Baselines

To compare our approaches with other existing methods, we propose several baselines:

• The classic bundle method with heuristic strategies to tune the regularization parameter. Further details can be found: https://lipn.univ-paris13.fr/~demelas/Manuscript_Final.pdf
• A Flux implementation of the classic gradient descent.
• A Flux implementation of Adam optimizer.

For each baseline, we consider an initial (step-size/regularization) parameter obtained through a grid search, and we save all the results in a JSON file.

Assuming that you have already downloaded or created your data and saved it in ./data/<your_folder>/, you can run the baselines as:

julia --project=. ./runs/runBaselines.jl --folder ./data/<your_folder>/ --maxIterDescentType 1000 --maxIterBundle 100 --TS 0.01 0.1 1.0 1 10 100 1000

Parameters Explanation

• --folder ./data/<your_folder>/: The folder containing your data.
• --maxIterDescentType 1000: The maximum number of iterations for Adam and Descent.
• --maxIterBundle 100: The maximum number of iterations for the Bundle methods.
• --TS 0.01 0.1 1.0 1 10 100 1000: The initial parameters to consider in the grid search.

Further References

• Bundle Network: A fully ML-based Bundle Method.
• Hyper-Parameter Learning: Learning an hyper-parameter in the (aggregated) Bundle method.