Efficient learning

The repository has codes of our online learning method based on optimizer-driven sampling.

To run the scripts in TableI and TableII. First you need to install mystic and its requirements: https://github.com/uqfoundation/mystic

mystic

highly-constrained non-convex optimization and uncertainty quantification

Adjustable settings are in _model.py, but are set at values used in manuscript. Results may be slightly different due to randomness from sampler and optimizer.

Executes with:

  $ python main_workflow.py

and may take anywhere from a few minutes to a few days, depending on the model and tolerance used, and randomness. Relevant information is printed to stdout, and also dumped into several files:

• cost.pkl: the objective function
• hist.pkl: historical testing misfit
• size.pkl: historical endpoint cache size
• score.txt: convergence of the test score versus model evaluations
• stop: a database of the endpoints of the optimizers
• func.db: a database of the learned surrogates
• eval: a database of evaluations of the objective

Any of the "pkl" files can be read like this:

 >>> import dill
 >>> cost = dill.load(open('cost.pkl', 'rb'))

while relevant results from the databases are plotted with:

python plot_func.py

and test score convergence is plotted with:

python plot_*_converge.py    (* = loose, tight)

with "loose" corresponding to loose tolerance, and tight to strict tolerance.

=======