augmented_lagrangian_method on purely computational / numeric functions?
Subject: augmented_lagrangian_method on purely computational / numeric functions?
From: Daniel Lakeland
Date: Tue, 16 Jan 2007 21:30:46 -0800
On Tue, Jan 16, 2007 at 10:04:39PM -0700, Robert Dodier wrote:
> On 1/16/07, Daniel Lakeland <dlakelan at street-artists.org> wrote:
>
> > Does augmented_lagrangian_method require that its FOM argument be
> > symbolically differentiable?
>
> Yes, augmented_lagrangian_method assumes that it is going to
> compute the necessary derivatives by symbolic differentiation.
> The point of that was to remove the derivatives from the problem
> formulation since they can be obtained mechanically.
It's a good idea if the argument is symbolic. If it's a
computational/numerical function then it might be possible for the
user to supply a numerical differentiator as you suggested.
However the user may have no way of supplying a derivative for some
very complex functions. I assume without derivatives the LBFGS can't
work because it's a quasi-newton method. perhaps it's possible to use
a secant method internally?
A minimizer without derivative computations would be
fantastic. perhaps it's a good topic for the student projects
mentioned in a previous email?
--
Daniel Lakeland
dlakelan at street-artists.org
http://www.street-artists.org/~dlakelan