About a decade ago, I created an experiment that tested the Elo value of a position eval.
Bob Hyatt independently verified the the results and the experiment. It came down to a modern
position eval is worth over 400 Elo.
AIUI, Leela is Stockfish with a neural net as the position eval. The rating difference between the
two according to CCRL is almost 800 Elo. So, I question is the ANN doing any good at all. IIUC, Leela
didn''t go the MCTS route and still uses negamax. If so, it may be that Google is right and using an
ANN requires MCTS.
Is Leela doing anything?
Moderators: hgm, Rebel, chrisw
-
CRoberson
- Posts: 2056
- Joined: Mon Mar 13, 2006 2:31 am
- Location: North Carolina, USA
-
Daniel Shawul
- Posts: 4185
- Joined: Tue Mar 14, 2006 11:34 am
- Location: Ethiopia
Re: Is Leela doing anything?
This is like doing an "experiment" where the result can be anything you want it to be -- and then
verify your claim with a "has been" chess programmer.
Where do you get your information from ?AIUI, Leela is Stockfish with a neural net as the position eval. The rating difference between the
two according to CCRL is almost 800 Elo. So, I question is the ANN doing any good at all. IIUC, Leela
didn''t go the MCTS route and still uses negamax. If so, it may be that Google is right and using an
ANN requires MCTS.
No, leela is not Stockfish with neural nets. It just reuses its move generator which is just
good engineering. Leelas has nothing to do with stockfish at all.
No, leela did go the MCTS route.
No, the elo difference is probably not 800 elo on a super hardware, GPU/TPU
-
jp
- Posts: 1470
- Joined: Mon Apr 23, 2018 7:54 am
Re: Is Leela doing anything?
No. Leela does not use negamax. Leela is trying to be as close as possible to Alpha Zero. Any differences are because the AZ team don't want to say anything about anything.