This Poker-Playing Supercomputer Is Getting Crushed By Humans

​A competition pitting a computer against some of the best human poker players in the world is underway in Pittsburgh, and the computer is getting slammed.

The tournament, hosted at Pittsburgh's Rivers Casino, is currently on its third day. It will continue until May 8th, when the AI, named Claudico, has played 80,000 hands of heads-up no-limit Texas Hold'em against its competition: four top players, including Doug Polk, who's made millions on the professional circuit. So far, 16,000 hands have been played, and Claudico is down $166,566. In fact, Claudico has been in the red since day one of the tournament. Although the AI made some gains on day two, they were obliterated by the third.

But Claudico's fortunes could still change with more than 60,000 hands to go. If Claudico does win, it would be an absolutely massive achievement for artificial intelligence research. Heads-up no-limit Texas Hold'em is one of the hardest games for a computer to win because it involves mind-boggling amounts of possible decisions, not all of them immediately logical, to compute.

"Poker is now a benchmark for artificial intelligence research, just as chess once was," Tuomas Sandholm, a computer science professor at Carnegie Mellon University who led Claudico's development, in a statement. "It's a game of exceeding complexity that requires a machine to make decisions based on incomplete and often misleading information, thanks to bluffing, slow play and other decoys. And to win, the machine has to out-smart its human opponents."

A big win would indicate not just proficiency at poker, but the ability to compute huge amounts of possible outcomes and make the right call—a boon to decision-making algorithms of the kind that could eventually have many applications, from day planning, to lawyering, to medical monitoring.

"You could use the same basic framework to do robust decision making like trying to come up with insulin and glucose monitoring plans [for diabetes patients]," said Neil Burch, a computer scientist at the University of Alberta who helped design a poker-playing AI earlier this year. "You get regular snapshots of glucose levels, and you have to decide how much insulin you should take, and how often."

No-limit Texas Hold'em doesn't cap the amounts that players can bet, making it a game of deceit since players can manipulate the perceived odds by betting a ton of chips and bluffing. Limit Texas Hold'em, by comparison, is a carefully structured game with set bet limits. The only variants are the cards on the table. Limit was "solved" by the AI designed by Burch and his colleagues, called Cepheus, meaning that the AI will always win eventually.

But while two player limit Texas Hold'em has roughly 1013 possible choices in play at any time, the game currently being played by Claudico has 10161, according to a CMU statement. That's 161 zeroes compared to 13.

"I would be astonished if [no-limit] was solved," said Neil Burch, one of Cepheus' designers. "But it would be a nice accomplishment beating humans. Actually solving the game of no-limit—I certainly don't think that's possible in the near future."

The Claudico team is attempting to overcome this complexity with two tricks. First, Claudico uses a variant of the algorithm developed by Burch and his colleagues that can account for gargantuan amounts of complexity by representing the number of possible decisions in a simpler, abstracted form. Even so, the computations were so complex that Claudico is running on a supercomputer in Pittsburgh called Blacklight.

Claudico also updates its strategy as it goes along, meaning it will get better and better as the rounds progress. But its basic approach to the game—called "limping"—involves getting into every hand by calling bets. For a human, limping into pots could mean losing money during an unlucky streak. But for a computer, it's merely a learning experience.

Will Claudico succeed, marking a serious step forward for decision-making AIs, or will it be taken for all its worth by human card sharks? That's a tough bet to make, even for a supercomputer.