Artificial Intelligence for the first time in the world champion won the first Bashny.Net

It was AlphaGo from the company Google h3>


Division DeepMind company Google заявило that artificial intelligence company was able to beat the European champion board game of . System AlphaGo beat man 5 out of 5 games. Prior to this reform was one of the few puzzle games, professional players who have won at computers.

One of the most vivid indicators of development of artificial intelligence is to win the game logic. AI can beat the champion of a puzzle game that will demonstrate that the algorithm can solve the problem better person. Over the years, the number of games is growing conquered: surrendered and checkers, and chess. In 1996, the algorithm was first won by the best player in chess: it was the computer Deep Blue match against Kasparov. And already in 2005, the last time a person won a better algorithm. Since then, computer programs can beat any player in chess. Subjects and other games: aybiemovsky Watson to play and win at Jeopardy, and in 2014 an artificial intelligence search giant Google independently mastered the 49 old arcade games Atari .

But some games have not yet conquered. One nevzyatyh long been second. This board game originated in ancient China several thousand years ago. The rectangular board 19 × 19 lines filled with black and white stones. Before each player is to isolate the problem on the board with stones larger territory than the opponent. The game has a few rules that complicate an effective system of artificial intelligence to win over a man. For example, the possible positions of the stones on the standard board more than googol (10 <sup>100 sup>) times greater than in chess. The number of possible positions larger than atoms in the universe. Just as it is impossible to count all the moves, and so far the best computer systems play at the level of fans.

Many passages in the second dictated by simple intuition, and it is difficult to put in a similar algorithm. It is this complexity of attracting the attention of specialists in artificial intelligence. DeepMind - a developer of artificial intelligence, which Google acquired in 2014. In DeepMind able to create software that is able to beat the champions.

Preparation of the search tree is not suitable, so the system was established AlphaGo. It is based on the search for Monte Carlo and deep neural networks. Neural networks passed description of the board of 12 different layers consisting of millions of neural connections. One of the networks, "network policy", selects the next move. Other, "net value", predicts the winner.

97,872,147



Neural network trained on the 30 million strokes lots of real people. There has been the result of a correct prediction of the next turn in 57% of cases. Before AlphaGo best result was 44% . But the goal was to win, not just imitating humans. AlphaGo learned it through thousands of games between their neural networks and by improving connections in обучения Reinforcement . The entire process required considerable computing power, so everything is to be started in the cloud Google Cloud Platform.

First, the resulting product is tested to other best solutions. AlphaGo won 499 games out of 500 were then invited to a judge of the British Federation of the editor of the journal Nature, and a three-time чемпион Europe Fan Hui. This has been a professional player of Asian board game of 12 years. The match was held behind closed doors in the London office of Google in October last year.

To run the algorithm, потребовался Computing cluster of 170 video card and 1200 processors (probably meaning the individual nuclei). Fan was surprised to find that he had lost a computer in the first game. Champion wrote off defeat on its own non-aggressive. He thought that it was just a warm-up, and began to play more aggressively. But Fan lost all four subsequent batches. Algorithm AlphaGo won five out of five games.

As they say at Google, this is the first time that the program was able to win a professional Go players. The next logical step is to match in Seoul in March against the legendary Korean go-pro Lee Sedol - the best player in the first decade. For this match, the system performance will improve, so it can be run on a more modest equipment.

The play of tens of millions of people worldwide. The gradual conquest of another puzzle game - this is important. But it is also interesting that AlphaGo has not been created by a set of rules manually. Winning helped machine learning.

At Google are hoping to use the experience to solve real-world problems. The fact that the use of general-purpose methods mean that these algorithms will find application in many systems from climate modeling and analysis of diseases to trade shares on the stock exchange.

On Wednesday the results of the study were published in the scientific journal Nature.

Engaged in such developments and to Facebook. Mark Zuckerberg сообщил on Wednesday that its researchers are close to conquer the Chinese game. Zuckerberg has literally been closely following the development process: the author of the project sits twenty feet from the desk executive director of Facebook.



Research Facebook: arXiv: 1511.06410 [cs.LG]

Source: geektimes.ru/post/269990/</sup>