Friday 30 March 2018

The fastest-improving chess player in the world?

Back in December, I woke up on a day off work, preparing to visit the London Chess Classic - my first time spectating at a super-GM chess tournament. I was expecting to see some top-level chess, but I wasn't ready for the news of a breakthrough in chess computing courtesy of Google: AlphaZero.

The Google-owned DeepMind team had put their absurdly powerful AI computing power onto the game of chess. And with that absurd computing power, had managed in a matter of hours to become the best chess player ever.

Lots of people didn't think that AI, neural networks would be very good at chess. We already had very powerful chess engines, powerful enough to beat any human with one second/move on a mobile phone. But they work by assessing millions of positions using human heuristics - counting up the point value of the pieces, the squares they're on etc. AlphaZero, on the other hand, was told only the rules of the game, and given a process to play millions of games against itself and figure out how to win by itself.

An impressive technological achievement but what was even more exciting were the example games given. Several of the games had beautiful, long-term positional sacrifices from AlphaZero. The kind of moves swashbuckling grandmasters used to play, before computer engines told them they didn't work.

There was only one drawback. No one got to play against AlphaZero. It was all hidden away in Google laboratories, only 20 games released, and there was no opportunity to have a go ourselves.

Enter Leela Chess Zero. Dedicated amateurs have recreated the kind of system that AlphaZero used - except free, open source, and reliant on solely on volunteers, their expertise and their computing power (and energy bills!).

You can play a version of it right now.

It's not as strong as AlphaZero yet. But estimates suggest that within a year or so, it could be competitive. In the little time since it has been launched, it's played about 1.5 million games against itself, all on people's computers. And it's already getting to around 2000 ELO strength - better than most amateur/club players.

This is a really exciting project and well worth dedicating a bit of computing power to, if you like chess or AI. You can find out how to contribute here and join the active discord channel here.