Orthogonal Predictive Differential Evolution

Abstract

In traditional differential evolution (DE) algorithms, the perturbation direction of mutation is not sophisticatedly designed, which performs ineffectively or inefficiently for optimizing some complex and large-scale problems. This paper designs an orthogonal predictive mutation scheme to solve this problem. The mutation investigates the landscape near the individuals by using orthogonal experimental design, and then applies factor analysis to predict a promising direction for the individuals to evolve. With a clear sense of search direction, the efficiency of DE is improved. Moreover, the step length of the proposed mutation is adaptively adjusted according to the effect of the prediction, which helps to balance the exploration and exploitation abilities of DE. By employing such a mutation scheme, a novel DE algorithm termed orthogonal predictive DE (OPDE) is proposed in this paper. As OPDE can adopt different kinds of classical mutation schemes for choosing the base vector and calculating the differential vector, we further develop an OPDE family including various OPDE variants. Experimental results demonstrate the effectiveness and high efficiency of the proposed algorithm.