A Variable Granularity Grouping Evolutionary Algorithm for Large-Scale Many-Objective Optimization
提出变粒度分组进化算法VGGEA，通过粗到细的粒度划分减少函数评估次数，并采用两阶段优化策略分别提升收敛性和多样性，解决大规模多目标优化问题。
Large-scale many-objective optimization problems (LSMaOPs) involve the large number of decision variables and objectives simultaneously. Most existing algorithms based on decision variable analysis perturb decision variables one by one, resulting in an unbearable number of function evaluations when solving LSMaOPs. Therefore, this paper proposes a variable granularity grouping evolutionary algorithm named VGGEA. Specifically, we firstly propose a variable granularity grouping strategy to reduce function evaluations greatly. According to the principle of decreasing the number of decision variables represented by a granularity, that is, from coarse granularity to fine granularity, the redundant decision variables of the objective function are gradually screened. Here, the decision variables are divided into two types of subsets: the preferred decision variable subset for each objective and the shared decision variable subset for multiple objectives. Then, we propose a two-stage optimization strategy. In the first stage, we optimize the preferred decision variable subset for each objective through a layered optimization strategy to improve the convergence of the population. In the second stage, we optimize the shared decision variable subset for multiple objectives to enhance the diversity of the population. Moreover, to better adapt to the increase in the number of objectives, we introduce the multiple populations method to balance the performance of the solutions on all objectives. The experimental results demonstrate the high performance and competitiveness of the proposed VGGEA.