How to scale stockfish NNUE score?

[maksimKorzh]

Hi guys, thanks to amazing library to probe stockfish NNUE by Daniel Shawul even a noob like me can now get NNUE score for each position, but it turns out that using that score directly leads to a VERY STRANGE behavior of my engine = it makes strange moves from time to time and sacrifices the queen for nothing in every game. I feel a bit confused at this point.

My questions are:
1. Do I need to evaluate material if I want to use pure NNUE eval?
2. Do I need to scale NNUE eval?
3. Do I need to tune material weights? (I heard SF uses around 200 for paen value while my engine does around 100)

Actually what should I do with NNUE score to make my engine play normal chess?

Please help.

[maksimKorzh]

Well, it seems like I've managed to fix this issue:

1. Scaled NNUE score like in PURE NNUE in CFish

Code: Select all

int score = eval_nnue(side, pieces, squares) * 5 / 4;

2. Assuming SF's relative pawn score is around 200 points (while my engine used around 100 points, depending on game phase) I've multiplied my piece weights by 2:

Code: Select all

// get opening/endgame material score
score_opening += material_score[opening][piece] * 2;
score_endgame += material_score[endgame][piece] * 2;

3. Finally I've added NNUE score to interpolated material score:

Code: Select all

// return final evaluation based on side
score += evaluate_nnue();

This resulted the following game (I used fixed depth 8 to make sure that search is the same while only eval difers):
[pgn][Event "Computer chess game"]
[Site "maksim-Inspiron-3582"]
[Date "2020.10.17"]
[Round "?"]
[White "BBC 1.2"]
[Black "BBC NNUE"]
[Result "0-1"]
[BlackElo "2200"]
[ECO "B45"]
[Opening "Sicilian"]
[Time "12:49:05"]
[Variation "Taimanov, Four Knights, 6.Nxc6 bxc6 7.e5 Nd5"]
[WhiteElo "2000"]
[Termination "normal"]
[PlyCount "136"]
[WhiteType "program"]
[BlackType "program"]

