Posted on Wednesday, February 01 2017 @ 13:51 CET by Thomas De MaesschalckThe four poker experts competed with Libratus in Heads-Up, No-Limit Texas Hold’em, the battle began on January 11 in the Rivers Casino in Pittsburgh and the AI got progressively better as it gained more experience. Every night, the AI analyzed what its human competitors did and trained itself on a Pittsburgh Supercomputing Center (PSC) system named Bridges:
Libratus is an artificial intelligence (AI) program for playing Heads-Up, No-Limit Texas Hold’em. It was developed at Carnegie Mellon University’s School of Computer Science by Prof. Tuomas Sandholm and Ph.D. student Noam Brown. Libratus’s strategy is not based on the experience of expert human players, so its game play could differ markedly from the pros. It uses algorithms to analyze the rules of poker and set its own strategy, based on approximately 15 million core hours of computation at the Pittsburgh Supercomputing Center (PSC). Libratus will continuously sharpen its strategy during the Brains vs. AI competition, performing computations with the PSC’s Bridges computer each night while the pros get some shuteye. During games, Bridges will perform live computations to aid Libratus with its end-game play. The algorithms that created Libratus are not specific to poker. The AI’s ability to reason when faced with incomplete or misleading information have a wide range of possible applications, including business negotiation, medicine, cybersecurity, auctions and more. Carnegie Mellon is a leading center for artificial intelligence research, with pioneering breakthroughs in self-driving cars, computer vision, automated translation, market design and machine learning. CMU’s DNA can be found in the Deep Blue program that defeated a chess grandmaster in 1997, the Watson AI that beat Jeopardy! champions in 2011, and Apple’s Siri digital assistant. Prof. Sandholm’s work has also included designing and fielding tens of billions of dollars of combinatorial sourcing auctions, and his optimization software runs the nationwide kidney exchange for UNOS.Bridges is rated at 1.35 petaflops, it features 752 regular memory nodes with 22,400 computational cores, a total of 274TB of memory and 48 GPU nodes. During the tournament, Libratus used 600 of the 752 regular memory nodes, each with 128GB RAM.
The first couple of days Libratus was unable to win but the AI improved over time and hit a huge win eight days into the competition. It kept getting better an eventually reached the finish line to win $1,776,250. The Register offers some coverage of the event:
The constant upgrade in difficulty is what made it challenging for the players. It’s “extremely tough as the AI keeps getting better,” [Dong] Kim told viewers while answering questions over a live stream on Twitch.In total, Libratus consumed 19 million core hours of computing, equivalent to 3,300 laptops generating over 2600TB of data throughout the tournament. Tuomas Sandholm, co-creator of Libratus and a machine-learning professor at Carnegie Mellon University (CMU), points out poker is a hard game to master for machines because it’s an imperfect information game. He also explains the algorithms behind Libratus are not game-dependent, they could be applied to other imperfect information situations to find the best strategy. For example, this could include cyber security, finance and military applications.
Libratus upped its game, crushing the chances of victory for team mortal, and charged to the finish line to win a whopping $1,776,250 – equivalent to 14.7 big-blinds per hundred or 147 milli-big-blinds per hand.
The large score is of “statistical significance” and a convincing win for the computer, the researchers say. It wasn’t down to a simple run of good cards, as the game was set up in a way to minimize the effect of luck. The four players were split into two teams of two people. One team plays in the open while the other team is locked in a room with no phones or outside communication. The locked-away team are dealt the same cards at the open team but with places switched: the open team humans gets the locked-away AI's hole cards, the locked-away humans get the open AI's hole cards, and so on. This is supposed to cancel out any run good effects.
