Re: bfgsmin iterations

[Jaroslav Hajek]

On Wed, May 13, 2009 at 10:27 AM, Levente Torok <address@hidden> wrote:
> Dear Michael Creel,
>
> I used your bfgsmin implementation and I found that it evaluates the
> objective function unnessessarily many times.
> I have recorded the points at which my function was evaluated
>
> Please find the enclosed png file. x axis - iteration, y - value of my 1 
> dimensional obj.func parameter.
> And this is the array in which I recorded the location of the evaluations
> z=[
>   3.00000000000000e+02
>    6.00000000000000e+02
>    1.00000000000000e-01
>    4.50000000000000e+02
>    4.50000000000000e+02
>    1.50000000000000e+02
>    1.00000000000000e-01
>    1.00000000000000e-01
>    2.25000000000000e+02
> ...
>
> As I see it usually calles the obj.func two times with exactly the same 
> argument.
> I'd like to know it if is normal and if so, isn't there an easy way of doing 
> a caching
> instead of double calling.
>
> Levente
>

Since I think exactly the same issue occurs in the new Octave's core
fsolve and fminunc codes,
you might be interested in how they cope with this:
If using finite difference derivatives, the duplicate evaluation is
avoided internally by the code.
However, if using user-supplied gradient/jacobian, this is not
technically possible, because the decision whether the derivatives are
needed is made *after* a function value request. So the user needs to
handle this if there's a significant redundancy to be exploited.  The
fsolve help includes a nice example how this can be efficiently done
using a persistent variable holding the intermediate state.

cheers

-- 
RNDr. Jaroslav Hajek
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz