A novel Two-layer Hierarchical Differential Evolution Algorithm for Global Optimization Yinzhi Zhou, Xinyu Li, Liang Gao* State Key Laboratory of Digital Manufacturing Equipment &Technology Huazhong University of Science and Technology Wuhan, China gaoliang@mail.hust.edu.cn Abstract: —This paper proposes a novel Two-layer Hierarchical differential evolution (THDE) algorithm to improve the search ability of differential evolution (DE) algorithm. Individuals are separated into bottom layer and top layer. In the bottom layer, individuals are divided into several groups. Modified DE/current-best/1/bin strategy is conducted to produce offspring, where the best individual comes from top layer. In the top layer, modified DE/rand/1/bin strategy is used to update individuals. A set of famous benchmark functions has been used to test and evaluate the performance of the proposed THDE. The experimental results show that the proposed algorithm is better than DE/current-best/1/bin and DE/rand/1/bin and better than or at least comparable to the self-adaptive DE (JDE) and intersect mutation differential evolution algorithm (IMDE) for most functions. Keywords- differential evolution, two-layer hierarchy, global optimization I. Introduction Differential evolution (DE) is a simple yet powerful evolutionary algorithm (EA) firstly introduced by Price and Storn [1] in 1997. DE attracts many researchers’ interesting in studying the improvement method and implication in many areas for its advantages of simplicity, fast convergence and less parameter. Many researchers suggest the proper parameter setting since DE is introduced [2][3][4]. This research work is supported by the National Basic Research Program of China (973 Program) under Grant no. 2011CB706804, and the Natural Science Foundation of China (NSFC) under Grant no. 51121002. Recently, some algorithms are proposed to guide the selection of operation individuals to improve the exploitation and exploration ability of DE. Kaelo and Ali [5] propose a modified DE algorithm called DE with random localization (DERL) by exhibiting local search ability around the base vector. Pant et al [6] employ the idea of parent centric crossover operators and proposes two versions of DE called DE with parent-centric crossover and DE (DEPCX) with probabilistic parent-centric crossover (Pro. DEPCX). Omran et al [7] use the concept of index neighborhoods and using weighted average of personal and neighborhood best position of mutation individual to produce next generation. Epitropakis et al [8] modify the random selection by assigning the selected probability inversely proportional to its distance from the individual. Zhou et al [9] separate the population into better part and worse part, and modified the mutation operation by intersect two groups. In addition, the combination of different mutation operation is another hot spot of the improvement research of DE recently. Qin et al [10] adaptive the mutation operations and control parameters by learning the experience from previous generation. Mallipeddi et al [11] proposed an ensemble of mutation strategies and parameters in DE (EPSDE). A candidate pool of mutation strategies and parameters are used to compete to produce offspring. In composite DE (CoDE) proposed by Wang et al [12], three trial vectors are generated by three different mutation strategy and randomly selected parameters from candidate pool. The best one among the three trail vectors and target vector are survived to next generation. Although different operations are conducted in the literature, these algorithms are not efficient for every 2013 IEEE International Conference on Systems, Man, and Cybernetics 978-1-4799-0652-9/13 $31.00 © 2013 IEEE DOI 2922 2013 IEEE International Conference on Systems, Man, and Cybernetics 978-1-4799-0652-9/13 $31.00 © 2013 IEEE DOI 2922 2013 IEEE International Conference on Systems, Man, and Cybernetics 978-1-4799-0652-9/13 $31.00 © 2013 IEEE DOI 10.1109/SMC.2013.497 2916