Abstract

This paper details the results of plying several machine learning techniques to the task of predicting an opponent's next action in the poker game Texas Hold'em. Hold'em is a game of imperfect information, deception and chance played between multiple competing agents. These complexities make it a rich game upon which to use machine learning models because the emergent statistical patterns are often subtle and difficult for a person to recognize. To this end, we explore a variety of features for predicting action, showing which appear to be the strongest. Furthermore, we show that as more rounds of betting are observed for a particular hand, it becomes easier to predict action. Finally, we propose a feature to differentiate players in terms of their style of play that is easily calculated for new opponents in real time.

To read our paper, click here.