1. e4 {(e2-e4 d7-d5 e4xd5 Ng8-f6 d2-d4 Nf6xd5 Bf1-b5+ Bc8-d7 Bb5-d3)
+0.11/8 0} c5 {(c7-c5 c2-c3 Ng8-f6 e4-e5 Nf6-d5 Ng1-f3 e7-e6 Bf1-d3)
-0.60/8 2} 2. Nc3 {(Nb1-c3 e7-e6 Ng1-f3 Nb8-c6 Bf1-c4 Bf8-d6 O-O Ng8-f6)
+0.50/8 0} Nc6 {(Nb8-c6 Ng1-f3 Ng8-f6 d2-d4 c5xd4 Nf3xd4 e7-e6 Nd4-b5)
-0.63/8 2} 3. Nf3 {(Ng1-f3 Ng8-f6 Bf1-c4 e7-e6 O-O d7-d5 e4xd5 e6xd5)
+0.51/8 0} Nf6 {(Ng8-f6 d2-d4 c5xd4 Nf3xd4 e7-e6 Bf1-e2 Bf8-b4 Nd4xc6)
-0.26/8 7} 4. d4 {(d2-d4 c5xd4 Nf3xd4 d7-d5 Nd4xc6 b7xc6 Bc1-f4 Bc8-g4)
+0.34/8 1} cxd4 {(c5xd4 Nf3xd4 e7-e6 Bf1-e2 Bf8-b4 Nd4xc6 b7xc6) -0.26/7 3}
5. Nxd4 {(Nf3xd4 e7-e5 Nd4xc6 b7xc6 Bf1-c4 d7-d5 e4xd5 Bc8-g4) +0.57/8 0}
e6 {(e7-e6 Bc1-f4 Bf8-b4 Nd4-b5 O-O a2-a3 Bb4-a5 Qd1-f3) -0.67/8 19} 6.
Nxc6 {(Nd4xc6 b7xc6 Bf1-d3 d7-d5 e4-e5 d5-d4 e5xf6 d4xc3) +0.62/8 0} bxc6
{(b7xc6 e4-e5 Nf6-d5 Nc3xd5 c6xd5 Bf1-d3 Qd8-c7 Bc1-f4) -0.05/8 3} 7. e5
{(e4-e5 Nf6-d5 Nc3xd5 c6xd5 Bc1-e3 Bc8-b7 Bf1-d3 d5-d4) +0.51/8 1} Nd5
{(Nf6-d5 Nc3xd5 c6xd5 Bf1-d3 Qd8-a5+ Bc1-d2 Qa5-b6 O-O Qb6xb2) +0.12/8 2}
8. Nxd5 {(Nc3xd5 c6xd5 Bf1-d3 Bf8-e7 Bc1-e3 Ra8-b8 b2-b3 O-O) +0.64/8 0}
cxd5 {(c6xd5 Bf1-d3 Qd8-a5+ Bc1-d2 Qa5-b6 Ra1-b1 Bc8-a6 O-O Ba6xd3) -0.26/8
4} 9. Bd3 {(Bf1-d3 Bc8-b7 O-O Bf8-c5 Qd1-h5 Ra8-c8 Bc1-g5 Qd8-c7) +0.69/8
0} Qa5+ {(Qd8-a5+ Bc1-d2 Qa5-b6 Ra1-b1 Bc8-a6 O-O Ba6xd3 c2xd3 Bf8-e7)
-0.16/8 6} 10. Bd2 {(Bc1-d2 Qa5-c7 Qd1-h5 Bf8-c5 O-O g7-g6 Qh5-g5 Bc5-d4
Ra1-e1) +0.77/8 1} Qb6 {(Qa5-b6 Ra1-b1 Bc8-a6 O-O Ba6xd3 c2xd3 Qb6-a6
Qd1-b3) +0.13/8 2} 11. Rb1 {(Ra1-b1 Bf8-c5 O-O O-O a2-a4 Bc5-d4 a4-a5
Qb6-c7) +0.70/8 1} Ba6 {(Bc8-a6 O-O Ba6xd3 c2xd3 Bf8-c5 b2-b4 Bc5-d4
Qd1-h5) +0.27/8 3} 12. O-O {(O-O Ba6xd3 c2xd3 Ra8-c8 Qd1-g4 Qb6-a6 d3-d4
Qa6xa2) +0.58/8 0} Bxd3 {(Ba6xd3 c2xd3 Qb6-a6 Qd1-b3 Bf8-e7 d3-d4 O-O
Rf1-c1) +0.48/8 2} 13. cxd3 {(c2xd3 Ra8-c8 Rb1-c1 Rc8xc1 Qd1xc1 Bf8-c5
