Over the years, we have seen how AI has advanced and it still is, to make things better for us than ever. Inching towards that very aim, Facebook’s AI Research team, in collaboration with Carnegie Mellon University (CMU), has developed an AI that can beat Pro Poker Players.
Winning Poker Against Human Players
As announced via a press release, the AI system named Pluribus was able to defeat 12 of the best Poker players in the popular six-player no-limit Texas Hold’em poker format.
Pluribus was put against 12 players in two settings. The first involved five human players and one AI, and the second one had five AIs and one human. The AI bot was able to win an average of $5 per hand, totaling up to $1,000 per hour — which is a breakthrough for an AI.
The match had over thousands of Poker hands in a time period of a couple of days.
The list of Poker players includes Jimmy Chou, Seth Davies, Michael Gagliano, Anthony Gregg, Dong Kim, Jason Les, Linus Loeliger, Daniel McAulay, Nick Pietrangelo, Sean Ruane, Trevor Savage, and Jake Toole.
It is suggested that Pluribus was able to win the game by applying the Nash Equilibrium theory that involves players to figure out equal probabilities in a game to win, much like it is done in the stone-paper-scissor game.
You can view the following video to have a look at the game-play:
Pluribus is the first AI capable of beating human experts in six-player no-limit Hold’em, the most widely played poker format in the world. This is the first time AI has beaten top human players in a complex game with more than two players or two teams.This video shows sample hands from Pluribus’s experiment against professional poker players. (Cards are turned face up to make it easier to see Pluribus’s strategy.)
Posted by Facebook AI on Wednesday, July 10, 2019
As a reminder, Pluribus’ achievement comes after various other experiments on Libratus AI, which was able to defeat players in a two-player Poker setup. The new breakthrough suggests that AI can attain superhuman capabilities with further challenges.
How Pluribus Was Trained?
To beat the top Poker players, Pluribus was first made to play in the self-play setting where it played with versions of itself. It helped the AI bot train itself with the use of trial and error methods.
Eventually, it upped its game by predicting the opponent’s move for some of their moves and not the entire game — by applying a real-time search to the game-play.
The AI bot was trained in just eight days with the use of a 64-core server, less than 512 GB of RAM, and no GPUs.
Another thing worth noting is that Pluribus is a way cheaper AI bot compared to the many expensive ones out there, and this just adds onto its pile of achievement.
Furthermore, Facebook and CMU hope to use the strategy showcased by Pluribus in areas such as cybersecurity, fraud prevention, online auctions or navigating traffic.
While you may think what’s great about this achievement, I would like to tell you that Poker is one of the most difficult card games as it’s all about bluffing and figuring out the opponent’s move, where at times even humans fail. Being able to follow this drill on the part of AI is something of a milestone.
Apart from Pluribus, there are other AI as well that have gained mastery over zero-sum games such as checkers, chess, Go, two-player poker, StarCraft 2, and Dota 2.