Promoting Objective Knowledge Transfer: A Cascaded Fuzzy System for Solving Dynamic Multiobjective Optimization Problems

Abstract

In this article, a novel dynamic multiobjective optimization algorithm (DMOA) with a cascaded fuzzy system (CFS) is developed, which aims to promote objective knowledge transfer from an innovative perspective of comprehensive information characterization. This development seeks to overcome the bottleneck of negative transfer in evolutionary transfer optimization (ETO)-based algorithms. Specifically, previous Pareto solutions, center- and knee-points of multisubpopulation are adaptively selected to establish the source domain, which are then assigned soft labels through the designed CFS, based on a thorough evaluation of both convergence and diversity. A target domain is constructed by centroid feed-forward of multisubpopulation, enabling further estimations on learning samples with the assistance of the kernel mean matching (KMM) method. By doing so, the property of nonindependently identically distributed data is considered to enhance efficient knowledge transfer. Extensive evaluation results demonstrate the reliability and superiority of the proposed CFS-DMOA in solving dynamic multiobjective optimization problems, showing significant competitiveness in terms of mitigating negative transfer as compared to other state-of-the-art ETO-based DMOAs. Moreover, the effectiveness of the soft labels provided by CFS in breaking the “either/or” limitation of hard labels is validated, facilitating a more flexible and comprehensive characterization of historical information, thereby promoting objective and effective knowledge transfer.