Research On Vehicle Following Decision Algorithm Based On Deep Reinforcement Learning

Advanced driving assistance system is one of the researching focuses in the field of automotive active safety technology,mainly including automotive adaptive cruise system and front collision avoidance system.Since the current productized automotive adaptive cruise system on the market uses traditional rule-based vehicle following decision-making algorithms to make decisions,so it cannot achieve efficient,comfortable and safe follow,and often needs to abandon the comfort to ensure safety.Aiming at this problem,a vehicle following decision-making algorithm based on deep reinforcement learning is proposed,which comprehensively considers vehicle dynamics and rigid body collision theory and solves the problem of poor comfort during vehicle following.Mainly include the following work:This thesis models the vehicle following process as a Markov decision process.The distance between the vehicles,the relative speed of the front and rear vehicles,and the acceleration of the following vehicles are used as the state set,which is the input of the decision system;the expected acceleration of the following vehicle is used as the action set,which is the output of the decision system.The principle of driving risk,optimization goals,and personalized design are used as the basis for decision-making in the decision-making system,thus constructing a vehicle-following decision-making framework based on deep reinforcement learning.The thesis combines the minimum safety distance model and comprehensively considers the safety,comfort and efficiency of the vehicle following process,and designs a modular reward function.Based on the deep deterministic policy gradient algorithm,a vehicle following autonomous decision-making algorithm is proposed.The trained vehicle following model can follow the vehicle ahead safely,comfortably and efficiently in a variety of environments.Aiming at the problem of low efficiency in the use of experience samples in the deep deterministic policy gradient algorithm,the size of the time difference error and the instant reward value are used as the criteria to classify the pros and cons of the samples,and a new compound priority experience replay mechanism is proposed to improve the utilization efficiency of the experience samples and speeds up the convergence speed of the algorithm.On the gym simulation platform,comparing the compound priority experience replay mechanism proposed in this thesis with the existing mechanisms,the method in this thesis not only has lower time complexity when storing experience samples,but also speeds up the convergence speed and improves the training stability of the model.In a simulation environment,the performance of the vehicle following decision algorithm proposed in this thesis is tested.The experimental results show that the trained model can follow the pilot vehicle safely,comfortably and efficiently in different test environments,and the performance is better than the existing ones.