FRIDAY, JULY 12, 2019 © 2019 Dow Jones & Company, Inc. All Rights Reserved. Computers Can Now Like a Champ. Better, Actually. A new artificial intelligence program is so advanced at a key human skill — deception — that it wiped out five human poker players with one lousy hand.

By Daniela Hernandez A super-powered computer performed trillions of calculations to hone an elusive skill: bluffing. Then it proceeded to wipe out five human poker champions with one lousy hand. “It [felt] very hopeless. You don’t feel like there’s anything you can do to win,” said Jason Les, who’s played poker professionally for 15 years and fell repeatedly to the poker bot’s bluffs. Researchers at Facebook Inc. and Carnegie Mellon University have built the most effective non-human bluffer to date. The bot, called Pluribus, is a leap forward in the capabilities of artificial intelligence — and also a cunning player of Texas Hold ’em. Artificial intelligence systems Carnegie Mellon University Professor of Computer Science Tuomas Sandholm, in partner- ship with researchers at Facebook Inc., developed Pluribus. The artificial intelligence developed at academic and corpo- system reasoned and bluffed in real time to best five human poker champions at their own rate labs have a disturbingly good game. PHOTO: CARNEGIE MELLON UNIVERSITY track record of beating humans at their own games. Pluribus, described two chips. By comparison, Libratus Brown, a Facebook AI research scien- in a paper published in the journal used 100 chips in its one-on-one tist and one of Pluribus’s creators. Science, follows in the digital foot- poker matches. AlphaGo, developed Multiplayer poker is considered less steps of superhuman AIs that have by Alphabet Inc.’s Google DeepMind, a game than an artform that requires vanquished humans at games like used 1,920 chips against a human a multitude of skills, especially the checkers, chess, “Jeopardy!”, Dota-2 Go player. International Business ability to read human interactions and and Go. Two years ago, another AI Machines Corp.’s Deep Blue used leverage that knowledge to exploit system developed at Carnegie Mellon, 480 custom-designed chips against mistakes and weaknesses. called Libratus, even bested a poker top chess champ Garry Kasparov, Pluribus developed its winning star. according to the paper. poker and superb bluffing But Libratus beat just one human at DeepMind declined to comment. skills by playing trillions of hands a time in a two-player game. Pluribus IBM did not immediately respond to a against five other clones of itself, said beat five opponents at once — without request for comment. Dr. Brown. breaking a sweat. Pluribus couldn’t try to predict After each round, it analyzed its According to its creators, the new the endgame; playing poker against decisions. If these resulted in wins, bot uses less than 128 gigabytes of multiple opponents meant it had to be the bot would be more likely to opt for memory while playing, and ran on able to reason in real time, said Noam such moves in the future. (over please) Pluribus’s digital brain realized it of getting a flush. It still had a slim Sandholm said, but it’s much harder could win by making a bet when it chance at a straight if it got lucky with to bluff successfully when juggling had a weak hand by forcing its oppo- the final card. It chose to bluff with multiple opponents. nent to fold, which also taught it a $2,400 bet, and the last standing Poker allows researchers to test that it should bluff in future plays, human, Linus Loeliger, called. The algorithmic for dealing said Dr. Brown. It then used those final community card was an eight with the unknown and to build the lessons to make real-time decisions of spades. The AI’s ace was most foundations for software that can when battling top human players, likely a losing hand, but it went all sleuth out fraud and deception in all of whom had earned more than in, betting all its chips, worth $6,550. real-world settings. Dr. Sandholm is $1 million playing professionally, Mr. Loeliger folded. He had a ten of involved in two startups, Strategy according to the paper. diamonds and a king of diamonds, a Robot Inc. and Optimized Markets “People have this notion that strong hand. He would have won. Inc., that are leveraging technology [bluffing] is a very human ability — Mr. Loeliger could not be reached similar to that baked into Pluribus that it’s about looking into the other for comment. for applications in defense, financial Unpredictability is what made services, gaming and health care. playing Pluribus difficult, profes- Facebook says it doesn’t have imme- sional poker players said. It’s also at diate plans to commercialize the tech- the core of the advance, experts said. nology. It exploits the very essence of poker, “A good AI has a ridiculously unfair uncertainty, by using math. advantage against humans: They Mr. Les, the other poker pro who don’t get tired. They don’t get hungry. lost to Pluribus and was also beaten They don’t deal with emotions,” said by its predecessor, Libratus, said the Michael “Gags” Gagliano, a profes- new bot’s moves were aggressive. AI WinnerBut can they takes learn to it love? all sional poker player with 11 years of development “is moving more rapidly experience who also lost to Pluribus. than people realize.” person’s eyes,” Dr. Brown said. “It’s Scientists are interested in building Playing successfully hinges on skill, really about math, and this is what’s AIs that can play games like poker and but also on how players deal with going on here. We can create an AI StarCraft where uncertainty abounds fatigue and stress, Mr. Gagliano said. algorithm that can bluff better than because they are microcosms of Exploiting their lack of mental and any human.” the real world, which is unpredict- physical stamina to force them to In one game against five human able. Traditionally, AI has struggled make a mistake is a big part of the players, Pluribus was dealt an ace of in uncertain situations, limiting the game. clubs and a two of clubs, not a great range of applications to which it can “The algorithm is not doing that. hand. It started by raising $250, a be applied, according to AI experts. It’s just sitting and waiting,” he said. standard move. Two humans called, In poker, “there’s hidden informa- “Any time you slip up ... it’s going to the other two folded, for a total tion, and to make matters worse, collect money in that situation. It’s size of $800. Three community cards, your adversary know things you don’t really difficult.” a jack of spades, a five of diamonds, know,” said Tuomas Sandholm, a Mr. Gagliano said the experience and a king of clubs, were dealt. (In Carnegie Mellon University professor made him more cognizant that poker, Texas Hold ’em, players share a set of and a Pluribus developer. “You have to like the rest of our lives, is becoming cards.) One of the humans checked. think about whether your adversary is increasingly more about the data. Pluribus bluffed and bet $800. trying to deceive you.” “It made me think about the theory Another player folded, while the other The more adversaries, the more behind poker, and the math of each called. hidden information an AI needs to situation,” said Mr. Gagliano. “Instead The next card was a three of contend with. Previous poker-playing of playing the player, you’re sticking hearts, which killed the bot’s chances bots also had the ability to bluff, Dr. to the stats.”

The Publisher’s Sale Of This Reprint Does Not Constitute Or Imply Any Endorsement Or Sponsorship Of Any Product, Service, Company Or Organization. Custom Reprints 800.843.0008 www.djreprints.com DO NOT EDIT OR ALTER REPRINT/REPRODUCTIONS NOT PERMITTED 56714