Human-like Behavior Strategy for Autonomous Vehicles Considering Driving Styles

Autonomous driving technology plays a crucial role in enhancing driving safety and efficiency, with the decision-making module being at its core. To achieve more human-like decision-making and accommodate drivers with diverse styles, we propose a method based on deep reinforcement learning. A driving simulator is utilized to collect driver data, which is then classified into three driving styles—aggressive, moderate, and conservative—using the K-means algorithm. A driving style recognition model is developed using the labeled data. We then design distinct reward functions for the Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC) algorithms based on the driving data of the three styles. Through comparative analysis, the SAC algorithm is selected for its superior performance in balancing comfort and driving efficiency. The decision-making models for different styles are trained and evaluated in the SUMO simulation environment. The results indicate that the aggressive model prioritizes efficiency over comfort, while the conservative model emphasizes comfort with reduced efficiency. This approach successfully accommodates the decision-making preferences of drivers with varying styles, demonstrating human-like decision-making capabilities.