Hybrid Attention Vision Transformer Framework for Cost Efficient Cloud Resource Allocation and Load Balancing

Abstract

This paper aims to develop FairViT-ShrikeNet , which is a hybrid system that combines attention motivated deep reinforcement learning, Vision Transformer workload encoding, and Shrike metaheuristic optimization to realise fairness sensitive and cost effective resource allocation in cloud computing systems. The approach includes the use of decentralised DRL agents with attention models, the use of Vision Transformers to code distributions of workloads into redistribution maps, and the optimization of allocation plans with Shrike in a simulated cloud control environment on Python, PyTorch, and CloudSim. Findings show that the proposed model reduced energy consumption to 168.54 J , execution time to 20.83 ms , and waiting time to 9.62 ms , while increasing throughput to 75.26 units/s , fairness index to 0.923 , and lowering migration cost to 8.95 units , outperforming RL-based, RMF-DNN, and F-RL baselines across all metrics. This work is novel because it integrates adaptive learning and transformer based workload encoding with metaheuristic refinement into a consistent architecture and provides an all-encompassing solution that improves fairness, scalability, and efficiency in dynamic cloud infrastructures.