A PRELIMINARY STUDY ON COOPERATIVE CONTROL USING DEEP REINFORCEMENT LEARNING AT NONSIGNALIZED INTERSECTION

Abstract

In this study, cooperative control methods at a nonsignalized intersection are examined. Assuming that all vehicles are connected cars, cooperative control models for their smooth merging is constructed using deep reinforcement learning and evaluated using micro-simulation. Using Deep Q-Network and Soft Actor Critic, a type of deep reinforcement learning, a cooperative control model at a nonsignalized T-intersection that minimizes travel time while avoiding collision was constructed. The analysis results showed that this method is capable of efficient traffic control with high average speed. In particular, it was shown that the cooperative control method constructed by Soft Actor Critic can control traffic more safely and stably than Deep Q-Network one. In addition, it was confirmed that the cooperative control model constructed by Soft Actor Critic can control vehicles without collision and with less time loss than the priority non-priority control.