NLPOpt-Net-Experiments

This repository includes the code and instructions to run the experiments provided in the article to reproduce the results.

The main repository includes the notebook to reproduce the result for Table 01.

Running the Experiments

To run the experiments you need to call:

python main.py --type 'type' --action 'action' --p 'no of parameter' --n 'no of variables' --me 'no of equality' --mi 'no of inequality' --train_frac 'training fraction'

We excecuted the followings to generate the results provided in the article.

Results in Table 02

python main.py --type qp --action compare --p 50 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type qcqp --action compare --p 50 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type nlp --action compare --p 50 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type nonconvx --action compare --p 50 --n 100 --me 50 --mi 50 --train_frac 0.8

Results in Table 03

python main.py --type qp --action compare_dc3 --p 10 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 25 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 75 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 100 --n 100 --me 50 --mi 50 --train_frac 0.8

Results in Table 04

python main.py --type qp --action compare_dc3 --p 50 --n 100 --mi 50 --me 10 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --mi 50 --me 30 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --mi 50 --me 50 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --mi 50 --me 70 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --mi 50 --me 90 --train_frac 0.8

Results in Table 05

python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 10 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 30 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 70 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 90 --train_frac 0.8

Results in Table 06

python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 50 --train_frac 0.8
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 50 --train_frac 0.6
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 50 --train_frac 0.4
python main.py --type qp --action compare_dc3 --p 50 --n 100 --me 50 --mi 50 --train_frac 0.2

Results in Table 07

For the active set agreement, the value is saved in all runs in the summary.json file under each problem run directory.

Citation

@misc{bimol2026nlpoptnet,
      title={NLPOpt-Net: A Learning Method for Nonlinear Optimization with Feasibility Guarantees}, 
      author={Bimol Nath Roy and Rahul Golder and MM Faruque Hasan},
      year={2026},
      eprint={2605.00260},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.00260}, 
}