Traffic Signal Control at Scramble Intersections by Deep Reinforcement Learning

Abstract

Shibuya Crossing has a large number of pedestrians and vehicles. To reduce the waiting time of intersection users, we tried to optimize the pattern and timing of changes by using deep reinforcement learning for signals. Existing studies often consider only vehicles. There are no studies on scrambled intersections as a destination for applying traffic control by deep reinforcement learning. In this study, we used a traffic simulation SUMO to construct the environment of the Shibuya Crossing. We then optimized the signals by learning them using three methods: DQN, A2C, and PPO. As a result, by using reinforcement learning, we succeeded in reducing the waiting time by a factor of four compared to the signal patterns currently used. We also clarified how the accuracy of the learning process changes when the method or state observation is changed and analyzed the behavior of the optimized signal.