Bd2-e3 Bc5xe3) +0.47/8 0} Qa6 {(Qb6-a6 Bd2-c3 Bf8-e7 Rb1-a1 O-O d3-d4
Rf8-b8 Qd1-g4) +0.96/8 3} 14. Qb3 {(Qd1-b3 Bf8-e7 Rf1-c1 O-O Rc1-c7 Ra8-b8
Qb3-c2 Rb8-c8) +0.71/8 0} Be7 {(Bf8-e7 Rf1-c1 O-O Rc1-c7 Rf8-b8 Qb3-c2
Rb8-c8 Rb1-d1) +1.35/8 3} 15. Rbc1 {(Rb1-c1 O-O a2-a3 Ra8-b8 Qb3-c3 Rb8-c8
Qc3-d4 Qa6-b6) +0.41/8 0} O-O {(O-O a2-a3 Ra8-b8 Qb3-c3 Rf8-c8 Qc3-d4
Rc8-c6 g2-g3) +1.83/8 2} 16. a3 {(a2-a3 Ra8-b8 Qb3-c3 d7-d6 Bd2-f4 d6xe5
Bf4xe5 Rb8-c8) +0.33/8 0} Rab8 {(Ra8-b8 Qb3-c3 Rf8-c8 Qc3-d4 d7-d6 e5xd6
Be7xd6 g2-g3) +1.85/8 3} 17. Qc3 {(Qb3-c3 Rf8-c8 Qc3-d4 Qa6-b6 Qd4xb6 a7xb6
Bd2-e3 b6-b5) +0.39/8 0} Rfc8 {(Rf8-c8 Qc3-d4 Rc8-c6 b2-b4 Qa6xa3 Rc1-a1
Qa3-b3 Qd4xa7) +1.86/8 5} 18. Qd4 {(Qc3-d4 d7-d6 Bd2-f4 Rc8xc1 Rf1xc1
Rb8-b3 Rc1-d1 d6xe5) +0.33/8 0} Bd8 {(Be7-d8 Rc1xc8 Rb8xc8 Rf1-c1 Rc8xc1+
Bd2xc1 Bd8-b6 Qd4-c3 Qa6-a4) +3.42/8 7} 19. Rxc8 {(Rc1xc8 Rb8xc8 Rf1-c1
Rc8xc1+ Bd2xc1 Bd8-b6 Qd4-c3 d5-d4 Qc3-c2) +0.11/8 0} Rxc8 {(Rb8xc8 Rf1-c1
Rc8xc1+ Bd2xc1 f7-f6 e5xf6 Bd8xf6 Qd4-e3 h7-h6) +1.82/8 1} 20. Rc1 {(Rf1-c1
Rc8xc1+ Bd2xc1 Bd8-b6 Qd4-c3 d5-d4 Qc3-b3 Bb6-c5 Bc1-f4) +0.26/8 0} Rxc1+
{(Rc8xc1+ Bd2xc1 f7-f6 e5xf6 Bd8xf6 Qd4-e3 Qa6-a4 b2-b4) +1.51/7 0} 21.
Bxc1 {(Bd2xc1 Qa6-c8 Qd4-c3 Qc8xc3 b2xc3 Bd8-e7 a3-a4 Be7-c5 Bc1-f4)
-0.00/8 0} f6 {(f7-f6 Bc1-f4 h7-h6 Qd4-c3 Qa6-a4 g2-g3 Qa4-d1+ Kg1-g2
Kg8-h7) +1.63/8 3} 22. Bf4 {(Bc1-f4 Bd8-b6 Qd4-c3 Qa6-a4 Qc3-c1 Kg8-f7
e5xf6 Kf7xf6) +0.62/8 0} Qa5 {(Qa6-a5 Qd4-e3 g7-g5 Bf4-g3 f6-f5 Qe3-c1
f5-f4) +3.23/7 2} 23. Qe3 {(Qd4-e3 Qa5-b5 b2-b4 Bd8-b6 Qe3-g3 Bb6-d4 e5xf6
Bd4xf6) +0.37/8 0} g5 {(g7-g5 Bf4-g3 f6-f5 Qe3-c1 Qa5-c7 Qc1-d2 Qc7-b6)
+1.86/7 0} 24. Bg3 {(Bf4-g3 Qa5-b6 Qe3-c1 Qb6-c6 Qc1-d1 Bd8-e7 Qd1-g4
Qc6-c1+) +0.51/8 0} f5 {(f6-f5 f2-f4 Bd8-b6 d3-d4 Qa5-a4 Bg3-f2 Qa4-d1+
