[OctDev] nonlinear constrained optimization work

[Joseph P. S. <js...@js...>]

Hi. In Octave 2.0.14, I've prototyped multmethod, a robust general
purpose nonlinear constrained optimization routine; it is an
implementation of the method of multipliers, based on my Mathematica
work; cf

  http://www.mathsource.com/Content/Applications/Mathematics/0207-289

multmethod incorporates a selection of unconstrained solvers: a very
robust but somewhat expensive [as expected] unconstrained non-gradient
optimizer based on Nick Higham's work (mdsmax.m), a much faster but less
robust [as expected] bfgs-based (and may include a trust region)
optimizer based on Tim Kelley's work (bfgswopt.m, cgtrust.m, and
related), and a numerical gradient routine based on Peter Spellucci's
work (part of donlp2), all written as Octave M-files.  I have secured
permission to use the work of Tim Kelley and Nick Higham as a starting
point.  The tiny reimplemented numgrad routine will not be an issue.

This should meet the needs of the "free NLP solver" project, which can
also be used as the "free QP solver".

As requested by http://www.octave.org/projects.html:

	Also, if you start working steadily on a project,
	please let bug...@be... know.

I'm sending this email to let you know.

I'm aware of the optim work at sourceforge, and will be looking into
that in an attempt to reconcile differences and unify the offering.  I
still need to refine, document, validate, format, rename, reconcile, and
package this work; I expect to work on this again in a couple of months,
and it will take a couple of months of work to make it happen.

FYI.  /Jskud