A Reason to Root for the Machines?

Followed that “Man-vs.-Machine” poker match with keen interest. After a rough start, it looks like Phil Laak and Ali Eslami managed to comeback yesterday to “win” the contest against Polaris, the poker-playing computer program developed by the University of Alberta’s Computer Poker Research Group.

The match took place at the Hyatt Regency Hotel in Vancouver, Canada as part of the annual meeting of the Association for the Advancement of Artificial Intelligence. As someone who has been to boring academic conferences before, this here meeting of the AAAI sounded like a hell of a good time, comparatively speaking.

There’s a neat live blog of the event, eagerly written by a reporter whose enthusiasm is pretty infectious. Complete hand histories from the four sessions has also been promised. Not sure if those will be appearing on the official match website or over on the Poker Academy website. (Someone from Poker Academy also has provided another live blog of the event, if you’re interested.) My understanding is that some of the CPRG’s research has been incorporated into Poker Academy’s poker-playing software already. Reading around on the Poker Academy site, it looks as though they will be making it possible for us low limit hacks to take a crack at Polaris sometime in the near future.

Of the four sessions, Laak and Eslami won two and lost one, with the other being declared a “statistical tie.” Each session involved 500 hands of $10/$20 limit Hold ’em, played as “duplicate” matches. That means the pair played a total of 4,000 hands against Polaris, and in the end came out with only a small advantage in terms of number of small bets won.

Both the humans and Jonathan Schaeffer -- chair of the University of Alberta’s Department of Computing Science and leader of the CPRG -- acknowledged that the match represented a much too small sample size from which to extrapolate any profound claims regarding how near or how far researchers are from “solving” two-player limit Hold ’em. Schaeffer does think it will happen, though, in the near future.

I think I get how studying games like checkers, chess, or poker affords useful information for those interested in “advancing” artificial intelligence as well as our understanding of how other systems subject to uncertainty tend to function. While contemplating the possibility of someone actually constructing a computer program that will always make the best choice in every possible situation in a two-player limit Hold ’em match, something occurred to me. Something regarding that tired old “skill-versus-luck” saw. Tell me if this makes any sense:

If it is possible to create a computer program that will always make the best choice in a given game, then that game necessarily requires “skill” in order for anyone to be consistently successful playing it.

Take tic-tac-toe -- a much easier example for a pea-brain like myself to consider. You and I learn the rules and some basic strategy, then play 100 games, with every game ending in a draw. We each possess enough “skill” to “play perfectly” (so to speak), which is why every game ends in a tie.

Here . . . try your luck against a computer to remind yourself what I’m talking about (click on ‘Computer vs. Player’ to begin):



If tic-tac-toe were a “a game subject to chance” -- I deliberately am using language from the UIGEA here -- skill-levels wouldn’t matter. But think about it. If you happened to lose a game up there against the computer, it certainly wasn’t bad luck that lost it for you. You made a wrong move. Tic-tac-toe is not a game subject to chance. There is such a thing as “perfect” play that ensures skill is always going to be more significant than luck in the play of the game. (By the way, this computer doesn’t always play perfectly -- you can beat it. Keep trying.)

Now consider limit Hold ’em. You and I learn the rules and some basic strategy, then play 100 hands. I win some hands, you win some hands, and in the end one of us likely has more chips than the other. Neither of us possesses enough “skill” to “play perfectly,” of course, a fact that does have some effect on how our little match ends. Also relevant here: limit Hold ’em is a game subject to chance. Thus, the cards we were dealt also influenced how we came out.

However, if the researchers at the University of Alberta are able to reach their goal of constructing a program that always makes the best choice in every possible situation in two-player limit Hold ’em, that program is going to come out ahead against us humans who don’t possess the same degree of “skill” -- given an adequate number of hands, of course. Would not such a finding definitively prove that poker -- or at least two-player limit Hold ’em -- is most certainly a “skill-based” game?

Of course, such a finding would not prove that poker is not “a game subject to chance,” I don’t believe. But it would help distinguish poker from, say, something like the lottery which can never be “solved.”

Wouldn’t it?