Solving Dynamic Traveling Salesman Problems With Deep Reinforcement Learning

This code solves dynamic trvaeling saleman problem with deep reinforcement learning. For more details, please see our paper Solving Dynamic Traveling Salesman Problems With Deep Reinforcement Learning which has been accepted at IEEE-TNNLS. If this code is useful for your work, please cite our paper:

@article{zhang2021solving,
  title={Solving Dynamic Traveling Salesman Problems With Deep Reinforcement Learning},
  author={Zhang, Zizhen and Liu, Hong and Zhou, MengChu and Wang, Jiahai},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2021},
  publisher={IEEE}
}

Dependencies

• python = 3.6.3
• NumPy
• Scipy
• PyTorch = 1.7
• tensorboard_logger

Quick Start

For training DTSP instances with 19 customers and using rollout as REINFORCE baseline with model M1:

python m1/train.py --baseline rollout --graph_size 19

For training DTSP instances with 19 customers and using rollout as REINFORCE baseline with model M2:

python m2/train.py --baseline rollout --graph_size 19

For testing DTSP instances with 19 customers with a trained model M1:

python m1/test.py --baseline rollout --graph_size 19 --resume trained_models/m1/normal_19.pt

For testing DTSP instances with 19 customers with a trained model M2:

python m2/test.py --baseline rollout --graph_size 19 --resume trained_models/m2/normal_19.pt

Acknowledgements

Our code is adpated from https://github.com/wouterkool/attention-learn-to-route.