Made from scratch - no pgn, no fen learning , strictly AI,
https://www.dropbox.com/s/kjohqezamh42s ... .nnue?dl=0
Net can be used with any SF clone - you can do whatever you want with it. I just hope you can have some fun with it
Probably at least on par with Zappa Mexico using 8 cores at tc 2 min + 2 sec
[pgn][Event "Computer Chess Game"]
[Site "VM-894787"]
[Date "2021.02.27"]
[Round "-"]
[White "Zappa Mexico II"]
[Black "Honey v13"]
[Result "0-1"]
[TimeControl "120+3"]
[Annotator "1. +0.18 1... +0.02"]
1. e4 {+0.18/16} c5 {+0.02/20 39} 2. d4 {+0.40/15 6} cxd4 {-0.09/17 2.6} 3.
Qxd4 {+0.26/15 9} Nc6 {+0.01/19 9} 4. Qa4 {+0.26/15 14} Nf6 {+0.11/18 9} 5.
Nc3 {+0.27/15 8} e6 {+0.16/20 12} 6. Nf3 {+0.26/15 11} Qc7 {+0.13/20 14} 7.
Bb5 {+0.26/14 13} h6 {+0.15/21 22} 8. O-O {+0.43/14 5} a6 {+0.13/21 9} 9.
Bxc6 {+0.32/14 5} dxc6 {+0.14/18 2.2} 10. e5 {+0.32/15 9} Nd5 {+0.13/19 3}
11. Bd2 {+0.36/15 9} b5 {+0.14/21 12} 12. Qg4 {+0.28/16 8} Nxc3
{+0.14/16 1.0} 13. Bxc3 {+0.30/17 6} c5 {+0.20/18 4} 14. Rfd1 {+0.19/16 4}
Bb7 {+0.19/18 5} 15. b3 {+0.26/15 4} h5 {+0.17/21 13} 16. Qg3 {+0.33/14 7}
Rc8 {+0.20/19 4} 17. Rac1 {+0.30/13 2.2} Be4 {+0.19/19 5} 18. Qf4
{+0.47/14 2.1} Qb7 {+0.18/19 3} 19. Ng5 {+0.32/14 2.1} Bg6 {+0.11/20 3} 20.
Ba5 {+0.15/14 4} c4 {+0.09/20 3} 21. Rd4 {+0.15/14 2.2} Ba3 {+0.19/22 3}
22. Rcd1 {+0.56/14 2.5} O-O {+0.18/21 1.0} 23. Rd7 {+0.40/15 4} Qc6
{+0.25/24 1.6} 24. c3 {+0.40/14 0.1} b4 {+0.23/28 1.8} 25. R7d6
{+0.28/14 4} Qb5 {+0.24/27 1.5} 26. Bxb4 {+0.28/14 0.1} Bxb4 {+0.23/32 1.8}
27. cxb4 {+0.28/14 0.1} c3 {+0.23/31 1.8} 28. a4 {+0.17/16 4} Qe2
{+0.26/27 2.0} 29. Rc1 {-0.33/17 12} Qb2 {+0.25/29 4} 30. a5 {-0.91/14 9}
Qxb3 {+0.26/23 1.3} 31. Rd4 {-1.68/13 4} c2 {+0.28/29 4} 32. f3
{-2.16/15 6} Rfd8 {+0.75/25 1.2} 33. Kf2 {-2.05/15 2.1} Qc3 {+1.57/23 2.0}
34. Rd6 {-2.05/14 0.1} Rxd6 {+1.87/28 4} 35. exd6 {-2.05/14 0.1} Rd8
{+2.45/24 2.4} 36. b5 {-3.80/13 8} axb5 {+4.72/24 2.3} 37. Ne4 {-4.34/13 3}
Bxe4 {+5.27/25 2.2} 38. fxe4 {-4.34/13 0.2} b4 {+5.61/23 2.4} 39. Qe3
{-8.44/13 3} Qxe3+ {+5.74/22 2.3} 40. Kxe3 {-8.44/13 0.1} b3 {+6.24/21 2.7}
41. Kd2 {-10.34/17 7} b2 {+6.63/20 2.6} 42. Kxc2 {-10.34/17 0.1} bxc1=R+
{+9.28/22 13} 43. Kxc1 {-10.50/18 5} Rxd6 {+8.91/19 2.7} 44. Kc2
{-10.50/18 6} Ra6 {+10.44/21 2.1} 45. Kd2 {-11.18/18 5} Rxa5
{+32.19/25 2.0} 46. Ke2 {-11.64/18 5} Ra2+ {+57.81/27 4} 47. Kf1
{-12.63/17 0.2} h4 {+1000.12/30 1.8} 48. h3 {-14.63/19 5} Kh7
{+1000.11/40 1.7} 49. Kg1 {-1000.10/18 3} Kg6 {+1000.10/44 1.9} 50. Kf1
{-1000.09/18 1.4} Kg5 {+1000.09/45 1.9} 51. Ke1 {-1000.08/16 0.1} Kf4
{+1000.08/50 1.9} 52. Kd1 {-1000.07/15 0.1} Kxe4 {+1000.07/56 2.3} 53. Kc1
{-1000.06/12 0.1} Rxg2 {+1000.06/71 2.0} 54. Kd1 {-1000.05/14 0.2} Kd3
{+1000.05/198 2.0} 55. Kc1 {-1000.04/15 0.1} e5 {+1000.04/255 0.5} 56. Kb1
{-1000.03/16 0.1} Kc3 {+1000.03/255 0.4} 57. Ka1 {-1000.02/63 0.1} Kb3
{+1000.02/255 0.4} 58. Kb1 {-1000.01/63 0.1} Rg1# {+1000.01/255 0.4}
{Xboard adjudication: Checkmate} 0-1
[Event "Computer Chess Game"]
[Site "VM-894787"]
[Date "2021.02.27"]
[Round "-"]
[White "Honey v13"]
[Black "Zappa Mexico II"]
[Result "1-0"]
[TimeControl "120+3"]
[Annotator "1. +0.13 1... -0.21"]
1. Nf3 {+0.13/18} d5 {-0.21/16 18} 2. d4 {+0.08/18 15} Nc6 {-0.14/15 7} 3.
c4 {+0.15/17 6} Bg4 {-0.16/15 7} 4. Nc3 {+0.09/18 7} e6 {-0.14/16 10} 5.
