Researchers at Carnegie Mellon University in Pittsburgh have developed software that outperformed five professionals at the same table in the Texas Hold'em poker game. The findings are published in the scientific journal Science .

In the Texas Hold 'em poker variant, players try to get the best combination of five cards from two cards in the hand and five cards on the table. Players can enter chips and apply bluff tactics to win rounds.

The Carnegie Mellon researchers developed a software program, Pluribus, that learned poker by playing the game against itself. In this way the software starts as a layman, but gradually 'learns' to know the game and apply tactics to get better.

When Pluribus was subsequently confronted with human poker players, the software adapted to the situation by looking in real time for a suitable strategy for the situation in which the computer was located.

The researchers also programmed Pluribus in such a way that the computer can take into account that an opponent can adjust his strategy during the poker game. In addition, the software was programmed to be unpredictable to some extent for opponents.

Software played against five people at the table

Pluribus played Texas Hold 'em at a table with five professional poker players who have won more than 1 million dollars (around 888,000 euros) in their careers.

A total of thirteen professional poker players participated in the experiment, in which ten thousand hands were played. In that setting, Pluribus managed to win an average of 48 big blinds per thousand hands after a correction that reduced the happiness factor.

"This is considered to be a very high win rate in Texas Hold 'em with six players, certainly against a group of professionals," the researchers write. "It implies that Pluribus is stronger than human opponents."

Poker is a relatively difficult skill to learn on a computer through artificial intelligence, because part of the information is hidden. This is in contrast to, for example, the board games chess or go, where the board and the pieces are visible to all participants.