Ordinal regression in evolutionary computation

Abstract

Surrogate ranking in evolutionary computation using ordinal regression is introduced. The fitness of individual points is indirectly estimated by modeling their rank. The aim is to reduce the number of costly fitness evaluations needed for evolution. The ordinal regression, or preference learning, implements a kernel-defined feature space and an optimization technique by which the margin between rank boundaries is maximized. The technique is illustrated on some classical numerical optimization functions using an evolution strategy. The benefits of surrogate ranking, compared to surrogates that model the fitness function directly, are discussed. © Springer-Verlag Berlin Heidelberg 2006.