rating adjustments of the NN testing

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

Hugo
Posts: 782
Joined: Tue Dec 01, 2009 11:10 am

rating adjustments of the NN testing

Post by Hugo »

Hi

does anyone know how it comes that one day the rating of tested 40 networks is arround 3200 +,
and a few hours later all versions are adjusted to 50 ELO less?
yesterday we had
41752 = 3172
41762 = 3183
41765 = 3192
41771 = 3200
41773 = 3206

now all of them have ~ 3140

C.K.
Hugo
Posts: 782
Joined: Tue Dec 01, 2009 11:10 am

Re: rating adjustments of the NN testing

Post by Hugo »

Hi

Again a rating adjustment on the Lc0 network site.
Almost all 40 NN have been reduced by - 20 ELO.
For example NN 41812 started with 3165 after > 31.000 games.
Then over a week it was by 3159.
Now it has been reduced to 3140.

seems nobody knows here why this is. I already asked in other forums but also nothing concrete yet.

C.K.
User avatar
Guenther
Posts: 4610
Joined: Wed Oct 01, 2008 6:33 am
Location: Regensburg, Germany
Full name: Guenther Simon

Re: rating adjustments of the NN testing

Post by Guenther »

Hugo wrote: Sun Mar 31, 2019 6:56 pm Hi

does anyone know how it comes that one day the rating of tested 40 networks is arround 3200 +,
and a few hours later all versions are adjusted to 50 ELO less?
yesterday we had
41752 = 3172
41762 = 3183
41765 = 3192
41771 = 3200
41773 = 3206

now all of them have ~ 3140

C.K.
Since a while (a few months meanwhile IIRC) they do regularily recalculations by adding match results from current nets vs. various
older nets (exact procedure for selecting the net number is surely written somewhere...) of the same NN. That's it.

Just check the matches site and you will see that not only N+1 vs. N nets are done since a certain time.
http://lczero.org/matches/

Edit:
Checked it precisely now. Since 2019-01-29 they added regularily cross matches between various older nets against the latest one.
(note that they did this also in the past, but just a very few times and randomly, probably just to check, if it is still progressing, or something went wrong)
N+1 vs. N matches have a different colour than N vs. N-x matches in the site quoted above.
https://rwbc-chess.de

trollwatch:
Talkchess nowadays is a joke - it is full of trolls/idiots/people stuck in the pleistocene > 80% of the posts fall into this category...
Hugo
Posts: 782
Joined: Tue Dec 01, 2009 11:10 am

Re: rating adjustments of the NN testing

Post by Hugo »

Thanks a lot for this explanation, Guenther !

kind regards, Clemens