SteBen: Steiner Tree Problem Benchmark for Neural Combinatorial Optimization on Graphs
Keywords: Dataset & benchmark, Neural Combinatorial Optimization, Steiner Tree Problem
TL;DR: We present the first large-scale STP benchmark for NCO methods, providing extensive datasets and across various NCO techniques.
Abstract: The Steiner Tree Problem (STP) is an NP-hard combinatorial optimization problem with applications in areas like network design and facility location. Despite its importance, learning-based solvers for STP have been hindered by the lack of large-scale, diverse datasets necessary to train and evaluate advanced neural models. To address this limitation, we introduce a standardized dataset comprising over a million high-quality instances with optimal solutions, spanning various problem sizes and graph structures. Our dataset enables benchmarking of neural combinatorial optimization methods across both supervised and reinforcement learning paradigms, encompassing autoregressive and non-autoregressive inference approaches. Our experiments show that supervised learning excels in in-distribution settings, while reinforcement learning generalizes better to unseen problem sizes, highlighting a trade-off between solution quality and generalization. We compare NCO methods across different STP scales and graph types, and demonstrate that solvers trained on our datasets generalize well to real-world instances without fine-tuning, proving its practical utility. We hope this benchmark promotes further STP research and advances NCO techniques for broader combinatorial optimization challenges.
Primary Area: datasets and benchmarks
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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: 5357
Loading