Multi-Task Learning for Routing Problem with Zero-Shot Generalization
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Multi-task learning, Vehicle routing problems, Zero-shot generalization, Combinatorial optimization, Neural combinatorial optimization
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Vehicle routing problems (VRPs) are widely studied due to their significant practical importance. In the last decade, leveraging neural networks to solve VRPs in an end-to-end manner has gained substantial research attention. However, current works require building separate neural models for each routing problem, which hinders its practicality in solving diverse problems. In this study, we treat the VRPs as different combinations of a set of shared underlying attributes and propose to solve them simultaneously as multi-task learning. By training a unified model on multiple VRPs with varying attributes, we can effectively solve unseen problems in a zero-shot manner. Our experimental results on eleven VRPs show that our unified model performs comparably to single-task models trained specifically for each problem. More importantly, our model exhibits promising zero-shot generalization to new VRPs, reducing the average gap to 4.6\% and 7.0\% for sizes 50 and 100, respectively, compared to over 20\% in the single-task approach.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7171
Loading