Re: alternative to fmins

[Olaf Till]

On Mon, Jan 18, 2016 at 11:25:33AM +0100, Dominique Richard wrote:
> Hello everyboby,
> 
> I am using fmins to estimate a set of parameters  by minimizing the
> error between modelized curves and experimental data.
> But I am afraid of not getting the best result and possibly falling into
> secondary minima.
> Is there a way to avoid this pitfall.
> More generally are they other functions that I could use instead ot
> 'fmins' for parameter estimation ?

Hi Dominique,

you should call 'optim_doc' from the Octave prompt to access overview
(and detailed) information on optimization functions. Menu item
'Residual optimization' is applicable to your case, specifically the
subitems (functions) 'nonlin_residmin' and 'nonlin_curvefit'.

'fmins' does scalar optimization; scalar optimization should not
normally be used for curve fitting since it 'wastes' some available
information. Furthermore, 'fmins' by default uses an algorithm which
is normally not in the first line.

If the problem should really be due to secondary minima (I wouldn't
take the failing of 'fmins' as a strong indication for this), and you
can't 'rule them out' by constraints, a stochastic optimizer could be
used.

Regards,

Olaf

-- 
public key id EAFE0591, e.g. on x-hkp://pool.sks-keyservers.net

signature.asc
Description: Digital signature