Applying Bandit-Based Monte Carlo Tree Search to Playing Big Two

Abstract

We investigate the application of a Monte Carlo Tree Search-based approach to AI game playing for the card game Big Two, a sequential, multiplayer game of imperfect information. After implementing several different types of AI players, we evaluate them in their performance relative to one another, focusing on the interplay between the game tree search strategy, multi-armed bandit selection policy, and heuristic playout strategy. We find that an MCTS approach is generally effective in this domain, particularly when paired with a strong heuristic playout strategy and a properly tuned selection policy.