Minic version 3

[Madeleine Birchfield]

I think many things are involved here, probably too much ! :
- data quality : I tried many different things (from games, from random play, from inside search tree), from depth 5 to 16, from 200M to 2B pack, only quiet position, only non capture next move, ...
- evaluation quality : clearly SF scored data looks better than Minic ones so probably RL might be a good idea !
- NNUE speed : for now the Seer style evaluation is twice as slow as the current SF one, partly because of float32 being used instead of int16. So a lot to be gained from quantization probably.

I wonder if improving Minic's handcrafted evaluation function would improve the evaluation quality of Minic scored data.

[Angle]

Hi, Vivien!
Nice to see you'r developing your own nets.
What is "5 minic_3.02"? Does it use HCE?
I am going to include some Minic versions to my nearest tournaments.
Happy New Year! Aleksey.

[xr_a_y]

Hi, yes this is HCE. I will continue to develop HCE because I like that and this is used for the Minic skills and features parameters.

[Angle]

Hi, yes this is HCE. I will continue to develop HCE because I like that and this is used for the Minic skills and features parameters.

How many changes in 3.02 HCE compared to 2.53 HCE?

[xr_a_y]

Hi, yes this is HCE. I will continue to develop HCE because I like that and this is used for the Minic skills and features parameters.

How many changes in 3.02 HCE compared to 2.53 HCE?

Almost none . Just some fixes and options.

[xr_a_y]

Minic 3.03 is release.

No strength increase for HCE again, but some code cleanup and little fixes (maybe the endgame issue viewed in some tourney where Minic cannot mate in won position).

This release is also associated with a new net : "niggling nymph" that was built on FRC data and is thus stronger than previous nets in FRC situations (https://github.com/tryingsomestuff/NNUE-Nets).

Short TC tests for standard chess

Code: Select all

Rank Name                             Elo     +/-   Games   Score    Draw 
   1 minic_2.53_napping_nexus         245      18    1251   80.4%   29.3% 
   2 minic_2.53_nascent_nutrient      151      17    1251   70.4%   31.7% 
   3 minic_3.02_nettling_nemesis       80      16    1251   61.4%   32.4% 
   4 minic_dev_niggling_nymph         -58      16    1251   41.7%   35.2% ****
   5 minic_3.02_narcotized_nightshift -70      16    1251   40.0%   33.6% 
   6 minic_3.02                       -86      16    1251   37.9%   30.0% 
   7 minic_dev                        -98      16    1251   36.3%   31.3% 
   8 minic_3.01_nefarious_nucleus    -132      17    1251   31.9%   28.4% 

Short TC tests for FRC

Code: Select all

Rank Name                             Elo     +/-   Games   Score    Draw 
   1 minic_2.53_napping_nexus         290      16    2272   84.1%   16.9% 
   2 minic_2.53_nascent_nutrient      203      14    2271   76.3%   19.0% 
   3 minic_3.02_nettling_nemesis      -30      13    2271   45.7%   20.7% 
   4 minic_3.02                       -30      13    2271   45.6%   22.9% 
   5 minic_dev                        -44      13    2271   43.7%   22.5% 
   6 minic_dev_niggling_nymph         -53      13    2271   42.4%   22.1% ****
   7 minic_3.02_narcotized_nightshift -79      13    2271   38.8%   23.2% 
   8 minic_3.01_nefarious_nucleus    -206      15    2272   23.4%   15.5% 

I'm now working on :
- HCE texel tuning using pure Minic data
- deep depth NNUE training (depth 16)
- NNUE quantization

Have fun

[Angle]

Hi, Vivien!
Thanks for the new version! It's strange for me that in your tests 3.0.1 NeNu is much weaker than 3.0.2 NaNi. In my tournament it plays even stronger.

[xr_a_y]

Hi, Vivien!
Thanks for the new version! It's strange for me that in your tests 3.0.1 NeNu is much weaker than 3.0.2 NaNi. In my tournament it plays even stronger.

Maybe it depends on the opponent set (and of course the number of games). Also my tests are fast TC at 10s+0.1

[Angle]

Hi, Vivien!
Thanks for the new version! It's strange for me that in your tests 3.0.1 NeNu is much weaker than 3.0.2 NaNi. In my tournament it plays even stronger.

Maybe it depends on the opponent set (and of course the number of games). Also my tests are fast TC at 10s+0.1

Anyways, both versions play very interesting positional chess (though with some oddities from 3.0.2 NaNi such as unreasonably high evals or lack of underpromotion knowledge etc.).

[xr_a_y]

I just had the opportunity to test Minic3 on a AMD Ryzen Threadripper 3990X ... 55Mnps for 128 threads, quite cool