LLM4Solver: Large Language Model for Efficient Algorithm Design of Combinatorial Optimization Solver

ICLR 2025 Conference Submission7319 Authors

26 Sept 2024 (modified: 13 Oct 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Combinatorial Optimization Solver, Large Language Models, Evolutionary Search
Abstract: The optimization of algorithms in exact combinatorial optimization (CO) solver plays a fundamental role in operations research. However, due to the extensive requirements on domain knowledge and the large search space for algorithm design, the refinement on these algorithms remains highly challenging for both manual and learning-based paradigms. To tackle this problem, we propose a novel machine learning framework---large language model for exact combinatorial optimization solver (LLM4Solver)---to $\textit{efficiently}$ design high-quality algorithms of the CO solvers. The core idea is that, instead of searching in the high-dimensional and discrete symbolic space from scratch, we can utilize the prior knowledge learned from large language models to directly search in the space of programming languages. Specifically, we first use a pre-trained LLM as the generator for high-quality algorithms. Then, to efficiently explore the discrete and non-gradient algorithm space, we employ a derivative-free evolutionary framework as the algorithm optimizer. Experiments on extensive benchmarks show that the algorithms learned by LLM4Solver $\textit{significantly}$ outperform all the state-of-the-art (SOTA) human-designed and learning-based policies (on GPU) in terms of the solution quality, the solving efficiency, and the cross-benchmark generalization ability. The appealing features of LLM4Solver include 1) the high training efficiency to outperform SOTA methods within ten iterations, and 2) the high cross-benchmark generalization ability on heterogeneous MIPLIB 2017. LLM4Solver shows the encouraging potential to efficiently design algorithms for the next generation of modern CO solvers.
Primary Area: optimization
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: 7319
Loading