Probability Matching is one of the most successful methods for adaptive operator selection (AOS), that is, online parameter control, in evolutionary algorithms. In this paper, we propose a variant of Probability Matching, called Recursive Probability Matching (RecPM-AOS), that estimates reward based on progress in past generations and estimates quality based on expected quality of possible selection of operators in the past. We apply RecPM-AOS to the online selection of mutation strategies in differential evolution (DE) on the bbob benchmark functions. The new method is compared with two AOS methods, namely, PM-AdapSS, which utilises probability matching with relative fitness improvement, and F-AUC, which combines the concept of area under the curve with a multi-arm bandit algorithm. Experimental results show that the new tuned RecPM-AOS method is the most effective at identifying the best mutation strategy to be used by DE in solving most functions in bbob among the AOS methods.
|Title of host publication
|15th Intl Conf. on Parallel Problem Solving from Nature
|Subtitle of host publication
| (PPSN 2018)
|Number of pages
|Published - Sept 2018
© Springer Nature Switzerland AG 2018. This is an author-produced version of the published paper. Uploaded in accordance with the publisher’s self-archiving policy. Further copying may not be permitted; contact the publisher for details.