liu.seSearch for publications in DiVA
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • harvard1
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • oxford
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Surrogate models composed of locally estimated neural networks
Linköping University, Department of Computer and Information Science, Statistics. Linköping University, Faculty of Arts and Sciences.
Linköping University, Department of Computer and Information Science, Statistics. Linköping University, Faculty of Arts and Sciences.
2008 (English)Report (Other academic)
Abstract [en]

When a computer code model is computationally expensive, or there is a strong demand for short execution times, it may be advantageous to invest in a computationally cheaper surrogate model that can provide almost the same output(s) as the original model. We examined the performance of surrogate models derived by first applying an adaptive or non-adaptive algorithm to generate a set of design points, and subsequently using locally estimated artificial neural networks (ANNs) to predict the output at previously untried inputs. We found that such surrogate models generally performed well, and indeed often much better than ANNs fitted to all data in the entire input domain. Furthermore, we observed that locally estimated ANNs can adapt to response surfaces exhibiting extreme features like sharp ridges, and that such prediction models can accommodate relatively high-dimensional inputs.

Place, publisher, year, edition, pages
2008.
Series
Report-LiU-IDA-STAT, 2
Keyword [en]
Artificial neural networks, response surface, experimental design, surrogate models, local fitting
National Category
Computer and Information Science
Identifiers
URN: urn:nbn:se:liu:diva-17112OAI: oai:DiVA.org:liu-17112DiVA: diva2:201977
Available from: 2009-03-06 Created: 2009-03-06 Last updated: 2009-03-06Bibliographically approved
In thesis
1. Computer Experiments Designed to Explore and Approximate Complex Deterministic Models
Open this publication in new window or tab >>Computer Experiments Designed to Explore and Approximate Complex Deterministic Models
2008 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Computer experiments are widely used to investigate how technical, economic, and ecological systems respond to changes in inputs or driving forces. This thesis is focused on design of computer experiments that can help us better understand the output from complex computer code models. The major part of our work was devoted to experiments involving derivation and application of computationally cheaper surrogate models of a given computer code model. We developed an adaptive sequential design algorithm that efficiently reveals nonlinearities in the model output, and we integrated this algorithm with methods for predicting model outputs at untried inputs. Compared to the methods currently in use, our sequential design has the advantage of not requiring any prior information about the response of the investigated model output to changes in the inputs. Of special interest, we found that our algorithm works satisfactorily even if the curvature of the response surface varies strongly over the input domain. Variance-based sensitivity analysis is a well-established technique to elucidate model outputs, but it can become prohibitively expensive to implement because it requires numerous model runs. Surrogate models can facilitate such analysis, and if our sequential design algorithm is utilized, it can supply useful information about both linear and nonlinear responses to model inputs. Experiments involving repeated runs of a model of the flow of water and nitrogen through a river basin showed that our approach can be applied to extract the essence of complex deterministic models. In addition, our research showed that computationally inexpensive surrogate models offer an ideal basis for interactive decision support tools and learning processes, because they can provide almost immediate responses to user-defined model inputs.

Abstract [sv]

Datorexperiment används allmänt för att undersöka hur tekniska, ekonomiska och ekologiska system reagerar på förändringar i tillförsel eller drivkrafter. Denna avhandling är inriktad på datorexperiment som kan hjälpa oss att bättre förstå beräkningar baserade på komplicerade numeriska modeller som bara är definierade av en datorkod. Huvuddelen av vårt arbete ägnades åt experiment som innefattar härledning och tillämpning av beräkningsmässigt billiga s.k. surrogatmodeller som ger nästan samma resultat som ursprungsmodellen. Vi utvecklade en adaptiv sekventiell designalgoritm som effektivt avslöjar icke-linjära reaktioner på ändrad input till modellen, och vi integrerade denna algoritm med metoder för att prediktera modellens output för nya indata. Jämfört med de metoder som nu används har vår algoritm fördelen att den inte ställer några krav på förhandsinformation om modellens struktur. Speciellt noterade vi att den fungerar tillfredsställande även om olika delar av modellens responsyta har helt olika statistiska egenskaper. Varians-baserad känslighetsanalys är en väl etablerad teknik för att belysa modellers output, men den kan leda till höga datorkostnader eftersom den kräver många modellkörningar. Surrogatmodeller kan i sådana fall underlätta analysen. Om vår sekventiella designalgoritm utnyttjas, kan man desutom få viktig information om både linjära och icke-linjära effekter av förändringar i modellens indata. Experiment som innefattade upprepade körningar av en model för flödet av vatten och kväve genom ett avrinningsområde visade att man kan klarlägga det centrala i stora komplexa modeller. Dessutom visade vår forskning att beräkningsmässigt billiga surrogatmodeller erbjuder en idealisk grund för beslutstöd och lärandeprocesser, eftersom de kan ge en nästan omedelbar respons på de data som användaren matar in i modellen.

Place, publisher, year, edition, pages
Linköping: Linköpings universitet, 2008. 58 + papers 1-4 p.
Series
Linköping Studies in Arts and Science, ISSN 0282-9800 ; 423Linköping Studies in Statistics, ISSN 1651-1700 ; 9
National Category
Computer and Information Science
Identifiers
urn:nbn:se:liu:diva-17115 (URN)978-91-7393-976-8 (ISBN)
Public defence
2008-02-29, Alan Turing, hus E, Campus Valla, Linköpings universitet, Linköping, 13:00 (English)
Opponent
Supervisors
Available from: 2009-03-06 Created: 2009-03-06 Last updated: 2014-09-23Bibliographically approved

Open Access in DiVA

No full text

Other links

Link to Ph.D. Thesis

Authority records BETA

Shahsavani, DavoodGrimvall, Anders

Search in DiVA

By author/editor
Shahsavani, DavoodGrimvall, Anders
By organisation
StatisticsFaculty of Arts and Sciences
Computer and Information Science

Search outside of DiVA

GoogleGoogle Scholar

urn-nbn

Altmetric score

urn-nbn
Total: 115 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • harvard1
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • oxford
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf