Multi-objective optimization with estimation of distribution guided by multi-criteria decision making

Abstract

This research project proposes extensions to meta-heuristics for multi-objectiveoptimization (MOO), employing multi-criteria decision making (MCDM) techniques toguide the synthesis of estimation of distribution models. MCDM techniques takeadvantage of a priori preferences of the decision maker to properly rank solutions thatare non-dominated among themselves. Notice that those non-dominated solutions areusually considered of an equivalent merit in the literature. By ranking non-dominatedsolutions, estimation of distribution models are expected to be capable of indicatingmore promising regions of the search space, thereby promoting further progress in thesearch when employing the same amount of computational resources. In principle, thewell-known MCDM technique called TOPSIS (Technique for Order Preference bySimilarity to Ideal Solution) is going to be applied here, whereas the techniques forMOO and estimation of distribution are going to be based on the more advanced andnewly-proposed approaches developed by the research group, as well as competitiveand state-of-the art solutions in the literature, such as NSGA-II for MOO and mixture ofGaussians for probability density estimation. The merit of the proposal will be evaluatedin terms of the required computational resources and the quality of the generatedsolutions, measured from the well-accepted hypervolume indicator. To make the mostreplicable performance comparison, we intend to take as a case study an extensive listof synthetic problems already proposed in the literature and widely used forperformance evaluation on MOO. It is not discarded the application to real problems ofpractical interest, but the knowledge of the Pareto frontier in the case of syntheticproblems facilitates the analysis of merit of the proposal. (AU)