CMU’s Artificial Intelligence Learns the Art of Bluff while Playing Poker

It was in January 2017 that Libratus - a computer program created by a research team in Carnegie Mellon University - proved to be able to beat the world's best poker players in a 20-day competition. Tuomas Sandholm, a Professor and Noam Brown, a Ph. D. student, both from the Computer Science Department at CMU, have created, in their own words, "superhuman artificial intelligence."

Now, twelve months later, the researchers revealed how they taught the computer program to play poker, in the no-limit Texas Hold'em variant. In an article published this week in the journal Science, Sandholm and Brown describe the challenges they faced with Libratus and show the process of creating the program. 

Poker is the perfect game to teach artificial intelligence to think furtively, and know how to omit information during strategic negotiations, the researchers say. Despite the victories of artificial intelligence in various strategy games - from chess to checkers to Go - poker has always been a more complicated area. Unlike many board games, there are hidden elements in card games. "In complete information systems, both players know the state of play at any point," reads the article. This is the case with chess, for example. In contrast, in games with incomplete information - as in the case of poker - part of the game is hidden in the cards of some players.

Read the article at Carnegie Mellon University News (December 17, 2017) and the Portuguese article at Público Online (December 18, 2017)