Darren Elias makes a living playing poker, to the tune of $7 million in tournament winnings as of 2019. He also holds the record for the most World Poker Tour titles, with four to his name.

That makes him one of the most famous players out there. But he says there’s one competitor, in particular, that he wouldn’t hedge his life’s earnings against.

Pluribus is the name — and this is one Texas hold’em champ that will quickly snatch up all of your chips. That’s because it’s an artificially intelligent computer program designed to win.

Representing a significant leap in AI tech, Pluribus has the ability to beat five human poker players, at once.

Advertise with us

It was developed by Tuomas Sandholm, a professor of computer science at Carnegie Mellon University, and Noam Brown, a research scientist at Facebook AI in New York City, who is also wrapping up his Ph.D in computer science at CMU.

Their aim is not to revolutionize poker, but to use it as a milestone for progress in AI research. “Using games as test beds gives you a way to score how strong the AI is,” said Mr. Brown.

Mr. Elias, 32 of Medford, N.J., said, as a poker expert, he helped the researchers find weaknesses in Pluribus’s performance in the system’s earliest days of training. He’d play against five versions of Pluribus online for five or six hours per day over a one-week period.

They quickly cleaned up the code when Mr. Elias made suggestions. What he saw thereafter shocked him.

“It would be a lot better the next day, which is kind of creepy,” he said. “It went from being a mediocre player to a world-class player in a matter of weeks.”

Advertise with us

Not the first, but the last

The research on poker and AI has been ongoing for some 16 years. Mr. Brown has been involved with such projects since 2012, he said.

The work on Pluribus was funded by the National Science Foundation in Washington, D.C., and the U.S. Army Research Office in Durham County, N.C.

By no means is Pluribus the first AI system to beat professional poker players.

In 2017, Mr. Bown and Mr. Sandholm teamed up on “Libratus,” another AI system that beat out quite a few players in one-on-one poker matches over a three-week span at the Rivers Casino in the North Shore. It was impressive, but it only illustrated the strengths of AI against one other component.

Advertise with us

To ratchet up the stakes, the duo took on a new challenge with Pluribus — the name means “many.” Texas hold’em is played with up to 10 at a table.

Most experts in the AI community, academic and industry, thought Mr. Brown and Mr. Sandholm could never build a system like Pluribus using their methodology. That entailed a different kind of machine learning, a method to teach an AI program how to function by feeding it information, called training data. Once a system understands relationships between those pieces of data, it can make predictions on its own.

With Pluribus, there was no outside data — meaning no information from actual Texas hold’em games that humans have played. Rather, the computer learned to master the game by playing against itself and using what it learned to create its own training data that self-informs the AI moving forward to the next game.

Over and over and over.

“It gradually improves over time ... it learns how to play poker by figuring out how to beat itself,” Mr. Brown said.

Advertise with us

Now, Pluribus can see all options on the table — literally — and come up with an idea of what makes sense at first glance. It’s just like seeing a chess board and noticing the first possible move that seems logical.

And it only takes its computer brain 10 to 15 seconds.

The prisoners’ dilemma

To come up with that matrix of choices, Pluribus relies on “game theory,” an area of study in economics that deals with the differing strategies and incentives of multiple participants in competitive situations. In other words, the outcome of one person’s choice is dependent on the action of the other person.

Game theory has been applied in the contexts of business, gaming, war and even biology.

Advertise with us

The classic example is called the “prisoners’ dilemma,” wherein two rational individuals facing criminal sentences might not cooperate, even if it’s in their best interest to do so.

Here’s the gist, as described by Albert William Tucker, a Canadian mathematician, who formalized the idea:

“Two members of a criminal gang are arrested and imprisoned. Each prisoner is in solitary confinement with no means of communicating with the other. The prosecutors lack sufficient evidence to convict the pair on the principal charge, but they have enough to convict both on a lesser charge. Simultaneously, the prosecutors offer each prisoner a bargain. Each prisoner is given the opportunity either to betray the other by testifying that the other committed the crime, or to cooperate with the other by remaining silent. The offer is:

- If A and B each betray the other, each of them serves two years in prison.

- If A betrays B but B remains silent, A will be set free and B will serve three years in prison (and vice versa)

Advertise with us

- If A and B both remain silent, both of them will serve only one year in prison (on the lesser charge).”

Click the Gif above to check out a larger version of the image.

Of course, there is luck involved in poker, so Pluribus hasn’t always won. Just most of the time. Maybe Pluribus has been dealt a bad hand, for example.

That’s at the heart of both poker and game theory, though: all sides have imperfect, incomplete information, making decisions harder. Players don’t know what cards the others have, for one — although Pluribus is pretty good at detecting humans who are bluffing.

To evaluate the millions of possibilities that could be used in actual games, it takes comparably little time to come up with its book of strategies compared to predecessor AI systems. Libratus took 15 million hours, while Pluribus only took eight days to hone its game plans.

Advertise with us

‘Beaten by a machine’

The days of poker AI research are coming to a sunset.

Mr. Brown said getting from a two-player to a six-player game has been the “final milestone” in that arena.

They didn’t just achieve that goal, either. “We really destroyed it,” he said with a laugh.

The results could inform real-world situations where those on multiple sides of the issue  have imperfect information.

Advertise with us

Mr. Brown says the strategies from Pluribus could potentially help detect credit card fraud, handle negotiations, deal with cybersecurity, and even help navigate traffic in self-driving cars.

If anything, it has already helped professional pokers reevaluate their strategies.

Mr. Elias said human poker players don’t typically bet on really huge pots of money, but Pluribus does. And wins. If a supercomputer wants to make that move, it’s something to think about, he said.

“It’s a humbling thing to be beaten by a machine,” he said. “To have something I dedicated my whole life to and to be beaten in a matter of weeks, is humbling.”

Courtney Linder: clinder@post-gazette.com or 412-263-1707. Twitter: @LinderPG.

Advertise with us