PI: Aaron Tuor
Efficient and robust control algorithms are needed for a wide range of mission-critical, real-world systems which are complex, uncertain, and must operate under specific constraints. Examples of such systems applications include automation of new manufacturing techniques for cost effective production of precision parts, refinement of processing techniques for desirable materials properties, and autonomous vehicle fleets. These dynamic environments all stand to benefit greatly from high-paced, real-time strategic decision-making. However, they are often characterized as having partial information and incomplete awareness of model state-space barring the effective application of standard control theoretic approaches. A broad range of new sensor data is becoming available offering the opportunity to extend automated control to a wider range of applications through deep learning-based modeling. Unfortunately, purely data-driven approaches are unsuitable due to their poor sampling efficiency and lack of operational safety guarantees.
To bridge this gap, researchers plan to design control methods for constrained decision-making under uncertainty which bridge advantages of data-driven and control theoretic approaches. Specifically, they are developing design-adaptable physics-informed neural network components for modeling dynamical systems, and deep learning reformulations of classical control theory algorithms which incorporate physics-informed components.
This application of deep learning extends control theoretic methods to a wider range of mission critical applications with incomplete knowledge of system dynamics. While the classical control basis of deep learning optimization offers constraint guarantees for operational safety and security, physics-informed components offer better sampling efficiency, and transparent domain aware.
Visit our GitHub repositories:
- NeuroMANCER: Neural Modules with Adaptive Nonlinear Constraints and Efficient Regularizations
- PSL: Python Systems Library