Autonomous Driving Policy Learning Based On Virtual Enviroment

Self-driving is an important area of artificial intelligence research.Self-driving is of great significance to traffic’s safety and efficiency.With the development of neural network in computer vison and natural language processing,increasingly companies and institutions are doving into deep learning for self-driving.The major domain of deep learning for self-driving research is learning of driving policy.Traditional driving policy is based on artificial design and mathematical modeling,which can not adapt to complex traffic environment.The goal of this paper is training more intelligent policy based on virtual traffic enviroment.The policy we get would guide opponent vehicles of self-driving training process in virtual driverless platform.This paper constructs a virtual environment based on driving theory and propose an algorithm that using deep reinforcement learning to obtain driving policy in virtual enviro-ment.This paper compares the results of different deep reinforcement learning models and obtain driving policy which exceeds traditional policy in mutiple traffic index.This paper gives constructive suggestions to future self-driving policy training based on experiments and analyses.This paper has two innovation points.Firstly,we design the state spaces of driving policy and corresponding neural networks and reward fuctions.Secondly,we analysis the effects of neural networks and traffic environments to driving performance.The driving simulation enviroment constructed by this paper could connect to different low-level traffic systems such as SUMO.We pay no attention to precise kinetic model and use simple kinetic model to express policy’s performance.