General framework for preserving passivity of deep neural network model
Abstract
Variety of neural networks such as MLP, LSTM, CNN have extensive applications in engineering domain solving classification and regression of real problems. Nevertheless, highly nonlinear systems such as hydraulic manipulators, friction-rich actuation, and polymer actuators behave unexpectedly depending on the circumstance conditions, which are hard to predict. The purpose of this project is to establish an idea on the hybrid use of physical-based and data-based dynamical modeling for robotics. A theoretical foundation for generalized passivity of a wide class of neural networks for time-series prediction of dynamical systems are considered. The fundamental framework for preserving passivity of network models will contribute to the reduce of training, the high robustness to noise and inputs outside the training domain.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 05, 2025
- Source ID
- FA23862414039
Entities
People
- Hyungpil Moon
Organizations
- Air Force Office of Scientific Research
- Sungkyunkwan University
- United States Air Force