This is the mail archive of the
gsl-discuss@sourceware.cygnus.com
mailing list for the GSL project.
Re: multidimensional optimization
- To: Fabrice Rossi <rossi at ufrmd dot dauphine dot fr>
- Subject: Re: multidimensional optimization
- From: Brian Gough <bjg at network-theory dot co dot uk>
- Date: Mon, 27 Sep 1999 15:18:34 +0100 (BST)
- Cc: gsl-discuss at sourceware dot cygnus dot com
- References: <37E79D47.A5213C06@ufrmd.dauphine.fr>
We don't have a design document for multidimensional optimization, so
this list is the right place to discuss the design.
Here is the general plan that I have followed in the past...
The starting point is to list and classify all the existing
techniques, and determine an appropriate API for each class of
problem.
(There is a nice little summary of minimisation techniques in the book
"Numerical Computation (Volume 2)" by Ueberhuber.)
Each API should aim to support a variety of different interchangeable
algorithms within the problem class. For iterative algorithms we have
adopted an initialise-iterate-test structure by convention.
The 1-d root finding and minimisation directories give the general
idea.
Fabrice Rossi writes:
> Hi All.
> I've read in "The Plan" that multidimensional optimization is
> considered as one of the priority one targets for GSL future. I'm
> working in the field of artifical neural networks and I need to
> implement simple multilayer perceptrons in order to test a new
> method I've designed.
> As I need multidimensional optimization to train my neural
> networks, I'll be pleased to help coding traditionnal algorithms in
> GSL (I've already implemented nn training algorithm in both C and
> C++). If I'm correct the work as not started on this part.
> I would like to know if there are any design documents about
> multidimensional optimization implementation in GSL and if the
> mailing list might be a correct place to discuss about this design.
> Fabrice Rossi