cxd5 {+0.11/18 8} exd5 {-0.09/15 9} 6. a3 {+0.10/17 10} a6 {-0.15/15 8} 7.
h3 {+0.07/18 9} Bxf3 {-0.11/15 5} 8. gxf3 {+0.16/16 2.4} Nge7 {-0.11/15 6}
9. h4 {+0.16/19 12} Nf5 {-0.13/14 5} 10. e4 {+0.20/19 9} dxe4 {-0.01/15 5}
11. d5 {+0.20/18 1.4} Ne5 {-0.01/14 0.1} 12. fxe4 {+0.18/19 2.7} Nxh4
{+0.00/15 5} 13. f4 {+0.19/17 1.7} Nef3+ {+0.32/14 5} 14. Ke2 {+0.20/20 3}
Bc5 {+0.28/15 7} 15. Kd3 {+0.19/20 4} c6 {+0.35/14 6} 16. Kc2 {+0.24/22 11}
cxd5 {+0.72/15 6} 17. Nxd5 {+0.00/21 3} f5 {+0.72/15 6} 18. Kb1
{+0.20/23 4} fxe4 {+0.59/15 10} 19. Bc4 {+0.23/23 6} Nf5 {+0.87/14 10} 20.
Qe2 {+0.24/22 7} Nd6 {+0.79/15 8} 21. Rd1 {+0.25/22 4} Rc8 {+0.74/15 6} 22.
Be3 {+0.25/24 4} Bxe3 {+0.96/15 7} 23. Nxe3 {+0.24/24 1.2} Qf6 {+0.90/15 6}
24. Nd5 {+0.22/25 1.9} Qf5 {+0.83/15 4} 25. Bb3 {+0.24/25 1.7} h5
{+0.81/15 2.9} 26. Ka2 {+0.27/29 4} Rh6 {+0.72/15 4} 27. Qf2 {+0.23/30 6}
h4 {+0.67/14 6} 28. Rac1 {+0.22/28 1.3} Qd7 {+0.36/14 7} 29. Nb6
{+0.21/29 5} Rxc1 {+0.36/13 0.2} 30. Rxc1 {+0.23/31 1.4} Qf5 {-0.74/14 7}
31. Rc7 {+1.81/25 6} g5 {-1.10/12 4} 32. Nd5 {+3.73/23 1.6} Kf8
{-1.44/13 2.2} 33. Qa7 {+4.20/24 2.2} Re6 {-3.30/12 4} 34. Rd7
{+4.48/25 2.2} Nd2 {-5.21/12 5} 35. Qd4 {+5.85/22 3} Nf7 {-7.73/13 4} 36.
Qb4+ {+6.21/20 2.3} Rd6 {-8.65/13 4} 37. Rxd6 {+7.02/23 2.4} Nxd6
{-13.40/12 4} 38. Qxd6+ {+8.13/21 2.3} Kg7 {-13.40/12 0.2} 39. Ne3
{+10.74/21 2.1} Qf8 {-13.11/12 2.8} 40. Qxd2 {+99.97/27 2.4} Kh7
{-16.40/12 4} 41. Qd7+ {+1000.11/30 2.5} Qg7 {-1000.09/16 2.4} 42. Qd8
{+1000.09/33 2.8} a5 {-1000.07/15 1.1} 43. Be6 {+1000.07/38 2.4} Qg6
{-1000.06/13 0.2} 44. f5 {+1000.06/61 2.3} Qxe6+ {-1000.05/13 0.1} 45. fxe6
{+1000.05/255 1.4} b5 {-1000.04/14 0.1} 46. Qxg5 {+1000.04/255 0.4} b4
{-1000.03/23 0.2} 47. Nf5 {+1000.03/255 0.4} b3+ {-1000.02/63 0.2} 48. Ka1
{+1000.02/255 0.4} h3 {-1000.01/63 0.1} 49. Qg7# {+1000.01/255 0.4}
{Xboard adjudication: Checkmate} 1-0[/pgn]
Much appreciation to Norman Schmidt. I was struggling with the command line method, his NNUE-GUI made it so much easier.
https://github.com/FireFather/nnue-gui/ ... readme.txt
nn-f0c1c3cbf2f1.nnue was made simply from depth 4 100 million generated fens
trainer bin ->gensfen depth 4 loop 100000000 > took about 2 to 3 minutes to generate on machine using 30 cores
The validation bin was made with depth 8 , 10 million generated fens
validation bin _>gensfen depth 8 loop 10000000 > a two to 3 minutes to generate
For the first learn run, I set it so that it would end fairly quickly
learn command -> learn targetdir training epochs 1000 batchsize 1000000 use_draw_in_training 1 use_draw_in_validation 1 lr 1 lambda 1 eval_limit 32000 nn_batch_size 1000 newbob_decay 0.5 eval_save_interval 1000000 loss_output_interval 1000000 validation_set_file_name validation/10MD8.binpack > ran just a few minutes
A NNUE that appears to be at least on par with Zappo Mexico using 8 cores at tc 2 min + 2 sec was made in less than 15 minutes of generation time, Most of that is a testimony of how good the SF core is, i.e., the use of my net is actually making SF play much weaker.
With one thread and at the faster tc of 1 min plus 1 second, appears to play almost like a anti-computer style and it does make some bizzaro moves. Appeared to be weaker than Zappa at TC 1 1 with just one thread. The game below is against the level World Champion in Honey( nodes searched are limited) . It just sort sits back and let the attacking side over extend. Note how black thought was winning around move 33 and it quickly turned. Seems to be very non aggressive in the early stage. Does seem to play the endgame decent, which I was surprised at. That is probably due to the SF code.
