LCZero: Progress and Scaling. Relation to CCRL Elo

[Joost Buijs]

You are aware that the cuDNN versions cannot run on just CPUs? They also have different code and behavior than the normal LCZero. It is not 'just faster'.

I have the CPU- version only. My graphic card wouldn't be bad neither, but it's AMD und doesn't support OpenCl.

When you have an AMD graphic card and it is not outdated it supports OpenCL, only the cuDNN version is nVidia specific.

[Albert Silver]

You are aware that the cuDNN versions cannot run on just CPUs? They also have different code and behavior than the normal LCZero. It is not 'just faster'.

I have the CPU- version only. My graphic card wouldn't be bad neither, but it's AMD und doesn't support OpenCl.

I don't think you understand: my comment on using PUCT=3.0 with good results only applies to the cuDNN version.

[Ponti]

Can LCZero learn tactics by playing against other "tactical and speculative" engines ( like Junior, Chess Tiger or OpenTal ) ?

[Uri Blass]

I see no reason to play other engines to learn tactics.
I am sure LCzero missed a lot of tactics in the games against itself.

[peter]

I don't think you understand: my comment on using PUCT=3.0 with good results only applies to the cuDNN version.

I think I understood, but I don't know, what good results you get at WAC, about which I was and am talking, do you understand?

[Albert Silver]

I don't think you understand: my comment on using PUCT=3.0 with good results only applies to the cuDNN version.

I think I understood, but I don't know, what good results you get at WAC, about which I was and am talking, do you understand?

I can test it, but I need a PGN of WAC, or for you to tell me what interface I can process it in. Fritz doesn't take EPD, only PGN, and I found no option in Arena.

[AdminX]

I don't think you understand: my comment on using PUCT=3.0 with good results only applies to the cuDNN version.

I think I understood, but I don't know, what good results you get at WAC, about which I was and am talking, do you understand?

I can test it, but I need a PGN of WAC, or for you to tell me what interface I can process it in. Fritz doesn't take EPD, only PGN, and I found no option in Arena.

I think Shredder might work.

[majortom]

but I need a PGN of WAC

I think it's easy to make something like this from these 300 positions, but I didn't see the PGN-files, capable for right move check and don't know what PGN tags should be used for this.

[Albert Silver]

but I need a PGN of WAC

I think it's easy to make something like this from these 300 positions, but I didn't see the PGN-files, capable for right move check and don't know what PGN tags should be used for this.

No worries, I found out how to do it in Arena:

https://sites.google.com/site/strategic ... pd-testing

Am testing now.

[Albert Silver]

I don't think you understand: my comment on using PUCT=3.0 with good results only applies to the cuDNN version.

I think I understood, but I don't know, what good results you get at WAC, about which I was and am talking, do you understand?

Ok, so I tested it with 10 seconds per move, which is what i understood you do, and got:

Analysis ended normally 164 of 300 matching moves Rated time: 24:30

However, I had never seen the suite before and seeing it now I think it has a lot of problematic positions that make this a very poor test to use. Not only do many have multiple solutions but many don't even need a solution, and will win in a variety of continuations. Just the first two positions are problematic. Look at position no.1:



Even this basic mate in two, if missed still wins easily. Leela's choice of 1.Nh5 may not be mate in two, but ask Komodo and it says:

1...Qxe5 2.dxe5 Rg8 3.Nf4 d4 4.Be4 Rce8 5.Ng6+ Kh7 6.Ne7+ Kh8 7.Nxg8 Kxg8 8.Bxc6 bxc6
White is clearly winning: +- (8.47 --) Depth: 29 00:00:23 130MN, tb=602

Then we have position no.2:



The given answer is 1...Rxb2, which Leela did not find and chose 1...c3! with a winning evaluation. Is she wrong? I tested Leela's choice of 1...c3 and fed it to Komodo 11.3, and got this:

1...c3 2.bxc3 Rxc3 3.e4 fxe4+ 4.Kxe4 Rxa3 5.Ke3 Kf5 6.Rd1 h5 7.h4 d2+ 8.Kxd2 Rh3 9.Kc2 Rh2+ 10.Kc3 Kxf4
Black is clearly winning: -+ (-4.40) Depth: 37 00:02:42 1276MN, tb=12718902

Don't get me wrong, I am not suggesting Leela is any stronger in tactics because of this, but this test does little to prove anything in my eyes. At the very least it would need some serious house-cleaning to remove all the positions with multiple solutions.