Artificial Intelligence-Based Transit Signal Priority Control Reduced Average Person Delay by 18.77 Percent in the Peak Hour and 23.37 Percent in the Off-Peak Hour Compared to Pretimed Signal Control.
Within a Simulated Setting, Researchers Developed a Controller For Transit Signal Priority Using Deep Learning Techniques and Data from Connected Vehicles.
Charlotte, North Carolina, United States
Transit Signal Priority Control with Connected Vehicle Technology: Deep Reinforcement Learning Approach
Summary Information
Transit Signal Priority (TSP) improves transit service by granting priority at traffic signals, while potentially affecting general traffic performance. The emergence of Connected Vehicle (CV) technology enables real-time, fine-grained traffic data for traffic signal optimization. This study developed adaptive TSP control frameworks and tested them for both isolated and multiple intersection scenarios, assuming the implementation of CV technology and comprehensive data obtained from CVs. The TSP control frameworks employed both single-agent Deep Reinforcement Learning and multiagent reinforcement learning (MARL) for signal optimization.
METHODOLOGY
An open-source traffic simulation software tool was used for the analyses.
- For the isolated traffic signal scenario, a four-legged signalized intersection in Charlotte, North Carolina (NC) was simulated for peak and off-peak hours using traffic volume data obtained from Charlotte Department of Transportation. The performance of six different signal control strategies, including Deep Q-Network (DQN) based TSP, were evaluated and compared. The performance of Pretimed Signal Controller (PSC) was taken as the baseline. Three metrics were used for isolated traffic signal scenario: average bus delay, average car delay, and average person delay.
- For the multiple intersection scenario, Central Avenue corridor which included five intersections, located east of downtown Charlotte, NC, was simulated. Eight MARL controllers were tested, which included Multi-Agent Proximal Policy Optimization with multi-discrete action (MAPPO-M) which enabled coordinated TSP by sharing information while making decentralized control decisions. Four metrics were used for evaluation: average travel time, average delay, average speed, and average number of stops (ANS).
FINDINGS
- In the isolated intersection scenarios, the DQN controller reduced average person delays by 18.77 percent and 23.37 percent in peak and off-peak hours, respectively, when compared to the baseline.
- In the multiple intersections scenarios, the MAPPO-M controller yielded the following benefits for buses: average travel time reduced by 21.78 percent, average delay reduced by 51.67 percent, average speed increased by 26.83 percent and the average number of stops decreased by 62.18.
- Similarly, the MAPPO-M controller yielded the following benefits for cars: average travel time reduced by 3.56 percent, average delay reduced by 9.00 percent, average speed increased by 2.81 percent and the average number of stops decreased by 9.49 percent.
