IUBio GIL .. BIOSCI/Bionet News .. Biosequences .. Software .. FTP

simulating evolution

JMILLER%VXBIO.SPAN at STAR.STANFORD.EDU JMILLER%VXBIO.SPAN at STAR.STANFORD.EDU
Fri Aug 30 13:11:19 EST 1991


 Ray's work is part of a much larger effort in what are called "genetic
algorithms". There was a meeting in San Diego a few weeks ago, with
many papers presented. In general, they all evolve programs by a
mutation/recombination/selection scheme. The reason Rays's work is 
interesting is because he didn't try evolving a set of machine language
instructions specific to his computer. Instead, he created his own
virtual cpu, with an instruction set quite different from typical
real cpus. One of the main things it supports is a diffusion/collision
model for string interaction, rather than pointers,JMPs, etc. The
instruction set is also quite small, and simple compared to a
typical cpu. It seems likely this is why he got such interesting results--
he seems to have bypassed the "brittleness" problem found in
some other Genetic algorithm studies, not to mention the automatic
programming attempts from a decade ago.
 Genetic algorithm work, in turn, is part of the emerging discipline
of "artifical life". A good introduction to this is Artifical Life,
Chris Langton editor, proceedings volume in the Santa Fe institute 
Studies in the sciences of Complexity (Addison-Wesley).
 By the way, Hillis is still doing very interesting work. His latest
work pretty much overcame the "local optima" problem in evolving a
program, by adding parasitic programs. The host-parasite interaction
keeps the algorithm from getting stuck at local optima,  and the
resulting programs compare favorably with those produced by conventional
means.
 I might add that Ray's work is unrealistic, because it is essentiall
y interacting genotypes, with little genotype/phenotype distinction.
-Peter Markiewicz




More information about the Bio-soft mailing list

Send comments to us at archive@iubioarchive.bio.net