Pluribus becomes first known bot to beat six-max NLH poker
Online Poker Report

Battle Of The Bots: Facebook’s Pluribus Poker AI Makes Giant Technological Leap

Facebook Pluribus poker AI

The World Series of Poker is just days away from crowning its 50th Main Event champion, but that may not be the most consequential news in the industry this week. That distinction likely goes to the unlikely duo of Facebook and Carnegie Mellon University.

The two have collaborated on a new artificial-intelligence poker program that appears to be able to consistently beat some of the best human players in the world. What makes this so noteworthy is that the computer isn’t just winning against a single heads-up opponent like CMU’s previous bot, Libratus.

The new code, named Pluribus, can play — and beat — six-handed no-limit hold ’em (NLH).

The software defeated professional poker players in two formats, one with five AIs and one human opponent and the other with one AI and five humans. Multiplayer poker games are essentially the final challenge facing AI poker researchers, and Facebook says Pluribus has pulled it off.

From the research:

“If each chip was worth a dollar, Pluribus would have won an average of about $5 per hand and would have made about $1,000/hour playing against five human players. These results are considered a decisive margin of victory by poker professionals.”

If the bot’s skills are as-advertised, online poker may never be the same again.

Pluribus beats six-max NLH poker

Facebook and CMU put Pluribus through two separate trials. One saw it play against five professionals selected at random from the following lineup:

  • Chris Ferguson
  • Greg Merson
  • Anthony Gregg
  • Darren Elias
  • Nick Petrangelo
  • Jason Les
  • Dong Kim
  • Daniel McAulay
  • Michael Gagliano
  • Jimmy Chou
  • Seth Davies
  • Linus Loeliger
  • Sean Ruane
  • Trevor Savage
  • Jake Toole

The second trial pitted one pro against five (non-colluding) versions of Pluribus. Researchers incentivized humans to play their best via shares of a $50,000 prize based on performance.

The bot performed well in both trials.

In matches against five humans, Pluribus posted an average win rate of five big blinds per 100 hands. Results varied in the matches with five AIs depending on the lone pro opponent. Human players lost 2.3 bb/100 in aggregate, though, with individual results ranging from -0.5 to -4 big blinds per hundred.

Pluribus bb/100

Opponents acknowledge bot’s chops

Pluribus’ opponents included a handful of poker pros widely considered to be elite, even world-class players.

Jason Les, for instance, is a two-time WSOP runner-up who participated in those matches against Libratus — as well as Claudico, an even earlier AI from CMU. Les was duly impressed with version three.

“I probably have more experience battling against best-in-class poker AI systems than any other poker professional in the world. I know all the spots to look for weaknesses, all the tricks to try to take advantage of a computer’s shortcomings. In this competition, the AI played a sound, game-theory optimal strategy that you really only see from top human professionals and, despite my best efforts, I was not successful in finding a way to exploit it. I would not want to play in a game of poker where this AI poker bot was at the table.”

Yikes.

Chris Ferguson specifically mentioned that Pluribus was adept at finding thin value on the river, while Sean Ruane noted the “relentless consistency” with which it plays. Daniel McAulay, who also played against Libratus, highlighted the speed and theoretical soundness of the bot’s decisions. And Darren Elias said it was satisfying to see a piece of software mirroring the play of pros at the highest level, but with superhuman perfection.

“To have your strategies more or less confirmed as correct by a supercomputer is a good feeling,” Elias said.

Pluribus isn’t just a poker bot for the 1%

The most remarkable aspect of the research wasn’t even the results. It’s the modest way in which they were achieved.

Pluribus is no supercomputer.

According to Facebook and CMU, it took just eight days, 512 GB of RAM, and $150 to program the machine. Previous high-level AI bots required millions of dollars of development and substantially more computing power.

Here are the specs from the researcher:

“Pluribus runs on two CPUs. For comparison, AlphaGo used 1,920 CPUs and 280 GPUs for real-time search in its 2016 matches against top Go professional Lee Sedol. Pluribus also uses less than 128 GB of memory. The amount of time Pluribus takes to search on a single subgame varies between one second and 33 seconds depending on the particular situation. On average, Pluribus plays twice as fast as typical human pros: 20 seconds per hand when playing against copies of itself in six-player poker.”

Pluribus machine learning graph

Unlike Libratus or the bots developed by the University of Alberta, the program powering Pluribus is both technologically and financially accessible. That should send a shiver up the spines of online poker players, and even those in the brick-and-mortar realm to some extent.

Read the Pluribus report

Facebook AI Research Scientist Noam Brown summarized some of his research in this report:

Facebook Pluribus poker AI research

You can read the full research paper from Brown and Tuomas Sandholm in Science Magazine.

Steve Ruddock
- Steve covers nearly every angle of online poker in his job as a full-time freelance poker writer. His primary focus for OPR is the developing legal and legislative picture for regulated US online poker and gambling.
Privacy Policy