Qe3-e1 Qd1xe1+) +8.01/8 0} 25. f4 {(f2-f4 Qa5-b6 Qe3-f2 g5-g4 a3-a4 Bd8-e7
h2-h3 Qb6xf2+) +0.19/8 0} Bb6 {(Bd8-b6 d3-d4 Qa5-a4 Bg3-f2 Qa4-d1+ Qe3-e1
Qd1xe1+ Bf2xe1 Bb6xd4+) +8.01/7 0} 26. d4 {(d3-d4 Qa5-a4 Kg1-h1 Qa4-d1+
Qe3-e1 Qd1xe1+ Bg3xe1 Bb6xd4 Be1-c3 Bd4xc3) -1.76/8 0} Qa4 {(Qa5-a4 Bg3-f2
Qa4-d1+ Qe3-e1 Qd1xe1+ Bf2xe1 Bb6xd4+ Be1-f2 Bd4xb2 Bf2xa7) +11.75/7 1} 27.
Kf1 {(Kg1-f1 Bb6xd4 Qe3-e2 g5xf4 Bg3-f2 Bd4xf2 Kf1xf2 Qa4-e4) -2.27/8 0}
Bxd4 {(Bb6xd4 Qe3-e2 g5xf4 Bg3-e1 Bd4-e3 Be1-c3 d5-d4 Qe2-h5) +15.42/8 0}
28. Qe2 {(Qe3-e2 g5xf4 Bg3-f2 Bd4xe5 Qe2xe5 Qa4-d1+ Qe5-e1 Qd1xe1+ Kf1xe1
a7-a5) -2.36/8 0} gxf4 {(g5xf4 Bg3-e1 Qa4-c6 b2-b4 Qc6-c7 Qe2-h5 Bd4xe5)
+15.22/7 0} 29. Bf2 {(Bg3-f2 Bd4xf2 Kf1xf2 Qa4-d4+ Kf2-f3 Kg8-f7 h2-h3
a7-a5 h3-h4) -2.31/8 0} Bxf2 {(Bd4xf2 Qe2xf2 Qa4-d1+ Qf2-e1 Qd1xe1+ Kf1xe1
Kg8-f7 Ke1-f2 Kf7-g7 Kf2-f3) +15.28/8 0} 30. Kxf2 {(Kf1xf2 Qa4-d4+ Kf2-f3
Kg8-f7 b2-b4 h7-h5 h2-h3 h5-h4 b4-b5) -2.31/8 0} Kg7 {(Kg8-g7 Qe2-d2 Qa4-e4
Qd2-c3 Qe4-c4 Qc3xc4 d5xc4) +14.36/7 0} 31. Kf3 {(Kf2-f3 Qa4-d4 b2-b4
Kg7-f7 h2-h4 h7-h5 b4-b5 Kf7-e7) -2.20/8 0} Qd4 {(Qa4-d4 h2-h3 Kg7-g6 b2-b4
h7-h5 h3-h4 Qd4-c3+ Kf3xf4 Qc3xa3) +16.80/8 0} 32. b4 {(b2-b4 Kg7-f7 b4-b5
h7-h5 h2-h3 h5-h4 Qe2-e1 Kf7-g6) -2.25/8 0} Kg6 {(Kg7-g6 h2-h4 Kg6-h5
Qe2-e1 a7-a6 a3-a4 Kh5-g6 b4-b5) +17.05/8 1} 33. h4 {(h2-h4 Kg6-h5 Qe2-e1
Qd4-e4+ Qe1xe4 f5xe4+ Kf3xf4 Kh5xh4 g2-g3+ Kh4-h3 b4-b5) -2.60/8 0} a6
{(a7-a6 Qe2-e1 h7-h6 a3-a4 Qd4-c4 a4-a5 Qc4-c2 b4-b5) +16.71/8 0} 34. Qe1
{(Qe2-e1 Kg6-h5 a3-a4 Kh5-h6 b4-b5 Qd4-d3+ Kf3xf4 a6xb5 a4xb5) -2.23/8 0}
Qd3+ {(Qd4-d3+ Kf3xf4 Qd3xa3 Qe1-d2 Qa3-a1 g2-g3 Qa1-f1+ Kf4-e3 h7-h6
Qd2-d4) +18.43/8 0} 35. Kxf4 {(Kf3xf4 Qd3-d4+ Kf4-f3 Qd4-g4+ Kf3-f2 Qg4xh4+
g2-g3 Qh4-e4 Qe1xe4 f5xe4 Kf2-e3 h7-h5) -2.46/8 0} Qxa3 {(Qd3xa3 Qe1-d2
