A Federated Reinforcement Learning Framework for Multi-Tenant Resource Allocation in Edge Computing Environments

Ultra-dense edge computing environments pose significant challenges for maintaining Service-Level Agreement (SLA) guarantees due to dynamic workloads, heterogeneous tenant requirements, and shared resource contention. While centralized reinforcement learning can adapt to changing conditions, it often suffers from scalability and privacy limitations, whereas fully independent learning strategies fail to coordinate resource usage across tenants sharing the same infrastructure. In this work, we propose a modular Federated Reinforcement Learning (FRL) framework for multi-tenant resource allocation in edge systems. The architecture operates as a closed-loop control pipeline consisting of real-time telemetry collection, graph-based state construction to capture tenant–node contention, federated policy learning using Proximal Policy Optimization (PPO) at each tenant with asynchronous aggregation, and an execution layer that enforces resource adjustments. To align learning with SLA objectives, we design an SLA-aware reward that balances latency violations, resource efficiency, and action stability. Extensive simulations in a large-scale ultra-dense edge scenario show that the proposed framework consistently reduces SLA violations, improves resource utilization balance, stabilizes control actions, and achieves higher global rewards compared to centralized, independent, and static allocation baselines.