[pgn][Event "Computer Chess Game"]
[Site "VM-894787"]
[Date "2021.02.27"]
[Round "-"]
[White "Bluefish v13 100MD4 V_1MD8 E-100"]
[Black "Honey v13 Level World Champion CCRL "]
[Result "1-0"]
[TimeControl "60+1"]
[Annotator "1. -0.19 1... -0.09"]
1. e3 {-0.19/17} Nf6 {-0.09/19 0.3} 2. Ne2 {+0.18/14 4} d5 {+0.32/18 0.3}
3. Nd4 {+0.26/15 7} e5 {+0.81/17 0.3} 4. Ne2 {+0.12/16 8} Nc6
{+0.91/15 0.3} 5. Na3 {-0.19/17 13} h5 {+1.58/15 0.3} 6. c3 {-0.20/16 6} h4
{+1.64/16 0.3} 7. h3 {-0.12/17 8} Bf5 {+1.74/16 0.3} 8. Ng1 {-0.13/15 5} d4
{+1.97/17 0.3} 9. cxd4 {-0.21/18 5} Bxa3 {+2.16/17 0.3} 10. bxa3
{+0.23/14 0.7} exd4 {+0.63/16 0.3} 11. Bb2 {-0.22/15 2.1} Qd7
{+0.62/16 0.3} 12. Nf3 {+0.20/12 0.3} d3 {+0.63/17 0.3} 13. Qa4
{+0.20/15 2.6} O-O-O {+0.10/16 0.3} 14. Rc1 {+0.00/15 1.3} Ne4
{+0.37/18 0.3} 15. Ne5 {-0.24/16 1.0} Nxe5 {+0.06/21 0.4} 16. Bxe5
{-0.24/16 0.3} Qxa4 {+0.18/23 0.3} 17. Rxc7+ {-0.19/17 0.6} Kb8
{+0.09/22 0.4} 18. Rc4+ {+0.00/16 0.5} Ka8 {+0.06/23 0.4} 19. Rxa4
{-0.21/17 0.8} Rd5 {+0.15/24 0.4} 20. Bb2 {-0.18/21 2.5} Rc8 {+0.49/18 0.4}
21. f3 {-0.20/18 0.3} Nxd2 {+0.61/20 0.4} 22. e4 {-0.20/21 0.5} Rb5
{+0.50/19 0.4} 23. Rb4 {-0.20/19 0.6} Rxb4 {+0.66/22 0.4} 24. axb4
{-0.02/22 0.7} Rc2 {+0.77/22 0.4} 25. Bxd3 {-0.07/21 1.0} Rxb2
{+0.41/20 0.4} 26. exf5 {+0.00/17 0.6} Rxa2 {+0.50/22 0.4} 27. Rg1
{-0.06/21 3} Rb2 {+0.51/21 0.4} 28. f6 {-0.09/18 0.8} gxf6 {+0.66/19 0.5}
29. Be2 {+0.00/20 1.5} Nb3 {+1.29/22 0.5} 30. Kf2 {+0.00/20 0.6} Nd4
{+1.66/19 0.5} 31. Rd1 {+0.00/18 1.0} Rxb4 {+1.61/20 0.4} 32. Bd3
{+0.12/18 1.6} Rb2+ {+1.65/19 0.5} 33. Ke3 {+0.19/18 0.3} Ne6
{+1.84/18 0.5} 34. Rb1 {-0.03/19 1.7} f5 {+2.18/19 0.4} 35. Rxb2
{+3.61/17 0.3} f4+ {-4.35/16 0.4} 36. Ke4 {+3.81/17 0.6} Nc5+
{-4.31/15 0.3} 37. Kd4 {+3.93/17 0.5} b6 {-4.46/15 0.3} 38. Bc4
{+3.99/19 0.8} Kb7 {-4.83/15 0.3} 39. Re2 {+4.60/19 0.8} Kc6 {-4.94/16 0.3}
40. Re7 {+4.79/18 0.7} Kd6 {-5.06/17 0.3} 41. Rxf7 {+4.99/18 0.8} Nd7
{-5.26/15 0.4} 42. Rxf4 {+5.62/18 0.8} a5 {-5.53/17 0.4} 43. Rxh4
{+5.98/20 0.9} a4 {-5.86/16 0.4} 44. Rh6+ {+6.45/16 0.6} Ke7 {-6.68/16 0.4}
45. Re6+ {+7.24/17 0.8} Kf8 {-6.64/15 0.4} 46. h4 {+25.33/20 0.7} a3
{-26.07/15 0.3} 47. h5 {+43.82/22 0.9} b5 {-33.92/17 0.3} 48. h6
{+99.92/22 1.2} Nb8 {-72.70/17 0.4} 49. Rf6+ {+1000.12/22 0.8} Ke7
{-99.93/20 0.4} 50. h7 {+1000.10/25 0.7} Kxf6 {-99.95/20 0.5} 51. h8=Q+
{+1000.09/27 0.9} Ke7 {-99.95/21 0.5} 52. Qe5+ {+1000.08/27 0.8} Kd7
{-99.95/20 0.5} 53. Bxb5+ {+1000.07/29 0.8} Kc8 {-99.96/21 0.5} 54. Qe8+
{+1000.06/33 0.7} Kc7 {-99.96/18 0.5} 55. Kc5 {+1000.05/48 0.8} Kb7
{-1000.04/52 0.3} 56. Qe7+ {+1000.04/186 0.7} Kc8 {-1000.03/251 0.9} 57.
Bc6 {+1000.03/255 0.4} Na6+ {-1000.02/255 0.4} 58. Kd6 {+1000.02/255 0.4}
a2 {-1000.01/250 0.6} 59. Qf8# {+1000.01/255 0.4}
{Xboard adjudication: Checkmate} 1-0[/pgn]
Long Term , will continue to try to get it better but only through the process from above - no learning from other engines, external pgn files or external fens. Only because I am curious how a strictly pure unadulterated AI approach will evolve overtime.
Obviously this is the result of all the others who blazed the trails before me , and to all of those , I tip my hat and give a nod of the head.
Obviously this net is not the quality of the standard SF NNUE, it's not even close.
My first NNUE nn-f0c1c3cbf2f1.nnue
Moderators: hgm, Rebel, chrisw
-
- Posts: 4889
- Joined: Thu Mar 09, 2006 6:34 am
- Location: Pen Argyl, Pennsylvania
-
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
Re: My first NNUE nn-f0c1c3cbf2f1.nnue
Nice work.
+1
+1
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".
-
- Posts: 4889
- Joined: Thu Mar 09, 2006 6:34 am
- Location: Pen Argyl, Pennsylvania
Re: My first NNUE nn-f0c1c3cbf2f1.nnue
Still working on the second generation, this game was interesting checkout White's move 6, starting to develop its own "AI" style. Reinforcement Training with 200M posititions at D5/ Time control was 2 3.
[pgn][Site "VM-894787"]
[Date "2021.02.28"]
[Round "-"]
[White "Bluefish-a7dd21a4d09d "]
[Black "Zappa Mexico"]
[Result "1-0"]
[Annotator "1. +0.10 1... -0.09"]
1. Nc3 {+0.10/19} d5 {-0.09/15 13} 2. Nf3 {-0.01/18 28} d4 {+0.08/14 20} 3.
Ne4 {+0.15/18 6} e5 {+0.15/14 10} 4. e3 {+0.15/19 9} Nf6 {+0.16/14 24} 5.
Neg5 {+0.16/17 3} h6 {+0.31/14 29} 6. Nxf7 {+0.16/20 3} Kxf7 {+0.31/13 0.2}
7. Nxe5+ {+0.00/23 5} Ke8 {+0.32/14 8} 8. Bb5+ {+0.13/21 11} c6
{+0.58/14 7} 9. Bc4 {+0.14/25 4} dxe3 {+0.31/13 11} 10. Bf7+ {+0.14/20 2.7}
Ke7 {+0.31/12 0.2} 11. Bb3 {+0.13/24 3} Qd4 {+0.17/13 12} 12. Ng6+
{+0.14/26 3} Ke8 {+0.17/13 0.2} 13. f3 {+0.14/27 10} Rh7 {+0.31/14 9} 14.
c3 {+0.15/20 5} Qc5 {+0.47/13 4} 15. Qe2 {+0.15/21 5} Qg5 {+0.38/14 8} 16.
Nxf8 {+0.14/22 8} Kxf8 {+0.38/13 0.2} 17. O-O {+0.12/20 4} Nbd7
{+0.43/13 4} 18. Bc2 {+0.13/23 17} Rh8 {+0.48/14 4} 19. d4 {+0.09/21 11}
Nd5 {+0.57/15 8} 20. f4 {+0.14/20 2.0} Qf6 {+0.47/15 7} 21. c4 {+0.14/20 7}
Nb4 {+0.37/14 4} 22. Bb3 {+0.13/21 7} Qxd4 {+0.46/14 4} 23. Bxe3
{+0.14/19 2.1} Qd3 {+0.33/15 7} 24. Qf2 {+0.12/24 7} Qe4 {-0.01/15 14} 25.
Rad1 {+0.15/24 7} a5 {-0.78/14 14} 26. f5 {+0.16/23 3} Kg8 {-0.19/12 4} 27.
Bd4 {+0.17/23 4} c5 {-2.05/13 10} 28. Qg3 {+3.40/25 3} Rh7 {-1.80/12 4} 29.
Rfe1 {+4.03/24 2.4} Qxf5 {-3.08/12 6} 30. Re8+ {+4.20/22 3} Kf7
{-5.14/12 4} 31. Rde1 {+4.31/25 6} cxd4 {-5.14/12 0.1} 32. c5+
{+6.20/23 2.4} Nd5 {-5.14/12 0.1} 33. R8e7+ {+6.38/23 2.5} Kg8 {-7.14/12 4}
34. Qd6 {+6.90/22 4} Nf8 {-7.14/11 0.2} 35. Bxd5+ {+7.25/21 4} Be6
{-12.48/13 4} 36. R1xe6 {+7.50/21 3} Kh8 {-11.47/13 4} 37. Re5 {+7.56/21 4}
Qf4 {-12.16/13 4} 38. Bxb7 {+7.77/21 3} g5 {-12.79/13 3} 39. Bxa8
{+8.11/21 3} Rxe7 {-12.79/13 0.1} 40. Rxe7 {+8.77/21 2.6} Qxd6
{-15.95/14 7} 41. cxd6 {+9.53/21 3} h5 {-15.96/14 4} 42. Kf2 {+11.72/22 3}
g4 {-15.53/13 4} 43. Ke2 {+28.49/22 4} h4 {-17.04/13 4} 44. Kd3
{+1000.10/32 6} h3 {-17.69/14 5} 45. g3 {+1000.09/32 1.9} a4
{-1000.09/16 4} 46. Be4 {+1000.08/39 2.7} Kg8 {-1000.07/16 1.1} 47. Kxd4
{+1000.07/45 2.4} a3 {-1000.06/15 0.1} 48. bxa3 {+1000.06/64 2.5} Kh8
{-1000.05/15 0.1} 49. Ke5 {+1000.05/100 2.5} Kg8 {-1000.04/17 0.1} 50. Kf6
{+1000.04/255 0.9} Kh8 {-1000.03/21 0.1} 51. d7 {+1000.03/255 0.9} Nxd7+
{-1000.02/16 0.1} 52. Rxd7 {+1000.02/255 0.8} Kg8 {-1000.01/63 0.1} 53.
Rd8# {+1000.01/255 0.8}
{White mates} 1-0[/pgn]
Running test now to see which of seven 2nd generation nets survives the selection process. With the second generation, draw rates are going up which is a good sign, they were ridiculously low on the first generation ( in the teens) . It almost feels like I am working with genetics here. Keep picking the best one.
[pgn][Site "VM-894787"]
[Date "2021.02.28"]
[Round "-"]
[White "Bluefish-a7dd21a4d09d "]
[Black "Zappa Mexico"]
[Result "1-0"]
[Annotator "1. +0.10 1... -0.09"]
1. Nc3 {+0.10/19} d5 {-0.09/15 13} 2. Nf3 {-0.01/18 28} d4 {+0.08/14 20} 3.
Ne4 {+0.15/18 6} e5 {+0.15/14 10} 4. e3 {+0.15/19 9} Nf6 {+0.16/14 24} 5.
Neg5 {+0.16/17 3} h6 {+0.31/14 29} 6. Nxf7 {+0.16/20 3} Kxf7 {+0.31/13 0.2}
7. Nxe5+ {+0.00/23 5} Ke8 {+0.32/14 8} 8. Bb5+ {+0.13/21 11} c6
{+0.58/14 7} 9. Bc4 {+0.14/25 4} dxe3 {+0.31/13 11} 10. Bf7+ {+0.14/20 2.7}
Ke7 {+0.31/12 0.2} 11. Bb3 {+0.13/24 3} Qd4 {+0.17/13 12} 12. Ng6+
{+0.14/26 3} Ke8 {+0.17/13 0.2} 13. f3 {+0.14/27 10} Rh7 {+0.31/14 9} 14.
c3 {+0.15/20 5} Qc5 {+0.47/13 4} 15. Qe2 {+0.15/21 5} Qg5 {+0.38/14 8} 16.
Nxf8 {+0.14/22 8} Kxf8 {+0.38/13 0.2} 17. O-O {+0.12/20 4} Nbd7
{+0.43/13 4} 18. Bc2 {+0.13/23 17} Rh8 {+0.48/14 4} 19. d4 {+0.09/21 11}
Nd5 {+0.57/15 8} 20. f4 {+0.14/20 2.0} Qf6 {+0.47/15 7} 21. c4 {+0.14/20 7}
Nb4 {+0.37/14 4} 22. Bb3 {+0.13/21 7} Qxd4 {+0.46/14 4} 23. Bxe3
{+0.14/19 2.1} Qd3 {+0.33/15 7} 24. Qf2 {+0.12/24 7} Qe4 {-0.01/15 14} 25.
Rad1 {+0.15/24 7} a5 {-0.78/14 14} 26. f5 {+0.16/23 3} Kg8 {-0.19/12 4} 27.
Bd4 {+0.17/23 4} c5 {-2.05/13 10} 28. Qg3 {+3.40/25 3} Rh7 {-1.80/12 4} 29.
Rfe1 {+4.03/24 2.4} Qxf5 {-3.08/12 6} 30. Re8+ {+4.20/22 3} Kf7
{-5.14/12 4} 31. Rde1 {+4.31/25 6} cxd4 {-5.14/12 0.1} 32. c5+
{+6.20/23 2.4} Nd5 {-5.14/12 0.1} 33. R8e7+ {+6.38/23 2.5} Kg8 {-7.14/12 4}
34. Qd6 {+6.90/22 4} Nf8 {-7.14/11 0.2} 35. Bxd5+ {+7.25/21 4} Be6
{-12.48/13 4} 36. R1xe6 {+7.50/21 3} Kh8 {-11.47/13 4} 37. Re5 {+7.56/21 4}
Qf4 {-12.16/13 4} 38. Bxb7 {+7.77/21 3} g5 {-12.79/13 3} 39. Bxa8
{+8.11/21 3} Rxe7 {-12.79/13 0.1} 40. Rxe7 {+8.77/21 2.6} Qxd6
{-15.95/14 7} 41. cxd6 {+9.53/21 3} h5 {-15.96/14 4} 42. Kf2 {+11.72/22 3}
g4 {-15.53/13 4} 43. Ke2 {+28.49/22 4} h4 {-17.04/13 4} 44. Kd3
{+1000.10/32 6} h3 {-17.69/14 5} 45. g3 {+1000.09/32 1.9} a4
{-1000.09/16 4} 46. Be4 {+1000.08/39 2.7} Kg8 {-1000.07/16 1.1} 47. Kxd4
{+1000.07/45 2.4} a3 {-1000.06/15 0.1} 48. bxa3 {+1000.06/64 2.5} Kh8
{-1000.05/15 0.1} 49. Ke5 {+1000.05/100 2.5} Kg8 {-1000.04/17 0.1} 50. Kf6
{+1000.04/255 0.9} Kh8 {-1000.03/21 0.1} 51. d7 {+1000.03/255 0.9} Nxd7+
{-1000.02/16 0.1} 52. Rxd7 {+1000.02/255 0.8} Kg8 {-1000.01/63 0.1} 53.
Rd8# {+1000.01/255 0.8}
{White mates} 1-0[/pgn]
Running test now to see which of seven 2nd generation nets survives the selection process. With the second generation, draw rates are going up which is a good sign, they were ridiculously low on the first generation ( in the teens) . It almost feels like I am working with genetics here. Keep picking the best one.
Code: Select all
PGN File: c:/cluster.mfb/pgn/02280330.pgn
Time Control: Time Control-> base+inc: 20+0.400
Games: 4200
Threads: 1
Hash: 128
Current date : time (EDST)
Date: 02/28/21 : 04:07:55
Projected-> Time: 1h:53m:6s
Run -> Time: 0h:37m:16s
1392 game(s) loaded
Rank Name Rating Δ + - # Σ Σ% W L D W% =% OppR
---------------------------------------------------------------------------------------------------------
1 Bluefish-a7dd21a4d09d 3517 0.0 30 30 398 210.0 52.8 161 139 98 40.5 24.6 3497
2 Bluefish-1fe3f556b52a 3510 7.3 29 29 398 205.5 51.6 151 138 109 37.9 27.4 3498
3 Bluefish-ad014ffac487 3509 0.7 29 29 398 205.0 51.5 152 140 106 38.2 26.6 3499
4 Bluefish-4ef1aa083e77 3501 8.4 29 29 398 199.0 50.0 144 144 110 36.2 27.6 3500
5 Bluefish-4e4fb00d3a85 3490 10.1 30 30 396 192.0 48.5 145 157 94 36.6 23.7 3502
6 Bluefish-1408e1e8963e 3489 1.1 30 30 398 191.5 48.1 144 159 95 36.2 23.9 3502
7 Bluefish-6709b398ac42 3484 5.0 31 31 398 189.0 47.5 154 174 70 38.7 17.6 3503
---------------------------------------------------------------------------------------------------------
-
- Posts: 4607
- Joined: Wed Oct 01, 2008 6:33 am
- Location: Regensburg, Germany
- Full name: Guenther Simon
Re: My first NNUE nn-f0c1c3cbf2f1.nnue
Well it did not turn quickly by normal play. Black simply let his Rb2 en prise for nothing.MikeB wrote: ↑Sat Feb 27, 2021 9:54 pm Made from scratch - no pgn, no fen learning , strictly AI,
https://www.dropbox.com/s/kjohqezamh42s ... .nnue?dl=0
Net can be used with any SF clone - you can do whatever you want with it. I just hope you can have some fun with it
...
With one thread and at the faster tc of 1 min plus 1 second, appears to play almost like a anti-computer style and it does make some bizzaro moves. Appeared to be weaker than Zappa at TC 1 1 with just one thread. The game below is against the level World Champion in Honey( nodes searched are limited) . It just sort sits back and let the attacking side over extend. Note how black thought was winning around move 33 and it quickly turned. Seems to be very non aggressive in the early stage. Does seem to play the endgame decent, which I was surprised at. That is probably due to the SF code.
[pgn][Event "Computer Chess Game"]
[Site "VM-894787"]
[Date "2021.02.27"]
[Round "-"]
[White "Bluefish v13 100MD4 V_1MD8 E-100"]
[Black "Honey v13 Level World Champion CCRL "]
[Result "1-0"]
[TimeControl "60+1"]
[Annotator "1. -0.19 1... -0.09"]
1. e3 {-0.19/17} Nf6 {-0.09/19 0.3} 2. Ne2 {+0.18/14 4} d5 {+0.32/18 0.3}
3. Nd4 {+0.26/15 7} e5 {+0.81/17 0.3} 4. Ne2 {+0.12/16 8} Nc6
{+0.91/15 0.3} 5. Na3 {-0.19/17 13} h5 {+1.58/15 0.3} 6. c3 {-0.20/16 6} h4
{+1.64/16 0.3} 7. h3 {-0.12/17 8} Bf5 {+1.74/16 0.3} 8. Ng1 {-0.13/15 5} d4
{+1.97/17 0.3} 9. cxd4 {-0.21/18 5} Bxa3 {+2.16/17 0.3} 10. bxa3
{+0.23/14 0.7} exd4 {+0.63/16 0.3} 11. Bb2 {-0.22/15 2.1} Qd7
{+0.62/16 0.3} 12. Nf3 {+0.20/12 0.3} d3 {+0.63/17 0.3} 13. Qa4
{+0.20/15 2.6} O-O-O {+0.10/16 0.3} 14. Rc1 {+0.00/15 1.3} Ne4
{+0.37/18 0.3} 15. Ne5 {-0.24/16 1.0} Nxe5 {+0.06/21 0.4} 16. Bxe5
{-0.24/16 0.3} Qxa4 {+0.18/23 0.3} 17. Rxc7+ {-0.19/17 0.6} Kb8
{+0.09/22 0.4} 18. Rc4+ {+0.00/16 0.5} Ka8 {+0.06/23 0.4} 19. Rxa4
{-0.21/17 0.8} Rd5 {+0.15/24 0.4} 20. Bb2 {-0.18/21 2.5} Rc8 {+0.49/18 0.4}
21. f3 {-0.20/18 0.3} Nxd2 {+0.61/20 0.4} 22. e4 {-0.20/21 0.5} Rb5
{+0.50/19 0.4} 23. Rb4 {-0.20/19 0.6} Rxb4 {+0.66/22 0.4} 24. axb4
{-0.02/22 0.7} Rc2 {+0.77/22 0.4} 25. Bxd3 {-0.07/21 1.0} Rxb2
{+0.41/20 0.4} 26. exf5 {+0.00/17 0.6} Rxa2 {+0.50/22 0.4} 27. Rg1
{-0.06/21 3} Rb2 {+0.51/21 0.4} 28. f6 {-0.09/18 0.8} gxf6 {+0.66/19 0.5}
29. Be2 {+0.00/20 1.5} Nb3 {+1.29/22 0.5} 30. Kf2 {+0.00/20 0.6} Nd4
{+1.66/19 0.5} 31. Rd1 {+0.00/18 1.0} Rxb4 {+1.61/20 0.4} 32. Bd3
{+0.12/18 1.6} Rb2+ {+1.65/19 0.5} 33. Ke3 {+0.19/18 0.3} Ne6
{+1.84/18 0.5} 34. Rb1 {-0.03/19 1.7} f5 {+2.18/19 0.4} 35. Rxb2
{+3.61/17 0.3} f4+ {-4.35/16 0.4} 36. Ke4 {+3.81/17 0.6} Nc5+
{-4.31/15 0.3} 37. Kd4 {+3.93/17 0.5} b6 {-4.46/15 0.3} 38. Bc4
{+3.99/19 0.8} Kb7 {-4.83/15 0.3} 39. Re2 {+4.60/19 0.8} Kc6 {-4.94/16 0.3}
40. Re7 {+4.79/18 0.7} Kd6 {-5.06/17 0.3} 41. Rxf7 {+4.99/18 0.8} Nd7
{-5.26/15 0.4} 42. Rxf4 {+5.62/18 0.8} a5 {-5.53/17 0.4} 43. Rxh4
{+5.98/20 0.9} a4 {-5.86/16 0.4} 44. Rh6+ {+6.45/16 0.6} Ke7 {-6.68/16 0.4}
45. Re6+ {+7.24/17 0.8} Kf8 {-6.64/15 0.4} 46. h4 {+25.33/20 0.7} a3
{-26.07/15 0.3} 47. h5 {+43.82/22 0.9} b5 {-33.92/17 0.3} 48. h6
{+99.92/22 1.2} Nb8 {-72.70/17 0.4} 49. Rf6+ {+1000.12/22 0.8} Ke7
{-99.93/20 0.4} 50. h7 {+1000.10/25 0.7} Kxf6 {-99.95/20 0.5} 51. h8=Q+
{+1000.09/27 0.9} Ke7 {-99.95/21 0.5} 52. Qe5+ {+1000.08/27 0.8} Kd7
{-99.95/20 0.5} 53. Bxb5+ {+1000.07/29 0.8} Kc8 {-99.96/21 0.5} 54. Qe8+
{+1000.06/33 0.7} Kc7 {-99.96/18 0.5} 55. Kc5 {+1000.05/48 0.8} Kb7
{-1000.04/52 0.3} 56. Qe7+ {+1000.04/186 0.7} Kc8 {-1000.03/251 0.9} 57.
