A Reinforcement Learning Traffic Signal Control Method Based on Traffic Intensity Analysis
Lu Wei, Lei Gao, Jian Yang, Jinhong Li
Abstract
To enhance the efficiency and practicality of traffic signal control methods based on reinforcement learning, a novel traffic intensity calculation method which inspired by Max-Pressure (MP) approach is proposed. Unlike the MP method that considers only queued vehicles, the proposed method takes into account both stopped and moving vehicles on the approaches of an intersection, providing a more accurate characterization of demand for traffic signal control. In addition, to improve the practicality of the method, a traffic signal phase duration optimization strategy based on deep deterministic policy gradient (DDPG) is designed under a fixed phase sequence. Comprehensive experiments conducted on SUMO traffic simulation tools demonstrate that the proposed method outperforms traditional and baseline RL-based traffic signal control approaches in terms of various measures of efficiencies such as queue length, delay, and average speed.