Implementation of the Policy Gradient algorithm for learning 2-opt improvement heuristics, following: http://proceedings.mlr.press/v129/costa20a/costa20a.pdf

https://arxiv.org/abs/2004.01608

Dependencies:

• Python 3.6.4
• Torch
• Numpy
• Matplotlib
• Apex
• tqdm
• pyconcorde

To train the model you can run:

For TSP instances with 20 nodes:

python PGTSP20.py

For TSP instances with 50/100 nodes (default 50 nodes):

python PGTSP50_100.py

To use the learned polcies reported in the paper you can run:

python TestLearnedAgent.py --load_path best_policy/policy-TSP20-epoch-189.pt --n_points 20 --test_size 1 --render 

where load_path can be replaced with one of the policies in /best_policy.

Learned policy on a TSP with 50 nodes: