Key points are not available for this paper at this time.
This work introduces a framework for systematically determining the sample size required to build accurate metamodels in Automated Machine Learning (AutoML) pipelines. The method employs a feedback control strategy that favors simple regression models when sufficient, using nonlinear models only when necessary. At each iteration, an adaptive sequential design of experiment techniques decides the number and placement of new samples, minimizing costly queries to the underlying process. All generated data sets are fully reused across iterations, ensuring efficiency and convergence until predefined error-based stopping criteria are satisfied. The framework contributes three innovations: (i) a proportional feedback controller to adaptively reduce error, (ii) a steady-state detection mechanism to stop training when no further improvement occurs, and (iii) an inclusive multiresponse strategy that simultaneously leverages all available data. Inspired by Process Systems Engineering, the approach is validated on benchmark cases, demonstrating an efficient, accurate metamodel construction with minimal computational cost.
FERNANDES et al. (Sat,) studied this question.