that when you calculate the local solution for a certain goal that you also take other goals within a radius of 2*t into account, so you will use the “Brick Algorithm” for multiple goals at the same time. Then you discard the result of the other goals and you will use only the result of the goal that you were calculating. This algorithm will be called “ILDB+ Algorithm”.<br />
<br />
==== Defense of Assumption ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is very logical, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>If the assumption that traffic jams going to a certain location only occur close to that location is a very wrong assumption, then this algorithm doesn’t work (if it is partially true then this algorithm might still work).</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine ourselves how much iteration rounds are needed and what the value for t is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius t of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the + part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is runned is limited). You can also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius t the total driving time of all traffic doesn’t decrease much if you continue with increasing it and strange enough the total driving time increases a bit sometime for higher radius t, because with a higher radius t you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is very high and therefore because we only use small graphs in our final implementation we don’t use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
Before reading and understanding this algorithm you have to read and understand the Brick algorithm. This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, we called this algorithm “Auction House Algorithm” because it is a little bit comparable with an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list:<br />
<br />
==== Definition List ====<br />
<ul><br />
<li>'''Route:''' A set of roads that will take you from your start location to your destination. </li><br />
<li>'''In Advance Best:''' A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etcetera). m is in this case some constant, which will will determine. </li><br />
<li> '''Capacity:''' The maximal amount of cars that can take a certain road each hour. </li><br />
<li> '''Claiming:''' If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road. </li><br />
<li> '''Neural:''' Neural(arg1, arg2, …, arg n) is the output of a Neural Network with n input arguments. </li><br />
</ul><br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 Lane: 1750 vehicles per hour </li><br />
<li> 2 Lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving/waiting time of a road is easily determined. We will assume that the spits (Dutch word for the moment that everyone needs to go to work etcetera) will have a duration of 2 hours. And that when x amount of people in the spits need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the the driving/waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location, destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<ul><br />
<li> '''Shortest path with road (a):''' The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams, etcetera into account. </li><br />
<li> '''Shortest path without road (b):''' The total driving time that someone needs to go from its location to their destination without taking traffic jams, etcetera into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.) </li><br />
<li> '''Own total time with road (c):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. </li><br />
<li> '''Own total time without road (d):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. Also given that the m in advance best routes might not use this road. </li><br />
<li> '''How much other claims (e):''' How much other start location/destination combinations make a claim on this road. </li><br />
<li> '''Sum shortest path others with road (f):''' The total sum of the shortest path with road for other start location/destination combinations (including other destinations from the same city). </li><br />
<li> '''Sum shortest path others without road (g):''' The total sum of the shortest path without road for other start location/destination combinations. </li><br />
<li> '''Sum total time others with road (h):''' The total sum of the own total time with road for other start location/destination combinations. </li><br />
<li> '''Sum total time others without road (i):''' The total sum of the own total time without road for other start location/destination combinations. </li><br />
<li> '''How much people claim the road (j):''' The total sum of all drivers that claim this road. </li><br />
<li> '''Capacity of a road (k):''' The maximal capacity of a road. </li><br />
</ul><br />
We will define: l = Neural(a, b, c, d, e, f, g, h, i, j, k). l is probably a value between 0 and 1 (because it is the output of a Neural Network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location/destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + … + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unlimited m? ====<br />
A good question is why not unlimit m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m doesn’t have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evoluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behavior of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behavior?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behavior of a solution, however don’t conclude that one of the solutions is bad because it haves different behavior. Because the total driving time might be the same for solutions with different behavior. To illustrate this we will show a graph of a 3D function: <br><br />
[[File:Different_Behavior_Same_Result.png]] <br><br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behavior for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like Tetraëder or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create. <br><br />
[[File:Brick Algorithm - Situation Representation.png|250px]]<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|150px]]<br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is: <br><br />
[[File:RandomGraph.png|300px]] <br><br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward): <br><br />
[[File:USE_-_Grid.png|250px]] <br><br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get: <br><br />
[[File:USE - RRGrid.png|250px]] <br><br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
=='''References'''==<br />
<references/></div>Wed, 05 Apr 2017 09:45:43 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37385
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37385<p>S143110: /* What is a traffic jam */</p>
<hr />
<div>David Dekker - 0936100<br />
<br />
Haico Dorenbos - 0959563<br />
<br />
Jori van Falier - 0910959<br />
<br />
Tim Muyrers - 0907928<br />
<br />
Thomas Schouten - 0938927<br />
<br />
Sten Wessel - 0941508<br />
<br />
<br />
<br />
== Introduction ==<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people go to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these: <br />
1. The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
2. The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that can not be predicted. Because of all these problems we will be using self learning software to deal with these problems. This program will try to compensate for all the unknowns. <br />
<br />
== Objectives ==<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams.<br />
<br />
=== Maximize overall utility ===<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time, this is the main objective as mentioned above.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprises ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices available currently. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to a situation instead trying to avoid certain situations.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will distribute traffic equally over the roads available, which may seem like the optimal situation. A self learning system however should have more roads available, since it can predict traffic jams that haven't even happened yet. It can reroute the user many kilometers before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Idea visualize prof.jpg|300 px]]</div><br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously. The group of researchers will be split where four members of the team will work on making the self learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representable as possible. <br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external personnel to use and understand the way the system operates, and what conclusions it provides.<br />
<br />
== User and Society ==<br />
<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish. As long as people are driving their car (instead of autonomous), we can investigate further if/how we want to incorporate this. <br />
<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey. When looking at literature we find something called: “risk aversion”<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favorable bet (overall travel time decrease) when it means they can directly lose out (more travel time) <ref>Youtube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
=== Traffic research ===<br />
== What is a traffic jam ==<br />
<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can indentify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: Over Breaking<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans). We do see though that accidents can be massively decreased when we decrease the amount of cars driving at a road. The busier it is the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the riskfactor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsed1PNG.png|380 px]]</div><br />
Too many drivers on the road also causes direct traffic jams through over breaking. When a car in front breakers the one after it breaks too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill.<br />
<br />
Road works can’t really be avoided. But if the system is connected to a database, it should be able to take this into account.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self learning system should be able to.<br />
<br />
<br />
<br />
So to make sure the navigational system works in society you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. <br />
More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focussed on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be benicial.<br />
<br />
=== Traffic data ===<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by “Kennisinstituut voor Mobiliteitsbeleid (KiM)” from the ministry of infrastructure and milieu<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3.924 respondents from car drivers of the “Mobiliteitspanel Nederland (MPN)”. <br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought because of traffic information. Most of the information comes from the internet and the radio and during the travel also information panels above the road (DRIPs). From the questions follows that easy in use and quality of information are important reason for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long distant travels. The system is also used for points of interest. The most used points of interest are parking lot (75%), gas stations (49%) and places of entertainment (38%). Furthermore the system is used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of important follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route, heard or seen a traffic jam, or because of road works. Because our system will work optimal when everyone using the system, follows the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can proof that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors into choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Drive is used to host the questionnaire, because it has all features (sort of questions/ set-up) required for this research. <br />
<br />
The link below will bring you to the questionnaire*:<br />
<br />
https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform<br />
<br />
* Please note that by choosing certain answers it is possible that the questionaire forwards you and thus skipping certain sections. In order to see the full questionaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone recieving the questionaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non frequent drivers. While the system is designed for non selfdriving cars it would also work on autonomous cars. The system will first be tested and deployed in The Netherlands and thus this is also the target group for distributing the questionaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigational system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCS is designed to be applicable into society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing and whatever the system proposes it should still be faster then not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistleblower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this he/she is taking to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily lives. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account etc. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know way in advance when users are going to drive. This section asks of different occasions (different destinations) the time a user would be able/willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigational system. In this section the participants are asked about possible approaches the system could take regarding traffic/travel time distribution.<br />
<br />
A participant is free to not answer a question if he/she does not want to answer, without having to provide an reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representable participants (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well recieved with up to 95% of the participants showing interrest in the product<br />
* 72.4 % of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) vs sending the users on secondairy roads)<br />
* Point based system or reward for driving longer routes is preferred amongst the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travels ahead of time and inform the navigation system of said plans.<br />
* 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until 7th of April). Seven people responded to the English version and fifty-six filled in the English version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent. <br />
<br />
The introduction questions focused on introducing the product and for us the investigators to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answer are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9% respectively.<br />
<br />
So what does this mean. To start it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter device communication. As for environmental nuisances it’s hard to say how the device would be optimized at this point since the answers are divided amongst all four categories. <br />
<br />
The second question category tests whether or not participants are willing to follow the system in four situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally one would be the best. The preferred options are a point based system that will sent you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not defined) sounds appealing to a large part of the participants with both sections respectively scoring 81.9% and 73.8% on Very good or Good in comparison with the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made). <br />
<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
<br />
<br />
The next section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people this fall in performance will most likely be compensated by the extra information the system is able to receive from more users. 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be more investigated, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
<br />
In the beginning of the questionaire participants were asked on their point of view on the important of environmental factors influencing navigation systems functioning. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take polution/noise into account or not aswell as having a option to change this setting or no option.<br />
64% would like standard incorperation of polution into the system, with 72% prefering to have an option to change the setting the system is on standard. <br />
For noise these values are slightly lower with 60% wanting a standard incorperation of noise into the system, and a slightly higher amount of participants (76%) prefering to have a option to change noise incorperation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency etc), but will also need a incorperation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
<br />
Making system interaction easier is a strong aim in future planning of travels. For this a section of the questionnaire is design specifically to understand how participants/drivers would like to do this. The aim is to design one maximally two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants which is the mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car aswell a option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regard work travel as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Goede Graph user research.JPG|666 px]]</div><br />
As you can see in the graph above on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interrested in the system with all possible goals achieving over 50% of interrested participants. The three most popular goals are "The overal travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overal travel time is lower", "The overal travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product. What this means for the design mostly means that the goal can be customized as long as users dont lose out on travel time themselves. <br />
<br />
'''Final conclusion'''<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in depth interviews, the information collected through this limited resources survey give a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of things could not be specified. For example the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. Same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things can not be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the population. Furthermore the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point and reward systems are favoured by the users and have to be researched even more, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things can not be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
==== Introduction ====<br />
We start by introducing a very easy model for our problem. It is not very expanded and therefore the model doesn’t really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etcetera is taken into account ). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is completely static and ironically it doesn’t take time into account. Moreover we assume that everyone uses our clever navigation system and that everyone is obedient to our clever navigation system and therefore they will always follow their directions.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like this: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
The vertex that is two times circled is the destination, which means that the drivers from cities B, C and D have to go to this location (we don’t take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each road there are two edges, because you can take the driveway forward or the driveway backward (for destination A however there are none driveway backwards, because if you are at destination A then you don’t have to drive back). Every driveway in this graph is indicated with a number and every city is indicated with a letter.<br />
