Improving Imperfect-Information Reasoning in Ludii through Constraint-Based Hypothesis Generation

Details

Supervisors

Piette, Eric

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enGeneral Game Playing (GGP) systems aim to support autonomous AI agents capable of reasoning across a wide variety of games without game-specific tuning. While classical GGP frameworks were primarily designed around perfect-information assumptions, many tabletop games involve hidden information that fundamentally changes the nature of the AI problem. In such settings, agents must reason under uncertainty rather than over a fully observable game state.This thesis investigates a concrete architectural limitation of the Ludii GGP system: simulation-based agents inadvertently access hidden game-state information through unrestricted context copies, producing artificially inflated performance figures in imperfect-information games. Battleship is used as a case study, a hidden-position game in which the opponent's fleet is completely invisible to the acting player.Three contributions are presented. First, a controlled imperfect-information benchmark is established in the Tabletop Games Framework (TAG), demonstrating that simulation consistency matters more than search depth in hidden-position environments; three determinization strategies are evaluated (random, constraint-based, and heatmap-guided), with the constraint-based strategy proving sufficient to restore coherent reasoning and the heatmap extension bringing no measurable additional benefit. Second, the information leakage problem in Ludii is formally characterized and empirically confirmed through controlled tournaments, showing that native simulation-based agents reach near-perfect win rates against non-simulating baselines purely through involuntary access to privileged state. Third, a generic determinization layer is designed for Ludii, intercepting simulation-time context copies and replacing hidden information with coherent hypotheses derived from accumulated observations, requiring no modification to existing agent logic and only a single backward-compatible extension to the engine.Experimental evaluation confirms that the determinization layer eliminates the leakage advantage, restoring win rates consistent with genuine reasoning under uncertainty. The re-implemented OSLA agent reaches 87.5% under constraint-based determinization in Ludii, closely mirroring the 87% observed in TAG under a matched time budget, providing empirical validation across both frameworks.