And this AI isn't just better than us at poker...
Well, if you were worried about that victory, then you will be positively petrified about the fact that an artificial intelligence algorithm is about to beat four of the world’s best poker players, as this is an achievement which surpasses anything we’ve previously seen in AI research.
Poker is a game of skill and strategy, but it is also a game of bluffing, courage and instinct. For a computer program, these factors make it incredibly complex to master, and of all the games which AI researchers have poured decades of research into (others include chess, checkers, Jeopardy and Go) it stands alone as the one in which humans can still claim to be best at.
But that is all about to change in a casino in Pittsburgh.
There, four of the world’s best no-limits, heads-up Texas Hold’em players — Jason Les, Daniel McAulay, Jimmy Chuo and Dong Kim — have been taking part in a 20-day tournament which ends on Monday. 120,000 hands of poker will have been played and unless there is some miracle on the final day of play, the algorithm — called Libratus — will have scored a significant victory over the elite of the poker world.
While the rapid advance of AI may be worrying for some, for the man who created the algorithm, this victory is a hugely important milestone.
“It is a huge deal,” Tuomas Sandholm, professor of computer science at Carnegie Mellon University told Newstalk.
“Performance against humans has been one of the main benchmarks for AI... Of all of the games that have been seriously worked on in the AI community, this is the only game where AI has not surpassed the best humans.”
At the end of day 19, the poker algorithm held a commanding $1.6 million lead over the players and given that the algorithm gets better with every hand, it is highly unlikely there will be some huge turnaround for the human race on the final day of play.
No-limits heads-up Texas Hold’em is what is known as an “imperfect information” game. Players don’t know what cards their opponents hold, making it massively more difficult for an algorithm to figure out what the correct strategy is in any given hand.
There are said to be approximately 10160 decision points in the game. That number is so high that the algorithm require a $10 million supercomputer to do the number crunching when playing each hand. Coming up with a strategy of how to play was not something Sandholm and his graduate student Noam Brown preprogrammed into the algorithm.
Instead Libratus played trillions of hands against itself in preparation for the tournament in order to come up with a winning strategy, a technique known as reinforcement learning.
The algorithm is not specific to poker however, it was written with the real world in mind. “We are testing whether the best AI's ability to do strategic reasoning under imperfect information has surpassed that of the best humans,” Sandholm said. “The AI's algorithms are not for poker: they are game independent.”
This is why beating no-limit heads-up Texas Hold’em is so important. Not because the game is beaten, but because the algorithm that does it will have much wider applications than one designed to specifically beat Go or chess or checkers.
Among the areas where Libratus can be applied are negotiation, cybersecurity, the military, auctions, finance, strategic pricing, as well as steering evolution and biological adaptation.
Advances in AI have been happening much more rapidly than even those working in the field had anticipated and with the increasing availability of supercomputing power, those advances are likely to happen at ever faster rates.