Lane Change Decision-Making through Deep Reinforcement Learning with Driver’s Inputs

Abstract

Drivers usually make better decisions than the driving system in complex driving situations at the current level of autonomous driving. Combined with the driver decision-making advantage, a new Deep Reinforcement Learning architecture named dc-DRL is proposed for the lane change decision-making tasks in this study. Through the input of external drivers’ control actions in experience replay buffer and corresponding revision in actor and critic network, safe and efficient lane change maneuvers can be achieved. The state and action space are notably treated as continuous, and the Deep Deterministic Policy Gradient algorithm is applied to our study. Extensive simulations are conducted in a realistic driving simulator, CARLA, to test the safety and efficiency performance in single-agent and multi-agent environment, and the results show that the proposed method outperforms the benchmark method in driving safety and efficiency for lane-change maneuvers, with a 32% reduction in collision rate and a 13% reduction in traveling time to complete a designed unit section.