New Bot Plays Perfect Poker: Researchers Its strategy involves the ability to 'regret' past moves By Matt Cantor, Newser User Posted Jan 11, 2015 2:10 PM CST 17 comments Comments A dealer shuffles a deck of cards during a round of Texas Hold 'em at the World Series of Poker, Friday, May 31, 2013, in Las Vegas. (AP Photo/Julie Jacobson) (Newser) – The world's greatest poker players have a formidable new foe. Scientists have developed a computer program they say plays an effectively perfect game of Fixed-limit Heads-up Texas Hold 'em, the BBC reports. The Cepheus system "can't be beaten with statistical significance within a lifetime of human poker playing," developers say. The game won't necessarily win every hand, the Guardian reports, but it should beat humans in the long run. It developed its abilities after "playing 24 trillion hands of poker every second for two months," says a researcher. The program reportedly learns from its mistakes, experiencing what might be called electronic "regret" over moves that don't work out perfectly. If you're skeptical, you can try playing against it right here. Its favored version of poker starts with players receiving two cards the other player can't see, and that's key to the significance of Cepheus. Computers have previously developed winning strategies for games like chess, but in those games, each player has all the information about what's happened so far—they're called "perfect information games." Prior to Cepheus, "no nontrivial imperfect-information game played competitively by humans" had been "solved," researchers write in Science. At the Guardian, poker writer Christopher Hall has some doubts: "One of the bot’s limitations appears to be that it did not seem to adapt against my change of style, something that could be its undoing. The most important thing in any heads-up battle is finding out your opponent’s flaws ... and relentlessly exploiting them until they change their style to compensate," he writes. He came out slightly ahead of Cepheus after 400 hands, though that's too small a sample size to draw a clear conclusion, he notes.