John Douglas Porter wrote:
> Juan J. Merelo wrote:
>
>>The ideal is supposed to be ~ 20/80, but some
>>purists say it should be ~ 1/99.
>>
>
> There is no theoretical ideal.
> Anyone who says there is is either wrong, or is speaking in the
> context of some very specific experimental protocol.
>
>
You are right... of course, there's no ideal, but the rule of thumb says
that mutation shouldn't be too high, or you're turning an EA into a
random search algorithm.
>
>>Genetic programming, actually, uses
>>only some very specific kind of mutations,
>>
>
> Well, sure; so does GA. Flipping a bit is very specific, yes?
>
Yep, right.
>>[GP] makes up for lost diversity by creating a huge and diverse initial
>>population.
>>
>
> No, that's not true -- at least, no more so than for GA.
> GP does not suffer from "loss of diversity" any more than GA.
> I would contend that it suffers from the opposite problem;
> without heuristic constraints, the genome space in GP is
> effectively unbounded (since individuals can grow without bound).
It's usually bounded, right? Not theoretically, but most experiments
limit tree depth.
> One simple way to see this is to consider that in (standard) GA,
Let's say that it does in canonical, Goldberg-style, GA. In many
problems, nowadays, size-changing genomes are used (for instance,
evolving neural nets).
>>OPEAL comes closest to it, if only I find the time
>>to upload it.
>>
>
> You probably already know this, but I'd point out that before
> you upload OPEAL to CPAN, you'll need to fix its naming scheme.
> Almost all your modules have top-level names.
>
Of course they do; I don't want to commit to a namespace before knowing
which namespace it is. I have already proposed Algorithm::Evolutionary
and AI::Evolutionary or AI::EC (for Evolutionary Computation).
Which one would you vote for? Evolutionary computation is not purely an
artificial intelligence technique, that is why it's better to just leave
it in the Algorithm namespace.
J
--
PPSN2002 => http://ppsn2002.ugr.es
Home => http://geneura.ugr.es/~jmerelo
Tutorial Perl => http://granavenida.com/perl
