Thesis
Intelligent traffic light via reinforcement learning
University of Iowa
Master of Science (MS), University of Iowa
Autumn 2021
DOI: 10.17077/etd.006294
Abstract
People travel with vehicles daily. The traffic light is one of the most critical signals in an intersection to prevent accidents and help all vehicles on all the roads pass the intersection safely and efficiently. However, most traffic lights used today are fixed-time traffic lights. The logic is predefined and sometimes this approach is inefficient and wastes drivers’ time. This research focuses on finding an optimal operation strategy for a traffic light to make it more efficient to pass the traffic flow. One of the artificial intelligence (AI) techniques, reinforcement learning, is employed to train the control agent that can optimally choose the traffic light phase. Consequently, the designed traffic light is intelligent. We use three reinforcement learning methods, including Deep Q-learning (DQN), Double Deep Q-Learning (DDQN), and Proximal Policy Optimization (PPO). The generated policies are compared, and PPO is found to be the best method to train the intelligent traffic light. Then, PPO is utilized to further design an intelligent traffic light, considering the light phases with variable intervals. In this design, the control agent can choose a traffic light phase and a time interval for the phase to last. Furthermore, environment and action disturbances are considered. Environment disturbance is a probability that traffic collisions may happen. Action disturbance is a probability that the traffic light may malfunction and switch its phase randomly. The results demonstrate that the intelligent traffic light is robust and has high efficiency.
Details
- Title: Subtitle
- Intelligent traffic light via reinforcement learning
- Creators
- Yue Zhu
- Contributors
- Shaoping Xiao (Advisor)Venanzio Cichella (Committee Member)Chris Schwarz (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Mechanical Engineering
- Date degree season
- Autumn 2021
- DOI
- 10.17077/etd.006294
- Publisher
- University of Iowa
- Number of pages
- viii, 77 pages
- Copyright
- Copyright 2021 Yue Zhu
- Language
- English
- Description illustrations
- color illustrations
- Description bibliographic
- Includes bibliographical references (pages 73-77).
- Public Abstract (ETD)
Traveling with vehicles is very popular. However, waiting for a traffic light from red to green is always a waste of time and pollutes the environment due to harmful gas. This research designs an intelligent traffic light via three different deep reinforcement learning methods, including Deep Q-learning (DQN), Double Deep Q-learning (DDQN), and Proximal Policy Optimization (PPO). It has been shown that the optimal policy derived from PPO has the best performance. Furthermore, the traffic light phases with variable time intervals are studied. In addition, both environment and action disturbances are considered. The introduction of variable time intervals results in a better performance than fixed-time intervals. The simulations demonstrate that the intelligent traffic light trained via PPO is robust.
- Academic Unit
- Mechanical Engineering
- Record Identifier
- 9984210944502771
Metrics
12 File views/ downloads
67 Record Views