On Nov 17, 2010, at 5:21 AM, Matteo Lombardi wrote:
A couple of questions:
-Does the magnitude of the gradient has to be correct or what is
important is just the direction?(my sensitivity gives just the
direction..)
The magnitude must be correct. (In principle you should always be
able to compute the correct magnitude and direction of the gradient
efficiently via an adjoint method.)
Having a substantially incorrect magnitude of the gradient will almost
certainly cause the methods to fail quickly.
-At the moment my sensitivity(gradient) is evaluated on every mesh
points and thus there is some sort of "interpolation" to project it
on my control points (which are linked to the shape
parametrization). Thus for sure in this projection process there is
some "fitting error" and for sure there is also some noise coming
from the CFD calculations. Do you think this is enough to cause the
methods to fail?
It varies. In general, the methods that try to construct second
derivatives (SLSQP, LBFGS, and the other variable-metric/quasi-Newton
methods) are going to be more sensitive to errors in the gradient.
MMA may get farther; in many cases I have found that small errors in
the gradient only cause problems with MMA at the very end of the
optimization, when you are trying to polish off the last few decimal
places of the optimum.
Steven
_______________________________________________
NLopt-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss
