janopig/hroch

Extremly fast c++/python symbolic regression library based on parallel local search.

Symbolic regression and classification library

High-Performance python symbolic regression library based on parallel local search

• Zero hyperparameter tunning.
• Accurate results in seconds or minutes, in contrast to slow GP-based methods.
• Small models size.
• Support for regression, classification and fuzzy math.
• Support 32 and 64 bit floating point arithmetic.
• Work with unprotected version of math operators (log, sqrt, division)
• Speedup search by using feature importances computed from bbox model

Supported instructions
math	add, sub, mul, div, pdiv, inv, minv, sq2, pow, exp, log, sqrt, cbrt, aq
goniometric	sin, cos, tan, asin, acos, atan, sinh, cosh, tanh
other	nop, max, min, abs, floor, ceil, lt, gt, lte, gte
fuzzy	f_and, f_or, f_xor, f_impl, f_not, f_nand, f_nor, f_nxor, f_nimpl

Sources

C++20 source code available in separate repo sr_core

Dependencies

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

Installation

pip install HROCH

Usage

Symbolic_Regression_Demo.ipynb

Documentation

from HROCH import SymbolicRegressor

reg = SymbolicRegressor(num_threads=8, time_limit=60.0, problem='math', precision='f64')
reg.fit(X_train, y_train)
yp = reg.predict(X_test)

Changelog

v1.4

• sklearn compatibility
• Classificators:
  • NonlinearLogisticRegressor for a binary classification
  • SymbolicClassifier for multiclass classification
  • FuzzyRegressor for a special binary classification

Older versions

v1.3

• Public c++ sources
• Commanline interface changed to cpython
• Support for classification score logloss and accuracy
• Support for final transformations:
  • ordinal regression
  • logistic function
  • clipping
• Acess to equations from all paralel hillclimbers
• User defined constants

v1.2

• Features probability as input parameter
• Custom instructions set
• Parallel hilclimbing parameters

v1.1

• Improved late acceptance hillclimbing

v1.0

• First release

SRBench

full results

hroch's Issues

Separate Fuzzy regression and classification

Add unittests

add tests for scikit-learn compatibility, random state and other

remove ci_test

Access all search equations from Python code

v 1.3 Return equations from all parallel hillclimbers.

Add finetuning step

By scipy.optimize or pytorch as new gpgomea.

https://github.com/users/janoPig/projects/4?pane=issue&itemId=40911391

improve fuzzy regression

Add nop to default 'fuzzy'
fix wrong initialization

Fix unittests

<test.features_test.TestFeatures testMethod=test_random_state_mt>

random fail on windows. After adding the debug logs, the problem did not occur. Interesting.

Add binary classificator

Add a binary classifier compatible with sklearn as a separate class.

Implement sample_weight

janoPig/sr_core#2

feature names

UserWarning: X has feature names, but NonlinearLogisticRegressor was fitted without feature names

Implement tree representation of equations

Const search space

v 1.3 Adding your own constant search space. The const_size parameter is now available to control the number of constants to use when searching hc. Added the ability to use a set of predefined immutable constants by the search process, even in combination with mutable constants.

Write documentation

Add srbench to CI

Add a short test (1s per sample), return a few relevant parameters, remove Strogatz and Feynman tests, keep only for windows.