A Kernelized Operator Approach to Nonlinear Data-Enabled Predictive Control
Thomas de Jong, Siep Weiland, Mircea Lazar · 2025
Read A Kernelized Operator Approach to Nonlinear Data-Enabled Predictive Control by Thomas de Jong, Siep Weiland, Mircea Lazar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-17500.txt.
Introduction
Thomas de Jong, Siep Weiland, and Mircea Lazar present a control method that learns from data in this 2025 paper. The method plans future moves for a system by looking at past behavior, rather than at a fixed model. It is aimed at systems whose behavior bends in non-straight ways, which are harder to control.
Older data-based methods use kernel tricks to build predictors step by step. The new method instead learns a whole-system operator in a special function space. That shape lets large data sets be handled well and makes the control sums much faster to compute. For the same data size, the authors report far shorter compute times, which allows larger data sets and better control.
Catalog record
| CREATOR | Thomas de Jong, Siep Weiland, Mircea Lazar |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-17500
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Kernelized Operator Approach to Nonlinear Data-Enabled Predictive Control Thomas de Jong, Siep Weiland, Mircea Lazar Abstract This paper considers the design of nonlinear data-enabled predictive control (DeePC) using kernel functions. Compared with existing methods that use kernels to parameterize multi-step predictors for nonlinear DeePC, we adopt a novel, operator-based approach. More specifically, we employ a universal product kernel parameterization of nonlinear systems operators as a prediction mechanism for nonlinear DeePC. We show that by using a product reproducing kernel Hilbert space (RKHS) to learn the system trajectories, big data sets can be handled effectively to construct the corresponding product Gram matrix. Moreover, we show that the structure of the adopted product RKHS representation allows for a computationally efficient DeePC formulation. Compared to existing methods, our approach achieves substantially faster computation times for the same data size. This allows for the use of much larger data sets and enhanced control performance. I INTRODUCTION Willems’ fundamental lemma underpins data-driven control by enabling the prediction of system behavior directly from data, without requiring an explicit model [ 1 ] . This concept forms the foundation of data-enabled predictive control (DeePC), which uses historical data for predicting future trajectories [ 2 ] . For linear time-invariant (LTI) systems and noise-free data, equivalences between Model Predictive Control (MPC), Subspace Predictive Control (SPC) and DeePC have been established [ 3 ] . Extending DeePC to nonlinear systems has been an area of active research. For example, in [ 4 ] , DeePC was generalized to nonlinear systems using general basis functions, while in [ 5 ] , the Koopman framework was employed to parameterize the DeePC predictor. The Koopman approach provides a way to represent nonlinear dynamics by mapping system states to a higher-dimensional space by means of observable functions, where a linear operator governs their evolution. Identifying suitable basis functions or Koopman observable functions can be done using for example neural networks as was done in [ 6 ] to construct a basis, or as done in [ 5 , 7 ] for learning Koopman observable functions. Alternatively, function estimation using regularized kernel methods offers a tractable solution for learning dynamical systems due to the well-posedness of function classes in RKHSs [ 8 , 9 ] . This has led to increased interest in kernel-based approaches for data-driven control [ 10 , 11 , 12 ] . Several kernel-based approaches have also emerged in the field of data-driven predictive control. In [ 13 ] a multi-step kernel-based predictor was identified and used within an MPC framework; therein, tight bounds on the kernelized multi-step predictors’ approximation error were derived. More recently, [ 14 ] presented a kernelized DeePC formulation for nonlinear systems, which corresponds to solving a system of equations similar to Willem’s fundamental lemma. Therein, methods for robustification and efficient online implementation based on kernels that are linear in future control inputs were presented. In [ 4 ] it was shown that the special structure of basis or kernel functions that are linear in the future control inputs corresponds in fact to a DeePC formulation of linear-in-control input Koopman MPC [ 15 ] . The above kernelized predictors for data-driven predictive control require in general usage of kernel functions defined on the stacked space of initial conditions and future control input