Comparison study of large-scale optimisation techniques on the LSMOP benchmark functions

Abstract

In this paper, we study the performance of three popular large-scale optimisation algorithms on the recently proposed large-scale many-objective optimisation problems (LSMOP). We briefly explain the three methods (MOEA/DVA, LMEA and WOF) and give an overview of their use and performance in the literature. For the Weighted Optimization Framework (WOF), we propose a new transformation function to eliminate the parameter needed in its previous version. In our experiments, we compare the three algorithms on the LSMOP1-9 functions with 2 and 3 objectives and up to 1006 decision variables. The special focus of our study is on the convergence speed and behaviour, since MOEA/DVA and LMEA, in contrast to WOF, need huge computational budgets to obtain variable groups prior to optimisation. Our experiments show that MOEA/DVA and WOF perform significantly better than LMEA on almost all instances and WOF further outperforms MOEA/DVA significantly in most of the 1006-variable problems, in solution quality as well as convergence speed. In most instances the WOF only needs 0.1% to 10% of the total evaluations to outperform the final solution sets obtained by LMEA and MOEA/DVA.