Research On The Method Of Feature Learning Of Go Based On K-means

Artificial intelligence is currently a hot research subject, and Computer Game is animportant branch of AI research. Since1950, the level of computer game programs is continuesincreasing.Most kinds of chess have been solved expect Go. With complexity, huge, and no clearfeature, it makes the level of computer go program is still at the same level of human amateurplayers.This paper innovatively use the idea of feature learning in the image recognition area, andcombine the current popular algorithm of Go which is UCB algorithm, and. This paperproposed a method of based on K-means clustering algorithm to learning features fromthousands of game records. The algorithm consists of two parts:1, K-means feature learning,which using K-means clustering algorithm to represent a huge number of board patches by asmall number of cluster centers. Due to the special property of Go, I make some appropriatemodifications to the K-means algorithm in order to get a reasonable correct results.2, win ratelearning, after vectorization the result of K-means clustering train the neural networks withthe win rate computed by UCB algorithm, use the output value to trim the game tree.By comparing to the baseline algorithm, the algorithm proposed in this paper can get up to70%of win rate with same amount of simulation, when the win rate is the same, it can save upto81%computation power.