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

Wu, Yuda; Lin, Qian; Zhou, Qi; Hu, Jiexiang; Wang, Shengyi; Peng, Yutong

doi:10.1016/j.ast.2021.106728

Cited by 7 publications

(1 citation statement)

References 36 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, the EI related to the LF model only reflects the sample effect on the model at this fidelity rather than on the MF model, and it cannot be used to indicate the sample fidelity. Therefore, some studies [22][23][24][25] evaluate the acquisition functions such as EI, LCB, and MES with the different parts of the prediction variance of MF model, which are brought by lack of corresponding fidelity samples, for comparison. In addition, He 26 simply considered the prediction variance associated with different fidelity models as the potential effect of the corresponding fidelity samples, and took the HF sample only when the potential effect of LF sample was low enough.…”

Section: Introductionmentioning

confidence: 99%

Decreased Max-value Entropy Search for Multi-fidelity Bayesian Optimization

Gong

Wang

Duan

et al. 2023

Preprint

View full text Add to dashboard Cite

Bayesian Optimization is a widely applied efficient framework for updating the surrogate model sequentially. To improve the efficiency, multi-fidelity Bayesian Optimization is developed to combine the information of samples in different fidelity levels. However, the multi-fidelity levels brings the challenge for sequential sampling. In multi-fidelity Bayesian Optimization, the sampling strategy is applied to determine not only the sample location but also the sample fidelity level for updating the model. To balance the benefit and the experiment cost, it is vital to measure the potential effect of each fidelity sample. Some sampling strategies, which is categorized as direct-type methods, can measure the potential effect of different samples in a easy way, but they cannot figure out the difference of samples that has little uncertainty and it has the risk for redundant sampling. Some other strategies, which is categorized as direct-type methods, can measure the potential effect appropriately, but the calculation of them are much complicated and time-consuming. In this paper, a new type method combining the direct-type and indirect-type methods is presented for measuring the potential effect of different fidelity sample. It is convenient enough and can avoid the problems occurring in the direct-type method. Based on this paradigm, a sampling strategy named decreased max-value entropy search(DMES) is proposed and applied in the multi-fidelity Bayesian Optimization framework. The characteristics of DMES and how it is different from direct-type method are detailed in some examples. Besides, two numerical experiments and one simulation experiment demonstrate the efficiency of DMES.

show abstract

Section: Introductionmentioning

confidence: 99%