Heuristic orientation adjustment for better exploration in multi-objective optimization

Abstract

Decomposition strategy which employs predefined subproblem framework and reference vectors has significant contribution in multi-objective optimization, and it can enhance local convergence as well as global diversity. However, the fixed exploring directions sacrifice flexibility and adaptability; therefore, extra reference adaptations should be considered under different shapes of the Pareto front. In this paper, a population-based heuristic orientation generating approach is presented to build a dynamic decomposition. The novel approach replaces the exhaustive reference distribution with reduced and partial orientations clustered within potential areas and provides flexible and scalable instructions for better exploration. Numerical experiment results demonstrate that the proposed method is compatible with both regular Pareto fronts and irregular cases and maintains outperformance or competitive performance compared to some state-of-the-art multi-objective approaches and adaptive-based algorithms. Moreover, the novel strategy presents more independence on subproblem aggregations and provides an autonomous evolving branch in decomposition-based researches.