A Novel Adaptive Sampling Method Based on Expected Improvement Reduction

This paper illustrates a multi-fidelity surrogate-based optimization (MFSBO) method based on a novel computation-aware expected improvement reduction (EIR) adaptive sampling strategy. EIR-based infill strategy selects model fidelity and samples of infill by hypothetical updates to measure the reduction of expected improvement. Batch low fidelity infill within a dynamically varying trust-region is also realized in the proposed strategy to enhance exploration. In order to construct multi-fidelity surrogate model for optimization, the Cokriging method is applied for the fusion of the data of different fidelities and computational costs. Eventually, two case studies are tested to verify the feasibility of the proposed method. Results show that the proposed method accelerates the convergence of MFSBO and thus obtains a more accurate optimum.