<br />
==== Data about roads and cities ====<br />
We will assume that 30.000 people live in city B who all need to go to A. We will furthermore assume that 20.000 people live in city C, who all need to go to city A. Moreover we will assume that 10.000 people live in city D, who all need to go to city A. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower then it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are huge highways (they can handle a lot of drivers), therefore they have the function 0.00025x + 15 (you need at least 15 minutes to drive over that certain road)</li><br />
<li>Roads 2, 4, 5, 6 and 7 are little, short roads (In dutch: N-wegen) (and they can only handle a small amount of drivers) and these roads have the function 0.0005x + 5 (you need at least 5 minutes to driver over that certain road).</li><br />
<li>Roads 8 en 9 are small, very short roads (you are soon at the new node by using this vertex) and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm. Suppose we have some driver that live in city C, which drives over the driveways 6, 5 and 1 (in this order) and finally arrives at his destination A. If this happens then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10.000 people have driven over road 1 and they have waited 17,5 minutes then the cost value for this road is 175.000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving” round a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This script will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit our script till 10 “driving” rounds for example (you can take a different number than 10). If after 10 “driving” rounds not everyone has arrived at destination A then it will get a very high cost value (and therefore this is automatically a very bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was for example to say that this value is equal to (Float_MaxValue) * (amount_drivers_at_start - amount_drivers_reached_destination) / (amount_drivers_at_start). This cost value has to be definitely higher than a cost value of a solution for which it is possible for everyone to arrive within 10 “driving” to destination A. If we define our cost value this way then a solution for which 10 “driving” rounds is too less to for everyone to get to their destination is worse than a solution in which everyone get to their destination in 10 “driving” rounds.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also for different times it can be used. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are not very logical (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time. Especially because this algorithm need to determine driveways * destinations * different_parts_of_the_day amount of constant. If you look for example at an area Germany and the Netherland then there is a huge amount of driveways and a huge amount of destinations. Hence you will have to determine a lot of constants.</li><br />
<li> Because of the reason before this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a sidenote, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
''Optimizing the distribution matrix''<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
''The program''<br />
<br />
The program, written in Python with Tensorflow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is very slow.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an expanded version of the “Brick Algorithm”. Therefore it is really important to have read the “Brick Algorithm” and understand how it works. This algorithm is an improve of the “Brick Algorithm” because it also work for a more larger scale. The idea of this algorithm is based on that traffic jams only occur at local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven.<br />
<br />
==== Situation Representation ====<br />
We are trying again to see the problem as a graph with nodes and vertices, likewise the “Brick Algorithm”. However this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm):<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 humans that need to go to A. If a city doesn’t have subscript A then 200 humans need to go to A from that city. For subscript G and J the same holds. Furthermore we assume that every black road has 1 lane and has a minimal time of 10 minutes. Every orange road has 2 lanes and has a minimal time of 20 minutes. And every green road has 3 lanes and has a minimal time of 30 minutes. Then we also have red road, which has a minimal time of 15 minutes and has 1 lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes for example. So every node that is in t minutes of a goal is taken into account. We take t is 30 minutes in this example and you might notice that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round:<br />
<br />
==== First iteration round ====<br />
In the first iteration round you will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within t minimal waiting time from its goal. You will discard every node in this example that is further than t time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example you will use the “Brick Algorithm” for three zones with subscript: A, G and J (note that zones doesn’t have to be disjunctive) and you will found the optimal solution for these zones. After that (same iteration round) you will take back the discarded nodes and vertices and take a look at humans that will travel to a goal and are outside of that zone (so for example someone from a city that hasn’t got subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travelers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as new inhabitants of that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account etc. This iteration is stopped after u iteration rounds have passed where we will determine u based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is, however if two goals are close to each other then it will not take the other goal into account, which might be bad. Therefore you might say that when you calculate the local solution for a certain goal that you also take other goals within a radius of 2*t into account, so you will use the “Brick Algorithm” for multiple goals at the same time. Then you discard the result of the other goals and you will use only the result of the goal that you were calculating. This algorithm will be called “ILDB+ Algorithm”.<br />
<br />
==== Defense of Assumption ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is very logical, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>If the assumption that traffic jams going to a certain location only occur close to that location is a very wrong assumption, then this algorithm doesn’t work (if it is partially true then this algorithm might still work).</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine ourselves how much iteration rounds are needed and what the value for t is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius t of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the + part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is runned is limited). You can also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius t the total driving time of all traffic doesn’t decrease much if you continue with increasing it and strange enough the total driving time increases a bit sometime for higher radius t, because with a higher radius t you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is very high and therefore because we only use small graphs in our final implementation we don’t use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
Before reading and understanding this algorithm you have to read and understand the Brick algorithm. This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, we called this algorithm “Auction House Algorithm” because it is a little bit comparable with an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list:<br />
<br />
==== Definition List ====<br />
<ul><br />
<li>'''Route:''' A set of roads that will take you from your start location to your destination. </li><br />
<li>'''In Advance Best:''' A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etcetera). m is in this case some constant, which will will determine. </li><br />
<li> '''Capacity:''' The maximal amount of cars that can take a certain road each hour. </li><br />
<li> '''Claiming:''' If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road. </li><br />
<li> '''Neural:''' Neural(arg1, arg2, …, arg n) is the output of a Neural Network with n input arguments. </li><br />
</ul><br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 Lane: 1750 vehicles per hour </li><br />
<li> 2 Lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving/waiting time of a road is easily determined. We will assume that the spits (Dutch word for the moment that everyone needs to go to work etcetera) will have a duration of 2 hours. And that when x amount of people in the spits need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the the driving/waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location, destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<ul><br />
<li> '''Shortest path with road (a):''' The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams, etcetera into account. </li><br />
<li> '''Shortest path without road (b):''' The total driving time that someone needs to go from its location to their destination without taking traffic jams, etcetera into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.) </li><br />
<li> '''Own total time with road (c):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. </li><br />
<li> '''Own total time without road (d):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. Also given that the m in advance best routes might not use this road. </li><br />
<li> '''How much other claims (e):''' How much other start location/destination combinations make a claim on this road. </li><br />
<li> '''Sum shortest path others with road (f):''' The total sum of the shortest path with road for other start location/destination combinations (including other destinations from the same city). </li><br />
<li> '''Sum shortest path others without road (g):''' The total sum of the shortest path without road for other start location/destination combinations. </li><br />
<li> '''Sum total time others with road (h):''' The total sum of the own total time with road for other start location/destination combinations. </li><br />
<li> '''Sum total time others without road (i):''' The total sum of the own total time without road for other start location/destination combinations. </li><br />
<li> '''How much people claim the road (j):''' The total sum of all drivers that claim this road. </li><br />
<li> '''Capacity of a road (k):''' The maximal capacity of a road. </li><br />
</ul><br />
We will define: l = Neural(a, b, c, d, e, f, g, h, i, j, k). l is probably a value between 0 and 1 (because it is the output of a Neural Network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location/destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + … + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unlimited m? ====<br />
A good question is why not unlimit m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m doesn’t have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evoluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behavior of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behavior?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behavior of a solution, however don’t conclude that one of the solutions is bad because it haves different behavior. Because the total driving time might be the same for solutions with different behavior. To illustrate this we will show a graph of a 3D function: <br><br />
[[File:Different_Behavior_Same_Result.png]] <br><br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behavior for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like Tetraëder or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create. <br><br />
[[File:Brick Algorithm - Situation Representation.png|250px]]<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|150px]]<br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is: <br><br />
[[File:RandomGraph.png|300px]] <br><br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward): <br><br />
[[File:USE_-_Grid.png|250px]] <br><br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get: <br><br />
[[File:USE - RRGrid.png|250px]] <br><br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
=='''References'''==<br />
<references/></div>Wed, 05 Apr 2017 09:33:34 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37376
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37376<p>S143110: /* Extra research */</p>
<hr />
<div>David Dekker - 0936100<br />
<br />
Haico Dorenbos - 0959563<br />
<br />
Jori van Falier - 0910959<br />
<br />
Tim Muyrers - 0907928<br />
<br />
Thomas Schouten - 0938927<br />
<br />
Sten Wessel - 0941508<br />
<br />
<br />
<br />
== Introduction ==<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people go to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these: <br />
1. The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
2. The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that can not be predicted. Because of all these problems we will be using self learning software to deal with these problems. This program will try to compensate for all the unknowns. <br />
<br />
== Objectives ==<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams.<br />
<br />
=== Maximize overall utility ===<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time, this is the main objective as mentioned above.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprises ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices available currently. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to a situation instead trying to avoid certain situations.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will distribute traffic equally over the roads available, which may seem like the optimal situation. A self learning system however should have more roads available, since it can predict traffic jams that haven't even happened yet. It can reroute the user many kilometers before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Idea visualize prof.jpg|300 px]]</div><br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously. The group of researchers will be split where four members of the team will work on making the self learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representable as possible. <br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external personnel to use and understand the way the system operates, and what conclusions it provides.<br />
<br />
== User and Society ==<br />
<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish. As long as people are driving their car (instead of autonomous), we can investigate further if/how we want to incorporate this. <br />
<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey. When looking at literature we find something called: “risk aversion”<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favorable bet (overall travel time decrease) when it means they can directly lose out (more travel time) <ref>Youtube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== What is a traffic jam ==<br />
<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can indentify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: Over Breaking<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans). We do see though that accidents can be massively decreased when we decrease the amount of cars driving at a road. The busier it is the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the riskfactor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsed1PNG.png|380 px]]</div><br />
Too many drivers on the road also causes direct traffic jams through over breaking. When a car in front breakers the one after it breaks too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill.<br />
<br />
Road works can’t really be avoided. But if the system is connected to a database, it should be able to take this into account.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self learning system should be able to.<br />
<br />
<br />
<br />
So to make sure the navigational system works in society you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. <br />
More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focussed on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be benicial.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by “Kennisinstituut voor Mobiliteitsbeleid (KiM)” from the ministry of infrastructure and milieu<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3.924 respondents from car drivers of the “Mobiliteitspanel Nederland (MPN)”. <br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought because of traffic information. Most of the information comes from the internet and the radio and during the travel also information panels above the road (DRIPs). From the questions follows that easy in use and quality of information are important reason for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long distant travels. The system is also used for points of interest. The most used points of interest are parking lot (75%), gas stations (49%) and places of entertainment (38%). Furthermore the system is used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of important follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route, heard or seen a traffic jam, or because of road works. Because our system will work optimal when everyone using the system, follows the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can proof that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors into choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Drive is used to host the questionnaire, because it has all features (sort of questions/ set-up) required for this research. <br />
<br />
The link below will bring you to the questionnaire*:<br />
<br />
https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform<br />
<br />
* Please note that by choosing certain answers it is possible that the questionaire forwards you and thus skipping certain sections. In order to see the full questionaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone recieving the questionaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non frequent drivers. While the system is designed for non selfdriving cars it would also work on autonomous cars. The system will first be tested and deployed in The Netherlands and thus this is also the target group for distributing the questionaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigational system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCS is designed to be applicable into society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing and whatever the system proposes it should still be faster then not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistleblower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this he/she is taking to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily lives. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account etc. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know way in advance when users are going to drive. This section asks of different occasions (different destinations) the time a user would be able/willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigational system. In this section the participants are asked about possible approaches the system could take regarding traffic/travel time distribution.<br />
