Differential evolution with M-fitness method

Abstract

Differential Evolution(DE) is a powerful algorithm to solve global optimization problems. Because the optimization process of original DE is quite easy to understand and code, it has been widely applied in many fields. In recent years, many adaptive parameters DEs have been proposed and achieved better performance on many problems. But simplicity and parallelism of DE have been decreased in those adaptive DE, so they can't be easily transferred to other fields. Moreover, adaptive parameter mechanisms don't always perform better compared with some popular parameter settings. To enhance the performance while maintaining the simplicity and parallelism of DE algorithm, in this paper, we introduce a m-fitness method. The method we proposed use distribution information of fitness value to tune p-value which is a parameter used in DE/pbest/1 to control convergence speed. Moreover, in the method, the information also has been used in selection phase by using half-meanfit selection we proposed in paper. DE with m-fitness method(mDE) is compared on benchmark functions with classical DE and some representative adaptive DE. The results show that the DE with m-fitness method is competitive with other various DE in performance.