New uncertainty handling strategies in multi-objective evolutionary optimization

Abstract

Since many real-world optimization problems are noisy, vector optimization algorithms that can cope with noise and uncertainty are required. We propose new, robust selection strategies for evolutionary multi-objective optimization in the presence of noise. We apply new measures of uncertainty for estimating the recently introduced Pareto-dominance for uncertain and noisy environments (PDU). The first measure is the inter-quartile range of the outcomes of repeated function evaluations. The second is based on axis-aligned bounding boxes around the upper and lower quantiles of the sampled fitness values in objective space. Experiments on real and artificial problems show promising results. © 2010 Springer-Verlag.