Active Opening Book Application for Monte-Carlo Tree Search in 19×19 Go

The dominant approach for programs playing the Asian board game of Go is nowadays Monte-Carlo Tree Search (MCTS). However, MCTS does not perform well in the opening phase of the game, as the branching factor is high and consequences of moves can be far delayed. Human knowledge about Go openings is typically captured in joseki, local sequences of moves that are considered optimal for both players. The choice of the correct joseki in a given whole-board position, however, is difficult to formalize. This paper presents an approach to successfully apply global as well as local opening moves, extracted from databases of high-level game records, in the MCTS framework. Instead of blindly playing moves that match local joseki patterns (passive opening book application), knowledge about these moves is integrated into the search algorithm by the techniques of move pruning and move biasing (active opening book application). Thus, the opening book serves to nudge the search into the direction of tried and tested local moves, while the search is able to filter out locally optimal, but globally problematic move choices. In our experiments, active book application outperforms passive book application and plain MCTS in 19×19 Go.