Design of large-scale metaheuristic component studies

Abstract

Metaheuristics employ a variety of different components using a wide array of operators to execute their search. This determines their intensification, diversification and all other behavioural features and is thus critical for success on different optimisation problems. Choosing the right components with the right operators remains a difficult task. In this paper we propose a design of experiments that should be used for extensive component studies. We demonstrate the applicability of this design by exploring the differences in operator specific performance in two closely related metaheuristic frameworks - -the well-known (+ ?)-Evolution Strategy and the strongly metaphor-focussed Invasive Weed Optimisation - -where operators show varying degrees of similarity in different components. This experiment shows that similarity of operators does not comprehensively account for similarity in performance. Presumably small changes of an operator can influence the algorithmic behaviour more than the utilisation of a completely different operator in another component. Even when employed in different combinations, these influences remain strong. This emphasises the need for a more detailed analysis of the specific effects of components and their respective operators on the search process.