EAGER: Real-Time: Learning-based Optimal Control of Stochastic Nonlinear Systems

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

The two main challenges in optimal real-time control of complex engineering systems arise from the computational complexity of high-fidelity fundamental models of such systems and the inherent uncertainty stemming from lack of exact understanding of the underlying physical, chemical, and biological phenomena governing the system behavior. High-fidelity models are often prohibitive for real-time decision making because they are too computationally demanding. On the other hand, model uncertainty can compromise the reliability of decision making thus can be detrimental to safe, reliable, and optimal operation of complex systems. In this exploratory research project, a new paradigm for learning-based optimal control that guarantees stability and robustness of uncertain nonlinear systems will be developed by using approximate models of the system and its uncertainty, while optimizing the system performance with respect to an updated model of the system learned online. The prototypical example will focus on cold atmospheric plasma jets that can have significant impact in materials processing and in the emerging field of plasma medicine. The proposed research is motivated by the growing importance of using high-fidelity models for optimal control of engineering systems as well as the theoretical challenges associated with addressing systematic handling of uncertainties, non-conservative control performance, and low computational complexity in a unified optimal control formulation. The ultimate objective is to develop a learning-based optimal control method that leverages high-fidelity knowledge of an uncertain system, ensures safe and robust system operation in the presence of uncertainties, mitigates conservative control performance, and is amenable to real-time computations. High-fidelity models will be used to systematically inform the design and verify the performance of learning-based optimal control via closed-loop simulations under uncertainty. The specific aims of the project are: (1) develop theory and formulations for learning-based optimal control, (2) develop an uncertainty propagation method that is especially suited for performance verification of learning-based optimal control using nonlinear high-fidelity models with arbitrary probabilistic uncertainties, and (3) demonstrate the potential benefits of learning-based optimal control on a complex engineering system, i.e., a cold atmospheric plasma jet testbed. The proposed methodology may find numerous applications beyond the ones involving low-temperature atmospheric plasma jets. In addition to training a graduate student in an emerging multi-disciplinary field, a new upper level undergraduate/graduate course will be developed that integrates basic concepts from data science, Bayesian inference, and robust optimization of complex chemical and biomolecular systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.