Real Routing NCO

Note: we are currently finalizing the repository. Stay tuned!

🗺️ Problem

Most NCO methods use simplified routing with 2D Euclidean distance. This is not realistic for real-world applications which can have complex 1) distance matrices and 2) duration matrices between locations because of road networks, traffic, and more.

Left: previous works with simplified routing. Right: RRNCO with real-world routing!

How can we bridge this gap between toy and real settings?

We need two things:

1. A dataset with real-world routing information
2. A model that can handle such data -- not only node but also edge information

✨ Solution 1: RRNCO Dataset

We introduce the RRNCO (Real Routing NCO) dataset, which contains real-world routing information for 100 cities around the world, from which instances can be subsampled and generated on the fly

RRNCO data generation pipeline

✨ Solution 2: The RRNCO Model

The RRNCO model efficiently processes topology information by leveraging several techniques including scale adaptive biases

RRNCO model architecture

🚀 Installation

We use uv (extremely fast Python package manager) to manage the dependencies:

uv venv --python 3.12 # create a new virtual environment
source .venv/bin/activate # activate the virtual environment
uv sync --all-extras # for all dependencies

Note that this project is also compatible with normal pip install -e . in case you use a different package manager.

Data download

Download data and checkpoints To download the data and checkpoints from HuggingFace automatically, you can use:

python scripts/download_hf.py

You may use the option "--no-data" to skip the dataset or "--no-models" to skip the checkpoints.

Data generation

Tip

This is already done when you download the dataset from HuggingFace. You can skip this step if you only want to use the pretrained models.

Dataset generation

Instructions on how to install the OSRM backend and generate (new) datasets data_generation folder.

After generating city data using the data generation pipeline, move the generated files to the following directory:

data/dataset/{city}/{city}_data.npz

For example, if the city is Seoul, the data file should be located at:

data/dataset/Seoul/Seoul_data.npz

Additionally, the file data/dataset/splited_cities_list.json contains a predefined split of cities into training and test sets. If you wish to modify the training cities, simply edit the list under the "train" key in this JSON file.

Test instances generation

To (re)generate the test dataset (included in HuggingFace already):

python scripts/generate_data.py

This will generate test instance from the dataset provided above.

How to run

To get started with running RRNCO, please follow the steps below:

Training

To train a model, use the train.py script. For example, to train a model for the ATSP problem:

python train.py experiment=rrnet env=atsp

Available environment options are:

• atsp (Asymmetric TSP)

• rcvrp (Real-world Capacitated VRP(ACVRP))

• rcvrptw (Real-world Capacitated VRP with Time Windows(ACVRPTW))

You can also configure experiment settings using the file config/experiment/rrnet.yaml.

Model Testing

You can evaluate a trained model using the test.py script. Make sure to provide the correct dataset path via --datasets and model checkpoint via --checkpoint.

Examples for different tasks:

ATSP

python test.py --problem atsp --datasets data/atsp/atsp_n100_seed3333_in_distribution.npz --batch_size 32 --checkpoint checkpoints/atsp/epoch_199.ckpt

RCVRP

python test.py --problem rcvrp --datasets data/rcvrp/rcvrp_n100_seed3333_in_distribution.npz --batch_size 32 --checkpoint checkpoints/rcvrp/epoch_199.ckpt

RCVRPTW

python test.py --problem rcvrptw --datasets data/rcvrptw/rcvrptw_n100_seed3333_in_distribution.npz --batch_size 32 --checkpoint checkpoints/rcvrptw/epoch_199.ckpt

🤩 Citation

If you find RRNCO valuable for your research or applied projects:

@article{son2025rrnco_neuralcombinatorialoptimizationrealworldrouting,
      title={{Neural Combinatorial Optimization for Real-World Routing}},
      author={Jiwoo Son and Zhikai Zhao and Federico Berto and Chuanbo Hua and Changhyun Kwon and Jinkyoo Park},
      year={2025},
      eprint={2503.16159},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://github.com/ai4co/real-routing-nco},
}