The Differentiable Feasibility Pump

This repository provides method to run the differentiable feasibility pump algorithm. It is based on the paper The Differentiable Feasibility Pump by Matteo Cacciola, Alexandre Forel, Andrea Lodi, and Antonio Frangioni. The paper is available here.

First, we recommend creating a virtual Python environment:

    python -m venv env_fp/
    source env_fp/bin/activate

Then, install the dependencies using:

    python -m pip install --upgrade pip
    pip install torch --index-url https://download.pytorch.org/whl/cpu
    python -m pip install .

The installation can be checked by running:

    python -m pytest

This project requires the gurobi solver. Free academic licenses can be obtained at:

• https://www.gurobi.com/academia/academic-program-and-licenses/
• https://www.gurobi.com/downloads/end-user-license-agreement-academic/

The experiments in the paper are based on instances form the MIPLIB library. The instances should be in .mps format. By default, the instances are read in the folder data\MIPLIB. A few small instances are included in this repository for test purposes.

Experiments can be run using the script script\main.py and specifying the instance name:

    python script\main.py instance_name

The differentiable feasibility pump can be configured through optional command line arguments. For instance, the variant called DP4 in the paper can be run using:

    python script\main.py instance_name --lr 0.6 --integ_loss 10 --feas_loss 0.001 --reg_loss 0.1 --p 2

By default, the feasibility pump of Fischetti et al. (2005) is run.