<br />
A participant is free to not answer a question if he/she does not want to answer, without having to provide an reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representable participants (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well recieved with up to 95% of the participants showing interrest in the product<br />
* 72.4 % of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) vs sending the users on secondairy roads)<br />
* Point based system or reward for driving longer routes is preferred amongst the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travels ahead of time and inform the navigation system of said plans.<br />
* 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until 7th of April). Seven people responded to the English version and fifty-six filled in the English version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent. <br />
<br />
The introduction questions focused on introducing the product and for us the investigators to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answer are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9% respectively.<br />
<br />
So what does this mean. To start it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter device communication. As for environmental nuisances it’s hard to say how the device would be optimized at this point since the answers are divided amongst all four categories. <br />
<br />
The second question category tests whether or not participants are willing to follow the system in four situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally one would be the best. The preferred options are a point based system that will sent you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not defined) sounds appealing to a large part of the participants with both sections respectively scoring 81.9% and 73.8% on Very good or Good in comparison with the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made). <br />
<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
<br />
<br />
The next section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people this fall in performance will most likely be compensated by the extra information the system is able to receive from more users. 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be more investigated, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
<br />
In the beginning of the questionaire participants were asked on their point of view on the important of environmental factors influencing navigation systems functioning. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take polution/noise into account or not aswell as having a option to change this setting or no option.<br />
64% would like standard incorperation of polution into the system, with 72% prefering to have an option to change the setting the system is on standard. <br />
For noise these values are slightly lower with 60% wanting a standard incorperation of noise into the system, and a slightly higher amount of participants (76%) prefering to have a option to change noise incorperation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency etc), but will also need a incorperation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
<br />
Making system interaction easier is a strong aim in future planning of travels. For this a section of the questionnaire is design specifically to understand how participants/drivers would like to do this. The aim is to design one maximally two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants which is the mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car aswell a option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regard work travel as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Goede Graph user research.JPG|666 px]]</div><br />
As you can see in the graph above on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interrested in the system with all possible goals achieving over 50% of interrested participants. The three most popular goals are "The overal travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overal travel time is lower", "The overal travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product. What this means for the design mostly means that the goal can be customized as long as users dont lose out on travel time themselves. <br />
<br />
'''Final conclusion'''<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in depth interviews, the information collected through this limited resources survey give a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of things could not be specified. For example the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. Same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things can not be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the population. Furthermore the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point and reward systems are favoured by the users and have to be researched even more, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things can not be researched by us, because we do not know enough about the system. Once the system is able to work on a real life situation, people can be interviewed asking about noise disturbance, and more research can be done about the impact on the environment caused by pollution.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
==== Introduction ====<br />
We start by introducing a very easy model for our problem. It is not very expanded and therefore the model doesn’t really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etcetera is taken into account ). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is completely static and ironically it doesn’t take time into account. Moreover we assume that everyone uses our clever navigation system and that everyone is obedient to our clever navigation system and therefore they will always follow their directions.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like this: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
The vertex that is two times circled is the destination, which means that the drivers from cities B, C and D have to go to this location (we don’t take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each road there are two edges, because you can take the driveway forward or the driveway backward (for destination A however there are none driveway backwards, because if you are at destination A then you don’t have to drive back). Every driveway in this graph is indicated with a number and every city is indicated with a letter.<br />
<br />
==== Data about roads and cities ====<br />
We will assume that 30.000 people live in city B who all need to go to A. We will furthermore assume that 20.000 people live in city C, who all need to go to city A. Moreover we will assume that 10.000 people live in city D, who all need to go to city A. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower then it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are huge highways (they can handle a lot of drivers), therefore they have the function 0.00025x + 15 (you need at least 15 minutes to drive over that certain road)</li><br />
<li>Roads 2, 4, 5, 6 and 7 are little, short roads (In dutch: N-wegen) (and they can only handle a small amount of drivers) and these roads have the function 0.0005x + 5 (you need at least 5 minutes to driver over that certain road).</li><br />
<li>Roads 8 en 9 are small, very short roads (you are soon at the new node by using this vertex) and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm. Suppose we have some driver that live in city C, which drives over the driveways 6, 5 and 1 (in this order) and finally arrives at his destination A. If this happens then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10.000 people have driven over road 1 and they have waited 17,5 minutes then the cost value for this road is 175.000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving” round a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This script will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit our script till 10 “driving” rounds for example (you can take a different number than 10). If after 10 “driving” rounds not everyone has arrived at destination A then it will get a very high cost value (and therefore this is automatically a very bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was for example to say that this value is equal to (Float_MaxValue) * (amount_drivers_at_start - amount_drivers_reached_destination) / (amount_drivers_at_start). This cost value has to be definitely higher than a cost value of a solution for which it is possible for everyone to arrive within 10 “driving” to destination A. If we define our cost value this way then a solution for which 10 “driving” rounds is too less to for everyone to get to their destination is worse than a solution in which everyone get to their destination in 10 “driving” rounds.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also for different times it can be used. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are not very logical (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time. Especially because this algorithm need to determine driveways * destinations * different_parts_of_the_day amount of constant. If you look for example at an area Germany and the Netherland then there is a huge amount of driveways and a huge amount of destinations. Hence you will have to determine a lot of constants.</li><br />
<li> Because of the reason before this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a sidenote, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
''Optimizing the distribution matrix''<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
''The program''<br />
<br />
The program, written in Python with Tensorflow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is very slow.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an expanded version of the “Brick Algorithm”. Therefore it is really important to have read the “Brick Algorithm” and understand how it works. This algorithm is an improve of the “Brick Algorithm” because it also work for a more larger scale. The idea of this algorithm is based on that traffic jams only occur at local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven.<br />
<br />
==== Situation Representation ====<br />
We are trying again to see the problem as a graph with nodes and vertices, likewise the “Brick Algorithm”. However this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm):<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 humans that need to go to A. If a city doesn’t have subscript A then 200 humans need to go to A from that city. For subscript G and J the same holds. Furthermore we assume that every black road has 1 lane and has a minimal time of 10 minutes. Every orange road has 2 lanes and has a minimal time of 20 minutes. And every green road has 3 lanes and has a minimal time of 30 minutes. Then we also have red road, which has a minimal time of 15 minutes and has 1 lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes for example. So every node that is in t minutes of a goal is taken into account. We take t is 30 minutes in this example and you might notice that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round:<br />
<br />
==== First iteration round ====<br />
In the first iteration round you will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within t minimal waiting time from its goal. You will discard every node in this example that is further than t time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example you will use the “Brick Algorithm” for three zones with subscript: A, G and J (note that zones doesn’t have to be disjunctive) and you will found the optimal solution for these zones. After that (same iteration round) you will take back the discarded nodes and vertices and take a look at humans that will travel to a goal and are outside of that zone (so for example someone from a city that hasn’t got subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travelers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as new inhabitants of that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account etc. This iteration is stopped after u iteration rounds have passed where we will determine u based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is, however if two goals are close to each other then it will not take the other goal into account, which might be bad. Therefore you might say that when you calculate the local solution for a certain goal that you also take other goals within a radius of 2*t into account, so you will use the “Brick Algorithm” for multiple goals at the same time. Then you discard the result of the other goals and you will use only the result of the goal that you were calculating. This algorithm will be called “ILDB+ Algorithm”.<br />
<br />
==== Defense of Assumption ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is very logical, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>If the assumption that traffic jams going to a certain location only occur close to that location is a very wrong assumption, then this algorithm doesn’t work (if it is partially true then this algorithm might still work).</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine ourselves how much iteration rounds are needed and what the value for t is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius t of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the + part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is runned is limited). You can also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius t the total driving time of all traffic doesn’t decrease much if you continue with increasing it and strange enough the total driving time increases a bit sometime for higher radius t, because with a higher radius t you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is very high and therefore because we only use small graphs in our final implementation we don’t use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
Before reading and understanding this algorithm you have to read and understand the Brick algorithm. This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, we called this algorithm “Auction House Algorithm” because it is a little bit comparable with an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list:<br />
<br />
==== Definition List ====<br />
<ul><br />
<li>'''Route:''' A set of roads that will take you from your start location to your destination. </li><br />
<li>'''In Advance Best:''' A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etcetera). m is in this case some constant, which will will determine. </li><br />
<li> '''Capacity:''' The maximal amount of cars that can take a certain road each hour. </li><br />
<li> '''Claiming:''' If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road. </li><br />
<li> '''Neural:''' Neural(arg1, arg2, …, arg n) is the output of a Neural Network with n input arguments. </li><br />
</ul><br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 Lane: 1750 vehicles per hour </li><br />
<li> 2 Lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving/waiting time of a road is easily determined. We will assume that the spits (Dutch word for the moment that everyone needs to go to work etcetera) will have a duration of 2 hours. And that when x amount of people in the spits need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the the driving/waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location, destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<ul><br />
<li> '''Shortest path with road (a):''' The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams, etcetera into account. </li><br />
<li> '''Shortest path without road (b):''' The total driving time that someone needs to go from its location to their destination without taking traffic jams, etcetera into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.) </li><br />
<li> '''Own total time with road (c):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. </li><br />
<li> '''Own total time without road (d):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. Also given that the m in advance best routes might not use this road. </li><br />
<li> '''How much other claims (e):''' How much other start location/destination combinations make a claim on this road. </li><br />
<li> '''Sum shortest path others with road (f):''' The total sum of the shortest path with road for other start location/destination combinations (including other destinations from the same city). </li><br />
<li> '''Sum shortest path others without road (g):''' The total sum of the shortest path without road for other start location/destination combinations. </li><br />
<li> '''Sum total time others with road (h):''' The total sum of the own total time with road for other start location/destination combinations. </li><br />
<li> '''Sum total time others without road (i):''' The total sum of the own total time without road for other start location/destination combinations. </li><br />
<li> '''How much people claim the road (j):''' The total sum of all drivers that claim this road. </li><br />
<li> '''Capacity of a road (k):''' The maximal capacity of a road. </li><br />
</ul><br />
We will define: l = Neural(a, b, c, d, e, f, g, h, i, j, k). l is probably a value between 0 and 1 (because it is the output of a Neural Network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location/destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + … + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unlimited m? ====<br />
A good question is why not unlimit m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m doesn’t have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evoluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behavior of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behavior?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behavior of a solution, however don’t conclude that one of the solutions is bad because it haves different behavior. Because the total driving time might be the same for solutions with different behavior. To illustrate this we will show a graph of a 3D function: <br><br />
[[File:Different_Behavior_Same_Result.png]] <br><br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behavior for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like Tetraëder or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create. <br><br />
[[File:Brick Algorithm - Situation Representation.png|250px]]<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|150px]]<br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is: <br><br />
[[File:RandomGraph.png|300px]] <br><br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward): <br><br />
[[File:USE_-_Grid.png|250px]] <br><br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get: <br><br />
[[File:USE - RRGrid.png|250px]] <br><br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
=='''References'''==<br />
<references/></div>Wed, 05 Apr 2017 09:17:56 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37375
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37375<p>S143110: /* Extra research */</p>
<hr />
<div>David Dekker - 0936100<br />
<br />
Haico Dorenbos - 0959563<br />
<br />
Jori van Falier - 0910959<br />
<br />
Tim Muyrers - 0907928<br />
<br />
Thomas Schouten - 0938927<br />
<br />
Sten Wessel - 0941508<br />
<br />
<br />
<br />
== Introduction ==<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people go to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these: <br />
