Online design of experiments by active learning for nonlinear system identification

TL;DR

This paper introduces active learning strategies for real-time input signal design in nonlinear system identification, improving sample efficiency over random excitation by adapting static AL methods to dynamic models with constraints.

Contribution

It adapts existing active learning techniques for static models to dynamic nonlinear system identification, integrating them with Kalman filtering and constraint handling.

Findings

Active learning improves sample efficiency in nonlinear system identification.

The proposed methods outperform random excitation on benchmark problems.

The approach effectively handles input and output constraints.

Abstract

We investigate the use of active-learning (AL) strategies to generate the input excitation signal at runtime for system identification of linear and nonlinear autoregressive and state-space models. We adapt various existing AL approaches for static model regression to the dynamic context, coupling them with a Kalman filter to update the model parameters recursively, and also cope with the presence of input and output constraints. We show the increased sample efficiency of the proposed approaches with respect to random excitation on different nonlinear system identification benchmarks.