LC0 output

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.
Post Reply
cma6
Posts: 145
Joined: Thu May 29, 2014 3:58 pm

LC0 output

Post by cma6 » Tue Dec 25, 2018 9:06 pm

I have discovered the Lc0 online chess engine. I was analyzing a game of mine after 6 moves. As a newbie, I'm not sure if I understand the output:

info depth 7 seldepth 14 time 167638 nodes 32171 score cp 13 hashfull 72 nps 191 tbhits 0 pv e1g1 e8g8 b1c3 b8c6 c1f4 c8f5 d1b3 d8d7 f3e5 c6e5 d4e5

What are "depth", "selfdepth", "hashfull"?

Can I assume that Lc0 thinks W has an edge of .13?

How do the node count and nps numbers compare with the same terms in a typical UI like Aquarium with SF?

What kind of depth/selfdepth numbers would be comparable with depth=40 ply shown in an A/B engine UI?

This is the url for Lc0 for the web: https://frpays.github.io/lc0-js/engine.html
To do the calculations, is that engine using my local system (great CPU, very poor GPU) or using its own GPU in the cloud?

Thanks to the Lc0 experts and Merry Christmas!

Nay Lin Tun
Posts: 546
Joined: Mon Jan 16, 2012 5:34 am

Re: LC0 output

Post by Nay Lin Tun » Wed Dec 26, 2018 1:03 am

It is 6x64 net( very small net). It use your local gpu. You can see nps in analysis. For rough strength comparision, 1 node of 20x256 net is approx equal to vs 1000 nodes of SF. But for this smaller net, i have no idea, may be 1 node = 100 node of SF. My personal estimated strength , somewhere between 2000-2400( depends on your local gpu)

For further detail, ask the programmer in leela discord.
Attachments
Capture 3.PNG
Capture 3.PNG (94.97 KiB) Viewed 1590 times

cma6
Posts: 145
Joined: Thu May 29, 2014 3:58 pm

Re: LC0 output

Post by cma6 » Wed Dec 26, 2018 1:29 am

Nay Lin,
Thank you for the clarification. Therefore, to run Lc0 at the highest level, should one plan on using one's own system with a very fast gpu, e.g., the online version is meant mostly for illustration purposes?

Redshift
Posts: 4
Joined: Wed Jan 31, 2018 2:51 pm

Re: LC0 output

Post by Redshift » Wed Dec 26, 2018 3:57 pm

Correct

cma6
Posts: 145
Joined: Thu May 29, 2014 3:58 pm

LC0 output

Post by cma6 » Wed Dec 26, 2018 8:48 pm

Redshift:
Then if one is using one's own local system and fast gpu, assuming one has the latest weights file, what role do the networks play in doing lc0 analysis on one's own system?
(I am not referring to a processing contribution to the ongoing Lc0 project.)

Werewolf
Posts: 1199
Joined: Thu Sep 18, 2008 8:24 pm

Re: LC0 output

Post by Werewolf » Thu Dec 27, 2018 8:40 am

The network is basically the evaluation function

edit: I meant the weights file is the EF - sorry

Sven
Posts: 3830
Joined: Thu May 15, 2008 7:57 pm
Location: Berlin, Germany
Full name: Sven Schüle
Contact:

Re: LC0 output

Post by Sven » Thu Dec 27, 2018 12:31 pm

cma6 wrote:
Tue Dec 25, 2018 9:06 pm
I have discovered the Lc0 online chess engine. I was analyzing a game of mine after 6 moves. As a newbie, I'm not sure if I understand the output:

info depth 7 seldepth 14 time 167638 nodes 32171 score cp 13 hashfull 72 nps 191 tbhits 0 pv e1g1 e8g8 b1c3 b8c6 c1f4 c8f5 d1b3 d8d7 f3e5 c6e5 d4e5

What are "depth", "selfdepth", "hashfull"?

Can I assume that Lc0 thinks W has an edge of .13?

How do the node count and nps numbers compare with the same terms in a typical UI like Aquarium with SF?

What kind of depth/selfdepth numbers would be comparable with depth=40 ply shown in an A/B engine UI?

This is the url for Lc0 for the web: https://frpays.github.io/lc0-js/engine.html
To do the calculations, is that engine using my local system (great CPU, very poor GPU) or using its own GPU in the cloud?

Thanks to the Lc0 experts and Merry Christmas!
I am not an Lc0 expert but most of your questions are easy to resolve for me since they are hardly related to Lc0 itself.

"depth" is the nominal search depth that was used when starting the part of the tree search ("iteration") that resulted in the given output line.

"seldepth" means "selective depth" and is the maximum search depth (including all kinds of search extensions, like searching one ply deeper when replying to check) that was reached during the current iteration.

"hashfull" indicates how much of the main transposition hash table was used at that time, given in percent.

As you assumed, "score cp 13" means that the engine evaluates the leaf node of the principal variation as +13 centipawns (i.e. +0.13) from the viewpoint of the moving side (here White).

Node counts and nps of NN-based chess engines are much lower than those of traditional engines of comparable strength. This is because NN access is very slow but the information retrieved from a very strong (i.e. well-trained) NN somehow replaces what a traditional engine like SF needs to find through deep search.

That the JS-based "LCZero for the web" engine uses your local GPU has already been written, you can simply check it with your Task Manager (Windows) or a program like top or htop (Linux).
Sven Schüle (engine author: Jumbo, KnockOut, Surprise)

cma6
Posts: 145
Joined: Thu May 29, 2014 3:58 pm

Re: LC0 output

Post by cma6 » Fri Dec 28, 2018 2:25 am

Sven:
Thanks for the great primer for an Lc0 newbie.

dkappe
Posts: 294
Joined: Tue Aug 21, 2018 5:52 pm
Full name: Dietrich Kappe

Re: LC0 output

Post by dkappe » Sat Dec 29, 2018 3:16 am

Under the parameters tab, you can select several other nets, including 11248.

Post Reply