CluAttention (PDTSP)

experiment.py is the central entry point. The four wrapper scripts fix the ablation setting:

• run_full.py: full model
• run_no_enc_cluster.py: encoder cluster attention disabled
• run_no_dec_cluster.py: decoder cluster attention disabled
• run_pomo.py: POMO baseline (cluster attention disabled in both encoder and decoder)

These wrappers only toggle the cluster-attention modules. The rest of the model structure, training setup, and evaluation options stay the same.

Requirements

• Python 3.9+
• PyTorch 1.12+ (install the CUDA build that matches your driver if needed)

Install the Python dependencies:

pip install -r requirements.txt

Example PyTorch installation for CUDA 11.8:

pip install torch --index-url https://download.pytorch.org/whl/cu118

Data Format

Evaluation data is stored as .pkl files. Two formats are supported:

• List format: [(depot_xy, node_xy), ...]
• Dict format: {'data': [(depot_xy, node_xy), ...], 'scale_factor': float}

Where:

• depot_xy is the depot coordinate (x, y)
• node_xy is the list of customer node coordinates

For .pkl instances, the code assumes the first half of node_xy are pickup nodes and the second half are their paired delivery nodes.

Evaluation expects each .pkl file to contain a single instance by default.

Generate a Test Set

Example: 100-node instances, 100 files, clustered distribution.

python generate_pdp_dataset.py --name test --problem pdp --data_distribution clustered --cluster_std 0.1 --graph_sizes 100 --dataset_size 1 --num_files 100 --seed 10000 --data_dir data

Example output path:

data/pdp/pdp100_test_clustered_std0.1_seed10000.pkl

Training

Full model:

python run_full.py --task train --problem_size 100 --distribution clustered --train_batch_size 128

Encoder cluster attention disabled:

python run_no_enc_cluster.py --task train --problem_size 100 --distribution clustered --train_batch_size 128

Decoder cluster attention disabled:

python run_no_dec_cluster.py --task train --problem_size 100 --distribution clustered --train_batch_size 128

POMO baseline:

python run_pomo.py --task train --problem_size 100 --distribution clustered --train_batch_size 128

Disable the learned gate:

python run_full.py --task train --problem_size 100 --distribution clustered --disable_gate

Evaluation

Greedy decoding:

python run_full.py --task test "data/pdp/pdp100_test_clustered_std0.1_seed*.pkl" --model_path /path/to/result_dir --epoch 800 --decode_strategy greedy

Sampling with width 1280:

python run_full.py --task test "data/pdp/pdp100_test_clustered_std0.1_seed*.pkl" --model_path /path/to/result_dir --epoch 800 --decode_strategy sampling --width 1280

Sampling with width 12800:

python run_full.py --task test "data/pdp/pdp100_test_clustered_std0.1_seed*.pkl" --model_path /path/to/result_dir --epoch 800 --decode_strategy sampling --width 12800

To evaluate another ablation, replace run_full.py with one of:

• run_no_enc_cluster.py
• run_no_dec_cluster.py
• run_pomo.py

Project Structure

• experiment.py: shared training and evaluation entry point
• run_full.py: full-model wrapper
• run_no_enc_cluster.py: wrapper without encoder cluster attention
• run_no_dec_cluster.py: wrapper without decoder cluster attention
• run_pomo.py: POMO baseline wrapper
• CVRPModel_training.py: model definition
• CVRPEnv.py: environment and state transitions
• CVRPTrainer.py: training loop
• CVRPTester.py: checkpoint loading and evaluation
• generate_pdp_dataset.py: test-set generator
• utils.py: logging and utility helpers
• data/: dataset directory (not version-controlled)

References

Cluster-Aware Attention-Based Deep Reinforcement Learning for Pickup and Delivery Problems Wentao Wang, Lifeng Han, and Guangyu Zou. arXiv preprint arXiv:2603.10053, 2026. [Paper]

BibTeX (Click to expand)

@misc{wang2026clusteraware,
      title={Cluster-Aware Attention-Based Deep Reinforcement Learning for Pickup and Delivery Problems},
      author={Wentao Wang and Lifeng Han and Guangyu Zou},
      year={2026},
      eprint={2603.10053},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={[https://arxiv.org/abs/2603.10053](https://arxiv.org/abs/2603.10053)},
}