Meet Libratus, the CMU bot trying to beat four poker pros at Pittsburgh’s Rivers Casino

Carnegie Mellon’s AI wants to be the best at Heads-Up No-Limit Texas Hold’em.

Poker pro Jason Les takes on Libratus at Heads-Up No Limit Texas Hold 'Em.

Poker pro Jason Les takes on Libratus at Heads-Up No Limit Texas Hold 'Em.

MJ slaby / the incline
MJ Slaby

On Wednesday morning, five poker players met for a 20-day, $200,000 tournament at Rivers Casino.

Four of them, Jimmy Chou, Dong Kim, Jason Les and Daniel McAulay, are considered some of the best in the world at Heads-Up No-Limit Texas Hold’em.

The fifth player, Libratus, wasn’t visible away from the computers where the game would be played. That’s because Libratus is Artificial Intelligence developed at Carnegie Mellon University and runs through a supercomputer and computers at Rivers Casino. 

It’s never publicly played, so the four human players couldn’t watch any videos to size up their competition, like they do for human players.

“This is like a player who’s been training in Antarctica and has come to the U.S. to play,” said Les, a poker pro of 10 years from Costa Mesa, Cailf.

How and why you can see it play

Heads-Up No-Limit Texas Hold’em is really the last frontier when it comes to developing AI to beat humans at games, said Tuomas Sandholm, a professor of computer science at CMU. Sandholm developed Libratus with Ph.D. student Noam Brown.

The game has eluded the AI community since the early 1990s, Sandholm said. AI has already beat humans at chess, checkers and other games, but this version of poker is different because of the decision-making skills needed, he said.

Les added that the game is typically played one-on-one on a computer, players have to be really involved in every hand, and the tough situations increase as the game goes on.

Anyone 21 and older can come watch the contest, dubbed “Brains vs. Artificial Intelligence: Upping the Ante” in the Rivers Casino poker room. It started Wednesday and runs 20 days through Jan. 30. Play goes from 11 a.m. to 7 p.m. daily.

The time commitment is unusual and will be the most grueling part, Les said about the contest.

They’ll play 120,000 hands total, which is 50 percent more than pros played against Claudico, an earlier AI. That’s to improve the statistical significance of the winner, according to organizers.

To make sure luck doesn’t shine on the winner, two pros will play Libratus while the other two pros are using the same hands in a separate room so their outcomes can be compared.

Although the game is played on computer screens, spectators can see the game on screens above the pros’ heads or you can watch here. (Maybe you’ll pick up on some game tips, too.)

The four humans are playing for shares of a $200,000 prize, according to CMU.

Who — or what — is Libratus?

The poker community often calls Libratus — and its predecessors, like Claudico — bots, but it’s Artificial Intelligence or AI, Sandholm said.

Tartanian8 came before Libratus, and a smaller version of Tartanian8, named Baby Tartanian8, won both categories of Heads-Up No-Limit Texas Hold’em at the 2016 annual Computer Poker Competition, which is just between AI programs, humans not welcome. Even before that was Claudico, which played poker pros at Rivers Casino in 2015 (and lost).

Libratus — which means balanced and powerful in Latin — is totally different AI from those that came before it, Sandholm said. It has improved algorithms and end-game solver, as well as a more accurate model of the game.

That model helps Libratus strategize.

Heads-Up No-Limit Texas Hold’em has 10160  — so that’s a 1, followed by 160 zeroes — possible scenarios for a player. That’s more scenarios than particles that exist, and too many for a computer to handle, Sandholm said. So, he said, the model helps because it is a smaller version of those scenario sets.

Plus, Libratus uses Bridges, a supercomputer at the Pittsburgh Supercomputing Center, a joint center between CMU and the University of Pittsburgh, during play and at the end of the day. (Claudico also used a supercomputer at the end of the day, but not during play.) Libratus has more computing power during the game and later computes to improve strategy.

But all of this won’t matter if Libratus doesn’t realize one thing that the human players already know.

“It can’t win unless it bluffs and accepts that the other side wants to deceive it, too,” said Andrew Moore, dean of the CMU School of Computer Science.

What is Libratus important?

Whether or not Libratus wins isn’t just about poker. If the AI proves it can adjust for incomplete information and an opponent that is providing misinformation, that could lead to other applications in areas like business, military cybersecurity and medicine, Sandholm said.

Moore agreed. He said one example is a smartphone app where a user could ask it to purchase a car by saying the limit price is $5,000, but sellers should be told that the limit is closer to $3,000. The AI would have to be able to adjust and bargain, Moore said.

Ken Chiacchia, senior science writer for the Pittsburgh Supercomputing Center, also added AI could help make more effective decisions with any complex real-world case where humans have to guess, such as the stock markets or in the case of looking for bombs and weapons.

But first, Libratus has to win at poker.