Uso de metamodelos na evolução diferencial para problemas envolvendo simulações de alto custo computacional

• Autor(a) principal: Silva, Eduardo Krempser da
• Data de Publicação: 2014
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/216
• topic: Otimização estrutural; Evolução estrutural; Metamodelos; Metaheurísitca; Structural optimization; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::MATEMATICA DA COMPUTACAO

In several situations, discontinuities and di_culties in the use of derivatives, as well as the presence of discrete or mixed design variables make the application of classic optimization methods less attractive. These dificulties suggest the application of population-based metaheuristics, and specially those with natural inspiration. Among them, we are interested in the so called Diferential Evolution (DE), which has obtained good results in the literature. Despite of the quality of solutions obtained by DE in several optimization problems, DE cannot be used when the solution evaluation involves a computational expensive simulation. Therefore, even considering the continuous improvement of computational power in recent decades with a cost reduction, the application of metaheuristics such as DE, is deficient in the mentioned cases due to the large number of evaluations/simulations required to obtain a solution. In this context, the application of metamodels, or surrogate models, together with metaheuristics has received a growing attention of researchers of several areas. The metamodels generate a computational model which is simpler to be used in parts of the optimization process, replacing the original model. Its construction is based in previously obtained simulation results during the current optimization process. This work presents a new application strategy of metamodels within DE, which allows for computational cost reduction, besides providing a mechanism to alleviate the user of the variant parameter definition, one of most important definitions in DE. The work presents also an evaluation of di_erent local approximation models used as metamodel in the proposed scheme. Finally, some improvements are proposed by means of Design of Experiments (DoE) techniques and the adaptive selection of metamodel parameters; the DoE are applied aiming for the best way to select the next points to be evaluation and, consequently, generating a better final solution with a lower cost. The other proposed use of Adaptive Operator Selection mechanism is to select the number of radial basis used in the Radial Basis Function Networks applied as metamodel.