Performance evaluation of counterfactual regret minimization and Monte Carlo tree search variants: A case study on belote

Serrao, Alexandre
(2025)

Files

SERRAO_81811600_2025.pdf
  • Closed access
  • Adobe PDF
  • 1.96 MB
SERRAO_81811600_2025_APPENDIX1.pdf
  • Closed access
  • Adobe PDF
  • 572.5 KB

Details

Supervisors
Piette, Eric ; Morenville, Achille
Faculty
Ecole polytechnique de Louvain
Degree label
Master [120] : ingénieur civil en informatique, à finalité spécialisée
Abstract
Imperfect Information Games provide a challenging setting for artificial intelligence due to stochastic events, large exploration costs and hidden information. While algorithms such as MCTS and CFR variants are already widely covered in the literature, their applicability remains under explored for more unconventional games such as Belote. This thesis implemented the Belote game in OpenSpiel, providing a new research tool for IIG’s, and evaluated the performances of known algorithms on this setting. Although traditional ISMCTS and MCCFR algorithms manage to provide a strong baseline and robust strategies, they still struggle with scalability. As for DeepCFR, although it shows promising learning signs, it still underperforms in a context as complex as the one of the Belote game. By providing tools and preliminary insights on the subject, this thesis opens the way for other researches on the adaptability of current IIG’s algorithms in high stress environments.