Adaptive AI-based Decentralized Resource Management in the Cloud-Edge Continuum

In the Cloud-Edge Continuum, dynamic infrastructure change and variable workloads complicate efficient resource management. Centralized methods can struggle to adapt, whilst purely decentralized policies lack global oversight. This paper proposes a hybrid framework using Graph Neural Network (GNN) embeddings and collaborative multi-agent reinforcement learning (MARL). Local agents handle neighbourhood-level decisions, and a global orchestrator coordinates system-wide. This work contributes to decentralized application placement strategies with centralized oversight, GNN integration and collaborative MARL for efficient, adaptive and scalable resource management.

💡 Research Summary

The paper addresses the challenging problem of resource management in the Cloud‑Edge Continuum, where dynamic infrastructure changes and highly variable workloads make centralized control difficult to scale and purely decentralized policies lack a global perspective. To bridge this gap, the authors propose a hybrid framework that combines Graph Neural Network (GNN) embeddings with collaborative Multi‑Agent Reinforcement Learning (MARL). The system is organized hierarchically: local agents operate on neighbourhood‑level subgraphs, making placement decisions for application components, while a global orchestrator aggregates abstracted information from all locals and coordinates system‑wide objectives.

The authors first formalize the problem with four models: a resource graph (computational, storage, and network nodes), an application graph (micro‑services and their communication constraints), a cost model (completion time, resource utilization, SLA violation rate), and an objective function that maximizes utilization while minimizing latency and violations. GNNs are employed during the “Analysis” phase of the classic MAPE autonomic loop to generate compact embeddings that capture both node attributes and topological dependencies. Crucially, the GNN uses a masking technique to handle dynamic addition or removal of nodes without retraining the entire model, reducing operational overhead.

Local agents receive partial observations consisting of the full application graph and a local resource subgraph. Their policies are trained with MARL (e.g., PPO or A3C) using a reward that balances utilization, inverse completion time, and inverse SLA violation, weighted by parameters α, β, γ. The global agent receives an aggregated view of all local embeddings and learns a higher‑level policy that decides which local agents to activate and how to align their actions with a global reward composed of a non‑local term and weighted local rewards (λ, μ_i). This dual‑reward structure mitigates non‑stationarity typical in multi‑agent settings and encourages cooperation toward a common objective.

Training proceeds in two phases. Phase 1 pre‑trains each local agent independently on its subgraph, allowing rapid convergence on localized decisions. Phase 2 jointly trains the global agent together with the already‑trained locals, using experience replay buffers to improve sample efficiency and stability. The authors also discuss practical optimization techniques such as GraphSAGE for embedding compression and experience replay for faster convergence.

For evaluation, the authors plan to use the ECLYPSE simulation platform, extending it to support MARL experiments in edge‑cloud environments. Metrics include the three core cost components (RU, CT, SVR) as well as training efficiency (time and memory consumption). Future work includes real‑world test‑bed validation, scaling experiments with larger topologies, and incorporation of economic factors such as energy consumption and operational costs into the objective function.

In conclusion, the paper presents a novel hierarchical architecture that leverages GNN‑based state representation and collaborative MARL to achieve scalable, adaptive, and globally coherent resource management across the Cloud‑Edge Continuum. While the design is promising, the current work lacks empirical results; future studies must quantify communication overhead of the global coordinator, assess the memory/computation trade‑offs of the GNN embeddings at scale, and demonstrate real‑time performance on heterogeneous edge devices. If these challenges are addressed, the proposed framework could become a cornerstone for next‑generation continuum resource orchestration.