Research On Decision Control Of Autonomous Driving Based On Deep Reinforcement Learning

With the rapid development of Internet technology and communication technology,the intelligent and networked technologies of vehicles are gradually improving.As an important part of intelligent transportation,unmanned tech nology is becoming more and more mature.At present,most of the driverless cars use traditional local path planning and vehicle control algorithms,while sensing,decision-making,and control are independent modules that cannot guarantee high-precision time synchronization and spatial synchronization.For autonomous driving,the closer the above three modules are combined,the higher the safety and accuracy of autonomous driving.With the development of artificial intelligence technology,deep reinforcemen t learning provides another solution to the decision-making control problem of complex systems.This paper studies the unmanned decision control method based on deep reinforcement learning,and analyzes the security of the dynamic driving environment in the open racing car simulator.The main research work of this paper is as follows:Firstly,a reinforcement learning interaction structure for unmanned driving is established.Based on the environmental parameters returned by the TORCS simulation environment,the states and actions in the reinforcement learning algorithm are defined.Based on the expected driving performance,the reward function of the reinforcement learning algorithm is designed and the termination condition of the training is designed.According to the principle of deep deterministic policy gradient algorithm and the state and action requirements of the unmanned environment,the policy,the value and the target neural network are established respectively.The training and parameter updating methods of the above network are analyzed,and the whole network framework of the reinforcement learning algorithm is built.According to the state of the vehicle during the driving process,the training method of empirical buffer pool separation is designed.Aiming at the problem that the initial random noise of the exploration policy is large,which leads to a large number of invalid explorations,an improved exploration policy is proposed.By means of guidance,the probability of the vehicle being biase d toward the correct direction is greater.With the goal of minimizing tracking error and heading error,the vehicle's lane keeping exploration policy is improved,so that the vehicle can quickly learn the steering policy in the correct direction at the beginning of training.Based on the artificial potential field method,the exploration policy of vehicle overtaking and collision avoidance conditions is improved,and the low return exploration process is reduced.The improved deep reinforcement learning algorithm is simulated and verified on the open racing car simulator.The simulation results show that the improved exploration policy avoids the local optimal situation and increases the proportion of high-return samples in the sample pool.The experience p ool separation method solves the problem that the sample distribution is uneven and the neural network training is unstable.