AlphaGo won a third decisive game with Lee Sedol

The third victory of artificial intelligence in the historical series of five games in go

Go - East Asian logic game with ancient history and rich cultural traditions. There are several rules and variations, but the main goal is the same: each of the two players needs to isolate the territory of the largest possible size with stones of their own color. The researchers of artificial intelligence in go are attracted by complexity. It's too late to conquer chess - computers have been playing better than people for several years. But in the number of possible positions is much higher (10 ¹⁰⁰ ) than in chess. Modern programs play well, but at the level of go-lovers. Before creating a system that can beat the best of people, at best, a dozen years.

So the experts said before the appearance of AlphaGo. Development from DeepMind (in 2014, the company bought Google) uses the Monte Carlo method. The best of modern computer go systems use this technology. But in AlphaGo, moves also help select neural network policies and values. They can be called close relatives of neural networks that recognize images and speech. The system was first taught the game on 160 thousand games from the KGS server. Then AlphaGo trained in games against herself.

For the first time, the attention of this sample of AI attracted in January. Then Google published a scientific paper on AlphaGo, where the October 2015 match was mentioned. Three-time European champion Fan Hui lost a new game in five out of five games.

It is believed that the European level of go proficiency is lower than in the homeland of the game, in Asia. Questions were raised separate moves and errors of Hui. Therefore, go players and AI enthusiasts eagerly awaited the AlphaGo seoul match against one of the best representatives of the team of people - the ninth professional dan Lee Sedol .
')
Until March, the question was as follows: does AlphaGo have a chance against Lee Sedol ? But Sedol lost both the first and second games. If after the first game the master of the ninth dan showed optimism, then after the second he left in confusion. Sedol said that he had not seen the moments where he had played the game.

Today was the third installment. The game was attended by the former head of Google, Eric Schmidt, one of the important engineers of the company, Jeff Dean, and Sergey Brin, who flew to Seoul the other day. Obviously, the conquest of go symbolizes the importance of machine learning technologies that are the heart of Google.


From left to right: DeepMind founder Demis Hassabis, ninth professional dan owner, Lee Sedol, Google co-founder Sergey Brin.

As in the first game, Sedol went black, that is, the first. Also, one of the good moments would be familiarity with the system - DeepMind engineers hardly have the resources to change something during the match days. At the very least, a rumor claims that Cedol and a team of go experts spent the night in search of AlphaGo's weaknesses. Black's first moves were aggressive. It seems that Sedol was trying to divert the game to another direction, trying to put pressure on the weak points of the system.


Video broadcast of the third installment of the match AlphaGo - Lee Cedol.

During the game, the AI ​​demonstrated the ability to work with complex moments in the form of ko positions. It is believed that computer go systems do poorly with co. By the end of the third game, black had run out of time, so the man had to think for a minute.

After 176 moves, the game ended in victory for the AlphaGo system. The AI ​​had 8 minutes and 31 seconds on the clock.



This is AlphaGo's third consecutive win in the series. The overall outcome of the historical five-game series has been determined. The fate of monetary remuneration has been determined - it is obvious that now he will go to charity and the development of go, as was intended in the event of a human defeat. The last two games on Sunday and Tuesday will determine the depth of the demonstration of the benefits of AlphaGo - despite the clear conclusion of the series, it will still play.


Last March, the system beat the player of the second professional dan, today she bypasses the ninth dan. Perhaps the fact that the AlphaGo system has been training against itself for the past five months has helped. According to the head of DeepMind Demis Hassabis, the current configuration uses about the same amount of hardware resources as the version that beat Fan Hui. Hassabis says that algorithms are more important, not “iron”.

Tuesday will be the last, fifth game. Then there will be summed up the battle of thousands of years of the human game of go and two-year development, launched on a cluster of several dozen commercially available server processors of graphics accelerators.

Useful materials on the topic:

• Party moves .
• DeepMind channel with broadcasts of games on YouTube . Commented on in English by Michael Redmond (9 professional dan) and Chris Herlock.
• The review of AlphaGo algorithms published yesterday on "Habré" .