HROCH 1.0.11

Symbolic regression

HROCH

High optimized symbolic regression.

• Zero hyperparameter tunning.
• Accurate results in seconds or minutes, in contrast to slow GP-based methods.
• Small models size.
• Good results with noise data.
• Support mathematic equations and fuzzy logic operators.
• Support 32 and 64 bit floating point arithmetic.
• Work with unprotected version of math operators (log, sqrt, division)
• CLI

Dependencies

• AVX2 instructions set(all modern CPU support this)
• numpy

Installation

pip install HROCH

Usage

from HROCH import PHCRegressor

reg = PHCRegressor(numThreads=8, timeLimit=60.0, problem='math', precision='f64')
reg.fit(X_train, y_train)
yp = reg.predict(X_test)
# print symbolic expression
print(reg.sexpr)

Performance

Feynman dataset

Approximate comparison with methods tested in srbench.

Algorithm	Training time (s)	Model size	R2 > 0.999	R2 > 0.999999	R2 > 0.999999999	R2 mean
MRGP	14893	3177	0.931	0.000	0.000	0.998853549755939
Operon	2093	70	0.862	0.655	0.392	0.990832974928022
AIFeynman	26822	121	0.785	0.689	0.680	0.923670858619585
HROCH	42	17	0.781	0.679	0.633	0.988862822072670
SBP-GP	28944	487	0.737	0.388	0.246	0.994645420032544
GP-GOMEA	3677	34	0.716	0.539	0.504	0.996850949284431
AFP_FE	17682	41	0.591	0.315	0.185	0.985876419645066
EPLEX	10599	56	0.470	0.121	0.082	0.991763792716299
AFP	2895	37	0.448	0.263	0.159	0.968488776363814
FEAT	1561	195	0.397	0.121	0.112	0.932465581448533
gplearn	3716	78	0.328	0.151	0.151	0.901020570640627
ITEA	1435	21	0.276	0.233	0.224	0.911713461958873
DSR	615	15	0.250	0.207	0.207	0.875784840006460
BSR	28800	25	0.108	0.073	0.043	0.693995349495648
FFX	19	268	0.000	0.000	0.000	0.908164756903951