Almost 20 years ago, the computer Deep Blue stunned the world by becoming the first machine to win a chess game against a world champion, then the Russian Garry Kasparov. Built just for this purpose by IBM - American company focused on the information technology area - Deep Blue was based mainly on the "brute force" to beat Kasparov.
The computer was able to calculate 200 million moves per second, its programming included a database with hundreds of thousands of openings, desenrolares and games closures, both classical and fed by other great masters, to set the destination of its parts. Still, the feat was hailed as a major breakthrough in the field of AI (artificial intelligence). Only now, however, scientists have been able to do the same with the oriental origin game "Go". In a study published in the journal Nature, researchers from Google DeepMind, research arm in IA Area technology giant, based in the UK, report how they used new approaches in neural networks to overcome the challenge.
It created about 2500 years in China, the "Go", also known as "Igo" in Japan and "Baduk" in Korea at first glance seems simple. Armed with white and black stones, two players must place them alternately at the intersections of a square board with 19 lines on each side so as to completely surround the largest possible area or the opponent's pieces, which are thus removed from the game. Who wins in the end have the largest amount of controlled intersections (fully enclosed) and taken stones.
As the tray begins to be filled, however, the complexity of the game increases exponentially. This requires a careful balance between defense and attack, that due to the large area of search and the difficulty of assessing positions and movements, was long considered one of the greatest challenges in the field of AI.
To overcome it, scientists from Google DeepMind created a program, called AlphaGo, which uses two neural networks operating in parallel to make their moves. The first, called "value network", evaluates positions on the board, while the second, "policy network", decide the movements.
Moreover, these two networks were trained with an unprecedented combination of "deep learning" with millions of plays of human experts to find out what would be the most likely answers to your every move, reinforced by machine starts against herself to focus on the result desired departure, the victory, and not in the efficiency of each single move.
In the game.
With this, instead of calculating all the movements and possible results until the end of the game, as did the Deep Blue, the program analyzes only the best momentary positions and in the near future that will give him an advantage in the game, limiting their movements and strategies the more favorable. Thus, the number of evaluations it does is thousands of times smaller than was the IBM computer.
Thus, last October the AlphaGo won a tournament against the current European triple champion of "Go", the professional player of Chinese origin Fan Hui, for five to zero. And next March, he will face South Korean Lee SEDOL, leader of the world rankings the past ten years, in a new series of five matches being seen as an even greater challenge.
"Our approach put AlphaGo much closer than humans do players than any previous attempt," says David Silver, head of the research team responsible for developing the program.
The performance AlphaGo in matches against Hui was so good that some experts even had difficulty identifying who the machine and who was human to be presented to the rolls without knowing beforehand who made them.
Nenhum comentário:
Postar um comentário