Montenegro, C. (ECE) – Control of Uncertain Hybrid Systems

Machine learning endows autonomous systems to uncover underlying structures and physical laws from measured data and to leverage these models for prediction and decision-making. As the costs of data acquisition, processing, and storage decline—and sensors become increasingly widespread alongside ever-improving algorithms—artificial intelligence has attracted significant attention in research and industry.

Machine-learning methods are particularly attractive when an analytical model is too difficult—or even impossible—to derive because the underlying principles are poorly understood. As control engineering enters such domains—for example, physical human-robot interaction and self-driving vehicles—data-driven models offer a practical alternative to classical system-identification techniques for model-based control. In addition, we know that robotic or control systems seldom work in ideal conditions. Sensor noise, incomplete state information, and uncertain parameters are everyday realities, and controllers must be robust—able to attenuate these disturbances—and be backed by formal guarantees of stability and safety.

Coupling physical dynamics with embedded computation and communication introduces new challenges. Hardware elements such as analog-to-digital converters, sample-and-hold circuits, and quantizers, together with events like timers, resets, and impacts, yield an even more complex class of control systems in which designing controllers that remain robust to unmodeled dynamics and disturbances—and providing formal certificates of stability and safety—becomes harder. Cyber-physical systems that have continuous dynamics with event-driven behavior, therefore, require control strategies that explicitly account for these events and stay robust to adversarial uncertainties.

Therefore, the focus of this proposal is to design learning-based certificates and control techniques for hybrid systems with uncertainties in the form of unmodeled dynamics and unknown disturbances. We propose four research thrusts in this proposal. The first one addresses the problem of learning a surrogate model of the unmodeled using learning-based models that are both statistically sound and directly usable for feedback design. In the second thrust, we develop a safety control framework for systems whose dynamics are learned with high probability using a set-valued and variational analysis. In our third thrust, we consider the problem of learning certificates—in particular, Lyapunov functions and cost upper-bound surrogates—for hybrid systems. Finally, we tackle the optimal control problem for hybrid systems under unknown disturbances in our fourth thrust.

Event Host: Carlos Montenegro, PhD Student, Electrical & Computer Engineering

Advisor: Ricardo Sanfelice