ARTIFICIAL MIND CARRYING OUT pitiful people in GO - Is the revolt of machines just around the corner?
ARTIFICIAL MIND CARRYING OUT pitiful people in GO - Is the revolt of machines just around the corner?
Anonim

Not so long ago, South Korean go master and one of the most titled players in the world, Lee Sedol, announced his retirement and made a dramatic statement: rating through insane efforts. Now there is an entity that cannot be overcome."

Lee talked about the AlphaGo computer, developed by DeepMind, which Google bought for $ 650 million five years ago. The Korean lost to the car back in 2016, but since then the artificial intelligence has only become stronger. In general, the victory of a computer over a person in Go is considered a real breakthrough, which can potentially lead to large-scale changes in the world. Is the Terminator on the horizon already? Let's figure it out.

Programmers have long tested the power of artificial intelligence in challenging games with the best of humans. The Deep Blue computer developed by IBM beat Garry Kasparov in chess back in 1997. Before the match Kasparov thought: “It's just a car. The machines are stupid."

But after the defeat he confessed: "I felt - smelled - that there was a new type of mind at the table."

To defeat Kasparov, Deep Blue used brute computational power: after each move, the program calculated all possible scenarios and made a decision based on this data. But with Go, this approach does not work due to the amount of data that needs to be processed. In go, players take turns placing black and white stones on the board 19 by 19. The objective of the game is to occupy as much territory as possible, while locking up the opponent's stones, preventing him from gaining an advantage. In general, go is similar to the dots game familiar to many from school - only more difficult.

Due to the size of the board, 361 variants are already possible for the first move made by the black stones (in chess - only 20). Accordingly, with each move, the tree of potential alignments only grows. After the first two moves, there are 400 possible developments in chess, and 129,960 in go. Mathematician John Tromp calculated that the number of possible combinations would be 171-digit numbers.

Therefore, in the game of Go people are required not only intelligence and the ability to calculate, but also powerful abstract thinking, strong intuition - qualities that are poorly developed in computers. One of the developers of AlphaGo, Demis Hassabis, said: “This is a very intuitive game. Go masters often say they made a move because it seemed right. " According to him, the masters develop a special aesthetic sense, and a good position just looks beautiful.

Despite the fact that processors became more powerful and faster every year, the search for moves on the tree of possibilities allowed artificial intelligence to reach only the level of a strong amateur in go. Computers beat people, but only got a head start in a few stones. In 2014, David Fotland, one of the pioneers of go for computers, said that programs face the same problem as humans:

“Many players reach a certain amateur peak and cannot get stronger. To overcome this plateau, you need to make some kind of mental leap, and programs have the same problems. You need to look at the whole board, not just local battles. " To overcome this intellectual barrier and simulate the intuition and aesthetic sense of professionals, the AlphaGo developers connected neural networks and deep learning algorithms.

First, AlphaGo's neural networks were fed a database of human games, which included about 30 million moves.After that, he learned to correctly predict the course of a person 57% of the time, although the previous AI record was 44%. Then the developers taught AlphaGo to play against itself - so the computer learned even better to highlight the most profitable moves and develop new strategies.

All this helped to rationalize the processes on which Deep Blue, who beat Kasparov, worked. Now the system does not just play all possible combinations, but also knows how to focus on the most promising scenarios for the development of events. In addition, she finds her bearings even in situations that she has never encountered before. And such, because of the scale of Go, remained. Due to the new mechanism, AlphaGo beat all previously created computer players (while giving them a head start of four stones) and began to defeat professional people.

In October 2015, AlphaGo defeated two-time European champion Frenchman Fan Hui. They played five games, no one got a head start, and the computer won all five. This was the first time that a professional person had been defeated by a machine. After the match, Hui said that he had learned a lot, and this knowledge helped him to add and rise in the international rankings.

Popular by topic