Libratus, the artificial intelligence powered by a multi-million dollar supercomputer that crushed teams of poker players in No-Limit Hold’em twice this year, was honored with the 2017 HPCwire Readers’ Choice Award for Best Use of AI last week. The Libratus Poker AI was developed by Tuomas Sandholm and Noam Brown at Carnegie Mellon University.
I would suspect almost nobody reading this has any idea what HPCwire is – I know I didn’t before I read about the award. HPCwire calls itself “the leading publication for news and information for the high performance computing industry,” and if you visit its site, you would understand why. If anyone knows about high performance computing, it is the contributors and owners of that website.
Tom Tabor, CEO of Tabor Communications Inc., the company that publishes HPCwire, said, “HPCwire’s readership is broadly diversified; it includes industry leaders from the private sector, innovators in academia, and end users that are bringing HPC to the enterprise. Being selected to win either a Readers’ or Editors’ Choice Award is no small feat.”
At the beginning of 2017, Libratus took on the poker pro team of Jimmy Chou, Dong Kim, Jason Les, and Daniel McAulay. Each human played 30,000 heads-up hands against the computer over the course of about three weeks.
In order to try to eliminate as much luck as possible so that skill could be measured, the heads-up matches were played with a few special rules. First, the 20,000 chip stacks (50/100 blinds) were reset after each hand so that nobody could gain an advantage by swinging a big stack around like a club. Second, if there was an all-in and a call before the river, no more cards were dealt. Instead, chips were split according to the players’ equity in the hand. This way, nobody could get lucky by slamming a two-outer on the river. And third, hands were mirrored. That is, in a pair of matches, the hands dealt were exactly the same, except Libratus received one set of hole cards in one of the matches, while the human opponent received those same hole cards in the other.
For instance – and I don’t know how the pairings were actually setup – Jimmy Chou may have been dealt Aces against Libratus’ Queens in Hand #15,306. This hand would have been mirrored in Dong Kim’s Hand #15,306, in that he would have been the one to receive the Queens and Libratus would have gotten the Aces. Thus, it could not be said that either the humans or the AI were the beneficiaries of a lucky streak of hole cards.
Libratus won $14.72 per hand on average from the four players, for a total of $1,766,250. Kim was the most successful, only losing $85,649. PokerListings.com calculated that the probability of the four players actually outplaying Libratus and still losing that much money was between 0.0001 and 0.54 percent.
In April, Libratus did it again, beating a team of six Chinese players in Hainan, China, led by 2016 World Series of Poker bracelet winner Tue Du. Libratus did even better, dominating them for $22.00 per hand. There was real money on the line this time, as well, as Libratus earned a $290,000 purse for the victory.
