G2MILP: Learning to Generate Mixed-Integer Linear Programming (MILP) Instances

This is the code for G2MILP, a deep learning-based mixed-integer linear programming (MILP) instance generator.

Page: https://miralab-ustc.github.io/L2O-G2MILP/

Publications

"A Deep Instance Generative Framework for MILP Solvers Under Limited Data Availability". Zijie Geng, Xijun Li, Jie Wang, Xiao Li, Yongdong Zhang, Feng Wu. NeurIPS 2023 (Spotlight). [paper]

"G2MILP: Learning to Generate Mixed-Integer Linear Programming Instances for MILP Solvers". Jie Wang, Zijie Geng, Xijun Li, Jianye Hao, Yongdong Zhang, Feng Wu. [paper]

Environment

• Python environment

  • python 3.7
  • pytorch 1.13
  • torch-geometric 2.3
  • ecole 0.7.3
  • pyscipopt 3.5.0
  • community 0.16
  • networkx
  • pandas
  • tensorboardX
  • gurobipy

• MILP Solver

  • Gurobi 10.0.1. Academic License.

• Hydra

  • Hydra for managing hyperparameters and experiments.

In order to build the environment, you can follow commands in scripts/environment.sh.

Or alternatively, to build the environment from a file,

conda env create -f scripts/environment.yml

Usage

Go to the root directory L2O-G2MILP. Put the datasets under the ./data directory. Below is an illustration of the directory structure.

L2O-G2MILP
├── conf
├── data
│   ├── mik
│   │   ├── train/
│   │   └── test/
│   ├── mis
│   │   ├── train/
│   │   └── test/
│   └── setcover
│       ├── train/
│       └── test/
├── scripts/
├── src/
├── src_hard/
├── README.md
├── benchmark.py
├── generate.py
├── preprocess.py
├── train-hard.py
└── train.py

The hyperparameter configurations are in ./conf/. The commands to run for all datasets are in ./scripts/. The main part of the code is in ./src/. The workflow of G2MILP (using MIS as an example) is as following.

1. Preprocessing

To preprocess a dataset,

python preprocess.py dataset=mis num_workers=10

This will produce graph data for instances and the statistics of the dataset to be used for training. The preprocessed results are saved under ./preprocess/mis/.

2. Training G2MILP

To train G2MILP with default parameters,

python train.py dataset=mis cuda=0 num_workers=10 job_name=mis-default

The training log is saved under TRAIN DIR=./outputs/train/${DATE}/${TIME}-${JOB NAME}/. The model ckpts are saved under ${TRAIN DIR}/model/. The generated instances and benchmarking results are saved under ${TRAIN DIR}/eta-${eta}/.

3. Generating new instances

To generate new instances with a trained model,

python generate.py dataset=mis \
    generator.mask_ratio=0.01 \
    cuda=0 num_workers=10 \
    dir=${TRAIN DIR}

The generated instances and benchmarking results are saved under ${TRAIN DIR}/generate/${DATE}/${TIME}.

4. Generating hard instances

To generate hard instances,

python train-hard.py dataset=mis \
    cuda=0 num_workers=10 \
    pretrained_model_path=${PRETRAIN PATH}

The ${PRETRAIN PATH} is the path to the pretrained model.

Citation

If you find this code useful, please consider citing the following papers.

@inproceedings{geng2023deep,
  title={A Deep Instance Generative Framework for MILP Solvers Under Limited Data Availability},
  author={Geng, Zijie and Li, Xijun and Wang, Jie and Li, Xiao and Zhang, Yongdong and Wu, Feng},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}

@article{wang2023g2milp,
  title={G2MILP: Learning to Generate Mixed-Integer Linear Programming Instances for MILP Solvers},
  author={Wang, Jie and Geng, Zijie and Li, Xijun and Hao, Jianye and Zhang, Yongdong and Wu, Feng},
  journal={Authorea Preprints},
  year={2023},
  publisher={Authorea}
}