Genetic programming that speaks scikit-learn natively

What it does

gplearn performs symbolic regression: it evolves populations of random mathematical formulas to find an expression that best fits your data. It wraps this genetic programming machinery in scikit-learn-compatible estimators—SymbolicRegressor, SymbolicClassifier, and SymbolicTransformer—so you can drop them into pipelines and grid searches. The project deliberately narrows the broad field of genetic programming to this single, practical task.

The interesting bit

The value is in the constraint. Rather than building a sprawling GP framework, gplearn treats evolved formulas as just another scikit-learn estimator, complete with fit and predict. That makes automated feature engineering via genetic programming feel like swapping out a linear model.

Key highlights

• Implements SymbolicRegressor, SymbolicClassifier, and SymbolicTransformer for regression, binary classification, and feature engineering.
• Works natively with scikit-learn Pipeline and grid search modules.
• Evolves formulas through successive generations of selection and genetic operations.
• Exposes many tunable parameters, though the documentation claims to clarify the relevant ones.

Caveats

• The README notes there are “a lot of parameters to tweak,” which suggests a learning curve.
• Classification is explicitly limited to binary problems.
• Requires a fairly recent scikit-learn version, which may complicate legacy environments.

Verdict

Worth a look if you want interpretable, evolved equations inside a standard scikit-learn workflow. Skip it if you need multiclass classification or a general-purpose genetic programming toolkit.