Among the most recent outstanding achievements of artificial intelligence, one can single out, perhaps, the undoubted successes in chess, and more recently in the exotic oriental game of go. As it turned out, the next step is gambling.
A program called Libratus, developed by a group of researchers at Carnegie Mellon University, literally ripped apart a team of four poker professionals in a 20-day gaming marathon, winning $ 1.76 million from them.
Libratus co-developer Tuomas Sandholm proudly stated that "the best AI, capable of strategic thinking, without having complete information, has surpassed the best people." During the game, opponents of the program tried to identify the weaknesses of Libratus, but this did not help them. The AI gained experience from game to game.
At the end of each game day, Sandholm said, he and his colleagues carefully analyzed the meta-algorithm and refined the strategy of the game, with an emphasis on discovering and exploiting the opponent's “weaknesses”.
And here is how one of Libratus rivals, Jimmy Chow, describes his impressions:
“The computer played better and better day by day. The first couple of days we were still hoping for something. Every day he learned about his shortcomings from us, and the next day he was overcoming the problem. "
Winning poker is all about bluffing. According to Sandholm's colleague Frank Pfenning, the computer's victory was also possible thanks to bluffing. Developing an AI capable of this is an undoubted success.
“Imagine, ” continues Pfenning, “that your smartphone can one day negotiate the best price for a new car. And this is just the beginning. "