Data-Driven Structured Controller Design Using the Matrix S-Procedure

Basic Information

• Paper ID: 2503.14949
• Title: Data-Driven Structured Controller Design Using the Matrix S-Procedure
• Authors: Zhaohua Yang, Yuxing Zhong, Nachuan Yang, Xiaoxu Lyu, Ling Shi
• Classification: math.OC (Mathematical Optimization and Control)
• Submission Date: October 14, 2025 (Preprint submitted to Automatica)
• Paper Link: https://arxiv.org/abs/2503.14949

Abstract

This paper focuses on data-driven optimal structured controller design for discrete-time linear time-invariant (LTI) systems considering both $H_2$ and $H_\infty$ performance. Three scenarios are investigated: (i) model-based structured control, (ii) data-driven unstructured control, and (iii) data-driven structured control. For $H_2$ performance, scenarios (ii) and (iii) are primarily studied, as scenario (i) has been extensively researched in the literature. For $H_\infty$ performance, all three scenarios are considered. For structured control, a linearization technique is introduced that transforms the original non-convex problem into a semidefinite programming (SDP) problem, based on which an iterative linear matrix inequality (ILMI) algorithm is developed.

Research Background and Motivation

Problem Background

1. Challenges in Networked Control Systems: In large-scale networked control systems, centralized control strategies impose substantial communication burdens, necessitating the design of controllers with sparse structures.
2. Difficulties in Structured Controller Design: Controller design with structural constraints exhibits NP-hard complexity, with existing methods primarily divided into LMI-based and gradient-based approaches.
3. Demand for Data-Driven Control: In practical applications, the dynamics of many systems are often unavailable, motivating research into data-driven control methodologies.

Limitations of Existing Methods

1. Traditional Methods Rely on Known Models: Existing structured controller design methods assume the system model is known.
2. Conservatism in Data-Driven Methods: Existing data-driven approaches such as Miller et al. 15 suffer from:
   • Over-approximation of instantaneous bounded noise as energy-bounded noise sequences, introducing additional conservatism
   • Designing structured controllers only within convex subsets of non-convex feasible regions
   • Lack of performance monotonicity guarantees

Core Contributions

1. Improved Data-Driven Framework:
   • Represents the set of all possible system matrices as the intersection of sets from each data point, which is smaller than the set in 15
   • Employs the matrix S-procedure to handle all possible system matrices, proving that the feasible region contains that of 15
   • Establishes monotonic non-decreasing performance with respect to data length
2. Novel Linearization Technique:
   • Develops a linearization technique that transforms the original non-convex problem into an SDP problem
   • Proposes the ILMI algorithm for computing structured controllers
   • Enables exploration of larger feasible regions compared to 15
3. Comprehensive $H_\infty$ Control Study:
   • Provides the first comprehensive investigation of model-based structured control, data-driven unstructured control, and data-driven structured control

Methodology Details

Problem Formulation

Consider a discrete-time LTI system:

x_{k+1} = Ax_k + Bu_k + Gd_k
y_k = Cx_k + Du_k + Hd_k

where $x_k \in \mathbb{R}^{n_x}$ is the system state, $u_k \in \mathbb{R}^{n_u}$ is the control input, $y_k \in \mathbb{R}^{n_y}$ is the controlled output, and $d_k \in \mathbb{R}^{n_d}$ is the external disturbance.

A state feedback controller $u_k = Kx_k$ is employed, where $K \in \mathbb{R}^{n_u \times n_x}$ must satisfy structural constraints $K \in S$.

Data Collection Mechanism

Assuming matrices $A, B$ are unknown but deterministic, while matrices $C, D, G, H$ are known. The collected data includes:

X = [x_0, x_1, ..., x_T] ∈ ℝ^{n×(T+1)}
U = [u_0, u_1, ..., u_{T-1}] ∈ ℝ^{m×T}

Key Assumption: Process noise is bounded, i.e., $\|w_k\|_2 \leq \epsilon$.

Core Technical Innovations

1. Improved System Matrix Set Representation

Based on data $(x_i, u_i, x_{i+1})$ at the $i$-th time step, define:

Σ_i = {(A,B) | x_{i+1} = Ax_i + Bu_i + w_i, w_i w_i^T ⪯ ε²I}

The set of all system matrices consistent with the data is:

Σ = ∩_{i=0}^{T-1} Σ_i

This is more precise than the set in 15, reducing conservatism.

2. Application of Matrix S-Procedure

For $H_2$ control, utilizing the matrix S-procedure, a sufficient condition for condition (5a) to hold for all $(A,B) \in Σ$ is the existence of non-negative $α_0, α_1, ..., α_{T-1}$ and $β > 0$ such that: