Systems and Industrial Engineering Seminar
Building an Artificial Poker Player - A Blend of Probability and Statistics,
Decision Making, Game Theory, and Psychology
Mary Domm &
Jeff Goldberg
The University of Arizona
Abstract
Games of strategy and games of chance have long been an avocation of mathematicians
and engineers. Efforts range from the study of probability by Pascal
(motivated by the problem of developing winning gambling strategies) to the
encoding of chess playing algorithms in hardware and software in IBM's "Deep
Blue". Generally these studies have been limited to deterministic and
stochastic games with perfect information - for example chess and checkers,
or dice and video poker. The studies have been largely successful
to the extent that computer chess and checker programs can compete with the
best human players in the world and we can compute strategies that maximize
the expected value of stochastic games.
In this talk we discuss the game of poker in the framework of a stochastic
game with imperfect information. The stochastic element comes from
the random deal of the cards, while the imperfect information arises from
the deception used by opponents. Our eventual goal is to develop a
computer poker player that can compete successfully with human players.
We use the game of "Texas Hold'em" as our study game due to its simple structure
and popularity. Now, Texas Holdem can be played in legal poker
rooms across the country (except a few non-gambling states) and in virtual
poker rooms on the Internet.
In the talk we first motivate why poker is in some sense more difficult for
a computer to play well than chess. We review earlier efforts at developing
computer players and show that their major shortcoming is in opponent modeling.
If one cannot develop strategies based on your opponents and their behaviors,
then you will generally be unsuccessful at the game (you will be making many
decisions that have negative expectation). We discuss the components
of an opponent model and initial efforts at categorizing players using data
mining on a data set from IRC Poker games. We conclude with some
directions for future research.
.
Note: Refreshments will be served.
DATE AND TIME: 2 PM, Thursday,
November 29, 2001
PLACE:
Engineering Building (Bldg. #20), Room 30
1
Sponsored by the Department of Systems and Industrial Engineering.
For more information contact:
Meg Askey
, 621-2274