AlphaZero Approach to Modern Tabletop Games with Imperfect Information

Sammanfattning

Artificial intelligence (AI) in games has achieved superhuman performance in perfect information games like Go and chess. However, AI for games with imperfect information remains an active area of research. This thesis develops an AI agent for BattleLore, a modern tabletop game featuring hidden information implemented in the Tabletop Games framework (TAG). Modern board games bring new challenges as they often incorporate some hidden information in their gameplay and have more complex game states than traditional games, such as chess. AI techniques that have been very successful in traditional games, such as Monte Carlo Tree Search (MCTS), struggle to make satisfying decisions in environments that are not fully observable, fail to utilize hidden information effectively, or require human expertise. To address these challenges, this thesis explores an AlphaZero approach that combines Information Set Monte Carlo Tree Search (IS-MCTS) with a neural network. The neural network is used in evaluating game states and in guiding IS-MCTS. The proposed agent is evaluated against baseline agents, including random, One Step Look Ahead (OSLA), Rolling Horizon Evolutionary Algorithm (RHEA), and MCTS agents, to determine its performance. The results demonstrate that the proposed agent outperforms the simpler agents: random and OSLA, and achieves similar performance compared to the more strategic agents: MCTS and RHEA. And the main factor contributing to the performance of the implemented AlphaZero-like algorithm is the use of IS-MCTS search during the decision step.