Live Updates of the Laak vs Computer Match HERE - Page 18

threeonefour · #**171** 07-24-2007, 09:43 PM

Marv,

i am not interested in any of the numbers for in game calculations, i would just like to evaluate my play after the fact, kind of like how people plug a hand history into pokerstove just to see where they were at throughout the hand. so needing complete info for divat is largely irrelevant.

i agree all those metrics outside of DIVAT should be easy to calculate, which is why i am hoping someone creates such a calculator because if i attempted it, it would be pretty clunky. i only have one semester of C++ under my belt from a CS course i took 5 years ago, stuff like this makes me wish i tacked on a CS major.

TSchauenberg · #**172** 07-24-2007, 10:01 PM

Sure,

I was a former student/researcher at the University of Alberta. Research there started with limit ring games and evolved into various flavors of Poki. These include formula-based versions and simulation-based versions.

To isolate some variables and focus more simply on the underlying problem of dealing with imperfect information. This research switched over to heads-up limit research which delays having to tackle the multiplayer complexity until later.

Research into the limit heads-up bots has primarily fallen among two main thrusts: solving/approximating best-response equilibrium play and exploitive play.

The two most widely known examples of these different lines of research are Sparbot for the equilibrium case and Vexbot for the exploitive case. I'm not going to get into a debate of their quality as that is best for scientific papers/contests/matches to assess but as technology they are now getting to be a few years old allowing many opportunities for improvement in the bots that are participating in this year's Man vs. Machine match and in the AAAI Poker Program Championship or just independently created and kept a secret.

The main idea behind the best-response equilibrium programs is that they can put a guarantee on worst case behavior. No matter what someone does there is only so much they can lose by or in a perfect world they could guarantee to not lose. They basically build their strategy around being an impenetrable fortress.

On the other hand, the idea with the exploitive bots is that they look through their opponent's tendencies and try to construct the good counter strategies. In essence, they risk being exploited in an attempt to exploit their opponent.

Aside from building the bots to play poker, there has been another additional line of research as you mentioned (in Morgan's and Darse's theses) which focuses on attempting to get a fairer assessment in poker quicker. This line of research attempts to lessen the noise of a match's results to try and isolate who was the likely winner despite the money line results.

That's a basic summary of the largest bodies of research that has been done to date. Recently, there has been additional interest in NL as it adds more variables into the mix, and since research in the area of poker has been increasing recently, I think it is a safe assumption that research will start to heads towards multiplayer and other poker games as well.

In terms of poker research, there will be lots of cool results coming out of AAAI this year so keep your ears to the ground on that front. Aside from this greatly entertaining Man vs. Machine championship, there was also a Poker Program Championship whose results should be made shortly since the conference is currently going on and it will be interesting to see how far the state of AI has advanced since last years AAAI results.

Sorry for the long post, but I tried to give you the cliff's notes as you requested, but there was quite a bit to breakdown and summarize.

Terence.

TSchauenberg · #**173** 07-24-2007, 10:10 PM

Marv was just dissecting the numbers for you like you requested. The last three numbers are ways of assessing the strength of a poker hands. DIVAT is a way of assessing match results but is based on perfect information hands (where all hole cards are known) so you need information about all hole cards to apply it which was his point about why it can be used in the Man vs Machine match. To work with hands with not all hole cards known, it would have to be extended which Marv pointed out. It currently is also only formulated for the heads-up case since that is the current case of interest and it would also have to be extended to the multiplayer case for use there.

Terence.

threeonefour · #**174** 07-24-2007, 10:10 PM

Terence awesome post. i hope you stick around here and post in the future. same goes for any other U of A guys lurking in this forum.

TSchauenberg · #**175** 07-24-2007, 10:13 PM

Marv and eastbay have some great posts here on 2+2 so if you are interested in the subject you might enjoy going through them as well.

Terence.

TSchauenberg · #**176** 07-24-2007, 11:24 PM

Just letting everyone who is interested know that the fourth and deciding final match is going on and is about a fifth of the way through. Everything was exactly knotted up going into this match with each team having 1 win, 1 loss, and 1 tie.

disjunction · #**177** 07-25-2007, 02:27 AM

Humans win, as I thought they would. Congrats to University of Alberta team for doing some great work and making headway on a very difficult problem, even though a lot of people here seem to think it should be easy to write a good program.

Especially encouraging from the human side is that they did better on the second day of play, leading one to believe that they adapted to their unique opponent, as any poker player should when going against someone new.

Maybe this will put an end to the talk of "omg computers are so smart they can add numbers and stuff, it's easy to write a bot that can beat humans in anything all the time." At least it will stop that talk for 3 days maybe?

TSchauenberg · #**178** 07-25-2007, 02:31 AM

Yup, big congrats to Phil and Ali. Very nicely done.

threeonefour · #**179** 07-25-2007, 02:31 AM

[ QUOTE ]
Humans win, as I thought they would. Congrats to University of Alberta team for doing some great work.

Especially encouraging from the human side is that they did better on the second day of play, leading one to believe that they adapted to their unique opponent, as any poker player should when going against someone new.

Maybe this will put an end to the talk of "omg computers are so smart they can add numbers and stuff, it's easy to write a bot that can beat humans in anything all the time." At least it will stop that talk for 3 days maybe?

[/ QUOTE ]

for the record, there is no question that this bot would likely clean up on-line. i don't think a bot has to be the absolute best to be respectable. Laak and particularly Ali (based on reports of his play, i have no first hand knowledge) are almost certainly better than ~98% of serious players.

/end pro-computer spin

disjunction · #**180** 07-25-2007, 02:34 AM

[ QUOTE ]

for the record, there is no question that this bot would likely clean up on-line. i don't think a bot has to be the absolute best to be respectable.

/end pro-computer spin

[/ QUOTE ]

I think it's very very premature to say this (well, depending on what level. I think they've been able to beat novices for awhile), but I am open to the possibility and will be very interested in the hand histories.

Edit: I would also love to see this program take on the better rules-based bots online that Mr. Gatorade has called out.