What's the difference if the generated algorithm has the desired
fitness? Intelligence is not a pre-requisite for filling a niche. The
solution doesn't need to be elegant, comprehensible or efficient, unless
it's part of the fitness criteria. Brute-force is a good strategy when
there are plenty of brutes (computers) around.
On the other hand, behavior that needs to be understood and maintained
by humans should be as expressive, elegant, etc as possible, unless
people are cheaper than computers. :)
On 10/15/10 7:10 PM, Chris Gahan wrote:
I think so too. Genetic Programming always struck me as quite wasteful
because your algorithm doesn't care about *why* any of your offspring
succeeded or failed.
Intelligence has been a huge evolutionary boon because it allows
critters to learn from the mistakes of others, and then over time
those learnings become baked into the genome. (see "how learning can
guide evolution":
http://htpprints.yorku.ca/archive/00000172/01/hinton-nowlan.htm )
Evolution is really really slow... Which isn't a problem when you're
an insect, and the cost of forking a child process is eating a little
bit of garbage juice.
When you're trying to program with expensive and limited computing
resources however, it makes more sense to encode some knowledge into
the solution-generating system.
On Friday, October 15, 2010, frank<[email protected]> wrote:
I think genetic programming is probably most interesting if you don't
write a mere battery of black-box tests for the fitness evaluation.
It would be much cooler to have code that reasons about the candidat
solutions, and see how much of each is proovably correct, IMHO. =)
Casey Ransberger wrote:
The previous thread about testing got me thinking about this again. One of the
biggest problems I have in the large with getting developers to write tests is
the burden of maintaining the tests when the code changes.
I have this wacky idea that we need the tests more than the dev code; it makes
me wish I had some time to study prolog.
I wonder: what if all we did was write the tests? What if we threw some kind of
genetic algorithm or neural network at the task of making the tests pass?
I realize that there are some challenges with the idea: what's the DNA of a
computer program look like? Compiled methods? Pure functions? Abstract syntax
trees? Objects? Classes? Prototypes? Source code fragments? How are these
things composed, inherited, and mutated?
I've pitched the idea over beer before; the only objections I've heard have been of the form
"that's computationally expensive" and "no one knows how to do that."
Computational expense is usually less expensive than developer time these days,
so without knowing exactly *how* expensive, it's hard to buy that. And if no
one knows how to do it, it could be that there aren't enough of us trying:)
Does anyone know of any cool research in this area?
_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc
_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc
_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc
_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc
