An adaptive space preselection method for the multi-fidelity global optimization

Abstract Multi-fidelity (MF) metamodels have been well applied to aerospace structure optimization problems to relieve the computation burden. However, most of the existing multi-fidelity optimization methods explore the optimum in the whole design space, which may lead to low-efficiency of the optimization search. In this paper, a space preselection-based multi-fidelity lower confidence bounding (SPMF-LCB) optimization method is proposed to solve this problem. Firstly, a bootstrap-assisted area selection algorithm is proposed, which can adaptively partition the whole design space and select the most potential area to facilitate the optimization process. Secondly, the lower confidence bounding (LCB) method is extended to the multi-fidelity level, which can adaptively determine both the fidelity level and the location of sample points, with the consideration of the low-fidelity (LF) simulations budget. Finally, the probability of feasible (POF) method is combined with the extended LCB method to handle the constrained optimization problems. Eight analytical examples and the optimization problem of the radome of the missile are utilized to illustrate the efficiency of the proposed SPMF-LCB method. The performance of the proposed approach is compared with four existing methods. Results show that the proposed SPMF-LCB method performs the best considering the efficiency and robustness.