Monte-Carlo tree search and computer go

Abstract

The difficulty in positional evaluation and the large branching factor have made Go the most challenging board game for AI research. The classic full-board game-tree search paradigm has been powerless on Go even though this classic paradigm has produced programs with expert level performances in many other games. Three decades of research on knowledge and search did not push computer Go above intermediate amateur level. The emerging Monte-Carlo Tree Search (MCTS) paradigm is bringing an exciting breakthrough in computer Go toward challenging human experts, especially on smaller Go boards. This chapter gives an overview of both classical and MCTS approaches to computer Go. © 2009 Springer-Verlag Berlin Heidelberg.