Created by W.Langdon from gp-bibliography.bib Revision:1.7954
In this thesis, we propose a general geometric framework that addresses these two important problems. The unification is made possible by surprisingly simple representation-independent geometric definitions of crossover and mutation using the notion of distance associated with the search space. This novel way of looking at genetic operators allows us to rethink various familiar aspects of evolutionary algorithms in a very general setting, simplifying and clarifying their relations. We show that many important genetic operators for the most frequently used representations fit this framework. This makes this framework highly relevant because it unifies pre-existing evolutionary algorithms. The abstract definitions of mutation and crossover can be used as formal recipes to build new mutations and crossovers for virtually any new solution representations and problems. We designed and tested new operators on a number of problems obtaining very good experimental results. The same abstract definitions of mutation and crossover can be used to build a truly general representation-independent theory of evolutionary algorithms. We started building such a theory and showed that all evolutionary algorithms with geometric crossover does the same type of search, convex search. This is a general and important result because it shows that a non-trivial representation-independent theory of evolutionary algorithms is possible.",
ISNI: 0000 0001 3418 9612 Supervisor: Riccardo Poli",
Genetic Programming entries for Alberto Moraglio