On Mar 8, 2011, at 2:14 PM, Gregory Maxwell wrote:
On Tue, Mar 8, 2011 at 2:05 PM, <[email protected]> wrote:
PS. You can probably save a few more function evaluations if you
check for repeated calls with the same arguments. The reason is
that, on the line search, SLSQP often calls your function twice at
the end of the line search: once without the gradient, and then once
more (when it realizes ex post facto that the line search is done)
with the gradient.
Yes, I have noticed that. Unfortunately, I cannot make use of
this fact. I use automatic differentiation for the gradients, so
computing the derivative involves computing the function anyway.
Perhaps a dumb question— but why are you using automatic
differentiation
instead of using a gradient free algorithm?
Are the gradient free algorithims in NLOPT really so poor for your
application that
automatic differentiation + a with-gradient algorithm is superior?
Assuming Peter is using standard terminology, automatic
differentiation does *not* mean numerical differentiation with finite
differences. Automatic differentiation means calculating the exact
derivative analytically, just using a computer program rather than
doing it by hand: an AD differentiaties your source code symbolically.
So, automatic differentiation can in principle be the same efficiency
as programming analytical derivatives by hand, and with it the
gradient-based algorithms are probably far superior to the derivative-
free algorithms for this many unknowns.
In principle, since AD takes source code to compute an objective and
generates source code to compute objective+gradient, you could
possibly edit the resulting code to remove stuff only needed for the
computation of the objective. But editing program-generated code by
hand is messy, and is fragile because you need to re-do it each time
you change your objective, so I can understand why Peter wouldn't want
to do it.
Steven
_______________________________________________
NLopt-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss
