WSJ publishes misleading article claiming a computer can beat all comers at poker – but a close look says not yet.
It repeats the story of the Deep Blue beating Kasparov in 1997 – also misleading, since Kasparov was evidently still the superior player per se, just overcome by nervous imagination.
[spoiler title=”Click for whole story” open=”0″ style=”1″] WSJ
Computer Conquers Texas Hold ‘Em, Researchers Say
Advances in Artificial Intelligence Allow Program to Play Poker Almost Perfectly, New Paper Asserts
University of Alberta Prof. Michael Bowling, seated, and Michael Johanson, with his back to camera, showing the Cepheus poker program on their laptops. ENLARGE
University of Alberta Prof. Michael Bowling, seated, and Michael Johanson, with his back to camera, showing the Cepheus poker program on their laptops. JOHN ULAN/UNIVERSITY OF ALBERTA
By ROBERT LEE HOTZ
Jan. 8, 2015 2:00 p.m. ET
Artificial intelligence experts said they have developed a computer card shark that plays poker almost perfectly, having mastered a version of a popular game called Texas Hold ’em.
While playful at heart, their advance in the computational mathematics of game theory may lead to broader innovations in military strategy, national security, medical decision-making, complex contract negotiations and auctions, experts said.
The basic poker game “has been essentially solved,” said Tuomas Sandholm, director of the Electronic Marketplaces Laboratory at Carnegie Mellon University in Pittsburgh, whose own poker-playing program won a 2014 world championship. He wasn’t involved in the project. “This is a breakthrough.”
In recent years, high-powered computer programs have outmatched top human players in chess, checkers, Scrabble and the quiz game Jeopardy, but the uncertainties of poker—where so much information about the state of play is hidden—until now had defied the efforts of dozens of research teams.
All told, an estimated 150 million people play poker world-wide, according to to the International Federation of Poker in Switzerland and the U.S.-based Poker Players Alliance, which describe Texas Hold ’em as the most popular version of the card game.
In their work announced Thursday, researchers led by Michael Bowling at Canada’s University of Alberta Computer Poker Research Group in Edmonton tackled a two-player variant of that game called “heads-up limit hold ’em” in which the size of bets and the number of rounds of play during each hand are fixed. It is the simplest version of the game that people usually play. Scientists are still working to solve more complex variations of the poker game.
“Our goal is to advance artificial intelligence, and poker, we believe, is a really good test bed for trying out new algorithms,” said Dr. Bowling. “Poker is an ideal game to capture all kinds of uncertainty.”
Already, computer experts are using such artificial-intelligence techniques to handle the real-world complexities of bidding on large commercial-trucking contracts, to arrange randomized schedules for air marshals guarding commercial airliners, and to ensure equitable distribution of human kidneys donated for transplant surgery, experts said.
MACHINES AT PLAY
Computer researchers have been using games to test theories of artificial intelligence for decades. The newest innovation is a poker-playing program called Cepheus which plays Texas Hold ‘em. You can try your hand against it online. Here is a sample of their earlier successes.
Backgammon. In 1979, a backgammon-playing program defeated world champion Luigi Villa, the first time that a computer program beat a human champion at a board game.
Chess. In 1997, a chess-playing computer developed by IBM called Deep Blue won a 6-game match against world champion Garry Kasparov.
Checkers. In 1994, a checkers-playing computer program called Chinook was declared the Man-Machine World Champion in checkers.
Scrabble. In 2007, a computer program called Quackle beat former Scrabble world champion David Boys.
Go. In 2010, a program called MoGoTW running on a supercomputer defeated a professional Go player.
Jeopardy! In 2011, an IBM computer system called Watson beat two human experts to win the quiz show’s first prize of $1 million.
Dr. Bowling and his colleagues are working now to apply their latest poker-based game theories to the treatment of diabetes.
To learn all the nuances of bluff, bet, hole cards and luck of the draw, the new poker algorithm ran on an array of 4,000 computer processors for 68 days, calculating the optimal outcomes for more than a trillion possibilities of play. The scientists published their research in Science.
The program began with only the rules of play, then worked through more than six billion hands every second, more poker than has been played by the entire human race, the researchers said.
With each hand, the system improved its play, refining its decisions closer and closer to the ideal solution.
At each decision point, the system strove for what is known as the Nash Equilibrium, a winning strategy that benefits competing players in a noncooperative game. It is named for John Nash Jr., a Princeton University mathematician whose work on chance and complex systems won him a share of the 1994 Nobel Prize in Economics.
“In every possible sequence, it will ask itself how much it regrets having just bet,” Dr. Bowling said. “There are mathematical ways of turning that regret into ways of playing in the future.”
