 |
What do transformers learn from solving routing problems ?
1 : SUTD - MIT International Design Centre
2 : Université dÁrtois
Paris School of Business
3 : Predictim Globe Ltd
Predictim Globe
4 : Université de Lille
Research Centre in Computer Science, Signal and Automatic Control of Lille (CRIStAL UMR 9189)
Despite promising improvement demonstrated by putting deep neural networks to work in solving combinatorial optimization problems, fundamental questions about the nature of knowledge these networks acquire in this course and how well they extend to instances beyond their training examples remains unanswered, In this paper we adapt explainable machine learning tools to explain transformers, a recent prominent neural architecture, applied to traveling salesman problem.
