Collaboration! Towards Robust Neural Methods for Vehicle Routing Problems
Primary Area: general machine learning (i.e., none of the above)
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: Learning to Optimize, Vehicle Routing Problem, Combinatorial Optimization
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: While enjoying desirable efficiency and less dependence on domain expertise, existing neural methods for vehicle routing problems (VRPs) are vulnerable to adversarial attacks -- their performance drops drastically on adversarial instances, i.e., clean instances with crafted perturbations. To enhance the robustness, we propose a Collaborative Neural Framework (CNF) w.r.t the adversarial defense of neural methods for VRPs, which is crucial yet underexplored in literature. Given a neural method, we adversarially train multiple models in a collaborative manner to synergistically promote the robustness against attacks, while maintaining (or even boosting) the standard generalization on clean instances. A neural router is designed to elegantly distribute instances to each model, which improves load balancing and collaborative performance. Extensive experiments verify the effectiveness and versatility of CNF to defend against various attacks for different neural methods. Notably, our trained models also achieve decent out-of-distribution generalization performance on real-world benchmark instances.
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: 167
Loading