h7-h6 Qd2-d4 Qa3-c1+ Kf4-g3 Qc1-c4 Qd4xc4) +16.79/7 0} 36. Qd2 {(Qe1-d2
Qa3-b3 Qd2-d4 Qb3-c4 Kf4-e3 Qc4xd4+ Ke3xd4 f5-f4 Kd4-d3) -2.30/8 0} Qb3
{(Qa3-b3 Qd2-e1 Qb3-c4+ Kf4-g3 Qc4-e4 Qe1xe4 f5xe4 Kg3-f4 Kg6-h5) +21.53/8
1} 37. Qd4 {(Qd2-d4 Qb3-c2 g2-g3 Qc2-e4+ Qd4xe4 f5xe4 g3-g4 h7-h6 h4-h5+)
-2.37/8 0} Qc4 {(Qb3-c4 Kf4-e3 Qc4xd4+ Ke3xd4 f5-f4 Kd4-c5 Kg6-f5 Kc5-d6
d5-d4) +24.76/8 0} 38. Qxc4 {(Qd4xc4 d5xc4 Kf4-e3 f5-f4+ Ke3xf4 c4-c3 g2-g4
c3-c2 Kf4-e4) -2.30/8 0} dxc4 {(d5xc4 g2-g3 Kg6-h5 Kf4-f3 h7-h6 Kf3-f4
c4-c3 Kf4-e3) +24.97/8 0} 39. g4 {(g2-g4 f5xg4 Kf4xg4 c4-c3 h4-h5+ Kg6-h6
Kg4-h4 c3-c2 Kh4-g4) -3.56/8 0} fxg4 {(f5xg4 Kf4xg4 Kg6-h6 h4-h5 c4-c3
Kg4-h4 c3-c2 Kh4-g4) +32.64/8 0} 40. Kxg4 {(Kf4xg4 c4-c3 h4-h5+ Kg6-h6
b4-b5 a6xb5 Kg4-h4 c3-c2 Kh4-g4) -5.39/8 0} c3 {(c4-c3 b4-b5 a6xb5 h4-h5+
Kg6-h6 Kg4-f4 c3-c2 Kf4-e4 c2-c1Q) +52.97/8 0} 41. b5 {(b4-b5 a6xb5 Kg4-f4
c3-c2 Kf4-e4 c2-c1Q Ke4-d4 Kg6-h6) -12.44/8 0} c2 {(c3-c2 b5xa6 c2-c1Q
h4-h5+ Kg6-h6 a6-a7 Qc1-g5+ Kg4-h3 Qg5-d8 Kh3-g4) +46.51/8 0} 42. h5+
{(h4-h5+ Kg6-h6 b5xa6 c2-c1Q a6-a7 Qc1-a3 a7-a8Q Qa3xa8 Kg4-h4) -10.34/8 0}
Kh6 {(Kg6-h6 b5xa6 c2-c1Q a6-a7 Qc1-h1 Kg4-g3 Qh1-a8 Kg3-h4) +50.24/7 0}
43. bxa6 {(b5xa6 c2-c1Q a6-a7 Qc1-a3 a7-a8R Qa3xa8 Kg4-h4) -10.34/7 0} c1=Q
{(c2-c1Q a6-a7 Qc1-a1 a7-a8Q Qa1xa8 Kg4-f4 Qa8-a2) +51.75/7 0} 44. Kf3
{(Kg4-f3 Qc1-a3+ Kf3-f4 Qa3xa6 Kf4-e3 Kh6xh5 Ke3-e4 d7-d5+ Ke4-e3 Qa6-f1)
-13.60/8 0} Kxh5 {(Qc1-f1+ Kf3-g3 Qf1xa6 Kg3-g4 Qa6-a2 Kg4-h4 Qa2-a3
Kh4-g4) +59.50/7 0} 45. Ke4 {(Kf3-e4 Qc1-c6+ Ke4-f4 Qc6xa6 Kf4-e4 d7-d5+
e5xd6/ep Qa6xd6 Ke4-e3 Qd6-d5) -14.04/8 0} Qa3 {(Kh5-g6 a6-a7 Qc1-a3 Ke4-f4
Qa3xa7 Kf4-f3 Qa7-a2) +59.77/7 0} 46. a7 {(a6-a7 Qa3xa7 Ke4-d3 Qa7-c5
Kd3-e4 Kh5-g4 Ke4-d3 Qc5xe5) -16.62/8 0} Qxa7 {(Qa3xa7 Ke4-f3 Qa7-a2 Kf3-g3
Kh5-h6 Kg3-h4 Kh6-g6) +59.61/7 0} 47. Kd3 {(Ke4-d3 Qa7-a3+ Kd3-e4 Qa3-b2
