Friday, March 30, 2018

Open source chess engine tries to reproduce AlphaZero algorithms

https://www.reddit.com/r/chess/comments/882jr8/leela_chess_zero_the_fork_of_leela_zero_for_chess/

Some people here probably followed Deepmind's AlphaGo, then AlphaGoZero and finally AlphaZero's achievements. However Google did not release the weights for all thoses networks, so a community has formed trying to reproduce and maybe improve the AlphaZero engine.
To do this, a distributed effort to play self-play games has started, and the training is done on a single powerful computer (It can update the weights in about 5 hours currently).

The developers also allow everyone to play against their engine:
http://play.lczero.org/

Further information can be found on the website  http://lczero.org/ which necessarily is rather technical.

And while people like Magnus Carlsen struggle to reach 2850 Elo, the engine struggles as well... to reach 4250 Elo. Yes, that is a 4, not a 3. But how "real" that number is, is the real question.

No comments:

Post a Comment