The publication Science Daily rummaged in the brains of artificial intelligence Libratus and figured out how miraculously he managed to take out a pack of famous players in poker. Poker - it's not Go or Chess, where it's easier to calculate the situation.
Why it's interesting
But to begin with a couple of words about why the work of Libratus is very cool, even if you do not care about card games. AI has long endured people in games, where everything is tied up in mathematics, the theory of probability and the need to calculate many moves forward. But when AI was put at a poker table with mother uncles who smoke and bluff a lot, he was lost and poured all the money. And all because in the same unlimited Texas poker a lot of players (from 3 to 10), each have 2 unknown cards on their hands and 5 community cards, which creates a huge number of possible variants of events. Therefore, AI for the time being played the textbook, but was lost, as soon as a difficult situation arose at the table, when one or several players go all-in. And this is emotion and cunning, not mathematics.
How it works
The solution of the developers is simple and ingenious. Instead of forcing the program to do a lot of complex calculations, they taught it to generalize incoming information such as the amount of bets and the number of players to reduce each action to a typical scenario. They also divided the game into 3 conditional stages. Due to this in the early stages of AI could play neatly, and already the final stage to include an aggressive regime and make large bets, for which an ordinary person will need iron eggs. At the same time, Libratus constantly self-taught, analyzing his mistakes and holes in the game opponents.
As a result, he won not only 2 million from opponents of people, but also easily beat Baby Tartanian8, the most powerful poker program at the time. And looking at this it becomes clear why the big Ilon Mask is afraid of AI, and game rooms like Poker Stars are gradually squeezing poker out in favor of casinos and sports betting exchanges.
Source: Science Daily