An On-Line Multi-Fidelity Metamodel Assisted Multi-Objective Genetic Algorithm

Multi-objective genetic algorithms, which require a large number of fitness evaluations before obtaining a Pareto set, can become computationally intractable when applied to practical engineering design problems involving computationally expensive simulations. In this work, an on-line multi-fidelity metamodel (MFM) assisted multi-objective genetic algorithm (OLMFM-MOGA) approach is proposed, in which the MFM that can integrate information from both low-fidelity (LF) and high-fidelity (HF) models is constructed to replace the fitness evaluations during the optimization process. Two model management strategies, an individual-based updating strategy considering the interpolation uncertainty from MFM and a generation-based updating strategy considering the discrete degree of the populations, are incorporated in the OLMFM-MOGA. Three numerical examples and an engineering case with different degrees of complexity are used to demonstrate the effectiveness of the proposed approach. Results show that the proposed OLMFM-MOGA is able to obtain similar convergence and diversity of the Pareto frontier to the ones obtained by MOGA with only HF information, while at the same time significantly reducing the number of evaluations of the expensive HF model.