Machine Learning for Scheduling: A Paradigm Shift from Solver-Centric to Data-Centric Approaches

TL;DR

This paper addresses the shift from solver-centric optimization to data-centric ML in scheduling. It introduces a unifying framework that separates hybrid integration and end-to-end learning, and within end-to-end, distinguishes generation methods and feasibility handling, alongside supervised, self-supervised, and RL training signals. It highlights trade-offs in data requirements, interpretability, generalization, and deployment, and outlines an agenda centered on scalability, reliability, and universality, advocating a staged path from governance-friendly hybrids to autonomous data-centric schedulers with verifiable guarantees. The work provides guidance for practitioners and researchers to build adaptive, trustworthy scheduling systems suitable for data-rich, dynamic operations.

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 toward a more data-centric approach that leverages experience and enables fast decision-making in dynamic environments. This paper offers a framework-based synthesis and perspective on this methodological transition. We use the paradigm shift from solver-centric optimization to data-centric learning as a unifying lens to organize and interpret a rapidly expanding literature. We first briefly revisit classical optimization-based approaches and discuss how machine learning has been integrated to improve computational efficiency and guide search while retaining solver-based feasibility and accountability. We then synthesize end-to-end learning approaches that generate scheduling solutions (or solution-generating policies) directly from data, clarifying the key design choices in solution generation and feasibility handling. Building on these organizing frameworks, we compare learning mechanisms and training signals (supervised, self-supervised, and reinforcement learning) in terms of scalability, interpretability, and generalization, and highlight the trade-offs that matter for reliable deployment. Finally, we outline an agenda 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.

Figures (3)

• Figure 1: Illustrations of ML-augmented decomposition approaches
• Figure 2: Illustration of end-to-end approaches: one-shot generation and policy-based RL; constructive autoregressive decoding lies in between, producing schedules sequentially without MDP interaction.
• Figure 3: Illustrations of disjunctive graph representations for a job-shop scheduling instance