Transit Signal Priority under Connected Vehicle Environment: Deep Reinforcement Learning Approach

Transit Signal Priority (TSP) is a traffic signal control strategy that can provide priority to transit vehicles and thus improve transit service and enhance transportation equity. Conventional TSP strategies often ignore the fluctuation of passenger occupancy in transit vehicles, leading to sub-optimal solutions for the entire system. The use of Connected Vehicle (CV) technology enables the adoption of a more fine-grained objective in optimizing traffic signals, such as person delay, by allowing real-time information on passenger occupancy to be obtained. In this study, a deep reinforcement learning algorithm, deep Q-network (DQN), is applied to develop a traffic signal controller that minimizes the average person delay. The proposed DQN controller is tested in a simulation environment modeled after a real-world intersection and compared with pretimed and actuated controllers. Results show that the proposed DQN controller has the best performance in terms of average person delay. Compared to the baseline, it reduces the average person delay by 18.77% in peak hours and 23.37% in off-peak hours. Furthermore, it also results in decreased average delays for both buses and cars. The sensitivity analysis results indicate that the proposed controller has the potential for practical applications, as it can effectively handle some dynamic changes.