Knockout Tournament – Starts at 6pm
As a reminder, we’ll be conducting our unrated knockout tournament tomorrow, 11/25 at 6pm. Please plan to arrive as close to 5:30pm as possible to allow for us to get everyone registered, etc. The tournament is free and open to all, regardless of membership status.
Developments in Chess AI – by David Hayes
For those of you who are following the advanced artificial intelligence AlphaZero by DeepMind. AZ was a topic of the presentation by David Hayes a few months back. AZ beat the best chess computer engines in 2018, but was told the rules of chess (legal piece moves) before it learned to dominate the best computers.
DeepMind’s has developed a new AI called MuZero. MuZero learned the rules of chess by simply watching it being played, and then exceeded AZ’s skill in the game. MuZero was not given any information about its environment.
When humans are born, we have little or no information about our environment. Despite our ignorance, our brains learn the ‘rules’ of our environment, and survive and often thrive. We have gone into very hostile environments (the moon, deep sea, arctic, etc.), and survived, and learned with our general purpose intelligence.
Similarly, MZ has taken a step closer to a general purpose artificial intelligence like human intelligence. MZ dominated chess faster and better than AZ, with less effort (energy). Likewise, it did the same with shogi, go, and 75 Atari games (pacman, asteroids, etc.). The Atari games are visually complex domains that require intelligent development of a model to decipher the blinking lights and pixels, before the intelligent development of a planning model.
MZ did this by predicting the qualities most relevant to each game’s planning. Intuitively, MuZero internally invents game rules or dynamics that lead to accurate planning. Not so long ago, words like predicting and intuition were use to describe only human intelligence, but not anymore.
DeepMind reports that MZ managed a 731% median normalized score compared to 192%, 231%, and 431% for previous state-of-the-art approaches IMPALA, Rainbow, and LASER, respectively, while requiring substantially less training time (12 hours versus Rainbow’s 10 days).
For more information see: https://venturebeat.com/2019/11/20/deepminds-muzero-teaches-itself-how-to-win-at-atari-chess-shogi-and-go/