[2M5-OS-37b] OS-37

[2M5-OS-37b-03] Learning Dynamics for Control in Situations with Significant Modeling Errors

〇Akira Osaka¹, Naoya Takeishi¹, Takehisa Yairi¹ (1. The University of Tokyo)

Keywords:Grey-box Modeling, Model-based Control

Accurate modeling is essential in model-based control. There is a discrepancy between real-world behavior and physical models, such as equations of motion, and these modeling errors are significant problems. Grey-box modeling is one of the ways to compensate for these errors by combining data-driven models with physical models, and it can improve modeling capabilities. Several previous works on grey-box models point out that regularizations on data-driven models are necessary when unknown parameters are also included in physical models because excessive flexibility of neural networks hinders the accurate estimation of physical parameters. However, the appropriate formalization of regularizers is not thoroughly known. We aim to elucidate how to introduce regularizers to grey-box models with unidentified physical parameters. We conducted experiments and verified that the regularized learning approach enhanced parameter estimation accuracy, and correlation-based regularizers showed robustness against changes in hyperparameter values.