A Novel Subspace Construction Method for Large-Scale Evolutionary Multi-Objective Optimization

Problem transformation based optimization for large-scale multi-objective problems usually have excellent con-vergence yet insufficient diversity. Improving the distributivity of the search subspace in the decision space may prompt diversity to some extent. Therefore, we intend to propose a novel search subspace construction method to enhance the diversity. Firstly, a set of reference solutions is obtained by the algorithm. New solutions can be generated by a linear combination of two reference solutions, and thus locating a search subspace by these two reference solutions. Furthermore, a set of reference solutions locates a series of subspaces well-distributed in the original decision space. Once subspaces are determined, the algorithm firstly searches the subspaces to find solutions close to the Pareto set (PS) efficiently. Then, it searches the original decision space based on these solutions to make the population cover the PS uni-formly. Compared with three state-of-the-art large-scale multi-objective algorithms on well-known benchmark suite with up to 2000 decision variables, the experimental results demonstrate the competitive performance of the proposed algorithm.