Some people claim that version 0604-experimental plays stronger than 0619 on the same net. That may theoretically happen both because algorithm differences (I changed some formulas to be able to save some memory and floats may be rounded differently sometimes) or maybe there is nps difference.
I wasn't able to find any differences in my tests though.
Maybe someone has time to do tests and confirm/refute that claim? That would be much appreciated!
LC0 finally?
Moderator: Ras
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: LC0 finally?
Leela seems close to make progress again.
-
- Posts: 1983
- Joined: Thu Sep 18, 2008 10:24 pm
Re: LC0 finally?
Hasn’t the current network saturated? Why haven’t we moved to a larger net?
-
- Posts: 1766
- Joined: Wed Jun 03, 2009 12:14 am
Re: LC0 finally?
am wondering the same thing, if it's still a test-run they may want to be absolutely sure of the limit.
-
- Posts: 1983
- Joined: Thu Sep 18, 2008 10:24 pm
Re: LC0 finally?
I did actually wonder if the reason was another restart was due, but thought this time it would be different.yanquis1972 wrote: ↑Fri Jun 22, 2018 8:39 pm am wondering the same thing, if it's still a test-run they may want to be absolutely sure of the limit.
Ho hum. Maybe it’s called LC0 because it’s Elo keeps going back to zero

-
- Posts: 7286
- Joined: Thu Aug 18, 2011 12:04 pm
- Full name: Ed Schröder
Re: LC0 finally?
ElswhereWerewolf wrote: ↑Fri Jun 22, 2018 10:58 pmI did actually wonder if the reason was another restart was due, but thought this time it would be different.yanquis1972 wrote: ↑Fri Jun 22, 2018 8:39 pm am wondering the same thing, if it's still a test-run they may want to be absolutely sure of the limit.
Ho hum. Maybe it’s called LC0 because it’s Elo keeps going back to zero![]()
Id396 is a new net without input from broken games:
Net has been replaced with a new bootstrap net with 50 move rule normalization and learnable batch normalization beta bias. It will likely be weaker to start but will hopefully grow stronger quickly. Thanks to scs-ben!
scs-ben - 06/02/2018
192x15 network is being bootstrapped. Batch size is 512. Using games 7 million to current but excluding 10.5-14 million due to likely bad data. Batch normalization, 50 move rule plane normalization, test/train data separation are all included.
90% of coding is debugging, the other 10% is writing bugs.
-
- Posts: 1983
- Joined: Thu Sep 18, 2008 10:24 pm
Re: LC0 finally?
It’s good they keep trying things, but this net doesn’t look that good, unless the numbers are skewed. Perhaps another restart?
-
- Posts: 7286
- Joined: Thu Aug 18, 2011 12:04 pm
- Full name: Ed Schröder
Re: LC0 finally?
Well, there was a bug, hence the restart.
As said -
Patience, patience and hoping for no new bugs
As said -
This seems to be exactly the case.It will likely be weaker to start but will hopefully grow stronger quickly. Thanks to scs-ben!
Patience, patience and hoping for no new bugs
90% of coding is debugging, the other 10% is writing bugs.
-
- Posts: 10770
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: LC0 finally?
what about having correct rating graph by calculating the rating in the graph not based on self play but based on games against known opponents?
-
- Posts: 1766
- Joined: Wed Jun 03, 2009 12:14 am
Re: LC0 finally?
that's going on in the background & not-so-in-the-background...
http://lczero.org/?full_elo=1 click elo estimates
http://lczero.org/?full_elo=1 click elo estimates