Edinburgh Napier University

  Although the use of ensemble methods in machine-learning is ubiquitous due to their proven ability to outperform their constituent algorithms, ensembles of optimisation algorithms have received relatively little attention. Existing approaches lag behind machine-learning in both theory and practice, with no principled design guidelines available. In this paper, we address fundamental questions regarding ensemble composition in opti-misation using the domain of bin-packing as a example; in particular we investigate the trade-off between accuracy and diversity, and whether diversity metrics can be used as a proxy for constructing an ensemble, proposing a number of novel metrics for comparing algorithm diversity. We find that randomly composed ensembles can outperform ensembles of high-performing algorithms under certain conditions and that judicious choice of diversity metric is required to construct good ensembles. The method and findings can be generalised to any meta-heuristic ensemble, and lead to better understanding of how to undertake principled ensemble design.