Research On Driverless Control Policy Based On Deep Reinforcement Learning

Autonomous driving has always been an important application in the field of artificial intelligence.The output of unmanned vehicle control strategies is one of the core issues that unmanned driving technology research needs to face.Therefore,this article applies deep reinforcement learning(DRL)technology to the control strategy research of unmanned driving technology,which can effectively solve the problem of decision control in the field of autonomous driving.This paper proposes the CEC-DDPG algorithm(DDPG with Classified Experience Conducted)based on the deep deterministic policy gradient(DDPG)algorithm.The core improvement is the addition of multi-experience pool classification storage mechanism and conductor network.So that unmanned driving can quickly get a reasonable control strategy to achieve unmanned autonomous driving.The CEC-DDPG algorithm defines Talent pool and General pool,and defines high-quality experience samples through TD-error comparison method and highreward episodes classified method.Based on the Talent pool,the CEC-DDPG algorithm specifically designed a Stochastic conductor network.The Stochastic conductor network uses the high-quality experience in its own Talent pool to guide learning,and jointly outputs the current control strategy with the actor network.The Stochastic conductor network is updated based on the constantly updated high-quality experience of the Talent pool,and changes the way the Stochastic conductor network is updated,which can effectively improve the learning efficiency of each network in the algorithm.This paper theoretically derives the network parameter updates of the CEC-DDPG algorithm.This article uses TORCS(The Opening Racing Car Simulator)as the simulation test platform for the unmanned driving control system.Aiming at the problem of unreasonable speed control of unmanned driving when turning,the reward function is redesigned to achieve the purpose of predicting curve items and braking in advance.The best sampling ratio of the Talent pool was determined through experimental debugging.Based on the TORCS simulation system,the CEC-DDPG algorithm was experimentally verified.Experimental data shows that the CEC-DDPG algorithm designed in this paper performs better than the DDPG algorithm in multiple aspects such as policy learning speed,tracking error performance,and control strategy generalization performance.The learning process is more stable and can complete unmanned autonomy.Exercise tasks.The feasibility and superiority of CEC-DDPG algorithm in the output of unmanned driving control strategy are verified.