An Efficient Batch K-Fold Cross-Validation Voronoi Adaptive Sampling Technique for Global Surrogate Modeling

Abstract

Surrogate models can be used to approximate complex systems at a reduced cost and are widely used when data generation is expensive or time consuming. The accuracy of these models is dependent on the samples used to create them. Therefore, proper selection of samples within the parameter space is of paramount importance. Adaptive sampling procedures have been developed to identify optimal locations for new samples by leveraging response information from existing samples. In this manner, adaptive sampling methods significantly reduce the number of samples required to build a surrogate model of a given accuracy. However, adaptive sampling techniques have a cost associated with determining ideal sample locations, which typically grows with the sample count. The present effort seeks to reduce the cost associated with the adaptive sampling procedure and thereby maximize the efficiency of surrogate model creation. A new K-fold cross-validation (KFCV)-Voronoi adaptive sampling technique is proposed to reduce the sample selection costs by adding a global KFCV filter to the cross-validation (CV)-Voronoi technique. The costs are further reduced through an innovative Voronoi batch sampling technique. The proposed adaptive sampling acceleration techniques are evaluated using benchmark functions with increasing parameter space dimension and aerodynamic loading data.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
May 01, 2019
Accession Number
AD1076676

Entities

People

  • Abraham Atachbarian
  • Andrew L. Kaminsky
  • Kapil Pant
  • Wendy N. Hashii
  • Yi Wang

Tags

Communities of Interest

  • Biomedical
  • Materials and Manufacturing Processes

DTIC Thesaurus Topics

  • Accuracy
  • Aerodynamic Loading
  • Air Force
  • Air Force Facilities
  • Circuit Boards
  • Complex Systems
  • Computational Fluid Dynamics
  • Cost Reductions
  • Engineering
  • Mach Number
  • Printed Circuits
  • Sampling
  • South Carolina
  • Stochastic Processes
  • Two Dimensional
  • United States
  • Validation

Readers

  • Adaptive Control and Estimation with Uncertainty in Dynamic Systems.
  • Regression Analysis.

Technology Areas

  • Space