Research
My group works at the intersection of control theory, optimization, and machine learning. We study how to design controllers and algorithms for large-scale interconnected systems under structural and informational constraints, with an emphasis on distributed decision-making, closed-loop guarantees, and scalable computational methods. Our recent work is organized around three closely related directions.
1) Distributed and Networked Control
Many control problems are inherently distributed: information, actuation, and computation are spread across a network, and controller design must respect this architecture. Our work studies how such structural constraints can be incorporated directly into synthesis and learning. This includes convex approaches to distributed control under sparsity and information constraints, learning-based methods for structured linear feedback, and recent extensions to nonlinear and reinforcement-learning-based control of networked systems.
Representative publications
- Sparsity Invariance for Convex Design of Distributed Controllers , IEEE Transactions on Control of Network Systems, 2020.
- Learning the globally optimal distributed LQ regulator , L4DC 2020.
- Distributed neural network control with dependability guarantees: a compositional port-Hamiltonian approach , L4DC 2022.
- Closing the gap to quadratic invariance: a regret minimization approach to optimal distributed control , ECC 2024.
- Scaling Up Stability: Reinforcement Learning for Distributed Control of Networked Systems in the Space of Stabilizing Policies , IFAC 2026.
2) Learning-Based Control with Closed-Loop Guarantees
A central theme in our research is how learning can be modularly embedded into feedback policy design enhancing performance while preserving closed-loop system properties. This includes neural-network controllers for nonlinear systems, reinforcement-learning-oriented parametrizations, and system-level descriptions of nonlinear stabilizing controllers. We are particularly interested in developing learning-based control methods in which various closed-loop guarantees are built into the controller class itself, rather than imposed only after training.
Representative publications
- Neural System Level Synthesis: Learning over All Stabilizing Policies for Nonlinear Systems , CDC 2022.
- MAD: A Magnitude And Direction Policy Parametrization for Stability Constrained Reinforcement Learning , CDC 2025.
- Characterizing All Locally Exponentially Stabilizing Controllers as a Linear Feedback Plus Learnable Nonlinear Youla Dynamics , IFAC 2026.
3) Algorithms for Control and Optimization
Control and optimization are tightly connected: many control methods rely on fast iterative solvers, and many optimization algorithms can be analyzed as dynamical systems. Our work uses system-theoretic tools to study and design optimization methods with rigorous convergence properties, with a particular emphasis on learning to optimize. The objective is to develop algorithmic architectures that can be adapted to classes of problems while retaining formal guarantees such as linear convergence. This direction is motivated both by large-scale computational bottlenecks in control and by broader questions on how feedback ideas can inform the analysis and design of optimization algorithms.
Representative publications
- Learning to optimize with convergence guarantees using nonlinear system theory , IEEE Control Systems Letters, 2024.
- Learning to optimize with guarantees: a complete characterization of linearly convergent algorithms , 2025 (preprint, with A. Martin and I. Manchester).
Talks and materials
A video presentation summarizing recent research across these directions: