A Many-objective Evolutionary Algorithm Based on Hybrid Dynamic Decomposition
Chunlei Li, Libao Deng, Wenyin Gong, Liyan Qiao · 2023
The decomposition-based many-objective evolutionary algorithms (MaOEAs) generally decompose a many-objective optimization problem into several scalar subproblems for collaborative optimization using the aggregation function. However, the performance of a single decomposition method may degrade in high-dimensional objective space due to the immutable contour lines. To address this problem, this paper investigates the hybrid version of the decomposition methods, i.e., the penalty-based boundary intersection decomposition and the family of$L_{p}$scalarization. These two kinds of methods are applied in different evolutionary stages to achieve an appropriate balance between convergence and diversity. The$p$value in the$L_{p}$scalarization method is dynamically updated online to adapt various evolutionary status. Moreover, a shift-based density estimation-aided mating selection strategy and a cosine distance-based offspring arrange mechanism are designed to cooperate with the decomposition method for generating more promising solutions. To evaluate the proposed algorithm, it's applied to solve a series of benchmark instances. The experimental results show its high competitiveness when comparing with five state-of-the-art algorithms for many-objective optimization problems.