Decision space information driven algorithm for dynamic multiobjective optimization with a changing number of objectives

Dynamic multiobjective optimization problems (DMOPs) with a changing number of objectives receive little attention, but they exist widely in real life. These type of dynamics not only lead to expansion or contraction of Pareto optimal front/set (PF/PS) manifold, but also pose great challenges to balancing diversity and convergence. However, the current dynamic response mechanism has difficulty adapting these kind of problems. To tackle these problems, a decision space information driven algorithm (DSID) is proposed. Once the number of objectives changes, an individual guidance strategy based on manifold learning (IGSML) is introduced to identify solutions suitable for changes. Then IGSML produces excellent solutions by learning the manifold of these solutions. Meanwhile, a variable layering reconstruction strategy (VLRS) is proposed to divide the decision variables into three layers: convergence, diversity and multi-functional variables. Afterwards, VLRS takes into account the different degrees of influence of variables at different layers in the process of objective change, and makes targeted operations on different variables to quickly respond to changes. These two strategies cooperate with each other to balance the diversity and convergence. Comprehensive experiments are conducted on 15 benchmark functions with a varying number of objectives. Simulation results verify the efficacy of the proposed algorithm.