sequences. Generating long persistently exciting input sequences that sample effectively the corresponding stacked space can be challenging for nonlinear systems, however. Persistency of excitation conditions that lead to exact kernelized representations for certain classes of nonlinear systems were recently presented in [ 16 ] , along with kernelized fundamental lemmas for linear, Hammerstein and flat nonlinear systems. Alternatively, a discrete-time dynamical system can be regarded as an operator defined on a Hilbert space [ 17 ] that maps sequences of inputs to sequences of outputs. More recently, [ 18 ] defined the domain of the operator as a product Hilbert space. The corresponding operator maps input sequences and states to output sequences. Therein, it was proven that for positive definite kernels [ 19 ] the resulting product RKHS is dense in the space of nonlinear systems operators and that its unique minimizer is a universal approximator. The developed results in [ 18 ] yield an alternative formulation of Willems’ fundamental lemma and a new insight in persistency of excitation for nonlinear systems. Specifically, [ 18 ] points out that a that a persistently exciting input sequence should be paired with a linearly independent/distinct set of points in the space of initial conditions. This allows for generating larger data sets of trajectories by using a short, persistently exciting input sequence that is applied to several distinct initial conditions. Hence, it would be very interesting to employ the product RKHS operator parameterization for designing nonlinear DeePC algorithms. To this end, in this paper, building on the framework from [ 18 ] , we develop a kernelized operator approach for data-driven control of nonlinear systems through the DeePC framework. By modeling discrete-time nonlinear systems as operators on a Hilbert space, our method generalizes beyond linear systems and incorporates specific classes of nonlinear systems not considered by previous methods that rely directly on the fundamental lemma such as [ 16 , 20 ] . A main challenge in extending the kernelized operator framework of [ 18 ] for data-driven control is the dependence of the predictor on the Gram matrix, whose dimension scales with both input and state data. We demonstrate that the product kernel structure allows separation of the input and state data, which reduces the dimension of the DeePC equality constraints significantly. Additionally, we introduce a systematic data-generation method, providing a structured approach for creating datasets required to learn the operator. Our product kernel formulation offers computational advantages over previous kernel-based DeePC approaches, such as [ 14 ] . Notation and basic definitions Let \mathbb{R} , \mathbb{R}_{+} and \mathbb{N} denote the set and field of real numbers, the set of non-negative reals, and the set of non-negative integers. For every c\in\mathbb{R} and \Pi\subseteq\mathbb{R} define \Pi_{\geq c}:=\{k\in\Pi|k\geq c\} . A square n\times n matrix where all elements of 1/n is denoted by \mathbb{1}_{n} and \mathbf{1}_{n} denotes a vector in \mathbb{R}^{n} with all elements equal to 1/n . For two matrices A\in\mathbb{R}^{n\times m} , B\in\mathbb{R}^{p\times q} , we denote their Kronecker product by A\otimes B , defined as A\otimes B=\begin{bmatrix}a_{1,1}B&\dots&a_{1,m}B\\ \vdots&&\vdots\\ a_{n,1}B&\dots&a_{n,m}B\end{bmatrix}. Throughout this paper, for any finite number q\in\mathbb{N}_{\geq 1} of column vectors or functions \{\xi_{1},\dots,\xi_{q}\} we will make use of the operator \operatorname{col}(\xi_{1},\dots,\xi_{q}):=[\xi_{1}^{\top},\dots,\xi_{q}^{\top}]^{\top} . II Preliminaries We consider nonlinear MIMO systems in state space form with inputs u\in\mathbb{R}^{m} , states x\in\mathbb{R}^{n} and outputs y\in\mathbb{R}^{p} , i.e., \displaystyle x_{k+1} \displaystyle=f(x_{k},u_{k}),\quad k\in\mathbb{N}, (1) \displaystyle y_{k} \displaystyle=h(x_{k}). The functions f:\mathbb{R}^{n}\times\mathbb{R}^{m}\rightarrow\mathbb{R}^{n} and h:\mathbb{R}^{n}\rightarrow\mathbb{R}^{p} are assumed to be unknown . We denote by \mathbb{X}\subseteq\mathbb{R}^{n} , \mathbb{U}\subseteq\mathbb{R}^{m} and \mathbb{Y}\subseteq\mathbb{R}^{p} the constraint admissible sets of states, inputs and outputs respectively. For simplicity of exposition we assume that the state is measured, but the derived methods extend directly to a surrogate state consisting of stacked past inputs and outputs. Definition II.1 A function k:\mathcal{Z}\times\mathcal{Z}\rightarrow\mathbb{R} , where \mathcal{Z} denotes the data space, is called a kernel function if it satisfies the following properties: ( i ) it is symmetric, i.e., k(z_{1},z_{2})=k(z_{2},z_{1}) for all (z_{1},z_{2})\in\mathcal{Z}\times\mathcal{Z}