A Survey of Machine-Learning-Based Scheduling: From Solver-Centric to Data-Centric Paradigms

Abstract: Scheduling problems are a fundamental class of combinatorial optimization problems that underpin operational efficiency in manufacturing, logistics, and service systems. While operations research has traditionally developed solver-centric methods emphasizing model structure and optimality, recent advances in machine learning are reshaping scheduling into a data-centric discipline that learns from experience and adapts to dynamic environments. This paper provides a comprehensive and comparative review of this methodological transition. We first revisit classical optimization-based approaches and summarize how ML has been integrated within them to improve computational efficiency. We then review end-to-end learning approaches that generate scheduling solutions directly from data, highlighting how they shift decision-making from explicit optimization to learned inference. Adopting a systematic, method-oriented perspective, we compare these paradigms and their underlying learning algorithms in terms of principles, scalability, interpretability, and generalization. Finally, we discuss key research challenges and outline future directions along three interdependent dimensions, scalability, reliability, and universality, that together define a pathway toward adaptive, intelligent, and trustworthy scheduling systems for data-driven operations management.