AI has now mastered chess, Go and poker – where will it end?

2 Feb 2017

A dark day for poker? Image: Pam Walker/Shutterstock

AI had a breakthrough year in 2016, with Google proving its technology could outsmart even the best Go player. But poker? We thought that was impossible.

Outsmarting chess players is something computers programs were destined to do. There is a finite number of moves and everything is out in the open. Run enough sequences, and a well-planned algorithm can prosper.

Think Deep Blue finally beating Garry Kasparov, widely thought of as one of the world’s best players, in 1997.

Outsmarting Go players is something computers programmes were destined to do, too, though with significantly more processing power. There is still a finite number of moves, but that figure is in the millions. Again, though, everything is out in the open.

Think AlphaGo stunning the Go community when it beat Lee Sedol in 2016.

All in

But poker? Poker is completely different, with limited information available and far more variables relating to the imperfect human mind. How can a computer program handle the relatively immeasurable thoughts of a top poker player?

It turns out it can, by acting in a similar way to AlphaGo and Deep Blue.

Libratus, an improved AI program based on a previous one called Claudico from 2015, just beat four professional poker players at Heads-Up No-Limit Texas Hold’em.

The ‘Brains v Artificial Intelligence’ repeat saw four human players face off against Liberatus – Dong Kim, Jimmy Chou, Daniel McAulay and Jason Les, the latter finishing 40th in last year’s World Series of Poker main event.

120,000 rounds, by four

They played separate heads-up games with Libratus, each logging 120,000 hands against the machine and ultimately losing by a cumulative $1.7m in chips. After four days, the margin was $50,000 but, from there, the algorithm proved its worth and the pros were crushed over three weeks.

The developers of Libratus – Prof Tuomas Sandholm and PhD student Noam Brown from Carnegie Mellon University (CMU)  – said the victory is statistically significant and not simply a matter of luck.

“The best AI’s ability to do strategic reasoning with imperfect information has now surpassed that of the best humans,” Sandholm said, with the study published here.

With 10 to the power of 160 (the number one followed by 160 zeroes) information sets, this type of poker is immensely complex for a computer to understand. A total speed of 1.35 petaflops is available, and Libratus drew on 600 of the 846 nodes sported by Pittsburgh Supercomputing Center’s Bridges computer.

Big moment

This evolution of AI is significant for many reasons, most notably the fact that the machine learned to ‘bluff’.

“The computer can’t win at poker if it can’t bluff,” said Frank Pfenning, head of the department at CMU’s School of Computer Science.

“Developing an AI that can do that successfully is a tremendous step forward scientifically, and has numerous applications. Imagine that your smartphone will someday be able to negotiate the best price on a new car for you. That’s just the beginning.”

Business negotiation, military strategy, cybersecurity and medical treatment planning could all benefit from automated decision-making using a Libratus-like AI. The latter benefit, in particular, could be fascinating.

Gordon Hunt was a journalist with Silicon Republic

editorial@siliconrepublic.com