A new methodology for the automatic creation of adaptive hybrid algorithms

Abstract

Hybrid Evolutionary Algorithms are a promising alternative to deal with the problem of selecting the most appropriate Evolutionary Algorithm for a specific problem. By means of the combination of different heuristic optimization approaches, it is possible to profit from the benefits of the best approach or, even more, to discover synergies between the algorithms that could improve the results of the best performing individual algorithm. Nowadays, there is an active research in the design of dynamic or adaptive combination strategies for hybrid algorithms. However, little research has been done in the automatic learning of the best hybridization strategy. This paper proposes a new methodology for developing intelligent adaptive hybrid algorithms that uses data mining techniques to analyze the results from past executions. The proposed methodology has been evaluated on a well-known benchmark on continuous optimization made up of 19 different functions and several dimensions (50, 100, 200 and 500). Several analyses have been conducted and statistical tests have been used for validating the results. The generated hybrid algorithm has achieved outstanding results, obtaining significantly better results than the MOS algorithm, the most performant algorithm on this benchmark, and the CMA-ES algorithm, one of the reference algorithms in continuous optimization.