Chess is a difficult game. At least by human standards. But for heavy-duty artificial intelligence this is a common hobby that can be mastered in a very short time. In an article published in the online library Arxiv.org engineers from Google share details about how their latest AI configuration AlphaZero were able to develop “superhuman capabilities” in chess just a few hours.
After the system has made only a basic set of rules (without strategies), AlphaZero took only four hours to master the game so well that it fell even the strongest in the world of the Stockfish chess engine.
After a series of 100 games against Stockfish AI algorithm AlphaZero has won 25 times a game with the white pieces three times in the game black. All the other parties both systems are tied. In the end, Stockfish could not win a single time, and AlphaZero, in turn, never lost.
“We have a new Lord of the chess. This is without a doubt revolutionaire the game, but it would be nice to think about how it can be applied outside of chess,” commented David Kramaley, a researcher of chess and editor of the website Chessable.
AlphaZero system is based on the work of AI algorithms AlphaGo and AlphaGo Zero, which are the creators of the engineers of the laboratory of artificial intelligence DeepMind owned by Google. DeepMind is working to improve these AI algorithms for several years, along the way winning with them best players in the world in the ancient logical game. The culmination of the winning series took place in October of this year, when new and fully standalone version of the AI-algorithm AlphaGo trained exclusively on the game itself, and not against human opponents, have won all their previous versions.
At the same time, the previous version of the algorithm AlphaGo Zero partially studied the game through the observation of how people play. Thus the developers wanted to help her in learning the strategies of the game, but as it turned out, in fact it only slowed down in development. Full all yourself when training affected significantly more effective results AlphaGo Zero in competitions one-on-one.
“It’s like watching an alien civilization, inventing their own mathematics,” — said in an interview with Gizmodo back in October, Nick Hines, a scientist at MIT working in the field of computer science.
“What we’re seeing here, is an independent, self-sufficient model created with the excluded factor human prejudice. She independently is able to learn what she deems optimal, which will certainly differentiate it from our own concepts.”
However, the development of AI happens so fast that what was achieved by October of this year, could already be out of date. In the published article, DeepMind engineers celebrate as the latest version AlphaZero has moved to a new level and is able to tackle a wider range of tasks. This means that now AlphaZero is able not only to play chess. It also copes with Shogi (Japanese Board game) and go in which to achieve perfection, she managed just 2 and 8 hours, respectively.
While neither Google engineers or engineers DeepMind openly made comments about the new results of their work, they expect third-party and independent expert analysis of the articles in Arxiv.org. However, it is clear that the ascent of this algorithm to the peaks of artificial intelligence is still far from complete, but now it is so cool that it is recognized even by grandmasters.
“I always wanted to see how it would have looked like a game of chess among the more advanced species landed on Earth. Now I know”, — commented on the results AlphaZero grandmaster Peter Heine Nielsen.
Only 4 hours Google “nurtured” the best chess players in history
Nikolai Khizhnyak