Page 1 of 1
Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 3:36 am
by tturgut
1) Correspondence chess: Does anyone have any idea what settings are best for Leela for long analysis?
would it be reliable in cc?
right now, I do not see any evaluation change regardless if it is 2-3 minutes or 2 hour analysis.. this should be related to the parameters I guess. (if it is an evolving tree, it should change with longer analysis)
2) I saw increasing puct number would be better for finding tactics.?Does anyone have anyexperience?
3) Endings: Syzygy does not work well with this engine for some reason. (it works, but...)
4) openings: alpha zero recommended English. Leela likes french, KI. any other opening recommendations? (it looks like locked centers mainly?)
Re: Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 4:55 am
by dkappe
I wasn’t too excited about lc0 multipv. It’s very easy for a single move to dominate the visits.
I added a random root mod to my bender fork of lc0. After an initial 4000 iterations of search, it explores random root moves 1/16th of the time, then random root moves proportional to Q 3/16ths of the time. Set the cpuct to 5, and you have much better multipv.
https://github.com/dkappe/lc0/tree/bender
https://github.com/dkappe/leela-chess-w ... iki/Bender
Re: Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 4:58 am
by dkappe
Sorry, the point being that it yields better, more useful results for analysis (and it may play a bit better than vanilla lc0).
Re: Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 5:01 am
by Nay Lin Tun
https://www.chess.com/forum/view/genera ... e-on-leela
For unlimited search/ visit "-v 10000000000000" in engine parameter.
Re: Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 7:04 am
by tturgut
Thanks. My computer /software skills are limited. (I was able to set it up, but that's it)
Can you clarify a little more?
1) "For unlimited search/ visit "-v 10000000000000" in engine parameter."
exactly which engine parameter? (I am using CB13)
2) "Set the cpuct to 5, and you have much better multipv."
I guess this increases the tree width, (catching more tactical shots), but decreases the depth. is this correct?
Then, would something higher than 5 be better?
3) weights: shall i just leave it as it is? (autodiscover)
I put 11248 in the same folder as lc0.
4) syzygy: I wrote
-s c:\TB
is this correct?
5) engine number: doesn't this matter at all? 2,4,6.... anything is OK?
thx!
Re: Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 9:06 am
by dkappe
Cpuct to 5 with the random root mod.
Re: Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 9:22 am
by George Tsavdaris
Unless he uses a GUI to analyze, as most probably does, so this is not needed as he just presses ANALYZE in the GUI and let it run as long as he wants.
Re: Leela: Would it be reliable in correspondence chess?
Posted: Wed Oct 10, 2018 9:38 am
by crem
George Tsavdaris wrote: ↑Wed Oct 10, 2018 9:22 am
Unless he uses a GUI to analyze, as most probably does, so this is not needed as he just presses ANALYZE in the GUI and let it run as long as he wants.
There's actually no "-v" parameter. It was in LCZero (v0.10), but in Lc0 (v0.15 and further) it's not there.
Re: Leela: Would it be reliable in correspondence chess?
Posted: Thu Oct 11, 2018 12:41 pm
by tturgut
http://talkchess.com/forum3/viewtopic.php?f=2&t=68623
I wasn't sure if i should add this discussion at this topic . 2 problems:
1) crashing, better now with the new driver.
2) 2070 may be enough. (2080 does not change evaluations after a minute or so) so 2070 may just be good enough if it runs just a little longer.