Research On Autonomous Driving Decision-making Based On Deep Reinforcement Learning

Autonomous driving involves many technologies,such as perception,decisionmaking,control,communication and so on.As the development direction of traffic in the future,Autonomous driving can ensure traffic safety,improve traffic efficiency,reduce air pollution and bring great economic benefits to the society.Driving decision is the key to achieve safe,efficient and harmonious driving behavior,and learning the optimal decision in complex and changeable scenes is a challenge in the field of unmanned driving.Reinforcement learning,as a frontier direction in the field of artificial intelligence,is an effective method to solve the problem of driverless decision control.However,it has higher requirements in sample learning efficiency,multi scene learning ability and multi-agent cooperation,which is also a difficult problem at present.Therefore,based on deep reinforcement learning,this paper studies driverless decisionmaking,mainly including the following points:Firstly,the autonomous driving behavior decision algorithm is deeply studied.In view of the low learning efficiency of the traditional deep reinforcement learning algorithm,this paper proposes a TD3WD(Twin Delayed Deep Deterministic Policy Gradient with Discrete)algorithm integrating different action space outputs to improve the exploration efficiency of agents.The pre trained network is used to extract image features instead of the original image as state input,which reduces the computational cost in the process of reinforcement learning training.Secondly,for driving task learning in multiple scenarios,a behavior decisionmaking method based on hierarchical reinforcement learning is designed to classify sub tasks and design each sub task model.At the same time,in order to improve the training speed and convergence performance of subtasks,a course learning method is proposed,which improves the learning efficiency and convergence performance of the model.Thirdly,for multi vehicle cooperative driving task learning,a multi-agent cooperative decision-making method based on MADDPG algorithm is designed.The information exchange between agents is considered in decision-making.Experiments show that the Multi-Agent Reinforcement learning method can effectively improve the stability of decision-making.Finally,for the real vehicle deployment of the algorithm,the model of training convergence in simulation is migrated to the real vehicle for testing,which verifies the effectiveness of the reinforcement learning algorithm.