Auto-selection mechanism of differential evolution algorithm variants and its application

• Published in: European journal of operational research Vol. 270; no. 2; pp. 636 - 653
• Main Authors: Fan, Qinqin, Yan, Xuefeng, Zhang, Yilian
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 16.10.2018
• Subjects: Differential evolution; Evolutionary computations; Multi-algorithm selection; Parameter estimation; Differential evolution; Evolutionary computations; Parameter estimation; Multi-algorithm selection
• ISSN: 0377-2217; 1872-6860
• DOI: 10.1016/j.ejor.2017.10.013

•The proposed approach can automatically assign the computational resource to each component algorithm.•Our proposal is used to estimate eight parameters of a heavy oil thermal cracking model.•Three famous test suites are utilized in the experiments. Each type of problems, such as unimodal/multimodal, linear/non-linear, convex/non-convex, and symmetrical/asymmetrical, has its own characteristics. Although various differential evolution (DE) variants have been proposed, several studies indicate that a DE variant may only exhibit high solution efficiency in solving a specific type of problems, but may perform poorly in others. Therefore, an important decision is made to automatically select a suitable DE variant among several chosen algorithms for solving a particular type of problems during the evolutionary process. To achieve this objective, an auto-selection mechanism (ASM) is introduced in this study. In the ASM, rankings attained using Friedman's test are adopted to assess the performances of DE variants. A learning strategy is employed to update the choice probabilities of DE variants, and an additional selection probability is used to alleviate the greedy selection issue. Three sets of benchmark test functions proposed in BBOB2012, IEEE CEC2005, and IEEE CEC2014 are used to evaluate the effectiveness of the ASM. The performance of the proposed algorithm is also compared with that of nine state-of-the-art DE variants and four non-DE algorithms. Statistical analysis results demonstrate that the ASM is an efficient and effective method that can take full advantages of multiple algorithms. Furthermore, the ASM is utilized to estimate the parameters of a heavy oil thermal cracking model. Experimental results indicate that the proposed algorithm outperforms the other compared algorithms in this case.