Monte Carlo tree search (MCTS) combined with the upper confidence bounds applied to trees (UCT) algorithm has brought forth many advances in game related AI. This includes general game players and programs for specific games such as Amazons, Arimaa, and Go. However, there often is a need for further enhancements beyond this to make the strongest player for a given game. Move groups are investigated in detail to determine their effect on MCTS as a general enhancement. The structures of move groups that increase performance in an artificial game are analyzed then applied without knowledge of move values. We noted that there is a speed increase in the tree while maintaining performance for a fixed number of simulations. Finally, move groups are applied to the game of Amazons successfully.
展开▼
