How to calculate piece weights with logistic regression?

[mar]

In my eyes, Reinventing the wheel is the only way to understand the wheel.

exactly, not only is deep understanding helpful but it also allows you to build a wheel that better fits your needs

[Tony P.]

Meh. I consider debugging others' wheels (and then writing and swatting new bugs for them) another good way to learn wheel construction. There are so many wheels in existence that there's one that suits the needs well. It just takes time to find. However, I also understand those who don't bother to make others' wheels out.

[Ferdy]

Put up a repo on github for piece value estimator. It uses different linear models from sklearn lib.

Sample output

Code: Select all

                                                       fen  P-p  ...  Q-q  result
0        r3r3/pbp1kp2/2p2bpp/2N1p3/8/5B2/PPP2PPP/2KRR3 ...   -1  ...    0       1
1        3rr1k1/pb4pp/2p1n3/1p1n4/1P1qN3/P7/B1Q2PPP/3R1...    0  ...    0       0
2          6k1/p1p3p1/1p5p/4p3/2P2n2/2P2B2/P4P1P/6K1 w - -   -1  ...    0       0
3              6k1/6p1/2p5/p1q4p/2P1n3/p4RNP/R5P1/7K b - -   -2  ...   -1       0
4          4rk2/1R5p/6p1/r2P1p2/p1PpB3/5P2/P5PP/4K2R w K -    1  ...    0       1
...                                                    ...  ...  ...  ...     ...
1206278                 8/pp6/2pk4/5R2/1P6/r7/5PK1/8 w - -   -1  ...    0       0
1206279                      1r6/8/8/8/R6p/6nk/5K2/8 w - -   -1  ...    0       0
1206280                    7R/6p1/6k1/4Kp2/6rP/8/8/8 w - -   -1  ...    0       0
1206281    1r2k3/1q3p1p/p3pQ2/8/2pR4/5Pr1/1PP3P1/R6K b - -   -1  ...    0       1
1206282                 8/8/8/4p1p1/1K1kP1P1/5P2/8/8 b - -    1  ...    0       0

[1206283 rows x 7 columns]

Features that are not 0:
=======================

piece: P-p, num: 579322, pct: 48.03%
piece: N-n, num: 254340, pct: 21.08%
piece: B-b, num: 251468, pct: 20.85%
piece: R-r, num: 117705, pct: 9.76%
piece: Q-q, num: 24378, pct: 2.02%


model 1: Linear Regression
=======================

Metrics:
mse: 0.16410313874489252
mae: 0.3518987466042799
r2_score: 0.3396676182349635

coefficients: [0.17450496 0.33598911 0.37350445 0.52490821 0.96500647]
pawn: 175, knight: 336, bishop: 374, rook: 525, queen: 965


model 2: Ridge
=======================

Metrics:
mse: 0.16410313539039326
mae: 0.35189949045291846
r2_score: 0.33966763173308734

coefficients: [0.1745041  0.33597642 0.37349147 0.5248878  0.96495136]
pawn: 175, knight: 336, bishop: 373, rook: 525, queen: 965


model 3: Lasso
=======================

Metrics:
mse: 0.16410447794342517
mae: 0.35203124553077647
r2_score: 0.33966222945225355

coefficients: [0.17432736 0.33381853 0.37129856 0.52166285 0.95753882]
pawn: 174, knight: 334, bishop: 371, rook: 522, queen: 958


model 4: Stochastic Gradient Descent
=======================

Metrics:
mse: 0.16415350143193788
mae: 0.35034513471558837
r2_score: 0.3394649645054669

coefficients: [0.17870765 0.33873901 0.37098609 0.53000329 0.96326733]
pawn: 179, knight: 339, bishop: 371, rook: 530, queen: 963


References:
mse      : mean squared error
mae      : mean absolute error
r2_score : coefficient of determination