Intelligent Agents for the Game of Go

Monte-Carlo Tree Search (MCTS) is a very efficient recent technology for games and planning, particularly in the high-dimensional case, when the number of time steps is moderate and when there is no natural evaluation function. Surprisingly, MCTS makes very little use of learning. In this paper, we present four techniques (ontologies, Bernstein races, Contextual Monte-Carlo and poolRave) for learning agents in Monte-Carlo Tree Search, and experiment them in difficult games and in particular, the Game of Go.