Loading...

Using a Deep Reinforcement Learning Agent for Lane Direction Control

Zare Hadesh, Ashkan | 2022

204 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55086 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Nasiri, Habibollah
  7. Abstract:
  8. In recent years with the progress of technology in different areas, the production of self-driving cars has been feasible. We can expect that vehicles of transportation networks will consist of both self-driving and regular cars in the future. In this research, a new method will be proposed for urban transportation networks to change the direction of reversible lanes according to the network's state. These reversible lanes are exclusive for self-driving cars. Human drivers are not allowed to enter these reversible lanes, considering the limitations of human ability compared to a computer in analyzing data and making decisions about moving direction. To achieve this goal, reinforcement learning is used, and a deep Q-learning agent is used to solve it. To express the state used in this study, we need the information of intersections at both ends of the reversible lanes. The data used in this state consists of the queue length, the number of vehicles in each lane, the current phase of the traffic lights and the remaining time of that phase, and the data related to the previous actions. The reward is considered as the amount of self-driving cars entering the reversible lanes. The actions are defined as different combinations of reversible lanes. After implementing this method in SUMO, the network stopped time and average travel time are reduced by about 12.2 % and 5.1 %, respectively
  9. Keywords:
  10. Deep Reinforcement Learning ; Self-Driving Car ; Intelligent Transportation System (ITS) ; Urban Transportation Network ; Reversible Lanes ; Transportation Systems Management

 Digital Object List

 Bookmark

No TOC