Applying first order logic decision tree induction to opponent modelling in No-limit Texas Hold'em Poker

Abstract

In this thesis methods will be discussed to attempt to improve the decision making for agents involved in playing Texas Hold'em poker. Poker is a card game which features uncertainty, hidden information, deception and an environment in which multiple agents compete with each other to win the game. Opponent modelling will play a key role in improving the decision making of the agent(s).

The opponent models will help the agents to adapt more quickly to the playing style of their opponents and help them to make a better prediction of their next actions. In order for the agents to cope better with changes in playing style or the environment, it is important that the models are dynamic and keep evolving during the game.

Tilde, a first-order logic decision tree induction algorithm, will be used to construct models from a dataset. In turn, Monte-Carlo tree search will yield the action with the highest expected value after simulation. The opponent models will guide the simulation.