Adaptive machine learning for intelligent resource allocation in cloud computing: a comparative study

Abstract

Efficient cloud resource allocation requires controllers that anticipate workload variation rather than respond only after demand changes. This study evaluates an adaptive machine learning (AML) controller that integrates one-step-ahead random forest forecasting, capacity-constrained resource adjustment, and 24-hour model retraining. A 30-day CloudSim experiment modeled 50 hosts and 200 virtual machines (VM) at five-minute intervals and compared AML with fixed and reactive dynamic allocation using resource utilization, simulated cost, and response time. Across 20 independent AML runs, mean resource utilization was 85.2%±1.8% (95% confidence interval: 84.4–86.0%), and mean high-load response time was 54.8±3.2 ms. In the descriptive baseline summaries, AML achieved approximately 85% utilization, compared with 75% for dynamic allocation and 60% for fixed allocation, while the corresponding high-load response times were approximately 55, 70, and 100 ms. AML also reduced the rounded simulated cost by 10% relative to the dynamic baseline and 40% relative to fixed allocation. Together, the results indicate that converting near-term forecasts into capacity-constrained actions can improve the utilization–latency–cost balance within the evaluated simulation. The study supports prediction-guided allocation as a practical basis for intelligent cloud resource decisions under controlled workload conditions.