« Back to top page

Dynamic

Dynamic

Multivariate TPE Sampler that considers dynamic value ranges

Abstract The original motivation can be found here. HPO is often iterative: a range that looked reasonable at the start can turn out to be too narrow once a few trials complete. For example, if reg_alpha for an XGBoost model is first searched over [0, 1] and the best trials keep landing near 0.99, the natural next step is to widen it to, say, [0, 2] and keep optimizing on the same study.