Adaptive Cloud Resource Allocation Using Attention-Driven Deep Reinforcement Learning

Abstract

Efficient resource allocation remains a critical challenge in dynamic Cloud Computing (CC) environments, where maintaining Quality of Service (QoS), minimizing latency, and ensuring fairness are paramount. This study proposes a novel Deep Reinforcement Learning (DRL)-based framework that models resource allocation as a multi-agent Markov Decision Process (MDP), with each Virtual Machine (VM) link treated as an autonomous agent. Leveraging a Deep Q-Network (DQN) architecture enhanced by an attention mechanism, the framework enables agents to refine state observations and coordinate decisions adaptively. A custom reward function balancing throughput, latency, and resource cost guides the learning process, whereas experience replay and temporal annealing strategies promote optimal policy convergence. Experimental results demonstrate significant improvements in energy efficiency, execution time, waiting time, fairness, and throughput when benchmarked against existing Reinforcement Learning (RL)-based, Resource Management Framework–Deep Neural Network (RMF-DNN), and Federated Reinforcement Learning (F-RL) models. The proposed system introduces architectural innovations, including decentralized agent-based learning, attention-guided state refinement, and fairness-aware scheduling, establishing a scalable and intelligent solution for cloud resource management.