Research And Application Of Monte Carlo Tree Search Network In Autonomous Driving

The problem of autonomous driving is a sequential decision-making problem of complex behavior.The perception,decision-making,and control of unmanned vehicles can affect the trajectory in the traffic road and the performance of autonomous driving.In the early days of the development of autonomous driving,the autonomous driving problem was usually decomposed into multiple sub-problems,using the idea of divide-and-conquer to separately solve the optimal problem in each sub-problem.In recent years,AlphaGo’s victory over human professional players in the field of Go can prove that the Monte Carlo tree search network can solve local optimal or even global optimal in sequential problems.In this thesis,the autonomous driving problem is modeled,and the Monte Carlo tree search and convolutional neural network are used to solve the autonomous driving problem model.This thesis designs and implements an end-to-end training and decision-making Monte Carlo tree search network algorithm,captures the trial data through the on-board camera,and outputs the steering wheel angle to realize the automatic driving of the vehicle.The algorithm is mainly divided into three parts,Monte Carlo tree search,strategy network and evaluation network.Based on the Monte Carlo tree search,some processes in the selection,expansion,simulation and update of the Monte Carlo tree search are carried out Networking,in order to reduce the breadth and depth of the search tree,improve search efficiency,and greatly improve the intelligent level of the algorithm in autonomous driving decision-making.At the same time,a convolutional neural network with only 7 layers of convolution and 4 layers of full connection is designed for the computing unit carried by the unmanned vehicle,so as to effectively process real-time data and avoid lag in decision-making.The inputs of the strategy and evaluation network are the unmanned vehicle road condition information from the first perspective,the output of the strategy network is the selection probability of the steering wheel angle,and the output of the evaluation network is the danger degree of the current state of the unmanned vehicle.In order to verify the effectiveness of the algorithm,this thesis uses the open source unmanned vehicle simulation platform to realize the collection and processing of autonomous driving training data.After several hours of training on the strategy and evaluation network,the algorithm can effectively avoid obstacles and reach the designated destination in the simulation platform.Finally,this thesis designs and implements a low-cost miniature smart car hardware test platform,and completes the data collection,network training,algorithm testing and other processes in the laboratory environment.The final experimental results show that the algorithm can learn human knowledge in cornering,obstacle avoidance and path planning,and can have good autonomous driving performance in special scenarios.