W. B. Langdon and R. Poli
1. In the text it is assumed, that the number of representations for the
same fitness increases with the length. Can this be estimated in any way?
And what happens, e.g. if a No-Operation-node is allowed in the search?
Will GP tend to use it more and more, as learning proceeds, and as your
explanation suggests, too?
2. Maybe, there is additional reason for bloat: the maximum number of
incoming connections to a node. Does it make a difference in bloat, if one
uses only node functions with max three inputs, or max four inputs? On
later stages of the genetic algorithm, the probability increases to get a
node function with a larger number of inputs into the population, hence the
trees are getting a more complex structure in order to "supply" the node
with subtrees.
Maybe, this will not cause bloat, but enhance the effect of bloating. Did
you ever hear about works related to this subject?
Best regards
Mario Koeppen
-- ---[]--[][][]--[]--[]-------- Mario Koeppen ---------------------------------- [] [] [] [] [] [] [][][] [][] Department for Pattern Recognition [] [] [] [] Fraunhofer IPK-Berlin [] [] [] [] Pascalstr. 8-9, 10587 Berlin, Germany ---[]--[]------[]--[]---------------------------------------------------------phone (++49)(0)30 390 06 200 fax (++49)(0)30 391 7517 email mario.koeppen@ipk.fhg.de homepage http://vision.fhg.de/~ipk/koeppen ------------------------------------------------------------------------------ "In Computer Vision, we do not solve the problems; we just get tired of them!" Rama Chellapa