Applying Machine Learning to the Game of Go

I consider the application of machine learning to the Japanese game of Go. The game of Go is an over 3500 year old game of strategy for two players based on very simple rules yet leading to highly complex strategies. Measured by various quantities such as number of possible positions and number of possible games, Go is by far more complex than Chess. As a consequence, the task of creating a computer program that plays Go well has been unsuccessful to date. This state of affairs is not likely to change very rapidly, because any brute force approach to the game is bound to fail for combinatorial reasons. It appears that only a complex interaction of pattern recognition, planning, and search would be able to match human opponents. Hence the game of Go appears to be an excellent testbed for techniques of artificial intelligence and machine learning in complex structured domains. Our work is driven by the idea that the hundred thousands of game records of games played by competent players avaiblable on the internet provide an invaluable resource for developing a strong computer Go program. For exploiting this data, we develop a graph based representation of Go positions and moves, and create feature spaces for learning the corresponding evaluation functions. Using kernel classifiers such as the support vector machine, we train the computer to find good moves for both artificial tsume Go problems and for 9x9 game play.

B.Hammer