Non-parametric model of the space of continuous black-box optimization problems

Abstract

Exploratory Landscape Analysis are data driven methods used for automated algorithm selection in continuous black-box optimization. Most of these methods follow strong assumptions that limit their characterization power, or loose information by compressing the data into a few scalar features. A more flexible approach is to avoid explicit measuring and comparing of specific structures. In this paper we present a proof-of-concept for a more general method, which produces non-parametric models of the space of problems. Using non-metric multidimensional scaling, we generate synthetic features for each problem, which could replace or complement the existing ones. We demonstrate approaches to prod..