chauncychtt / iclr_2019

Code to reproduce the experiments in the paper "Learning concise representations for regression by evolving networks of trees"

Experiments

Experiments can be run using analysis/ml-analyst/submit_jobs.py. See the command line options (python submit_jobs.py -h) for help.

As example, this command would launch the main experiment from the paper:

python submit_jobs.py --r -ml Feat,FeatCN,FeatCorr,RandomForest,MLPmod,Kernel,Linear,XGBoostLong -n_trials 5 ../penn-ml-benchmark/datasets/regression/

Notebooks

analysis/ contains these notebooks:

• results_iclr.ipynb produces the main results figures.
• results_appendix.ipynb produces the results comparing different stochastic optimization approaches.
• stats.ipynb depends on results_iclr.ipynb and produces the statistical tests.
• archive.ipynb contains code to reproduce the illustrative example.

Dependencies

• FEAT
• scikit-learn
• xgboost
• datasets come from PMLB