Otimização multiobjetivo com base em processo gaussiano de regressão (Kriging)

Abstract

Objective and constraint functions in engineering optimization problems are, usually, calculated with the aid of complex computational tools such as finite elements or computational fluid dynamics. Thus, each evaluation of these functions can take a significant amount of time. In order to speed up the optimization process involving such time-consuming functions, surrogate models are commonly used. Nowadays, a standard technique to optimize computationally costly functions is the Efficient Global Optimization (EGO). The EGO algorithm was developed in the late 1990s and it is based on the iterative building and improvement of the Kriging surrogate model. At each iteration, a new design, which holds the maximum expected improvement, is sampled. For multiobjective problems, analogous algorithms have been developed from 2005 on. Among those, it can be highlighted the ParEgo (or MEGO) and the EGO based on the expected hypervolume indicator (or just called HEGO). However, such algorithms have some drawbacks. For instance, MEGO has difficulties on finding Pareto fronts that are convex (or with a complex shape) and HEGO has a relatively higher computational cost due to the calculations of the expected hypervolume. Recent works (2011 – 2017) present some alternatives to mitigate these and other limitations, as well as more robust filling criteria (i.e., the choice of the point to be sampled in the next iteration), making the algorithms more efficient. The present thesis is inserted within this context. Here, new multiobjective optimization algorithms are proposed for high computational cost functions based on the Kriging metamodel. In the initial phase of the research, the MVPF (minimization of the variance of the Kriging-predicted front) algorithm was developed, which at each iteration creates a Pareto front using only the metamodels and chooses the project with the highest variance to be evaluated. Then, the SME (sequential minimization of entropy) algorithm was developed, which, instead of selecting the project with the highest variance, chooses the one with the highest Shannon entropy. The main advantages of SME in comparison to classic algorithms are the low computational cost (which does not increase significantly with the number of sampled points) and the speed of convergence (in obtaining a Pareto front). Different test problems are solved and, in almost all of them, the proposed algorithms are superior to MEGO and HEGO. In addition, some engineering problems are solved using the proposed algorithms, such as the optimization of curved fiber orientations in airplane panels and the optimization of geometric parameters in a snap-fit joint. Finally, an important by-product of this work was the publication of a computational package in the R language. This package, called moko (acronym for MultiObjective Kriging Optimization), can be found in the official repository CRAN (The Comprehensive R Archive Network) and easily installed by any user.