PokerBot World Series 2004

This summer, the sixth International Conference on Cognitive Modeling will take place in Pittsburgh, Pennsylvania. While you might think this would be of limited interest to poker players, this year's conference will also host the PokerBot World Series, a competition to develop the best computer model of a human poker player. Organized by Christian Lebiere of Micro Analysis and Design and Dan Bothell of Carnegie Mellon University (Lebiere also has an appointment at CMU), the competition seeks to encourage the development of sophisticated models of complex cognitive function in the best possible testbed environment – poker.

Poker is actually a natural choice for a conference on cognitive modeling. The competition announcement lists nine cognitive domains taxed by poker, including everything from reasoning under uncertainty to pattern recognition. Depending on how you slice it, you could easily fit a few more in there, including mathematical reasoning, categorization, short-term and long-term memory, inhibitory control, and evaluation of risk. The organizers hope the competition will not only be an interesting addition to the conference program, but will also spur interest in a new and richly productive domain for cognitive modeling research.

The competition comes at an interesting time. These days, it's almost impossible to write an article on poker without referring to the boom in poker's popularity, reflected in both cardroom and online poker, and most visibly in the number of tournaments on television. It's almost inevitable that poker would expand into other areas, as well. While we probably shouldn't expect to see poker-inspired snacks in grocery stores or a big surge in poker-themed music, academic study of poker is an idea whose time has come – or rather, come again.

Academic study of poker isn't as big as, say, physics, but it's hardly new. Game theoretic analysis of poker (or at least pokerlike games) has been around since the beginning of the field of game theory. While application of game theory to actual game play has lagged behind, a few books have made the connection, the most familiar being David Sklansky's Theory of Poker, which draws on game theory and includes a chapter on game theory and bluffing. At least two other authors have tried to provide detailed game theoretic analyses of poker. Norman Zadeh's Winning Poker Systems in 1974 and Nesmith Ankeny's 1981 classic, Poker Strategy (subtitled Winning With Game Theory), were both written by academicians with an interest in bringing mathematical rigor to the game. And more currently, Darse Billings and colleagues at the University of Alberta have developed a broad research program that ranges from game theory to artificial intelligence, geared toward developing intelligent poker-playing computer programs.

Cognitive modeling adds an interesting twist to the academic study of poker. Where game theory is concerned with developing "optimal" strategies, and artificial intelligence is to some extent concerned with developing the strongest possible poker-playing computer program, cognitive modeling is more about developing humanlike programs – programs that not only make the same mistakes humans make, but also make them for analogous reasons. Of course, not all people play alike. A model of weak players may be just as valid as a model of strong players, or perhaps even more so, since strong players are harder to find. But since the PokerBot World Series is a competition, it's a fair bet that most of the models will be built more to win than to emulate human behavior.

Outside of academic interest in cognitive processes, what use is a poker-playing bot? Developing better models of human poker play is essential to answering difficult questions about value in poker. We can argue the relative merits of (for example) J-9 suited or A-10 in various situations until we're blue in the face, but while some arguments are more compelling than others, proving your case can be very difficult. Ideally, it would be great to play against typical opponents for a few thousand hours, keeping detailed statistics. The more accurately those models approximate the behavior of human opponents, the more valid the results will be. While there are commercial packages that attempt to solve this problem, there's nothing like a friendly competition to really drive progress.

Of course, writing a good poker-playing program is not trivial. You need to have good programming skills, detailed knowledge of poker, and the ability to translate that knowledge into a working pokerbot. If that sounds like you, or if you'd just like to read a bit more, visit the web site for the competition at http://simon.lrdc.pitt.edu/~iccm/pokerbot.html. The conference runs from July 30 through Aug. 1, but the actual pokerbot competition will be run in advance, so that the results can be discussed at a special symposium. It remains to be seen how many and what kinds of entries this first competition will draw, from both the academic and the poker-playing communities, but I've already started working on my entry. So far, it does a great job with the trash talk. I just have to add in the poker-playing part.

Daniel Kimberg is the author of Serious Poker and maintains a web site for serious poker players at www.seriouspoker.com.