Game-Specific vs. Generalized MCTS Approaches : A Study Using Ludii’s Game Library

Details

Supervisors

Piette, Eric

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enThis thesis investigated whether measurable game properties can be used to predict effective Monte Carlo Tree Search (MCTS) configurations in General Game Playing, and whether such predictions can be exploited to improve Best Agent Identification procedures. To support this study, a large-scale experimental framework was developed within the Ludii General Game System, enabling systematic evaluation of MCTS variants across hundreds of structurally diverse games under controlled experimental conditions. Both One-Factor-at-a-Time and multi-component combination experiments were conducted, producing datasets for supervised prediction and warm-start transfer analyses. The results show that game-property-based prediction captures meaningful coarse per- formance trends across games. Tree-based models consistently outperform linear baselines, indicating that the relationship between game structure and MCTS performance is at least partially nonlinear. However, predictive accuracy remains insufficient for reliable direct best-agent recommendation. Multiple analyses suggest that the primary limitation lies not only in data quantity, but also in the expressiveness of the available game representations. Learning curves plateau at larger training sizes, simple feature engineering provides lit- tle improvement, and PCA and clustering analyses reveal only coarse separability with substantial overlap in feature space. These findings reinforce the broader challenge of adaptive search in General Game Playing: selecting effective search behavior from limited prior knowledge about a game remains difficult even when large-scale structural descriptors are available. Despite these limitations, the experiments also show that predictive models remain practically useful when used as weak priors rather than as final decision mechanisms. Warm- start Best Agent Identification experiments demonstrate that fractional prior initialization can improve early-budget search efficiency while still allowing rapid online correction when predictions are inaccurate. This suggests that offline learning and online adaptation should be viewed as complementary components rather than competing alternatives. Overall, a main contribution of this work is methodological. Rather than relying on direct zero-shot prediction of the best MCTS configuration, the results support a hybrid approach in which predictive models provide low-commitment guidance while adaptive online search remains responsible for robust final selection. More broadly, the results suggest that future progress in adaptive MCTS configuration may depend less on increasingly complex predictive models than on richer representations of game dynamics and tighter integration between offline learning and online adaptation.