Research On Autonomous Driving Intelligent Decision Control Based On Deep Reinforcement Learning Method

Decision-making control of autonomous driving is the core of autonomous driving technology.It needs the perceived information of road scenes to make safe and reasonable decisions,especially for the uncontrollable emergencies in various situations.Faced with this problem,an intelligent decision-making method with strong autonomous learning and generalization ability is urgently needed.Because deep reinforcement learning can not only learn from zero autonomy but also has strong generalization performance.Moreover,it can realize direct control from original input to output through end-to-end way,this advantage is very suitable the automatic driving from perception to decision control situation.Therefore,the application of deep reinforcement learning technology in autonomous driving decision-making control,according to the travel vehicle situation provide intelligent strategic decision,it has a very broad application prospect and important research significance.Because the output action of autonomous driving is continuous and random exploration in autonomous driving action space may lead to unexpected consequences,and this thesis adopts Deep Deterministic Policy Gradient(DDPG)algorithm in deep reinforcement learning for intelligent decision-making control of autonomous driving.In view of the slow convergence and instability of the traditional DDPG algorithm,this thesis proposes a DDPG with Two Samples(DDPGwTS)algorithm based on the improved DDPG algorithm for autonomous driving decision-making control.On the one hand,the second sampling method is introduced in the experience replay stage.Firstly,the sequence in the experience pool is sampled according to the distribution of cumulative returns of the sequence,then do the sample is sampled according to the TD bias distribution of the sample in the above sampled sequence.Then,the sampled sample is used to train the algorithm to improve the convergence speed and the quality of the strategy.On the other hand,it is proposed that online network and target network use dynamic parameter tracking method to carry out weight parameters transfer to improve the stability of the algorithm and convergence speed.Finally,based on the artificial intelligence field outstanding open source racing simulator(TORCS)to make a precise design for the whole DDPGwTS algorithm frame detailed segment,to make it applicable to autonomous driving vehicles and performance is simulated and verified.This thesis has 21 figures,12 tables and 51 references.