Code for ICLR2022 Paper: Pareto Set Learning for Neural Multi-objective Combinatorial Optimization

It contains the training and testing codes for three multi-objective combinatorial optimization (MOCO) problems:

• Multi-Objective Travelling Salesman Problem (MOTSP)
• Multi-Objective Capacitated Vehicle Routing Problem (MOCVRP)
• Multi-Objective Knapsack Problem (MOKP)

This code is heavily based on the POMO repository, and it has been reorganized accroding to the new POMO version. The main changes include:

• Graph embedding has been removed.
• BatchNorm has been replaced by InstanceNorm.

Quick Start

• To train a model, such as MOTSP with 20 nodes, run train_motsp_n20.py in the corresponding folder.
• To test a model, such as MOTSP with 20 nodes, run test_motsp_n20.py in the corresponding folder.
• Pretrained models for each problem can be found in the result folder.

Reference

If our work is helpful for your research, please cite our paper:

@inproceedings{lin2022pareto,
  title={Pareto Set Learning for Neural Multi-Objective Combinatorial Optimization},
  author={Xi Lin, Zhiyuan Yang, Qingfu Zhang},
  booktitle={International Conference on Learning Representations},
  year={2022},
  url={https://openreview.net/forum?id=QuObT9BTWo}
}

If you find our code useful, please also consider citing the POMO paper:

@inproceedings{Kwon2020pomo,
  title = {POMO: Policy Optimization with Multiple Optima for Reinforcement Learning},
  author = {Yeong-Dae Kwon, Jinho Choo, Byoungjip Kim, Iljoo Yoon, Youngjune Gwon, Seungjai Min},
  booktitle = {Advances in Neural Information Processing Systems},
  pages = {21188--21198},
  volume = {33},
  year = {2020}
}