In a Pittsburgh casino, four of the world’s best poker players are taking part in a tournament with very high stakes. They’re not competing for a huge pot of money or international acclaim, but for some something much more important: the pride of the human race.
The players are playing against an artificial intelligence algorithm running on a supercomputer located 15 miles away from the Rivers Casino at the Pittsburgh Supercomputing Center.
The game — heads-up no-limits Texas Hold’em — is the last great challengefor AI researchers where humans still hold sway, but if early indications are anything to go by, poker is about to go the way of chess, checkers, Jeopardy and, most recently, Go.
“It is a huge deal,” Tuomas Sandholm, professor of computer science at Carnegie Mellon University and creator of the algorithm, told VICE News. “Performance against humans has been one of the main benchmarks for AI from the beginning. Of all of the games that have been seriously worked on in the AI community — that people have put serious effort into for decades — this is the only game where AI has not surpassed the best humans.”
Poker is fundamentally different from checkers or chess or even the incredibly complex game Go. Poker is what is known as an “imperfect information” game. In Texas Hold’em, players don’t know what cards their opponents are holding, meaning they can’t be certain of how much they should bet. There is also the highly complex process of bluffing to contend with, all of which adds up to a game requiring reasoning and intelligence — something machines have so far found difficult to manage.
There’s said to be something like 10160 decision points in the game. That number — which is described as “ten duoquinquagintillion” — is so large that only a $10 million HP supercomputer with over 11,000 processing cores can handle the number-crunching needed to take on the best players in the world.
With just one day left in the 20-day tournament, the writing’s on the wall, with the algorithm — called Libratus — holding a commanding lead of more than $1.6 million over the human players. The algorithm is also getting better with every hand it plays, thanks to something called reinforcement learning.
Libratus learned the game by playing trillions of hands against itself, figuring out its own unique strategy for winning along the way. But there are some glimmers of hope for us humans. On the 13th day of play for example, the humans made a comeback, winning the day’s play by $93,000.
Advances in artificial intelligence research have been progressing at a pace even those in the industry have been surprised by in recent years, but beating heads-up no-limit Texas Hold’em will stand as achievement above the rest because of its real world implications.
“I would say from an AI perspective, [beating Go and poker] are both great achievements, but from a practical perspective, this is much more important in that very few real world situations are perfect information games, while most real world situations are imperfect information games,” Sandholm said.
It means that the algorithm could have much greater implications for solving problems in the real world. Sandholm, who developed the algorithm with graduate student Noam Brown at Carnegie Mellon University, believes that Libratus could be applied to a wide range of industries, including cybersecurity, the military, negotiations and medical diagnoses.
“These are not algorithms for solving poker, they are algorithms for solving imperfect information games in general and poker is just a benchmark through which the different research groups around the world compare performance against each other and build up on each other’s algorithmic techniques,” Sandholm said.
Sandholm and his team at CMU are not the only group seeking to beat this game. A group of researchers from the University of Alberta, Prague’s Charles University and the Czech Technical University have created an algorithm called DeepStack which claims to be “the first computer program to beat professional poker players in heads-up no-limit Texas hold’em.”
Michael Bowling, one of the researchers who created DeepStack told VICE News that he cannot discuss the details of the algorithm at the moment, as the work is under peer review.
Sandholm says that while it may be technically correct that DeepStack is the first to beat a professional player, the professional players who took part in that test are not among the best in the world. “There is a huge difference between the best heads up, no limit Texas Hold’em players and a typical poker pro,” Sandholm said.
Beyond whether DeepStack or Libratus ends up making human decision making obsolete, there’s the the issue of how AI impacts the game of poker itself, a game played by millions around the world. But rather than destroying the game, Joe Barnard, COO of the International Federation of Poker told VICE News, he believes it could end up helping improve competition by leveling a very uneven playing field.
“The landscape at the moment is to seek out the weakest opposition and exploit them the most and make a good living,” Barnard said “There is no other sport that is really the same.”
Barnard compared it to top tennis professionals playing school kids and being ranked as a result of their victories. AI algorithms like Libratus will help change this by allowing organizations to accurately rank players on their ability rather than on who makes the most money.
“It gives us a chance to actually change the whole landscape of how poker is played so that anybody can really play, it is not about how much money they put in, it is just the most skillful players play against each other to find out who is the best,” Barnard said.
AI could also be used for training and already the top professional poker players have adapted their game as a result of seeing how AI is playing. “To me poker is an intellectual exercise and these AIs are enriching the game a lot,” Sandholm said. “The top pros that played [the AI] already have incorporated many of the moves it makes into their own games.”
In 1979 Doyle Brunson, the first man to win $1 million playing poker, published the first ever book on poker strategy. Called Super/System, the book changed the way the game of poker was played forever. In the next couple of years, artificial intelligence is going to do that all over again.