A Survey About the Decision-making of Self-driving Based on Deep Reinforcement Learning

Self-driving mainly includes four parts: perception, planning, decision-making and control. The decision system is used to output safe and reasonable driving behaviors according to vehicle's state and surrounding environment, so as to autonomously control the vehicle to driving, which is the key part and core of self-driving. With the development of machine learning and AI, the self-driving technology gets further improvement. Using the deep reinforcement learning (DRL) algorithms to solve the behavior decision-making problems of self-driving vehicles has become one of the most concerned research hot spot. Traditional rule-based method lacks the ability to adapt the dynamic environments, DRL can overcome this drawback very well, it has unique advantage for processing those problems. In this paper, we summarize recent years research on DRL for decision making in the field of self-driving and provide a classification of driving tasks using these methods. And analyze in detail the advantages and disadvantages of DRL methods proposed by researchers for solving these driving tasks, and finally summarize some of the challenges and difficulties in applying DRL to complete decision-making problems and look forward to future research works and directions. The study contributes to a deeper understanding of the theories and methods related to DRL and decision-making of self-driving, and provides a reference for future research.