You know how annoying the web is? A few years back I came across a paper
from a group in France or Swizerland (I think) I cant remember their names.
Anyway, I have had no luck in finding it or anything similar and wondered if
anyone else has a clue to who or where I should look.
The idea was to evolve NN's using GA's. The key part was that these guys had
developed a language which created the NN itself, but the GA operated on the
language string. The language, they could show, could encode any valid NN,
and only valid NN's would result from transformations of the string. The
langauage was also a significant compression of the NN description and so
provided for a very efficient GA search. Neat idea (well I thought so).
Alan Penn
UCL
|