Using achievement scalarization function for adaptive pareto ranking
Lavinia Eugenia Ferariu · 2016
Conciliating between multiple objectives involves managing partially sorted solutions. In this regard, the main difficulties consist in finding proper ranking strategies able to preserve the genetic diversity, while guiding the search towards the areas preferred by practitioners (e.g. the middle of the best Pareto-fronts). This paper presents a new approach based on adaptive ranking. The adaptive techniques analyze the layout of fronts specific to the current population, in order to extract necessary information about the conflicting relationship between objectives and the behavior of genetic search. This feedback is used for configuring the preferred area and the ranking policies, as well. In this context, one contribution refers to the procedure proposed for investigating the distribution of fronts. It monitors the layout of fronts via distances based on activation scalarization function, which are likely non-conflicting with the closeness to the Pareto-optimal solutions. The benefits provided by this metric are also exploited by two other new procedures. One of them labels the preferred solutions by directly controlling the extent of the best fronts included in the preferred area. The other one corrects the spread of the best fronts via ranks' adjustment. The performances of the approach are illustrated on several bi-objective optimizations involving weakly/strongly conflicting objectives, as well as different layouts of Pareto-fronts.