Journal article 338 views 53 downloads
Predicting effective control parameters for differential evolution using cluster analysis of objective function features
Journal of Heuristics, Volume: 25, Issue: 6, Pages: 1015 - 1031
PDF | Version of Record
Released under the terms of a Creative Commons Attribution 4.0 International License (CC-BY).Download (1.21MB)
A methodology is introduced which uses three simple objectivefunction features to predict effective control parameters for differential evolution. This is achieved using cluster analysis techniques to classify objectivefunctions using these features. Information on prior performance of variouscontro...
|Published in:||Journal of Heuristics|
Springer Science and Business Media LLC
Check full text
No Tags, Be the first to tag this record!
A methodology is introduced which uses three simple objectivefunction features to predict effective control parameters for differential evolution. This is achieved using cluster analysis techniques to classify objectivefunctions using these features. Information on prior performance of variouscontrol parameters for each classification is then used to determine which control parameters to use in future optimisations. Our approach is compared tostate–of–the–art adaptive and non–adaptive techniques. Two accepted benchmark suites are used to compare performance and in all cases we show thatthe improvement resulting from our approach is statistically significant. Themajority of the computational effort of this methodology is performed off–line, however even when taking into account the additional on–line cost ourapproach outperforms other adaptive techniques. We also study the key tuning parameters of our methodology, such as number of clusters, which furthersupport the finding that the simple features selected are predictors of effectivecontrol parameters. The findings presented in this paper are significant becausethey show that simple to calculate features of objective functions can help toselect control parameters for optimisation algorithms. This can have an immediate positive impact the application of these optimisation algorithms on realworld problems where it is often difficult to select effective control parameters.
Originality: The first results showing that a statistically significant improvement in performance can be achieved for optimisation algorithms using automatic problem classification.Rigour: The paper went through a lengthy and rigorous peer review process because our results were so significant. A total of 67,800 experiments were run over a variety of established benchmark functions, repeated for different random seeds. Significance: Our results show that you can use information about a function to predict how to tune your optimisation algorithm for it. This is a step towards fixing a key limitation of many gradient free optimisation algorithms – problem dependent tuning.
tuning evolutionary algorithms, differential evolution, unsupervised learning, optimization methods, evolutionary algorithms