[3H1-OS-9a] OS-9

9:30 AM - 9:45 AM JST(12:30 AM - 12:45 AM UTC)

[3H1-OS-9a-03] Neural Networks for Dynamical System Modeling Based on Kernel Representation of Dirac Structures

〇Reiho Li¹, Razmik Arman Khosrovian², Takaharu Yaguchi^3,4, Hiroaki Yoshimura⁵, Takashi Matsubara^1,4 (1. Hokkaido University, 2. The University of Osaka, 3. Kobe University, 4. RIKEN AIP, 5. Waseda University)

Keywords:

System Identification,Deep Learning,Dirac Structure,Bond Graph

Deep learning has been widely used as a data-driven modeling method for dynamical systems. However, existing modeling approaches are limited by the fixed input-output causality of system components, which makes it difficult to apply them to complex tasks such as inverse design and interconnection of subsystems. To address this limitation, we propose a modeling method for dynamical systems based on the kernel representation of Dirac structures using deep learning. Since this method describes the system in a form that does not fix input-output causality, it can be flexibly applied to inverse design. Furthermore, by algebraically coupling pre-trained models without retraining, we can efficiently handle interconnection of subsystems. Experimental results demonstrate that the proposed method is effective for tasks that were outside the scope of existing methods.