Front guided surrogate-assisted evolutionary algorithm with adaptive selection of three-branch infill criteria for expensive many-objective optimization

Abstract

Expensive optimization problems are those where a single function evaluation takes a large amount of time, and are typically only allowed for a finite number of function evaluations. Thus, algorithms are extremely limited in the computational resources, and need to balance their ability of promoting convergence with of maintaining diversity. Especially when coping with many-objective optimization problems where the conflict between the two demands is further exacerbated, adaptive strategies that can promptly adjust computational resources according to the state of the population are of particular importance. To effectively solve expensive many-objective optimization problems with different characteristics, we propose a front guided surrogate-assisted evolutionary algorithm with adaptive selection of three-branch infill criteria (FGSAEA). FGSAEA consists of the front guided adaptive selection strategy and three infill criteria, all of which are featured by the positional relationship between the front of the candidate population and the archive storing truly evaluated solutions. FGSAEA uses the front guided adaptive selection strategy to determine the evolutionary state. To reduce the interference from inaccurate predictions of surrogate models, the strategy always takes the front of an archive storing all evaluated solutions as a reference. Then, FGSAEA utilizes one of three infill criteria respectively focused on promoting the convergence, the diversity of population, or the accuracy of surrogate models according to the demands. Experiments on benchmark problems demonstrate that FGSAEA is very competitive compared to state-of-the-art surrogate-assisted expensive many-objective optimization algorithms.