1. The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
2. The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that can not be predicted. Because of all these problems we will be using self learning software to deal with these problems. This program will try to compensate for all the unknowns. <br />
<br />
== Objectives ==<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams.<br />
<br />
=== Maximize overall utility ===<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time, this is the main objective as mentioned above.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprises ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices available currently. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to a situation instead trying to avoid certain situations.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will distribute traffic equally over the roads available, which may seem like the optimal situation. A self learning system however should have more roads available, since it can predict traffic jams that haven't even happened yet. It can reroute the user many kilometers before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Idea visualize prof.jpg|300 px]]</div><br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously. The group of researchers will be split where four members of the team will work on making the self learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representable as possible. <br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external personnel to use and understand the way the system operates, and what conclusions it provides.<br />
<br />
== User and Society ==<br />
<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish. As long as people are driving their car (instead of autonomous), we can investigate further if/how we want to incorporate this. <br />
<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey. When looking at literature we find something called: “risk aversion”<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favorable bet (overall travel time decrease) when it means they can directly lose out (more travel time) <ref>Youtube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== What is a traffic jam ==<br />
<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can indentify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: Over Breaking<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans). We do see though that accidents can be massively decreased when we decrease the amount of cars driving at a road. The busier it is the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the riskfactor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsed1PNG.png|380 px]]</div><br />
Too many drivers on the road also causes direct traffic jams through over breaking. When a car in front breakers the one after it breaks too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill.<br />
<br />
Road works can’t really be avoided. But if the system is connected to a database, it should be able to take this into account.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self learning system should be able to.<br />
<br />
<br />
<br />
So to make sure the navigational system works in society you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. <br />
More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focussed on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be benicial.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by “Kennisinstituut voor Mobiliteitsbeleid (KiM)” from the ministry of infrastructure and milieu<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3.924 respondents from car drivers of the “Mobiliteitspanel Nederland (MPN)”. <br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought because of traffic information. Most of the information comes from the internet and the radio and during the travel also information panels above the road (DRIPs). From the questions follows that easy in use and quality of information are important reason for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long distant travels. The system is also used for points of interest. The most used points of interest are parking lot (75%), gas stations (49%) and places of entertainment (38%). Furthermore the system is used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of important follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route, heard or seen a traffic jam, or because of road works. Because our system will work optimal when everyone using the system, follows the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can proof that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors into choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Drive is used to host the questionnaire, because it has all features (sort of questions/ set-up) required for this research. <br />
<br />
The link below will bring you to the questionnaire*:<br />
<br />
https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform<br />
<br />
* Please note that by choosing certain answers it is possible that the questionaire forwards you and thus skipping certain sections. In order to see the full questionaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone recieving the questionaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non frequent drivers. While the system is designed for non selfdriving cars it would also work on autonomous cars. The system will first be tested and deployed in The Netherlands and thus this is also the target group for distributing the questionaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigational system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCS is designed to be applicable into society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing and whatever the system proposes it should still be faster then not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistleblower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this he/she is taking to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily lives. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account etc. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know way in advance when users are going to drive. This section asks of different occasions (different destinations) the time a user would be able/willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigational system. In this section the participants are asked about possible approaches the system could take regarding traffic/travel time distribution.<br />
<br />
A participant is free to not answer a question if he/she does not want to answer, without having to provide an reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representable participants (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well recieved with up to 95% of the participants showing interrest in the product<br />
* 72.4 % of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) vs sending the users on secondairy roads)<br />
* Point based system or reward for driving longer routes is preferred amongst the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travels ahead of time and inform the navigation system of said plans.<br />
* 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until 7th of April). Seven people responded to the English version and fifty-six filled in the English version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent. <br />
<br />
The introduction questions focused on introducing the product and for us the investigators to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answer are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9% respectively.<br />
<br />
So what does this mean. To start it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter device communication. As for environmental nuisances it’s hard to say how the device would be optimized at this point since the answers are divided amongst all four categories. <br />
<br />
The second question category tests whether or not participants are willing to follow the system in four situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally one would be the best. The preferred options are a point based system that will sent you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not defined) sounds appealing to a large part of the participants with both sections respectively scoring 81.9% and 73.8% on Very good or Good in comparison with the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made). <br />
<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
<br />
<br />
The next section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people this fall in performance will most likely be compensated by the extra information the system is able to receive from more users. 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be more investigated, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
<br />
In the beginning of the questionaire participants were asked on their point of view on the important of environmental factors influencing navigation systems functioning. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take polution/noise into account or not aswell as having a option to change this setting or no option.<br />
64% would like standard incorperation of polution into the system, with 72% prefering to have an option to change the setting the system is on standard. <br />
For noise these values are slightly lower with 60% wanting a standard incorperation of noise into the system, and a slightly higher amount of participants (76%) prefering to have a option to change noise incorperation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency etc), but will also need a incorperation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
<br />
Making system interaction easier is a strong aim in future planning of travels. For this a section of the questionnaire is design specifically to understand how participants/drivers would like to do this. The aim is to design one maximally two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants which is the mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car aswell a option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regard work travel as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Goede Graph user research.JPG|666 px]]</div><br />
As you can see in the graph above on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interrested in the system with all possible goals achieving over 50% of interrested participants. The three most popular goals are "The overal travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overal travel time is lower", "The overal travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product. What this means for the design mostly means that the goal can be customized as long as users dont lose out on travel time themselves. <br />
<br />
'''Final conclusion'''<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in depth interviews, the information collected through this limited resources survey give a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of things could not be specified. For example the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. Same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things can not be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the population. Furthermore the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point and reward systems are favoured by the users and have to be researched even more, looking into the type and size of the reward, and the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things can not be researched by us, because we do not know enough about the system.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
==== Introduction ====<br />
We start by introducing a very easy model for our problem. It is not very expanded and therefore the model doesn’t really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etcetera is taken into account ). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is completely static and ironically it doesn’t take time into account. Moreover we assume that everyone uses our clever navigation system and that everyone is obedient to our clever navigation system and therefore they will always follow their directions.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like this: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
The vertex that is two times circled is the destination, which means that the drivers from cities B, C and D have to go to this location (we don’t take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each road there are two edges, because you can take the driveway forward or the driveway backward (for destination A however there are none driveway backwards, because if you are at destination A then you don’t have to drive back). Every driveway in this graph is indicated with a number and every city is indicated with a letter.<br />
<br />
==== Data about roads and cities ====<br />
We will assume that 30.000 people live in city B who all need to go to A. We will furthermore assume that 20.000 people live in city C, who all need to go to city A. Moreover we will assume that 10.000 people live in city D, who all need to go to city A. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower then it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are huge highways (they can handle a lot of drivers), therefore they have the function 0.00025x + 15 (you need at least 15 minutes to drive over that certain road)</li><br />
<li>Roads 2, 4, 5, 6 and 7 are little, short roads (In dutch: N-wegen) (and they can only handle a small amount of drivers) and these roads have the function 0.0005x + 5 (you need at least 5 minutes to driver over that certain road).</li><br />
<li>Roads 8 en 9 are small, very short roads (you are soon at the new node by using this vertex) and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm. Suppose we have some driver that live in city C, which drives over the driveways 6, 5 and 1 (in this order) and finally arrives at his destination A. If this happens then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10.000 people have driven over road 1 and they have waited 17,5 minutes then the cost value for this road is 175.000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving” round a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This script will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit our script till 10 “driving” rounds for example (you can take a different number than 10). If after 10 “driving” rounds not everyone has arrived at destination A then it will get a very high cost value (and therefore this is automatically a very bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was for example to say that this value is equal to (Float_MaxValue) * (amount_drivers_at_start - amount_drivers_reached_destination) / (amount_drivers_at_start). This cost value has to be definitely higher than a cost value of a solution for which it is possible for everyone to arrive within 10 “driving” to destination A. If we define our cost value this way then a solution for which 10 “driving” rounds is too less to for everyone to get to their destination is worse than a solution in which everyone get to their destination in 10 “driving” rounds.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also for different times it can be used. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are not very logical (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time. Especially because this algorithm need to determine driveways * destinations * different_parts_of_the_day amount of constant. If you look for example at an area Germany and the Netherland then there is a huge amount of driveways and a huge amount of destinations. Hence you will have to determine a lot of constants.</li><br />
<li> Because of the reason before this algorithm will also take a lot of memory for larger areas.</li><br />
</ul><br />
<br />
=== Matrix Algorithm (A first attempt at an implementation) ===<br />
This is a way of optimizing to a very specific problem, with a given situation. It optimizes for that situation, so it will never be a real, general solution. It works by changing the input to something better, instead of learning how to convert any input to a certain solution.<br />
<br />
Please remember that the problem solved here is mainly to get a feel for both the programming language and the type of problem. We know this problem is so simple you could easily solve it by other means, with much less effort, but that is a good way of checking our own results.<br />
<br />
We assume there are two cities, A and B, and two roads from A to B. We assume cars leave in driving rounds, so in each round there is a certain amount of cars on the road. For example, a typical distribution of cars over one road would be [12, 3] which means that the first round 12 cars leave city A, and the second round 3 cars do so.<br />
<br />
Of each road we know exactly how much time a car needs to travel the road given how much cars are on the road in total. This is a function which we will call the travel time function. We assume here it is of the form max(a * cars - b, min_travel_time), which means that up to a certain amount of cars on the road the travel time equals the minimum travel time needed to traverse the road, and above that the travel time increases linearly. The latter case is of course meant to signify a traffic jam.<br />
<br />
One road is a highway, and the other road a secondary road which means that on the secondary road the minimal travel time is larger, a traffic jam occurs with less cars, and increases more per extra car. This is easily extendable with more roads, but for every road you would have to know this travel time function.<br />
<br />
So for two roads together we have again a distribution matrix, for example [[8, 3], [3, 1]]. The idea is that we predetermine the amount of driving rounds, because changing the size of this matrix while optimizing is a bit of a problem still. (As a sidenote, in real life you would probably want to limit this because of user preferences anyway) <br />
<br />
''Optimizing the distribution matrix''<br />
<br />
We calculate the cost (travel time) of a certain distribution matrix by adding up the travel times of all the cars. This is the value that is going to be minimized. For the cars leaving the first round, the travel time for each car is given by the travel time function. For the cars in the second round, the same holds but we add the time that they had to wait while the first round was driving. We continue this for how many rounds there are. Then, if there are still cars left after all the rounds, we assume all these cars depart in the last round.<br />
<br />
The program optimizes this matrix by minimizing the costs. Therefore, if there are sufficiently many rounds, no cars will be left behind, because if the number of rounds goes to infinity the cost of leaving cars behind will go to infinity as well.<br />
<br />
''The program''<br />
<br />
The program, written in Python with Tensorflow, has as input the matrix initialized with some values which should not matter too much. It then uses a Gradient Descent Optimizer to minimize the cost. The optimizer is run for a certain amount of steps with a certain learning rate, and then it outputs the optimized distribution matrix.<br />
<br />
''Drawbacks of this approach''<br />
*We have to predetermine the amount of driving rounds.<br />
*We have to know for each road the travel time function.<br />
*This generalizes and scales badly, because the matrix will get very large, and you need a matrix for every possible trajectory between two cities.<br />
<br />
''Problems with the current program''<br />
