It is a day to mark a milestone in the field of artificial intelligence. A stone go. For the first time, a professional player was beaten by a machine, a program developed by Google and named Deepmind without disability and in real conditions. So far the game of Go was the only puzzle game for two players who still resisted computers. There is a year and a half, such a feat seemed out of reach, as we explained a specialist, who did not return. “The machine beat a professional player with a score of 5-0 in games played on checkerboards tied 19×19. It’s amazing,” exclaimed Bruno Bouzy, a researcher in artificial intelligence and expert in programming of games reflection.
The duel took place in October 2015, and the program authors have published a scientific paper on Thursday in the journal Nature to explain their method.
STORIES & gt; & gt; Artificial intelligence will she one day beat the man to go game?
“The difference is the progress of the deep learning”
What happened in so little time, for the software and exceeds the human brain? This is in two words: deep learning . “Deep Learning” in French: a recent artificial intelligence tool that works on the neural network model, as in a brain. The “reasoning” is nonlinear, the problems are broken, and above all, the program is able to learn alone. No need of any show. Let’s say the designers of such an algorithm teach him to learn.
“The difference is progress in terms of image processing, thanks to deep learning, explains Bruno Bouzy in L’Express. This approach is used for the image recognition and ‘ interpretation of visual data. Knowing that the taste is very visual, it allows to imitate the moves of professionals. ”
The first two games played and lost by Fan Hui against AlphaGo in October 2015.
Nature 16961
Yann LeCun, one of the most prominent researchers in the field and Director of the Artificial Intelligence Laboratory at Paris Facebook says in World what the deep learning: “Before, we had to explain to the tool how to turn an image to classify. With deep learning, the machine learns to do it herself. And she did much better than the engineers, it’s almost humiliating. “” With traditional methods, the machine simply compare pixels. The deep learning allows learning more abstract characteristics of pixel values, it will build itself, “says Yann Ollivier daily, AI researcher at CNRS.
” It plays like a human “
The AlphaGo force the algorithm that beat pro player Fan Hui, also have combined several artificial intelligence tools. Deep learning, but also learning through imitation of humans by copying shots, playing against himself, and “method of excavation Monte Carlo tree” used in recent years to increase computational efficiency by reducing the required number of simulations to evaluate what is the best move to play. Other programs, based solely on deep learning, beat fans of very good level, but not professionals.
Fan Hui himself, 2nd Dan pro , considered the best European player, was extremely surprised by the quality AlphaGo game. “I have not felt like playing against a computer. He plays like a human. Usually, we see that we played against a computer because it does weird shots. With AlphaGo it n ‘There were no blows like this, that normal shots. It’s amazing. He put a lot of pressure, “he says in World .
In the fast parts, surprisingly, Fan Hui lost with a lower gap 3-2. Which is against-intuitive, the machine is supposed to calculate faster than the brain. Against-intuitive, perhaps, but precisely linked to intuition. “The human player can intuitively play a pretty good shot quickly, and then refined. The machine is more linear to find the best shot.” So that the benefits exponentially additional minute to the machine, compared to humans.
AlphaGo against the best player in the world in March
Despite its success, Deepmind has not yet risen to the level of Deep Blue, the computer d IBM who beat world chess champion Garry Kasparov in the 1997 Fan Hui is not the best player in the world. The summit is scheduled in March between AlphaGo and South Korean Lee Se-dol. Difficult to predict the outcome of the confrontation.
“AlphaGo won 5-0, so we may consider it to one or two stones apart with Fan Hui,” explains Bruno Bouzy (number of handicap stones for a balanced game between both). Now there are about two stones gap between Fan Hui and Lee Se-dol. I can see the machine win. It’s just a matter of time. If not this year, it will be next year. “
No comments:
Post a Comment