On Space-Filling Input Design for Nonlinear Dynamic Model Learning: A Gaussian Process Approach

Yuhan Liu, Máté Kiss, Roland Tóth, Maarten Schoukens

Abstract

While optimal input design for linear systems has been well-established, no systematic approach exists for nonlinear systems where robustness to extrapolation/interpolation errors is prioritized over minimizing estimated parameter variance. To address this issue, we develop a novel space-filling input design strategy for nonlinear system identification that ensures data coverage of a given region of interest. By placing a Gaussian Process (GP) prior on the joint input-state space, the proposed strategy leverages the GP posterior variance to construct a cost function that promotes space-filling input design. Consequently, this enables maximization of the coverage in the region of interest, thereby facilitating the generation of informative datasets. Furthermore, we theoretically prove that minimization of the cost function implies the space-filling property of the obtained data. Effectiveness of the proposed strategy is demonstrated on both an academic and a mass-spring-damper example, highlighting its potential practical impact on efficient exploration of the dynamics of nonlinear systems.

Introduction

In system identification—the process of building mathematical models of dynamic systems from observed data—linear systems are a well-understood domain. Established "optimal" designs, such as D-optimality, effectively minimize the variance of estimated parameters. However, nonlinear systems present a unique challenge: model accuracy depends less on parameter variance and more on generalization error. If training data is clustered within a narrow operating range, the model will likely fail when the system moves into unexplored regions of the state space.

Designing experiments to explore these regions creates a "chicken-and-egg" dilemma: effective exploration requires knowledge of the system's dynamics, yet those dynamics are precisely what the experiment aims to learn. This research addresses the offline version of this problem: given a prior (potentially crude) model, how can we design an input sequence that forces the system to visit as much of the relevant state space as possible?

Methodology: A Gaussian Process Approach

The proposed strategy treats unknown system dynamics as realizations of a Gaussian Process (GP). The core innovation lies in using the GP posterior variance as a differentiable proxy for geometric distance.

  1. GP Prior and Anchor Points: A GP prior is placed on the latent system functions. The user defines a grid of "anchor points" across a region of interest (the state-input space to be covered).
  2. Cost Function Optimization: The objective function is the average GP posterior variance across these anchor points. Because GP posterior variance at a point depends on the proximity of existing data rather than observed values, minimizing this variance effectively "pushes" the system's trajectory toward the anchor points.
  3. Theoretical Guarantees: The authors formally prove that minimizing this cost function guarantees an "$\epsilon$-space-filling" property, provided the kernel is monotonically decreasing with distance. This ensures that the optimizer has a clear path to improve coverage.

Experimental Evaluation

The method was primarily tested on a nonlinear mass-spring-damper system, representing a mass moving on a rail. This setup compared the GP-based approach against traditional methods, such as Schroeder phase multisines, which are commonly used but do not explicitly optimize for state-space coverage.

Key Results

The GP-based optimization significantly improved the "filling distance" ($\rho$), a metric where lower values indicate better coverage. As shown in the results below, the optimized data points (orange) fill the designated region of interest much more uniformly than the initial data (blue).

Comparison of initial and optimized data coverage in the 3D state-input space for the mass-spring-damper system.

Quantitatively, for $M=512$ anchor points, the proposed method achieved a filling distance of 0.3029, a substantial improvement over the 1.1811 achieved by standard multisine designs.

Numerical comparison of filling distance (ρ) for different numbers of anchor points (M).

The method proved robust even under practical constraints, such as limited input frequency ranges, and demonstrated efficiency by achieving high coverage with relatively small sample sizes.

Conclusion and Future Work

This GP framework provides a powerful, distribution-agnostic tool for "filling the holes" in experimental data. While the $O(N^3)$ computational complexity of GP variance calculations may pose challenges for very large datasets, it remains highly effective for the offline design of moderately sized experimental sequences.

Future research will explore iterative approaches for scenarios where no prior model exists. By starting with a generic experiment and iteratively refining the model and the space-filling design, researchers can systematically explore and learn the dynamics of complex nonlinear systems from scratch.

Summary updated

Abstract

While optimal input design for linear systems has been well-established, no systematic approach exists for nonlinear systems where robustness to extrapolation/interpolation errors is prioritized over minimizing estimated parameter variance. To address this issue, we develop a novel space-filling input design strategy for nonlinear system identification that ensures data coverage of a given region of interest. By placing a Gaussian Process (GP) prior on the joint input-state space, the proposed strategy leverages the GP posterior variance to construct a cost function that promotes space-filling input design. Consequently, this enables maximization of the coverage in the region of interest, thereby facilitating the generation of informative datasets. Furthermore, we theoretically prove that minimization of the cost function implies the space-filling property of the obtained data. Effectiveness of the proposed strategy is demonstrated on both an academic and a mass-spring-damper example, highlighting its potential practical impact on efficient exploration of the dynamics of nonlinear systems.
BibTeX
@article{liu2025space,
  author = {Yuhan Liu and Máté Kiss and Roland Tóth and Maarten Schoukens},
  title = {{On Space-Filling Input Design for Nonlinear Dynamic Model Learning: A Gaussian Process Approach}},
  journal = {IEEE Control. Syst. Lett},
  year = {2025},
  doi = {10.1109/lcsys.2025.3582509},
}
Figure