Opponent modelling in multi-agent hidden-information games
| Témavezető: | Király Tamás |
| ELTE TTK, Operációkutatási Tsz. | |
| email: | tamas.kiraly@ttk.elte.hu |
Projekt leírás
Opponent modelling studies how an agent can use observed actions to form predictions about another agent's behavior in a game. In certain settings (e.g., in a poker tournament with both weaker and stronger players), game-theoretically optimal (GTO) play might be regret-free, but strategies that focus on exploiting weaker players might have an advantage over GTO players, because they could use their winnings to create an asymmetric situation against stronger opponents later in the tournament. In this project, we consider a hierarchical family of simple opponent models to represent increasingly complex decision rules. The models will be combined using complexity-based prior weights and updated according to observed play, following ideas related to Solomonoff induction and Bayesian model averaging. The resulting predictions will be used for action selection in repeated imperfect-information games. The goal is to test the approach experimentally on a small game where this strategy might be suitable, such as Leduc poker.