About   |   Articles archive   |   Contact us!   |      |     

Artificial intelligence has learned to bluff

24.07.2019
Business, Science, Technology

Specialists at Facebook and Carnegie Mellon University have reached a new stage in the development of artificial intelligence (AI). If earlier he had already defeated people in games such as chess and go, where strategic thinking is required, now the Pluribus bot they have created has defeated professional poker players. The computer not only calculated possible situations, but also deceived rivals when he had to bluff.
Another bot developed by Carnegie Mellon bot Libratus two years ago had already defeated professional poker players. But then the game was played one on one, and now AI confronted five players at once, each of whom earned more than $ 1 million in his professional career in poker, notes The Wall Street Journal.

AI is usually lost in unpredictable situations, which so far limits its use, experts say. Therefore, games such as poker, where there is a characteristic uncertainty in real life, especially attract AI developers, writes WSJ. “[In poker] there is hidden information, and the situation is complicated by the fact that the opponent knows what you do not know,” says the publication of Tuomas Sandholm, a professor at Carnegie Mellon University and one of the Pluribus developers. “You have to wonder if your opponent is trying to deceive you.”

When playing with multiple opponents, it’s even harder to bluff yourself. In order to hone his bluffing skills and develop his own strategy, Pluribus played trillions of times against five of his own clones, each time analyzing his decisions, told WSJ one of its creators, Noam Brown from Facebook AI Research. If they were successful, it was more likely that he would select them again in the future. Thus, Pluribus realized that he could win even with weak cards in his hands if he forced his opponent to fold. “People believe that the ability to bluff is an exclusively human quality and for that you need to look into the eyes of another,” notes Brown. - In fact, this is math. We can create an AI algorithm that can bluff better than any person. ”

For example, in one of the draws, Pluribus got an ace and a deuce of clubs. He started the game as standard, raising the rate to $ 250. Two people took it, the rest of them folded. The first three common cards laid out on the table were the peak jack, the five of diamonds and the king of clubs. Pluribus decided to bluff and bet $ 800. One player passed, but the other continued to play. The fourth common card was the troika of worms. The computer had only a small chance of a straight (the fifth best combination consisting of five consecutive cards), but he again decided to bluff, putting another $ 2,400. The last remaining player, Linus Leliger, made the call. The eight of the peak was the fifth common card, and Pluribus would most likely have lost. But he went all-in, putting all his $ 6,550 he had. Leliger gave up, although he had a stronger combination.

WSJ failed to get comments from Leliger. But another loser player, Jason Les, said that Pluribus was playing aggressively: “I had a feeling of hopelessness. It seemed impossible to win. ”

"AI has an unfair advantage over people: he does not get tired, does not experience hunger and emotions," says another loser, Michael Galliano, a player with 11 years of experience. Therefore, you cannot use your opponent’s mental and physical fatigue to make him make a mistake, which is an important part of the game between people. According to Galliano, this experience brought him to the idea that in poker, as in other areas, data analysis is becoming increasingly important.

Another sign of AI progress is that Pluribus has only two chips, thanks to which it used less than 128 GB of memory. While its predecessor Libratus, who played against one opponent in 2017, had 100 chips, the WSJ said. For comparison: IBM’s Deep Blue computer from IBM, who beat Garry Kasparov in chess in 1997, had 480 special chips. There are many more possible combinations in go than in chess, since the game is played on a 19x19 board. Therefore, the computer-developed AlphaGo computer for Google had 1920 chips. In January 2016, he was able to outright (5: 0) defeat the European champion in go, Fan Hui, a Frenchman of Chinese origin. In March of the same year, AlphaGo played against Korean Lee Sedol, who is considered one of the best players in the world. But he did not perform much better, having managed to take only one batch out of five from the computer.

IN.info ©, 2017