Etude du couplage entre un algorithme génétique et des méthodes d'optimisation locale

Abstract

The global optimization of expensive functions is the field in which this work is carried out. The industry is highly interested in solving such problems bacause it needs to handle more and more with functions of this kind. They are the result of high fidelity computer experiments. This master's thesis was carried out in partnership with the research center Cenaero. The purpose of this master's thesis is apply hybrid algorithms to optimize the computationnaly expensive functions. Hybrid algorithms combine a global search method, such as a genetic algorithm, with a local search method. There are many ways to build hybrid algorithms due to the choice of the local method used in the coupling and also because of the way of combining the two methods. Hybrid methods have been developped in the software Minamo. The surrogate-assisted genetic algorithm (using RBF surrogates) of this software takes part in our hybrid methods as global method. We have tested the coupling of the genetic algorithm with four local methods which belong to the trust-region family. We implemented two of these local method in the framework of the work while existing softwares have been used for the two other ones. To compare the performance of the hybrid methods, against each other but also compared with the surrogate-assisted genetic algorithm used alone, some tests on mathematical functions were performed. The hybrid methods' effectiveness has been proved, as far as the quality of the solution is concerned but also about the execution time of the algorithms. We also remark that for objective function of different kind, different hybrid methods were the most efficient.

Date of Award	24 Jun 2011
Original language	French
Supervisor	Annick Sartenaer (Supervisor), Caroline SAINVITU (Co-Supervisor), Philippe TOINT (Jury) & Anne Lemaitre (Jury)