Personalized Algorithm Generation: A Case Study in Meta-Learning ODE Integrators

Requirements

This project uses python 3.8. Set up the project using the following step:

Install

$pip install -r requirements.txt

Test on Different Targets

Look at examples.

• Go to the folder

  $cd examples

• Run the training process (eg. linear traget)

  $python main_linear.py

• Change the parameters (such as learning rate, integrator stage, targeted order and so on) in configuration file
  • For "linear target", a huge weight is multiplied on Taylor-based regularizer as a scale because this value is too small.
  • For "square_nonlinear target", we multiply an increasing weight on MSE loss and a decreasing one on Taylor-based regularizer in order to focus on different loss in the training.

See the evaluation

• Ordercheck Run ipynb file ordercheck.
• Plot the MSE on Van der Pol Oscillator and the Brussealtor. Run ipynb file plot error bar.
• See the figures according to above evaluation in results.