Dimensional window method: A plug-in-style large-scale handling technique for evolutionary algorithm

Large-scale optimization constitutes a pivotal characteristic of numerous real-world problems, where large-scale evolutionary algorithms emerge as a potent instrument for addressing such intricacies. However, existing methods are typically tailored to address only a particular class of problems and lack the versatility to be readily adapted to other evolutionary algorithms or generalized across diverse problem domains. To address the issue above, this paper proposes the dimensional window method, a simple yet effective enhancement that can be seamlessly integrated into low-dimensional evolutionary algorithms to bolster their performance in large-scale optimization. Specifically, the dimensional window method involves grouping a subset of randomly selected dimensions into a window during each iteration, restricting the population’s evolution to the dimensions within this window. Furthermore, the effectiveness of the dimensional window method is analyzed, and the window is improved based on the insights gained, including the isometric segmentation individual-level window length and the neural network-guided window element. Extensive experiments on single-objective, multi-objective, constrained multi-objective, and discrete test problems with large-scale attributes demonstrate that the proposed method significantly mitigates the curse of dimensionality and enhances the performance of evolutionary algorithms in large-scale settings. A more significant advantage lies in the fact that the proposed plug-ins not only demonstrate remarkable performance when tackling real-world challenges, such as ratio error estimation problems, but also offer easily integration into existing evolutionary algorithm platforms, all while being highly user-friendly for evolutionary algorithm users.