This repository contains scripts for generating partition-based formulations for trained ReLU neural networks and a few test instances implemented in Gurobi. More details on the method here: https://arxiv.org/abs/2102.04373.

Please cite this work as:

@article{tsay2021partition,
  title   =   {{Partition-based formulations for mixed-integer optimization of trained ReLU neural networks}},
  author  =   {Tsay, Calvin and Kronqvist, Jan and Thebelt, Alexander and Misener, Ruth},
  journal =   {ArXiv},
  volume  =   {2102.04373},
  year    =   {2021}
}

The solver software Gurobi is required to run the examples. Gurobi is a commercial mathematical optimization solver and free of charge for academic research. It is available on Linux, Windows and Mac OS.

Please follow the instructions to obtain a free academic license. Once Gurobi is installed on your system, follow the steps to setup the Python interface gurobipy.

• Calvin Tsay (tsaycal) - Imperial College London
• Jan Kronqvist (jkronqvi) - KTH Royal Institute of Technology
• Alexander Thebelt (ThebTron) - Imperial College London
• Ruth Misener (rmisener) - Imperial College London

This repository is released under the Apache License 2.0. Please refer to the LICENSE file for details.

This work was supported by Engineering & Physical Sciences Research Council (EPSRC) Fellowships to CT and RM (grants EP/T001577/1 and EP/P016871/1), an Imperial College Research Fellowship to CT, a Royal Society Newton International Fellowship (NIF\R1\182194) to JK, a grant by the Swedish Cultural Foundation in Finland to JK, and a PhD studentship funded by BASF to AT.