"Line search failed" in lsquares



"Robert Dodier" <robert.dodier at gmail.com> schrieb am 09/11/2008 03:54:07 
AM:

> On Wed, Sep 10, 2008 at 9:44 AM, Jan Ploski <Jan.Ploski at offis.de> wrote:
> 
> > I can see by plotting that the initial parameters hit the data pretty 
well.
> 
> Hmm. Really? Looks fairly noisy to me ...

Hi,

What I meant is that the curve does not lie light years away from the 
data, which I suspected might cause trouble. I know that the actual values 
decrease exponentially (I'm estimating per process memory consumption of a 
distributed memory application), so the noise does not bother me much.

> 
> > Can you give me a clue why it fails and what I could do to correct it?
> 
> Try dividing all the y values by a large number such as 1e8.
> 
> After that I get a = 4.8, b = 0.26 approximately.
> I guess you'll have to rescale a.

Scaling the data to [0,1] worked perfectly for all cases where it was 
failing before. Thanks!

> > Also, what are the meanings of the undocumented "iprint" parameter and
> > of the "tol" parameter, which is mentioned, but not explained in the
> > documentation?
> 
> iprint and tol are parameters which are recognized by the LBFGS
> algorithm. ? lbfgs says more about that. Sorry that the documentation
> is incomplete.
> 
> I am glad to see that you find some use in the lsquares code.
> If you have any comments about what would make it more useful,
> I would be interested to hear about it.

Obviously, making it automatically cope with large-valued data would be 
nice. But it becomes a cosmetic detail once you know that the scaling is 
required. Maybe you can just mention it in the documentation.

Regards,
Jan Ploski

--
Dipl.-Inform. (FH) Jan Ploski
OFFIS
FuE Bereich Energie | R&D Division Energy
Escherweg 2  - 26121 Oldenburg - Germany
Phone/Fax: +49 441 9722 - 184 / 202
E-Mail: Jan.Ploski at offis.de
URL: http://www.offis.de