Risk-Aware Distributed Edge Service Orchestration Using Graph-Attentive Multi-Agent Deep Reinforcement Learning

Distributed edge computing must coordinate concurrent service requests over limited computation and communication resources while meeting strict service-level objectives. This paper presents a risk-aware multi-agent deep reinforcement learning framework for SLA-driven task offloading and resource orchestration in distributed edge computing systems. We formulate edge orchestration as a chance-constrained control problem that enforces, for each service class, probabilistic end-to-end latency and reliability requirements while maximizing long-term computational utility. To support scalable operation, we develop a centralized-training decentralized-execution architecture with graph-attentive state aggregation, enabling edge agents to capture inter-node workload, queueing, and resource-coupling effects. A distributional critic is integrated to learn return distributions and estimate tail-risk behavior, allowing the controller to address rare but critical delay spikes. Constraint satisfaction is handled through a primal–dual Lagrangian mechanism that adapts per-slice risk budgets online and calibrates decisions to SLA violation probability. We further introduce a lightweight risk-aware action refinement step that improves exploration safety without sacrificing decentralized execution. Simulations over heterogeneous edge topologies and non-stationary workloads show that the proposed method reduces SLA violation probability and tail delay compared with constrained MARL, graph-based DRL, and risk-neutral baselines while preserving competitive overall service-completion performance.