Ke4-d3 Qb2xe5 Kd3-c4 Qe5-d5+ Kc4-c3 Kh5-h4) -16.94/8 0} Qa6+ {(Qa7-a6+
Kd3-c3 Qa6-b5 Kc3-d4 h7-h6 Kd4-e4 Qb5-d5+ Ke4-f4 Qd5-b3 Kf4-e4) +61.16/8 0}
48. Ke3 {(Kd3-e3 Qa6-b6+ Ke3-e4 Qb6-c5 Ke4-d3 Qc5xe5 Kd3-c4 Qe5-d5+ Kc4-b4
Kh5-h4 Kb4-c3) -16.94/8 0} Qa1 {(Qa6-a1 Ke3-f4 h7-h6 Kf4-g3 Qa1-b2 Kg3-f4
Qb2-b3 Kf4-e4) +61.16/8 0} 49. Ke4 {(Ke3-e4 Kh5-h4 Ke4-e3 Qa1xe5+ Ke3-f3
Qe5-d5+ Kf3-f4 Qd5-d4+ Kf4-f3 Qd4-d3+ Kf3-f4 Kh4-h3) -17.03/8 0} Qb2
{(Qa1-b2 Ke4-f4 Qb2-b3 Kf4-e4 h7-h6 Ke4-f4 Qb3-h3 Kf4-e4) +61.21/8 0} 50.
Ke3 {(Ke4-e3 Qb2xe5+ Ke3-d3 Qe5-d5+ Kd3-e3 Kh5-h4 Ke3-f4 Qd5-d4+ Kf4-f3
Qd4-d3+ Kf3-f4 Kh4-h3) -17.03/8 0} Qxe5+ {(Qb2xe5+ Ke3-f2 Qe5-c5+ Kf2-g2
Qc5-f5 Kg2-h2 Qf5-f4+ Kh2-g2 Qf4-g4+ Kg2-h1 h7-h6 Kh1-h2) +62.33/8 0} 51.
Kd3 {(Ke3-d3 Kh5-h4 Kd3-d2 Qe5-e4 Kd2-c3 Kh4-g3 Kc3-d2 Kg3-f3 Kd2-c3)
-17.24/8 0} Kg6 {(Kh5-g6 Kd3-d2 d7-d6 Kd2-d3 Kg6-g7 Kd3-c4 Qe5-f5 Kc4-b4)
+62.29/8 0} 52. Kc4 {(Kd3-c4 Kg6-f5 Kc4-d3 Qe5-e4+ Kd3-c3 Kf5-f6 Kc3-d2
Kf6-e5 Kd2-c3) -17.32/8 0} Kg7 {(Kg6-g7 Kc4-b3 Qe5-d5+ Kb3-a4 Kg7-f6 Ka4-a3
Qd5-a5+ Ka3-b3 Qa5-a1 Kb3-b4) +62.19/8 0} 53. Kd3 {(Kc4-d3 Qe5-f4 Kd3-c3
Qf4-e3+ Kc3-c4 Kg7-f6 Kc4-b5 Kf6-e5 Kb5-c4) -17.12/8 0} Kf8 {(Kg7-f8 Kd3-c2
h7-h6 Kc2-b3 Kf8-f7 Kb3-b4 d7-d6 Kb4-c4) +62.13/8 0} 54. Kc4 {(Kd3-c4
Qe5-d5+ Kc4-c3 Kf8-e7 Kc3-b4 Qd5-d4+ Kb4-b5 Ke7-d6 Kb5-a6 Qd4-e4) -17.27/8
0} Ke7 {(Kf8-e7 Kc4-b4 Qe5-d5 Kb4-a3 h7-h6 Ka3-b4 Qd5-a2 Kb4-c5) +62.46/8
0} 55. Kd3 {(Kc4-d3 Ke7-d6 Kd3-d2 Qe5-e4 Kd2-c3 Qe4-d5 Kc3-c2 Qd5-e4+)
-17.30/8 0} Kd6 {(Ke7-d6 Kd3-c2 Qe5-c5+ Kc2-b3 Qc5-d5+ Kb3-a4 Qd5-f5 Ka4-b4
Qf5-c2 Kb4-a3) +62.55/8 0} 56. Kd2 {(Kd3-d2 Qe5-e4 Kd2-c3 Kd6-e5 Kc3-b3
Qe4-d3+ Kb3-b4 Ke5-d6 Kb4-a5) -17.35/8 0} Qd4+ {(Qe5-d4+ Kd2-c2 Kd6-e7
Kc2-b3 Qd4-g4 Kb3-b2 Qg4-b4+ Kb2-a1 Qb4-b5 Ka1-a2) +62.71/8 0} 57. Ke2
