Artificial intelligence again defeated man. This time the AI won the poker game. And this, it seems, is not surprising, because in this game, artificial intelligence was the first before. But now everything is different: the car immediately won five players in Texas Hold'em and at the same time was aimed at the highest possible gain.
What is Texas Hold'em?
If you do not go into the details of the rules of the game, thenThis is one of the most common types of poker. In those very scenes that you saw in a lot of movies when players sit around the table and take turns betting, it is hold'em that appears. There are 6 players at the table and they gradually make bets. At any time, you can exit the game, but then you will not get back the invested money. This continues until a tense "battle" one on one, in which the winner takes everything.
In a recent experiment, development specialistsThe AI from Facebook and Carnegie Mellon University put artificial intelligence at the table to the players, who had one task: to win as much money as possible.
Why artificial intelligence could beat a man
Developing a profitable poker algorithmsimultaneously with several players (and in theory and at several tables at the same time) - the task is much more difficult than teaching the AI to play chess or go. Basically, this is because in poker only very limited information is given on what happens at the poker table. That is, if in the same chess the computer “sees” the whole picture, including the disposition of the opponent’s pieces, in AI the poker knows only what cards are in his hands and which are laid out on the table. It is also possible, as they say, to “count the cards”, having foreseen which of them have already “gone”, which will make it possible to predict possible combinations. But it is still more difficult than analyzing the “open” chess field.
Thanks to the new learning algorithms, the system isThe name Pluribus for 12 days and 10,000 hands fought with 12 professional poker players. As a result, Pluribus won an average of $ 5 per hand and about $ 1000 per hour.
See also: How artificial intelligence beat the best poker players.
How AI learned to play poker
For starters, the developers forced the AI to playagainst yourself and by trial and error to choose the most appropriate tactics. This is, in general, the standard way of learning. Another thing is interesting here: Pluribus was created in just eight days using a 64-core server station equipped with only 512 GB of RAM. This training cost the company only about $ 150.
But the AI training did not end there. Programmers came up with a way to work artificial intelligence, which allowed him to play extremely effectively. Instead of trying to predict how his opponents will behave until the end of the game, Pluribus was designed to “look” only two or three moves ahead. This allowed the system resources to be used for more useful actions. For example, on the development of opportunities to bluff.
"The ability to deliberately bluff is consideredthe exclusive prerogative of the people, "- the authors of the publication The Verge. “At the same time, Pluribus devoid of long-term strategic thinking, as noted by almost all the participants in the experiment, very successfully bluffed. For the algorithm, this was just one of the ways to win the distribution, but far from the most basic. ”
Researchers have announced that they will not lay outthe algorithm and everything related to its development is freely available. At the same time, the developments gained during the recent experiment are very important: they will allow, for example, to identify financial frauds in the securities market, help automatic traffic control systems and will be useful even in developing more efficient autopilots.
You can read even more interesting materials in our Telegram news channel.