'Absolute monster bluffer': Poker AI smashes elite human players after 8 days of training
The list of things that AI can do better than us mere humans has a new addition: bluffing in a Texas Hold ’em poker game. And it was amazingly cheap to train and operate, unlike some previous algorithmic champion players.
Human players simply didn’t stand a chance against Pluribus, an AI system developed by researchers at Facebook Inc. and Carnegie Mellon University. Even the elite of the poker landscape, each pocketing at least a million dollars in pro competitions, couldn’t beat it.
“It is an absolute monster bluffer. I would say it’s a much more efficient bluffer than most humans. And that’s what makes it so difficult to play against,” said one of the humans, Jason Les. “You’re always in a situation with a ton of pressure that the AI is putting on you and you know it’s very likely it could be bluffing here.”
Another one bites the dust
Computers have been surpassing humans in one game after another, with checkers, chess, go and some competitive video games like Starcraft 2 already conquered. The predecessor of Pluribus called Libratus could play poker better than professional human players two years ago, but it could only handle two-player games. But in multiplayer games, especially with hidden information like in poker, finding a winning strategy is much harder for an algorithm.
Pluribus spent about eight days playing trillions of poker hands against five clones of itself, finding decisions that bring greater wins and gradually evolving its “blueprint strategy.”
In actual games it used a predictive algorithm to fine-tune its decision to cards in play. When one computer was facing against five live elite players, it was winning about 5 big blinds per 100 hands, which translates into a thousand dollars per hour if the chips were worth a dollar each. There were also games in which one human played against five AIs.
Google News, Bing News, Yahoo News, 200+ publications
“Its major strength is its ability to use mixed strategies,” Darren Elias, the holder of the most poker titles in the world, commented. “That’s the same thing that humans try to do. It’s a matter of execution for humans – to do this in a perfectly random way and to do so consistently. Most people just can’t.”
The achievement came at a surprisingly low computational cost, the researchers say. Training of Pluribus on a cloud service would cost just $150 while running it required a computer with two processors and 128 GB of memory, which was enough to play twice as fast as an average human player. Noam Brown, one of its creators, writes that this low cost gives hope that future research in AI technology is not limited to teams with multibillion-dollar funding.
The approaches used in developing the AI poker genius may find application in fraud prevention, cybersecurity, and taking action on harmful content, all of which involve multiple actors and hidden information. But knowing that a machine is better at bluffing than the best that humanity can offer may seem disconcerting. Who knows, maybe in a few election cycles AIs will be advising candidates what to promise voters on the campaign trail.
If you like this story, share it with a friend!