Surrogate-Losses-for-Decision-Focused-Learning-using-Differentiable-Optimization

Code for the ECAI 2025 paper "Minimizing Surrogate Losses for Decision-Focused Learning using Differentiable Optimization"

Summary

This repository contains the implementation of our novel approach to Decision-Focused Learning (DFL) using differentiable optimization layer DYS-Net. We explore various surrogate losses and demonstrate their effectiveness across multiple LP/ILP/MILP optimization problems.

Citation

If you use this code in your research, please cite our paper (will update after the proceeding is publicly available):

@misc{mandi2025minimizingsurrogatelossesdecisionfocused,
      title={Minimizing Surrogate Losses for Decision-Focused Learning using Differentiable Optimization}, 
      author={Jayanta Mandi and Ali İrfan Mahmutoğulları and Senne Berden and Tias Guns},
      year={2025},
      eprint={2508.11365},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.11365}, 
}

Installation

We recommend using a virtual environment to avoid conflicts with other Python packages:

# Create a virtual environment
python3 -m venv env_dfl

# Activate the virtual environment
# On Linux/Mac
source env_dfl/bin/activate
# On Windows
# env_dfl\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Instructions for Running Experiments

To run the experiments, execute the Exp_run.sh script from the command line:

./Exp_run.sh

Make sure the script has executable permissions:

chmod +x Exp_run.sh

The Exp_run.sh script will execute all the necessary experiment scripts in the correct order, running various models with different parameters.

Model Name in Exp_run.sh

Here's how the math notation in our paper translates to actual model names in the code:

• CVX Models

  • $Regret^{CVX}$ corresponds to CVX-Regret
  • $SqDE^{CVX}$ corresponds to CVX-Squared
  • $SPO_{+}^{CVX}$ corresponds to CVX-SPO
  • $SCE^{CVX}$ corresponds to CVX-SCE

• DYS-Net Models

  • $Regret^{DYS}$ corresponds to DYS-Regret
  • $SqDE^{DYS}$ corresponds to DYS-Squared
  • $SPO_{+}^{DYS}$ corresponds to DYS-SPO
  • $SCE^{DYS}$ corresponds to DYS-SCE

Modifying Configuration Parameters

To run experiments with different configurations, you can modify the parameter values directly in the Exp_Run.sh file:

1. For Shortest Path experiments, modify the --grid_size parameter (e.g., change from 15 to another value)
2. For Knapsack experiments, modify the --num_items parameter (e.g., change from 400 to another value)
3. For Facility Location experiments, modify the --num_customers parameter (e.g., change from 200 to another value) and --num_facilities if needed

Hyperparameter Configurations

Important: The experiment scripts read hyperparameter configurations from the pkg/configs folder. These configuration files include:

• shortestpath_config.json and shortestpath_DYSconfig.json: For Shortest Path experiments
• knapsack_config.json and knapsack_DYSconfig.json: For Knapsack experiments
• facilitylocation_config.json and facilitylocation_DYSconfig.json: For Facility Location experiments

If you need to modify hyperparameters such as learning rates, model architectures, or optimization settings, edit these JSON configuration files rather than changing the command-line arguments in the Exp_Run.sh script.

Acknowledgments

Our implementation builds upon PyEPO, a benchmarking library for End-to-End Predict-then-Optimize techniques.