Design search and optimisation using radial basis functions with regression capabilities
Design search and optimisation using radial basis functions with regression capabilities
Modern design search and optimisation (DSO) processes that involve the use of expensive computer simulations commonly use surrogate modelling techniques, where data is collected from planned experiments on the expensive codes and then used to build meta-models. Such models (often termed response surface models or RSMs) can be built using many methods that have a variety of capabilities. For example, simple polynomial (often linear or quadratic ) regression curves have been used in this way for many years. These lack the ability to model complex shapes and so are not very useful in constructing global RSM's for non-linear codes such as the Navier Stokes solvers used in CFS - they are, however, easy to build. By contrast Kriging and Gaussian Process models can be much more sophisticated but are often difficult and time consuming to set up and tune. At an intermediate lvel radial basis function (RBF) models using simple spline functions offer rapid modelling capabilities with some ability to fit complex data. However, as normally used such RBF RSM's strictly interpolate the available computational data and while acceptable in some cases, when used with codes that are iteratively converged, they find it difficult to deal with the numerical noise inevitably present. This paper describes a modification to the basic RBF scheme that allows a systematic variation of the degree of regression from a pure linear regression line to a fully interpolating cubic radial basis function model. The ideas presented are illustrated with data from the field of aerospace design.
1852338296
39-49
Keane, A.J.
26d7fa33-5415-4910-89d8-fb3620413def
2004
Keane, A.J.
26d7fa33-5415-4910-89d8-fb3620413def
Keane, A.J.
(2004)
Design search and optimisation using radial basis functions with regression capabilities.
In,
Parmee, I.C.
(ed.)
Adaptive computing in design and manufacture VI.
Springer, .
Record type:
Book Section
Abstract
Modern design search and optimisation (DSO) processes that involve the use of expensive computer simulations commonly use surrogate modelling techniques, where data is collected from planned experiments on the expensive codes and then used to build meta-models. Such models (often termed response surface models or RSMs) can be built using many methods that have a variety of capabilities. For example, simple polynomial (often linear or quadratic ) regression curves have been used in this way for many years. These lack the ability to model complex shapes and so are not very useful in constructing global RSM's for non-linear codes such as the Navier Stokes solvers used in CFS - they are, however, easy to build. By contrast Kriging and Gaussian Process models can be much more sophisticated but are often difficult and time consuming to set up and tune. At an intermediate lvel radial basis function (RBF) models using simple spline functions offer rapid modelling capabilities with some ability to fit complex data. However, as normally used such RBF RSM's strictly interpolate the available computational data and while acceptable in some cases, when used with codes that are iteratively converged, they find it difficult to deal with the numerical noise inevitably present. This paper describes a modification to the basic RBF scheme that allows a systematic variation of the degree of regression from a pure linear regression line to a fully interpolating cubic radial basis function model. The ideas presented are illustrated with data from the field of aerospace design.
Text
kean_04.pdf
- Accepted Manuscript
More information
Published date: 2004
Additional Information:
Out of print
Identifiers
Local EPrints ID: 22793
URI: http://eprints.soton.ac.uk/id/eprint/22793
ISBN: 1852338296
PURE UUID: 4355cb8a-ebfb-4b82-b9b2-099a24577f40
Catalogue record
Date deposited: 29 Mar 2006
Last modified: 16 Mar 2024 02:53
Export record
Contributors
Editor:
I.C. Parmee
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics