Intelligent Driving Path Planning Based On Deep Reinforcement Learning

In recent years,with the traffic congestion and road safety caused by the increase of vehicle ownership,the intelligent automobile has become the focus of the future development of the automobile industry.Intelligent driving vehicle is a multi-disciplinary complex of sensors,network communication,navigation positioning,artificial intelligence,among which navigation positioning,path planning,behavior decision and vehicle control are the key technologies of intelligent driving.This paper studies the path planning of intelligent driving,and uses the deep reinforcement learning algorithm to complete the experiment,algorithm improvement and simulation verification of intelligent driving path planning in virtual environment.Firstly,it analyzes that the enhanced learning is to match the state and action strategy by the trial and error learning of the agent in the environment.The paper describes the elements and process of reinforcement learning,and analyzes the theory,framework and network structure of reinforcement learning algorithm.DDPG algorithm can produce deterministic action strategy,and it has excellent performance in continuous control problem,and can be used in intelligent driving path planning.Secondly,the paper presents a sample demonstration of global path planning for global path planning selection map software.The map can provide the shortest path and the most unobstructed path from the starting point to the destination.For local path planning,the TORCS platform is selected to simulate the local path planning experiment of intelligent vehicles.According to the sensor and controller information provided by TORCS,the network structure of DDPG algorithm is designed,and the reward function is designed considering the safety and stability of the vehicle.The whole system obtains the original sensor input from the simulation environment,outputs continuous acceleration,steering and braking behavior,realizes the control of intelligent vehicle,makes it can drive safely according to the desired track,and completes the simulation experiment of intelligent driving local path planning based on open source platform TORCS using DDPG algorithm.Finally,the experience playback mechanism of DDPG algorithm is analyzed.It excellent experience samples,which leads to low learning efficiency.The idea of giving priority to experience samples by TD error is used for reference.The priority evaluation of experience samples is carried out by using immediate reward value,which makes the probability of more valuable experience samples being sampled is greater and the training efficiency is improved.In the contrast experiment,both algorithms can complete the local path planning of intelligent driving.The RP-DDPG algorithm based on reward value priority experience playback mechanism is more efficient than the traditional DDPG algorithm,and the driving stability of intelligent vehicle is better.