Back to search

Virginia Tech

Kernel-Based Metamodeling for Heteroscedastic Simulation: Theory and Distribution-Free Uncertainty Quantification

Abstract

dc:description.abstract

Simulation metamodels approximate expensive stochastic simulators for downstream analysis and decision-making. They can become unreliable under strong heteroscedasticity, large or high-dimensional datasets, and tight simulation budgets for uncertainty quantification. This dissertation develops scalable prediction, convergence theory, and distribution-free uncertainty quantification. Nested heteroscedastic Gaussian process (NHGP) provides a scalable divide-and-conquer scheme: each disjoint subset fits a sub-stochastic-kriging model, and the fits are fused into one metamodel. Theoretical results show that the NHGP rule is the best linear unbiased predictor among such aggregations, prove consistency, and demonstrate empirical performance comparable to a benchmark method. A second thread studies stochastic-kriging prediction error under kernel misspecification and growing dimension: it characterizes near-optimal regimes, explains how noise-variance scaling governs worst-case performance, and proves high-probability pointwise bounds whose dimension dependence appears only through a log n factor. The third methodological piece is heteroscedastic weighted kernel ridge regression, which uses precision-based weights for input-dependent noise. Using the equivalence between Gaussian-process prediction and kernel ridge regression, we derive sharp mean squared error rates across kernel spectrum decay regimes, identify rate-optimal regularization, and validate the theory empirically. Finally, we combine conditional kernel mean embeddings with distributional conformal prediction to estimate conditional cumulative distribution functions and build calibrated prediction sets with finite-sample coverage, together with an adaptive two-stage budget allocation between estimation and calibration.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Industrial and Systems Engineering
Department dc:contributor.department
Industrial and Systems Engineering
Grantor dc:publisher
Virginia Tech
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Zhao, Jin
Chair dc:contributor.committeechair
  • Chen, Xi
Committee members dc:contributor.committeemember
  • Xu, Jie
  • Sarin, Subhash C.
  • Bansal, Manish

Subjects

dc:subject × 5

Rights

dc:rights
Statement dc:rights
  • In Copyright
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Other
vt_gsexam:47056

Chain of custody

source
Harvested from
Virginia Tech
Base URL
vtechworks.lib.vt.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Zhao, Jin. Kernel-Based Metamodeling for Heteroscedastic Simulation: Theory and Distribution-Free Uncertainty Quantification. doctoral thesis, Virginia Tech, 2026. https://hdl.handle.net/10919/143194