Neural network hyperparameter tuning for online model parameter updating using inverse mapping models

To decrease the mismatch between a model and a physical system, physically interpretable model parameter values of nonlinear systems can be updated in real-time by using the Inverse Mapping Parameter Updating (IMPU) method. In this method, an Inverse Mapping Model (IMM), constituted by an Artificial Neural Network (ANN), is trained, offline, using simulated data that consists of features of output responses (ANN inputs) and corresponding parameter values (ANN outputs). In an online phase, the trained ANN can then be used to infer parameter values with high computational efficiency. The (non-trivial) choice of ANN-hyperparameters, e.g., ANN structure and training settings, may significantly influence the accuracy of the trained ANN. Therefore, this work discusses multiple ANN-hyperparameter tuning techniques to increase the accuracy of the IMPU method, of which the Bayesian search technique is the most promising considering accuracy and efficiency as it learns from previously evaluated hyperparameter values.