fynnsu / AI_for_TSP

Winning solution to the AI for TSP Competition: Track 2

AI_for_TSP

This Repository contains the winning solution to the AI For TSP Competition Track 2 (https://github.com/paulorocosta/ai-for-tsp-competition).

Important Scripts

• generate_train_instances.py - creates a dataset of training instances
• Train.py - trains a base POMO model on a dataset of instances
• active_search_instance.py - performs efficient active search (EAS) on a single test instance, creating improved node embeddings
• create_tour_instance.py - uses EAS node embeddings, creates a tour for a single test instance, using MC tree searches to improve performance
• combine_tours.py - combines tours from multiple instances into the required submission format

Team Members

• Fynn Schmitt-Ulms
• André Hottung
• Kevin Tierney
• Meinolf Sellmann

Acknowledgements

Our solution is built using code from the original POMO paper by Kwon et al.
Paper: https://arxiv.org/abs/2010.16011
Code: https://github.com/yd-kwon/POMO

We extend our solution using the Efficient Active Search method described by Hottung et al.
Paper: https://arxiv.org/abs/2106.05126
Code: https://github.com/ahottung/EAS