{(Kd2-e2 Qd4-e4+ Ke2-d2 Kd6-c5 Kd2-d1 Qe4-d4+ Kd1-e2 Kc5-d5 Ke2-f3 Qd4-e4+
Kf3-g3 Kd5-e5) -17.34/8 0} Qf4 {(Qd4-f4 Ke2-d3 Kd6-e7 Kd3-e2 Qf4-h4 Ke2-f3
h7-h6 Kf3-g2) +62.74/8 0} 58. Kd3 {(Ke2-d3 Kd6-d5 Kd3-e2 Qf4-e4+ Ke2-f2
Kd5-d6 Kf2-g3 Kd6-e5 Kg3-h3) -17.34/8 0} Kc7 {(Kd6-c7 Kd3-c3 Qf4-a4 Kc3-d3
Kc7-b7 Kd3-c3 Qa4-b5 Kc3-d2) +62.73/8 0} 59. Kc3 {(Kd3-c3 Qf4-e4 Kc3-d2
Kc7-d6 Kd2-c3 Qe4-d5 Kc3-c2 Qd5-f3) -17.23/8 0} h6 {(h7-h6 Kc3-d3 Kc7-d6
Kd3-c3 Kd6-e7 Kc3-b2 Qf4-e3 Kb2-a1) +62.74/8 0} 60. Kd3 {(Kc3-d3 Kc7-d6
Kd3-c3 Qf4-e4 Kc3-d2 Kd6-e5 Kd2-c3 Ke5-d5) -16.78/8 0} Qb4 {(Qf4-b4 Kd3-e2
Kc7-b7 Ke2-f3 h6-h5 Kf3-g3 Kb7-a8 Kg3-h3) +62.48/8 0} 61. Ke3 {(Kd3-e3
Qb4-g4 Ke3-d3 Kc7-d6 Kd3-e3 Qg4-c4 Ke3-f3 Qc4-d5+) -16.65/8 0} Qc4 {(Qb4-c4
Ke3-f2 Qc4-e4 Kf2-g3 h6-h5 Kg3-h2 Qe4-f3 Kh2-g1) +62.63/8 0} 62. Kf3
{(Ke3-f3 Qc4-d4 Kf3-e2 Qd4-e4+ Ke2-d2 Kc7-d6 Kd2-c3 Qe4-d5 Kc3-c2) -16.75/8
0} h5 {(h6-h5 Kf3-g3 Kc7-d6 Kg3-g2 Kd6-e7 Kg2-g3 Qc4-d4 Kg3-g2) +62.63/8 0}
63. Ke3 {(Kf3-e3 Kc7-d6 Ke3-d2 Kd6-e5 Kd2-e1 Qc4-e4+ Ke1-d1 Qe4-d3+ Kd1-e1
Ke5-d6) -16.62/8 0} Kd6 {(Kc7-d6 Ke3-f3 Qc4-b4 Kf3-e3 Kd6-e7 Ke3-d3 Ke7-f8
Kd3-c2) +62.44/8 0} 64. Kf3 {(Ke3-f3 Kd6-e5 Kf3-f2 Qc4-h4+ Kf2-g2 Qh4-d4
Kg2-g3 Qd4-e4 Kg3-h3) -16.64/8 0} Qc3+ {(Qc4-c3+ Kf3-f4 Qc3-c1+ Kf4-g3
Qc1-d1 Kg3-g2 Kd6-e7 Kg2-g3 Qd1-d2 Kg3-h4) +62.45/8 0} 65. Ke4 {(Kf3-e4
Qc3-e5+ Ke4-d3 Kd6-e7 Kd3-c4 Qe5-d5+ Kc4-c3 Qd5-e4 Kc3-d2 Ke7-d6 Kd2-c3)
-16.58/8 0} Ke7 {(Kd6-e7 Ke4-f4 Qc3-c5 Kf4-e4 Qc5-c1 Ke4-d4 Ke7-f8 Kd4-e5)
+62.19/8 1} 66. Kf4 {(Ke4-f4 Qc3-d4+ Kf4-g3 h5-h4+ Kg3-h3 Ke7-d6 Kh3-g2
Qd4-g4+ Kg2-f2 Qg4-e4 Kf2-g1) -16.70/8 0} Kf6 {(Ke7-f6 Kf4-e4 d7-d5+ Ke4-f4
e6-e5+) +M3/7 0} 67. Ke4 {(Kf4-e4 d7-d5+ Ke4-f4 e6-e5+) -M3/8 0} d5+
{(d7-d5+ Ke4-f4 e6-e5+) +M2/8 0} 68. Kf4 {(Ke4-f4 e6-e5+) -M2/8 0} e5#
{(e6-e5+) +M1/8 0} 0-1
[/pgn]

