Niall Stevenson Undergraduate Dissertation 2017/18

Evolutionary Algorithms for Noisy Optimisation

Supervised by D.Sudholt

Abstract

The ability of evolutionary computing to adapt robustly to noisy and uncertain environments is an increasingly popular subsection of the overall discipline as algorithms are more and more commonly being applied to stochastic and uncertain real world situations. Previous work has shown that recombination across the population can provide more accurate results than mutation alone. There is room for expansion on previous work with regards to the explicit usefulness of crossover and the implicit usefulness of population, both of these with regards to the (mu +1) and Compact Genetic Algorithms- and to a lesser extent the (1+1) algorithm. Simple exploration is made into basic problems complicated through the use of additive posterior Gaussian noise of the form N(theta, sigma^2). The aim of this project is to further identify the key features of evolutionary strategies that make them robust in the presence of noise through developing software which can deploy these strategies in problem models.