Abstract:

Literature: 

Abstract:

Literature: 

Abstract:

This work addresses the absolute stability analysis of feedback systems with slope-restricted and repeated nonlinearities.  Under the integral quadratic constraint framework, we can employ static O’Shea-Zames-Falb multipliers to obtain an LMI-based criterion that ensures absolute stability.  However, this criterion is only a sufficient condition for the absolute stability.  In other words, if this “primal” LMI is infeasible, no definitive conclusion can be drawn.  To overcome this limitation, we consider the corresponding dual LMI that becomes feasible if and only if the primal LMI is infeasible.  Our main result establishes that if the dual solution satisfies a certain rank condition, then it is possible to construct both a destabilizing slope-restricted nonlinearity and a non-zero equilibrium state, thereby proving that the system of interest is not absolutely stable.

Abstract:

Data-driven control has attracted growing attention as modeling complex systems becomes increasingly difficult. A central question is whether the available data are sufficient to guarantee a desired control objective such as stabilization. The concept of data-informativity provides a theoretical framework to address this question.

Most existing studies assume time-series trajectory data. However, in many applications, such as single-cell RNA sequencing, population-level biological measurements, and certain economic or social systems, individual trajectories cannot be tracked and only snapshot population distributions at discrete time points are observed. We extend data-informativity to such snapshot data for stabilization. A key difficulty is that associating snapshots across adjacent time points leads to factorially many permutations, resulting in severe computational complexity. In this talk, after reviewing several recent results on the density steering of ensemble dynamics and its system identification counterpart via Schrödinger bridges, we derive a naive condition requiring exhaustive permutation checks and a computationally tractable sufficient condition whose complexity is independent of the number of permutations, and numerically illustrate the trade-off between conservativeness and computational efficiency.

Abstract:

This talk addresses the problem of identifying informative nodes in network systems where the network topology is known, but the individual node dynamics are unknown except for their stability. In this context, informative nodes are defined as those whose convergence ensures the convergence of the entire network, serving as key locations for observation and control. We identify specific topological characteristics of the network for characterizing these nodes.

Literature:

N. Akiyama, S. Azuma, and I. Banno: Structure-based Sensor Placement for Network Systems, CDC2025.

Abstract:

The majority of recent work on data-driven control assumes that the parameters of the underlying true system are completely unknown. However, in practice, we often have some information on these parameters, even though their exact values are unknown. Taking into account such prior knowledge can weaken the requirements on the data, and may enable data-driven control in situations where control design is impossible using data alone. The purpose of this talk is to develop conditions for data-driven stabilization using prior knowledge on system-theoretic properties, specifically stabilizability and controllability. For this, we will extend the data informativity framework to include prior knowledge. 

We begin the talk by reviewing existing methods for direct data-driven stabilization. Thereafter, we discuss the inclusion of prior knowledge on stabilizability and controllability. In the case of controllability, we prove that the conditions on the data required for stabilization are equivalent to those without the inclusion of any prior knowledge. As such, controllability as prior knowledge does not help in weakening the conditions on the data. However, in the case of stabilizability as prior knowledge, we show that the conditions on the data are, in general, weaker. In this case, we also develop new control design methods taking into account the stabilizability of the true system. We close the talk by discussing experiment design methods. These methods construct suitable inputs for the unknown system, in such a way that the resulting data are informative for data-driven stabilization (again taking into account the prior knowledge).

Literature:

Book on data-driven control: https://henkvanwaarde.github.io/dblsct.html
Data informativity with prior knowledge: https://arxiv.org/pdf/2510.25452
Experiment design with prior knowledge: https://arxiv.org/pdf/2512.01876

Abstract:

This talk begins by presenting our recent studies on the vulnerability of data-driven control methods to adversarial perturbations in collected data. We show that carefully crafted attacks can subtly manipulate training data to destabilize the closed-loop system while remaining difficult to detect. We further demonstrate the feasibility of embedding backdoors through data perturbation attacks, illustrated by an example involving automated vehicles.

