An enhanced LSHADE with generalized Pareto distribution selection for escaping local optima

Zhe Xu, Jiatianyi Yu, Baohang Zhang, Lin Yang, Yanting Liu, Shangce Gao

Abstract: Local search hybridized with adaptive differential evolution (LSHADE), as one of the effective variants of differential evolution, has successfully spawned numerous variations that achieved victory in numerous IEEE congress on evolutionary computation (CEC) competitions. The successful strategies have yielded outstanding performance, with these LSHADE variants consistently showcasing rapid convergence. However, they have grappled with the challenge of escaping the dilemma of converging into local optima. To address this issue, this paper proposes an enhanced version of LSHADE, incorporating a generalized Pareto distribution selection mechanism (LSHADE-GPS). The objective is to leverage the power-law long tail characteristics of the Pareto distribution to generate emerging individuals capable of breaking free from the current impasse, thereby significantly enhancing the algorithm’s exploratory performance. Paired with a novel adaptive selection mechanism, effective control is applied at opportune moments, substantially elevating the likelihood of the algorithm escaping local optima. Extensive experimentation and data analysis on CEC2017 benchmark functions demonstrate that LSHADE-GPS exhibits superior performance compared to other state-of-the-art competitors.