A review published in Nature Reviews Electrical Engineering establishes that contemporary machine learning methods do not replace classical system identification but rather extend its foundational scope. Authored by Gianluigi Pillonetto, Tianshi Chen, and Lennart Ljung, the analysis examines system identification as the formal scientific discipline responsible for constructing dynamical models directly from experimental data. Although system identification is sometimes characterized as a mature discipline whose primary theoretical and practical challenges have already been resolved, the review demonstrates that the statistical foundations established by the field remain directly relevant to contemporary machine learning research.
The authors emphasize that the core questions presently guiding modern machine learning have long served as the central pillars of system identification. Key among these challenges is the problem of inferring reliable, robust models from experimental data that are inherently finite and contaminated by noise. In addition, researchers must address how to effectively balance model complexity against predictive capability, preventing overfitting while maintaining sufficient flexibility. A further critical challenge lies in quantifying uncertainty within dynamical systems, in which small prediction errors can continuously propagate and accumulate over extended operational time horizons.
To resolve these enduring questions, the review traces the historical trajectory of the discipline from the classical identification framework and the fundamental generalization problem to a modern predictor-oriented viewpoint. This predictor-oriented perspective unifies classical methodologies with contemporary machine learning algorithms under a shared mathematical structure. Across both historical and modern paradigms, a dynamical model is systematically identified and refined by directly optimizing its predictive performance on observed physical data, showing that predictive optimization forms a universal bridge across disparate modelling traditions.
The review also highlights a major conceptual evolution in how researchers manage model complexity, moving from discrete model-order selection to continuous functional regularization. Historically, investigators relied on discrete model-order selection criteria to determine appropriate model dimensions. The shift toward functional complexity control and continuous regularization now enables researchers to exploit large basis expansions, kernel methods, and Gaussian-process priors without overfitting noisy observations. As illustrated in the review, sample impulse responses derived from kernel-defined Gaussian-process priors demonstrate how continuous regularization maintains predictive rigor.
Beyond kernel-based approaches, contemporary system identification increasingly incorporates deep neural architectures to model complex dynamical processes. These deep learning frameworks effectively extend system identification methodology to strongly nonlinear, high-dimensional dynamics that challenge traditional linear or low-order models. However, the authors caution that the deployment of deep neural networks requires careful, ongoing attention to system stability, rigorous uncertainty quantification, and the demands of long-horizon prediction, where accumulating errors can degrade performance if left unmanaged.
Throughout the review, system identification is conceptualized as an iterative process linking physical observations, predictive modelling, and modern learning frameworks. Examining the interactions among paradigms in modern system identification reveals how computational software tools, statistical learning theory, and experimental measurements converge into a cohesive operational workflow. By treating physical observations and predictive learning within an integrated iterative loop, the discipline provides a principled framework for evaluating how models perform when confronted with real-world dynamical behaviors.
Looking toward future developments, the authors outline several promising research frontiers that build upon these unified foundations. These emerging directions include uncertainty-aware learning, transferable dynamical representations capable of operating across diverse settings, and physical artificial intelligence. In physical artificial intelligence, the core principles of system identification serve as the essential foundation for autonomous technologies that must interact directly and reliably with the physical world, ensuring that learned dynamical systems remain grounded in rigorous predictive and statistical methods.
Subject of Research: Predictor-oriented unification of classical system identification and machine learning for dynamical systems
Article Title: A unified view of system identification in the learning era
Article References: Pillonetto, G., Chen, T., & Ljung, L. (2026). A unified view of system identification in the learning era. Nature Reviews Electrical Engineering. https://doi.org/10.1038/s44287-026-00345-z
Image Credits: AI Generated
DOI: 10.1038/s44287-026-00345-z
Keywords: system identification, dynamical models, machine learning, predictor-oriented framework, regularization, kernel methods, Gaussian processes, deep neural architectures, uncertainty quantification, model-order selection, physical artificial intelligence, transferable dynamical representations, uncertainty-aware learning, stability, long-horizon prediction, functional complexity control, nonlinear dynamics
Cite Scienmag News
Blake Davidson. (November 16, 2026). Predictor-Oriented Framework Unifies Classical System Identification and Modern Machine Learning. Scienmag. https://scienmag.com/predictor-oriented-framework-unifies-classical-system-identification-and-modern-machine-learning/
Blake Davidson. "Predictor-Oriented Framework Unifies Classical System Identification and Modern Machine Learning." Scienmag, 16 November 2026, https://scienmag.com/predictor-oriented-framework-unifies-classical-system-identification-and-modern-machine-learning/. Accessed 16 November 2026.
Blake Davidson. "Predictor-Oriented Framework Unifies Classical System Identification and Modern Machine Learning." Scienmag. November 16, 2026. https://scienmag.com/predictor-oriented-framework-unifies-classical-system-identification-and-modern-machine-learning/








