First-order methods and automatic differentiation: A multi-step systems identification perspective

This paper presents a tool for multi-step system identification that leverages first-order optimization and exact gradient computation. Drawing inspiration from neural network training and Automatic Differentiation (AD), the proposed method computes and analyzes the gradients with respect to the par...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Donati, Cesare, Mammarella, Martina, Dabbene, Fabrizio, Novara, Carlo, Lagoa, Constantino
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:This paper presents a tool for multi-step system identification that leverages first-order optimization and exact gradient computation. Drawing inspiration from neural network training and Automatic Differentiation (AD), the proposed method computes and analyzes the gradients with respect to the parameters to identify by propagating them through system dynamics. Thus, it defines a linear, time-varying dynamical system that models the gradient evolution. This allows to formally address the "exploding gradient" issue, by providing conditions for a reliable and efficient optimization and identification process for dynamical systems. Results indicate that the proposed method is both effective and efficient, making it a promising tool for future research and applications in nonlinear systems identification and non-convex optimization.
DOI:10.48550/arxiv.2410.03544