Análise experimental de operadores de recombinação para o algorítimo de evolução diferencial

Abstract

The search for improved solutions is a constant goal in all fields of engineering. Consequently, there is a great interest in the development of robust and efficient optimization techniques, to enable the solution of a wide range of problems in a fast and effective fashion. Amongst the algorithms commonly employed for the solution of optimization problems in engineering, Evolutionary Algorithms (EAs) are widespread across several disciplines. More specifically, the method known as Differential Evolution (DE) has yielded good performance for continuous optimization problems, particularly in low-dimensional spaces, and representsan interesting balance between simplicity of implementation and quality of the solutions found. As is the case with the vast majority of EA approaches, the performance of DE is strongly influenced by the specific variation operators (recombination and mutation) employed, aswell as by the values used for the algorithm parameters. This work presents a systematic investigation of the effects of sixteen recombination operators for real-coded spaces on the performanceof differential evolution. A unified notation is adopted for all operators, based on vector operations, and a modular, standardized implementation is presented in the form of an open-source package in R. The goal is to simplify the analysis of similarities among different operator variants, as well as the understanding of their effects on a population of candidate solutions. Another objective is to provide a standardized platform for the development and evaluation of new operators for Differential Evolution. To investigate and compare the performance of the recombination operators investigated in this work, a 28-problem benchmark set is used with dimensions 5 and 20. The results are discussed in terms of promising directions for the development of new operators for DE. The results also suggest that some recombination variants so far unused in the DE literature may be recommended in terms of improving the expected performance of the algorithm in problems of the two dimensions tested.