Adaptive Traffic Signal Control Through Time Period Division and Deep Reinforcement Learning

Abstract

In this paper, we propose an adaptive traffic signal control model combining time period division and deep reinforcement learning to improve the efficiency of traffic by dynamically changing the traffic phase duration according to the real-time situation. In our model, a day-time period is divided into two overlap period parts representing the morning situation and the evening situation, then the deep reinforcement learning algorithm-TD3 is selected to train the corresponding agent in each part, and finally a fuzzy method is used to coordinate these two agents at different time. In order to get better performance, we make some improvements in TD3. We improve the algorithm’s experience-replay mechanism and use some tricks in training. Simulation results shows that our model can effectively reduce vehicles’ accumulative waiting time, queue length and alleviate CO2 emission.