Traffic Simulation for Autonomous Driving Society

Abstract

This study presents a basic investigation of the possible behaviors of autonomous vehicles under mixed autonomy conditions. Specifically, the author focuses on two behaviors: acquisition of acceleration strategies as a local behavior and altruistic route selection as a global behavior. For the acquisition of acceleration/deceleration strategies, a reinforcement learning algorithm was employed to reduce both CO2 emissions and traffic jams. Simulations on a road segment with a traffic light showed that an appropriate strategy can be learned. For altruistic route selection, autonomous vehicles, considered to have a relatively low time value, were assumed to tolerate suboptimal routes. Simulations on a medium-sized city road network showed that the performance of the whole transportation system, including human-driven vehicles, was improved. It is expected that the control algorithm proposed in this study, operating under the assumption that traffic safety is ensured for autonomous vehicles, will provide additional benefits as well as prevent traffic accidents.