A comparison of three large-scale global optimizers on the CEC 2017 single objective real parameter numerical optimization benchmark

The scalability of optimization algorithms is an important issue that has been thoroughly studied in the past. However, these studies were normally conducted by gradually increasing the dimensionality of the benchmark and analyzing how an algorithm exhibiting a good performance on low-dimensional problems degrades as the problem size increases. In this contribution we follow the opposite approach: we take some well-known large-scale global optimizers based on the MOS framework and specifically designed for problems of thousands of variables and evaluate them on much smaller problems (up to 100 dimensions). The results show that, surprisingly, these algorithms are able to find good solutions to many of the functions of the benchmark, systematically reaching the global optimum for some of them. Furthermore, the differences in performance among the three considered algorithms are also analyzed and compared with statistical methods. Finally, several hypothesis are given to explain these differences in performance among the three algorithms.