- author = "Alberto Moraglio",
- title = "Towards a Geometric Unification of Evolutionary Algorithms",
- school = "Department of Computer Science, University of Essex",
- year = "2007",
- address = "UK",
- month = nov,
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://eden.dei.uc.pt/~moraglio/Thesis_final.pdf",
-
URL = "
http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.446045",
- size = "392 pages",
-
abstract = "Evolutionary algorithms are successful and widespread
general problem solving methods that mimic in a
simplified manner biological evolution. Whereas most of
evolutionary algorithms share the same basic
algorithmic structure, they differ in the solution
representation - the genotype - and in the search
operators employed - mutation and crossover - that are
representation specific. In the research community
there is a strong feeling that the evolutionary
computation field needs unification and systematisation
in a rational framework to survive its own exceptional
growth of the last two decades. The lack of a common
formal framework encompassing all solution
representations have prevented EC researchers to build
a truly general theory of evolutionary algorithms, as
well as to develop a formal theory of search operators
design for new representations and problems.
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.",
-
notes = "Some chapters on GP trees.
ISNI: 0000 0001 3418 9612 Supervisor: Riccardo Poli",
