Improving GP Hyperparameter Fitting for Bayesian Optimization with a Goal-Oriented Criterion

Bobby Huggins, Roman Garnett

Published: 29 Aug 2025, Last Modified: 29 Aug 2025AutoML 2025 Non-Archival Content TrackEveryoneRevisionsBibTeXCC BY 4.0
Submission Type: Short paper
Tldr: We introduce the Probability Density of Improvement (PDI) criterion function for GP surrogate hyperparameters and show that it achieves a statistically significant improvement over traditional Type-II MLE on a benchmark set of Bayesian optimization tasks.
Submission Number: 31