Building a computer poker agent with emphasis on opponent modeling

Abstract

In this thesis, we present a computer agent for the game of no-limit Texas Hold'em Poker for two players. Poker is a partially observable, stochastic, multi-agent, sequential game. This combination of characteristics makes it a very challenging game to master for both human and computer players. We explore this problem from an opponent modeling perspective, using data mining to build a database of player styles that allows our agent to quickly model the strategy of any new opponent. The opponent model is then used to develop a robust counter strategy. A simpler version of this agent modified for a three player game was able to win the 2011 MIT Poker Bot Competition.