Autonomous Vehicles' Decision-Making Behavior in Complex Driving Environments Using Deep Reinforcement Learning

Automated vehicles (AV) are considered the key element of intelligent transportation systems in the future. Most studies about AV's behavior decision are based on oversimplified driving environments. With multiple conflicting objects and behaviors in the complex driving environment, traditional modeling methods cannot make effective behavior decisions. This paper proposes a deep reinforcement learning (DRL) method to solve the problem of AV's decision behavior modeling in complex environments. The DRL model is based on a deep deterministic policy gradient (DDPG), considering rule-based constraints. DDPG can make an action choice in a continuous-value space. Rule-based constrains are composed of safety constraints and dynamic constraints, added into the learning process to avoid unreasonable situations. Three reward functions are discussed in this paper. Compared with model convergence and validity, the reward function that integrated both safety and efficiency factors performs best. This study validated the effects of constraints and the validity of the model.