Talk of the Day- Power of Self-Learning Systems

Demis Hassabis, Co-Founder & CEO, DeepMind talks about Self-Learning Systems.

Grandmasters might train for an upcoming match by spending many hours and days researching the latest openings and chess developments, and adopt a strict diet and exercise regime. They will also prepare specifically for the opponent they are expecting to face. This specific preparation involves collecting all of the prospective opponent’s games using a huge database of tournament games from all over the world, looking for weaknesses in the opponent’s play, and particularly in their openings set-ups. In the last 20 years or so, top grandmasters have found they have to remember much more than previously, as they feel the need to adopt a wide range of openings to avoid the preparation of their opponents.

By contrast, AlphaZero’s training before the Stockfish match took nine hours. It began training from a clean slate with no chess knowledge other than the rules of the game, and it didn’t look at Stockfish’s play at all. AlphaZero also did not use any available chess openings knowledge, and instead worked out its own openings as it trained and played against itself. At the start of these crucial nine hours, AlphaZero did play chess, but not as we know it. As anyone who has taught chess to a small child will know, random play will get you nowhere when playing chess.

Game Changer by Natasha Regan, Matthew Sadler