Prototype Design Of Autonomous Vehicle Simulation Based On Reinforcement Learning

With the continuous improvement of intelligent network,5G technology and the gradual decline of production costs,autonomous vehicle will usher in a period of comprehensive development.For a long time in the future,human driving vehicles,autonomous vehicles and auxiliary driving vehicles will coexist in the urban road network,forming a new pattern of mixed traffic flow.Therefore,for the related research of the modeling and simulation of mixed traffic flow,autonomous vehicle virtual simulation test will become a hot issue in the field of intelligent transportation.However,the existing traffic simulation software is lack of autonomous vehicle model,and autonomous vehicle simulation software can not restore the real traffic flow scene.In order to solve the above problems,this paper combines the advanced deep reinforcement learning algorithm in the field of artificial intelligence,carries out the secondary development based on the traffic simulation software SUMO,constructs the micro control and macro path planning model of autonomous vehicle,and carries out the related research on the co-simulation test of SUMO and Carla.Firstly,aiming at the problems of low accuracy,low efficiency and road network errors in the existing simulation road network construction,this paper proposes an intelligent road network optimization method based on multi-source map data,including lane number modification algorithm based on SHP and OSM data and a novel road ID modification algorithm based on road network topology.On this basis,combined with the theory of activity travel chain,the multi-mode travel input based on activity is completed.On the basis of realizing the intelligent construction and optimization of the simulation road network,a more real traffic scene is restored.Secondly,the paper analyzes the principle of deep reinforcement learning algorithm and the model mechanism of classic car following and lane changing model.On this basis,the traffic domain knowledge is integrated into the definition of reinforcement learning problem for micro control of autonomous vehicles,and the model is built by combining with deep reinforcement learning PPO algorithm.Combined with SUMO simulation scene,the results show that the autonomous model can complete the driving task more safely,quickly and efficiently.Then,based on the in-depth analysis of the classic path planning problems,the paper constructs the macro path planning model of autonomous based on deep reinforcement learning,and completes the model construction and training by combining the DQN algorithm of deep reinforcement learning and customized SUMO simulation scene.The results show that the autonomous vehicle macro control model can select the optimal path according to the change of road traffic after obtaining the information of the starting and ending points,and reach the destination faster.Finally,the paper integrates the micro control model of autonomous with the macro route selection model,and the ancient city of Suzhou is selected as a case city to test.The results show that the model has high generalization and can be widely used in many traffic simulation scenarios.On this basis,the co-simulation principle of SUMO and Carla is studied,the real-time co-simulation of the case city is realized,and the virtual simulation test of the whole chain is completed.