|Faculty Advisor:||Dr. Mitch Marcus|
|Paper:||Final Paper [PDF]|
|Presentation:||Project Poster [PDF]|
This design project explores a fascinating application of artificial intelligence and machine learning to one of today’s most popular mental challenges: Texas Hold’em poker. Recently, IBM’s Deep Blue computer chess program managed to defeat the grandmaster Gary Kasparov utilizing an advanced brute-force search—a method dependent upon chess being a game of deterministic perfect information. This technique cannot be applied to the game of poker because poker is a game of incomplete information and chance outcomes, among other substantial difficulties. In this case, one does not know the cards that his opponents are holding and one does not know whether your opponents are accurately representing the strengths of their hands or if they are just bluffing. Furthermore, future situations are non-deterministic and modeled probabilistically. This raises several interesting complexities regarding optimal strategy in the game of poker.
The approach taken in this project is to identify a set of attributes that can define the important features of a single game state. Given a value for each of these attributes, the bot, like any strong human player, has enough information to decide what the proper action is to take. I use a database of game states and corresponding actions with which to train the bot, after which the engine is capable of classifying new game states into the proper action classes. An obvious platform upon which to administer this project is to interact with the well-known online poker application, PartyPoker.net®. This allows the bot to play with real people on a well-established platform for play money.
Following the strategies described in this paper, the bot that I developed achieves impressive results. Extensive play over time shows that the bot is losing an average of 3.27 big bets per 100 hands played. Regardless of the fact that it is a losing player, it is quite impressive that this value is so close to zero. There are plenty of foreseeable improvements, so its decent performance to this point shows promise that it has potential to succeed at these full tables pending future development.