Computers are top-notch gamers. Since the 1970s, artificial intelligence has shamed our greatest chess players, outgunned our top trivia junkies and dominated at Connect Four. But all of these upsets had one thing in common: They involved “perfect information games.” That is, each player already knew everything that had happened in the game before they made any decisions. In poker, that’s not possible.
But scientists presented the first algorithm ever to solve an imperfect information game on Thursday, using the old standby of closed-handed gameplay: Texas Hold’em. In a study published in the journal Science, researchers showed that their computer program could “solve” Texas Hold’em poker with a strategy of bets and bluffs so optimal that it virtually never loses.
Given the sheer amount of money that flows through the gambling industry, it is conceivable that a program like this one could be quite disruptive on the casino floor. For now, however, there are no plans for commercialization.
Beyond gameplay, the program could eventually be used to develop optimal strategies for medicine and security, specifically for situations in which decision-makers must act quickly based on limited information. “Even after playing a lifetime of poker, there’s still a high probability that you would lose, even if you found an edge against the program,” says Michael Bowling, a computer scientist at the University of Alberta and lead author of the paper. “But the techniques that we used to solve the game apply even more broadly than entertainment activities. I’m talking about any decision-making scenario. Politics becomes a game. Auctions become a game. Security becomes a game.”
Training Cepheus, AI Poker Champion
Forget everything you ever knew about losing to a computer. In checkers or chess, each player has all of the information that he or she needs to make an informed decision—which makes it easy for a computer, a veritable database of the best possible moves, to sweep every round. But when it comes to Texas Hold’em, “That whole process of search doesn’t work,” Bowling says. “It’s hard to reason through how the other player might respond without knowing what cards they’re going to hold.”
“Solving” poker has long been the goal of many artificial intelligence researchers. But ambition often turned to frustration. In the late 1990s, Koller and Pfeffer of the University of California Berkeley declared, “We are nowhere close to being able to solve huge games such as full-scale poker, and it is unlikely that we will ever be able to do so.”