A Framework for Multi-fidelity Modeling in Global Optimization Approaches

Zelda B. Zabinsky, Giulia Pedrielli, Hao Huang

Research output: Chapter in Book/Report/Conference proceedingConference contribution

1 Scopus citations

Abstract

Optimization of complex systems often involves running a detailed simulation model that requires large computational time per function evaluation. Many methods have been researched to use a few detailed, high-fidelity, function evaluations to construct a low-fidelity model, or surrogate, including Kriging, Gaussian processes, response surface approximation, and meta-modeling. We present a framework for global optimization of a high-fidelity model that takes advantage of low-fidelity models by iteratively evaluating the low-fidelity model and providing a mechanism to decide when and where to evaluate the high-fidelity model. This is achieved by sequentially refining the prediction of the computationally expensive high-fidelity model based on observed values in both high- and low-fidelity. The proposed multi-fidelity algorithm combines Probabilistic Branch and Bound, that uses a partitioning scheme to estimate subregions with near-optimal performance, with Gaussian processes, that provide predictive capability for the high-fidelity function. The output of the multi-fidelity algorithm is a set of subregions that approximates a target level set of best solutions in the feasible region. We present the algorithm for the first time and an analysis that characterizes the finite-time performance in terms of incorrect elimination of subregions of the solution space.

Original languageEnglish (US)
Title of host publicationMachine Learning, Optimization, and Data Science - 5th International Conference, LOD 2019, Proceedings
EditorsGiuseppe Nicosia, Panos Pardalos, Renato Umeton, Giovanni Giuffrida, Vincenzo Sciacca
PublisherSpringer
Pages335-346
Number of pages12
ISBN (Print)9783030375980
DOIs
StatePublished - 2019
Event5th International Conference on Machine Learning, Optimization, and Data Science, LOD 2019 - Siena, Italy
Duration: Sep 10 2019Sep 13 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11943 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference5th International Conference on Machine Learning, Optimization, and Data Science, LOD 2019
Country/TerritoryItaly
CitySiena
Period9/10/199/13/19

Keywords

  • Gaussian processes
  • Global optimization
  • Meta-models
  • Multi-fidelity models
  • Probabilistic Branch and Bound

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint

Dive into the research topics of 'A Framework for Multi-fidelity Modeling in Global Optimization Approaches'. Together they form a unique fingerprint.

Cite this