ShashChess

[amchess]

https://github.com/amchess/ShashChess/releases/tag/18.1
18.1
Nnue eval aligned with the classical one: it's crucial on Shashchess and I proposed
it also to the official community to avoid misleading evaluations by the gui.
There is a great boost in strength as the tests will show.

Stockfish patch
Author: SFisGOD
Date: Tue Aug 31 12:56:19 2021 +0200
Timestamp: 1630407379

Update default net to nn-735bba95dec0.nnue

[Cornfed]

I've noticed on the two most recent SFisGOD nets (in Stockfish) something odd.

I compare each with Stockfish Development version from 8/05 version in analysis (don't care about game play) for 'best moves' and more often than not, the SFisGOD nets are far slower to find and settle on a 'best' move which 8/05 has decided on earlier at less ply.

[Cornfed]

I've noticed on the two most recent SFisGOD nets (in Stockfish) something odd.

I compare each with Stockfish Development version from 8/05 version in analysis (don't care about game play) for 'best moves' and more often than not, the SFisGOD nets are far slower to find and settle on a 'best' move which 8/05 has decided on earlier at less ply.

EDIT - which reminds me - does anyone know of a board/forum where people discuss (only) Stockfish Development versions?

[peter]

https://github.com/amchess/ShashChess/releases/tag/18.1
18.1
Nnue eval aligned with the classical one: it's crucial on Shashchess and I proposed
it also to the official community to avoid misleading evaluations by the gui.
There is a great boost in strength as the tests will show.

Stockfish patch
Author: SFisGOD
Date: Tue Aug 31 12:56:19 2021 +0200
Timestamp: 1630407379

Update default net to nn-735bba95dec0.nnue

Thanks for the new version!
There are SC18 compiles together with 18.1 ones in download, which makes the anyhow big data even double. Non- embedded compiles would spare much download- time btw., but that's complaining on a really high level.

Another thing I noticed, bmi2 compile is faster with bench- command in console over avx2, which is unusual for my Ryzen 5950. So I installed both and in SMP the thing changes. Command bench will be single-threaded, I guess, can that be just inverted by multi- threading?
Thanks again regards

[Graham Banks]

I've noticed on the two most recent SFisGOD nets (in Stockfish) something odd.

I compare each with Stockfish Development version from 8/05 version in analysis (don't care about game play) for 'best moves' and more often than not, the SFisGOD nets are far slower to find and settle on a 'best' move which 8/05 has decided on earlier at less ply.

EDIT - which reminds me - does anyone know of a board/forum where people discuss (only) Stockfish Development versions?

Discord?

[peter]

Command bench will be single-threaded, I guess,

Or to ask somewhat easier, how's the console- command for setting number of threads before bench?

[amchess]

https://github.com/amchess/ShashChess/releases/tag/18.1
18.1
Nnue eval aligned with the classical one: it's crucial on Shashchess and I proposed
it also to the official community to avoid misleading evaluations by the gui.
There is a great boost in strength as the tests will show.

Stockfish patch
Author: SFisGOD
Date: Tue Aug 31 12:56:19 2021 +0200
Timestamp: 1630407379

Update default net to nn-735bba95dec0.nnue

Thanks for the new version!
There are SC18 compiles together with 18.1 ones in download, which makes the anyhow big data even double. Non- embedded compiles would spare much download- time btw., but that's complaining on a really high level.

Another thing I noticed, bmi2 compile is faster with bench- command in console over avx2, which is unusual for my Ryzen 5950. So I installed both and in SMP the thing changes. Command bench will be single-threaded, I guess, can that be just inverted by multi- threading?
Thanks again regards

Sorry. I eliminated 18 builds. Thanks!

[amchess]

Tests results:

LTC Match on positions based on characteristics
https://github.com/amchess/ShashChess/wiki/Matches

Hard positions test suite
https://github.com/amchess/ShashChess/w ... te-results

[connor_mcmonigle]

Tests results:

LTC Match on positions based on characteristics
https://github.com/amchess/ShashChess/wiki/Matches

Hard positions test suite
https://github.com/amchess/ShashChess/w ... te-results

What's the point of a 20 game test? This is a meaningless result.

[amchess]

No since it is an LTC test.
If I were to run 1000 games at long times (at least half an hour each), maybe on completely random opening positions, it would be the equivalent of at least 41 days!
There are so many patches to test and the development of a chess engine would be impossible.
Since the purpose of Shashchess is to be the best at long times, the criterion that is used, even in other branches of software engineering (testing strategy), is to use particularly significant samples. This concept is also used in statistics when we talk about "projections", for example, in the case of electoral surveys.
In our case, the chess concept of characteristic fits perfectly with this purpose: 20 games are played in one day and are based on the 10 characteristics (to simplify, center types), most common in chess. Since, if several positions share the same feature, they also share game plans, in this way, we can cover the range of possible situations without wasting too much time unnecessarily. We have explained this in the wiki.