Re: Genetic Operators for Neuroevolution

On Tue, 22 Aug 2006 04:08:25 +0000, I. Myself wrote:

/.../

In 3, two random points are picked, and then one end of each chromosome
is moved to the middle of the other. The original middle portions are
moved to become the new ends. The end that is moved to the middle can
be the top or bottom end.

If, in the fitness evaluation of your chromosome (perhaps via some
"phenotype"), the /absolute position/ of alleles on the chromosome is
significant - as is most often the case in GA encodings - then sequences
of alleles that are shifted along the chromosome as you describe are going
to be seen "out of context" by the evaluation procedure and amount to
little more than random noise; a "macromutation", if you like.

/.../

All of the above are suggested by biology. Real biological chromosomes
are very long, and typically break and are reattached in many places as
a result of mating. Most of the segments are reattached in
corresponding positions, as in my types 1 and 2 above, but sometimes
pieces get attached elsewhere, as in 3 above.

Yes; but biological expression of chromosome is in general not absolute
position-dependent; rather, it may depend on "marker" sequences and
suchlike. If your evaluation procedure works in a similar way then you may
disregard my argument above.

/.../

There is one major difference between operator 3 and the others, and
that becomes evident when the population becomes homogeneous. If all
chromosomes are identical, then 1 and 2 have no effect whatsoever! This
actually happens in practice; after many generations of evolution all
diversity is lost except the small changes due to mutation.

This is a general feature of GAs; genetic diversity is lost due to a
combination of selection pressure (new high-fitness genotypes tend to
"swamp" a population quickly) and genetic drift (genotypes are "lost" by
randomness in the selection procedure). This loss of diversity, as you
have probably discovered, can be disconcertingly rapid, particularly in
smallish populations. And yes, of course recombination becomes ineffective
in a homogeneous population.

However, operator 3 will create major changes to chromosomes even in
this case. This means that diversity of a population cannot be totally
lost.

But if my argument above applies then these "major changes" are equivalent
to randomisation... maybe a good thing, maybe not, but why not just cut
out the middleman and randomise a few chromosomes instead...?

--
Lionel B

.