A knowledge-guided dimensionality reduction multiobjective evolutionary algorithm for large-scale sparse multiobjective optimization

Large-scale sparse multiobjective optimization problems widely arise in practical applications such as critical node detection, feature selection, and pattern mining. These problems usually involve a large number of decision variables, whereas their Pareto-optimal solutions are composed of only a small subset of active variables. As a result, an algorithm is required to achieve both search efficiency and solution quality in a high-dimensional and sparse search space. Existing large-scale sparse multiobjective evolutionary algorithms usually reduce the dimensionality of binary decision vectors to alleviate the curse of dimensionality. However, their dimensionality reduction schemes are often designed based on human experience or fixed rules, making them difficult to adapt to different problem characteristics and evolutionary stages. This may lead to the loss of important variable interactions or inefficient use of function evaluations. To address this issue, this paper proposes KGDR-MOEA, a knowledge-guided dimensionality reduction multiobjective evolutionary algorithm. In KGDR-MOEA, deep reinforcement learning is employed to model the dynamic relationship between the sparse evolutionary state of the population and candidate dimensionality reduction schemes. Based on evolutionary feedback, the algorithm adaptively selects a suitable dimensionality reduction strategy for the current search stage. By dynamically switching between full-space search and variable-grouping-based reduced search, KGDR-MOEA maintains global exploration while improving local exploitation efficiency, thereby approximating sparse Pareto-optimal solutions more effectively. Experimental results on benchmark test suites and real-world applications demonstrate that KGDR-MOEA achieves overall superior or at least competitive performance compared with several state-of-the-art large-scale sparse multiobjective evolutionary algorithms in terms of convergence efficiency, solution quality, and cross-problem adaptability.