Almost-parameter-free Differential Evolution
Xiaowei Zhang, Sanyang Liu · 2011
A fit parameter setting usually improves greatly the performance of Differential Evolution, hence various strategies for parameter setting have been proposed. However, how to set these is nuisance because each of the strategies only outperforms the other one in some specific aspect. In the paper, an almost-parameter-free Differential Evolution is presented. The algorithm thinks of each individual as a charged particle and utilizes the attraction-repulsion mechanism among them to decide on the step length of the motion of the individual in the direction of the difference. Moreover, Taguchi's parameter design method with the two-level orthogonal array is used to execute the crossover operation for the purpose of obtaining the better combination of factor levels. The proposed algorithm has small population size, and avoids the settings of the scale factor and the crossover probability. Numerical experiments show that the proposed algorithm outperforms the other compared algorithms.