Dynamic auxiliary reference vector-based many-objective evolutionary algorithm with adaptive multi-population collaboration

Abstract

Reference-vector-based many-objective evolutionary algorithms exhibit limited performance when addressing problems with irregular Pareto fronts. To overcome this limitation, this paper proposes a dynamic reference vector adjustment strategy that incorporates auxiliary vectors to enhance adaptability to complex Pareto front geometries. This strategy combines uniformly distributed reference vectors with adaptively adjusted auxiliary vectors, which effectively improves the approximation capability for irregular fronts. To further balance convergence and diversity, an adaptive multi-population evolutionary framework is designed, in which subpopulations with different search tendencies are defined and computational resources are dynamically allocated to achieve effective coordination between exploration and exploitation. In addition, a dimension-aware environmental selection mechanism is introduced, which adaptively switches selection strategies according to the number of objectives, thereby enabling more refined control over the trade-off between convergence and diversity. Based on these components, a unified algorithmic framework, termed A-RVEA-LS, is constructed. Comparative experiments on benchmark problems including DTLZ, MaF, and WFG against nine advanced algorithms demonstrate that A-RVEA-LS exhibits significantly superior overall performance and robustness in the majority of test cases.