ENHANCING CLOUD RESOURCE ALLOCATION WITH VISION TRANSFORMER, DEEP REINFORCEMENT LEARNING, AND IMPROVED SHRIKE OPTIMIZATION ALGORITHM

Abstract

Cloud computing relies on efficient resource management to deliver scalable, on-demand services while minimizing costs, energy consumption, and response delays in Infrastructure as a Service (IaaS) environments. Traditional approaches often struggle with dynamic workloads, leading to resource underutilization, high migration overheads, and suboptimal Quality of Service (QoS). This paper introduces a novel hybrid framework that synergistically combines Vision Transformer (ViT) for adaptive load pattern recognition and balancing, Deep Reinforcement Learning (DRL) integrated with an Enhanced Shrike Optimization Algorithm (ESOA) for intelligent resource allocation, and Twin Fold Moth Flame Optimization (TFMFO) for cost-effective virtual machine (VM) migration. The ESOA improves upon the standard Shrike Optimization Algorithm by incorporating adaptive pheromone evaporation, Lévy flights for better global exploration, and stagnation detection to escape local optima, enabling faster convergence and superior handling of multimodal optimization problems typical in cloud scenarios. Extensive simulations using CloudSim with heterogeneous workloads (100–500 tasks) demonstrate significant performance gains: makespan reduced to 770 ms (26–36% improvement), average response time to 280 ms, energy consumption to 6160 J (30–35% savings), operational cost to $9.1, and throughput increased to 119 tasks/sec (20–25% higher) compared to state-of-the-art methods such as DRL-based placement, dual-threshold MBFD, multi-objective RL for edge, and VMP-ER. The proposed framework offers a robust, sustainable solution for modern cloud data centers, advancing efficient load balancing, resource utilization, and energy-aware operations.