SFNNUE Fails!! In this Positions!!

[Master Om]

Ranked second on Pohl's list:
https://www.sp-cc.de/files/ergebnisse_nnue_schnell.txt

With the latest SFNNUE it fails with that net too to find the key move. Something wrong in new SFNNUE.

[peter]

Ranked second on Pohl's list:
https://www.sp-cc.de/files/ergebnisse_nnue_schnell.txt

With the latest SFNNUE it fails with that net too to find the key move. Something wrong in new SFNNUE.

It's the hybrid- eval- üatch from end of July, not the net.
Try one of July, if not the very latest one, or one of Jörg Oster's NNUE-always versions or Cfish NNUE or latest ShashChess with "Pure" NNUE- Eval or dkappe's NetRunner with the scaleable treshold- parameters (all to be found on github).
Guess SF- developers should think about a choice like that in UCI options too, if SF is to be developed as an analyis tool onwards too and not as a toy for engine- engine game- playing only.

[Dann Corbit]

Fails to solve x, but solves y,z,....

we must be careful to condemn based on an observation.
We can laugh at science when it fails to predict the passage of mercury, but still predicts clearly all the other planets, and meanwhile the missed position fades into obscurity because it never happens again

IOW, an error is not important unless it matters in practice.
Sure, we want 100 % solutions, but we know that is not possible.
So, let's hope for 99%

[peter]

Fails to solve x, but solves y,z,....

Of course you're right, Dann.
There's positions enough, where classical SF- eval solves better (faster) than NNUE- eval does.

If I had to choose for a single one eval, I'd probably take some kind of hybrid too, but not with so low treshold- boundaries and if I can't have a scaling per UCI, like I have at dkappe's NetRunner e.g., I'd be happy with NNUE pure or NNUE hybrid as a UCI- choice. As a matter of fact in interactive analysis both helps in many positions. Switching doesn't clear the hash (or at leasts it doesn't have to), so trying both with Forward- Backward is a good thing often.

[zullil]

Not sure "fails" is the right word here. For example, in the first posted position, it seems that Bxa7 is hardly failing:

info depth 54 seldepth 69 multipv 1 score cp 548 nodes 85074237208 nps 30965621 hashfull 1000 tbhits 3333976 time 2747377 pv b8a7 c6e4 d3e4 e6e5 f1a1 b5b4 f3e3 b7b5 b3a4 b4c4 g2f1 c4c2 a4b5 c5c4 e3e2 c2c3 a1a2 d8c8 b5b6 c3b4 a2b2 b4a5 b6b7 a5a7 b7c8n e7c8 e2c4 c8d6 c4d3 f6e7 h1g2 a7a8 f2f3 h7h5 d3c3 e7f6 b2d2 d6c8 h2h4 c8e7 f1c4 a8a7 c4d5 e7d5 d2d5 a7a6 c3b2 a6c4 d5d7 c4a4 b2d2 a4a6 d2c2 a6e6 d7d2 e6b6 c2a2 b6c7 d2d5 f6e7 a2d2 e7a3 d5d7 c7c4

[Master Om]

Ranked second on Pohl's list:
https://www.sp-cc.de/files/ergebnisse_nnue_schnell.txt

With the latest SFNNUE it fails with that net too to find the key move. Something wrong in new SFNNUE.

It's the hybrid- eval- üatch from end of July, not the net.
Try one of July, if not the very latest one, or one of Jörg Oster's NNUE-always versions or Cfish NNUE or latest ShashChess with "Pure" NNUE- Eval or dkappe's NetRunner with the scaleable treshold- parameters (all to be found on github).
Guess SF- developers should think about a choice like that in UCI options too, if SF is to be developed as an analyis tool onwards too and not as a toy for engine- engine game- playing only.

I agree. This Hybrid patch is one that shifts from classical eval and NN eval ?
Also can u provide the links u mentioned above ?

[Master Om]

Not sure "fails" is the right word here. For example, in the first posted position, it seems that Bxa7 is hardly failing:

info depth 54 seldepth 69 multipv 1 score cp 548 nodes 85074237208 nps 30965621 hashfull 1000 tbhits 3333976 time 2747377 pv b8a7 c6e4 d3e4 e6e5 f1a1 b5b4 f3e3 b7b5 b3a4 b4c4 g2f1 c4c2 a4b5 c5c4 e3e2 c2c3 a1a2 d8c8 b5b6 c3b4 a2b2 b4a5 b6b7 a5a7 b7c8n e7c8 e2c4 c8d6 c4d3 f6e7 h1g2 a7a8 f2f3 h7h5 d3c3 e7f6 b2d2 d6c8 h2h4 c8e7 f1c4 a8a7 c4d5 e7d5 d2d5 a7a6 c3b2 a6c4 d5d7 c4a4 b2d2 a4a6 d2c2 a6e6 d7d2 e6b6 c2a2 b6c7 d2d5 f6e7 a2d2 e7a3 d5d7 c7c4

Qxf6 is clearly better move.
The Nodchip last 19/7/2020 one is fast in finding it but not the 11/8/2020 one.

Code: Select all

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[FEN "1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1"]
[SetUp "1"]

1. Qxf6+ Kxf6 2. Be5+ Kg5 3. Bg7 Bxe4 4. f4+ Kh5 5. Bxe4 g5 (5... Qc6 6. g4+
Kh4 7. Bf6+ g5 8. Ne5 h6 9. fxg5 Ng8 10. Bxd8 Qd6 11. gxh6+ Qxd8 12. Rg1 Kg5
13. Nxf7+ Kh4 14. h7 Nf6 15. Nxd8 Nxh7 16. Bxh7 {(0:01:24) 657778kN}{[%eval
1259,35,Stockfish+NNUE x64][%meval TB:1k]} 16... Nc3 {(0:01:21) 634334kN}
{[%eval 1223,35,Stockfish+NNUE x64][%meval TB:1k]} 17. Nxe6 {(0:01:21)
633151kN}{[%eval 1197,35,Stockfish+NNUE x64][%meval TB:1k]} 17... Ne2
{(0:01:20) 631619kN}{[%eval 1177,35,Stockfish+NNUE x64][%meval TB:1k]}) 6. Ne5
Qc6 7. g4+ Kh4 8. Bf6 h6 9. fxg5 Ng8 10. Bxd8 Ra8 (10... Qd6 11. gxh6+ Qxd8
12. Rg1 Kg5 13. Nxf7+ Kf4 14. Nxd8 Nxh6 15. bxa4 c4 (15... cxb4 16. Nxe6+ Ke3
17. g5 Ng8 18. Rg3+ Kd2 19. Kg2 Ne7 20. g6 Kxc2 21. d4+ Kd2 22. d5 b3 23. Rxb3
Rxa4 24. Rxb7 {(0:01:41) 749690kN}{[%eval 1301,36,Stockfish+NNUE x64][%meval
TB:12k]} 24... Nxg6 {(0:01:35) 713304kN}{[%eval 1288,36,Stockfish+NNUE
x64][%meval TB:7k]} 25. Bxg6 {(0:01:34) 710620kN}{[%eval
1279,36,Stockfish+NNUE x64][%meval TB:7k]} 25... Ra2 {(0:01:29) 692503kN}
{[%eval 1288,35,Stockfish+NNUE x64][%meval TB:3k]})16. g5 Ng8 17. Re1 Kxg5 18.
Nxe6+ Kf6 19. Nc7 cxd3 20. Bxd3 Rxa4 21. Nd5+ Kf7 22. Bc4 Ra3 23. Nb6+ Kg6 24.
Bxg8 Rh3 25. Bc4 Rh4 26. Bd3+ Kg7 27. b5 Kf8 28. Re6 {(0:02:53) 1182584kN}
{[%eval 1330,40,Stockfish+NNUE x64][%meval TB:145k]} 28... Rf4 {(0:02:33)
1059143kN}{[%eval 1321,39,Stockfish+NNUE x64][%meval TB:98k]}) 11. Bxc6 Nc3
(11... Rxd8 12. Rg1 Kxg5 13. Nxf7+ Kf6 14. Nxd8 Nc3 15. Bxb7 cxb4 16. Re1 Nb5
17. Rxe6+ Kg7 18. Nc6 Nf6 19. h3 Kf7 20. Re1 Nc3 21. Nd4 Nh7 22. Rf1+ Ke7 23.
Ba6 Nd5 24. Nf5+ Kd8 25. Bc4 Ndf6 {(0:02:26) 1022963kN}{[%eval
1322,39,Stockfish+NNUE x64][%meval TB:83k]} 26. Ne3 {(0:02:26) 1022377kN}
{[%eval 1313,39,Stockfish+NNUE x64][%meval TB:83k]} 26... Nd7 {(0:02:21)
993444kN}{[%eval 1322,38,Stockfish+NNUE x64][%meval TB:71k]}) 12. Nxf7 bxc6
(12... Rxd8 13. Nxd8 bxc6 14. g6 Kg5 (14... Kxg4 15. Nxc6 Kg5 16. Rf8 Kxg6 17.
Rxg8+ Kf7 18. Rg1 cxb4 19. Rf1+ Ke8 20. Nxb4 Ke7 21. Nc6+ Kd6 22. Nd4 Ke5 23.
Nf3+ Kf4 24. Nd2+ Ke3 25. Nc4+ Ke2 26. Rf6 Nd1 27. Rxh6 Kf1 28. Rf6+ Nf2+ 29.
Rxf2+ Kxf2 {(0:02:03) 874492kN}{[%eval 1322,37,Stockfish+NNUE x64][%meval
TB:41k]} 30. c3 {(0:01:59) 851324kN}{[%eval 1309,37,Stockfish+NNUE x64][%meval
TB:38k]} 30... Kf3 {(0:01:54) 818251kN}{[%eval 1300,37,Stockfish+NNUE
x64][%meval TB:31k]} 31. Kg1 {(0:01:46) 776845kN}{[%eval
1290,36,Stockfish+NNUE x64][%meval TB:16k]})15. Rf8 Kxg6 16. Rxg8+ Kf6 17. Re8
Ke5 18. Rxe6+ Kd5 19. bxc5 Na2 20. Re4 Nc3 21. Re1 Kxc5 22. b4+ Kd5 23. Ne6
Na2 24. c4+ Kd6 25. b5 Nb4 26. bxc6 Kxc6 27. g5 hxg5 28. Nd4+ Kd6 29. Nb3
{(0:02:20) 987469kN}{[%eval 1327,38,Stockfish+NNUE x64][%meval TB:69k]} 29...
Nxd3 {(0:02:08) 909997kN}{[%eval 1317,37,Stockfish+NNUE x64][%meval TB:48k]})
13. gxh6+ Rxd8 14. h7 Nf6 15. Rxf6 Ra8 16. g5 Nd5 17. Rf1 Kh5 18. h8=Q+ Rxh8
19. Nxh8 Kxg5 20. Nf7+ Kg6 21. bxc5 Kg7 22. Rf3 Nb4 23. Ne5 Kh7 24. Rf1 Kg7
25. Rf7+ Kh6 26. Rf6+ Kg7 27. Rxe6 Nd5 28. Nxc6 Nf4 29. Nd4 Nxe6 30. Nxe6+ Kf7
{(0:03:10) 1288904kN}{[%eval 1338,40,Stockfish+NNUE x64][%meval TB:186k]} *

eval is +13.50

[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "?"]
[Black "?"]
[Result "*"]
[Analysis "Stockfish+NNUE x64"]
[AnalysisTime "322"]
[FEN "1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1"]
[SetUp "1"]

1. Qxf6+ Kxf6 2. Be5+ Kg5 3. Bg7 Bxe4 4. f4+ Kh5 5. Bxe4 g5 6. Ne5 Qc6 7. g4+
Kh4 8. Bf6 h6 9. fxg5 Ng8 10. Bxd8 Qd6 11. gxh6+ Qxd8 12. Rg1 Kg5 13. Nxf7+
Kf4 14. Nxd8 Nxh6 15. bxa4 Ra8 16. g5 Ng8 17. Nxe6+ Ke3 18. Nc7 Rc8 19. Nd5+
Kd4 20. c3+ Ke5 21. bxc5 Rxc5 22. g6 Rxd5 23. Rg5+ Kf4 24. Bxd5 Ne7 25. g7
Kxg5 26. g8=Q+ Nxg8 27. Bxg8 Kf6 28. a5 Ke7 29. Bc4 Kd7 30. a6 Kc8 31. a7 Kc7
32. a8=Q Kd7 {(0:05:22) 2248323kN}{[%eval 1342,42,Stockfish+NNUE x64][%meval
322s|TB:656k]} *

[peter]

Also can u provide the links u mentioned above ?

https://github.com/joergoster/Stockfish ... nue_always

https://github.com/syzygy1/Cfish/

Of Cfish you'll have to build your own compile of the latest version, releases are from July only.

https://github.com/amchess/ShashChess/releases/tag/13.1

https://github.com/dkappe/NetRunner/releases/tag/Nr0.1

[zullil]

Qxf6 is clearly better move.

Can't imagine the engine would ever lose or draw after Bxa7.

Though Qxf6 might save some energy, I suppose.

[Master Om]

Qxf6 is clearly better move.

Can't imagine the engine would ever lose or draw after Bxa7.

Though Qxf6 might save some energy, I suppose.

Thats why we called it "best" not good or better