[Beowulf] Re: Beowulf Digest, Vol 15, Issue 16

[Lombard, David N]

From: Jim Lux on Sent: Monday, May 09, 2005 12:03 PM
> At 09:49 AM 5/9/2005, David Mathog wrote:
> 
> >
> >So you have a calculation problem that's embarrassingly parallel
> >but an infinite parameter space to search.  Seems to me that if
> >this process is to be automated you will need to define a goal
> >function, presumably based primarily on the far field results,
> >and then use some search strategy or other to try to find at
> >least a local "best" design in your parameter space. For instance,
> >this probem might be amenable to a genetic algorithm approach.
> 
> 
> Actually, the ideal "goal evaluator" is me, looking at the results of
> several runs and comparing them, then telling the "box" which way to
go
> next. As you say, if you could define a goal function with sufficient
> clarity, then any manner of optimizers could grind away on the problem
> overnight.  Unfortunately, most real design problems have requirements
> that
> are a bit fuzzy:  Don't make it "too big" or "too flimsy".  terms like
> "flimsy" are hard to encapsulate succinctly in a mathematical
formulation
> (although, gosh, we certainly try, by requiring certain mechanical
> resonance properties and failure strengths). Much like other things,
you
> know them when you see them.

Hmm, too "big" or "flimsy" have very precise definition based on exactly
the measures you describe.  Such goal functions have been in use in the
various structural optimizers for years now, e.g., weight, stress
distributions, deflections, modal responses, &etc.

> >I know essentially nothing about antenna design so take the following
> >suggestion with the requisite large crystal of salt.  Can you
> >subdivide the available (flat?) radiating area into a grid of
> >identical squares which are classified as antenna/non-antenna?
> >At that point your parameters may reduce to:  1) number of squares,
> >2) their distribution.  The first is a single integer and the second
> >is a bit vector (ie, MxN bits, 1 for cells that are
> >antenna, 0 for cells that are not.) This is a simple enough
> >parameter space that a genetic algorithm should be relatively
> >simple to implement.  Hopefully you can make this work with so
> >many itty bitty squares that the little squares are much smaller
> >than the shortest wavelength so that the jaggedy edges won't
> >change the results significantly.
> 
> Aha... your idea has been anticipated!  Several people have done just
this
> (using a Beowulf, even, for the optimizing).  Randy Haupt did a fair
> amount
> of it with wire antennas (and others, I'm sure).  There was also
someone
> at
> UCLA who designed wireless antennas using just what you describe
(adding
> and removing small patches of conductive surface).  They then
fabricated
> the antennas and tested them.

Yes, there are a number of extant solutions using stochastic
optimization in the automotive and aerospace manufacturers among others
for these very purposes.  Clusters running in throughput mode are a
really good engine for this work.

-- 
David N. Lombard

My comments represent my opinions, not those of Intel Corporation.