Reinforcement Learning applied to the game of Poker
José Pedro Neto dos Santos Marques · 2013
Games have always been a field of interest for Artificial Intelligence, as Games have usually simple rules. However in order to play them at a competent level, some degree of complex strategies are required. Despite this, some good (and well known) advances were made, specially in games like Chess or Checkers. But while these advances were important, results in games like Chess or Checkers deterministic games with complete information are hard to adapt to real world problems, as it is rare to have situations with complete information, and real world problems usually have stochastic variables. Thus the research has shifted into stochastic games with incomplete information, like Poker. There have been some excellent results, and recently a computer agent was able to win against a world class human player, the culmination of years of research and many different approaches. While these are indeed excellent results, the focus of this dissertation is somewhat different. It is not intended to create a world class Poker player, but one agent that can learn how to play a game from scratch. This approach could lead to more generic architectures, allowing them to be ported to another games more easily. In this work, a Reinforcement Learning approach was pursued, and the Poker variant Texas Hold’em was used as a test bed. This work shows two different agents that were capable of learning the game from scratch, achieving positive results versus a great variety of adversaries. It shows the learning structure and the decisions made during the development of such agents, concluding that while this is a viable approach to the game of Poker, the more generic the agent is, the worse it performs.