Adaptive knowledge transfer based on machine learning method for evolutionary multitasking optimization

In recent years, evolutionary multitasking has exhibited significant potential in solving multiple optimization tasks synergistically by the evolution of a single population. The paradigm enables different tasks to share underlying similarities by transferring information to each other, which has been shown to accelerate the convergence of similar tasks. In the absence of prior knowledge of the relationships between optimization tasks, it is not trivial to control the degree of intertask knowledge transfer, thus negative knowledge transfer between tasks frequently occurs to impede convergence behavior. In this paper, we propose a multifactorial evolutionary algorithm (MFEA) based on the machine learning method, termed MFEA-ML, to learn to adaptively transfer online at the individual level to alleviate negative transfer and boost positive transfer. Different from most of the existing algorithms that measure intertask similarities for adaptive knowledge transfer, the proposed method collects training data by tracing the survival status of the individuals generated by intertask transfer and accordingly constructs a machine learning model to guide the transfer of genetic materials from the perspective of individual pairs. The efficacy of MFEA-ML is demonstrated on a series of benchmark problems as well as a practical engineering design scenario involving simultaneous consideration of two mission requirements. In the future, modifying the proposed method to handle expensive multitask optimization problems is a promising direction.