Nonstationary and Noisy Function Optimisation

Abstract

Unlike most of the examples we have used so far, real-world environments typically contain sources of uncertainty. This means that if we measure the fitness of a solution more than once, we will not always get the same result. Of course, biological evolution happens in just such a dynamic environment, but there are also many EA applications in environments featuring change or noise when solutions are evaluated. In these nonstationary situations the search algorithm has to be designed so that it can compensate for the unpredictable environment by monitoring its performance and altering some aspects of its behaviour. An objective of the resulting adaptation is not to find a single optimum, but rather to select a sequence of values over time that maximise or minimise some measure of the evaluations, such as the average or worst. This chapter discusses the various sources of unpredictability, and describes the principal adaptations to the basic EA in response to them.