A Deep Reinforcement Learning-Based Cooperative Traffic Signal System Through Dual-Sensing Max Pressure Control

Improving the quality of traffic signal control (TSC) at urban intersections can alleviate congestion and environmental problems. To achieve better management of the traffic flows, current researches tend to apply deep reinforcement learning (DRL) for TSC in a traffic road network. However, most of the existing DRL-based methods often ignore the impact of differences in road capacities, particularly in heterogeneous intersections where traffic congestion is more likely to occur. To fill this gap, a noval traffic signal control strategy is proposed under the framework of DRL. In this paper, a new dual-sensing pressure of lane capacity and average lane velocity is developed based on the improved max pressure method. On this basis, the coordination between adjacent intersections is considered to suppress the problem of intersection congestion and even the spread to adjacent intersections caused by different road capacities. Moreover, a dynamically adjustable phase duration selection strategy based on realtime traffic information is designed to further optimize the average travel time. Through comprehensive experimental evaluation, proposed approach has been proven to surpass the performance of the previous TSC methods.