after 37 million games i didnt see any saturation ,maybe that network is better than A0?
can anyone explain this?
Lc0 saturation.
Moderators: hgm, Rebel, chrisw
-
- Posts: 228
- Joined: Sun Mar 12, 2006 3:11 pm
Re: Lc0 saturation.
I certainly cannot comment on the question.
Practically, they would would have to run leela on the same hardware as A0 to know. Although scaling arguments (the leela ratio) have been made on the leela forum to suggest it is not there yet, but also not very far away.
IIRC, David Silver in one of his lectures indicated that A0 was still getting stronger when they stopped the A0 (Go, chess, shogi) training. So it seems A0 could have been, or could be, made even stronger. Their publically stated goal was to demonstrate the potential of their AI approach in these closed systems and that goal had been achieved, so they stopped - apparently.
Practically, they would would have to run leela on the same hardware as A0 to know. Although scaling arguments (the leela ratio) have been made on the leela forum to suggest it is not there yet, but also not very far away.
IIRC, David Silver in one of his lectures indicated that A0 was still getting stronger when they stopped the A0 (Go, chess, shogi) training. So it seems A0 could have been, or could be, made even stronger. Their publically stated goal was to demonstrate the potential of their AI approach in these closed systems and that goal had been achieved, so they stopped - apparently.
Last edited by frankp on Sat Aug 18, 2018 11:25 am, edited 1 time in total.
-
- Posts: 10311
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: Lc0 saturation.
A0 scaled better than stockfish8 based on what I read(lost against stockfish at fast time control only to win at longer time control).
Does lc0 scale better than stockfish8?
If not then it seems that lc0 and A0 do not do the same thing.
-
- Posts: 1470
- Joined: Mon Apr 23, 2018 7:54 am
Re: Lc0 saturation.
No, it was supposedly still getting better for Go, not for chess. Their figure showed it had flatlined for chess after about 12 million games. Or 21 million if you want to be cautious.frankp wrote: ↑Sat Aug 18, 2018 11:02 am IIRC, David Silver in one of his lectures indicated that A0 was still getting stronger when they stopped the A0 (Go, chess, shogi) training. So it seems A0 could have been, or could be, made even stronger. Their publically stated goal was to demonstrate the potential of their AI approach in these closed systems and that goal had been achieved, so they stopped - apparently.
-
- Posts: 165
- Joined: Tue Dec 02, 2014 1:29 am
Re: Lc0 saturation.
what i mean saturation is lower elo gain per week for example until no further improving
-
- Posts: 1080
- Joined: Fri Sep 16, 2016 6:55 pm
- Location: USA/Minnesota
- Full name: Leo Anger
-
- Posts: 165
- Joined: Tue Dec 02, 2014 1:29 am
Re: Lc0 saturation.
from testserver.lczero.org there is a lot regression, see from 10854 to 10879
http://testserver.lczero.org/matches
http://testserver.lczero.org/matches
-
- Posts: 565
- Joined: Thu Nov 13, 2014 12:03 pm
Re: Lc0 saturation.
it will eventually saturate the 20x256 NN.. might saturate at a higher ELO with a bigger net, but that's not clear
as of now saturation can be checked for with some statistical time-series test, but just from looking at it, I doubt it reached that point.
as of now saturation can be checked for with some statistical time-series test, but just from looking at it, I doubt it reached that point.
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: Lc0 saturation.
I heard bigger net was coming anyway. I don't think Lczero will reach the level of the big 3 with this size of net, but it might with the next one. Anyway, in one year, Lczero will be the best engine in the world.
-
- Posts: 1080
- Joined: Fri Sep 16, 2016 6:55 pm
- Location: USA/Minnesota
- Full name: Leo Anger