A hybrid evolutionary algorithm with dual populations for many-objective optimization

Many-objective optimization has posed great challenges to existing evolutionary algorithms that are designed for solving two- or three-objective problems. Most of the algorithms do not scale well with the number of objectives due to the expansion of the objective space. In this paper, a hybrid evolutionary algorithm with dual populations (HEA-DP) is proposed to tackle many-objective problems. The algorithm combines the advantages of decomposition-based and indicator-based approaches by maintaining two populations. The fitness values of individuals in the first population are determined by an aggregation function, while individuals in the second population are evaluated according to an efficient performance indicator. The information about the objective space is shared by employing a reproduction strategy that chooses parents from both populations. In this way, the algorithm can explore the objective space more thoroughly and can have more stable performance. Several state-of-the-art many-objective algorithms are adopted as peer algorithms to validate the proposed algorithm. We test the algorithms on two commonly used many-objective problem suites using different numbers of objectives. Numerical results indicate that HEA-DP is highly competitive in most of the problem instances.