AI Poker Agent for 5-Card Stud Miniatures

Abstract

Poker is a prime subject for artificial intelligence and in the last 30 years, various bots and new techniques inspired by computer solvers have gained prominence in the AI community. The reason this topic has risen in popularity is because of the difficulty inherent in an AI approach to poker. Unlike chess, go, etc., poker is a game of incomplete information, which makes machine learning approaches significantly more difficult. At any given point in a poker game, there is a high level of uncertainty from the opponent's hidden cards to the opponent's strategy which is constantly changing. To make matters worse, the poker game tree (branching factor over $10^{18}$) is several orders of magnitude larger than chess (branching factor of 47) and returns for even the best human players are highly variable.

In this paper, we introduce a new approach to 5-card stud poker AI. At a high level, our strategy consists of two AIs: an opponent model and an action model. The opponent model attempts to estimate a probability distribution over our opponent's possible hands and the action model uses this hand distribution to create a mixed strategy at any given node of the game tree.