Authors: Rupal Singh Thakur, Navneet Kaur
Abstract: Elastic clouds must continuously decide how much compute, memory and storage capacity to grant to workloads whose inten-sity and mix change from minute to minute, and classical threshold, queueing and forecasting controllers struggle to hold that balance without either over-provisioning or violating service level agreements. Reinforcement learning (RL) has therefore be-come the dominant learning paradigm for intelligent resource allocation, because it treats provisioning as sequential decision making under uncertainty and improves its policy directly from interaction rather than from a hand-tuned model of the data centre. This review synthesises and re-analyses past work on RL-driven cloud resource allocation published between 2006 and 2025, spanning tabular value methods, deep Q-networks, actor-critic and policy-gradient controllers, multi-agent edge-cloud schemes, and hybrid designs that combine learning with forecasting, heuristics or meta-learning. Ninety-four primary studies were screened against explicit inclusion criteria and forty-one were retained for quantitative synthesis, from which twelve re-ported effect sizes with sufficient statistical detail for a random-effects meta-analysis. Evidence is organised along three evalua-tion axes that the literature routinely conflates: accuracy, meaning the fidelity of demand estimation and SLA compliance; adapt-ability, meaning the speed of recovery after workload drift or migration to a new environment; and efficiency, meaning energy, monetary cost and computational overhead. The pooled improvement of RL controllers over non-learning baselines is 23.1 percent, but heterogeneity is substantial, and gains shrink markedly once studies are stratified by benchmark realism, baseline strength and reporting quality. The review identifies reproducibility gaps, weak baselines, simulator monoculture and neglected safety guarantees as the principal obstacles, and outlines a common evaluation protocol for future comparative work.
