An Evolutionary Algorithm for Many-Objective Optimization Based on Indicator and Vector-Angle Decomposition

The evolutionary algorithms for many-objective optimization based on reference-point decomposition are widely concerned since they generally maintain good performance on many optimization problems, however, most of these algorithms show insufficient versatility on optimization problems with various types of Pareto fronts. To address this issue, we propose an evolutionary algorithm for many-objective optimization based on indicator and vector-angle decomposition, termed IVAD. In the proposed algorithm, the objective vectors of current population, as a set of reference vectors, are used to dynamically partition the whole objective space. And the max-min-vector-angle selection strategy, by calculating the vector angles between each pair of solutions, is constructed to select well-diversity solutions. Furthermore, to enhance the balance between convergence and diversity, the elite replacement, based on
$I_{\varepsilon +}$
indicator and vector angle, is proposed for each cluster that the selected individuals belong to. The proposed algorithm is compared with state-of-the-art many-objective evolutionary algorithms based on reference-point and vector-angle decomposition on three test suites with up to 15 objectives. Experimental results demonstrate that the proposed IVAD obtains more competitive performance on many-objective optimization problems with various types of Pareto fronts, and enhances the ability to balance convergence and diversity.