DeepMind Announces AlphaGo Match With World Go Champion Ke Jie

DeepMind , now part of the Alphabet holding, has been working on the AlphaGo software platform for several years. Last year, she won the go against one of the world champions, Lee Sedol. Thus, the developers of this neural network were able to prove that the computer is quite able to compete with a person in areas that were previously considered the prerogative of people.

The same game was included in the rating of the most difficult for “understanding” computer games. Experts believed that in the near future, the AI ​​(its weak form) will not be able to win not only the world champions, but also ordinary high-level players. However, what happened happened, Alpha Go beat Lee Sedol in four games out of five. And now the turn has come to fight with another champion, Ke Jie , a 9th dan player. Jie takes the first line of the ranking of China and the world by points scored.

Another professional player approved the contribution of artificial intelligence to go, saying that Alpha Go provides a great deal to the development of strategy and tactics. “The game Alpha Go allowed us all to feel free, to understand that any move is possible. Now everyone is trying to play in a new way, in a style that has not yet been tested before, ”said Zhou Ruyang, player 9 is given in go.
')
The creators of the software platform have long stated the desire to bring together AlphGo and Ke Jie, now these plans come true . Two champions will meet, a car and a man at the feast day, which will be held for five days from May 23 to May 27 inclusive in the homeland of this game, in China, in the city of Wuzhen, Zhejiang Province.

Dates - May 23-27. The event will be held jointly by the China Go Association and the Chinese government. The championship, which will be held during the festival, will be attended by famous go players, as well as experts in artificial intelligence.

As for AI, AlphaGo will play not only against one person. The organizers proposed several stages of the competition, including the following:

• “Pair Go” is a game where one of the go champions from China will play against the other ... Everything is traditional, except that both will have their own partner AlphaGo. It is expected that in this format the computer and the person will be able to learn from each other;
• “Team Go” is a competition in which five pros will play against AlphaGo. Professionals will try to beat the system by examining the features of its game. They are allowed to discuss the moves between themselves. The team includes, in addition to the computer, Shi Yue, Chen Yaoe, Mi Uting, Tang Weihing and Zhou Ruyang;
• “Ke Jie vs. AlphaGo” is perhaps the central element of the whole go festival. Two champions will fight in it - a computer and a man.

By the way, Ke Jie is known for his balanced game. He is not too keen on attacking the enemy, remembering the "rears" and developing a reasonable system of attack and defense. When Lee Sedol played with AlphaGo, Jie explained the advantages of some or other disadvantages to Google’s chief, Sundaru Pichai.

On the feast day, 3 games with a time control of 3 hours per game (+ 1 minute) will be played . The winner will receive good prizes - $ 1.5 million US. The loser is $ 300 thousand. This is the largest gaming fund for the game of go, and not only of go. For example, chess grandmasters Karjakin and Carlsen played for 1 million for two (600,000 for the winner, 400,000 for the loser).



According to TheVerge, Jie lost several unofficial games to Alpha Go. Perhaps, in the official game, he is the champion from China, who gathers and shows the best of what he is capable of. In any case, playing with AlphaGo is not a way to show the superiority of AI over a person, but the ability to prove that a car is no worse than a person can cope with solving a number of problems. “Instead of the game“ to destroy, ”as many fear, the AI ​​helps its opponents become stronger by displaying original ways of thinking,” said one of the founders of DeepMind, Demis Hassabis.

AlphaGo consists of a combination of the Monte Carlo method and policy neural networks (value networks) and value networks. To the current level of its professionalism, the system came out a long time after its “birth”. She had to play go hundreds of thousands of times, sorting the games from the KGS server, where masters of various levels play - from the sixth to the ninth dans. Gradually, the system taught itself, and the game was played not only with people, but also with itself. Gradually, AlphaGo began to beat the professionals, and after that won the European champion in go, and then - Lee Sedol.

The game itself is considered more difficult for computers because there are more possible positions of stones on a standard board in googol ( ^10,100 ) than in the same chess. This is more than the number of atoms in the universe. At the very beginning of the game, opponents place stones of various colors on the board and begin to attack each other in the hope of taking an area of ​​their own color larger than the opponent.



By the way, DeepMind is working not only on learning its AI for playing board games, but on training a software platform in a virtual environment, in particular , when playing StarCraft. DeepMind has chosen this game because for 20 years it has been the benchmark among cybersportsmen.