University of Alberta Prof. Michael Bowling, showing some of the computations that went into developing the Cepheus computer poker program that he describes as “the perfect program.” ENLARGE
University of Alberta Prof. Michael Bowling, showing some of the computations that went into developing the Cepheus computer poker program that he describes as “the perfect program.” JOHN ULAN/UNIVERSITY OF ALBERTA
The end result is a program called Cepheus that not only outplays skilled human players in a fair game, but always plays an essentially perfect game, the researchers said.
“If you played against it, no matter what you do, you still wouldn’t be able to eke out any more than a tiny advantage over millions and millions of hands,” Dr. Bowling said. Among other things, the researchers discovered the dealer almost always has an edge.
The researchers said they don’t plan to commercialize their new system. In fact, they are making the computer source code public and have set up a university website where people can try their hand against the system free.
“People can try their hand against the perfect program,” said Dr. Bowling.
Write to Robert Lee Hotz at firstname.lastname@example.org
There are 15 comments.
OldestReader RecommendedTimothy HarrellTimothy Harrell Jan 9, 2015
Er.. this only relates to a very restricted (and essentially pointless) version of Poker known as heads-up fixed limit hold-em, where the other guy can call you every time regardless because you don’t have the ability to bet them out of the pot.
In this restricted form, it’s merely a case of calculating how the odds stack rather than psyching out your opponent, so what’s so amazing that a computer can work out how to play it?
Jerry ColeJerry Cole Jan 9, 2015
Humans? We don’t need any stinkin Humans – signed your friend, the computer 😀
MICHAEL EDMONDSMICHAEL EDMONDS Jan 9, 2015
Dumb! Poker is far more than the odds of the cards. Yes, if you play millions of games, this thing could probably do ok. People don’t play millions of games. Instead, they play a few dozen games and stop when the money runs out. In that environment, odds matter, but bluffing and intuition make more difference.
Anybody who has ever played against a really good player would laugh at this.
Elliott Widaski Jan 9, 2015
What was the ‘rake?’ Wouldn’t it still lose? At $5 a game, a trillion games, that’s a loss of $5 Trillion. Dumb machine. But it thinks it won.
Travis BrunsonTravis Brunson Jan 9, 2015
Agree with most commenters. Not only is this a WAAAY over-simplified version of a game that features infinite random factors when played at high levels, its not even replicating most of the factors involved when playing Hold ’em for “real” at almost any level.
What I want to see programmed in is that since the computer “regrets” (itself an asinine concept for something that has no bills to pay, morals or soul) having made a losing decision, when it “loses” a big hand, the program goes on “tilt” until smoke starts pouring out of the servers. When a computer program “loses” it doesn’t “really” worry about dealing with stress, missing the rent, losing a car payment, having a run of bad luck. Can it really “adjust” for sitting across from a random madman?
“Great” poker (forget perfect, I’d define the goal as “profitable”) is more than just statistics and math.
sam bealsam beal Jan 9, 2015
Isn’t “Texas Hold ’em” no limit, which is not what this AI can “play & win”.
Victor CookVictor Cook Jan 9, 2015
Brute forcing a potential solution to a finite series is NOT something to break open the Champagne for.
Madelene TepersonMadelene Teperson Jan 9, 2015
Scientists are still working to solve more complex variations of the poker game…
Not so solved after all…
Ken JackmanKen Jackman Jan 9, 2015
I’ll bet Captain Kirk could beat it ala the cobiashi maru (sp).
Tony GibsonTony Gibson Jan 8, 2015
Keep that computer away from the WOPR at all costs.
Reg NelsonReg Nelson Jan 8, 2015
” . . .a two-player variant of that game called “heads-up limit hold ’em” in which the size of bets and the number of rounds of play during each hand are fixed.”
That’s a subset of a subset of a subset of a subset of the game. Only using endgame (two players) with fixed bets and a fixed number of bets is hardly mastering Hold ’em. Totally misleading article.
JOHN VIGUERIEJOHN VIGUERIE Jan 8, 2015
@Reg Nelson yeah – it mastered the ‘tic-tac-toe’ version of poker… great.
Thomas CooperThomas Cooper Jan 8, 2015
Didn’t G2 Game Design come out with an essentially unbeatable poker machine a couple of years ago?
Madelene TepersonMadelene Teperson Jan 9, 2015
@Thomas Cooper You can’t bluff a computer, so a computer playing pure mathematical odds would eventually win all the money. Provided, of course, human players would be dumb enough to keep throwing money in a pot a computer was avidly betting.
Sy CorensonSy Corenson Jan 9, 2015
@Thomas Cooper Pit these two against each other and see what happens. That should settle it.