Research On Self-driving Decision Algorithm Based On DDPG Algorithm

With the rapid development of computer technology,unmanned driving technology is gradually entering people’s field of vision and becoming possible.However,the realization of unmanned driving still needs many ways to go,such as safety,stability,versatility and other issues.Deep reinforcement learning algorithms combine the strategy generation ability of reinforcement learning with the approximation and fitting ability of deep learning to the model.,Can efficiently handle various complex and high-dimensional problems,and provide a way forward for unmanned driving.Many scholars have invested in the research of deep reinforcement learning algorithms.As a result,many excellent deep reinforcement learning algorithms have emerged.This article focuses on improving the DDPG algorithms that have been excellent in recent years,and conducted three sets of experiments based on this.The first set of experiments is the demonstration and simulation of the DDPG algorithm.First,the theoretical basis of the DDPG algorithm is discussed,and then the simulation platform TORCS that will be used in this subject is introduced,and some environment settings and mode settings of the simulation platform are given.The state data was analyzed,and finally the DDPG algorithm was simulated and analyzed on the platform,and improved ideas were put forward for the problem of too long training time and not many laps in the final completion of the DDPG algorithm.Next,we introduced the concept of entropy,designed an improved DDPG algorithm based on maximum entropy reinforcement learning,realized the breakthrough of the original algorithm from a single-step strategy to multiple distribution strategies,improved the problem of insufficient environmental exploration by deterministic strategies,and improved After the second set of experiments is completed,the data is sorted in the form of graphs and tables,and the algorithms before and after the improvement are rationally compared and analyzed,verifying the improved DDPG algorithm based on maximum entropy reinforcement learning.Effectiveness.Finally,in the second set of experiments,the problem of continuous collisions in certain corners in the initial training of the trolley was proposed.The idea of combining discrete and continuous was proposed.The original Actor network was modified and added to the original output layer.A set of parallel output layers of 6neurons,and the introduction of Gumbel-Softmax Trick method,designed an unmanned driving decision method based on gear control,and conducted a third set of experiments,which proved that the method can effectively improve The early exploration efficiency of the trolley has realized the diversification of algorithm control.