Research On Intelligent Vehicle Autonomous Driving Strategy Based On Deep Reinforcement Learning

Autonomous driving is the most potential solution to improve traffic network efficiency and driving safety.Many companies and researchers have done a lot of research on all aspects of autonomous driving,and learning about driving strategies is the most important.The traditional learning method of driving strategy is to build mathematical model of vehicles and roads through artificial experience,which can deal with simple traffic scenes,but can’t do anything in the face of complex traffic environment,and the generalization ability of the model is weak.Inspired by the outstanding achievements of deep learning and reinforcement learning in various fields,this paper mainly studies the autonomous driving decision-making system of intelligent vehicles.Through the Deep Deterministic Policy Gradient algorithm in the depth reinforcement learning algorithm without model and self-learning,intelligent vehicles can drive autonomously.The environment information perceived by the interaction between intelligent vehicles and traffic scenes is used as driving strategy control According to the input of the controller,the controller makes driving decisions according to the perceived environmental information,judges the driving behavior with the reward value function,and then feeds back the results to the neural network,through constantly updating the neural network parameter value until learning the excellent driving strategy.Then the autonomous driving strategy learned by the intelligent vehicle is used to control the traffic flow state in the road network to achieve the desired traffic flow stability state and ensure the maximum average speed of vehicles in the road network,so as to improve the traffic efficiency of the road network.Firstly,the deep learning methods and reinforcement learning methods in the field of machine learning are analyzed,so the DDPG algorithm is selected as the algorithm for autonomous driving strategy learning of smart cars.The Markov decision process is used to design the state space of the decision-making controller for autonomous driving of intelligent vehicles,and the information of two aspects of the vehicle is used as the input of the driving strategy;the action space of the decision-making controller for autonomous driving is designed by analyzing the driving behavior of the vehicle.To avoid the conflict between the accelerator pedal and the brake pedal of the vehicle,an acceleration interval is set as the output of the driving strategy;the reward and punishment mechanism of the driving strategy is designed by considering the quantitative indicators such as the average speed of the vehicle,the headway,the waiting time at the intersection.Secondly,the theoretical realization process of applying DDPG algorithm to autonomous driving decision-making system is studied,and artificial neural network in deep learning is used as driving decision-making controller.In the process of the controller learning the driving strategy,the experience playback mechanism is adopted,which reduces the correlation between training samples.At the same time,when calculating the strategy gradient of the model,the mini-batch method is used to improve the training speed,which avoids the vehicle controller learning the local optimal strategy.Finally,build a ring road scene and intersection scene in the SUMO traffic simulation platform to realize the intelligent car autonomous driving decision simulation experiment.First introduced the traffic simulation platform SUMO and the reinforcement learning platform required in the experiment.then deployed the ring road and intersection traffic scenarios in the SUMO traffic simulation platform,designed the vehicle controller structure and the signal light phase.and finally carried out the data obtained by the experiment processing,through specific analysis of quantitative indicators such as reward value,vehicle position,average speed,and waiting time,to verify the feasibility and effectiveness of deep reinforcement learning algorithm in the intelligent vehicle autonomous driving decision-making system.