An artificial intelligence strategic reasoning have been limited to
program developed by two-party competition. The ability to beat five other players in such a Carnegie Mellon complicated game opens up new University in opportunities to use AI to solve a wide variety of real-world problems." collaboration with "Playing a six-player game rather than Facebook AI has head-to-head requires fundamental defeated leading changes in how the AI develops its playing strategy," said Brown, who professionals in joined Facebook AI last year. "We're six-player no-limit elated with its performance and believe some of Pluribus' playing strategies Texas hold'em poker, might even change the way pros play the game." the world's most Pluribus' algorithms created some popular form of poker. surprising features into its strategy. For instance, most human players avoid The AI, called Pluribus, defeated poker "donk betting" -- that is, ending one professional Darren Elias, who holds the round with a call but then starting the record for most World Poker Tour titles, next round with a bet. It's seen as a and Chris "Jesus" Ferguson, winner of weak move that usually doesn't make six World Series of Poker events. Each strategic sense. But Pluribus placed pro separately played 5,000 hands of donk bets far more often than the poker against five copies of Pluribus. professionals it defeated. In another experiment involving 13 pros, "Its major strength is its ability to use all of whom have won more than $1 mixed strategies," Elias said last week as million playing poker, Pluribus played he prepared for the 2019 World Series five pros at a time for a total of 10,000 of Poker main event. "That's the same hands and again emerged victorious. thing that humans try to do. It's a "Pluribus achieved superhuman matter of execution for humans -- to do performance at multi-player poker, this in a perfectly random way and to which is a recognized milestone in do so consistently. Most people just artificial intelligence and in game theory can't." that has been open for decades," said Pluribus registered a solid win with Tuomas Sandholm, Angel Jordan statistical significance, which is Professor of Computer Science, who particularly impressive given its developed Pluribus with Noam Brown, opposition, Elias said. "The bot wasn't who is finishing his Ph.D. in Carnegie just playing against some middle of the Mellon's Computer Science Department road pros. It was playing some of the as a research scientist at Facebook AI. best players in the world." Michael "Gags" Gagliano, who has game. It looks ahead several moves as it earned nearly $2 million in career does so, but not requiring looking ahead earnings, also competed against all the way to the end of the game, Pluribus. which would be computationally prohibitive. Limited-lookahead search is "It was incredibly fascinating getting to a standard approach in play against the poker bot and seeing perfect-information games, but is some of the strategies it chose" said extremely challenging in Gagliano. "There were several plays imperfect-information games. A new that humans simply are not making at limited-lookahead search algorithm is all, especially relating to its bet sizing. the main breakthrough that enabled Bots/AI are an important part in the Pluribus to achieve superhuman evolution of poker, and it was amazing multi-player poker. to have first-hand experience in this large step toward the future." Specifically, the search is an imperfect-information-game solve of a All of the AIs that displayed limited-lookahead subgame. At the superhuman skills at two-player games leaves of that subgame, the AI considers did so by approximating what's called a five possible continuation strategies Nash equilibrium. Named for the late each opponent and itself might adopt Carnegie Mellon alumnus and Nobel for the rest of the game. The number of laureate John Forbes Nash Jr., a Nash possible continuation strategies is far equilibrium is a pair of strategies (one larger, but the researchers found that per player) where neither player can their algorithm only needs to consider benefit from changing strategy as long five continuation strategies per player as the other player's strategy remains at each leaf to compute a strong, the same. Although the AI's strategy balanced overall strategy. guarantees only a result no worse than a tie, the AI emerges victorious if its Pluribus also seeks to be unpredictable. opponent makes miscalculations and For instance, betting would make sense can't maintain the equilibrium. if the AI held the best possible hand, but if the AI bets only when it has the In a game with more than two players, best hand, opponents will quickly catch playing a Nash equilibrium can be a on. So Pluribus calculates how it would losing strategy. So Pluribus dispenses act with every possible hand it could with theoretical guarantees of success hold and then computes a strategy that and develops strategies that is balanced across all of those nevertheless enable it to consistently possibilities. outplay opponents. Though poker is an incredibly Pluribus first computes a "blueprint" complicated game, Pluribus made strategy by playing six copies of itself, efficient use of computation. AIs that which is sufficient for the first round of have achieved recent milestones in betting. From that point on, Pluribus games have used large numbers of does a more detailed search of possible servers and/or farms of GPUs; Libratus moves in a finer-grained abstraction of used around 15 million core hours to hand — rather than selling competitors develop its strategies and, during live on the strength of what it’s holding. game play, used 1,400 CPU cores. “The bot doesn’t view it as deceptive or Pluribus computed its blueprint strategy lying in any way, it just views it as ‘This in eight days using only 12,400 core is the action that’s going to make me hours and used just 28 cores during live the most money in this situation.’” play. Brown said.
games people can play it requires strategy and perfect decision making. Then there’s this article showing an Artificial Intelligence playing poker. The article says that the Artificial Intelligence aka Pluribus was able to defeat a number of pro poker players that has a number of titles. Though poker is an incredibly complicated game, Pluribus made efficient use of computation. AIs that have achieved recent milestones in games have used large numbers of servers and/or farms of GPUs; Libratus used around 15 million core hours to develop its strategies and, during live game play, used 1,400 CPU cores. Pluribus computed its blueprint strategy in eight days using only 12,400 core hours and used just 28 cores during live play. I myself play poker with my friends but never expected that a Artificial Intelligence could conquer the league of poker. Pluribus is different because, more or less, it is analyzing the effect of bluffing — that is, betting with a weak