Monte-Carlo opening books for amazons

Abstract

Automatically creating opening books is a natural step towards the building of strong game-playing programs, especially when there is little available knowledge about the game. However, while recent popular Monte-Carlo Tree-Search programs showed strong results for various games, we show here that programs based on such methods cannot efficiently use opening books created using algorithms based on minimax. To overcome this issue, we propose to use an MCTS-based technique, Meta-MCTS, to create such opening books. This method, while requiring some tuning to arrive at the best opening book possible, shows promising results to create an opening book for the game of the Amazons, even if this is at the cost of removing its Monte-Carlo part. © 2011 Springer-Verlag.