LC0 finally?

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

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Werewolf
Posts: 956
Joined: Thu Sep 18, 2008 8:24 pm

LC0 finally?

Post by Werewolf » Wed Jun 20, 2018 8:35 pm

This looks really promising with the latest networks, is it as good as it looks?

User avatar
Laskos
Posts: 8054
Joined: Wed Jul 26, 2006 8:21 pm

Re: LC0 finally?

Post by Laskos » Wed Jun 20, 2018 10:42 pm

Werewolf wrote:
Wed Jun 20, 2018 8:35 pm
This looks really promising with the latest networks, is it as good as it looks?
No. It's not bad generally for 11th net (ID4011), but the rating is misleading. ID4011 is some 500 Elo points weaker than the best nets of the run 1 and run 5 against AB engines, but shows some 400 Elo points better rating in self-games than the best nets of those runs, which is misleading. The rating of the latest run are very inflated, but the real progress against AB engines is still not bad for only 11th net.

Werewolf
Posts: 956
Joined: Thu Sep 18, 2008 8:24 pm

Re: LC0 finally?

Post by Werewolf » Thu Jun 21, 2018 7:58 am

4018 reads at 3400 elo, so is that about 2900-3000 in real terms?

Not bad....

User avatar
Laskos
Posts: 8054
Joined: Wed Jul 26, 2006 8:21 pm

Re: LC0 finally?

Post by Laskos » Thu Jun 21, 2018 9:44 am

Werewolf wrote:
Thu Jun 21, 2018 7:58 am
4018 reads at 3400 elo, so is that about 2900-3000 in real terms?

Not bad....
About 2900 in CCRL 40/4' conditions on GTX 1060 GPU. ID4018 is about the level of ID1028 of the previous run (or shown 2500 Elo level of that run), which means a significantly faster progress (18 nets compared to 28 nets, each needing about 40,000 games). The shown rating in self-games now is very inflated, for obscure to me reasons.

crem
Posts: 16
Joined: Wed May 23, 2018 7:29 pm

Re: LC0 finally?

Post by crem » Thu Jun 21, 2018 10:20 am

Laskos wrote:
Thu Jun 21, 2018 9:44 am
About 2900 in CCRL 40/4' conditions on GTX 1060 GPU. ID4018 is about the level of ID1028 of the previous run (or shown 2500 Elo level of that run), which means a significantly faster progress (18 nets compared to 28 nets, each needing about 40,000 games). The shown rating in self-games now is very inflated, for obscure to me reasons.
I think it's just id4000 vs id4001 match which gave very low precision Elo delta with just 1 lost game (result was +423 -1 =0).
If it was 2 lost games instead of 1, it would be 400 points lower. And if it was 0 lost games instead of 1, it would be +inf points higher.

yanquis1972
Posts: 1648
Joined: Tue Jun 02, 2009 10:14 pm

Re: LC0 finally?

Post by yanquis1972 » Thu Jun 21, 2018 4:25 pm

is the testserver about ready to switch to a larger net then? think on the first one it was quite quickly; i suppose the thinking is too quickly.

User avatar
Rebel
Posts: 4156
Joined: Thu Aug 18, 2011 10:04 am

Re: LC0 finally?

Post by Rebel » Thu Jun 21, 2018 6:17 pm

Update -> http://rebel13.nl/zero.html

After the w396 bootstrap lczero (w429) is climbing and climbing again.

Note that the NPS is 33 only.

SzG
Posts: 2463
Joined: Fri Mar 10, 2006 6:20 am
Location: Szentendre, Hungary
Full name: Gabor Szots

Re: LC0 finally?

Post by SzG » Fri Jun 22, 2018 6:57 am

Does the graph that I see on the main page indicate that whatever you do this self-learning will saturate at some point so no further progress is to be expected?
Gabor Szots

CCRL testing group

User avatar
Laskos
Posts: 8054
Joined: Wed Jul 26, 2006 8:21 pm

Re: LC0 finally?

Post by Laskos » Fri Jun 22, 2018 7:07 am

SzG wrote:
Fri Jun 22, 2018 6:57 am
Does the graph that I see on the main page indicate that whatever you do this self-learning will saturate at some point so no further progress is to be expected?
By ID4039, in my gauntlet at fast TC against AB engines, it's already similar in strength to best 6x64 nets of the previous runs, scoring 61.0/200. Probably close to saturation, but let's see (if they won't start from scratch again).

crem
Posts: 16
Joined: Wed May 23, 2018 7:29 pm

Re: LC0 finally?

Post by crem » Fri Jun 22, 2018 8:11 am

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!

Post Reply