[Insight-users] Re: vnl_powell: three parameter fit (Mathieu Malaterre)

[m.weigert at fz-juelich.de]

Hi Mathieu,

this is also what I know so far about LM that
it is a method for performing least squares fits.
I don't think (I dont know, to be honest) if it can be used to perform unbounded optimization as with
Powell or other optimization algorithms 
(although there are optimization schemes that seem to combine LM and gradient descent for example).
However, if you can express your target function
in case of a sum of differences, you can use it to fit you model parameters.
For example, it should be possible to do image registration for sum of squared differences as target function with LM when you look at the pixelwise difference instead of the sum and you can
introduce further cost functions as LM is multivalued.


Best regards,
Markus





----- Original Message -----
From: Mathieu Malaterre <mathieu.malaterre at gmail.com>
Date: Tuesday, May 15, 2007 3:19 pm
Subject: Re: vnl_powell: three parameter fit (Mathieu Malaterre)

> Hi Markus,
> 
>  Thanks for the comments. I am also currently investigating LM, 
> as it
> seems very promising. Unfortunately I cannot figure out how to use it.
> According to:
> 
> http://paine.wiau.man.ac.uk/pub/doc_vxl/books/core/book_6.html#SEC44
> 
>  I think the vnl implementation lack general implementation and only
> provide least square API (need to decompose my function as a sum of
> square function...). See my post:
> 
> http://sourceforge.net/mailarchive/forum.php?thread_name=bf0c3b3f0705150506q527466a4jc530bf42455f540c%40mail.gmail.com&forum_name=vxl-users
> 
>  I am waiting for a vnl guru to confirm the vnl - LM optimizer cannot
> do general cost function minimization.
> 
>  In which case I am going now toward a lbfgs approach. I can compute
> the first derivative of my function, and it would be dumb not to use
> this information (powell method is too generic).
> 
>  A final note. I could get powel to converge to the correct solution
> when giving the optimizer a relative close solution. I was not able
> (read: I don't have the time to dig in powell theory) to figure out
> 'how good' my initial solution should be.
> 
> I'll let you know how it goes,
> 
> Thanks for your interest,
> -Mathieu
> 
> 
> On 5/15/07, m.weigert at fz-juelich.de <m.weigert at fz-juelich.de> wrote:
> 
> > Hi Mathieu,
> >
> > I think there is also a Levenberg-Marquardt implementation in 
> ITK which may be better suited to
> > perform non-linear least squares fits.
> > I don't know how to do this with vnl_powell, I am currently 
> investigating Levenberg-Marquardt for
> > parameter fitting.
> >
> > Kind regards,
> > Markus
> >
> >
> 
> 
> -- 
> Mathieu
> 
>