MID-ESM: An Adaptive Median-Based Ensemble of Surrogate Models

Abstract

Along with the development of surrogate models, there is a growing need to use surrogate models instead of computationally intensive simulations to estimate real system responses. Compared with individual surrogate models, the ensemble of surrogate models is gradually drawing more attention due to its better applicability and robustness. Thus, this paper proposes an adaptive median-based ensemble of surrogate models (MID-ESMs). At first, construct a reference model using the median of the predicted values of several surrogate models. Then an adaptive weight ensemble strategy is proposed based on the reference model to integrate global trends and local features. Thirty test functions and a practical engineering case are used to evaluate the model performance. In addition, this paper investigates the effect of homoscedasticity noise and test functions of different dimensions on the proposed model. The results demonstrate that MID-ESM has higher accuracy and robustness than individual surrogate models and other ensembles of surrogate models, offering better applicability in engineering problems.