Scientists at Carnegie Mellon University created artificial intelligence and taught him to play Texas Hold'em, the most popular form of poker. Similar programs existed before, but they were usually good in one-on-one games. The new AI managed to furnish five professionals at once.
“In the past two decades, artificial intelligence systems have mastered more and more complex forms of poker. However, previous breakthroughs were limited to the participation of only two players, ”the developers said.
Pluribus studied Texas Hold'em with five copies of it. At first, the system acted at random and checked to what result a particular move leads. Subsequently, the AI began to understand what actions and with what probability lead to victory. This method of teaching artificial intelligence was previously used in the board game of Go , Dota 2 , StarCraft 2, and other games.
Pluribus skills were evaluated in two formats: five players against AI and one player against five copies of the program. The experiment involved professional players who earned more than one million dollars from poker. Among them are Jimmy Chow, Seth Davis, Michael Galliano, Anthony Gregg, Dong Kim, Jason Les, Linus Leliger, Daniel McAuley, Greg Merson, Nicholas Petrangelo, Sean Rouen, Trevor Savage and Jacob Thule.
Every day, researchers selected top five random professionals who did not know who they were playing with. Each was assigned a pseudonym, which remained unchanged throughout the experiment. This made it possible to track the progress of individual participants. For motivation, people were divided into $ 50,000. In addition, they were guaranteed at least $ 0.40 per hand with an increase of up to $ 1.60, depending on the outcome of the game.
According to the results of the 12 days of the experiment, the Pluribus program showed the best results in more than 10,000 hands, both in games against five professionals, and in the format of five copies of AI against one person.