Libratus AI Beat the Pros in No-limit
As for human superiority in Poker, maybe it's time To pass
In cards, you need to Know when to play a Hand and when to foldCarnegie Mellon University researchers put Their cards on the table By publishing a paper in The journal Science. In the paper, the researchers Explained how they taught the Libratus AI program to beat Professional players in no-limit Hold'em. It looks like another Domino Fell in the series of Experiments "man vs machine" - at First this is there were Checkers and chess, then the Ancient game of go. Here, the contestant does not Know exactly what cards his Opponents have.
Therefore, a loophole appears in The form of a bluff.
Poker is different from these games
For quite a long time, There was an opinion that A game based on incomplete Information would be very difficult For machine learning. Professor Tuomas Sandholm and PhD Candidate Noam Brown have shown How this can be done.
Libratus won at the end Of a -day competition with Four poker pros held at Rivers Casino in Pittsburgh.
The program beat each player One - on-one in Head'S-Up, No-Limit Texas Hold'em and collected more Than $. million in chips, playing, hands. The scientists said: "Libratus technology Does not use expert knowledge Or human data, and is Not tailored to poker. The program is applicable for A wide range of games With incomplete information." And not just for games. Making decisions based on incomplete Information is a key skill For businesses, Finance, cyber security And military strategy development with tactics. How did the researchers succeed? Using a three-step approach. First, they developed an algorithm That simplified the solutions used In a typical poker game. The algorithm produced a template For the game, detailed for The first rounds, but more General for subsequent rounds. "It is intuitively clear That there is a big Difference between the flash king And flash with the lady. If you interpret these two Hands as identical, it reduces The difficulty of the game And makes calculations easier." As the game progresses To its climax, a second Module is added to improve The template. It defines a real-time strategy. If the opponent makes an Unexpected move, the strategy is Reworked to take into account The opponent's decision. The third module analyzes how Much the opponent puts Libratus, And thereby tries to find Gaps in his strategy. This way, the program gets More information to choose from.
Libratus won over similar programs, Like Baby Tartanian, and then Won outright against Jason Les, Dong Kim, Daniel McCauley, and Jimmy Chou.
The latter commented on the Event: "The most amazing thing Is that the program can adapt. She is constantly learning and improving. She tested us to find Our weaknesses. You should be happy about Every chip you get out Of Libratus hands." However, Professor Sandholm himself Considers the consequences more serious. In his opinion, the dramatic Consequences will be that the Best AI programs can outperform People in strategic thinking based On incomplete information.