2025 ERNSI – Program

Lara Thomas – Safran Data Systems, France

Title: Identification approaches for accurate mechanical parameter characterization applied to ground antenna systems

Abstract:

Parabolic antennas have many applications in the aerospace industry, for example, to transmit and receive satellite data or to track aircraft during flight tests. SAFRAN designs and manufactures such antennas, with ongoing efforts to enhance their performances. This involves advanced and precise control systems that ensure both accurate trajectory tracking and effective disturbance rejection. This presentation will address three key identification topics.First, precise and robust control laws require a good knowledge of the antenna mechanical system. The Model Predictive Control (MPC) approach relies on an accurate physical model. Consequently, the achievable performance depends on the model reliability and on the discrepancies between the real plant and the mathematical representation used in the algorithm. The robust CRONE control methodology is based on a holistic antenna modeling including coupling, flexibility and the evaluation of parametric variations. Second, realistic performance assessment demands the integration of nonlinear mechanical effects, such as friction, backlash, and wear, into simulation models, enabling accurate prediction of tracking errors. Third, wind turbulence represents the dominant external disturbance for outdoor mechanical systems. To mitigate the risk of signal loss, wind effects must be incorporated into the dynamic model, allowing the design of control laws capable of rejecting them effectively. To this end, we have developed a real-time wind simulator for parabolic tracking antenna models, based on spectral analysis of turbulence, providing a powerful tool for robust control design and validation.

Marco Forgione – IDSIA, Switzerland

Title: Meta Learning for System Identification

Abstract:

In recent years, system identification has greatly benefited from machine learning, particularly for modeling complex, nonlinear dynamical systems with partially unknown physics. However, learning black-box models with standard approaches often requires large datasets and substantial computational resources. This talk explores meta-learning as a promising alternative. Rather than relying on manually designed system identification algorithms, we use meta-learning techniques to learn the learning procedure itself from a collection of datasets, so that it is optimally adapted to an entire class of problems. This approach has the potential to reduce computational demands and enhance generalization across related dynamical systems.

Alberto Bemporad – IMT School for Advanced Studies Lucca, Italy

Title: Efficient Numerical Optimization Methods for Learning Nonlinear State-Space Models

Abstract:

Model-based control design from input/output data relies critically on the ability to learn high-quality state-space models in a numerically efficient manner. Over the past few decades, a rich body of theoretical results and algorithms has been developed for learning linear and nonlinear models from data, with recent momentum driven by major advances in supervised learning for regression, particularly in training feedforward and recurrent neural networks (RNNs). An identification method is truly effective not only when the problem is carefully formulated — in terms of model structure, loss function, and regularization — but also when it issolved using numerical methods that deliver both high accuracy (to ensure model quality) and computational efficiency (to enable rapid hyper-parameter tuning).
In my talk, I will present batch and incremental methods for learning nonlinear state-space models, potentially with nonsmooth L1 and group-Lasso penalties, that belong to the family of quasi-Newton optimization algorithms. These methods achieve both high accuracy and fast convergence, outperform popular gradient-descent approaches, and, in the linear case, are numerically more stable than classical subspace identification techniques. I will illustrate how algorithms based on L-BFGS-B optimization — implemented in the Python package « jax_sysid » — can tackle a range of challenging identification problems, including nonlinear system identification with noise models and the simultaneous identification of self-scheduled LPV models and robust control invariant sets.