Re: Lineage selection

Greetings,

On Tue, 08 Nov 2005 00:13:04 +0200, hmm wrote:
> Hi
>
> I am planning to use "lineage" selection in my EA for pairing cross-overed
> individuals.
> Idea is my own, so I am asking if somebody here knows any research allready
> done this area.
> I think it is very biologically oriented way to pair individuals.
>
> I came this idea when I thoug that I should not cross-over ants and
> elephants. So if I want to have one big population in my simulation but
> support different species only way I see to do it is calculate structural
> difference or do it lot more simple way and calculate "how far" relatives
> two individuals are.

Can you not take a cue from the "conventional" biological definition of
species? That is, two organisms belong to the same species iff they are
able to produce viable offspring. In an EA scenario, a "viable offspring"
might be interpreted as a "crossed-over individual of better than terrible
fitness" (if - as is likely the case - your crossover is
non-deterministic, you might have to average over the fitness of
several/all possible crossed-over offspring of a pair of individuals). So
in this sense, the "distance" between two individuals is the
inverse of the (average) fitness of crossed-over offspring of those
individuals.

> Any ideas how should I calculate fast/easy/small_memory_usage way how
> far two "relatives" are?

Admittedly, the scheme described above may be computationally expensive...

> In my current implementation http://sourceforge.net/projects/narugo I
> try to keep individual relative trees but I am considering simpler and
> more effective way.

....but perhaps less so than maintaining a full record of ancestry?

> I see it biologically realistic crossover individuals close to 3-100
> relative steps so that possibility goes very small after 100.

If I understand you correctly, you are using something like "number of
generations since most recent common ancestor" as a measure of
"biological distance" between individuals. That's a nice idea, but I'm not
quite convinced that it is "biologically realistic"... in nature, for
example, a species that becomes geographically split may evolve into two
species (in the sense of individuals from the two "demes" being unable to
produce viable offspring) relatively rapidly in evolutionary time - that
is, in a comparatively small number of generations. Thus two individuals
may have a relatively recent most recent common ancestor while
yet belonging to different species.

I suspect another potential problem in your approach: in an EA scenario,
if you restrict mating to be between individuals which are "related" -
read "similar" - then they are likely to differ at a comparatively small
number of loci. In which case, crossover is likely to become less
effective at producing "novel" genotypes. This perhaps reflects a
fundamental difference between the roles of crossover in artificial and
natural evolution; in the former recombination tends to be viewed as a
(primary?) mechanism for the production of novel, potentially
high-fitness, genotypes. In natural evolution the role of recombination is
still hotly debated but, since natural species tend to be *very* highly
genetically converged, recombination is perhaps less likely to fulfil a
similar function.

Yet another (related) caveat: there is an inherent difficulty in EAs in
maintaining distinct species within a single population. Any successful
(i.e. high fitness) individual will have a tendency to take over the
population surprisingly quickly and thoroughly. That is, subsequent to a
successful innovation, the *entire population* will very quickly be in its
entirety descended from that one successful (common, recent) ancestor and
you are left with a single "species"... the literature abounds with
schemes for addressing this issue (cf. "nicheing", geographical GAs, etc.).

Regards,

--
Lionel B


.

Relevant Pages

  • Re: Defense requested
    ... >>splitting into two species. ... divergence we actually observe. ... >>humans, chimps, and gorillas (which are usually placed in three separate ... > primates share a common ancestor, and then use that as a platform to ...
    (talk.origins)
  • Re: Defense requested
    ... >>splitting into two species. ... assertions of evolution. ... Humans and chimps share a common ancestor. ...
    (talk.origins)
  • Re: Missing links? Eh?
    ... Once you get down to the species level there are ... >>>all the descendants of that most recent common ancestor. ... > Phylogeny and Cladistics in Ancestral Primate Taxonomy." ... Why higher taxa are needed in evolutionary studies. ...
    (talk.origins)
  • Re: A "Dover" sidetrack
    ... >>> others, isn't there some midway point for certain species, as Pandas ... >>common ancestor. ... Because if evolution doesn't stop, ... a plant, or even a bacterium. ...
    (talk.origins)
  • Re: A "Dover" sidetrack
    ... All species are related by common ... >>common ancestor. ... expectation that evolution would predict a different result. ... equidistant from bacteria? ...
    (talk.origins)