Researchers have developed a computer program they say can beat any human on the planet at a particular variant of Texas Hold'em poker.
The scientists aren't planning to clean up with their powerful poker bot. Instead, they hope it can help computers become better decision-makers in the face of uncertainty. The work is published Thursday in the journal Science.
Poker players know the game is a mixture of luck, statistics and psychology. You need to know your odds and know whether your opponent is bluffing. In the words of Kenny Rogers, "You've got to know when to hold 'em, know when to fold 'em."
The variety of strategies one can use to play poker is truly immense, according to Michael Bowling, a computer scientist at the University of Alberta in Canada who invented the new program.
"There are ... 10 trillion different decision points in the game," Bowling says. "So the actual number of ways of playing is an astronomically large number."
But in the case of Heads-up Limit Texas Hold'em (a variant in which there are only two players and the betting is capped) one of these trillions of strategies can beat all the others.
It doesn't matter if the human opponent is a master bluffer, or a blundering amateur, Bowling says — a computer using that single best strategy can eventually win.
It took Bowling years to find that winning approach. His team had to crunch the strategies with supercomputers. The final push for the ultimate strategy was crunched by a computing cluster made of nearly 5,000 CPUs, according to Michael Johanson, a graduate student on the project. It took the cluster more than two months to find the answer, Johanson says. "Almost a thousand years of computation went into this."
But at the end of all that, they had the perfect strategy, the scientists say. Well, almost perfect.
"It's just a tiny bit off," Bowling says. "And that tiny bit is so small that even if you played a lifetime — 12 hours a day, 200 hands an hour for 70 years — you still wouldn't be able to tell it apart from having played a perfect game."
The strategy confirms several long-held beliefs among poker players. For example, the program will always raise the stakes if it has a reasonable hand and is the first to bet.
But the computer program also showed some surprises. It often bets even when it holds a pair of low-ranked cards like fours and threes, according to Bowling. Many professional players would fold, or at least call, with such a weak hand.
Bowling says the program isn't much of a threat to online gamblers. Heads-up Limit Hold'em is not the variety of poker most people play. But he does believe that "poker bots" are trying to win in online game rooms. "My guess is there are probably quite strong poker bots out there," he says. "But you're not going to hear a lot of talk about them."
He hopes the program he's developed can help make computers generally better at making decisions, when faced with uncertainty.
Oren Etzioni, the head of Seattle's Allen Institute for Artificial Intelligence, says the new program is breakthrough. Etzioni notes that computers now have beaten humans at chess, checkers, backgammon and even Jeopardy!
But he says humans still have an edge: Each game-winning computer involves "very, very different computer software," he says. In other words, they're highly specialized tricks for winning at particular games.
"The human," Etzioni says, "is still eons ahead of these computer programs in general intelligence."
ROBERT SIEGEL, HOST:
Today in the journal Science, researchers introduced a computer program that can outwit any human on the planet at poker. To help explain how this gambling robot beats the odds, NPR's Geoff Brumfiel goes back to last year's World Series of Poker.
(SOUNDBITE OF 2014 WORLD SERIES OF POKER)
UNIDENTIFIED ANNOUNCER: And there it is, the pocket nines for Jacobson.
GEOFF BRUMFIEL, BYLINE: The final was broadcast on ESPN. This particular type of poker is called Texas Hold'em. Don't worry if you don't know the rules, players have cards, they bet, then they show their hands. And one person takes the pot.
(SOUNDBITE OF 2014 WORLD SERIES OF POKER)
UNIDENTIFIED ANNOUNCER: Martin Jacobson has his breakthrough win. He has climbed the mountain and is poker's 2014 world champion.
BRUMFIEL: Poker isn't just a game of chance. It's a mixture of luck, statistics and psychology. You need to know the odds, know whether your opponent is bluffing. You've got to know when to hold 'em, (singing) know when to fold 'em, know when to walk away. Enough. Michael Bowling is a computer scientist at the University of Alberta in Canada who invented the new program. He's played poker in Vegas once.
MICHAEL BOWLING: I actually did phenomenally well out of pure luck and so I'm just going to never gamble again and be up on Vegas.
BRUMFIEL: Bowling loves thinking about all the strategies for playing poker.
BOWLING: Well, there are 10 to the 14, which is 10 trillion different decision points in the game. And so the actual number of ways of playing is an astronomically large number.
BRUMFIEL: And it turns out, one of these trillions of strategies can beat all the others. It doesn't matter if your opponent is a master bluffer or blundering amateur; a computer with that strategy can eventually win. It took Bowling years to find it. His team had to crunch the strategies with supercomputers. But eventually, they came up with the perfect strategy; well, almost perfect.
BOWLING: It is just a tiny bit off, and that tiny bit is so small that even if you played a lifetime -12 hours a day, 200 hands an hour for 70 years - you still wouldn't be able to tell it apart from actually having played a perfect game.
BRUMFIEL: It does have limitations. It can only play one opponent at a time. And the version of poker it wins at that isn't exactly like what's on TV. But that doesn't matter to Bowling. Remember, he's not a gambler.
He hopes the program he's developed can make computers generally better at making decisions when faced with uncertainty. Still, this left me wondering - are there any games for humans to win? I called Oren Etzioni, the head of the Allen Institute for Artificial Intelligence and asked him, maybe Connect Four?
OREN ETZIONI: The computer will never lose in Connect Four.
ETZIONI: The computer will beat you.
ETZIONI: The computer is playing at the level of the world champion.
ETZIONI: I don't know Parcheesi. It sounds kind of cheesy. (Laughter).
BRUMFIEL: I don't think anyone plays Parcheesi anymore. Maybe I can win on that ground.
BRUMFIEL: Actually, Etzioni says humans still have an edge. Chess computers, poker computers, Parcheesi computers, if they exist, they're all completely different.
ETZIONI: It's not the same program that's doing this. It's very, very different computer software.
BRUMFIEL: A single human brain, on the other hand, can play all those games even if it loses. Geoff Brumfiel, NPR News.
(SOUNDBITE OF SONG, "THE GAMBLER")
KENNY ROGERS: (Singing) You've got to know when to hold 'em, know when to fold 'em, know when to walk away and know when to run. You never count your money when you're sitting at the table. There'll be time enough for counting... Transcript provided by NPR, Copyright NPR.