Researchers use a GeForce GTX 1080 for poker supremacy

Games have been used as a measure of AI capabilities for quite some time. Despite the idea starting off in the eighteenth century with fraudulent chess-playing machines, computer algorithm approaches started to pop up in the 1940s. Twenty years ago, IBM supercomputer Big Blue made headlines for beating reigning world champion Garry Kasparov in a set of matches.

Texas Hold 'Em poker represents a different kind of challenge than chess, since each player has access to only a limited part of the game's total state. Despite this challenge, researchers published work in 2015 claiming that an algorithm could defeat human players one-on-one in the limited-bet form of Texas Hold 'Em. Recently, a group of scholars announced DeepStack, an algorithm that can win one-on-one no-limit Texas Hold'em whilst harnessing the power of an Nvidia GeForce GTX 1080 graphics card.

The researchers used a number of heuristics to reduce the computational complexity of the game to a manageable level. The approach reviews each round as an independent game, discarding knowledge of previous moves. The computer also restricts itself to a limited number of different bets in every round in order to reduce possible future outcomes by fourteen orders of magnitude. The Nvidia GeForce GTX 1080 graphics card requires about five seconds to evaluate the reduced level of information revealed in one round of a game.

The researchers tested the approach against 33 real players in a tournament. Eleven of those played a full 3,000 games against Deep Stack, and the AI beat 10 of tem by a statistically-significant margin. Arstechnica has a more detailed overview of the algorithm here, for those interested in reading more in a friendlier format than the original Science article.

Comments closed
    • AnotherReader
    • 3 years ago

    ArsTechnica notes [quote<] Both were designed to face single players one-on-one. Adding a full table of players would up the complexity and re-up the computational challenge [/quote<] So not quite poker supremacy yet.

    • blastdoor
    • 3 years ago

    I was struck by this passage in the write-up at Ars:

    [quote<]To avoid getting stuck in an infinite recursion, DeepStack simply forgets the past. "Our goal is to avoid ever maintaining a strategy for the entire game," its developers write. Instead, each time DeepStack needs to act, it performs a quick search to pick out a strategy based on the current state of the game.[/quote<] So this AI forgets the past, ignores the future, and is entirely focused on winning in the moment. I'm assuming the entire research team is made up of Americans?

      • ImSpartacus
      • 3 years ago

      Of course, that’s why this ai wins.

        • blastdoor
        • 3 years ago

        Yup — at poker.

    • VincentHanna
    • 3 years ago

    Am I the only person who wants to see a GTX 1080 attempt to run this program, going full tilt on a version of texas hold’em?

    You bought a [s<]$700[/s<] $500 GPU? What do you do with it? I run a texas hold'em game that I bought. Oh yeah? I've got a little [url=https://images-na.ssl-images-amazon.com/images/I/71cLo-pkJLL._SY355_.jpg<]10 in one poker game[/url<] that runs on a calculator chip.

      • chuckula
      • 3 years ago

      Yes.

      The rest of us want to see Vega do it.

      After all, there’s a reason that Vega is only one letter short of Vegas.

        • derFunkenstein
        • 3 years ago

        There’s only one Vegas, but there are multiple Vegas? Oh, wait. I messed that up…

        • ImSpartacus
        • 3 years ago

        Who needs Vegas when you can buy $AMD.

        SuBae will lead us to the gains!

      • psuedonymous
      • 3 years ago

      All GPUs should now be reviewed by having them play poker against each other.

    • chuckula
    • 3 years ago

    [quote<]The approach reviews each round as an independent game, discarding knowledge of previous moves. [/quote<] If that's true then it's interesting because the algorithm is not trying to "learn" the tendencies of a particular human opponent based on a history of previous hands. [quote<]The computer also restricts itself to a limited number of different bets in every round in order to reduce possible future outcomes by fourteen orders of magnitude.[/quote<] I have a feeling that the management of bets & how far to go before folding or the river is a big part of what makes this AI tick. It doesn't take a crazy good AI to guess your odds at any given hand of poker just by looking at your own cards and the other available cards, with the obvious uncertainties that come with incomplete information. However, the real strategy comes in to managing how much money & risk you are willing to take with each hand so that you can eventually come out on top in the long run.

      • psuedonymous
      • 3 years ago

      Could be interesting to actually play some real-world games using this AI as an assistant, particularly on sites that use webcam video of players. A lot of poker strategy is based on ‘reading’ your opponent: facing what appears to be a human player who responds like an AI would likely be baffling as hell to an experienced player.

        • VincentHanna
        • 3 years ago

        Nah, the experienced player would recognize the guy for what he is….

        A cheater.

          • Beahmont
          • 3 years ago

          Actually, they would recognize him as a mark. Because if you read up on this AI, it only one in 1v1 situations. Right now, if you were to take this thing into an actually ‘real world’ game, it would get creamed because it can’t deal with multiple opponents.

      • DPete27
      • 3 years ago

      [quote<]It doesn't take a crazy good AI to guess your odds at any given hand of poker just by looking at your own cards and the other available cards[/quote<] They display those odds in real-time on many televised WPT events I've seen. So yes, the odds alone are certainly not the reason for the complexity of the AI.

        • UberGerbil
        • 3 years ago

        But from what I’ve seen, those broadcasts also have complete information — via in-table cameras showing the hole cards of all the players — so calculating the real odds is trivial. The situation for the actual players — and for this algorithm — is altogether different when they have incomplete information, knowing only their own hole cards. In that case you have to calculate a range of possible outcomes based on all the cards that [i<]might[/i<] be in the other players' hands.

      • VincentHanna
      • 3 years ago

      [quote<]If that's true then it's interesting because the algorithm is not trying to "learn" the tendencies of a particular human opponent based on a history of previous hands.[/quote<] Kindof a misconception floating around, started, I think during the alpha-go coverage last year, but most of these learning machines, as in I'm not aware of one that does, don't learn on the fly. They study hundreds of thousands of games/moves, as well as playing millions of actual games against itself, to develop an algorithm that works, and then the program is frozen, and a version of the working algorithm is published.

Pin It on Pinterest

Share This