Solver Free Decision-Focused Learning for Linear Optimization Problems

Repository Structure

This repository contains the code accompanying our NeurIPS 2025 paper on Solver-Free Decision-Focused Learning for Linear Optimization Problems. The repository contains all data generators, loss functions, and experiment scripts used in the paper.

• experiments/ – main experiment script, along with configuration files used in the paper.
• architectures/ – PyTorch models used to predict cost coefficients.
• data/ – data generators, including code for precomputation of adjacent vertices.
• methods/ – implementations of losses not included in PyEPO (LAVA and CaVE).
• optimization_problems/ – Gurobi implementations of random LP, shortest path, and knapsack instances.
• utils/ – training utilities, regret computation and helper functions.
• plotting/ – scripts for visualizing regret/time curves for each method.

Requirements

This project depends on the following requirements. The listed version numbers were used in our experiments, but other versions may also work. We used Python 3.13 in our experiments.

• torch==2.8.0
• pyepo==1.0.2
• cvxpy==1.7.3
• gurobipy==12.0.3
• numpy==2.3.5
• scipy==1.16.3
• scikit_learn==1.7.2
• matplotlib==3.10.7
• tqdm==4.67.1
• pathos==0.3.4

Running Experiments

Every experiment is driven by a JSON config that specifies:

• benchmark: one of random_lp, shortest_path, knapsack, california_house_price_knapsack.
• Data/learning parameters (num_data, num_feat, deg, e, batch_size, num_epochs, etc.).
• Solver/model parameters (num_vars, num_constrs, model, etc.).
• experiments: list of method dictionaries, e.g., { "name": "LAVA", "method": "lava", "learning_rate": 0.001, "hyperparameters": {"threshold": -0.1} }.
• seeds: list of random seeds for repeated runs.

Example usage:

cd experiments
python main_experiment.py --config configs/first_experiment/config_random_lp.json

Outputs:

• results/<timestamp>/config.json – the config that was executed.
• results/<timestamp>/results.json – train/val/test regret arrays and wall-clock training times.
• results/<timestamp>/plots/*.png – figures generated by plotting/plot_results.py.

Adjust --config to point to any file under experiments/configs/**.

Citation

@article{berden2025solver,
  title={Solver-Free Decision-Focused Learning for Linear Optimization Problems},
  author={Berden, Senne and Mahmuto{\u{g}}ullar{\i}, Ali {\.I}rfan and Tsouros, Dimos and Guns, Tias},
  journal={arXiv preprint arXiv:2505.22224},
  year={2025}
}

Contact

For any questions or issues, please contact Senne Berden ([email protected]) or open an issue in this repository.