A Two-Stage Optimization Framework for Sparse Large-Scale Multiobjective Optimization

Sparse large-scale multiobjective optimization problems (LSMOPs) that widely exist in real life are characterized by a large number of zero-valued decision variables. Detecting non-zero variables in sparse LSMOPs has recently garnered significant attention and is considered a crucial research area. However, most of the current algorithms follow an optimizing-while-detecting way, which may cause the optimization of zero variables and non-zero variables to interact with each other, hindering the evolution of the population. To this end, we suggest a two-stage optimization framework that follows a detecting-then-optimizing way. In the first stage, we propose a template perturbation to detect non-zero variables and initialize the population hierarchically by the detection results. In the second stage, we propose a simple but effective decision variable re-testing technique to correct the results of the first stage and optimize the decision variables continuously. With this two-stage optimization strategy, the cooperation between the optimizations of zero and non-zero variables can be significantly strengthened. According to the experimental results on eight benchmark problems and three real-world applications, the proposed algorithm has shown a promising performance compared with the recent peer algorithms.