Summary

This thesis investigates a fundamental limitation in the Ludii General Game Playing system: simulation-based agents can involuntarily access hidden game-state information through unrestricted context copies. In imperfect-information games, this gives agents privileged knowledge and produces artificially inflated performance results that do not reflect genuine reasoning under uncertainty.

Battleship is used as a controlled hidden-position benchmark. Experiments first compare determinization strategies in the Tabletop Games Framework, then characterize the information leakage in Ludii. Native Ludii agents are shown to behave almost omnisciently because simulations expose the true positions of hidden pieces.

The main contribution is a generic determinization layer for Ludii. It intercepts simulation-time context copies, reconstructs the observations available to the player, and uses constraint-based hypothesis generation to replace hidden information with coherent possible game states. Existing search agents can use this layer without changes to their internal decision logic.

Experimental evaluation confirms that the layer removes the privileged-information advantage and restores meaningful reasoning under uncertainty. The determinized one-step look-ahead agent reaches an 87.5% win rate, closely matching the 87% observed in the reference framework under equivalent constraint-based determinization. The resulting agent rankings reflect search quality rather than access to hidden information.

The work establishes a foundation for reliable imperfect-information AI experiments in Ludii. Future extensions include hidden-identity games such as Stratego, richer belief-state representations, incremental history reconstruction, and support for games with more complex observation structures.

Suggested citation

Lechat, J. (2026). Improving Imperfect-Information Reasoning in Ludii through Constraint-Based Hypothesis Generation. Master's thesis, Université catholique de Louvain (UCLouvain).