RADAR: Learning to Route with Asymmetry-aware Distance Representations

March 5, 2026 · View on GitHub

Hang Yi, Ziwei Huang, Yining Ma, Zhiguang Cao

📄 Paper: RADAR

Overview

RADAR is a neural combinatorial optimization framework designed for solving asymmetric routing problems, such as the Asymmetric Traveling Salesman Problem (ATSP).

The proposed framework enhances neural routing solvers with the ability to effectively model asymmetric distance matrices. RADAR leverages Singular Value Decomposition (SVD) to initialize compact embeddings that capture static asymmetry, and introduces Sinkhorn normalization to model dynamic asymmetry during attention interactions.

Extensive experiments on both synthetic and real-world benchmarks demonstrate strong generalization ability and superior performance across various asymmetric VRPs.

Framework

Dataset and Checkpoints

The datasets and pretrained checkpoints used in our experiments can be downloaded from the following links.

Datasets

Pretrained Models

After downloading the files, please unzip them and place them into the corresponding folders.

For example:

RADAR
│
├── atsp
│   ├── dataset
│   │   └── (ATSP dataset files)
│   │
│   └── result
│       └── radar_official_checkpoint

Training and Testing

ATSP Example

Below we provide an example for training and evaluating RADAR on the ATSP task.

First, navigate to the atsp directory:

cd atsp

Training

To train the model:

python train.py

Testing

Before testing, modify the parameter problem_cnt in test.py to select the dataset size.

Supported problem sizes include:

  • problem_cnt = 100
  • problem_cnt = 200
  • problem_cnt = 500
  • problem_cnt = 1000

After setting the desired problem size, run:

python test.py

Acknowledgements

This repository is built upon MatNet. We sincerely thank the authors for making their code publicly available, which greatly facilitated the development of this project.

Citation

If you find this work or code useful in your research, please consider citing our paper:

@inproceedings{yi2026radar,
  title={RADAR: Learning to Route with Asymmetry-aware Distance Representations},
  author={Yi, Hang and Huang, Ziwei and Cao, Zhiguang and Ma, Yining},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}