*It is very slow.<br />
*When normalizing the distribution matrix values to a fraction of the total amount of cars instead, it should be more stable because it does not have to change values to whatever big amount of cars there may be so the learning rate can be smaller, but instead the results get very unstable.<br />
*Currently when on the first round each cars take for example ten minutes to clear the road, the next round can depart after one minute if that is the minimal travel time. Changing this leads to unstable behaviour.<br />
<br />
=== ILDB(+) Algorithm ===<br />
==== Introduction ====<br />
ILDB Algorithm stands for “Iterating Limited Distance Brick Algorithm”. This algorithm is an expanded version of the “Brick Algorithm”. Therefore it is really important to have read the “Brick Algorithm” and understand how it works. This algorithm is an improve of the “Brick Algorithm” because it also work for a more larger scale. The idea of this algorithm is based on that traffic jams only occur at local scale, which means that if a lot of traffic has to go to Eindhoven then the traffic jams occur close to Eindhoven.<br />
<br />
==== Situation Representation ====<br />
We are trying again to see the problem as a graph with nodes and vertices, likewise the “Brick Algorithm”. However this time the graph is more complex (to illustrate the difference between the “Brick Algorithm” and this algorithm):<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|thumb]]<br />
In this example A, G and J are goals. And to make the situation less complex we will assume that cities with subscript A have 2000 humans that need to go to A. If a city doesn’t have subscript A then 200 humans need to go to A from that city. For subscript G and J the same holds. Furthermore we assume that every black road has 1 lane and has a minimal time of 10 minutes. Every orange road has 2 lanes and has a minimal time of 20 minutes. And every green road has 3 lanes and has a minimal time of 30 minutes. Then we also have red road, which has a minimal time of 15 minutes and has 1 lane.<br />
<br />
==== Calculation Waiting Time ====<br />
The calculation of the waiting time for roads is the same as the Auction House algorithm.<br />
<br />
==== How does the algorithm works? ====<br />
In the “Bricks Algorithm” you have different driving rounds. In this algorithm you get some new type of round called iteration round. We first take a look at a certain radius from every goal by only looking at minimal time and we take 30 minutes for example. So every node that is in t minutes of a goal is taken into account. We take t is 30 minutes in this example and you might notice that every city that has a subscript of that certain goal is within 30 minutes of that goal. Then we start with the first iteration round:<br />
<br />
==== First iteration round ====<br />
In the first iteration round you will use the “Brick Algorithm” for every goal separately and for everyone that has to go to his goal and is within t minimal waiting time from its goal. You will discard every node in this example that is further than t time from its goal and every road that goes to a discarded node or comes from a discarded node is also discarded. So in this example you will use the “Brick Algorithm” for three zones with subscript: A, G and J (note that zones doesn’t have to be disjunctive) and you will found the optimal solution for these zones. After that (same iteration round) you will take back the discarded nodes and vertices and take a look at humans that will travel to a goal and are outside of that zone (so for example someone from a city that hasn’t got subscript A wants to travel to A). In this case you will use the shortest path algorithm to that certain zone taken into account the new waiting time due to the traffic jams that drivers within a zone have created. If travelers from outside the zone enter the zone with their shortest path algorithm, they will use the “Brick Algorithm” to get to their destination.<br />
<br />
==== Other iteration rounds ====<br />
The second iteration round is executed after the first iteration round is executed for every goal zone and the second iteration round will check where the drivers from outside the zone have entered the zone in the first iteration round and will take them into account as new inhabitants of that node that have to go to that certain goal for the “Brick Algorithm”. So suppose 500 outsiders have entered zone J by going to H, then in the “Brick Algorithm” drivers that have to go from H to J is increased by 500. After this a third iteration round is executed and will take the results of the second iteration round into account etc. This iteration is stopped after u iteration rounds have passed where we will determine u based on the calculation time of this algorithm.<br />
<br />
==== What if two goals are close to each other? ====<br />
This algorithm calculate separately for each goal what the optimal local solution is, however if two goals are close to each other then it will not take the other goal into account, which might be bad. Therefore you might say that when you calculate the local solution for a certain goal that you also take other goals within a radius of 2*t into account, so you will use the “Brick Algorithm” for multiple goals at the same time. Then you discard the result of the other goals and you will use only the result of the goal that you were calculating. This algorithm will be called “ILDB+ Algorithm”.<br />
<br />
==== Defense of Assumption ====<br />
The ILDB Algorithm is based on the assumption that traffic jams for traffic going to a certain location only occur close to that location. This assumption is very logical, because if for example there is some event going in Amsterdam and a lot of drivers from Berlin go to Amsterdam then no traffic jam will occur in Germany because of the drivers that will go to Amsterdam (traffic jams in Germany might happen, but they have a different cause).<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It can also solve the problem for larger problems more efficiently than the “Brick Algorithm”.</li><br />
<li>It needs less memory than the “Brick Algorithm” to solve the problem.</li><br />
<li> It doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem. </li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>It is harder to see if this algorithm is complete and always find the most optimal solution than for the “Brick Algorithm”, because a lot of path distributions are discarded in this solution.</li><br />
<li>If the assumption that traffic jams going to a certain location only occur close to that location is a very wrong assumption, then this algorithm doesn’t work (if it is partially true then this algorithm might still work).</li><br />
<li> This algorithm is more complex than the “Brick Algorithm” and therefore harder to understand and implement. </li><br />
<li> In this algorithm we must determine ourselves how much iteration rounds are needed and what the value for t is. </li><br />
</ul><br />
<br />
==== Results ====<br />
The ILDB algorithm has been implemented in Java and has been tested. A bit changes were made to make this algorithm run more faster and make the code less complex to make and to understand. These changes were:<br />
<ul><br />
<li>Set the amount of iteration rounds to 1 instead of some other constant higher than 1. </li><br />
<li> Only take local traffic into account when training. So traffic that is within radius t of that goal and need to go to that goal is taken into account. Therefore traffic that needs to go to the same goal that are outside radius t and traffic that needs to a different goal is not taken into account when training (but is used afterwards to show the total driving time for all drivers given a certain solution). </li><br />
<li> The normal part of this algorithm is implemented, so the + part is not implemented. </li><br />
<li> Random graphs were used with one goal to test the efficiency of this algorithm. </li><br />
</ul><br />
The time complexity of this algorithm to train is constant time compared to the amount of nodes and vertices in the graph that are not goals (because the distance in which the Brick Algorithm is runned is limited). You can also see why this algorithm is very efficient compared to other algorithms. For a graph with 120 nodes and 180 vertices and 1 goal it takes about 6 minutes and 29 seconds to run. It is a bit harder in this case to see why the solution the algorithm returns is the best one. However you can see that for a certain radius t the total driving time of all traffic doesn’t decrease much if you continue with increasing it and strange enough the total driving time increases a bit sometime for higher radius t, because with a higher radius t you have a lot more distribution constants that have to be determined for different roads and therefore you have more local minima in which this algorithm get stuck. All considered we have decided to not use this algorithm for our final implementation. This might sound strange because other algorithms have a higher time complexity than constant time, however the constant time in this algorithm is very high and therefore because we only use small graphs in our final implementation we don’t use this algorithm for our final implementation.<br />
<br />
== High Level Implementation Ideas ==<br />
=== Auction House Algorithm ===<br />
==== Introduction ====<br />
Before reading and understanding this algorithm you have to read and understand the Brick algorithm. This is the first algorithm that uses a Neural Network, therefore this algorithm might be hard to understand and difficult to read. So let's start with the algorithm, we called this algorithm “Auction House Algorithm” because it is a little bit comparable with an auction. Every road that you might take to get to your final destination, will be auctioned to everyone that might as well take the same road. An auction has some similarities to this algorithm. In an auction you will bid on something that has some value for you, but when the bid of someone is much higher than you wanted to offer, then it might be a good idea to stop with bidding any higher, because it is most likely the case that the item that is being auctioned is more valuable for the other than for you. Something similar happens in this algorithm, we will take a look at a certain road and check how much it is worth for a certain city. This value will be compared with the value of this road for other cities. This might sound very complex, therefore we will start by introducing a definition list:<br />
<br />
==== Definition List ====<br />
<ul><br />
<li>'''Route:''' A set of roads that will take you from your start location to your destination. </li><br />
<li>'''In Advance Best:''' A route is in advance best if it is the shortest route to your destination, without taking traffic jams and other things that might happen in the future into account. When we talk about m in advance best routes then we talk about the best m in advance routes (so the in advance best route, the second in advance best route, etcetera). m is in this case some constant, which will will determine. </li><br />
<li> '''Capacity:''' The maximal amount of cars that can take a certain road each hour. </li><br />
<li> '''Claiming:''' If in the m in advance best routes from a city to a destination, you will take a certain road then that city is claiming that road. </li><br />
<li> '''Neural:''' Neural(arg1, arg2, …, arg n) is the output of a Neural Network with n input arguments. </li><br />
</ul><br />
<br />
==== Situation Representation ====<br />
The situation representation is the same as the Brick Algorithm: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
<br />
==== Data about roads and cities ====<br />
The data about the cities will stay the same, however the data about the roads will change. We will assume that there is a maximum capacity for every road. And this is the maximum capacity for these kind of roads:<ref>Wegenwiki (2015). ''Capaciteit.'' Retrieved March, 2017 from https://www.wegenwiki.nl/Capaciteit</ref><br />
<ul><br />
<li> 1 Lane: 1750 vehicles per hour </li><br />
<li> 2 Lanes: 4200 vehicles per hour </li><br />
</ul><br />
We will define that roads 1 and 3 have 2 lanes. All the other roads only have 1 lane. The minimal driving time will stay the same:<br />
<ul><br />
<li> Roads 1 and 3 have a minimal driving time of 15 minutes. </li><br />
<li> Roads 8 and 9 have a minimal driving time of 3 minutes. </li><br />
<li> All other roads have a minimal driving time of 5 minutes. </li><br />
</ul><br />
The driving/waiting time of a road is easily determined. We will assume that the spits (Dutch word for the moment that everyone needs to go to work etcetera) will have a duration of 2 hours. And that when x amount of people in the spits need to go over a certain road, then the demand for that road each hour is x / 2. When x / 2 is smaller than the maximal capacity or equal, the the driving/waiting time is equal to the minimal driving time of that road. When x / 2 is larger the maximal capacity of that road then the driving/waiting time is determined by: (x * minimal driving time) / (maximum capacity * 2)<br />
<br />
==== Determine the value of a road ====<br />
From every start location, destination combination we will take the m in advance best routes. For every start location, destination combination you can determine what the value is for a certain road which will be claimed by that start location, destination combination. This will be done by assuming that the road might not be taken to go to your destination. To determine that value we will first define some other values:<br />
<ul><br />
<li> '''Shortest path with road (a):''' The total driving time that someone needs to go from its location to their destination with the in advance best route without taking traffic jams, etcetera into account. </li><br />
<li> '''Shortest path without road (b):''' The total driving time that someone needs to go from its location to their destination without taking traffic jams, etcetera into account, using the in advance best route, given that the route might not use this road. (If there is no possible route anymore to the destination then this value is equal to infinity. In the case you use a limit value, then this value is equal to the limit value instead of infinity.) </li><br />
<li> '''Own total time with road (c):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. </li><br />
<li> '''Own total time without road (d):''' The total sum of the times of different routes that someone from a city needs when using the m in advance best routes without taking traffic jams, etcetera into account. Also given that the m in advance best routes might not use this road. </li><br />
<li> '''How much other claims (e):''' How much other start location/destination combinations make a claim on this road. </li><br />
<li> '''Sum shortest path others with road (f):''' The total sum of the shortest path with road for other start location/destination combinations (including other destinations from the same city). </li><br />
<li> '''Sum shortest path others without road (g):''' The total sum of the shortest path without road for other start location/destination combinations. </li><br />
<li> '''Sum total time others with road (h):''' The total sum of the own total time with road for other start location/destination combinations. </li><br />
<li> '''Sum total time others without road (i):''' The total sum of the own total time without road for other start location/destination combinations. </li><br />
<li> '''How much people claim the road (j):''' The total sum of all drivers that claim this road. </li><br />
<li> '''Capacity of a road (k):''' The maximal capacity of a road. </li><br />
</ul><br />
We will define: l = Neural(a, b, c, d, e, f, g, h, i, j, k). l is probably a value between 0 and 1 (because it is the output of a Neural Network). The value of this road is equal to: [[File:Formula_Road_Value.jpg|100px]]<br />
<br />
==== Why such complex function? ====<br />
Because this function has nice properties:<br />
<ul><br />
<li> When l goes to 0 then this function converges to 0. </li><br />
<li> When l goes to 1 then this function diverges to infinity. </li><br />
<li> [[File:Formula_Neutralize.jpg|250px]] if a = 0.5 – x and b = 0.5 + x (so opposites neutralize each other) </li><br />
<li> [[File:Formula_Middle_One.jpg|130px]] if l = 0.5 </li><br />
</ul><br />
<br />
==== How does the algorithm furthermore work? ====<br />
For each start location/destination combination a value is assigned the m in advance routes. This value is calculated by multiplying all values of the roads which this route takes. This way you get a r1 value for the in advance best route, r2 for the second in advance best route, etc. Moreover you will distribute different citizens in a city based on their destination over different routes. Route i will take ri / (r1 + … + rm) relative amount of the drivers from the city that need to go to that destination.<br />
<br />
==== Why not an unlimited m? ====<br />
A good question is why not unlimit m in taking the in advance best m routes into account, because taking every route into account is better than taking not every route into account. This is done to limit the running time of our algorithm. For example if you take a look at all possible routes from Berlin to Amsterdam then it takes ages for our program to find all possible routes from Berlin to Amsterdam, because our script then also takes routes into account that go through Italy. Therefore you must limit the amount of routes that is taken into account by a certain constant m, because it is not very likely that the best path from Berlin to Amsterdam goes through Italy, not even when you try to prevent traffic jams. Thus logically the route you must take to prevent traffic jams and to go to your destination is within the in advance best m routes.<br />
This m doesn’t have to be an independent constant, but could for example also be dependent on the distance between your location and your destination. However we are not sure how to do this. So we take an independent m that is large enough to get the best solution, but small enough to get a good running time.<br />
<br />
== Further thoughts on the problem ==<br />
<br />
Solving the problem in a ‘classical’ reinforcement learning way, would be defining actions like delaying cars (sending them away later) or sending cars over different roads, and then define a network which attaches probabilities to these actions. Then you train it with quite some situations, and then it develops a policy of probabilities of when to send cars. Basically it learns what to do with all the cars that want to leave.<br />
<br />
We want to use the NEAT algorithm, in the form of the NEAT-Python library. In very short terms, this algorithm evoluates a neural network with an evolutionary algorithm, and therefore you do not have to design such a neural network yourself.<br />
<br />
We want to find out how we can solve this reinforcement learning problem with NEAT-Python.<br />
<br />
== NEAT Algorithm ==<br />
The reinforcement learning approach enables us to not only get a solution to the problem of rerouting traffic, but also gives us part of the solution strategy: the weights of the neural network connections. Thus, we do not have to determine the optimal network weights in the network beforehand. For solving the traffic rerouting problem, this is essential. Because of the many input factors involved for which the effects and their influence thereon are unknown or uncertain, we cannot possible construct a good model for this system.<br />
