Minic version 3

[xr_a_y]

Nope, ok with 3-4-5 too

Code: Select all

info string Info  2020-12-22 13:22:40-949: Init tb from path /ssd/Minic/TB/tablebase.sesse.net/syzygy/3-4-5
info string Info  2020-12-22 13:22:40-953: MAX_TB_MEN: 5

Also tested under Windows OS. I really wonder what makes this crash.

[Carlos777]

I am not sure if it is related, but I set up Arena to adjudicate the result according to 5 men Nalimov TB (when there is only 5 pieces left).
Would that be the issue?

Nope, ok with 3-4-5 too

Code: Select all

info string Info  2020-12-22 13:22:40-949: Init tb from path /ssd/Minic/TB/tablebase.sesse.net/syzygy/3-4-5
info string Info  2020-12-22 13:22:40-953: MAX_TB_MEN: 5

Also tested under Windows OS. I really wonder what makes this crash.

[Gabor Szots]

I installed Minic 3.02 under Arena as a WB engine using this command line:

-ttSize 256 -fullXboardOutput 1 -syzygyPath F:\Sakk\syzygy -NNUEFile narcotized_nightshift.bin

When I started it the protocol was xboard version 1, and it did not move, nor could it analyze. When I set the start position Arena told me "engine not found".

When I also inserted -xboard in the command line, it started as a UCI2 (!) engine and worked as expected.

[xr_a_y]

Gabor, Minic can use xboard or uci. The default is uci. If you specify -xboard as first command line argument it switches to xboard.
Sometimes, to use more arguments it is necessary to use -uci as first argument.

So in your case, I think

Code: Select all

-xboard -ttSizeMb 256 -fullXboardOutput 1 -syzygyPath F:\Sakk\syzygy -NNUEFile narcotized_nightshift.bin

Also, please note "-ttSize" is not a command line argument for Minic, it is -ttsizeMb
This hash size, as well as syzygyPath can be set using command or protocol option in your favorite GUI.

[xr_a_y]

I am not sure if it is related, but I set up Arena to adjudicate the result according to 5 men Nalimov TB (when there is only 5 pieces left).
Would that be the issue?

Shouldn't be an issue, I guess this is only a GUI related option.

[xr_a_y]

Merry Christmas everyone.

Here's another gift !

I'm releasing this morning, for fun only (not for rating list of course) a Stockfish data based net for Minic 3.02 named "nettling nemesis" (available here https://github.com/tryingsomestuff/NNUE-Nets as usual). It is a lot strong than Minic data based one (I mean A LOT ) !

And this will be a guide line for the next weeks for me, why this net based on SF data is so much stronger than the ones based on Minic data ? choice of positions ? number of data ? quality of eval ? endgame evaluation ? ... Maybe I'll find the key for improve Minic based net generation by studying this ...

[Werner]

Hi Vivien,
I wish you good luck. Here my results so far:

Minic 3.02NaNi x64 1CPU - Winter 0.9 x64 1CPU 45.0 - 55.0 +19/=52/-29 45.00%
Minic 3.02NaNi x64 1CPU - Orion 0.8NN x64 49.0 - 51.0 +27/=44/-29 49.00%
Minic 3.02NaNi x64 1CPU - Halogen 9.0 x64 1CPU 58.5 - 41.5 +36/=45/-19 58.50%
Minic 3.02NaNi x64 1CPU - Pedone 2.1 x64 1CPU 48.0 - 52.0 +26/=44/-30 48.00%
Minic 3.02NaNi x64 1CPU - Nirvana Chess 2.5 x64 1CPU 42.5 - 57.5 +25/=35/-40 42.50%
Minic 3.02NaNi x64 1CPU - Demolito 2020-10-10 x64 1CPU 57.0 - 43.0 +33/=48/-19 57.00%

[xr_a_y]

Hi Vivien,
I wish you good luck. Here my results so far:

Minic 3.02NaNi x64 1CPU - Winter 0.9 x64 1CPU 45.0 - 55.0 +19/=52/-29 45.00%
Minic 3.02NaNi x64 1CPU - Orion 0.8NN x64 49.0 - 51.0 +27/=44/-29 49.00%
Minic 3.02NaNi x64 1CPU - Halogen 9.0 x64 1CPU 58.5 - 41.5 +36/=45/-19 58.50%
Minic 3.02NaNi x64 1CPU - Pedone 2.1 x64 1CPU 48.0 - 52.0 +26/=44/-30 48.00%
Minic 3.02NaNi x64 1CPU - Nirvana Chess 2.5 x64 1CPU 42.5 - 57.5 +25/=35/-40 42.50%
Minic 3.02NaNi x64 1CPU - Demolito 2020-10-10 x64 1CPU 57.0 - 43.0 +33/=48/-19 57.00%

Thanks, this is quite good! I am happy with that

[xr_a_y]

Here are some short TC results (again this net is here just for the experiment, not for rating lists)

Code: Select all

Rank Name                          Elo     +/-   Games   Score    Draw 
   0 minic_3.02_ne_ne               28      11    2567   54.1%   27.0% 
   1 Defenchess_2.2                 54      28     428   57.7%   30.8% 
   2 demolito                        7      28     428   51.1%   28.3% 
   3 Vajolet2_2.8                  -51      28     427   42.7%   27.9% 
   4 Wasp375-linux                 -54      28     428   42.3%   26.6% 
   5 combusken                     -56      29     428   42.1%   24.3% 
   6 komodo-2.0.3                  -72      29     428   39.8%   24.1% 

[xr_a_y]

The "big Minic 500Elo split".

Here is the results of a small short TC tourney with all version of Minic and associated nets since august :

Code: Select all

Rank Name                            Elo     +/-   Games   Score    Draw 
   1 minic_2.53_nn-97f742aaefcd         296      32     519   84.6%   21.2% 
   2 minic_2.53_napping_nexus           172      26     521   72.9%   30.3% 
   3 minic_2.53_nascent_nutrient         84      26     520   61.8%   28.7% 
   4 minic_3.02_nettling_nemesis         36      25     521   55.2%   32.4% 
   5 minic_3.02                        -105      26     519   35.4%   26.4% 
   6 minic_2.46                        -113      27     520   34.3%   23.7% 
   7 minic_3.02_narcotized_nightshift  -128      27     520   32.3%   26.5% 
   8 minic_3.01_nefarious_nucleus      -205      29     520   23.5%   21.9% 

Minic 2.53 is using NNUE from Stockfish
Minic >=3 is using NNUE from Seer

What to learn ?
- 500Elo self play difference between "nefarious nucleus" (Seer style net) and one of the best SV net (Nodchip style net).
- "nefarious_nucleus", that just played at TCEC (at much longer TC of course), is quite weaker than HCE !
- "narcotized nightshift" is much stronger than "nefarious_nucleus", and looks like same level as HCE (at short TC)
- "nettling nemesis" (built with SF data) is far stronger than "narcotized nightshift"
- but "nettling nemesis" is still much weaker than the homebrew SF data net "napping nexus"

So quite a big variety of performances around Minic !

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'm currently generating some data from depth 9 FRC games.
I'll try to work on quantization during next week.

Have fun.