I'm actually interested in the "hybrid" notion that Frederik was asking
about (as opposed to the co-evolutionary solution suggested by Sean).
My work involves neural networks that have both qualitative and
quantitative characteristics. That is, they have topological structure that can
be evolved with the traditional GP Tree-based approach. However, the
high-dimensional weighted connectivity between nodes, and parameterized
transforms within the nodes, can be efficiently optimized using other
traditional ANN techniques. The way I’ve been dealing with this in the
past, is to evolve networks using GP to “explore” the structural solution space,
and then rely on training (evaluation) to exploit the available solution
sub-space which can be done using GA, gradient descent, and so on. This is
obviously computationally expensive, no matter how you dice it.
I wonder if an implementation of HyperNEAT
evolution could be added to ECJ that would efficiently blend these things in an
elegant way? Has anyone out there tried this?
The robotics sensor issues Frederik mentioned, despite my lack of knowledge
in that problem domain, sound very familiar to me.
Sent: Monday, November 14, 2011 11:49 AM
Subject: A Mixed GA/GP individual
I've been experimenting with ECJ for some weeks now. I have a fellow
code who already succeeded to do develop robot controllers using GP
and an extern simulator.
What I want to try now is to evolve some kind of indivual where the
is a GP individual (subpopulation 0?) but the senors are coded into a
vector of bits
or integers (subpopulation 1?) . Breeding should be between vectors or
of the same individuals. Is there an easy way to do such a thing? Is there
similar predifined parameter file for this?
Thank you in advance,
With kind regards,