Pluribus, Facebook’s AI beats 6 poker champions

Developed by Facebook and Carnegie Mellon University, Pluribus is the first artificial intelligence capable of beating the best humans in the six-player No-Limit Hold’em, the most-played poker format in the world. This is the first time that the AI ​​has dominated the best players in a game with more than two players or two teams.

The artificial intelligence has undergone major changes in recent years, thanks to deep learning, which allows the AI to learn alone, without any prior information. Games are a particularly interesting training exercise, offering complex and varied situations. Since DeepBlue’s chess victory in 1996, artificial intelligence has been breaking new records, such as  AlphaStar, which has mastered the StarCraft game.

Researchers at Facebook and Carnegie-Mellon University have managed to create artificial intelligence capable of beating five professional poker players at the same time in Texas Hold ’em, in a six-player, unlimited version.

Tuomas Sandholm, a researcher at Carnegie-Mellon University, has been working on artificial intelligence and poker for 16 years. He had already created in 2015 an AI called Claudico; it was able to compete in one-on-one, without getting a clear victory, unlike Cepheus, another AI developed by other researchers the same year. An update of Claudico in 2017, this time called Libratus, had been much more successful.

A world-first in a six-player game

This previous artificial intelligence could only compete in one against one, and a six-player game is a much more complex challenge. Unlike games like chess or the game of Go where all the pieces are visible, the poker has many unknowns with the cards in hand of other players, but also misleading information when his opponents bluff.

The team created Artificial Intelligence Pluribus based on previous work with Claudico and Libratus, with some new innovations, such as the ability to evaluate only a few upcoming actions rather than analyze all the possibilities until the end of the year. part. This makes it easier to change strategy against unpredictable players and the inability to know all the cards. To reduce the complexity of the game, they also used a process called abstraction, where some actions are ignored, while others are grouped together and considered identical.

A very inexpensive AI to train

The researchers faced their AI with five professional players, randomly selected from 12 of the world’s best players, each having already won more than a million dollars. They played a total of 10,000 games over 12 days. Three of the human players also played 5,000 games each against five independent copies of Pluribus. The IA has not only given an idea of the optimal player, being able for example to use truly random strategies but also appealed to unknown strategies professionals.

Thanks to the efficiency of the algorithms used, the training of the AI ​​has been particularly economical in computing power. Researchers estimate a cost of around $ 150, compared to the millions usually spent on training other artificial intelligence. Such an AI, which requires little training, could revolutionize areas such as drug discovery or cybersecurity.