AI Scientists Achieve ‘Superhuman Performance’ In Poker

The computer scientists behind the first poker bot to win a match against world-class poker professionals are revealing more information about their hand.

Carnegie Mellon University’s Tuomas Sandholm and Noam Brown said in an articlewith the American Association for the Advancement of Science that their bot called Libratus was able to achieve “superhuman performance” in heads-up no-limit hold’em.

In January, Libratus played against a team of four heads-up no-limit hold’em specialists in a 120,000-hand “Brains vs. AI” match over 20 days at a casino in Pittsburgh. The players were Jason Les, Dong Kim, Daniel McCauley and Jimmy Chou. When the dust settled, the pros were down 1,776,250 chips (about 14 big blinds per 100 hands), a decisive victory for the machine. Each human lost to the AI. An earlier version of Libratus lost to a different group of pros in 2015.

Defeating poker pros in a rematch was a tall ask, thanks to the complexity of poker. Sandholm and Brown compared how different limit is versus no-limit.

“An iterative algorithm […] was used to near-optimally solve heads-up limit Texas hold’em, a relatively simple version of poker, which has about 10^13 unique decision points,” they wrote. “In contrast, HUNL has 10^161 decision points, so traversing the entire game tree even once is impossible. Pre-computing a strategy for every decision point is infeasible for such a large game.”

Sandholm and Brown discussed in detail Libratus’ “three main modules” that allowed it to treat each hand of the match individually and find a strategy in real time. Despite 10^161 decision points in HUNL, the computer scientists said that they had to be wary of over-simplifying the game of poker. The humans could have exploited it if they had done so.

“Intuitively, there is little difference between a king-high flush and a queen-high flush. Treating those hands as identical reduces the complexity of the game and thus makes it computationally easier. Nevertheless, there are still differences even between a king-high flush and a queen-high flush. At the highest levels of play, those distinctions may be the difference between winning and losing.”

Many observers credited the bot’s river strategy as the primary reason for why it had so much success against its human counterparts. Libratus used a very balanced and powerful (literally what Libratus means in Latin) river over-bet strategy with both bluffs and value bets that kept the humans generally confused.

“We really got a beat-down,” the poker pro team said after the match.

Specifically, Libratus was able to take into account what are known as “blockers” in a poker hand. For example, if you held the ASpade Suit4Heart Suit on a board reading KSpade Suit7Spade Suit6Heart Suit2Club SuitJSpade Suit, you would have just ace-high but you’d know your opponent can’t have the best possible hand. That helps provides opportunities for massive over-bets on the river with bluffs.

Unlike previous poker AIs, Libratus was let loose to recompute on the river based on what the humans bet, as opposed to having to be pre-computed for the river scenarios.

“We ran our algorithm on an abstraction that is very detailed in the first two rounds of heads-up no-limit hold’em, but relatively coarse in the final two rounds,” Sandholm and Brown wrote. “However, Libratus never plays according to the abstraction solution in the final two rounds. Rather, it uses the abstract blueprint strategy in those rounds only to estimate what reward a player should expect to receive with a particular hand in a subgame. This estimate is used to determine a more precise strategy during actual play.”

Source :

You May Also Like