Genetic Algorithms

[Robert G. Brown]

On Sun, 27 Jun 1999, Eugene Leitl wrote:

> 
> Obviously you leave the embarrasingly parallel regime when you attempt 
> to distribute an individual over several nodes. I'm curious: how large 
> a population do you have? What prevents you from simply taking a
> larger amount of nodes, and assigning an individual/subpopulation to
> each node? In that case you even don't need a Beowulf -- any NOW will
> do. 

Actually, the evaluation of the error function (basically the fitness)
is what I'm currently distributing, which does involve a bit of IPC's.
I've thought about distributing the GA itself -- there are several
amusing possibilities that open up that emulate the development of
isolated populations followed by mixing to achieve hybrid vigor, but I
suspect that this will require a carefully metered degree of homogeneity
among the genotypes to prove beneficial -- one wishes to re-introduce
some genetic diversity, but dogs and cats won't profitably crossbreed,
so to speak.  I therefore have avoided an embarrassingly parallel
implementation of the GA section so far, but will probably try one in
the future.

    rgb

> 
> Robert G. Brown writes:
>  > ago.  Although there is some advantage to beowulfery in parallel
>  > evaluations of the fitness, there is enough serial work that the most
>  > gain one is likely to be able to realize -- as determined by Amdahl's
>  > Law -- appears (for our code, at any rate) to be between 10 and 20,
>  > depending on the size and scale of the training set.  I'm hoping to
>  > parallelize the GA itself more directly and up the gain to a higher
>  > fraction, but there are definitely some serial parts to the code.
> 

Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb@phy.duke.edu