procedure.md

Procedure to train a model for new function class

1. Fix error tolerance and pick adequate step sizes.
   • The function one_fun() in comparison_ode.py and using a PredictorConstODE might be helpful.
   • From my experience about 5-10 step sizes with linear or logarithmic spacing works well.
   • Alternatively, we can also start by fixing step sizes and then picking the error tolerance accordingly.
2. It might be helpful to use an input-scaler for certain function classes.
   • Fit the scaler using build_scaler() in scaling.py and save it as a file.
   • Load the fitted scaler before starting the training.
3. Execute the training of the model using the main() function in train_predictor_ode.py. The following tweaks may help to improve performance:
   • slowly reduce the learning rate
   • play with the eps parameter to control exploration
   • use a memory
   • use a different rewards function (see reward_functions.py)
   • different NN model parameters (loss function, optimizer, regularizer, network topology, ...)
   • ...
4. Evaluate the performance of the trained model.
   • The function one_fun() in comparison_ode.py can be used to provide an overview of the model's behavior.
   • For a precise performance evaluation use the function pareto_model() in comparison_ode.py.
   • Compare the performance to RK45 by evaluating RK45 using pareto_ode45() and plot the comparison with plot_pareto().
5. Save the model as well as all the used parameters (error tolerance, step sizes, etc.) in a new directory.