Open Access

Helena Stegherr

• Metaheuristic optimisation algorithms are often benchmarked and compared based purely on their performance. Other information that results from executing these algorithms has been discarded for a long time but within the last years, this detailed data on the search behaviour of metaheuristics gains more and more attention, emphasising its usefulness for a variety of research aspects. It increases the understanding of the search process of metaheuristics and informs on the effects of algorithmic components. This knowledge can then be utilised to develop improved strategies, to provide details for automated algorithm design approaches and to explain the search process to users, which can ultimately increase their trust in metaheuristics. As research on metaheuristic search behaviour is in parts still in its infancy, it is not yet established what is required to get comprehensive insights into the algorithms. There is no consistent strategy to describe and implement metaheuristics toMetaheuristic optimisation algorithms are often benchmarked and compared based purely on their performance. Other information that results from executing these algorithms has been discarded for a long time but within the last years, this detailed data on the search behaviour of metaheuristics gains more and more attention, emphasising its usefulness for a variety of research aspects. It increases the understanding of the search process of metaheuristics and informs on the effects of algorithmic components. This knowledge can then be utilised to develop improved strategies, to provide details for automated algorithm design approaches and to explain the search process to users, which can ultimately increase their trust in metaheuristics. As research on metaheuristic search behaviour is in parts still in its infancy, it is not yet established what is required to get comprehensive insights into the algorithms. There is no consistent strategy to describe and implement metaheuristics to make extensive conceptual and empirical comparisons possible. For an empirical behaviour analysis, appropriate corresponding measures of the behaviour are required and although there is a wide variety of such measures, their usefulness is not evaluated in a wider context. Additionally, an empirical analysis of several behaviour characteristics and relating those to specific metaheuristic components produces large amounts of data, which causes many common analysis methods to reach their limits. This work advances the analysis of metaheuristic behaviour by exploring strategies to alleviate those problems. For this, a unified, component-based view on metaheuristics and criteria that can discern these algorithms are investigated to enable a conceptual examination. The unified, component-based view is then utilised as the basis for the development of the software framework MAHF that enables extensive experiments with metaheuristics to analyse their search behaviour in relation to the algorithmic components, but also to determine useful behaviour characteristics. As this process generates large amounts of data, approaches based on unsupervised learning techniques, namely clustering and dimensionality reduction, are analysed that can aggregate the information without loosing important details. The developed foundation for metaheuristic behaviour analysis in relation to the algorithmic components is then applied on a range of algorithmic configurations to examine their behaviour as well as to verify the usefulness of the approaches. It is shown that interesting and unexpected behaviour can be connected to the specific algorithmic components but also that basing behaviour analyses on a combination of different measures provides a more complete understanding. Furthermore, the information from this experiment is used in an additional study where it is used to inform improvement strategies for a common metaheuristic algorithm. Overall, this work highlights the importance of researching metaheuristic behaviour but also strategies that can be applied in this context. There is, however, still much to be done and some important aspects of future research are highlighted.…