.png){kind=icon}
.png){kind=icon}
Return
GNN-enhanced Multi-Agent Reinforcement Learning for joint model caching and task offloading in collaborative Mobile Edge Intelligence networks
.png){kind=icon}
PRE
AI
Z
Y
J
Y
Z
Y
郭
DOI：10.1016/j.future.2026.108741.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
While Mobile Edge Intelligence (MEI) provides crucial low-latency Artificial Intelligence (AI) services for Internet of Things (IoT) and 5G/6G networks, individual edge servers struggle to host large scale AI models due to limited storage and compute capabilities. To overcome this Quality of Service (QoS) bottleneck, this paper investigates joint resource optimization in collaborative MEI networks. We aim to minimize the weighted sum of latency and terminal side energy consumption costs for all users across the network, while satisfying AI model availability and caching constraints. Specifically, we first consider a wired backhaul edge layer scenario based on graph topology. By integrating heterogeneous user distributions, time varying wireless channels, and model storage overheads, the coupled optimization problem of dynamic model caching, collaborative task offloading, and resource allocation is formulated as a Decentralized Partially Observable Stochastic Game (Dec-POSG). This provides a systematic decision making framework for multi agent coordination in distributed dynamic environments. Subsequently, to mitigate the decision bias caused by the limited local observations of edge nodes, we propose a topological feature extraction algorithm based on GraphSAGE. By utilizing neighborhood sampling and information aggregation mechanisms, the algorithm effectively captures the spatial correlation features and potential load states among edge servers, achieving deep perception of complex network topology information. Furthermore, we design and implement a Graph Neural Network enhanced Multi Agent Proximal Policy Optimization (GNN-MAPPO) algorithm. Through an end to end distributed learning strategy, the framework demonstrates the potential for near real time and fine grained dynamic management of caching locations, offloading routes, and physical resources, while maintaining low computational overhead. Finally, simulation experiments conducted across metropolitan-scale scenarios validate the proposed framework. The results show that GNN-MAPPO achieves a peak cache hit rate of 0.92 under a 50% model storage ratio. Compared to the state-of-the-art GA-MARL algorithm, GNN-MAPPO reduces the average task completion latency by 12.8% and the total weighted system cost by 15.1% under high-concurrency workloads of 105 requests per second. Furthermore, our learned heuristic achieves a near-optimal cost within 3.4% of the absolute global optimum computed by an exact mathematical solver in small-scale benchmarks. The entire control-plane decision-making process is executed within 8 ms, which consumes less than 8% of the 100 ms scheduling slot duration, leaving more than 92% of the slot fully available for data plane operations. These quantitative results confirm that GNN-MAPPO successfully balances high decision accuracy and low operational latency in large-scale collaborative mobile edge intelligence systems.
Journal
F
IF:
6.1
Papers:
6.8K
Citations:
2.3W
Organization
Cited Papers