Motivated by those findings, we investigate data informativity under data perturbation, a generalized noise model additive to the whole input-state data matrices constrained within a linear subspace and characterized by a quadratic matrix inequality (QMI). We derive necessary and sufficient conditions formulated as tractable linear matrix inequalities for data informativity under data perturbation with respect to stabilization and performance guarantees via state feedback, as well as stabilization via output feedback. Our proposed framework encompasses and extends existing analyses that consider exogenous disturbances and measurement noise, while also relaxing several restrictive assumptions commonly made in prior work. A central challenge in the data perturbation setting arises from the non-convexity of the set of systems consistent with the data, which renders standard matrix S-procedure techniques inapplicable. To resolve this issue, we develop a novel matrix S-procedure that does not rely on convexity of the system set by exploiting geometric properties of QMI solution sets. Furthermore, we derive sufficient conditions for data informativity for structured data perturbation by approximating the combined noise effect through the QMI framework. The effectiveness is demonstrated through several numerical examples.

Literature:

Data Informativity under Data Perturbation: https://arxiv.org/abs/2505.01641
Adversarial Destabilization Attacks to Direct Data-Driven Control: https://arxiv.org/abs/2507.14863

Abstract:

The Linear Programming (LP) approach to dynamic programming is one of the established methods for approaching infinite horizon optimal control problems. When applied to systems with continuous state or action spaces it gives rise to infinite dimensional LP that need to be approximated for computation, in the spirit of approximate dynamic programming (ADP). One advantage of doing this is the possibility of using data sampled for the system to generate finite dimensional approximations of the LP whose quality improves as the amount of data increases. Implementing this approach computationally, however, is far from straightforward, a fact that has arguably prevented the LP approach from becoming as popular as other ADP methods such as variants of policy iteration and value iteration. Besides the curse of dimensionality, additional difficulties such as obtaining bounded solutions consistently, boot-strapping small amounts of data to generate additional samples, etc. need to be addressed. We first introduce the general LP method for approximate dynamic programming, then discuss conditions on the system and the data under which such computational difficulties can be addressed.

Literature:

L. Falconi, A. Martinelli, and J. Lygeros, “Data-driven optimal control via linear programming: Boundedness guarantees,” IEEE Transactions on Automatic Control, vol. 70, no. 3, pp. 1683–1697, 2025.

Abstract:

Data-driven controller design such as  IFT, VRFT, FRIT, and so on, is known as a method that yields suitable controller parameters which achieve  nice tracking property for a given reference model.  However there are two critical problems: one is that we need a reference model, the other is that it is difficult to guarantee the stability of the closed loop system. To overcome them, for the former one, we provide a data-driven design method for a servo system without reference model by utilizing data-driven prediction. For the later one, we combine data-informativity for the stability and data-driven controller design for servo systems. Finally, we unify these two different approaches and also present the proposed design procedure as linear matrix inequalities. 

Abstract:

Typical large-scale networked systems such as brain, social, and biological networks require principled choices of where to intervene and how strongly under limited actuation resources. Controllability scores address this need by quantifying node importance through controllability-Gramian–based metrics. We introduce two scores, the Volumetric Controllability Score (VCS) and the Average Energy Controllability Score (AECS), whose optimization objectives mirror the log-det and trace-inverse criteria, respectively, and thus exhibit a close structural analogy to D- and A-optimal experimental design (OED). We then explain where the analogy breaks. Controllability scores allocate virtual actuation to rank intervention locations for fixed network dynamics and emphasize unique, interpretable solutions and robustness to modeling conventions, whereas OED primarily targets statistical efficiency for parameter estimation under a specified model and design space. Building on this positioning, we present optimization-based formulations with uniqueness guarantees and discuss computational strategies for large networks. We also showcase applications including brain-network analysis and present extensions to the target controllability score for settings where actuation is available only on a limited subset of nodes and objectives concern specific target states. Finally, we outline a data-driven pathway to extend controllability scores to unknown nonlinear systems via Koopman-operator–based linear representations identified from time-series data.

Literature:

https://arxiv.org/pdf/2205.03032
https://arxiv.org/pdf/2408.03023
https://arxiv.org/pdf/2501.13345
https://arxiv.org/pdf/2510.13354
https://arxiv.org/pdf/2601.10260
https://arxiv.org/pdf/2602.11921

Abstract:

With the development of data science technologies, data-driven control, which directly designs controllers from plant response data, has been attracting significant attention.  A challenging issue in this area is data-driven control design of large-scale multi-agent systems, which may require substantial data. 

This presentation addresses data-driven control design for output consensus in a multi-agent system consisting of heterogeneous linear agents.  In the model-based approach, an output synchronization control scheme has been proposed by combining a synchronous reference generator with a tracking controller derived from an algebraic equation associated with the internal model principle.  However, such an algebraic equation cannot be applied to the data-driven design using noisy data. 

In this presentation, we consider the output consensus, where the reference generator and the tracking controller are an integrator and a type-1 servo controller, respectively.  In this case, we can incorporate the reference generator model into the controller without resorting to the aforementioned algebraic equation.  

We present an output consensus controller with an LMI condition for data-driven servomechanism design using noisy data.  An advantage of our method is that the consensus controller can be designed at each agent without collecting data from other agents.  The effectiveness of our method is verified through simulations.

Abstract:

I consider two classes of data-driven problems for quarter-plane causal 2D systems. In the first one ("data-driven simulation") we are given sufficiently informative input-output data from such a system; boundary conditions; and an input- sequence over a finite extent set S of ZxZ. We compute from such data a local (i.e. over S) solution to the (unknown) system of partial difference equations governing the system. In the second class of problems (optimal output-tracking problem) we are given a quadratic cost functional and a reference trajectory. We compute from the data an input-output trajectory locally (i.e. over S) satisfying the system equations, that minimizes the quadratic cost.

Abstract:

This presentation introduces an operator-theoretic framework for derivative-free data-driven stabilization of continuous-time systems. A major challenge in this field is the direct use of continuous-time data without relying on sampling or filtering. To address this, we propose an operator-based data-embedding method to handle continuous-time data directly, which enables a seamless extension of discrete-time data-matrix techniques to the continuous-time setting.

We start with introducing a synthesis-operator approach for state-feedback continuous-time systems. By embedding trajectories into synthesis operators, we characterize the set of systems consistent with given data. Based on this operator-based characterization, we discuss data informativity for stabilization and system identification. The stabilization result is further extended to output-feedback system systems with noisy input-output data via behavioral theory. The core contribution is the derivation of necessary and sufficient conditions under which noisy data are informative for quadratic stabilization. These conditions are formulated as linear matrix inequalities by exploiting the finite-rank structure of the synthesis operators. 

The synthesis-operator approach for continuous-time systems is inspired by a study on data-driven control of discrete-time infinite-dimensional systems, where infinite-length data are generally required for system analysis and controller design. In this talk, we will highlight the theoretical connections between these two domains.

Literature:

https://arxiv.org/abs/2511.21041
https://arxiv.org/abs/2602.02992
https://arxiv.org/abs/2511.16008

Abstract:

In this presentation, we first review the concept of compressed sensing (also known as sparse modeling), a fundamental method for data compression, model reduction, feature selection, and super-resolution. We then demonstrate applications of compressed sensing to problems in systems and control, such as sparse control design and sparse system identification. The topics are selected from the presenter’s recent book, Compressed Sensing in Systems and Control (2025).

Literature:

https://nagahara-masaaki.github.io/spm_en.html

Abstract:

This talk investigates the sample complexity of active learning for linear systems, which is a key metric for evaluating the data-efficiency of system identification. Sample complexity refers to the minimum number of data samples required to estimate the system matrices with the prescribed accuracy and confidence levels. We first establish lower bounds on the sample complexity for any active learning algorithm. Next, we propose an active learning algorithm with near-optimal excitation inputs. Our approach, based on ordinary least squares and semidefinite programming, attains near-optimal sample complexity while allowing for efficient computation of an estimate of a system matrix. Consequently, the obtained sample complexity bounds characterize the fundamental limits of active excitation in linear system identification.