Abstract: |
A powerful approach to search is to try to learn a distribution ofgood solutions (in particular of the dependencies between theirvariables) and use this distribution as a basis to sample new searchpoints. Existing algorithms learn the search distribution directly onthe given problem representation. We ask how search distributions canbe modeled indirectly by a proper choice of factorial geneticcode. For instance, instead of learning a direct probabilistic modelof the dependencies between variables (like BOA does), one canalternatively learn a genetic representation of solutions on whichthese dependencies vanish. We consider questions like: Can everydistribution be induced indirectly by a proper factorialrepresentation? How can such representations be constructed from data?Are specific generative representations, like grammars or L-systems,universal w.r.t. inducing arbitrary distributions? We will considerlatent variable probabilistic models as a framework to address suchquestions and thereby also establish relations to machine learningconcepts like ICA. |
