Skip to content

Latest commit

 

History

History
162 lines (106 loc) · 6.55 KB

README.md

File metadata and controls

162 lines (106 loc) · 6.55 KB

[AAAI 2024] GLOP: Learning Global Partition and Local Construction for Solving Large-scale Routing Problems in Real-time

Welcome! This repository contains the code implementation of paper GLOP: Learning Global Partition and Local Construction for Solving Large-scale Routing Problems in Real-time. GLOP is a unified hierarchical framework that efficiently scales toward large-scale routing problems. It partitions large routing problems into Travelling Salesman Problems (TSPs) and TSPs into Shortest Hamiltonian Path Problems. We hybridize non-autoregressive neural heuristics for coarse-grained problem partitions and autoregressive neural heuristics for fine-grained route constructions.

diagram


🐛 Jul 2024: Thanks to Wenzheng Pan, we detected a bug in the insertion for ATSP and fixed it. After the bug fix, GLOP achieves better performance on ATSP; see the below table for the updated results. Results on other problems are unaffected.

Version ATSP150 ATSP250 ATSP1000
Before 1.89 (6.4s) 2.10 (9.6s) 2.79 (39s)
After 1.89 (8.2s) 2.04 (9.3s) 2.33 (15s)

Highlights

  • Hybridizing non-autoregressive solvers for problem partitions and autoregressive solvers for solution constructions.
  • Competitive performance across large-scale TSP, ATSP, CVRP, and PCTSP.
  • State-of-the-art scalability and efficiency: reasonable solutions for TSP100K, etc.

Dependencies


How to Use

Resources

Evaluation

To evaluate our method on your own datasets, add --path PATH_OF_YOUR_DATASET.

For TSP

# For TSP500:
python main.py --problem_size 500 --revision_iters 20 25 5 --revision_lens 100 50 20 --width 10 --eval_batch_size 64 --val_size 128 --decode_strategy greedy

# For TSP1000:
python main.py --problem_size 1000 --revision_iters 20 25 5 --revision_lens 100 50 20 --width 10 --eval_batch_size 32 --val_size 128 --decode_strategy greedy

# For TSP10k:
python main.py --problem_size 10000 --revision_iters 50 25 5 --revision_lens 100 50 20 --width 1 --eval_batch_size 16 --val_size 16 --decode_strategy greedy

# For TSP100k:
python main.py --problem_size 100000 --revision_iters 50 25 5 --revision_lens 100 50 20 --width 1 --eval_batch_size 1 --val_size 1 --decode_strategy greedy

# To conduct cross-distribution evaluation, e.g.:
python main.py --problem_size 100 --revision_lens 100 50 20 10 --revision_iters 20 10 10 5 --width 140 --eval_batch_size 100 --val_size 10000 --decode_strategy sampling --path data/tsp/tsp_uniform100_10000.pkl --no_aug --no_prune

# To reproduce the results of 49 TSPLib instances:
python eval_tsplib.py --eval_batch_size 1 --val_size 49 --path data/tsp/tsplib49.pkl --width 128 --decode_strategy greedy --no_prune

To reduce the inference duration, try:

# set
--width 1
# add
--no_aug
# less revisions, e.g.,
--revision_iters 5 5 5

For ATSP

Please refer to ./eval_atsp/

For CVRP

# For CVRP1K using LKH-3 as sub-solver: 
python eval_cvrp.py --cpus 12 --problem_size 1000

# For CVRP1K using neural sub-TSP solver
python main.py --problem_type cvrp --problem_size 1000 --revision_lens 20 --revision_iters 5

# For CVRP2K using LKH-3 as sub-solver: 
python eval_cvrp.py --cpus 12 --problem_size 2000

# For CVRP2K using neural sub-TSP solver
python main.py --problem_type cvrp --problem_size 2000 --revision_lens 50 20 --revision_iters 5 5

# For CVRP5K using LKH-3 as sub-solver
python eval_cvrp.py --cpus 12 --problem_size 5000 --ckpt_path pretrained/Partitioner/cvrp/cvrp-2000.pt

# For CVRP5K using neural sub-TSP solver
python main.py --problem_type cvrp --problem_size 5000 --ckpt_path pretrained/Partitioner/cvrp/cvrp-2000.pt --revision_lens 20 --revision_iters 5

# For CVRP7K using LKH-3 as sub-solver
python eval_cvrp.py --cpus 12 --problem_size 7000 --ckpt_path pretrained/Partitioner/cvrp/cvrp-2000.pt

# For CVRP7K using neural sub-TSP solver
python main.py --problem_type cvrp --problem_size 7000 --ckpt_path pretrained/Partitioner/cvrp/cvrp-2000.pt --revision_lens 20 --revision_iters 5

# For CVRPLIB using LKH-3 as sub-solver
python eval_cvrplib.py

# For CVRPLIB using neural sub-TSP solver
python eval_cvrplib_neural.py

For PCTSP

# e.g., for PCTSP500
python main.py --problem_type pctsp --problem_size 500 --n_subset 10 --eval_batch_size 50 --val_size 100 --revision_iters 10 10 5 --revision_lens 100 50 20

# set n_subset = 1 for greedy mode
--n_subset 1

Training

Please refer to READMEs in ./local_construction/ and ./heatmap/*/.


Citation

🤩 If you encounter any difficulty using our code, please do not hesitate to submit an issue or directly contact us!

😍 If you do find our work helpful (or if you would be so kind as to offer us some encouragement), please consider kindly giving a star, and citing our paper.

@inproceedings{ye2024glop,
  title={GLOP: Learning Global Partition and Local Construction for Solving Large-scale Routing Problems in Real-time},
  author={Ye, Haoran and Wang, Jiarui and Liang, Helan and Cao, Zhiguang and Li, Yong and Li, Fanzhang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2024},
}

Acknowledgements