High Stakes

Carnegie Mellon University's Tuomas Sandholm, professor of computer science, knows when to hold 'em, when to fold 'em and when to let an artificial intelligence do the heavy lifting.

Tartanian7, the poker-playing AI he designed with Ph.D. candidates Noam Brown and Sam Ganzfried took top honors at the Annual Computer Poker Competition this summer in Quebec.

The Association for the Advancement of Artificial Intelligence (AAAI) launched the Annual Computer Poker Competition as a challenge competition in 2006. The quest to solve poker brings together researchers working to solve games of imperfect information to collaborate and advance the work of the field.

"Developing a strong poker agent involves a fascinating mix of theoretical advances and system building," Ganzfried said. "Despite the fact that the game is so simple to describe and pretty much everyone knows the rules and has played it, the best computer programs are still worse than the best humans in many popular variants."

Those variants include no-limit Texas Hold 'em with more than two players and pot-limit Omaha with any number of players, but the verdict is still out on no-limit Texas Hold' em with two players said Ganzfried.

In the competition that occurs prior to AAAI's annual conference, teams submit programs to event organizers who simulate millions of hands of poker, with computer facing off against computer in matches that take place on servers, not felt.

After a lengthy session at the virtual "table," organizers announce the winners. Many teams publicize the algorithms that generated their strategies.

"The poker tournament itself has really driven progress in solving incomplete-information games in general because you can benchmark against others, compare your results, and improve over time," Sandholm said.

Teams can compete in six different variants of poker, but for the past three years many of the competition's stalwarts, some of whom have been working on poker AIs for two decades, have chosen to slug it out in the Two-Player No-Limit Texas Hold 'em instant run off and total bankroll competitions.

Tartanian7 took first place in both scoring formats for Two-Player No-Limit Texas Hold 'em. The AI was the most successful CMU entry in the competition's history.

AI's such as Tartanian7 aren't programmed with poker strategy. They learn about the game from the ground up. The strategy the AI learns is certainly effective but it's also unlike anything you would see from a human, professional No-Limit Hold 'em player.

"I think it's easier for humans to learn from each other and they've learned to play in a way which may be sub-optimal," Sandholm said. "When our program learns a game, it has never seen a human play, so it is determining its strategy from first principles."

Tartanian7 can clean out middling or worse players in comparatively few hands. Against elite competition, like the other AIs on the podium at AAAI-14, reaching a decisive result required playing tens of thousands of hands. In the end Tartanian7 defeated all opponents by a statistically significant margin.

The AI also offers a new tool for better understanding and exploring imperfect information in the real world.

"Our research isn't on poker specifically; it's on imperfect-information games in general, any strategic situation where one person knows something that another person doesn't," Brown said, using business negotiations as an example. "Most real-world situations fall into this category. The ideal 'Deep Blue' scenario, where every part of a problem is laid out nicely like pieces on a chessboard, rarely comes up. I bet 20 years from now this research will be applied to real-world scenarios that we'd never think of today."

Homepage Story Archives