Homologie en Programmation Génétique Application à la résolution d'un problème inverse
Michael Defoin Platel · HAL (Le Centre pour la Communication Scientifique Directe) · 2004
Evolutionary Algorithms (EA) are search methods working iteratively on a population of potential solutions that are randomly selected and modified. Genetic Programming (GP) is an EA that allows automatic search for programs, usually represented as syntax trees (TGP) or linear sequences (LGP). Two mechanisms perform the random variations needed to obtain new programs : the mutation operator (local variation) and the crossover operator (programs recombination). The crossover operator blindly exchanges parts of programs without taking the context into account, this is a "brutal" operation that may be responsible of the uncontrolled growth of programs during evolution. Mainly inspired by the homologous recombinaison of DNA strands, we introduce the Maximum Homologous Crossover for LGP. The MHC ensures, thanks to a measure of similarity, that recombinaison of programs is respectful. We show on classical GP benchmarks, e.g. the symbolic regression problem, that when using MHC the search process is less brutal and that an accurate control of programs size is also possible. These results are used to address a real world problem : the inversion of atmospheric components. We show that, with a constant computational effort, it is also possible to find teams of inversion predictors that outperform standard models.