On the Ensemble of Surrogate Models by Minimum Screening Index

Abstract

The ensemble of surrogate models has attracted more and more attention due to their more stable performance than individual models. This work proposes a novel adaptive ensemble of surrogate models based on the minimum screening index. Constructing the screening index to realize the definition and elimination of a global poor model to further update the model library. Compared with the cross-validation error, this index presents better reliability in the model library update. The baseline model is determined by the minimum screening index to further propose a new weight calculation strategy based on this baseline model; 35 test functions are used to evaluate the performance of the proposed model. The results show that this model presents better accuracy and robustness than the individual surrogates and the other ensemble of surrogate models. More importantly, in engineering applications, the same results are also obtained, indicating that the proposed model has a higher priority than the other ensemble of surrogate models. This effective model gives a new way for the design of engineering problems.