Bc6 {+1000.03/255 0.4} Na6+ {-1000.02/255 0.4} 58. Kd6 {+1000.02/255 0.4}
a2 {-1000.01/250 0.6} 59. Qf8# {+1000.01/255 0.4}
{Xboard adjudication: Checkmate} 1-0[/pgn]
What was the node limitation handicap for Honey here (even with nodes limited this move seems inexplicable)?
Black played 34...f5??? and the game was over
[d]k7/pp3p2/4np2/8/7p/3BKP1P/1r4P1/1R6 b - - 5 34
https://rwbc-chess.de
trollwatch:
Talkchess nowadays is a joke - it is full of trolls/idiots/people stuck in the pleistocene > 80% of the posts fall into this category...
trollwatch:
Talkchess nowadays is a joke - it is full of trolls/idiots/people stuck in the pleistocene > 80% of the posts fall into this category...
-
- Posts: 4889
- Joined: Thu Mar 09, 2006 6:34 am
- Location: Pen Argyl, Pennsylvania
Re: My first NNUE nn-f0c1c3cbf2f1.nnue
Interesting , not able to reproduce ... Looks like a bug, even with using UCI_Limit StrengthGuenther wrote: ↑Sun Feb 28, 2021 10:37 amWell it did not turn quickly by normal play. Black simply let his Rb2 en prise for nothing.MikeB wrote: ↑Sat Feb 27, 2021 9:54 pm Made from scratch - no pgn, no fen learning , strictly AI,
https://www.dropbox.com/s/kjohqezamh42s ... .nnue?dl=0
Net can be used with any SF clone - you can do whatever you want with it. I just hope you can have some fun with it
...
With one thread and at the faster tc of 1 min plus 1 second, appears to play almost like a anti-computer style and it does make some bizzaro moves. Appeared to be weaker than Zappa at TC 1 1 with just one thread. The game below is against the level World Champion in Honey( nodes searched are limited) . It just sort sits back and let the attacking side over extend. Note how black thought was winning around move 33 and it quickly turned. Seems to be very non aggressive in the early stage. Does seem to play the endgame decent, which I was surprised at. That is probably due to the SF code.
[pgn][Event "Computer Chess Game"]
[Site "VM-894787"]
[Date "2021.02.27"]
[Round "-"]
[White "Bluefish v13 100MD4 V_1MD8 E-100"]
[Black "Honey v13 Level World Champion CCRL "]
[Result "1-0"]
[TimeControl "60+1"]
[Annotator "1. -0.19 1... -0.09"]
1. e3 {-0.19/17} Nf6 {-0.09/19 0.3} 2. Ne2 {+0.18/14 4} d5 {+0.32/18 0.3}
3. Nd4 {+0.26/15 7} e5 {+0.81/17 0.3} 4. Ne2 {+0.12/16 8} Nc6
{+0.91/15 0.3} 5. Na3 {-0.19/17 13} h5 {+1.58/15 0.3} 6. c3 {-0.20/16 6} h4
{+1.64/16 0.3} 7. h3 {-0.12/17 8} Bf5 {+1.74/16 0.3} 8. Ng1 {-0.13/15 5} d4
{+1.97/17 0.3} 9. cxd4 {-0.21/18 5} Bxa3 {+2.16/17 0.3} 10. bxa3
{+0.23/14 0.7} exd4 {+0.63/16 0.3} 11. Bb2 {-0.22/15 2.1} Qd7
{+0.62/16 0.3} 12. Nf3 {+0.20/12 0.3} d3 {+0.63/17 0.3} 13. Qa4
{+0.20/15 2.6} O-O-O {+0.10/16 0.3} 14. Rc1 {+0.00/15 1.3} Ne4
{+0.37/18 0.3} 15. Ne5 {-0.24/16 1.0} Nxe5 {+0.06/21 0.4} 16. Bxe5
{-0.24/16 0.3} Qxa4 {+0.18/23 0.3} 17. Rxc7+ {-0.19/17 0.6} Kb8
{+0.09/22 0.4} 18. Rc4+ {+0.00/16 0.5} Ka8 {+0.06/23 0.4} 19. Rxa4
{-0.21/17 0.8} Rd5 {+0.15/24 0.4} 20. Bb2 {-0.18/21 2.5} Rc8 {+0.49/18 0.4}
21. f3 {-0.20/18 0.3} Nxd2 {+0.61/20 0.4} 22. e4 {-0.20/21 0.5} Rb5
{+0.50/19 0.4} 23. Rb4 {-0.20/19 0.6} Rxb4 {+0.66/22 0.4} 24. axb4
{-0.02/22 0.7} Rc2 {+0.77/22 0.4} 25. Bxd3 {-0.07/21 1.0} Rxb2
{+0.41/20 0.4} 26. exf5 {+0.00/17 0.6} Rxa2 {+0.50/22 0.4} 27. Rg1
{-0.06/21 3} Rb2 {+0.51/21 0.4} 28. f6 {-0.09/18 0.8} gxf6 {+0.66/19 0.5}
29. Be2 {+0.00/20 1.5} Nb3 {+1.29/22 0.5} 30. Kf2 {+0.00/20 0.6} Nd4
{+1.66/19 0.5} 31. Rd1 {+0.00/18 1.0} Rxb4 {+1.61/20 0.4} 32. Bd3
{+0.12/18 1.6} Rb2+ {+1.65/19 0.5} 33. Ke3 {+0.19/18 0.3} Ne6
{+1.84/18 0.5} 34. Rb1 {-0.03/19 1.7} f5 {+2.18/19 0.4} 35. Rxb2
{+3.61/17 0.3} f4+ {-4.35/16 0.4} 36. Ke4 {+3.81/17 0.6} Nc5+
{-4.31/15 0.3} 37. Kd4 {+3.93/17 0.5} b6 {-4.46/15 0.3} 38. Bc4
{+3.99/19 0.8} Kb7 {-4.83/15 0.3} 39. Re2 {+4.60/19 0.8} Kc6 {-4.94/16 0.3}
40. Re7 {+4.79/18 0.7} Kd6 {-5.06/17 0.3} 41. Rxf7 {+4.99/18 0.8} Nd7
{-5.26/15 0.4} 42. Rxf4 {+5.62/18 0.8} a5 {-5.53/17 0.4} 43. Rxh4
{+5.98/20 0.9} a4 {-5.86/16 0.4} 44. Rh6+ {+6.45/16 0.6} Ke7 {-6.68/16 0.4}
45. Re6+ {+7.24/17 0.8} Kf8 {-6.64/15 0.4} 46. h4 {+25.33/20 0.7} a3
{-26.07/15 0.3} 47. h5 {+43.82/22 0.9} b5 {-33.92/17 0.3} 48. h6
{+99.92/22 1.2} Nb8 {-72.70/17 0.4} 49. Rf6+ {+1000.12/22 0.8} Ke7
{-99.93/20 0.4} 50. h7 {+1000.10/25 0.7} Kxf6 {-99.95/20 0.5} 51. h8=Q+
{+1000.09/27 0.9} Ke7 {-99.95/21 0.5} 52. Qe5+ {+1000.08/27 0.8} Kd7
{-99.95/20 0.5} 53. Bxb5+ {+1000.07/29 0.8} Kc8 {-99.96/21 0.5} 54. Qe8+
{+1000.06/33 0.7} Kc7 {-99.96/18 0.5} 55. Kc5 {+1000.05/48 0.8} Kb7
{-1000.04/52 0.3} 56. Qe7+ {+1000.04/186 0.7} Kc8 {-1000.03/251 0.9} 57.
Bc6 {+1000.03/255 0.4} Na6+ {-1000.02/255 0.4} 58. Kd6 {+1000.02/255 0.4}
a2 {-1000.01/250 0.6} 59. Qf8# {+1000.01/255 0.4}
{Xboard adjudication: Checkmate} 1-0[/pgn]
What was the node limitation handicap for Honey here (even with nodes limited this move seems inexplicable)?
Black played 34...f5??? and the game was over
[d]k7/pp3p2/4np2/8/7p/3BKP1P/1r4P1/1R6 b - - 5 34
13 +1.90 122673 0:00.09 Rxb1 Bxb1 Kb8 Ke4 a5 Kf5 a4 Ba2 b5 Kxf6 Kc7 Ke5 Kb6 f4 Nd8 Kf6
13 +1.90 122673 0:00.09 Rxb1 Bxb1 Kb8 Ke4 a5 Kf5 a4 Ba2 b5 Kxf6 Kc7 Ke5 Kb6 f4 Nd8 Kf6
12 +1.90 93918 0:00.07 Rxb1 Bxb1 Kb8 Ke4 a5 Kf5 a4 Ba2 b5 Kxf6 Kc7 Ke5 Kb6 f4 Nd8 Kf6
11 +2.78 25977 0:00.02 Rxb1 Bxb1 b5 Ke4 Kb7 Kf5 a5 Kxf6 Nf4
10 +2.16 14584 0:00.01 Rxb1 Bxb1 b5 Ke4 Ng7 Ba2 Kb7 Bxf7 Kb6
9 +1.85 8519 0:00.01 Rxb1 Bxb1 Kb8 Ke4 Ng7 Ba2 a5 Bxf7 Kc7
8 +1.62 7939 0:00.01 Rxb1 Bxb1 b5 Ke4 Ng7 Ba2 Kb7 Bxf7 Kc6 Kf4
7 +2.40 1815 0:00.00 Rxb1 Bxb1 Kb8 g4 Kc7 Be4 Kd6 Bxb7
6 +3.33 986 0:00.00 Rxb1 Bxb1 Kb8 g4 hxg3 h4 Kc7
5 +3.45 262 0:00.00 Rxb1 Bxb1 a5 f4 Kb8
4 +3.22 176 0:00.00 Rxb1 Bxb1 a5 f4 Kb8
3 +3.56 109 0:00.00 Rxb1 Bxb1 a5 f4
2 +2.59 64 0:00.00 Rxg2 f4
1 NNUE evaluation using eval.bin enabled
1 +2.29 28 0:00.00 Rxg2