Dynamic Cloud Resource Optimization Using Reinforcement Learning
Abstract
Cloud computing offers flexible and scalable access to computational resources, but selecting a suitable resource configuration for a given workload remains a difficult task. High-Performance Computing (HPC) and Artificial Intelligence (AI) workloads have varying computational requirements, and their performance depends on complex interactions between workload characteristics and the available cloud configurations. Traditional resource recommendation methods generally rely on predefined rules, repeated test executions, or large volumes of historical data, which are poorly suited to new or changing applications. This paper presents a study of dynamic cloud resource optimization based on reinforcement learning, using the Oikonomos-II+ framework as its primary technical reference. In this approach, resource selection is treated as a contextual multi-armed bandit problem in which workload characteristics form the context and the available cloud configurations form the possible actions. A Variational Autoencoder (VAE) produces a compact representation of workload characteristics, Neural-LinUCB selects configurations while balancing exploration and exploitation, and Gaussian Process (GP) regression estimates expected performance together with its uncertainty. RHO-LOSS supports the evaluation of resource-selection outcomes. After a configuration is selected, the workload is executed, its performance is observed, and the resulting feedback is used to update the learning models and improve future recommendations. The paper reviews traditional, machine-learning-based and reinforcement-learning-based allocation approaches, explains the methodology and working cycle of the framework, and discusses its applications, advantages, challenges, limitations and future scope. Overall, the study concludes that learning-based, workload-aware recommendation is a promising route to better resource utilization and suitable performance for computationally intensive workloads, although computational overhead, data quality, implementation complexity and scalability require further evaluation.