Thread Subject: Have fminunc show the gradient?

I have been looking through the help for a long while and
cant seem to find how to make fminunc display the gradient
at each iteration - maybe I am just missing it??

Thanks,

Dave

Dave,

See the output function in the doc, which provides a number of
optimization quantities at each iteration, e.g. the gradient in
fminunc:
http://www.mathworks.com/access/helpdesk/help/toolbox/optim/ug/f19175.html#f11022

-Marcelo

David Doria wrote:
> I have been looking through the help for a long while and
> cant seem to find how to make fminunc display the gradient
> at each iteration - maybe I am just missing it??
>
> Thanks,
>
> Dave

Ah, great - that does the trick - although I was expecting
something a bit more built in.. so I don't have to write my
own output function, I could just toggle what the built in
output function outputs!

Thanks!

Dave

Ok, now I would like to see the hessian that it calculates
at every step... I don't see this in the OptimValues list...
is there a way to get at it?

Dave

There is a built-in plot function that plots the norm of the gradient.

One way to run it is to open the optimtool, and select FIrst Order
Optimality under Plot Functions. There is a way to run this from the
command line as well.

I assume you're using the medium-scale algorithm. This algorithm
does not compute the Hessian, but rather a so-called quasi-Newton
approximation to the inverse of the Hessian. This matrix is only
an approximation (often crude), and to the inverse of the Hessian
(not to the Hessian itself), so it's not made available via the
output function - typically it's not useful.

The output function gives you access to each iterate x; you can
compute the Hessian (or an approximation to it via finite differences)
at x inside the output function.

-Marcelo

David Doria wrote:
> Ok, now I would like to see the hessian that it calculates
> at every step... I don't see this in the OptimValues list...
> is there a way to get at it?
>
> Dave

Hi!
Thanks for the help, it was very useful! I have a related question: Is it possible to save the gradient so that I can see it in the workspace, rather than just display it for each iteration? I defined it as an output of outfun by:

function [stop,Y] = outfun(x, optimValues, state)
stop=false;

   switch state
       case 'iter'
            Y(optimValues.iteration+1,:)=optimValues.gradient'
       otherwise
   end
end

But still could not save it in the workspace. I'd appreciate any help!
Thanks

Marcelo Marazzi <mREMOVEmaALLraCAPITALSzzi@mathworks.com> wrote in message <g84pmc$rle$1@fred.mathworks.com>...
> There is a built-in plot function that plots the norm of the gradient.
>
> One way to run it is to open the optimtool, and select FIrst Order
> Optimality under Plot Functions. There is a way to run this from the
> command line as well.
>
> I assume you're using the medium-scale algorithm. This algorithm
> does not compute the Hessian, but rather a so-called quasi-Newton
> approximation to the inverse of the Hessian. This matrix is only
> an approximation (often crude), and to the inverse of the Hessian
> (not to the Hessian itself), so it's not made available via the
> output function - typically it's not useful.
>
> The output function gives you access to each iterate x; you can
> compute the Hessian (or an approximation to it via finite differences)
> at x inside the output function.
>
> -Marcelo
>
> David Doria wrote:
> > Ok, now I would like to see the hessian that it calculates
> > at every step... I don't see this in the OptimValues list...
> > is there a way to get at it?
> >
> > Dave