ABSTRACT
Go (Iwamoto, 1976) is one of the hardest known games to play by machine. Its very large branching factor (up to 361) makes search to even a moderate depth problematic, and precludes the use of massive search that has proved successful in games like checkers and chess. As a result, researchers have turned to pure pattern recognition and pattern-based reasoning approaches, which concentrate all their effort in evaluating the current board position, and make no explicit attempt to predict the evolution of the game. However, systems of this type remain far from being able to compete with even a moderately proficient human player. This extended abstract reports on GP, an artificial Go player that combines pattern recognition with limited but flexible search, and won several games against a novice human player.
