Component-level study of a decomposition-based multi-objective optimizer on a limited evaluation budget

Decomposition-based algorithms have emerged as one of the most popular classes of solvers for multi-objective optimization. Despite their popularity, a lack of guidance exists for how to configure such algorithms for real-world problems, based on the features or contexts of those problems. One context that is important for many real-world problems is that function evaluations are expensive, and so algorithms need to be able to provide adequate convergence on a limited budget (e.g. 500 evaluations). This study contributes to emerging guidance on algorithm configuration by investigating how the convergence of the popular decomposition-based optimizer MOEA/D, over a limited budget, is affected by choice of component level configuration. Two main aspects are considered: (1) impact of sharing information; (2) impact of normalisation scheme. The empirical test framework includes detailed trajectory analysis, as well as more conventional performance indicator analysis, to help identify and explain the behaviour of the optimizer. Use of neighbours in generating new solutions is found to be highly disruptive for searching on a small budget, leading to better convergence in some areas but far worse convergence in others. The findings also emphasise the challenge and importance of using an appropriate normalisation scheme.