Massively Parallel Environments for Large-Scale Combinatorial Optimizations Using Reinforcement Learning

Ming Zhu, Xiao-Yang Liu

Keywords: Combinatorial Optimizations, Massively Parallel Environments, Reinforcement learning, distribution-wise approach

Abstract: Most combinatorial optimization (CO) problems are NP-hard and difficult to find high-quality solutions. Reinforcement learning (RL) is a promising technique due to its powerful search capability; however, sampling speed is a common bottleneck. Current benchmark works only provide instance-wise approaches, while our work cover both instance-wise and distribution-wise approaches, especially in large-scale CO problems. In this paper, we build 24 GPU-based massively parallel environments for 12 CO problems, i.e., each problem has two environments; and use them to train RL-based approaches. We reproduce benchmark RL algorithms, including instance-wise and distribution-wise approaches especially in large-scale CO problems, on both synthetic datasets and real-world datasets. Take the graph maxcut problem as an example. The sampling speed is improved by at least two orders over conventional implementations, and the scale (i.e., number of nodes) of trained problems in a distribution-wise approach is up to thousands of nodes, i.e., improved by one order. The objective value obtained by inference (100 $\sim$ 200 seconds) in the distribution-wise scenario is almost the same as the state-of-the-art (SOTA) solver Gurobi (running for 1 hour), and better than the SOTA RL-based approach. The code is available at: https://github.com/OpenAfterReview.

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 381