Collaborative Research: Scalable & Communication Efficient Learning-Based Distributed Control

Intelligent infrastructures (e.g., transportation, energy) are poised to play an integral role in the ongoing societal transition towards a more sustainable future. These systems must operate reliably, robustly, and efficiently in uncertain and dynamic environments. Feedback control is the enabling technology for providing such guarantees. Centralized control, where one system is controlled by a single decision maker, is a mature technology with well-developed theory and efficient algorithms, and has enabled engineering successes across many applications, such as commercial aviation, process control, and robotics. In contrast, distributed control, wherein multiple subsystems are controlled by multiple decision makers, is much more challenging. While the last ten years have produced a wealth of new theory and computational tools for addressing the distributed control problem, it is nevertheless observed that the practical impact of distributed control in emerging areas such as smart infrastructure remains minimal. This project seeks to address this issue and move distributed control from theory to practice by building a foundational and integrated theory of distributed learning-enabled control and approximated distributed optimization. On the educational front, the research outcomes of this project will be integrated into graduate-level courses on learning-enabled control (Penn) and distributed-optimization (Columbia). Longer term, this project aims to create a new community of researchers working at the intersection of distributed learning, control, and optimization, and departmental efforts will be leveraged to recruit a diverse group of PhD students for the project.This project is motivated by the observation that there remain significant barriers to the practical use of safety-constrained real-time distributed control: (i) Existing methods are much too slow for real-time control; (ii) Distributed optimal control has mainly focused on linear models while many systems of interest are nonlinear; and (iii) It is often assumed that high-quality structured models reflecting system topology are available. Thrusts will cover the full control engineering pipeline to address these gaps. In Thrust I, federated and statistical learning are incorporated into structured system identification. Thrust II seeks to speed up distributed predictive control by developing distributed imitation and federated learning tools. Finally, Thrust III focusses on the design of distributed controllers robust to uncertainty from learning, numerical methods, and communication failures. In contrast to existing work, this proposal offers the first integrated approach to designing controllers that can be realistically deployed to societal-scale systems. Experimental validation of developed methods will be conducted on robotic platforms at Penn and on data from an HBT-EP plasma fusion tokamak at Columbia.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.