ESBO-FDDPG: A Federated Reinforcement Learning-Based Continuous Resource Scheduler for Cross-Domain Multi-Cloud Systems

Abstract

With the rapid development of cloud computing, the demand for efficient resource management across multiple data centers has become increasingly critical. Cloud platforms typically span geographically distributed data centers, offering scalability, flexibility, and on-demand resource provisioning. This research proposes a unified cross-domain continuous resource scheduling method for various data centers within the cloud platforms, aiming to tackle challenges such as uneven workload distribution, dynamic service demands, and energy inefficiency across multiple data centers. This research introduces a unified cross-domain continuous resources scheduling framework based on an Efficient Stain Bowerbird Optimizer-driven Federated Deep Deterministic Policy Gradient (ESBO-FDDPG) model. Workload traces from the Cross-Domain Cloud Scheduling Dataset were preprocessed using Min-Max normalization, and Statistical Pattern Recognition (SPR) with Fast Fourier Transform (FFT) and Radial Inverse Distance (RID) kernels was applied for feature extraction. The proposed method enables privacy-preserving, real-time, and adaptive scheduling by integrating federated learning with deep reinforcement learning. The proposed method supports continuous monitoring and coordination of workloads across multiple geographically distributed data centers, dynamically adjusting resource allocation based on service request patterns and power usage profiles. The system embeds vertical federated learning for cross-domain model updates without data sharing and uses reinforcement learning to continuously adapt to fluctuating workloads and infrastructure constraints. Simulation experiments were conducted using Python. Simulations conducted with different load and network conditions show that the suggested ESBO-FDDPG approach lowers average latency (63ms) and enhances energy efficiency (4.6 ×〖10〗^5bits/joules), and task completion rate (98.8%). This research offers important insights for sustainable cloud computing practices and a scalable, energy-conscious solution to unified resource scheduling in contemporary cloud systems.