Dark Horse Update

[dkappe]

I’ve had some requests to update Dark Horse to the new larger HalfKA architecture. I haven’t done any testing beyond making sure the net is the strongest of the training run. Same data as DH 0.3 (less than 1b positions).

https://www.patreon.com/posts/dark-horse-0-3-52384228

[Eduard]

Thank you Dietrich! I am a fan of White Rose. Would it be possible to produce White Rose as a large network?

[Raphexon]

No mention of the license this time?

[dkappe]

Thank you Dietrich! I am a fan of White Rose. Would it be possible to produce White Rose as a large network?

That one relied on the old training software rather than the pytorch one.

[dkappe]

No mention of the license this time?

Thanks for reminding me. Updated.

[peter]

I’ve had some requests to update Dark Horse to the new larger HalfKA architecture. I haven’t done any testing beyond making sure the net is the strongest of the training run. Same data as DH 0.3 (less than 1b positions).

https://www.patreon.com/posts/dark-horse-0-3-52384228

Thanks a lot, Dietrich, I was looking forward impatiently to this

[peter]

I’ve had some requests to update Dark Horse to the new larger HalfKA architecture. I haven’t done any testing beyond making sure the net is the strongest of the training run. Same data as DH 0.3 (less than 1b positions).

https://www.patreon.com/posts/dark-horse-0-3-52384228

Here's ShashChess 1.7 with Dark Horse 0.3xl.bin, option Tal and MultiPV=4 at Vincent's short set of 114 out of Hard Talkchess- suite

http://talkchess.com/forum3/viewtopic.p ... 39#p884039

Code: Select all

Bisher gelöst: 81 von 114  ;  52:51m
         1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17  18  19  20
 -------------------------------------------------------------------------------------
   0 |   3  15   -   -  39   0   1   9  24   -   3   -   -   -   -   8  28   4   2   2
  20 |   2  13  24  15   1   -   0   1  36   0   -  17   2   6  41   0   3   -   -   2
  40 |   3   0   -   -  20  49   1   0   0  27   -   -   -  60  54  10   1   3  10   2
  60 |   0   8  14   0  60   2  33   4  60   0   -  12   0   3  26   5   2   0   0  13
  80 |   -   -   -   2   -   -   4   3  11  24  60   -   -   -   -  46  52  38   0   1
 100 |   3   -   -   0   4   -   -   9   1   -  42  58   -   -

1 min./position is my personal SMP- TC for the 12x3GHz Xeon CPU, only able of SSE4.1 popcnt- compiles, 8G hash.

81/114 still isn't one of the very best results in my list, but it's better then the one with default SF- net was, ShashChess had its best runs with the smaller nets before at my trials, probably due to my not quite up to date hardware. As far as I remember (always storing best runs only) 17.1 had at least 10 soulutions less with default big net then at single primary variant.

Maybe Vincent could give new Dark Horse NNUE a try with Honey and or ShashChess for his sheet at single- core conditions too, regards

[peter]

Here's ShashChess 1.7

Typo, 17.1 was meant of course...

[peter]

Here's ShashChess 1.7 with Dark Horse 0.3xl.bin, option Tal and MultiPV=4 at Vincent's short set of 114 out of Hard Talkchess- suite

http://talkchess.com/forum3/viewtopic.p ... 39#p884039

Code: Select all

Bisher gelöst: 81 von 114  ;  52:51m
         1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17  18  19  20
 -------------------------------------------------------------------------------------
   0 |   3  15   -   -  39   0   1   9  24   -   3   -   -   -   -   8  28   4   2   2
  20 |   2  13  24  15   1   -   0   1  36   0   -  17   2   6  41   0   3   -   -   2
  40 |   3   0   -   -  20  49   1   0   0  27   -   -   -  60  54  10   1   3  10   2
  60 |   0   8  14   0  60   2  33   4  60   0   -  12   0   3  26   5   2   0   0  13
  80 |   -   -   -   2   -   -   4   3  11  24  60   -   -   -   -  46  52  38   0   1
 100 |   3   -   -   0   4   -   -   9   1   -  42  58   -   -

1 min./position is my personal SMP- TC for the 12x3GHz Xeon CPU, only able of SSE4.1 popcnt- compiles, 8G hash.

81/114 still isn't one of the very best results in my list, but it's better then the one with default SF- net was

And here the direct comparison with same version, setting, hardware and TC, only difference: new green SF dev.- NNUE (nn-8e47cf062333.nnue):

Code: Select all

Bisher gelöst: 76 von 114  ;  54:11m

         1   2   3   4   5   6   7   8   9  10  11  12  13  14  15  16  17  18  19  20
 -------------------------------------------------------------------------------------
   0 |   1   4   -   -   -  44   4   -   -   -   -   -  21   -   -   0   0   9   1   1
  20 |   -   -  42   5   2   -   0  14  33   1   -  15   1   -   -   8   0   -   -   2
  40 |   2   0  13  57   5   -   0   0   0   1  24   -   -  15   -   2   0   0   4   1
  60 |   0   8   4   0   -  13   -   3  60   0   -  14   0   3  24   4   7   0   0   -
  80 |   -   7   -   0   -   -   6   4   -   4   -  44  60   -  54   3  21  24  54   2
 100 |   3  31  50   0  11   -  53   7   -   0   5   -   -   -

[dkappe]

Someone notified me that someone in the SF community is planning on training the Dark Horse XL net with SF data. That’s perfectly fine as long as you don’t distribute this derivative work. See the license.

Code: Select all

License - Attribution-NoDerivatives 4.0 International (CC BY-ND 4.0) https://creativecommons.org/licenses/by-nd/4.0/legalcode