Poker AI

[Russ]

I am surprised at many of your responses which relate to application of computer technology, particularly as you are in the field yourself.

Considering the computing power and algorithms recently put to use against the world chess master, and drawing, the only thing stopping the same happening in poker is focus and access to the leading edge technologies that IBM have employed. And then again, who would know if someone is not already doing it. They certainly wouldn't publicise it, and they wouldn't be be showing players winning millions. It would be a gaggle of golden geese hatching several small golden eggs, a few hundred $'s at a time.

[Wake up CALL]

It is precisely because I am in the industry and it's applications to gaming that I hold my opinion. Programming Big Blue to bring chess Grand Masters to a draw was certaily quite an achievement.

If you spend enough money most anything can be accomplished, most anything other than a successful poker bot that is. [img]/images/graemlins/smile.gif[/img]

I did state that it will occur someday, just not today. To adequately argue this point a pure definition of winner needs to be accepted. I simply am using the one quoted by Mason Malmuth (1 BB per hour winner in mid limits and higher). If you wish to offer me your definition as being more liberal my response would likely change.

EDITED BELOW:

Actually now I seem to recall that Mason's definition might have been for an expert. If so I must reconsider my position.

[brianmarc]

If that's what MM says, and it's OK for you, then I suggest you both need to get out more. That's old B&M thinking. All that expertise; all that practice and the best you can do is $40 or $60/hr. There are so many online games that are much easier than that where a more reasonable goal is more like 3-5 BB/hr.

Actually, BB/hr is not a good way to measure success. $/hr is the only thing that matters.

[Dylan Wade]

I've been working on such a project on/off for about a month.

Designing an AI pre/flop and flop strategy based on EV calculations and heuristics is pretty trivial. Turn and river actions are even more straightforward.
I find the most difficult portion to design is the "tracker". The toughest part (and it's a major problem) is the difficulty in finding "maniacs", "calling stations", etc. and finding a precise adjustment to the bot's default play in these scenarios.

[Mano]

The thing is, if you could program a bot that could win even a fraction of a bet/hr, say 1/3 BB/hr you could make a fortune because you could run multiple bots playing multiple tables, and they could play 24 hrs a day, every day. Because of the greater volume of hands the bots could play, you might not mind sacrificing the small edge plays that would increase the difficulty in making the program (but would probably also increase the variance) because even if their hourly take is small compared to a typical winner they would be capable of playing so many more hands than a human that they would still make more. Of course this could be used to detect them, but you could get around that by running your computer players in shifts (even adding in random starting times and session lengths with in a reasonable range) and always having a number of them going at the same time.

[jknupp]

I'm a Computer Science student at Carnegie Mellon. I am 100% positive an AI poker player could be written to consistenly beat a LLHE game, and what's more, could be written relatively easily. Here is why:

We all seem to be forgetting the FTOP (which I'll reinterpret for the sake of this thread) At any point in a hand, there is a play that is absolutley the correct play Only knowing the other players' cards prevent us from making this play every time. In other words, a lack of information. If a bot were able to catalog the play of a large number of players in a meaningful way (which it easily could) it would be better suited than any human player to determine the correct play. Remember, most opponents in LLHE do not vary their play enough not to be easily read, if you are willing to pay enough attention. Everything else is just programming heurisitic rules, and this is the simplest form of AI.

Anyone who thinks that the processing power doesn't exist for this is out of their mind. I'm sure a program such as the one described above could easily run on a 500mhz pentium with 64m RAM. In fact, I had a highschool programming project to create an AI hearts player, and it was easy to make it able to beat most human opponents. Those were run on a system simmilar to the one described above.

Just reading/replying to this thread has almost given me the incentive to write one. Let me reiterate, it's not only possible, it's easy.

[gren]

I have to agree with jknupp. I'm still a poker newbie, just started the 1/2 tables at Party.

But, I still I believe it would be possible for a program to consistently beat the low limit games on the net. I would not be surprised if there are 'bots' that play and win.

A chess program and a poker program are very different. After playing the standard opening moves (the same ones human chess players play), chess playing programs use Min/Max algorithm find the "correct" move. This involves a brute force method of searching through a tree of possible moves to some depth.

While the nodes grow exponentially, it does not take much computational power to defeat a mediocore player. My Palm Pilot would beat me at chess pretty regularly if I set the depth high enough.

That being said, a poker bot doesn't need super computer. I don't think it would be that difficult, certainly not impossible to create a Bayesian Network to make correct plays to beat mediocore players.

Gren

[Wake up CALL]

"I'm a Computer Science student at Carnegie Mellon. "

After your first sentence above I lost interest.

[brianmarc]

Now what sentence in what post might that be?

[jdarwin]

I've come to believe that many people do not understand how easy it is to build a poker bot. I have built a poker bot and well to be quite honest, its been losing money. I tried letting it play the 3/6 and 2/4 tables with just good hands but that won't cut it... what limit will? .50/1.

For the last few days, its been playing .50/1 and the variance as been quite huge.

My bot plays on PartyPoker and it has the current abilities:
-Can tell how many people are seated at a table
-Can Autojoin tables if (numplayers <= 6)
-Plays only good cards (this REALLY needs to be worked out)

My next version will support the following:
-Has the ability to do players tracking (I can track a users play [fold,call/check,bet] on the preflop,flop,turn,and river - then analyze that play and determine the best strategy)
-Neural Network Analysis - Based on http://spaz.ca/aaron/poker/nnpoker.pdf - Neural Networks have really good prediction capabilities and this paper describes that

Honestly, I'm trying to build a winning poker bot... but so far I've come up short... I haven't played .50/1 long enough to see if it can win but only time will tell.

-jdarwin.