Improving Play of Monte-Carlo Engines in the Game of Go

Abstract

We explore the effects of using a system similar to an opening book to improve the capabilities of computer Go software based on Monte Carlo Tree Search methods. This system operates by matching the board against clusters of board configurations from games played by experts. It does not require an exact match of the current board to be present in the expert games. Experimentation included results from over 120,000 games in tournaments using the open source Go engines Fuego, Orego, Pachi, and Gnugo. The parameters of operating our matching system were explored in over thirty different combinations to find the best results. We find that this system through its filtering or biasing the choice of a next move to a small subset of possible moves can improve play even though this can only be applied effectively to the initial moves of a game.