A Decision-making Method for Self-driving Based on Deep Reinforcement Learning. (June 2020)

• Record Type: Journal Article
• Authors: Ke, Pang; Yanxin, Zhang; Chenkun, Yin
• Is Part Of: Journal of physics. Volume 1576(2020)
• Publication Date: 2020-06
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1576/1/012025
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: L5-level autonomous driving is the development trend of the future in the automotive industry, and the realization of autonomous driving through deep reinforcement learning algorithms are one of the research directions. Soft Actor-Critic the algorithm adds the maximum entropy term to the original deep reinforcement learning the objective function, and it shows great advantages in continuous control problems. Here, based on the open-source platform TORCS, this algorithm will be used to conduct automatic driving simulation experiments, design a reasonable reward function, add relevant constraints, use vehicle radar sensor information to make automatic driving decisions, and compare experiments with the Deep Deterministic Policy Gradient Algorithm. SAC can effectively extend training time, improve stability, and improve generalization ability.