Richard Clark wrote:
On Mon, 18 Jun 2007 12:08:09 -0700, Jim Lux
wrote:


I think this is true of all optimizers..
The more complex the thing being optimized, the more difficult it is to
have a generic interface that provides a way for the user to tell the
optimizer what the constraints and desires are. In the limit, you wind
up writing a program to implement the constraints, and another program
to implement the evaluation function, and yet another program to take
the optimized variables and turn it into a model which the modeling
engine can run.


Hi Jim, Arie,

It goes WAY beyond that (as if there weren't enough problems). In the
GA community, the introduction of unlimited variables leads to (very
quickly with even a few of them) what is called "combinatorial
explosion."

A second issue (and it seems that Arie probably has at least one
algorithm to tackle this) is with an engine becoming stuck at a local
minima or maxima. This false solution ignores a better one nearby (or
further down the road) simply because it satisfied the criteria within
a restricted region of a curve.

Arthur's models quite obviously exhibit this last problem when his
designs can be bested with mediocre examples that have been drawn from
the dusty bookshelves. The science of GA has moved well beyond the
Paleolithic era of AO.

73's
Richard Clark, KB7QHC
My comment was not so much about the details of the optimizers (for
which you've given some typical problems) but with the difficulty of
coming up with a generic optimizer that can manage generalized models,
generalized constraints, etc. Since the underlying optimizer algorithms
are available as canned packages/library routines in most cases, the
hard work is in formulating the stuff that goes to the optimizer.