My test for Minic 2.50 NNUE-Nascent Nutrient

[Sylwy]

Test conditions:
-TC=4'+2"
-Hash=256 MB
-GUI: Arena 3.5.1
-Book: Perfect_2010.abk for both engines
-default settings for both engines
-1 thread-CPU=Intel i5-7400-3GHz (Kaby Lake)
-OS: Windows 10 Home
-6-men Syzygy bases for both engines.




The previous test:


Texel 1.07 x64 has .......................3114 Elo points-CCRL Blitz
Minic 2.50 NNUE-Nascent Nutrient...around 3355 Elo points (3114+241)
Minic 2.50 NNUE-SV1520.............around 3458 (3451-Komodo 14 x64 +7)

Nascent Nutrient is about 100 Elo points below the best Segio Vieri nets.
A strong chess entity and an original play ! Congratulations !

[xr_a_y]

Thanks a lot for testing.

Did you really see some "personalty" in this net ? Does it recall Minic playing style (if such thing exists ...) ? It is different from SV net playing style to your eyes ?

[Sylwy]

It is different from SV net playing style to your eyes ?

Yes, it's definitely something different ! Almost all my NNUE tests used various versions of SV net. In Igel and Minic the difference is obvious. A very aggressive Minic.
I understand the frustration of the masters of handcrafted evaluation. A lot of years. A lot of work. But the finality is chess and here the progress is great......and inevitable. Someday the end of the new NNUE age will come. That's life !

[xr_a_y]

This is interesting. A fear I have is that in time all nets may more or less converge to the same with LR.
This doesn't seem to be the case yet, but Nascent Nutrient hasn't be RL yet.
I'm glad this one shows some aggressiveness.

[carldaman]

NNUE may converge only if the goal is strictly Elo gains. As long as it's possible to play differently, with various degrees of risk aversion, they don't have to converge.