LCZero: Progress and Scaling. Relation to CCRL Elo

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

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Albert Silver
Posts: 2867
Joined: Wed Mar 08, 2006 8:57 pm
Location: Rio de Janeiro, Brazil

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by Albert Silver » Thu May 17, 2018 5:44 pm

mhull wrote:
Thu May 17, 2018 5:37 pm
Laskos wrote:
Thu May 17, 2018 4:14 pm
Yes, I use a very short 3-mover book of solid, balanced GM openings. I think LC0 must play well from all those short, solid lines, so this procedure is probably more sound than using no book at all in matches for determining strength (not in learning).
Would it be more sound to measure human strength that way too?
It would, though humans are able to prepare in advance openings to try and get an edge, which is the basis of modern play. I remember a master class with Valery Salov, given here in Brazil, who complained that Gelfand owed at least 100 Elo of his rating to his amazingly good opening preparation.

Nowadays a number of players try to sidestep all this by deliberately entering 'harmless' openings that also sidestep all theory to just have a generic playable position in which it is just their chess skill against the opponent's. The current World Champion is notorious for this.
Last edited by Albert Silver on Thu May 17, 2018 5:48 pm, edited 1 time in total.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."

Milos
Posts: 3387
Joined: Wed Nov 25, 2009 12:47 am

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by Milos » Thu May 17, 2018 5:47 pm

mhull wrote:
Thu May 17, 2018 5:37 pm
Laskos wrote:
Thu May 17, 2018 4:14 pm
Yes, I use a very short 3-mover book of solid, balanced GM openings. I think LC0 must play well from all those short, solid lines, so this procedure is probably more sound than using no book at all in matches for determining strength (not in learning).
Would it be more sound to measure human strength that way too?
Ofc it would but humans are too stupid and too proud to accept anything like that.
Plus LC0 is way too stupid to play anything but pre deterministic moves so in reality LC0 is weaker than any rated human free of Alzheimer's.

mar
Posts: 2010
Joined: Fri Nov 26, 2010 1:00 pm
Location: Czech Republic
Full name: Martin Sedlak

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by mar » Thu May 17, 2018 6:21 pm

mar wrote:
Thu May 17, 2018 6:20 am
I plan to play 200 games to get a rough idea of how strong the current engine + net is, I'll post the results here.
I have to agree with Albert - after 100 games perf is -60 elo compared to ID 24x so fanboys can cry now, they should revert to 0.7 ASAP!
Martin Sedlak

jkiliani
Posts: 143
Joined: Wed Jan 17, 2018 12:26 pm

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by jkiliani » Thu May 17, 2018 7:27 pm

mar wrote:
Thu May 17, 2018 6:21 pm
mar wrote:
Thu May 17, 2018 6:20 am
I plan to play 200 games to get a rough idea of how strong the current engine + net is, I'll post the results here.
I have to agree with Albert - after 100 games perf is -60 elo compared to ID 24x so fanboys can cry now, they should revert to 0.7 ASAP!
Reverting to an earlier version just because something went wrong is not smart. The correct procedure is to figure out what exactly changed, and only react to (and possibly revert) that. Rather high on the list of suspects is the change to PUCT, from 0.85 to 0.6, and (temporarily) the introduction of FPU reduction. Since there are also issues with the 50-move plane and training on individual positions happens too often, we don't know for sure if PUCT is really responsible, and will investigate it along with other possible causes for the regression.

yanquis1972
Posts: 1762
Joined: Tue Jun 02, 2009 10:14 pm

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by yanquis1972 » Thu May 17, 2018 7:33 pm

if you run an old network vs current match, is it important/necessary to run the old network on the contemporaneous client?

i ran 260 games (meant to get to 300) LC0-CUDA vs rybka 3 1-cpu @ 1/s, close to default (maybe less than optimal) settings & got something like -50 elo, which pretty well shocked me. i've heard people say 3000 elo (CCRL) a few times but the play is sporadic it was hard to believe.

mar
Posts: 2010
Joined: Fri Nov 26, 2010 1:00 pm
Location: Czech Republic
Full name: Martin Sedlak

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by mar » Thu May 17, 2018 7:36 pm

jkiliani wrote:
Thu May 17, 2018 7:27 pm
Reverting to an earlier version just because something went wrong is not smart.
You're actually right - letting random people commit random stuff and changing random constants without testing is not smat but rather incompetent.
Problems have been pointed out by Ronald, Alvaro and other people who actually wrote chess engines and who know the hell what they're talking about,
but you fanboys keep repeating the same nonsense over and over again, well good luck with that. You'll bury a very intersting project.
Martin Sedlak

jkiliani
Posts: 143
Joined: Wed Jan 17, 2018 12:26 pm

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by jkiliani » Thu May 17, 2018 7:51 pm

mar wrote:
Thu May 17, 2018 7:36 pm
jkiliani wrote:
Thu May 17, 2018 7:27 pm
Reverting to an earlier version just because something went wrong is not smart.
You're actually right - letting random people commit random stuff and changing random constants without testing is not smat but rather incompetent.
Problems have been pointed out by Ronald, Alvaro and other people who actually wrote chess engines and who know the hell what they're talking about,
but you fanboys keep repeating the same nonsense over and over again, well good luck with that. You'll bury a very intersting project.
If you have anything constructive to contribute, feel free to do so on Github. Or fork the project and run your own version. There has never been "random people committing random stuff" here, ever. Code has always been tested, by the more experienced devs on the project. But as you should probably know if you ever developed anything yourself, overlooking something can always happen. In this particular instance, lowering PUCT raised the self-play strength, consistently across networks, and by a considerable margin as well, so there was nothing inconsistent with merging it. It's a possibility that it makes tactics unknown to the policy harder to find though, but that is difficult to actually prove conclusively.

You may have loads of chess programming experience, but unless you desist from name-calling and contribute in a positive way, neither I nor anyone else on this project will have the slightest interest in what you have to say.

yanquis1972
Posts: 1762
Joined: Tue Jun 02, 2009 10:14 pm

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by yanquis1972 » Thu May 17, 2018 8:40 pm

wow. tried 223 (cuda) & it's incredibly aggressive & smooth on the attack. somehow this got all fragmented in the newer builds. rybka 3 just immediately looks amateurish @1s/move.

JJJ
Posts: 1287
Joined: Sat Apr 19, 2014 11:47 am

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by JJJ » Thu May 17, 2018 9:48 pm

mar wrote:
Thu May 17, 2018 8:04 pm
jkiliani wrote:
Thu May 17, 2018 7:51 pm
If you have anything constructive to contribute, feel free to do so on Github.
Your contribution is load of crap in this thread.
Go write hello world then come back and lecture me.
another one to the foe list.

David Xu
Posts: 47
Joined: Mon Oct 31, 2016 8:45 pm

Re: LCZero: Progress and Scaling. Relation to CCRL Elo

Post by David Xu » Fri May 18, 2018 1:02 am

mar wrote:
Thu May 17, 2018 8:04 pm
jkiliani wrote:
Thu May 17, 2018 7:51 pm
If you have anything constructive to contribute, feel free to do so on Github.
Your contribution is load of crap in this thread.
Go write hello world then come back and lecture me.
This is a great way to make sure that what you have to say is 100% ignored by the devs. Of course, you may not actually care about that, since you seem more interested in inflating your own ego rather than contributing.

Post Reply