Graph Attention-Based Cooperative Multi-Agent Reinforcement Learning for Low-Latency Reliable Task Replication in Mobile Edge Computing

In distributed Mobile Edge Computing (MEC) systems, task replication enhances reliability by duplicating offloaded tasks across multiple Base Stations (BS), mitigating disruptions from user mobility, server failures, and network dynamics. Nevertheless, redundant replicas consume scarce computational and bandwidth resources, thereby elevating total latency in constrained MEC environments. Existing replication methods are often driven by independent local-threshold triggers or rely on unrealistic assumptions of timely global state visibility, under partial observability and coupled resources, multiple BS can react to the same uncertainty and generate redundant replicas, leading to over-replication. To address this distributed coordination challenge, we propose GATv2-CMASAC, a cooperative multiagent soft actor-critic framework formulated as a Cooperative Partially Observable Stochastic Game (C-POSG). BS agents leverage a dynamic heterogeneous graph enriched with contextaware state representations derived from Graph Attention Transformer v2 (GATv2) embeddings to make informed decisions on task replication and target selection. In addition, we propose a reliability-weighted reward redistribution mechanism that maps global rewards to agent-specific signals using attention coefficients and BS reliability metrics to address credit assignment and delay reduction. Extensive simulations under varying conditions show that GATv2-CMASAC outperforms heuristic and deep-RL baselines, reducing system-wide latency and lowering overreplication while maximizing reliability. Ablation results verify the contributions of GATv2-based representations and reliabilityweighted reward redistribution.