Research On The Game Of Go Based On Deep Learning And Mente Carlo Tree Search

Perfect information game has always been the main research direction of computer game,and computer game is an important field of artificial intelligence.It is one of the important indicators to verify the development of artificial intelligence.In the perfect information game,the simple and deep game of Go has a very high complexity and is an important means to test the level of computer game.The research results of the game of Go can be applied to other areas of artificial intelligence,including financial decision-making and motion control.In 2016,the Alpha Go launched by the team of Deepmind beat Lee Sedol who was the world championship of Go.However,this does not mean that the issues of the game of Go has been solved perfectly.An important step in the Monte Carlo tree search algorithm is simulation.The more times the simulation is performed in a given time,the more accurate the results of the Monte Carlo tree search algorithm are and the more accurate the returned results are.The traditional method uses a pattern-based approach to simulate,but there are some problems with pattern-based approach: the patterns need to be stored in the memory,and each time you need to compare the patterns in the board;the accuracy of the pattern-based method is not high,which has an impact on the simulation results.In order to solve the above problems,this paper uses a combination of deep learning and Monte Carlo search algorithm to play chess.We use Deep Learning to learn the game records and then we get a strategy network with more layers and another with fewer layers which named as rollout policy network.In this paper,we use the policy network and Monte Carlo search algorithm to search the game tree,and put forward the method of using the rollout network to make a quick move in view of the shortcoming of the traditional method.This method makes use of the rollout network to be fast so it can meet the time requirements in the simulation of Monte Carlo tree search algorithm.Compared with the traditional pattern-based method,this method has a higher accuracy in predicting the game records,as a result,the simulation results are more valuable.In addition,using the rollout policy network can make better use of GPU resources without the need for statistical computation of the patterns,overcoming the drawbacks of the traditional pattern-based approach.In this thesis,we apply the rollout policy network to simulate in Monte Carlo search algorithm.We implement the agent based on Deep Learning and Monte Carlo tree search,then we use the agent to play the game with the traditional pattern-based agent,and the results show that our method outperforms the traditional method.