Optimization using surrogates for engineering design
John Dennis (Rice University)
MSI Advanced Computation seminarDATE: 2006-02-20
TIME: 11:00:00 - 12:00:00
LOCATION: John Dedman GD35
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
This talk will outline the surrogate management framework, which is presently built on the filter MADS method for general nonlinear programming without derivatives. The focus is on the numerical results, with a brief introduction to the MADS algorithm and a slight mention of the convergence results.
This line of research was motivated by industrial applications, indeed, by a question I was asked by Paul Frank of Boeing Phantom Works. His group was often asked for help in dealing with very low dimensional design problems driven by expensive simulations. Everyone there was dissatisfied with the common practice of substituting inexpensive surrogates for the expensive ``true'' objective and constraint functions in the optimal design formulation. I hope to demonstrate in this talk just how simple the answer to Paul's question is.
The surrogate management framework has been implemented successfully
by several different groups, and it is unreasonably effective in
practice, where most of the application are extended valued and
certainly nondifferentiable. This has forced my colleagues and me
to begin to learn some nonsmooth analysis, which in turn has led
to MADS, a replacement for the GPS infrastructure algorithm.
