Artificial intelligence just defeated the player in poker game, won 1.7 million

From poker games, this AI can also be used in many different areas such as financial transactions, political negotiations or auction.

Yesterday, after 20 days playing in a Pittsburgh casino, an AI built by two researchers at Carnegie Mellon officially defeated the top 4 players in unlimited Texas Hold'Em games - a particularly complex form. the practice of poker games, which are based largely on long-term betting strategies and game theory.

For more than 20 years, computers have beaten the best of many games, like checkers, chess, Scrabble, and Jeopardy! or even the ancient Finn chess game. But no AI has ever beaten the best in games with "imperfect information" , games with certain hidden elements - for example, cards that are on the table. Even for humans, to win in Texas Hold'Em games without limitation requires a certain level of intuition in the player, which is not to mention luck.

For more than 20 years, computers have beaten the best in many games.

Meanwhile, Professor Tuomas Sandholm of Carnegie Mellon University and his graduate student Noam Brown designed an AI, called Libratus , meaning "Balance" in Latin. Nearly two years ago, these two researchers challenged some of the top players in this Texas Hold'Em game with an AI similar to Libratus and failed. But this time they won easily: after 20 days of playing, Libratus overcame four other players for more than $ 1.7 million, and all four players ended up with negative chips.

Maybe for you, poker is just a game. But according to University of Washington professor Michael Wellman, a game theory expert and a close observer of the AI ​​world in poker, Libratus' method of playing this game can help solve many problems, from Financial transactions until political negotiations, or even auctions.

In unlimited Hold'Em games, players do not have to win every small hand. Instead, they will try to win the most money, and that means they will have to develop a betting strategy to stay on dozens of cards until the end. A computer that can host this unlimited Texas Hold'Em game means that it can mimic the kind of human intuition that these strategies require.

Computers can host unlimited Texas Hold'Em games.

According to human players who failed to defeat this AI, Libratus has a very reasonable way to call. It knows when to cheat, or when to make a low bet with good cards, and when to change bets just to give up the game."It can divide your bet into rounds of 3 cards, 4 cards or 5 cards" . According to Daniel McAulay, 26, one of the players is defeated by this AI. "No human is capable of doing so."

So far, Sandholm still hesitates to share how Libratus works, but he promises to give details in the coming days. Along with McAulay, there are Dong Kim, Jason Les and Jimmy Chou - human players who believe this AI play has changed from day to day. Because when they feel there is a gap in this machine's strategy, that hole will be closed.

Jason Les, one of four failed players against Libratus in Pittsburg on January 11, 2017.

"It seems to be learning what we are doing and exploiting it," McAulay said. Sandholm and Brown may have studied to change the behavior of the machine from day to day, as they did with their first version, Claudiro, the AI ​​has been challenging people for nearly two years. . But this AI can also learn from the game itself when it is happening.

It may seem unfair for researchers at Carnegie Mellon to edit their AI between loops, but it is clear that players can also use all tactics at their disposal. Even though Hold'Em games are played in a "Heads-up" way - that is, the game is played separately between players and computers - but players can share strategies for each other every night.

"We spend a few hours every night discussing it." McAuley said. "We played together in pairs. But we still tried to win for humanity."