Artificial Intelligence in Poker

Markian Hlynka writes "The University of Alberta's research into Poker AI is featured in this New York Times article. There is also detailed discussion of the game of Poker, and the 'new breed' of players who have honed their abilities online. See the U of A's poker project for more information."

Comments

Poker AI? riight...

[KDan]

Poker is not a card game, it's a people game (aka don't play the cards, play the people). It's all about bluffing and reading other people's bluffs. I'm baffled that people even bother playing poker on the internet. Even with webcams the game wouldn't be the same at all.

Daniel

Re:Poker AI? riight...

[AssFace]

reading the article is too much I see.

it references that and points out how much of "reading the player" is overhyped and easily faked out.

whereas the real information is in the trade at hand - the exchange of money. watching the bets and the amounts in them at varying spots in the game.

I have a few friends that have won online tournaments and they approach it from a very mathematical point of view. They do very well in person or over the net.
Using the "read" approach, unless the read is of the play on the table, is only going to work with people that aren't aware of the read and therefore not faking the attributes.

I personally prefer to look for the security holes in the online software :)
(There was a famous one in '96 or so where the system was using the random function built in - I think in Turbo Pascal IIRC - they had it exposed by posting their random code on the net to prove that they were being fair. A consulting firm then exploited that to show that they only needed to see one or two cards beyond what was in their card to then show what everyone else in the game was holding... there is much higher security in it all these days, and better/smarter programming).

Another firend in college found a site that had a hole, not in the security, but in the method at which they gave out tokens - as long as you kept playing, there was a reward of some number of tokens as an incentive to keep you playing.
He then ran some numbers and proved that with that, they were open to an exploit of the Martingale system. He ran it on them for a good amount of time and it failed - he basically proved that their code was cheating on the inside.
He called them on it and after a few heated e-mails, got all his money back and was banned from the site.
I could go on and on - but that is going off topic.

Gem

[mao che minh]

I was following various links about the topic (artificial intelligence and poker) when I found this little gem. Wicked awesome site design, so I can only assume that his software building skills are as magnificent: POKER WITH AI-LEARNING

I say we help him beta test not only his program, but also help him stress-test his web server.

Re:Tell me...

[rkent]

You RTFA:

Peter Muller, a friend of Mr. Rao's who has played against the same bot, said the approximations in the game-theory model left a weakness and limited the bot's chances to do more than break even. Game-theory models usually assume that every player uses the best possible strategy, something that rarely if ever happens with humans.

"An optimal game theoretic strategy might ensure that you don't lose, but it won't be effective at exploiting an opponent's weaknesses," Mr. Muller said. "The best players learn how to exploit predictability, but don't do it often enough so that the opponents catch on."

In other words, it's easy to bluff a computer; you just play strongly and it'll assume you have a good hand and probably fold to you. Unless it's got a good hand, in which case you're screwed. Or if it has adoptive modelling that remembers how often you bluff, then you're REALLY screwed. Generally, though, it sounds like the Alberta AI just plays tightly, using "classes" of hands to avoid getting confused by the billions of possible hands, which does limit losses, but doesn't generally win big.

For non-robots, a simplified poker method

[RobertB-DC]

I have never been any good at poker... in high school, playing nickel-ante poker, I lost about $25 to just one of my friends. Typically, after about 15 minutes of play, everyone was playing with "my" money.

But recently, I spent some quality time with a hand-held poker game, and played the "hundreds or thousands" of games as described in the article. Not enough to become an expert, but I did come up with a technique to make my 100 credits last longer.

I hacked away as much complexity as I could. The heart of my method is to forget about the effect of getting two cards you need. The chances of getting two specific cards is something like 1/52 * 1/52 = 1/2704 -- too small to care about. So the entire method is about the next card.

Of course, I put it online: How To Lose Less At Video Poker. At the risk of slashdotting my own server, I'm curious if anyone can find any obvious flaws in the method.

I found this Java-based tutorial that purports to generate the "optimum payout" -- it often disagrees with me, presumably because it's trying for big payouts. My method doesn't promise profit, only smaller losses.

An important disclaimer: I've never used my method with any non-trivial amount of actual cash. Here in Texas, there are video poker machines in every Quickie Mart, but I just don't see the appeal. Now, if they would put in a Pac-Man machine...