Petruf, Michal (2026) Extending Multiple-Observer Information Set Monte Carlo Tree Search with Bayesian Opponent Modeling in Phantom Tic-Tac-Toe. Bachelor's Thesis, Artificial Intelligence.
|
Text
BachelorThesisMichalPetruf.pdf Download (1MB) | Preview |
Abstract
Standard search algorithms for imperfect information games, such as Information Set Monte Carlo Tree Search (ISMCTS), often assume uniform random opponent play, leading to suboptimal evaluations and wasted computational resources—a fundamental flaw known as Nonlocality. To mitigate this, this thesis proposes OM-MO-ISMCTS, an architecture that inte- grates explicit Bayesian opponent modeling into the Multiple-Observer ISMCTS framework. By maintaining a dynamic belief distribution over a portfolio of heuristic opponents, the algorithm generates an aggregated mixed strategy to bias both the determinization of hidden states and the tree selection phase for opponent nodes. The framework is evaluated in the partially observable domain of Phantom Tic-Tac-Toe through a formal ablation study of 480,000 independent games. Results demonstrate that the complete OM-MO-ISMCTS agent achieves statistically significant exploitative improvements against structured adversaries, peaking at a 90.42% win rate. Fur- thermore, the architecture surpasses the absolute peak performance of the baseline algorithm using only a quarter of the iteration budget, operating more than twice as fast in real-time ex- ecution. The ablation study suggests that Biased Selection is the more influential component in this setup, indicating that mitigating forward-looking nonlocality has a larger exploitative impact than resolving backward-looking state determinization alone. Ultimately, despite limited Bayesian belief convergence attributed primarily to the environment’s severe information sparsity, the synergistic combination of probabilistically biased determinization and selection provides superior foundation for real-time online planning against sub-optimal adversaries.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Dibangoye, J.S. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 08 Jul 2026 13:42 |
| Last Modified: | 08 Jul 2026 13:42 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37396 |
Actions (login required)
 |
View Item |