This repository contains code for deep reinforcement learning to solve the Traveling Salesman Problem with Drone (TSPD). For details, please see our paper A Deep Reinforcement Learning Approach for Solving the Traveling Salesman Problem with Drone. If this code is useful for your work, please cite our paper:

@article{bogyrbayeva2022deep,
      title={A Deep Reinforcement Learning Approach for Solving the Traveling Salesman Problem with Drone}, 
      author={Aigerim Bogyrbayeva and Taehyun Yoon and Hanbum Ko and Sungbin Lim and Hyokun Yun and Changhyun Kwon},
      year={2022},
      journal={Transportation Research Part C: Emerging Technologies},
      volume={To Appear}
}

For the optimization heuristic algorithms used in the paper, please see TSPDrone.jl.

• Python>=3.8
• NumPy
• SciPy
• PyTorch>=1.7

Training data is generated on the fly with the batch size and node numbers specified in /utils/options.py. If test data is not given in the data folder, the test data will be generated randomly as well.

For training TSPD, just run the following line. Any other training parameters can also be set in /utils/options.py such as the number of nodes, batch size, the number of epochs, decode lengths, etc.

python main.py --train=True

The trained weight files will be saved in the /trained_models directory.

Pre-trained weights files for random data as described in the paper are located in the /trained_models directory for some sizes, n = 11, 15, 20, 50, 100.

To perform only inference, please set train to False in /utils/options.py or just run:

python main.py --train=False

By default, the greedy decoding will run.

To run batch sampling, please set sampling to True and specify the number of samples n_samples in /utils/options.py.

The results of both greedy and batch sampling decoding will be stored in the results folder.

The /data directory includes random test instances used in the paper for n=11, 15, 20, 50, 100. Each file includes 100 instances.

Each row represents an instance, in the form of

x_1 y_1 d_1 x_2 y_2 d_2 ... x_n y_n d_n

where x_i y_i d_i represents the x-y coordinate of customer i and demand. All demands are set to be 1.0 for customers. The last components x_n y_n d_n represents the depot and d_n is set to 0.0 for the depot.

A sample solution of TSPD for 11 nodes is depicted below:

This repository heavily benefited from the following repositories:

• https://github.com/wouterkool/attention-learn-to-route
• https://github.com/OptMLGroup/VRP-RL