Learning TSP Requires Rethinking Generalization

by   Chaitanya K. Joshi, et al.

End-to-end training of neural network solvers for combinatorial problems such as the Travelling Salesman Problem is intractable and inefficient beyond a few hundreds of nodes. While state-of-the-art Machine Learning approaches perform closely to classical solvers for trivially small sizes, they are unable to generalize the learnt policy to larger instances of practical scales. Towards leveraging transfer learning to solve large-scale TSPs, this paper identifies inductive biases, model architectures and learning algorithms that promote generalization to instances larger than those seen in training. Our controlled experiments provide the first principled investigation into such zero-shot generalization, revealing that extrapolating beyond training data requires rethinking the entire neural combinatorial optimization pipeline, from network layers and learning paradigms to evaluation protocols.


page 19

page 21

page 22


How Good Is Neural Combinatorial Optimization?

Traditional solvers for tackling combinatorial optimization (CO) problem...

On Learning Paradigms for the Travelling Salesman Problem

We explore the impact of learning paradigms on training deep neural netw...

Generalization of Machine Learning for Problem Reduction: A Case Study on Travelling Salesman Problems

Combinatorial optimization plays an important role in real-world problem...

Generalization of Neural Combinatorial Solvers Through the Lens of Adversarial Robustness

End-to-end (geometric) deep learning has seen first successes in approxi...

Combinatorial optimization and reasoning with graph neural networks

Combinatorial optimization is a well-established area in operations rese...

Evaluating Curriculum Learning Strategies in Neural Combinatorial Optimization

Neural combinatorial optimization (NCO) aims at designing problem-indepe...

Neuro-algorithmic Policies enable Fast Combinatorial Generalization

Although model-based and model-free approaches to learning the control o...

Code Repositories


Code for the paper 'An Efficient Graph Convolutional Network Technique for the Travelling Salesman Problem' (INFORMS Annual Meeting 2019)

view repo


Code for the paper 'Learning TSP Requires Rethinking Generalization' (arXiv Pre-print)

view repo


Code for the paper 'On Learning Paradigms for the Travelling Salesman Problem' (NeurIPS 2019 Graph Representation Learning Workshop)

view repo