Using Machine Learning to Predict Winners of MCTS Variants in Games

Monte Carlo Tree Search (MCTS) is a widely utilized search algorithm for developing agents that can play board games intelligently. Numerous variants of MCTS have been proposed by researchers. However, it remains challenging to ascertain which variants are best suited for specific types of games. The effectiveness of certain MCTS variants across different game types is still unclear. This study employs machine learning to develop several models that predict the performance of one MCTS variant relative to others in a given game, based on the game's features. Among the individual models, LightGBM demonstrates superior predictive performance, while multi-model ensembles exhibit even greater efficacy. This research can assist game developers in screening MCTS variants to adapt to various game scenarios, thereby enhancing the efficiency of game development and design.