A decomposition and statistical learning based many-objective artificial bee colony optimizer

Abstract

Encouraged by the strong global search ability of the artificial bee colony (ABC) optimizer in single-objective problems (SOPs), we extend ABC to solve many-objective optimization problems (MaOPs) by exploiting the potential of direction vectors. Specifically, direction vectors are used not only to transform the original MaOP into a set of SOPs, but also to divide the bee colony into multiple subpopulations, while ABC serves the purpose of SOPs optimization. However, search equation of ABC is not efficient in convergence speed due to the overrated exploration property, to alleviate this issue, we suggest a search path statistical learning mechanism to predict the potential solutions, which is utilized to strengthen the exploitive search and accelerate the convergence rate. In addition, a new adaptive scalarization approach by exploiting population entropy and knee point information is developed to determine the elite in each subpopulation. For onlooker bee stage, a new fitness assignment scheme is proposed to achieve the computational effort allocation among food sources. The proposed algorithm is compared with several popular evolutionary optimizers on a wide range of test problems covering varying complexity, and empirical results show that it is very competitive and promising.