The First Man-Machine
Humans win man-vs-computer poker test
While computers have the edge over humans in chess, the first man-vs.-machine poker championship is over, with a pair of humans taking the crown.
Sponsored by the University of Alberta in Canada, the test pitted two top poker players, Phil Laak and Ali Eslami, against the twelve-member human team behind Polaris, the winner of the 2006 competition for poker-playing bots.
Over the course of several days, Laak and Eslami played independent heads-up games of $10/$20 Texas Hold' Em versus Polaris, playing 500 hands apiece. The results: a draw for the first match (one human win, one computer win), an overall win for Polaris in the second match, and victories for both Laak and Eslami in both rounds three and four, giving the humans the overall victory.
The contest had a twist: each match was essentially a duplicate, with the exact same deck. In one match, the computer sat in the "north" chair; in the other, the computer sat in the "south" chair. Theoretically, each player would interact with the same cards. Both matches were isolated from one another, naturally. Each player's bankroll would be added after the session to determine the overall winner, either the computer or the human players.
Unlike chess, which is largely a matter of prediction, poker involves pure chance and bluffing, as well as statistical analysis. Although the computer lacks any biological "tells" that can tip off an opponent, the Alberta team thought that the two human's mathematical aptitude would allow them to sniff out weaknesses in the program.
That apparently was borne out, as this commentary from Poker Academy's live blog reveals:
"[Eslami] explained that during his huddle with Phil Laak they discussed how they learned from the bot as opposed to learning to play the bot. For exmple [sic] he explained that he used to play 3-5 4-5 type low hands to bluff and represent big hands. He's taken this out of his strategy saying 'if the bot was near equilibrium then he would take the math they had done on it and use it for himself'. He also added that he's playing K low card a lot harder then he typically would as he's sensing Polaris has a slight vulnerability to it."
The blog also notes that both human competitors used Poker Academy's Poker Academy Prospector software to analyze the play, some of which may be found here. The victory also netted the two men some cash: $5,000 awarded to the two for the matches they won by more than 25 bets, and $2,500 for each session that was won or lost by fewer than 25 bets.