Investigação de técnicas eficientes para algoritmos evolutivos multiobjetivo baseados em decomposição

Abstract

Evolutionary algorithms (EAs) based on decomposition have been successfully applied in the optimization of problems with two or three merit functions. Over the last few years, this potential has been also investigated in the context of multi-objective problems. In this sense, this dissertation investigates two promising approaches to increase the performance of algorithms based on decomposition: (i) a systematic model for generating weighted vectors (reference vectors) uniformly distributed; and (i) a transformed weighted Tchebycheff scalarization strategy, which provides a simple and parameter-free control of both convergence and scattering of the approximated alternatives. These techniques are incorporated into the general structure of the Multiobjective Evolutionary Algorithm Based on Decomposition (MOEA/D) and the performance of each is evaluated against other well-known strategies, considering known benchmark problems, i.e., DTLZ1 to DTLZ4 with 3, 5, 8, 10 and 15 objectives. The results indicate that the proposed techniques are competitive when compared to the other approaches evaluated, mainly in relation to the quality indicators Inverted Generational Distance (IGD) and Hypervolume (HV).