Comparing Approaches to Noisy Optimization

Abstract

In the previous chapters, we have studied the performance of various evolution strategies on the noisy sphere. We have seen the beneficial effects of overvaluation that follow from failure to reevaluate parental fitness along with the problems for success probability-based mutation strength adaptation rules that result. We have investigated the gain that can be achieved by using populations of candidate solutions that are distributed in search space and the benefits of genetic repair in the presence of noise. With the analysis of the behavior of the (µ/µ,λ)-ES with cumulative mutation strength adaptation in Section 3 of Chapter 5, the dynamics of a complete evolution strategy in the presence of noise are now understood. However, evolution strategies are by no means the only algorithms used for optimization in the presence of noise, and nothing has yet been said with regard to their capabilities relative to those competing approaches. It is therefore the goal of this chapter to contrast the performance in the presence of noise of evolution strategies with that of other algorithms that are used frequently or devised explicitly for noisy optimization.