Traffic Simulation with Multi-agent and Deep Reinforcement Learning

Abstract

Series of traffic accidents and traffic congestions happen every day in big cities like Tokyo. Therefore, it’s necessary to simulate the traffic condition and research on a method to control vehicles’ behavior well. Besides, autonomous driving is being developed rapidly nowadays and researchers often use deep learning to study trajectory prediction and path planning for autonomous vehicles. In this research, we use the shortest path search algorithm and deep reinforcement learning to control vehicles’ behavior in a traffic simulator SUMO. Regarding the local behavior which contains their speed and acceleration, we utilized deep reinforcement learning to control it. Regarding global behavior, which is path planning, we used a method combining Dijkstra algorithm and deep reinforcement learning. The vehicle agents in the simulator have better behavior after training. They can have acceleration and path selection that shorten their driving time when they encounter different traffic situations.