Page 4 of 4

Re: PHCEL and LC0

Posted: Sat Aug 25, 2018 7:03 pm
by Rubinus

Re: PHCEL and LC0

Posted: Thu Aug 30, 2018 12:05 pm
by Rubinus
Texel 1.08a11 - Lc0 v0.17.0-blas-id11149_12m+5s +9=2-1
games
Texel use own binary book.
Unlike the rc2 version, the new 0.17 again twice exceeded the time limit, probably in a straight line. If I considered them to be a draw, performance network 11149 on my laptop would be about 3060 ELO - level Cheng, Alfil, Craft ... A Hiarcs range - unfortunately, when it is obviously a problem with this program, maybe 10 years ago perhaps the world two , overwrite to 64 bit.
all my LC0 games

Re: PHCEL and LC0

Posted: Mon Sep 03, 2018 2:02 pm
by Rubinus
Pedone 1.7 - Lc0 v0.17.0-blas-id11149_12m+5s +16=4-0 !!!
games
Pedone use my binary book - ph-book2010-small.bin
Pedone is probably not a good opponent for LC0. As the gauntlets have shown, LC0's are clever motors - they are not so easily overpowered and tactical errors will decide. Besides, another problem is that official 0.17 is probably worse than rc2. Another five times over, one in the winning position of a pair of draws to mat, twice in a straight line. Besides, the weird distribution of time, a few more games are scuffled with a bonus game. The match had two parts. First half (10: 0!) LC0 uci options: NN backend to use = 1 (blas, opencl crashed), Minibatch size for NN inference = 16. In the second half (2: 8) I still reduced Scale thinking time = 2.400000 to 1.600000. Maybe too, my estimate for the next match is 1.8-1.9. Until version 0.10, I did not have the LC0 problems, I have not seen one time. Now it seems that the compilation is created by somebody who understands the speed but not the stability and the time consuming ... I really do not want to participate in the learning of the network, it goes down in total amateur. And we have another restart ...
all my LC0 games

Re: PHCEL and LC0

Posted: Mon Sep 03, 2018 3:07 pm
by jdart
I think you really have to run it on a GPU. Pedone would not fare so well then.

I only looked at the games a little bit. Here is one position (round 8) where LC0 played a losing move in the endgame (Kf5). Stockfish thinks Rd6 draws:



--Jon

Re: PHCEL and LC0

Posted: Mon Sep 03, 2018 5:35 pm
by Rubinus
opecl will start creating error messages after a while. Poor compilation, I did not have a problem with version 0.10 (gpu).
In the given position, he was white at the hour 0: 00: 07.1 - too aggressive use of time, so I reduced the think thinking time to the second half by one third.