The topic of this thesis is traffic simulation including the traffic network and the motorists within. The system is designed to help measure traffic and its characteristics of an existing, or a future urban traffic network, facilitating the design of such a network even in practice.
While developing this application, the primary concern was to make it capable of simulating the complete traffic of a small town. Thus it was important to algorithmically optimize it, both in terms of running time and storage space. For example an optimized A* searching algorithm was developed for this special case.
Another important aspect was making the simulation as realistic as possible with making some reasonable simplifications in the given problem. On the one, hand the designed network contains the important aspects of urban transport; on the other hand, the behavior of drivers corresponds to certain real-life driving styles.
For determining driving styles, I have applied a neural network-based reinforcement learning method. We develop our driving style based on the behavior of the drivers around us and our own reactions to the other drivers. Drivers are also aware of the part of the Road Traffic Act (KRESZ) that is important in this simulation, and they find the optimal route according to different parameters. All these form drivers’ adaptivity, that is their behavior based on their surroundings.
Finally, concrete results can be found to show the scalability of the system and to prove that we can run simulations even in large networks. The thesis also contains two different driving attitudes based on different reinforcements as a demonstration for the adequacy and the applicability of the combined learning algorithm presented in this thesis.