Minic version 3

[Guenther]

Yes Guenther,
I sent you an email - hope the adress is working.

I just downloaded the attachment of your email, thanks!
At A first quick glance it looks like the eval for Minic in those games looks quite instable?

BTW I also noticed after downloading a compilation for my hardware that it seems nnue is disabled
in the uci options by default - did you enable it in the test with Clover?

Code: Select all

info string This is Minic version 3.16 (NNUE available)
uci
id name Minic 3.16
id author Vivien Clauzon
...
option name NNUEFile type string default embedded
option name forceNNUE type check default false
option name NNUEScaling type spin default 80 min 32 max 256
option name NNUEThreshold type spin default 600 min -1500 max 1500
...

[Werner]

option name forceNNUE type check default false

is correct -
if you enable this, there is always NNUE even in endgames. And this weakens the engine.
I use the same setting now in the other compile with good results:
Minic 3.16 x64 1CPU - Clover 2.4 x64 1CPU : 37,0/69 18-13-38 54% +28

[Guenther]

option name forceNNUE type check default false

is correct -
if you enable this, there is always NNUE even in endgames. And this weakens the engine.
I use the same setting now in the other compile with good results:
Minic 3.16 x64 1CPU - Clover 2.4 x64 1CPU : 37,0/69 18-13-38 54% +28

Ok. So it seems there might have been a conversion problem for avx512 and it read
the wrong values from nnue perhaps? May be Vivien will enlighten us later.

[xr_a_y]

option name forceNNUE type check default false

is correct -
if you enable this, there is always NNUE even in endgames. And this weakens the engine.
I use the same setting now in the other compile with good results:
Minic 3.16 x64 1CPU - Clover 2.4 x64 1CPU : 37,0/69 18-13-38 54% +28

Ok. So it seems there might have been a conversion problem for avx512 and it read
the wrong values from nnue perhaps? May be Vivien will enlighten us later.

I'd love to "enlight" you, but I'm still working on this I'll publish as soon as I find the clue

In the meantime, please stick to the avx2 build.

Indeed "forceNNUE" will use "pure" NNUE eval, by default this is not the case and "hybrid" is used, this means that HCE is used when material score is too big.

[xr_a_y]

Minic 3.17 is available for downlad on github (https://github.com/tryingsomestuff/Mini ... ed2/Minic3). Same net as previous version but some search tweaks giving around 20Elo at short TC.
I really don't know it holds at longer TC.

[matejst]

Merci, Vivien ! J'aime beaucoup les différentes possibilités de configuration du moteur. Pourtant, le moteur reste de temps en temps non-responsif sur mon ordinateur, quand je l'arrête après une période d'analyse. Ce n'est pas grave (je perds quelques secondes).

[xr_a_y]

Merci, Vivien ! J'aime beaucoup les différentes possibilités de configuration du moteur. Pourtant, le moteur reste de temps en temps non-responsif sur mon ordinateur, quand je l'arrête après une période d'analyse. Ce n'est pas grave (je perds quelques secondes).

Merci ! C'est peut-être un souci que je peux regarder rapidement. Cela se produit en UCI ou en CECP ? et avec quel GUI que j'essai de reproduire la situation ?

[matejst]

Merci, Vivien ! Ça se produit sous Shredder GUI, UCI. J'ai essayé les moteurs minic-nehalem et minic-bdver1 qui fonctionnent sous mon système. Le moteur met plusieurs secondes à réagir à la fin d'une analyse.

[xr_a_y]

Waiting for Minic 3.18 ...

Here are two new unofficial nets just to have fun with

- Nifty Nebulosity V2 for Minic >=3.06 (11/2021): just another training session using the T74 LC0 data. It is not an "official" net. It plays 80Elo above "nucleated neurulation".
- Nifty Nebulosity V3 for Minic >=3.06 (12/2021): a training session using the T60T70wIsRight LC0+SF data. It is not an "official" net. It plays 110Elo above "nucleated neurulation".

Get them here : https://github.com/tryingsomestuff/NNUE-Nets

[Graham Banks]

Minic 3.17 1CPU and 4CPU gauntlets are running for the CCRL 40/15 list.