Artificial intelligence for the first time in the world won a professional go

It was Google's AlphaGo

Google ’s DeepMind division stated that the company's artificial intelligence was able to defeat the European champion of the go board game. The AlphaGo system beat a man in 5 of 5 games. Before that, it was one of the few logical games, professional players in which they won over computers.

One of the visual indicators of the development of artificial intelligence is a victory in logic games. The AI ​​can beat the champion in any logic game, which will demonstrate that the algorithm can solve a problem better than a human. Over the years, the number of conquered games has grown: both checkers and chess have surrendered. In 1996, for the first time, the algorithm won against the best chess player: it was the fight of the computer Deep Blue against Kasparov. And in 2005, the man last won the best algorithm. Since then, computer programs can beat any chess player. Other games are amenable: Aibiemsky Watson played and won in Jeopardy, and in 2014, the artificial intelligence of search giant Google independently mastered 49 old arcade games Atari .

But some games are not conquered yet. One of the nevzyatyh long time remained go. This is a board game originated in ancient China several thousand years ago. A rectangular board of 19 × 19 lines is filled with black and white stones. Each of the players has the task to isolate a larger area on the game board with stones than the enemy. The game has several rules that complicated the creation of an effective artificial intelligence system for defeating a person. For example, the possible positions of stones on a standard board are more than in a googol ( ^10,100 ) times more than in chess. The number of possible positions is greater than the atoms in the universe. It’s just impossible to calculate all the moves, and so far the best computer systems are played at the amateur level.
')
Many moves are dictated by simple intuition, and this is difficult to put into the algorithm. It is this complexity that attracts the attention of specialists in artificial intelligence. DeepMind is a developer of artificial intelligence systems, which Google acquired in 2014. DeepMind was able to create software that is able to beat the champions.

Creating a search tree is not suitable here, so the AlphaGo system was created. It is based on the search for Monte Carlo and deep neural networks. Neural networks pass the description of the state of the board through 12 different layers consisting of millions of neural-like connections. One of the networks, the “policy network”, chooses the next move. The other, the “value network,” predicts a winner.



The neural network was trained for 30 million moves of lots of real people. The result of a correct prediction of the next move was achieved in 57% of cases. Before AlphaGo, the best result was 44% . But the goal was victory, and not just imitation of man. AlphaGo has learned this through thousands of batches between its own neural networks and by improving connections in the reinforcement learning process. In this case, the whole process required considerable computational power, so everything started in the Google Cloud Platform cloud.

First, the resulting product was tested with other best solutions. AlphaGo won 499 matches out of 500. Then, a judge from the British Go Federation, editor of Nature magazine and three - time European champion Fan Hui were invited. This professional player has been involved in the Asian board game since 12 years old. The match behind closed doors was held in the London office of Google in October last year.

To run the algorithm, a computational cluster of 170 video cards and 1200 processors was required (probably, separate cores were meant). Fan was surprised to find that he lost the computer in the first game. The champion wrote off a defeat on his own non-aggressive style. He considered that it was only a warm-up, and began to play more aggressively. But Fan lost all of the four subsequent games. AlphaGo algorithm won in five out of five games.



As they say at Google, this is the first time that the program has been able to beat a professional go player. It should be noted that Fen has a champion title in 2013, 2014 and 2015 only in Europe, where his level of possession is not very high. The next logical step is the match in Seoul in March against the legendary Korean go-pro Lee Sedol , the best go player in the last decade. The level of play of this person will be much higher. For this match, the performance of the system will be improved so that it can be run on more modest hardware.

In go go tens of millions of people around the world. The gradual conquest of another logical game - this is important. But it is also interesting that AlphaGo was not created using manually defined rules. Winning helped machine learning.

Google hopes to use this experience to solve real-world problems. The fact of using general-purpose methods means that such algorithms can be used in many systems: from climate modeling and disease analysis to stock trading on the exchange.

On Wednesday, the results of the study were published in the scientific journal Nature. You can download recordings of games with Fane from the DeepMind website , or you can watch them in the web widget .

Similar development involved in Facebook. Mark Zuckerberg said on Wednesday that his researchers are close to conquering the Chinese game. Zuckerberg literally keeps a close eye on the development process: the author of the project sits six meters from the desk of the executive director of Facebook.



Facebook Research: arXiv: 1511.06410 [cs.LG]