<br />
To apply reinforcement learning, we need a neural network to manipulate the weights for to optimize its performance. Thus, we need to construct the topology of the network beforehand. Conducting a study in different variants of neural network topologies gives a lot of options to choose from, each applicable for a specific problem domain. See for instance The Neural Network Zoo by the Asimov Institute<ref>The Asimov Institute (2016). The Neural Network Zoo. Retrieved March, 2017 from http://www.asimovinstitute.org/neural-network-zoo/</ref> for an extensive list of possibilities. Determining a suitable topology for our problem is challenging, because of the unknown underlying structure and the scope of relevant factors. However, if it is possible to evolve network weights, why would it not be possible to evolve network topology? We turn to the NEAT algorithm, short for NeuroEvolution of Augmenting Topologies, specifically designed to do that. It was originally developed Ken Stanley in 2002<ref>Stanley, K.O. & Miikkulainen, R. ''Efficient Evolution of Neural Network Topologies.'' Austin, The University of Texas. Retrieved March, 2017 from http://nn.cs.utexas.edu/downloads/papers/stanley.cec02.pdf</ref> and modified and extended over the last couple of years.<br />
<br />
=== Double Pole Balancing ===<br />
To understand how this works, we introduce a famously investigated problem in the area of artificial intelligence: the double pole balancing problem. This problem has been widely used as a measure for the performance of an artificial intelligent system. It involves a cart moving in a bounded 1-dimensional space (either to the left or right) on which two poles of a different length are fixed at one end at the same point on the cart (see figure).<br />
<br />
In the initial position, the poles stand upright and the cart has a nonzero initial velocity. The goal is to keep the poles balanced or prevent it to fall by exerting a force on the cart to make it move. The state of the system on any given point in time is described by:<br />
* The angular position of the poles \theta_{1,2};<br />
* The angular velocity of the poles \dot\theta_{1,2};<br />
* The position of the cart on the track x, which is bounded within [0, X];<br />
* The velocity of the cart \dot x;<br />
* Acting force on the cart F.<br />
<br />
These values are known to the system and can be used to determine what force to use next on the cart to keep the poles balanced. The time it is able to, within the bounds of the track, balance the poles is used as a performance measure.<br />
<br />
When the NEAT algorithm was developed, it showed that it could solve this problem more effectively and efficiently than that has been done before with other approaches, showing that it had potential for solving more complex problems and was distinct from other already existing approaches.<br />
<br />
=== Algorithm Description ===<br />
<br />
The main idea of the NEAT algorithm is that it combines a genetic evolution strategy with neural networks. Both genetic algorithms and neural networks have been used for solving problems where the behavior of the system investigated is not fully known.<br />
<br />
For genetic algorithms, one defines a genome structure, relating input to the solution output in a general way. Then, by means of a performance measure and random influences, successful genomes are combined in new generations to optimize results. There are many variants of genetic algorithms, but all are related in applying evolution theory as studied in the field of biology. Neural networks have a problem solving strategy that is based on ‘training’ the networks, on which we will not elaborate on here.<br />
<br />
Both approaches separately have their limitations however. For genetic algorithms, it is often not efficient and without a clear goal on what to optimize as a solution, it cannot perform well. For training neural networks, a topology has to be defined beforehand, which is already a challenging task for complex problems.<br />
<br />
With the NEAT algorithm however, the genetic algorithm is not applied to solving the problem directly. It does not evolve a solution of the problem, but rather a solver of the problem. In this case, that is a neural network. The algorithm starts with a minimal neural network possible for this problem. It is minimal in the sense that it only contains nodes for the input and output. Depending on the specific variant of the algorithm used, the initial neural network is fully connected between the input and output or only minimally connected. Then the evolution process starts with generating new neural networks which are then evaluated to determine how the network scores against performance measure. The evolution of networks consists of both topology (adding and removing nodes or connections) and altering connection weights and node biases.<br />
<br />
=== Application to simple problem ===<br />
To check how this will work on our problem, we applied the NEAT algorithm with a Python implementation to our test model introduced above with three starting nodes and a single goal node. Between these nodes, there are nine roads. For a first implementation, we define our problem with discrete time steps, where we assume a driver completes a trip over a road connection within exactly one time step. We define our Neural Network for this problem with three input nodes, representing the starting distribution of cars in the starting nodes, and 9*T output nodes, where T is the number of time steps we look into the future. These output nodes represent a matrix where for each time step a number determines how many cars we send onto that road. For now, to keep it simple and to enable manual checks, we define the road capacity to be 1 car per time step. This means that the matrix will consist of ones and zeros, indicating for each time step and each road whether the system will send a car on that road or not.<br />
<br />
For each generation of networks the algorithm produces, we test its fitness. We do this by evaluating the network a number of times with random starting distributions, and then simulating those starting distributions over time with the output the network provided. The network is considered to be more fit than another when the average time (over all test runs) it took to get the cars to the goal is less than the average time the other network clocked. When a solution failed to send all cars to the goal in T time steps, it is penalized in such a way that its fitness will always be less than a solution that was able to send all the cars to the goal.<br />
<br />
When running the algorithm this way, we indeed saw that it learned a strategy and that average (and best) fitness was increasing over the generations.<br />
<br />
=== Further application with traffic simulation ===<br />
Now that we see we get reasonable results with the simple test model, we turn to more realistic simulations. For this, we want to replace our simple 4-state model with a traffic simulation program, SUMO. SUMO, Simulation of Urban MObility, is a road traffic simulation package. It has a large variety of applications, from simulating new traffic light algorithms to simulating way larger, more global networks. More specifically, with SUMO we can simulate traffic flows and traffic jams to deal with, instead of with restricted discrete distributions of cars.<br />
<br />
The model changes a little bit. The Neural Network still has three input and 9*T output nodes and is still evaluated with NEAT. However, the fitness can now be simulated more in a more realistic way using SUMO. It uses the 9*T matrix to know how many cars should be sent in which direction for each time step. Using e.g. the capacity of the roads and possibly arising traffic jams, it knows how long it takes until all cars reach the goal. This can serve as a more realistic fitness which is used to evaluate the neural networks.<br />
<br />
=== Extending the model with more goals ===<br />
The next step in making the model slightly more realistic, is adding more goal cities. In the current model with four cities and nine roads, only one of the cities served as a goal. It changes the topology of the neural network: it has 4*G - G input nodes (G: amount of goal cities), telling how many cars need to move from one of the four cities to one of the G goals (without cars moving from a goal to the same goal). Furthermore, the amount of roads increases to 12 since people need to move in both directions on all roads now. Lastly, the amount of output nodes increases drastically: it will now have 12*T*G output nodes, telling at each time step T how many cars of the cars to goal G need to go to which road. The rest is still the same: SUMO will make a simulation based on the output nodes of the neural network, which will determine the fitness.<br />
<br />
== Testing an implementation ==<br />
=== Introduction ===<br />
Testing an implementation of an algorithm is sometimes very hard. It returns a certain solution but how do you know when a certain solution is the best one and what kind of graph do you have to use in your test cases. In this part of the wiki we will explain more about testing our implementation.<br />
<br />
=== Comparing multiple solutions from algorithms ===<br />
To check if a solution is indeed the optimal, you can compare it with other solutions of the same algorithm or with the solutions of a different algorithms. <br />
<br />
==== How to compare ====<br />
When comparing solutions of different algorithms and the same algorithm you must be sure that the comparison is valid. A comparison is only valid if they refer to the same situation (thus the same graph, the same amount of drivers and the same destination). If for example you conclude that your algorithm is bad because it returns a different total driving time for France than for Germany, then your conclusion is wrong because indeed the optimal driving time is different for France than for Germany, so it has nothing to do with your algorithm. This doesn’t mean that you cannot use different situation. Of course you can use different situations, however only compare the solutions for France with other solutions for France and only compare solutions for Germany with other solutions for Germany.<br />
<br />
==== Different behavior?! ====<br />
When comparing different solutions for the same/different algorithms you have to look at the total driving time of the optimal solution. Furthermore for comparing the same/different algorithms you might also take a look at the behavior of a solution, however don’t conclude that one of the solutions is bad because it haves different behavior. Because the total driving time might be the same for solutions with different behavior. To illustrate this we will show a graph of a 3D function: <br><br />
[[File:Different_Behavior_Same_Result.png]] <br><br />
Suppose our Artificial Intelligent algorithm has to find the solution with the lowest value and suppose it concludes that (0, -5) is the best solution and suppose some other algorithm concludes that the best solution is (0, 5). Is one of these algorithms wrong because the solution of both algorithms is different? Definitely not, they both found the lowest value 0, however the lowest value occur at multiple places. The same holds for our algorithm, there can be multiple solutions for which the result is optimal. In some of our algorithm you can see that a scalar multiplication of the distribution function that the algorithm returns is also a valid solution of the problem, because it has the same total driving time.<br />
<br />
==== Same Algorithm Comparison ====<br />
This sound strange, but is a very simple test to check if your algorithm is functioning. Most of our algorithms are stochastic and not deterministic so different total driving times might return as value for the same graph as input. This is not a problem if the total driving time doesn’t differ significantly very much (< 1%), however if they differ a lot then either your implementation is wrong or your idea might not work (however there might also be a local minimum, which is not very likely). If your algorithm passes this test, it is an indication that is behaves well. However you cannot directly conclude that your algorithm works.<br />
<br />
==== Different Algorithm Comparison ====<br />
Also you might test different algorithms, but be definitely sure that you compare same situations. Because a different formula for traffic jams might change the total driving time a lot. For comparing different algorithms the same holds as for comparing the same algorithms. If the total driving time is different it might not be a problem, however the total driving time may not differ significantly very much. If they do then either one of the algorithms is wrong, but they can also be both wrong. If both of your algorithms pass this test, it is a very good indication that both behaves well. Especially when both algorithms differ very much from implementation.<br />
<br />
=== Test Assumptions ===<br />
To check if a solution is optimal you might make some assumptions about behavior for the optimal solution. If your solution passes these assumptions it indicates that your implementation is good (but you cannot directly conclude These assumptions are:<br />
==== One direction traffic ====<br />
Every road consist of a driveway forward and a driveway backward. If drivers are both send on the driveway forward and the driveway backward that need to go to the same goal then something is wrong with your algorithm. You can easily conclude this by saying that either one of the cities is closer to the goal or it is equal, in which case it is never a good idea to go to the city that is further away from the goal.<br />
==== More roads is better travel time ====<br />
Using the same graph with more roads won’t increase the sum of the travel times if everyone follow the orders of the navigation system. If the navigation system uses the same distribution over the roads and not sending anyone over the new roads it would have the same sum of the travel times. Hence if the travel time increases then something is wrong.<br />
==== Shortest Path Without Traffic Jams ====<br />
Suppose you assume that everyone takes the shortest path and assume that no traffic jams occur (even when a lot of drivers go over the same road) then you get some total driving time A. This total driving time A is always lower than the total driving time B that the artificial intelligent algorithm will find and moreover the total driving time B doesn’t differ very much from the total driving time A. If the new found total driving time B is 20 times larger than total driving time A then something is wrong.<br />
<br />
=== Choosing a graph ===<br />
Choosing a graph to run your algorithm on is also important for testing your algorithm and can also help to determine the efficiency of your algorithm.<br />
<br />
==== Predetermined Graph ====<br />
Using an existing simple graph, like Tetraëder or Complex Rectangle:<br />
Advantages are that these graphs are very small and simple to understand and therefore also more easily to see if a solution is indeed the optimal one. Disadvantage are that these graphs are relatively simple compared to real graphs, also creating a predetermined graph that is very large will cost a lot of time for you to create. <br><br />
[[File:Brick Algorithm - Situation Representation.png|250px]]<br />
[[File:ILDB(+) Algorithm - Situation Representation.png|150px]]<br />
<br />
==== Total Random Graph ====<br />
You will generate an amount of nodes {a1, a2, …, an} and select an amount of vertices {b1, b2, …, bm}. Every vertex vi will be a connection between {ac,ad} and {ad, ac} where c and d are random integers between 1 and n, if this connection has already been made then new c and d will be generated. An example of this is: <br><br />
[[File:RandomGraph.png|300px]] <br><br />
Graphs created by this are indeed total random, however most of the time the properties of this graph are not very good. Like depth of the graph and often a graph is not connected, furthermore it is often not realistic because you have crossing vertices without the option to go from one of the vertices to the other. You can make them more realistic by checking each random graph if it is connected and has a certain depth, however this make this method slower and often they won’t be realistic even if you have some requirements.<br />
<br />
==== RR-Grid ====<br />
RR-Grid stands for random removal grid. A grid graph is very simple and looks like a square (where each vertex stands for two vertices one forward and one backward): <br><br />
[[File:USE_-_Grid.png|250px]] <br><br />
What a RR-grid basically do is remove some vertices. So let’s remove 10 random vertices (from the 40 vertices), then we get: <br><br />
[[File:USE - RRGrid.png|250px]] <br><br />
The advantages of this method are that the generated graphs are random and have nicer properties than total random graphs and the probability that such a graph is not connected is lower than for a total random graph, however you need a lot of nodes to get a high depth.<br />
<br />
=='''References'''==<br />
<references/></div>Wed, 05 Apr 2017 09:15:25 GMTS143110https://cstwiki.wtb.tue.nl/wiki/Talk:PRE2016_3_Groep13PRE2016 3 Groep13
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37374
https://cstwiki.wtb.tue.nl/index.php?title=PRE2016_3_Groep13&diff=37374<p>S143110: /* Future user research */</p>
<hr />
<div>David Dekker - 0936100<br />
<br />
Haico Dorenbos - 0959563<br />
<br />
Jori van Falier - 0910959<br />
<br />
Tim Muyrers - 0907928<br />
<br />
Thomas Schouten - 0938927<br />
<br />
Sten Wessel - 0941508<br />
<br />
<br />
<br />
== Introduction ==<br />
Waiting for a very long time in traffic jams even when your destination is in the same city. Solving traffic problems is something that has already inspired a lot of engineers and will also inspire a lot of new engineers. “Only 10 percent of cars would have to be connected for it to work” <ref>National Electronic Library for Health (2017). ''New AI Algorithm Beats Even the World's Worst Traffic.'' Retrieved March, 2017 from https://motherboard.vice.com/en_us/article/new-ai-algorithm-beats-even-the-worlds-worst-traffic</ref> was a recent quote in an article about solving traffic jams with artificial intelligence. In our project we are also interested in solving traffic problems using artificial intelligence. This wiki page therefore will display the possibilities of self learning navigational software. Throughout the report several points of attention will be introduced, investigated and processed in the prototype navigational system. The goal of the system is to create a cooperative/inter-vehicular, high-level (for example, no traffic lights) and based on the current situation (hardly any self driving cars, not all people have navigational systems). This research is set up trying to answer the following research question:<br />
<br />
''How can the travel time be minimized while maximizing overall utility, looking at user and society, by using a Self Learning, Inter-Vehicular Cooperative Navigation System (SLIVCNS)?<br />
''<br />
<br />
== Problem description ==<br />
<br />
Every day in the Netherlands people go to work with their cars, and every day the same thing happens, traffic jams causing a lot of pollution and waiting time. Most of the time people are traveling a lot longer than necessary to reach their goal. This gave us the idea to create a system that will reroute this traffic through secondary roads to minimize the overall waiting time or maybe even prevent traffic jams. By doing a lot of user research we hope to create a user friendly system that will help the user and society with this problem.<br />
