Authors: Nimal Perera
Abstract: As cloud computing infrastructures transition from passive resource providers to "Intelligent Clouds," the complexity of managing heterogeneous, bursty, and globally distributed workloads has rendered traditional heuristic scheduling insufficient. This review examines the paradigm shift toward machine learning-driven optimization for cloud workload scheduling. We analyze the evolution from static rule-based systems to autonomous, data-driven frameworks that leverage reinforcement learning, deep neural networks, and multi-agent systems. The article categorizes contemporary ML-based scheduling techniques, evaluates their performance against multi-objective criteria—such as energy efficiency, Service Level Agreement (SLA) compliance, and cost—and identifies critical bottlenecks like model drift and interpretability. By synthesizing recent breakthroughs in 2025 and 2026, including the rise of "Agentic AI" in resource orchestration and federated learning for privacy-preserving scheduling, this review provides a roadmap for researchers and practitioners aiming to navigate the complexities of next-generation autonomous cloud environments.
DOI: https://doi.org/10.5281/zenodo.19437787
