https://github.com/ilyaorson/directcollocationneuralodes
A direct collocation approach for more effective training of Neural ODEs
https://github.com/ilyaorson/directcollocationneuralodes
collocation julia neural-ode optimal-control
Last synced: 6 months ago
JSON representation
A direct collocation approach for more effective training of Neural ODEs
- Host: GitHub
- URL: https://github.com/ilyaorson/directcollocationneuralodes
- Owner: IlyaOrson
- License: mit
- Created: 2023-11-10T11:30:59.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2025-01-17T08:03:29.000Z (12 months ago)
- Last Synced: 2025-01-17T08:29:53.871Z (12 months ago)
- Topics: collocation, julia, neural-ode, optimal-control
- Language: HTML
- Homepage: https://ilyaorson.github.io/DirectCollocationNeuralODEs/
- Size: 759 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Train Neural ODEs with Direct Collocation
This repository implements a method to optimize Neural ODEs by approximating their integrated paths with local polynomials.
Instead of relying on backpropagation through the integrator, we jointly optimize the neural network weights and the coefficients of these polynomials.
This approach increases the optimization space and includes constraints to approximate the underlying differential equations and any state constraints.
**Key Idea:**
* Joint optimization of neural network parameters and polynomial path approximations.
* Avoid backpropagation through the numerical ODE solver.
* Include constraints for ODE satisfaction and state limits.
**Example:**
This repository showcases the constrained control van der Pol problem from the paper *Neural ODEs as Feedback Policies for Nonlinear Optimal Control* ([IFAC 2023](https://doi.org/10.1016/j.ifacol.2023.10.1248)).