Generalizable Deep RL-Based TSP Solver via Approximate Invariance

Han Fang, Zhihao Song, Paul Weng

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: Travelling Salesman Problem, Transformer, Invariance, Augmentation, Monte-Carlo Tree Search

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Recently, deep reinforcement learning (DRL) has shown promising results for learning fast heuristics to solve traveling salesman problems (TSP). Meanwhile, most existing state-of-the-art (SOTA) DRL methods yield solvers that do not generalize well on TSP instances larger than those seen during training. However, such generalization ability is crucial in practice since training on large instances is impractical. To tackle this issue, we propose a novel DRL method, called TS$^3$, which is designed to enforce a variety of (possibly approximate) invariances to promote the generalizability of the learned solver. More specifically, TS$^3$ applies a modified policy gradient algorithm enhanced with data augmentation to train a Transformer-based model to select the next city to visit among the k-nearest neighbors of the last visited city by integrating a local view and global view of a TSP instance. To further validate the capability of TS$^3$, we also propose its combination with Monte-Carlo Tree Search. Abundant experiments on random TSP and TSPLIB instances demonstrate that our propositions achieve a dominant performance when generalizing to large-sized TSPs.

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: 1815