SymILO: A Symmetry-Aware Learning Framework for Integer Linear Optimization

This repository contains the code for the NeurIPS 2024 paper: SymILO: A Symmetry-Aware Learning Framework for Integer Linear Optimization.

Environment Setup

To run this code, you need the following dependencies:

• Python 3.9.19
• CPLEX 22.1.1
• gurobipy 11.0.3
• pyg 2.5.3
• pytorch 2.4.0

Data preparation

Follow instructions here to prepare the data.

Training the Model

To train the model, you can use the following bash commands:

# symmetry-agnostic
python train.py --expName IP-r --dataset IP  --opt none --epoch 50
python train.py --expName SMSP-r --dataset SMSP  --opt none --epoch 50

# symmetry-aware
python train.py --expName IP-rs --dataset IP  --opt opt --epoch 50
python train.py --expName SMSP-rs --dataset SMSP  --opt opt --epoch 50

Evaluation on Top-m% error

By running the following scripts, the Top-m% error of different experiments will be calculated and presented in ./$expName/prediction_error.txt, respectively.

# symmetry-agnostic
python get_logits.py --expName IP-r --dataset IP --mod valid
python accuracy.py --expName IP-r --dataset IP

python get_logits.py --expName SMSP-r --dataset SMSP --mod valid
python accuracy.py --expName SMSP-r --dataset SMSP

# symmetry-aware
python get_logits.py --expName IP-rs --dataset IP --mod valid
python accuracy.py --expName IP-rs --dataset IP

python get_logits.py --expName SMSP-rs --dataset SMSP --mod valid
python accuracy.py --expName SMSP-rs --dataset SMSP

Evaluation on relative primal gaps

Before evaluation, the prediction of the test instances should be generated by running

# symmetry-agnostic
python get_logits.py --dataset IP --expName IP-r --mod test
python get_logits.py --dataset SMSP --expName SMSP-r --mod test

# symmetry-aware
python get_logits.py --dataset IP --expName IP-rs --mod test
python get_logits.py --dataset SMSP --expName SMSP-rs --mod test

To evaluate the model, you can use the following commands:

(Note: Due to device difference, the hyperparameter --perc should be adaptively tuned.)

#-----IP-----

python eval.py --expName IP-r --dataset IP --method fixTop --perc 0.0 --maxtime 3600 --nWorker 5 # CPLEX -- this step may take several days

# symmetry-agnostic
python eval.py --expName IP-r --dataset IP --method fixTop --perc 0.1 --maxtime 1000 --nWorker 5
python eval.py --expName IP-r --dataset IP --method PS --perc 0.1 --maxtime 1000 --nWorker 5

# symmetry-aware
python eval.py --expName IP-rs --dataset IP --method fixTop --maxtime 1000 --perc 0.1 --nWorker 5
python eval.py --expName IP-rs --dataset IP --method PS --maxtime 1000 --perc 0.1 --nWorker 5

#-----SMSP-----

python eval.py --expName SMSP-r --dataset SMSP --method fixTop --perc 0.0 --maxtime 3600 --nWorker 5 # CPLEX -- this step may take several days

# symmetry-agnostic
python eval.py --expName SMSP-r --dataset SMSP --method fixTop --perc 0.5 --maxtime 1000 --nWorker 5
python eval.py --expName SMSP-r --dataset SMSP --method PS --perc 0.5 --maxtime 1000 --nWorker 5

# symmetry-aware
python eval.py --expName SMSP-rs --dataset SMSP --method fixTop --perc 0.5 --maxtime 1000 --nWorker 5
python eval.py --expName SMSP-rs --dataset SMSP --method PS --perc 0.5 --maxtime 1000 --nWorker 5

Visualization

After evaluation, we plot the figure with primal gap vs. solving time. First, convert the solving history data into matlab format by running

python get_primals.py

next, use matlab to draw the result figure by running sciprt draw_prim_curves.m. The corresponding values of relative primal gaps will be displayed in the console.

Citation

If you find this code helpful in your research, please consider citing our paper:

@article{chen2024symilo,
  title={SymILO: A Symmetry-Aware Learning Framework for Integer Linear Optimization},
  author={Chen, Qian and Zhang, Tianjian and Yang, Linxin and Han, Qingyu and Wang, Akang and Sun, Ruoyu and Luo, Xiaodong and Chang, Tsung-Hui},
  journal={NeurIPS},
  year={2024}
}