Autonomous Driving Decision-Making Method with Deep Reinforcement Learning Integrating Attention Mechanism and Human Value Feature

Although Deep Reinforcement Learning (DRL) methods are generally considered effective solutions for designing intelligent decision-making systems, they still face challenges in solving efficiency and safety when applied to complex dynamic driving tasks. This paper proposes an efficiency-enhanced and safety-aware decision-making framework (AH-DDQN) for autonomous driving on highways to enhance both the solution efficiency of driving strategies and driving safety. The framework first utilizes an attention mechanism for environmental state feature representation learning, which captures critical environmental information and improves data efficiency during training. Subsequently, it incorporates human value feature into the policy during runtime to enhance decision-making safety. Experimental results demonstrate that the proposed framework exhibits superior performance compared to baseline algorithms.