Research On Move Prediction In Go Based On Convolutional Neural Networks

Computer Game is an important research topic in artificial intelligence,to a large extent,the development of artificial intelligence has benefited from the development of Computer Game.As a major subject of Computer Game research,the game of Go owing to its enormous search space and the difficulty of evaluating board positions has made it difficult to acquire satisfying playing strength.In recent years,move prediction in Go based on convolutional neural network has gradually been an effective approach to master the game of Go.This is because of several reasons,first,the move prediction approach predicts human expert moves by supervised learning,which doesn't need deep search,thus it can combat the high branching factor in the game of Go;second,as the convolutional neural network can recognize visual patterns from the raw input image,and improve the accuracy of classification by feature extraction in each layer from the previous layer,it can evaluate the Go board positions effectively,which conquers the difficulty of modeling for ambiguous concepts computationally in the game of Go.Modeling a move predictor by using convolutional neural network has greatly promoted the development of Computer Go,but which has a huge development space in the respective of the current research.In this paper,we carry out the following three aspects of research which is in terms of move prediction in Go based on convolutional neural network:1)This paper presents a survey of move prediction in Go based on convolutional neural network.Firstly,we simply introduces the difficulties of Computer Go and points out that move prediction in Go is an effective way to solver Go AI by means of analyzing monte carlo tree search;then,in the aspect of the structure of convolutional neural networks,the structure of each layer and the training process of the network,we illustrates the basic knowledge of convolutional neural network correlated with move prediction in Go;finally,from the point of the feature representation of Go board positions,the characteristics of the network structure,the method of assessment and the analysis of performance,we summarizes the convolutional neural network applied to move prediction in Go,and the prospects of future studies in this field are presented.2)To improve the accuracy of move prediction in Go and its playing strength,this paper presents a new method of influence function based convolutional neural network for move prediction in Go.Firstly,we assess the influence distribution in a Go board position by using influence functions;then,we normalize these influence values and build up influence features for training a convolutional neural network;finally,we train a move predictor based on convolutional neural network with both influence features and others.Based on Zobrist's influence function,Chen's influence function and Yen's influence function,an experimental study is conducted in the paper,the results show that the strategy of combining with influence functions may improve the accuracy of move prediction in Go based on convolutional neural network and accordingly increase the playing level of the Go program.3)Based on the method of convolutional neural network for move prediction in Go,we design and develop DeepStone which is a software of human-computer Go machine.DeepStone can play the game of Go in 9 × 9 board,13 × 13 board and 19 × 19 board.Firstly,we simply introduces the GUI and the functional structure of DeepStone,then,illustrates key technologies in DeepStone including deciphering SGF file,convolutional neural network,monte carlo tree search,mathematical morphology and so on;finally,we assess the performance of DeepStone from actual games.The results show that DeepStone can beat stronger Go programs based on monte carlo tree search.Besides,DeepStone won the first prize(first place)in 9 × 9 Go,the first prize(first place)in 13 × 13 Go and the first prize(second place)in 19 × 19 board.