Minic version 2

[Modern Times]

Minic 2.47 is released. This is a quite special version, as NNUE (from Stockfish) as be integrated inside.
This release has NNUE technology enabled (by #define WITH_NNUE inside definition.hpp), but has to be considerate as "MinicNNUE" only if you specify a network file using the NNUEFile GUI option.

Hi there, so if you don't specify an NNUE file then it just runs as traditional Minic ?

[xr_a_y]

Yes it does.

[Frank Quisinsky]

Perfect!!

Best
Frank

[Gabor Szots]

Yes it does.

I suggest making NNUE the default if you think that makes your engine stronger. Tests are usually performed using default settings and if we use those the eventual result may be misleading.

Also, a link to your preferred nn would be useful. BTW, are SF nn's free to use?

[Guenther]

Yes it does.

I suggest making NNUE the default if you think that makes your engine stronger. Tests are usually performed using default settings and if we use those the eventual result may be misleading.

Also, a link to your preferred nn would be useful. BTW, are SF nn's free to use?

Well, he answered all of this already here, which will be true obviously for all further versions until he has trained his own net:
http://talkchess.com/forum3/viewtopic.p ... 50#p855313

[Gabor Szots]

Yes it does.

I suggest making NNUE the default if you think that makes your engine stronger. Tests are usually performed using default settings and if we use those the eventual result may be misleading.

Also, a link to your preferred nn would be useful. BTW, are SF nn's free to use?

Well, he answered all of this already here, which will be true obviously for all further versions until he has trained his own net:
http://talkchess.com/forum3/viewtopic.p ... 50#p855313

OK, thanks.

[Jamal Bubker]

Merci Vivien pour la sortie de cette nouvelle version !

[xr_a_y]

Some good news about Minic :

Very short TC (5s+0.1) versus various opponents shows some progress, especially for the NNUE part. The network used here is still SV net nn-97f742aaefcd.nnue for all versions (thus a Stockfish based net).

Code: Select all

Rank Name                          Elo     +/-   Games   Score    Draw 
   1 komodo-11.01-linux            244      18    1466   80.3%   24.1% 
   2 Ethereal                      222      17    1465   78.2%   23.5% 
   3 minic_2.50_uci_nnue           105      15    1465   64.7%   28.5% 
   4 minic_2.49_uci_nnue           102      15    1465   64.2%   30.9% 
   5 RubiChess                      76      15    1465   60.8%   30.2% 
   6 minic_2.48_uci_nnue            68      15    1465   59.7%   31.9% 
   7 Defenchess_2.2                 62      15    1466   58.8%   31.0% 
   8 demolito                        2      15    1465   50.3%   27.6% 
   9 nemorino                      -55      15    1465   42.2%   27.6% 
  10 texel                         -87      15    1465   37.7%   27.9% 
  11 Vajolet2_2.8                  -96      15    1465   36.5%   29.8% 
  12 minic_2.50                   -128      16    1465   32.4%   27.2% 
  13 minic_2.49                   -132      16    1466   31.9%   28.4% 
  14 minic_2.48                   -142      16    1465   30.6%   27.4% 
  15 gogobello_linux_bmi2         -223      18    1465   21.7%   21.6% 

Version 2.50 will be release soon. But let's first comment on release 2.49 (a milestone before integrating learning code from Nodchip repo and adapt it to Minic). This 2.49 version shows great progress on the NNUE side. It includes merges of recent SF cleanup in NNUE code and use of a scaling factor of NNUE eval versus standard eval to avoid retuning all the search ... This tuning factor is computed each time you run Minic based on 50000 random (fixed seed) positions. There is some progress in the standard eval also. Maybe +15Elo at longer TC.

My next step was to integrate learning code into Minic. This was not so difficult, and not that intrusive, and I still think that NNUE (use and learn parts) can be a shared library by all the community. Release 2.50 includes this learning code without any loss of strength. I didn't ported the genfen part because it does not feel necessary as I already implemented some genfen features for both random games and search tree and the conversion tool from pgn to genfen binary format in the previous release of Minic.

So what's next ? Well I trained by myself a quite good Stockfish based net using learner from Nodchip repo (but still very far from the powerfull SV nets !) and I'm now trying to build a Minic based net using the learning code ported to Minic and pure Minic data ! But I'm not in success yet ! I'll release a Minic based net as soon as I have one that is scoring +100 versus standard eval.

My point of view hasn't changed since August and I still think that NNUE eval (from SF based net or Minic based net) shall not be considered official Minic.

I'm trying hard to better understand this world but this technology still look a little unstable to me.

And please, do not forget to give credits to the net builder when you are posting results from MinicNNUE !

[xr_a_y]

And here is a 1 thread versus 4 threads short match at 10s+0.1.
Everything seem to be scaling well.

Code: Select all

Rank Name                          Elo     +/-   Games   Score    Draw 
   1 minic_2.50_uci_nnue_4t        294      99      61   84.4%   21.3% 
   2 minic_2.50_uci_nnue            23      72      61   53.3%   34.4% 
   3 minic_2.50_4t                 -57      76      61   41.8%   27.9% 
   4 minic_2.50                   -236      77      61   20.5%   34.4% 

[alex67a]

Minic don't work on Win 7 and Arena gui...