Comparative evaluation of large-scale many objective algorithms on complex optimization problems

In the field of optimization, there has been an enormous surge in interest in addressing large-scale many-objective problems. Numerous academicians and practitioners have contributed to evolutionary computation by developing a variety of optimization algorithms tailored to tackle computationally challenging optimization problems. Recently, various largescale many-objective optimization algorithms (LSMaOAs) have been proposed to address complex large-scale many-objective optimization prob lems (LSMaOPs). These LSMaOAs have shown remarkable performance in addressing a variety of LSMaOPs. However, there is a pressing need to further investigate their performance in comparison to each other on different classes of LSMaOPs. In this study, we conduct a comparative investigation of three established LSMaOAs namely, LMEA, LMOCSO and S3CMAES over rigorous benchmarking on DTLZ, LSMOP, UF9-10, WFG test suites, encompassing problem sets with three to ten objectives and varying numbers of variables between 100 and 500. Additionally, we assess the algorithm’s efficacy on a test suite specifically designed for large-scale multi/many-objective problems (100-1000 decision variables). In addition, we propose Hybrid-LMEA, a light hybrid that integrates decision-variable clustering with competitive learning to improve both convergence and diversity. The hybrid works especially well on high-dimensional large-scale many-objective optimization problems with better performance in 8 and 12 out of 27 test cases for IGD and GD, respectively. The outcomes of the experiments indicate the relative efficacy and effectiveness of the different algorithms in addressing large-scale many-objective problems. Researchers can leverage this comparative data to make informed decisions about which algorithms to employ for particular optimization problem domains.