Washington (AFP)

After checkers, chess, go and two-man poker, an artificial intelligence beat poker champions in six-party games, researchers at Carnegie Mellon University and Facebook said Thursday.

The feat of this artificial intelligence, called Pluribus, is described in the prestigious American journal Science.

"Pluribus has achieved a superhuman performance in multiplayer poker, which is a major step for artificial intelligence and game theory," said Tuomas Sandholm, professor of computer science at Carnegie Mellon, and co-developer of Pluribus with PhD student Noah Brown, currently in Facebook's artificial intelligence unit.

"So far, the major stages of superhuman artificial intelligence in terms of strategic reasoning were limited to two games," he said in a statement from the university.

According to the two designers of Pluribus, the way is open for artificial intelligence to solve more real-world problems where, like poker, information is hidden or unavailable and players bluff.

Pluribus first beat two great poker champions, Darren Elias and Chris Ferguson. Everyone has played more than 5,000 games against Pluribus, which has won a lot.

The machine then opposed thirteen professionals - five at a time - for a total of 10,000 games. Here again, Pluribus emerged victorious.

She first trained to play Texas Hold'em poker against herself, learning the subtleties of the game through random decisions. Her inventors let her develop her tactics, with surprises at the end.

"His main strength is his ability to use different strategies," commented player Darren Elias. "Humans try to do the same, but for humans, the execution is poor, most people can not really do it in a perfectly random and consistent way."

Pluribus has resorted to surprising blows, for example a decision often considered bad by the pros and called in English "donk betting", when a player bet aggressively when it has no tactical advantage.

Noah Brown goes so far as to think that Pluribus's victory could "change the way the pros play poker".

© 2019 AFP