Systematic construction of continuous-time neural networks for linear dynamical systems
This article presents a systematic methodology for constructing continuous-time neural networks (specifically, dynamic neural networks or DyNNs) tailored to simulate linear time-invariant (LTI) systems. The approach avoids traditional gradient-based training and instead derives both the architecture and parameters of the network directly from the LTI system's state-space representation.\N\NThe proposed framework relies on a preprocessing algorithm that block-diagonalizes the system matrix using Schur decomposition and eigenvalue clustering. This transformation enables the construction of sparse neural networks whose hidden layers are organized horizontally, a departure from the standard vertical layer design. Each neuron in the hidden layers solves either a first-order or second-order ODE, determined by the spectral properties of the system matrix.\N\NThe main theoretical contribution is a mapping from the LTI system parameters to those of the DyNN, ensuring equivalence of input-output behavior. An upper bound on the numerical approximation error is also derived. The methodology is demonstrated on several numerical examples to show accuracy and sparsity.\N\NThe work contributes to the field of scientific machine learning by offering a principled, model-based alternative to neural architecture search and backpropagation training for a specific but important class of dynamical systems.
- A 3(2) pair of Runge-Kutta formulas
- A data-driven approximation of the koopman operator: extending dynamic mode decomposition
- A family of embedded Runge-Kutta formulae
- A New Certified Hierarchical and Adaptive RB-ML-ROM Surrogate Model for Parametrized PDEs
- A proposal on machine learning via dynamical systems
- Adaptive deep learning for high-dimensional Hamilton-Jacobi-Bellman equations
- Algorithm 432 [C2]: Solution of the matrix equation AX + XB = C [F4]
- Automatic Selection of Methods for Solving Stiff and Nonstiff Systems of Ordinary Differential Equations
- Construction of reduced-order models for fluid flows using deep feedforward neural networks
- Data-driven discovery of closure models
- Data-driven learning of nonautonomous systems
- Deep learning‐based reduced order models for the real‐time simulation of the nonlinear dynamics of microstructures
- Deep vs. shallow networks: an approximation theory perspective
- Dynamical Systems–Based Neural Networks
- Extended dynamic mode decomposition with dictionary learning: A data-driven adaptive spectral decomposition of the Koopman operator
- Fourth- and Fifth-Order, Scaled Rungs–Kutta Algorithms for Treating Dense Output
- Generative Adversarial Network for Probabilistic Forecast of Random Dynamical Systems
- scientific article; zbMATH DE number 940566 (Why is no real title available?)
- scientific article; zbMATH DE number 7626756 (Why is no real title available?)
- Learning the Dynamics for Unknown Hyperbolic Conservation Laws Using Deep Neural Networks
- Linearly recurrent autoencoder networks for learning dynamics
- Model order reduction: Theory, research aspects and applications. Selected papers based on the presentations at the workshop `Model order reduction, coupled problems and optimization', Leiden, The Netherlands, September 19--23, 2005.
- Neural architecture search: a survey
- Neural Control of Parametric Solutions for High-Dimensional Evolution PDEs
- On a quasi-linear parabolic equation occurring in aerodynamics
- On swapping diagonal blocks in real Schur form
- On the Koopman operator of algorithms
- Order Barriers for Continuous Explicit Runge-Kutta Methods
- Research on gain scheduling
- Solving Ordinary Differential Equations I
- Stabilized neural ordinary differential equations for long-time forecasting of dynamical systems
- Subspace model identification Part 2. Analysis of the elementary output-error state-space model identification algorithm
- Subspace model identification Part 3. Analysis of the ordinary output-error state-space model identification algorithm
- Symplectic Model Reduction of Hamiltonian Systems on Nonlinear Manifolds and Approximation with Weakly Symplectic Autoencoder
- The MATLAB ODE Suite
- The partial differential equation ut + uux = μxx
- Time series forecasting of COVID-19 transmission in Canada using LSTM networks
This page was built for publication: Systematic construction of continuous-time neural networks for linear dynamical systems
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6959695)
