Model Selection for Symbolic Regression

This repository contains the code used to generate results for the paper "A Comparative Study of Model Selection Criteria for Symbolic Regression".

1) Dataset Construction

cd src
python generate_datasets.py

Datasets are written to src/data.

2) Symbolic Model Generation

Generate perturbed models from a ground-truth function:

python export_pop.py --n_m <num_mutations> --n_f <num_features> --f_n <function_name>

• --n_m: number of perturbed models to generate
• --n_f: number of features (e.g., f1 uses 10 features)
• --f_n: ground-truth function name (f1–f7)

Example (100 models for f1):

python export_pop.py --n_m 100 --n_f 10 --f_n f1

To generate all functions:

chmod +x export_pop_all.sh
./export_pop_all.sh

Outputs are .operon files named <function>_<mutations>_<features>.operon in src/functions.

3) Compute Selection Metrics

Compute metrics for one or more functions:

python compute_metrics.py [f1 f2 ...]

If no function names are provided, metrics are computed for all seven functions. Results are written to results/ with names like model_selection_methods_f1_100_10.csv.

Example:

python compute_metrics.py f1

4) Evaluate Metric Performance

python eval_modelsel.py <k> <results_csv> <test_column>

• <k>: number of top-ranked models to evaluate
• <results_csv>: CSV from the previous step
• <test_column>: ground-truth metric (e.g., MSE_test_opt)

Example:

python eval_modelsel.py 50 results/model_selection_methods_f1_100_10.csv MSE_test_opt

Batch evaluation:

chmod +x eval_modelsel.sh
./eval_modelsel.sh

Outputs are saved to results/.

5) Plot Results

python perf_plot.py

Plots are saved to results/plots as PDF files.