We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Statistics > Methodology

Title: On the optimization of hyperparameters in Gaussian process regression with the help of low-order high-dimensional model representation

Abstract: When the data are sparse, optimization of hyperparameters of the kernel in Gaussian process regression by the commonly used maximum likelihood estimation (MLE) criterion often leads to overfitting. We show that choosing hyperparameters (in this case, kernel length parameter and regularization parameter) based on a criterion of the completeness of the basis in the corresponding linear regression problem is superior to MLE. We show that this is facilitated by the use of high-dimensional model representation (HDMR) whereby a low-order HDMR representation can provide reliable reference functions and large synthetic test data sets needed for basis parameter optimization even when the original data are few.
Comments: 16 pages, 2 figures, 2 tables
Subjects: Methodology (stat.ME); Numerical Analysis (math.NA)
Journal reference: Journal of Mathematical Chemistry, 61, 7-20 (2023)
DOI: 10.1007/s10910-022-01407-x
Cite as: arXiv:2112.01374 [stat.ME]
  (or arXiv:2112.01374v2 [stat.ME] for this version)

Submission history

From: Sergei Manzhos [view email]
[v1] Tue, 30 Nov 2021 08:17:44 GMT (580kb)
[v2] Thu, 6 Jan 2022 00:32:46 GMT (597kb)

Link back to: arXiv, form interface, contact.