Keywords: Control of Partial Differential Equations, Control Design for Hybrid Systems, Numerical Methods for Optimization
Abstract: This paper presents a switched observer design methodology for distributed parameter systems governed by partial differential equations. The approach combines Proper Orthogonal Decomposition (POD) model reduction with adaptive sensor configuration switching to address computational challenges in infinite-dimensional state estimation. A finite-dimensional reduced-order model is constructed through POD basis extraction, enabling design of a robust switched Luenberger observer that dynamically selects among multiple measurement configurations. Sufficient conditions for exponential stability and prescribed H_infty performance are formulated as Linear Matrix Inequalities (LMIs), ensuring computational tractability. The methodology accommodates practical considerations including sensor failures and operational flexibility. Validation through a counter-current tubular heat exchanger demonstrates accurate temperature distribution estimation under distributed uncertainties, confirming effectiveness for industrial distributed parameter system monitoring.

Keywords: Optimization for Learning and Control, Real-Time Control Problems, Large Scale Optimization Problems
Abstract: This paper discusses a dynamics-aware distributed optimization over a network of linear agents subject to input saturations, which commonly affect actuators in various physical plants. We formulate a coupled optimization problem and design a novel distributed controller that guarantees every agent asymptotically converges to the unique optimizer of the problem in a cooperative manner. We analyze local synchronizability by using local sector conditions, and derive a sufficient condition for solvability of the dynamics-aware distributed optimization problem under the input saturation in terms of the linear matrix inequalities. Furthermore, we derive a possible alternative that does not require any prior information about the optimizer, albeit with increased conservativeness. Through a numerical simulation, we validate the effectiveness of the proposed criteria.

Keywords: Applications in Energy and Power Systems, Optimization for Learning and Control, Robust Control and Stabilization
Abstract: We consider a class of power systems composed of multiple interconnected nodes with a time-varying network topology. For these systems, we aim to augment a pre-existing technique, which is responsible for designing switch-stabilizing control laws, such that the resulting controller possesses a pre-imposed sparsity pattern. In order to compensate for the bilinearity of the resulting optimization problem, we employ an iterative numerical procedure that relaxes the original problem into a sequence of convex ones. Moreover, in order to enhance closed-loop performance with respect to steady-state response, we also leverage a classical result from control theory for the distributed setting considered in our proposed design framework.

Keywords: Optimization for Learning and Control, Data-driven Models and Decision-Making
Abstract: This paper presents a data-driven framework for feedback linearization of nonlinear control-affine systems that promotes geometric validity through structured parameterization and manifold constraints. The state diffeomorphism varphi(x) is parameterized with a linear core constrained to mathrm{SL}(n) and augmented by polynomial and Fourier residual features. The input transformation comprises a drift cancellation term alpha(x) and a decoupling matrix beta(x), yielding the control law u = alpha(x) + beta(x)v. The map alpha(x) is parameterized using polynomial and trigonometric features, while beta(x) is defined as a matrix exponential of traceless, state-dependent matrices, guaranteeing detbeta(x) = 1 by construction. The state diffeomorphism varphi(x) is regularized to maintain well-conditioned Jacobians, with invertibility validated empirically over the operating region. Learning is carried out via tangent matching with geometric regularization, optimized using Riemannian gradient descent on manifolds. Validation on a four-dimensional nonlinear two-input system demonstrates that the learned transformations remain globally invertible and well conditioned, with alpha(x) accurately canceling the nonlinear drift. Integrated with a pole-placement controller in Brunovsk'y coordinates, the method achieves accurate reference tracking. These results demonstrate that manifold-constrained learning enables practical and reliable data-driven feedback linearization.

Keywords: Data-driven Models and Decision-Making, Optimization for Learning and Control, Multi-Objective Control and Optimization
Abstract: This paper introduces MLAgents, a multi-agent system for end-to-end ML model synthesis, emphasizing data privacy and adaptive optimization. In its closed-loop control architecture, LLM agents generate pipelines for secure local execution, protecting sensitive data. The core contribution is a non-LLM Improvement Agent that treats pipeline modification as a multi-armed bandit problem. This agent uses Thompson sampling, a Bayesian adaptive policy, to balance exploration-exploitation based on empirical feedback. We demonstrate that this principled approach achieves competitive performance while adhering to strict privacy-by-design principles.

Keywords: Optimization for Learning and Control, Real-Time Control Problems, Control Design for Hybrid Systems
Abstract: The report presents a non-linear formula for controlling an active magnetic suspension system. Its component functions are dependent on the position of the levitating object. Their analytical forms were proposed and the parameters optimized on the basis of linear model analysis. The controller was designed to obtain an overdamping property of the closed loop system for a considered operating range. A compromise regarding the position of the poles were presented. The results of simulation studies and real-time experiments were presented, illustrating the operation of the system, especially in extreme positions of the permissible range.