A Distributional Soft-Actor-Critic-Discrete-Based Decision-Making Algorithm for Autonomous Driving

Abstract

In the decision-making process of autonomous driving, deep reinforcement learning can help vehicles better adapt to complex traffic environments, thereby improving driving safety and efficiency. We propose a Distributional Soft-Actor-Critic-Discrete (DSAC-D) algorithm based on Soft-Actor-Critic-Discrete (SAC-D) and the improved driving risk field, aiming to address the problems of poor stability, insufficient learning accuracy, and Q-value overestimation in existing deep reinforcement learning-based autonomous driving decision-making processes. The proposed algorithm has achieved a 44.26% improvement in reward value compared to the SAC-D algorithm in a complex environment with the same dimensions. Based on the algorithm, we designed a whole-vehicle autonomous driving system with efficient and accurate control capability, offering novel approaches for the development of next-generation autonomous driving,