BayesHalvingSearchCV

Date: Aug 14, 2026 Version: 0.1.1

Useful links: Source Repository | Issues & Ideas

Two scikit-learn-compatible hyperparameter search estimators, sharing one multi-fidelity “bullseye” data-growth mechanism and one scatter-search multi-start layer:

• PatternSearchCV — Hooke-Jeeves pattern search.

• BayesHalvingSearchCV — a from-scratch Gaussian Process + Expected Improvement Bayesian search (no Optuna, no torch), on the exact same multi-fidelity infrastructure.

Both start on a small, representative subsample of the training data and buy more data as their own search trajectory shows them converging on an optimum, confirming every reported result only on full data.

Getting started

Information on installing the package and a first working example — see Getting started.

API reference

Every parameter each estimator accepts, generated from the classes’ own NumPyDoc docstrings — see API Reference.