Integrating Machine Learning with Control and System Identification for Enhanced Performance and Interpretability

The recent developments in the research area highlight a significant shift towards integrating machine learning techniques with traditional control and system identification methods to enhance performance, interpretability, and efficiency. A notable trend is the use of neural networks and deep learning models to approximate complex dynamics and operators, particularly in the context of model predictive control (MPC) and partial differential equations (PDEs). Innovations include frameworks for bounding approximation errors in imitation learning of MPC, leveraging neural networks' universal approximation capabilities, and developing physics-informed neural networks for real-time predictions of complex systems. Additionally, there's a growing emphasis on meta-learning and gradient-based approaches for rapid adaptation of neural state-space models, indicating a move towards more flexible and generalizable system identification methods. The integration of physical knowledge into deep learning models through techniques like orthogonal projection-based regularization and the development of novel symbolic model discovery methods further underscore the field's direction towards more interpretable and physically consistent models.

Noteworthy Papers

  • Imitation Learning of MPC with Neural Networks: Error Guarantees and Sparsification: Introduces a method for bounding approximation errors in neural network-based MPC, enhancing stability and performance guarantees.
  • Stochastic Process Learning via Operator Flow Matching: Proposes a novel framework for learning stochastic process priors on function spaces, outperforming state-of-the-art models in functional regression.
  • Orthogonal projection-based regularization for efficient model augmentation: Presents a technique to improve parameter learning and model accuracy by integrating physical knowledge into deep learning models.
  • Meta-Learning for Physically-Constrained Neural System Identification: Demonstrates a meta-learning framework for rapid adaptation of neural state-space models, improving downstream performance in practical applications.
  • Physics-Informed Latent Neural Operator for Real-time Predictions of Complex Physical Systems: Introduces a physics-informed latent operator learning framework that significantly reduces data requirements while maintaining high accuracy.
  • ELM-DeepONets: Backpropagation-Free Training of Deep Operator Networks via Extreme Learning Machines: Offers a scalable and efficient alternative for operator learning by leveraging the backpropagation-free nature of Extreme Learning Machines.

Sources

Imitation Learning of MPC with Neural Networks: Error Guarantees and Sparsification

Approximation Rates in Fr\'echet Metrics: Barron Spaces, Paley-Wiener Spaces, and Fourier Multipliers

Stochastic Process Learning via Operator Flow Matching

Orthogonal projection-based regularization for efficient model augmentation

Nonlinear port-Hamiltonian system identification from input-state-output data

Meta-Learning for Physically-Constrained Neural System Identification

A Comparison of Strategies to Embed Physics-Informed Neural Networks in Nonlinear Model Predictive Control Formulations Solved via Direct Transcription

Finite Sample Identification of Partially Observed Bilinear Dynamical Systems

NOMTO: Neural Operator-based symbolic Model approximaTion and discOvery

Physics-Informed Latent Neural Operator for Real-time Predictions of Complex Physical Systems

ELM-DeepONets: Backpropagation-Free Training of Deep Operator Networks via Extreme Learning Machines

ARMAX identification of low rank graphical models

Built with on top of