Research & Writing

Academia

Writing and research on software engineering, AI, computing, and digital culture.

Conference Papers

Artificial Intelligence · November 2025

Opponent Modelling for Training Monte Carlo Tree Search-Based Models in the Board Game Citadels

This work investigates how opponent modelling influences the training of a hybrid MCTS-RL framework applied to the multiplayer board game Citadels, which features hidden information, multiple players, and time constraints. The proposal decouples planning and execution through a persistent decision structure built offline from simulations, reducing the online computational cost of MCTS. In a simulation environment with expert agents and a random agent, different training sessions (including a varied environment) are compared by win rate. The results indicate that the diversity of strategies in training tends to produce a more robust and generalist agent than training against individually strong opponents.

Theses & Monographs

Artificial Intelligence · 2025

Persistent Monte Carlo Tree Search for Agents Operating in Complex Environments: A Case Study on Citadels

Decision-making in multiplayer, partially observable games poses significant challenges due to hidden information and large branching factors. This study introduces a persistent Monte Carlo Tree Search (MCTS) framework designed to operate efficiently in such environments, using the strategic board game Citadels as a case study. The key contribution is the decoupling of the tree construction phase from the application phase: decision trees are generated offline through environment simulations, and relevant statistics are stored in tabular form, enabling rapid querying during execution without the need for additional simulations. Experimental evaluation demonstrates that exposure to diverse opponents during training produces models with robust and generalizable strategies, capable of strong performance across a wide range of game scenarios. In particular, models trained in an environment where opponents were stochastically sampled consistently outperformed those trained against single opponent types and exhibited stable performance across training runs. These findings underscore the effectiveness of combining MCTS with persistent knowledge storage to produce fast, reliable, and adaptable agents in complex domains.