The Association for the Advancement of Artificial Intelligence (AAAI) wrapped up its Man vs Machine Poker Challenge yesterday, and, as we had suspected, the human element won out, although the margin of victory was not overly large.
Poker is a tricky game to teach a machine. Computers are inherently deterministic and uncreative, which is why a game of "imperfect information", such as poker, is challenging to reduce to mathematical formulae sufficiently adaptable to handle chaotic real life scenarios yet reducible to stable computer code. Although a computer can calculate to perfection the odds of receiving a particular hand of cards, which is certainly a big part of the game, it cannot form judgments about whether a certain player is lying or not - which many poker devotees argue is even more important.
It can, however, approximate such judgments by storing an internal record of how a particular player plays certain kinds of hands to search for tendencies. This is where most of the AI work on computerized poker games goes: developing algorithms that track opponents' betting habits.
Games such as chess or checkers can be replicated well on a computer since such playing style judgments, though occasionally helpful in chess, for example, sheer number crunching can map out an overwhelming number of possible outcomes. The chess board is a closed, determinative place, easily modeled in code.
The Man vs Machine Poker Challenge sought to eliminate as much of the luck as possible by playing a poker variant known as duplicate poker, in which one teammate plays the same hand as the other teammate's opponent, and vice versa. By eliminating luck, duplicate poker reduces the game to its strategic fundamentals.
Phil "the Unabomber" Laak and Ali "the Prince" Eslami, professional players known for their mathematical dexterity, edged out the University of Alberta's computer poker program Polaris by about $400 over the course of the two day low limit match.
The more interesting work is yet to come, however; playing Texas Hold'em against one person, where only two cards are unknown, and one player's tendencies are to be studied, has been challenging enough for those keen on developing championship caliber poker software. Developing software that deals with multiple lying individuals and ten or more unknown cards will be exponentially more difficult.
Duplicate poker may be helpful in measuring scientifically how different individuals react to comparable situations, but the variant itself is something of a red herring, at least when it comes to developing strategic poker software. For one thing, by eliminating luck, the variant rewards the mathematically gifted, which plays to the computer's strength.
The variant also eliminates one of the most powerful psychological components of the game itself - the winning (or losing) streak. Such a psychological component may not exist in the program itself, but when it comes to the analyzing the psychological habits of an opponent, buffeted as that player may be by the randomness of the cards, the complexity of the game magnifies with the good or bad fortune that befalls that player.
Players may change their strategies - not to mention the amount of the wager - completely based on how confident or wealthy they feel at that time, and what cards have been good to them. Gamblers are a superstitious lot, and superstition has no rhyme or reason.
The machine has a long way to go. ®
Burke Hansen, attorney at large, heads a San Francisco law office