Simplifying Hyperparameter Derivation for Integration Neural Networks Using Information Criterion*
Yoshiharu Iwata, Hidefumi Wakamatsu · 2024
When optimal design is performed using simulation, attempts have been made to construct highly accurate approximators using machine learning to address the conflicting issues of simulation accuracy and time. In response, an integration neural network (INN and INN2), which combines deductive and inductive knowledge, has been proposed to obtain highly accurate approximate solutions with a small amount of data. However, creating this evaluation data requires much time and effort. Therefore, this study focused on the information criterion, which statistically evaluates the balance between the input data's diversity and the model's accuracy. The possibility of optimizing the hyperparameters using only this information criterion (AIC, BIC) and training data, thus eliminating the need for evaluation data, was investigated. The results showed that INN and INN2 behaved differently. In INN, the appropriate structure using the information content criterion and evaluation data was consistent, and hyperparameter optimization without evaluation data was possible. On the other hand, for INN2, it was found necessary to use evaluation data.