Efficiency of Static Knowledge Bias in Monte-Carlo Tree Search

Monte-Carlo methods are currently the best known algorithms for the game of Go. It is already shown that Monte-Carlo simulations based on a probability model containing static knowledge of the game are more efficient than random simulations. Some programs also use such probability models in the tree search policy to limit the search to a subset of the legal moves or to bias the search. However, this aspect is not so well documented. In this paper, we describe more precisely how static knowledge can be used to improve the tree search policy. We show experimentally the efficiency of the proposed method by a large number of games played against open source Go programs.