Abstract
This paper proposes a new approach to solving tsumego and playing games in computer go. The approach is based on the Monte-Carlo tree search (MCTS) algorithm. To be more specific, the proposed method improves how to execute a single playout on a leaf node. Our method brings diversity to playout so that only a particular part of the Monte-Carlo tree will not grow in the depth direction. The diversity of playout can be brought to realization by embedding tabu lists into all leaf nodes. Here, the tabu list means a short memory which is very popular in the tabu search algorithm. Tabu lists impose restriction on several moves from the first move during the playout so as to control the points of prohibited moves. If a candidate move corresponds with an element contained in the tabu list, then the move is treated as the prohibited move. It is possible that the diversity of playout could induce the behavior of MCTS to the best-first search on a more uniformly-broadened tree. This works well for solving tsumego which has the only path without branching. Furthermore, the variable-length tabu lists embedded into the MCTS would have an advantage in playing 9×9 go.
