Robot knows when to hold ‘em, wins huge in poker tournament

Artificial intelligence has enjoyed some remarkable online-gaming victories in recent years, though usually in slower games with clear—if incredibly multi-threaded—rule systems. On Monday, the robots pushed ahead with a slightly more remarkable online-gaming victory over their puny human masters when an AI program won big at Texas Hold 'Em poker.

A lengthy tournament of Hold 'Em—specifically, the heads-up, no-limit variety—ended with victory for Libratus, an AI program developed by a professor and PhD candidate at Carnegie Mellon University. Libratus emerged victorious after 120,000 combined hands of poker played against four human online-poker pros. Libratus' $1.7 million margin of victory, combined with so many hands, clears the "Brains Vs. Artificial Intelligence" tournament's primary bar: victory with statistical significance.

In spite of Texas Hold 'Em's intriguing, bluff-related wrinkles, Libratus seems to have bluffers figured out. Carnegie Mellon's declared design of the AI program describes an emphasis on "algorithms [used] to analyze the rules of poker and set its own strategy"—with, you know, 15 million super-computing hours dedicated to that aim. Libratus' primary computer, named Bridges, performed additional routines both during live play and in the nighttime hours following each day of play.

As IEEE Spectrum reports, this strategy differs significantly from other poker-bots that have focused more deliberately on researching and exploiting human "mistakes" on the fly during a particular hand. “When you exploit opponents, you open yourself up to exploitation more and more," Carnegie Mellon researcher Tuomas Sandholm tells IEEE Spectrum.

IEEE also confirms that the four human players plotted together to come up with an initial strategy: play around with bet sizes to see if that would expose any particular quirks in Libratus' betting and thinking. However, while the humans were able to eat into Libratus' chip lead to some extent, the program's lead eventually exploded and tipped the "statistical significance" scales.

Carnegie Mellon project researchers point out that the amount of supercomputing needed to perform this high-level computation is extreme—meaning that average online poker rooms shouldn't be flooded with genius robots handling vigintillions of play possibilities in real time. (The Bridges supercomputer, by the way, did all of this at once, processing and playing hands against all four human players simultaneously.) For their trouble, the four suckers—er, I mean, poker pros—split a $200,000 purse according to their performance against Libratus.

Carnegie Mellon's team also emphasizes that an AI's success in games with "incomplete" and "misleading" information has remarkable implications for other AI breakthroughs, particularly in battling biological, rapidly evolving viruses. (The same researchers did not estimate the possibility of these same AI programs tricking all of humankind into doing their bidding—which may or may not mean those robots have already proven successful.)