Can I now say NNUE scores are now scaled correctly?

[hgm]

Let me get this straight: NNUE is just one term of the Stockfish evaluation, and there are other terms?

Is material the only other term (and is it tapered from opening to end-game)?

[Daniel Shawul]

Hi guys, thanks to amazing library to probe stockfish NNUE by Daniel Shawul even a noob like me can now get NNUE score for each position, but it turns out that using that score directly leads to a VERY STRANGE behavior of my engine = it makes strange moves from time to time and sacrifices the queen for nothing in every game. I feel a bit confused at this point.

My questions are:
1. Do I need to evaluate material if I want to use pure NNUE eval?
2. Do I need to scale NNUE eval?
3. Do I need to tune material weights? (I heard SF uses around 200 for paen value while my engine does around 100)

Actually what should I do with NNUE score to make my engine play normal chess?

Please help.

FYI I have now added support for the incremental evaluation of NNUE in the library.
This gives a speed bump of 14 to 18%, so maybe worth it for those in need of the highest performance.
The engine has to keep track of dirtyPiece and Accumulator using the same data structure as what the library use.
Make sure accumulator is cache aligned i.e. 64 bytes.
It maybe possible to move some of the code for incremental evaluation from the engine to the library in the future.

For an example implementation of "NNUE incremental" in an engine, look here
https://github.com/dshawul/Scorpio/comm ... 7b7322bf84
If you want "NNUE without incremental evalution", the implementation is much easier
https://github.com/dshawul/Scorpio/comm ... 94cd4b2c5c

I got more that 400 elos out of my implementation with a shared library.

[maksimKorzh]

Let me get this straight: NNUE is just one term of the Stockfish evaluation, and there are other terms?

Is material the only other term (and is it tapered from opening to end-game)?

Yes, material is the ONLY other term and YES it's tapered from opening to end-game.

[maksimKorzh]

Hi guys, thanks to amazing library to probe stockfish NNUE by Daniel Shawul even a noob like me can now get NNUE score for each position, but it turns out that using that score directly leads to a VERY STRANGE behavior of my engine = it makes strange moves from time to time and sacrifices the queen for nothing in every game. I feel a bit confused at this point.

My questions are:
1. Do I need to evaluate material if I want to use pure NNUE eval?
2. Do I need to scale NNUE eval?
3. Do I need to tune material weights? (I heard SF uses around 200 for paen value while my engine does around 100)

Actually what should I do with NNUE score to make my engine play normal chess?

Please help.

FYI I have now added support for the incremental evaluation of NNUE in the library.
This gives a speed bump of 14 to 18%, so maybe worth it for those in need of the highest performance.
The engine has to keep track of dirtyPiece and Accumulator using the same data structure as what the library use.
Make sure accumulator is cache aligned i.e. 64 bytes.
It maybe possible to move some of the code for incremental evaluation from the engine to the library in the future.

For an example implementation of "NNUE incremental" in an engine, look here
https://github.com/dshawul/Scorpio/comm ... 7b7322bf84
If you want "NNUE without incremental evalution", the implementation is much easier
https://github.com/dshawul/Scorpio/comm ... 94cd4b2c5c

I got more that 400 elos out of my implementation with a shared library.

Thanks Daniel!
I've annoyed you with couple more issues on github...

Can you please clarify one more question - I still can't figure it out from CFish's source:
Should I use pawn material weight relatively equal to 100 or 200
the reason why I'm asking is here:
[url]view-source:https://hxim.github.io/Stockfish-Evaluation-Guide/[/url]
see line 1153
It says

Code: Select all

var PAWNVALUE = 206;

and this confuses me... may be it's specific to stockgish evaluation guide GUI...

[hgm]

OK, thanks. I ask because I am interested in trying out this NNUE technique in a new engine, for Janggi. I know absolutely nothing about that game, but I do have a PGN file with 25000 games in it. And perhaps this is enough to train it, and get somewhat reasonable play.

It could also be interesting to try it for Paco Shako.

[maksimKorzh]

OK, thanks. I ask because I am interested in trying out this NNUE technique in a new engine, for Janggi. I know absolutely nothing about that game, but I do have a PGN file with 25000 games in it. And perhaps this is enough to train it, and get somewhat reasonable play.

It could also be interesting to try it for Paco Shako.

I am always getting excited when you're trying to implement engines for games other but chess. I'm just wondering what motivates you?
I mean usually people developing one engine for years and it's difficult to master chess programming at a decent level...
But you're trying many other games! This is incredible, so what motivates you? Thanks in advance.

[maksimKorzh]

After a consultation with Daniel Shawul and fixing the bugs I can now say for sure:
1. NNUE scores already incorporate material weights in them
2. Daniel's library returns positive score for whites and negative for black, so no need to alter that
3. Scaling is not needed if pure NNUE eval is used, only for hybrid solutions

May be it helps someone

[Daniel Shawul]

Some may want to do scaling to get better compatibility with search prunings such as
futility pruning and razoring, etc. where the margins are tuned probably with your old centi-pawn scores.
Other than that, you don't need scaling unless you use classic evaluation in some way.