Empirical Investigations on the Potential of Differential Evolution Based Algorithm Portfolios

Abstract

In spite of the proliferation of effective evolutionary algorithms, there are no guidelines on choosing a suitable algorithm for a given optimization scenario. Consequently, a typical EA practitioner resorts to a trial-and-error search of appropriate algorithms. In addition, EA practitioners prefer an algorithmic framework that is capable of solving a diverse range of problems. To address these problems, population-based algorithm portfolio framework has been proposed in EA literature. Often such proposals identify an effective portfolio and benchmark it against existing algorithm to demonstrate its superiority. However successful design of robust algorithm portfolios demands a thorough understanding of their search dynamics. Towards this, the thesis proposes to observe and understand the robust potential as well as the performance efficacy of combining classical Differential Evolution (DE) variants with diverse characteristics in a distributed framework (i.e. Differential Evolution algorithm portfolios) through systematic empirical analyses. The empirical comparative performance analysis of DE portfolios formed by combining 5 classical DE variants has been carried out. The robustness of those portfolios towards variations in algorithmic structure and problem characteristics has also been observed. Further, simulation experiments to empirically evaluate the performance efficacy of different minimal DE portfolios (portfolios with 2 constituent algorithms), formed by varying mutation, crossover and selection operations of the constituent DE variants have been carried out. The portfolios displayed a very competitive performance by outperforming their constituent DE variants thus revealing the robustness manifesting from combining algorithms with different search characteristics. It is imperative that this understanding of minimal DE portfolios will facilitate successful construction of robust algorithm portfolios. newline