Offline Decision Transformers for Neural Combinatorial Optimization: Surpassing Heuristics on the Traveling Salesman Problem

This repository implements the paper Offline Decision Transformers for Neural Combinatorial Optimization: Surpassing Heuristics on the Traveling Salesman Problem, presented at the NeurIPS 2025 Differentiable Learning of Combinatorial Algorithms Workshop.

Installation

conda env create --file environment.yml

Data preparation

Place all datasets in the data/ directory.

1. Download or generate the TSP instance data from learning-paradigms-for-tsp.
2. Then, generate the corresponding solution data.

Training

Train a model on Farthest Insertion (FI) for N=20:

python train.py \
    --train_dataset=data/tsp20_train_fi.txt \
    --val_dataset=data/tsp20_val_fi.txt \
    --log_dir=dt_runs/dt_n20_fi \
    --num_epochs=2000 \
    --save_iters=10

Evaluation

Evaluate the trained model:

python eval.py \
    --test_dataset=data/tsp20_test_fi.txt \
    --model_dir=dt_runs/dt_n20_fi_[%y%m%d%H%M%S] \
    --chk_pt_name="best.pt" \
    --eval_optimal_gap \
    --optimal_dataset=data/tsp20_test_concorde.txt

Acknowledgement

This implementation is based on Elastic-DT and learning-paradigms-for-tsp.

License

This project is licensed under the MIT License - see the LICENSE file for details. For details on the licenses of third-party dependencies used in this project, please refer to the NOTICE.md file.