.png){kind=icon}
.png){kind=icon}
Return
Machine Learning-Driven Combinatorial Optimization: A Systematic Review
.png){kind=icon}
PRE
AI
DOI：10.1007/s11831-026-10679-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This survey focuses on routing and scheduling, two cornerstone Combinatorial optimization problems (COPs) domains in logistics and manufacturing that are increasingly required to operate under large-scale and dynamic constraints. Traditional methods often struggling with scalability and dynamic constraints. The integration of machine learning (ML) has introduced powerful new solvers. This survey comprehensively reviews ML advances for COPs from 2020 to 2026, focusing on four key paradigms: supervised learning, reinforcement learning, unsupervised learning, and large language models (LLMs). We categorize methods into three frameworks: end-to-end neural solvers that directly generate solutions, hybrid solvers that use ML to guide classical optimization algorithms, and generative approaches that model solution distributions. Each paradigm is analyzed in terms of model architectures, integration strategies, and applications to routing, scheduling, and other COPs. We also examine the emerging role of LLMs as optimizers, programmers, and collaborative tools. To support rigorous cross-paradigm comparison, we establish a unified comparative framework that evaluates methods along seven dimensions—solution quality, scalability, generalization, computational efficiency, data independence, feasibility assurance, and theoretical guarantees—and provide coarse ratings for representative approaches across all paradigms, including LLM-based methods. Building on this analysis, the paper identifies core challenges such as generalization across scales and distributions, the efficiency–quality trade-off, theoretical guarantees, data dependency, dynamic constraint handling, feasibility assurance for generative models, and robustness under non-stationary environments, and outlines future directions for building more robust and practical ML-driven optimization systems.
Journal
IF:
12.1
Papers:
1.8K
Citations:
1.2W
View more organizations