From octave-sources-request at bevo dot che dot wisc dot edu Mon Nov 26 12:54:52 2001
Subject: Re: Levenberg-Marquardt algorithm
From: Paul Kienzle <pkienzle at users dot sourceforge dot net>
To: Francesco Potorti` <pot at gnu dot org>
CC: Paul Kienzle <pkienzle at users dot sourceforge dot net>,        octave-sources@bevo.che.wisc.edu,        Matthias Jueschke <matthias@physik3.gwdg.de>
Date: Mon, 26 Nov 2001 13:52:47 -0500

Francesco Potorti` wrote:

>    > I have this: <ftp://fly.cnuce.cnr.it/ftp/pub/software/octave/leasqr/>.
>
>    I asked all the authors, and it is now under the GPL.  I've included it in
>    octave-forge  (http://octave.sf.net).
>
> Thanks, Paul!
>
> I wanted to remove then my version form my ftp server and replace it
> with a pointer to octave.sf.net, but I can't find it there.  Can you
> provide a less generic pointer?

leasqr.m, leasqrdemo.m and dfdp.m are in the CVS directory
   http://cvs.sourceforge.net/cgi-bin/viewcvs.cgi/octave/octave-forge/main/optim/
I won't point to specific files since the reference changes with each revision.

>    I see that NIST has some data sets for non-linear regression with expected
>    results (http://www.itl.nist.gov/div898/strd/nls/nls_main.shtml).  You can use
>    these to check both versions, and if both are correct we should use the one
>    which is faster.
>
> I do not currently have the interest to test these versions, but as far
> as I understand, the problem is not only correctness, but robustness and
> similar things.  If the dataset is well chosen, however, those things
> should come up by themselves.

These issues are discussed on the Background Information link
(http://www.itl.nist.gov/div898/strd/nls/nls_info.shtml).

Paul Kienzle
pkienzle at users dot sf dot net