<br />
=== The simple problem ===<br />
<br />
The problem in the simplest form is as follows, there is a city A and city B, in the morning people move from A to B to go to work. There is a highway (which is the fast route) and two secondary roads. When the people move from A to B a traffic jam occurs, this causes people to wait and increases their travel time. By rerouting a part of these people to the secondary roads this traffic jam can be minimized in such a way that two things can happen, or a mix of these: <br />
1. The traffic jam is prevented and therefore the overall travel time is a lot lower, this is hard to implement as people who had to take the secondary road have longer travel time. <br />
2. The traffic jam has become smaller and all the people whether they take the secondary road or the highway have the same travel time. This option seems more fair but still causes the problem of people waiting in traffic jams.<br />
Which one is to be preferred has to be researched as the user is very important in this decision.<br />
<br />
=== The full problem ===<br />
<br />
As seen above the problem can be made very simple but also a bit unrealistic, because in the end the system has to run in the real world, where there are thousands of ways to get from A to B. Also in the real world a lot of problems arise. Traffic is very unpredictable and can come from anywhere and go anywhere. It is not simply people moving from A to B. Also a lot of people will not be using our system, we will get no information from them, but they will have impact on our system. The biggest problem is all the unknowns, a lot of things can happen on the road. For example an accident can cause a traffic jam that can not be predicted. Because of all these problems we will be using self learning software to deal with these problems. This program will try to compensate for all the unknowns. <br />
<br />
== Objectives ==<br />
=== Minimize travel time ===<br />
The main goal of the system is to minimize travel time by rerouting users, to prevent longer waiting times caused by traffic jams.<br />
<br />
=== Maximize overall utility ===<br />
The utility can be split in three groups.<br />
==== User ====<br />
*Minimize travel time, this is the main objective as mentioned above.<br />
*Estimate arrival time by predicting traffic jams so the user knows the optimal route and best moment of departure to reach the goal location on time.<br />
*Making the system fair for all users. This will require user research in finding the fairest solution.<br />
*Taking the privacy of the user into account by doing user research.<br />
==== Society ====<br />
*Traffic jam prevention to minimize unnecessary pollution, by rerouting to minimize unnecessary waiting time.<br />
*Minimize disturbance by rerouting through less populated locations.<br />
<br />
==== Enterprises ====<br />
*Creating a system that is better than the present navigation systems, so potential users will be interested in buying this system.<br />
<br />
== State of the Art ==<br />
There are a lot of different navigation devices available currently. The most simple one is only able to tell you the shortest route. A small upgrade also allows the user to exclude certain routes, like toll roads and tunnels. These often also give a choice for the user to either take a fastest (time) or shortest (distance) route. Some devices also allow traffic information incorporation which ask you to reroute if there is a traffic jam. When combining such a navigation system with a self driving car it is also possible to make the decision automatically. These all however respond to a situation instead trying to avoid certain situations.<br />
<br />
<br />
== Improvement over state of the art == <br />
To improve the state of the art it is important to know what the implications are of the current system. These systems will reroute users as soon as there is a faster route available and thus will distribute traffic equally over the roads available, which may seem like the optimal situation. A self learning system however should have more roads available, since it can predict traffic jams that haven't even happened yet. It can reroute the user many kilometers before the actual traffic jam and thus prevent new traffic jams in reroute roads.<br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Idea visualize prof.jpg|300 px]]</div><br />
<br />
== Approach ==<br />
The goal of the research is to improve the current traffic situation. This can only be considered an improvement when the changes do not diminish the overall utility. This means that within the navigation system the user as well as the society have to play an important role. As the time window is quite a tight one this research will be conducted simultaneously. The group of researchers will be split where four members of the team will work on making the self learning navigation work. The remaining two members will focus on the users and society. By frequent interaction, the team will be able to create a working navigational system in which the user and society will be strongly embedded. This also allows the prototype to be as representable as possible. <br />
<br />
The completion of the program will also include a visual simulation which allows all team members as well as external personnel to use and understand the way the system operates, and what conclusions it provides.<br />
<br />
== User and Society ==<br />
<br />
People behave differently when wearing something to hide their identity<ref>New Republic (2014). ''Sunglasses Make You Less Generous.'' Retrieved March, 2017 from https://newrepublic.com/article/117152/how-sunglasses-and-masks-affect-moral-behavior</ref>. This also happens in a car. People are not directly visible and thus act more selfish. As long as people are driving their car (instead of autonomous), we can investigate further if/how we want to incorporate this. <br />
<br />
<br />
Furthermore our navigation idea will request people to drive different roads, which at points will probably make them lose time on their journey. When looking at literature we find something called: “risk aversion”<ref>Tversky, A. & Kahneman, D. (1991). Loss Aversion in Riskless Choice: A Reference-Dependent Model. ''The Quarterly Journal of Economics'', 106, 1039-1061. Retrieved March, 2017 from http://www.jstor.org/stable/2937956?seq=1#page_scan_tab_contents</ref> <ref>Scott, A. & Witt, J. (2015). Loss Aversion, Reference Dependence and Diminishing Sensitivity in Choice Experiments. ''Melbourne Institute Working Paper'', 15/16, 1-39. Retrieved March, 2017 from https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2655266</ref>. Risk aversion is a human trait where someone would avoid taking a favorable bet (overall travel time decrease) when it means they can directly lose out (more travel time) <ref>Youtube (2015). ''Would You Take This Bet?'' Retrieved March, 2017 from https://www.youtube.com/watch?v=vBX-KulgJ1o</ref>.<br />
<br />
== What is a traffic jam ==<br />
<br />
In order to say anything about the efficiency of our system we need to first look at what the possible causes of a traffic jam are. In general we can indentify three leading causes in the Netherlands for traffic jams which are:<br />
<br />
* Accidents<br />
* Too many drivers on the road: Over Breaking<br />
* Road works<br />
<br />
A solution for the first cause has to be investigated more. While accidents are bound to happen as long as humans control cars (as making mistakes is common amongst humans). We do see though that accidents can be massively decreased when we decrease the amount of cars driving at a road. The busier it is the more common it is for accidents to happen.<br />
A study performed in 2010 <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> found a relation between business on roads and accidents. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsedPNG.png|300 px]]</div><br />
E(λ) being the estimated number of crashes, Q is the measurement for traffic volume, xi is the risk factor, γi is the corresponding coefficient to the riskfactor and finally β is the effect of the volume of traffic on crashes. <br />
<br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:FormulaUsed1PNG.png|380 px]]</div><br />
Too many drivers on the road also causes direct traffic jams through over breaking. When a car in front breakers the one after it breaks too but a little harder thus going a little slower. The person driving behind this person does the same and thus it will eventually reach a standstill.<br />
<br />
Road works can’t really be avoided. But if the system is connected to a database, it should be able to take this into account.<br />
<br />
While reaction systems are able to deal with most of these issues, it cannot prevent accidents from happening. While a self learning system should be able to.<br />
<br />
<br />
<br />
So to make sure the navigational system works in society you have to give it a certain goal of not exceeding a density of cars on the highways. For testing purposes you could look at a particular road and figure out the relation between accidents and car density. <br />
More research is needed for an accurate relation between traffic density and accidents. As the study from Kirsten Duivenvoorden <ref>Duivenvoorden, K. (2010). ''The relationship between traffic volume and road safety on the secondary road network. A literature review.'' Leidschendam, SWOV Institute for Road Safety Research. Retrieved March, 2017 from https://www.swov.nl/sites/default/files/publicaties/rapport/d-2010-02.pdf</ref> explains current research is mostly focussed on geometric variables, meaning the median and width of roads, the presence of turning lanes and presence of traffic lights and traffic speed. The reason for this is that in the past the volume could not be changed, and thus this wasn't the aim of researchers. With this new technology this can however be influenced and thus more research would be benicial.<br />
<br />
== User research ==<br />
<br />
=== Already done research ===<br />
<br />
We base already done research on a questionnaire about navigation systems done by “Kennisinstituut voor Mobiliteitsbeleid (KiM)” from the ministry of infrastructure and milieu<ref> Kennisinstituut voor Mobiliteitsbeleid (KiM), September 2015, Retrieved March, 2017 from https://epublicatie.minienm.nl/navigatiesystemen-wie-wanneer-en-waarom#/slide_navigatiesystemen-wie-wanneer-en-waarom </ref>. This questionnaire has 3.924 respondents from car drivers of the “Mobiliteitspanel Nederland (MPN)”. <br />
==== Use of navigation systems ====<br />
<br />
Current navigation system are rarely bought because of traffic information. Most of the information comes from the internet and the radio and during the travel also information panels above the road (DRIPs). From the questions follows that easy in use and quality of information are important reason for people not to use the navigation system solely for traffic information. That is where our navigation system can make the difference. By making an easy to use system with reliable information, we hope our system will be of interest for the user.<br />
<br />
Navigation systems are mostly used for non-frequent or long distant travels. The system is also used for points of interest. The most used points of interest are parking lot (75%), gas stations (49%) and places of entertainment (38%). Furthermore the system is used to avoid toll roads and ferries. All these options might be interesting to implement in our system.<br />
<br />
==== Route taking ====<br />
<br />
Something else of important follows from the research done here. Most users will choose for the fastest route (60-70%), also the default setting is used a lot which is the fastest route in most cases. Also a relatively large amount of users deviate from the route given by their navigation system, mostly because they know a better route, heard or seen a traffic jam, or because of road works. Because our system will work optimal when everyone using the system, follows the system, it might be very important to inform the user that our system has noticed these traffic jams and road works, and that we can proof that our system can find the most optimal route for the user.<br />
<br />
=== Our user research ===<br />
<br />
To make sure all user preferences are taken into account a user research is conducted, focusing on the key aspects which should provide a strong basis for the product to be released to the public. The aim of the research, and thus leading factors into choosing the approach, is to gain as much information about as many different drivers as possible. Since the aspects of the research are predetermined, the best research method available is a questionnaire. Google Drive is used to host the questionnaire, because it has all features (sort of questions/ set-up) required for this research. <br />
<br />
The link below will bring you to the questionnaire*:<br />
<br />
https://docs.google.com/forms/d/e/1FAIpQLScDPnmNoF6QXpMX930HC-V1ydxndn6kBt-_NZO-LiEaaBGuWA/viewform<br />
<br />
* Please note that by choosing certain answers it is possible that the questionaire forwards you and thus skipping certain sections. In order to see the full questionaire, make sure to not answer the questions and just go on to the next section.<br />
<br />
The questionnaire is divided in several sections. Just like any research the questionnaire begins with a brief introduction of our idea followed by asking the participants permission to use the data they provide through an inform consent form. In all following sections the key issues are presented to possible users. To make sure everyone recieving the questionaire will be able to understand and participate an English and Dutch version is available.<br />
<br />
The target group contains all drivers, both frequent and non frequent drivers. While the system is designed for non selfdriving cars it would also work on autonomous cars. The system will first be tested and deployed in The Netherlands and thus this is also the target group for distributing the questionaire.<br />
<br />
==== ''Section 1: Introduction questions'' ====<br />
<br />
The system investigated is a navigational system. Thus it is important that the participants are actual drivers, and thus the introduction asks the participants this question. It also briefly looks into the preferences a car owner has.<br />
<br />
====''Section 2: Willingness to follow the system''====<br />
<br />
A navigation system could work very well, but in the current situation it is still the users controlling the car. Since SLIVCS is designed to be applicable into society at this moment it thus is dependent on the willingness of the user to follow its instructions. If an user would stray from the path the system proposes often the system would no longer work. This section proposes several situations to the participant and asks the user whether or not he would follow the system (thus believing and whatever the system proposes it should still be faster then not listening to the system).<br />
<br />
====''Section 3: Fairness''====<br />
<br />
The system can decrease travel time and traffic jams in several ways. The participants are presented with this way in a honest way. This means that we do present options where a user of the system is always equal or worse of to a person not using SLIVCNS. This sections is designed to get a better understanding of the target group, and to give a better insight into the goal our system should aim for.<br />
<br />
====''Section 4: Privacy''====<br />
<br />
The current state of the world means that almost all information gets shared <ref>Everyone is under surveillance now, says whistleblower Edward Snowden. (2014, May 3). ''The Guardian.'' Retrieved March, 2017 from https://www.theguardian.com/world/2014/may/03/everyone-is-under-surveillance-now-says-whistleblower-edward-snowden</ref>. This however does not mean that all users all necessarily willing to share all information. Since our system needs to know certain things about users it is important to know that people do not mind sharing this information (anonymously). We ask the participants if they would be willing to share their current location as well as their destination. If the participant is willing to share this he/she is taking to the next section with follow-up questions. If the participant is not willing to share this information he is taken to the final stage of the questionnaire.<br />
<br />
====''Section 5: System interaction''====<br />
<br />
In this section the participants are asked to think about the idea of SLIVCNS, and how they would see such a system fit in their daily lives. It contains three subsections. The first subsection asks the participants what inputs they would like to have on the system. There are six predetermined options, like what type of roads, should the system take pollution into account etc. It is also possible to write your own input, if this is not present in the existing questions. <br />
The second question makes the participant think about how they would like to communicate with the system. As explained in an earlier chapter it is important that the system knows when a user wants to travel from a to b. This means that a user would need to interact with the system quite frequently, which means that this interaction should be easy. Possible communications are for example a mobile app, an online website or a custom made interface. Once again the questionnaire provides room for a written answer. <br />
The final subsection is regarding the route planning of possible users. For the system to work optimally it is beneficial to know way in advance when users are going to drive. This section asks of different occasions (different destinations) the time a user would be able/willing to provide the system with its travel plans before departing.<br />
<br />
====''Section 6: Finalizing questions''====<br />
<br />
This section is focused on the aim of the navigational system. In this section the participants are asked about possible approaches the system could take regarding traffic/travel time distribution.<br />
<br />
A participant is free to not answer a question if he/she does not want to answer, without having to provide an reason. All information helps towards optimizing SLIVCS for the user, and thus there are no mandatory questions (excluding accepting the inform consent form).<br />
<br />
=== Outcome user research ===<br />
<br />
=====Outcomes summed up =====<br />
* 63 participants<br />
* Representable participants (drivers)<br />
* No clear answer to environmental factors that should be taken into account<br />
* The idea is well recieved with up to 95% of the participants showing interrest in the product<br />
* 72.4 % of all instructions by the navigation system will be followed (clear distinction between routing via the fastest way (highway) vs sending the users on secondairy roads)<br />
* Point based system or reward for driving longer routes is preferred amongst the participants<br />
* 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination<br />
* Important to know how the navigation system will be limited by allowing users to choose or decline certain road types <br />
* Mobile app is preferred device to interact with the navigation system<br />
* Over 75% of participants are willing to plan their travels ahead of time and inform the navigation system of said plans.<br />
* 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product<br />
<br />
Per section of the questionnaire the general outcomes will be explained and what this means for our design. <br />
<br />
In total 63 people participated in this study (and it is still open for responses until 7th of April). Seven people responded to the English version and fifty-six filled in the English version. Out of the 63 participants one person was not qualified to participate and thus only declined the inform consent. <br />
<br />
The introduction questions focused on introducing the product and for us the investigators to get familiar with the target group. Approximately 74% of the participants indicated to use a car on a regular basis. After a short introduction 95% is interested in the product. When the participant is asked what environmental nuisances the system should take into account a wide array of answer are given. The possible answers “No pollution nor noise disturbance”, “pollution”, “noise disturbance”, “both pollution and noise disturbance” got 25.8% 26.4% 9.9% and 37.9% respectively.<br />
<br />
So what does this mean. To start it means that the target group intended for the research is reached. The majority does regularly drive a car, and thus are potential users of our system. With 95% of the participants being interested in the system shows that the potential for a successful device is high. It also means that if developed well it would gain a large market share and thus function well through inter device communication. As for environmental nuisances it’s hard to say how the device would be optimized at this point since the answers are divided amongst all four categories. <br />
<br />
The second question category tests whether or not participants are willing to follow the system in four situations. Over all four categories an average of 72.4% was willing to follow the system. There are however differences between categories. Participants indicated they were 98.5% willing to follow the system if a traffic jam was already on their route, while only 50% would follow the system if it sends you on a secondary road with the aim of avoiding a traffic jam for other drivers. While (as we will see later in the questionnaire results) the system would be optimized to get the optimal drive time for all drivers it would be important for users to understand and believe the effectivity of the system. While a willingness of 72.4% is a strong basis for the system the function well, a higher percentage would increase its effectivity. <br />
The next section is about how our system should function to decrease travel time. There are four proposed approaches and ideally one would be the best. The preferred options are a point based system that will sent you on a longer route, but sends you on a shorter route the next time. Also the idea of a reward (not defined) sounds appealing to a large part of the participants with both sections respectively scoring 81.9% and 73.8% on Very good or Good in comparison with the sum of all participants. For the design of the system this would mean that there is more room for the designers to reroute cars to avoid traffic jams completely. It does however also mean that it is important to include regulatory bodies like the government into the process to discuss possible rewards (if this choice is made). <br />
<br />
As explained earlier the system has a couple of areas which require users of the system to facilitate the system certain information to function optimally. The section privacy investigates if users are willing to provide general navigation information for the system to use and process. As expected from the current situation of navigation systems the hypothesis of users willing to provide the system with some of their information is confirmed. Respectively 82.5% and 79.3% of all participants indicate they are willing to share their current location and destination. <br />
<br />
<br />
<br />
The next section is aimed at making the product more appealing to users. Here several inputs are proposed which will decrease the system performance. However if this makes the product reach more people this fall in performance will most likely be compensated by the extra information the system is able to receive from more users. 36% of the participants would like to see an option to tell the system if they are in a hurry or not. A design choice is made that only options which more than half of the participants would love to see in the system are considered (because previously mentioned reasons). Approximately 71.4% of the participants however are interested in being able to tell the systems what roads they like to take or avoid. This means that this option should be more investigated, since the system should have a large degree of freedom in deciding which route to take. Having users choose their own paths defeats the purpose of the system. <br />
<br />
<br />
In the beginning of the questionaire participants were asked on their point of view on the important of environmental factors influencing navigation systems functioning. In this section further questions are posed to get a better understanding of how these factors have to be taken into account. The participants could either choose for the system to automatically take polution/noise into account or not aswell as having a option to change this setting or no option.<br />
64% would like standard incorperation of polution into the system, with 72% prefering to have an option to change the setting the system is on standard. <br />
For noise these values are slightly lower with 60% wanting a standard incorperation of noise into the system, and a slightly higher amount of participants (76%) prefering to have a option to change noise incorperation into the system. <br />
What this means for the system is that it is important for people to think about the environment. Driving is not purely a form of travel from a to b but also about doing so responsibly. The system requires certain inputs in the forms of roads (capacity, events that influence road efficiency etc), but will also need a incorperation of other environmental properties to guide the system to a positive driving experience. <br />
<br />
<br />
Making system interaction easier is a strong aim in future planning of travels. For this a section of the questionnaire is design specifically to understand how participants/drivers would like to do this. The aim is to design one maximally two devices to keep the costs low and the usability high. Only one device is liked a lot by the participants which is the mobile app. Since the app is an excellent way to also inform the device of you travel plans over a large distance it would be able to provide pre planned travel trips. For the end product this is a good outcome. While the phone can be used as an interface in the car aswell a option can be chosen to combine it with a purely visual interface.<br />
<br />
The next section of the questionnaire continued of the planning and information exchange between user and device before departure. The most important input is regard work travel as this is the most frequent reason for travel <ref>Centraal Bureau voor de Statistiek (2016). ''Transport en mobiliteit.'' Retrieved March, 2017 from https://www.cbs.nl/-/media/_pdf/2016/25/tm2016_web.pdf</ref><br />
. <br />
<div class="center" style="width: auto; margin-left: auto; margin-right: auto;">[[File:Goede Graph user research.JPG|666 px]]</div><br />
As you can see in the graph above on any occasion more than 75% is willing to inform the system ahead of time with the users intentions. This is a positive outcome, since future route planning could be incorporated in the system. Besides self-learning traffic patterns it can now also use data provided by users to make conclusions about the traffic situation.<br />
<br />
Finally like in the beginning the participants are once again presented with several possible goals of the system and if they would use the system accordingly. In general participants indicate to be still heavily interrested in the system with all possible goals achieving over 50% of interrested participants. The three most popular goals are "The overal travel time of all traffic is lower, but your travel time is the same", "There are still some traffic jams but the overal travel time is lower", "The overal travel time stays the same but your travel time is lower" with respectively 88.7% 91.9% and 95.1% of participants indicating to be interrested in this goal/product. What this means for the design mostly means that the goal can be customized as long as users dont lose out on travel time themselves. <br />
<br />
'''Final conclusion'''<br />
<br />
The research was designed to get a better understanding of the potential users and the stakeholders in transport. While several aspects of the study could be elaborated through more in depth interviews, the information collected through this limited resources survey give a good indication on what the important aspects of the product are.<br />
<br />
This questionnaire gives a good basis for an early user influenced design. Once this prototype is created more elaborate user and society research can be conducted providing a smooth implementation into society and achieving the goal of creating a user friendly system.<br />
<br />
=== Future user research ===<br />
<br />
==== Improvement over the done research ====<br />
<br />
In the research we have done, a lot of things could not be specified. For example the amount of increase in travel time in order to decrease pollution and noise disturbance is not known to us. Same goes for the increase in travel time to decrease traffic jams for other users. In this stage of the development these things can not be specified. The questions we asked were more for a general understanding of the user needs, and help us choose the right direction for the system. Once the system is further developed more specific questions can be asked, as a lot more of the system is known. <br />
The research we have done has been on a selective group that is not a good representation of the population. Furthermore the amount of reactions is only a very small fraction of the population, in order to get reliable results the questionnaire should be send to more users. This will make the representation of the population more accurate and maybe some compensations have to be made, because not every type of user is willing to fill in a questionnaire. <br />
<br />
==== Extra research ====<br />
Future research could focus on whether or not letting the participants know the reason for a certain route planning, this could increase the willingness of users to follow the system. As we saw the point and reward systems are favoured by the users and have to be researched even more, looking into the type and size of the reward, or the balance in the point system. <br />
<br />
The system will require an interface that has to be developed. This interface has to be user friendly, by being easy to use and understandable for all users. This is why during the development of this interface a lot of user research has to be done to create a user friendly interface. This is not the focus of this project, but has to be done if the system would be developed further. <br />
<br />
Other research like society research, is also of great importance in the development of this system and if the development is nearing its end, research about bringing the system on the market has to be done. For this project where we are in the early stages of development these things can not be researched by us, because we do not know enough about the system.<br />
<br />
== Low Level Implementation Ideas ==<br />
=== Brick Algorithm ===<br />
==== Introduction ====<br />
We start by introducing a very easy model for our problem. It is not very expanded and therefore the model doesn’t really look like what will happen in the reality. In this model we only take into account the waiting time of driver (so no privacy, pollution, etcetera is taken into account ). Furthermore in this model we look at one part of the day: the morning rush and we will assume that everyone travels to the same destination. Our model is completely static and ironically it doesn’t take time into account. Moreover we assume that everyone uses our clever navigation system and that everyone is obedient to our clever navigation system and therefore they will always follow their directions.<br />
<br />
==== Situation Representation ====<br />
We will represent the current situation as a graph with vertices and edges, thus it will look like this: <br><br />
[[File:Brick Algorithm - Situation Representation.png]] <br><br />
The vertex that is two times circled is the destination, which means that the drivers from cities B, C and D have to go to this location (we don’t take drivers from city A into account, because they already arrived at their goal). The edges are driveways with a specific direction. So for each road there are two edges, because you can take the driveway forward or the driveway backward (for destination A however there are none driveway backwards, because if you are at destination A then you don’t have to drive back). Every driveway in this graph is indicated with a number and every city is indicated with a letter.<br />
<br />
==== Data about roads and cities ====<br />
We will assume that 30.000 people live in city B who all need to go to A. We will furthermore assume that 20.000 people live in city C, who all need to go to city A. Moreover we will assume that 10.000 people live in city D, who all need to go to city A. Last of all we will assume that the waiting time for each road is determined by a simple linear function: ax + b, where x is the amount of driver over a driveway (we will see later how x is determined), a is the capacity constant of the road (if it is lower then it can handle more drivers) and b is the constant, which indicated how much time drivers take minimal by taken a driveway. The result of this function is the amount of minutes that drivers have to wait before arriving at the new node. Now we have some more information about the roads:<br />
<ul><br />
<li>Roads 1 and 3 are huge highways (they can handle a lot of drivers), therefore they have the function 0.00025x + 15 (you need at least 15 minutes to drive over that certain road)</li><br />
<li>Roads 2, 4, 5, 6 and 7 are little, short roads (In dutch: N-wegen) (and they can only handle a small amount of drivers) and these roads have the function 0.0005x + 5 (you need at least 5 minutes to driver over that certain road).</li><br />
<li>Roads 8 en 9 are small, very short roads (you are soon at the new node by using this vertex) and have the function 0.0005x + 3 (you take at least 3 minutes to drive over that certain road).</li><br />
</ul><br />
<br />
==== How to determine x (amount of drivers over a certain road) ====<br />
Driver can take different roads. Of course this will influence the value x. In this algorithm the value of x will start at 0, because at the start of our algorithm. Suppose we have some driver that live in city C, which drives over the driveways 6, 5 and 1 (in this order) and finally arrives at his destination A. If this happens then the x value for driveways 6, 5 and 1 is incremented by 1. Therefore the x value for driveways 6, 5 and 1 is now equal to 1. This is executed for all drivers, they keep driving over the driveways until they reach their final destination A and for every driveway we will increase the amount x by one if a new driver uses this driveway. <br />
<br />
==== Comparing Solutions ====<br />
Each way of driving must be comparable with each other way of driving (it must be a total order, so you must be able to say whether a < b, a > b or a = b). A value (cost value) is assigned to every solution (route drivers might take). The cost value that we use in our case is the sum of all waiting times for each driveway multiplied with the amount of drivers that take that certain road. So for example if 10.000 people have driven over road 1 and they have waited 17,5 minutes then the cost value for this road is 175.000, this will be summed with the cost values of all roads and then we will have our final cost value. A solution is better than another solution if it has a lower final cost value. Therefore our goal is to minimize the final cost value.<br />
<br />
==== How will the drivers behave in optimal case? ====<br />
That is something this artificial algorithm will determine (by assuming everyone will follow our clever navigation system). The main idea is that we will use a distribution table, where every driveway gets a certain value (tried by our artificial algorithm could be for example a gradient optimizer or a genetic algorithm). But how do drivers behave when all values are assigned. Suppose for example that driveway 1 gets a value of 20, driveway 4 gets a value of 40 and driveway 9 gets a value of 70 (those are all outgoing driveway from B) then this algorithm will say that (20/(20+40+70)) = (20/130) relative amount of all drivers from city B will take driveway 1, moreover (40/130) will take driveway 4 and last of (70/130) will take driveway 9. This will be done at the start of this algorithm for every city. Of course after this is done, not everyone has arrived at destination A. So we have to do this again and use the same distribution table. Hence city B has after the first “driving” round a new amount of drivers, which is equal to the amount of driver that went from city D to city B (by using driveway 5) added with the amount of drivers that went from city C to city B by using driveway 8. After this we will get a new “driving” round. This script will continue until everyone arrives at destination A.<br />
<br />
==== What will happen if none or a few drivers to go city A? ====<br />
We will limit our script till 10 “driving” rounds for example (you can take a different number than 10). If after 10 “driving” rounds not everyone has arrived at destination A then it will get a very high cost value (and therefore this is automatically a very bad solution, worse than solutions in which all drivers arrive at destination A). Our idea was for example to say that this value is equal to (Float_MaxValue) * (amount_drivers_at_start - amount_drivers_reached_destination) / (amount_drivers_at_start). This cost value has to be definitely higher than a cost value of a solution for which it is possible for everyone to arrive within 10 “driving” to destination A. If we define our cost value this way then a solution for which 10 “driving” rounds is too less to for everyone to get to their destination is worse than a solution in which everyone get to their destination in 10 “driving” rounds.<br />
<br />
==== Advantages ====<br />
<ul><br />
<li>It is a very simple model, because it doesn’t need Neural Networks or other complex ways to find the solution. A gradient solver or genetic algorithm is enough to solve this problem.</li><br />
<li>It can be easily expanded to multiple goals, where from each city they need to go to different nodes. Also for different times it can be used. Thus it will always find the best solution for every case, so it is complete and optimal.</li><br />
</ul><br />
<br />
==== Disadvantages ====<br />
<ul><br />
<li>Assumptions were made which are not very logical (waiting time is a linear function is not a very good assumption for example, likewise assuming that everyone will use our clever navigation system).</li><br />
<li>For larger areas this algorithm will take a lot of time. Especially because this algorithm need to determine driveways * destinations * different_parts_of_the_day amount of constant. If you