Fine Tuning a Data-Driven Estimator

Abstract: In recent years, many industries have developed high-fidelity simulators, such as digital twins, to represent physical systems, although their parameters must be calibrated to accurately reflect the true system. This need has paved the way for the creation of data-driven parameter estimators to calibrate such simulators. These estimators are constructed by generating synthetic observations for various parameter settings of the simulator and then establishing a mapping from these observations to the corresponding parameter settings using supervised learning. However, if the true system's parameters fall outside the range of the sampled parameter set used to construct the mapping, the resulting predictions will suffer from out-of-distribution (OOD) issues. In this paper, we introduce a fine-tuning approach for a specific data-driven estimator, known as the Two-Stage estimator, designed to mitigate the problems associated with OOD and improve its accuracy.