Wednesday, February 04, 2015

[Rant] I've Solved Poker, Says the Fish

Anyone who has ever gambled knows a guy who claims to have a system. Everyone has a buddy who "always wins at blackjack". If we're friends, then you have a buddy with irrational confidence in his football betting system. But nobody really has a system, or at least nobody has one for long. This is what came to mind when there was a big todo about someone having a sure bet computer program for winning poker.

There is no such thing and I'll tell you why. First, when they say poker, they mean a very specific variation of poker which nobody plays. They mean Texas Hold'em, which is fine because it is the most common game; with a limit on amounts bet per hand -- OK, most Hold'em games are played with no limit, but limit games are legitimate and not uncommon; and the game has to be heads-up -- meaning just two players at the table, which never happens except at the end of tournaments or in dedicated heads-up tournaments, and then it is almost always no limit. I know of no such thing as a Texas Hold'em limit heads-up tournament, nor do I know any casino that runs a limit heads-up money game. So to start with they have solved poker by reducing it to a pretty much non-existent mutation of the game.

Next, they make liberal use of the term "long run." Since even this program cannot win every hand, it is easy for a player to gain an advantage in the short term and then just quit. Computer loses. "That's not fair you, didn't play long enough" does not fly. There is no whining in poker, unless you are Phil Hellmuth. Even if it can be proven that given enough time the computer's winnings will be greater than zero, I'd still prefer to be the guy who had a hot run, won big, then walked away. That is to say, you don't "win" poker by being up a $100 over three years of steady play, you win by being up $100,000 after one crazy night and then spending it on hookers and blow.

Lastly, it appears this program constantly hones its play via a feedback mechanism by which it determines what actions have proven right and wrong in previous hands and then adjusts its behavior accordingly -- how often and in what circumstances a bluff should work, for instance. But I cannot see how this optimization can't depend on expectations of what the opponent will do based on the opponent's past behavior, and any human being, and particularly a skilled poker player, can alter his behavior at will. If they've gotten around this issue I'd very much like to know how.

I strongly suspect this will be debunked in the upcoming weeks. I also suspect a skilled player could have this computer smoking at the ears and crying "Norman, coordinate!" without much effort.