The fact that the poker world might change drastically has been there, where perhaps not everyone suspected it. Noam Brown and Tumas Sandholm published an article in early July in the journal Science . Their robot, they wrote, has won against the best poker players in the world. Pluribus is the artificial intelligence (AI), a bot, and unlike many of its conspecifics, it only needs a small amount of RAM: 128 gigabytes. The two researchers wrote of a "milestone".
Especially the poker scene caused Pluribus to get excited. Because the bot had won against 13 of the best players in the world permanently, including the Swiss Linus Löliger, perhaps the best online players in the world. 10,000 hands were played in twelve days. In a second experiment, in which a professional competed against five bots, the AI was also better. In the end, she had 480 times her mission again. "If it had been real money, Pluribus would have earned about $ 1,000 an hour," says one of its creators.
AI and the sport
Poker is the next sport in which an AI has made significant achievements and defeated humans. In chess, it succeeded with Deep Blue. Even the best Go player was sensational and defeated much sooner than expected by a Google computer. Well, poker, a game many still think is based on a lucky coincidence or the right leg getting up. But the success of Pluribus had been suggested, the two researchers at Carnegie Mellon University had already programmed other successful poker bots. And they promise research (and the poker scene) a lot of benefit from their discovery.
So far, the successes of AI in poker limited to the heads-up, the duel of two players. New to Pluribus is that the AI holds its own against several players. The biggest hurdle was: unlike chess, poker does not know all the information, the opponents' hand, and the new cards added in the course of a round in the most popular variant, Texas Hold'em. This makes it impossible for an AI to calculate all variants.
The researchers did not try this with pluribus. Instead, they managed to teach the AI a supposedly human behavior: the bot is bluffing. "He's a monster bluffer," said Jason Les, one of the poker pros who faced him. Kristian Kersting, KI researcher at the TU Darmstadt, says: "The most exciting thing is that Pluribus shows: human traits, such as a bluff, which we describe as intuition, can be mapped algorithmically, and we learn more about ourselves than about the AI. "
The researchers let their AI play offline against themselves for eight days. She obviously learned that it was worth bluffing. At first, the computer behaved like a beginner, but he was accumulating more and more knowledge in the trial-and-error process. That is how his data treasure was created, which continues to grow with every game. "Blueprint" the researchers call Pluribus' strategy framework, which came with astonishingly little computational effort.
Pluribus then repeatedly resorts to this "blueprint" and looks for solutions. And thus became an inscrutable opponent. "It was pretty hard because you did not recognize a pattern," says Darren Elias, who competed against five bots at the same time. Sometimes Pluribus played the same hand, sometimes like that, and used much more money to bluff than people would. "The strategy is very different from humans because Pluribus never learned how to play human," says inventor Brown.