Traffic-driven task offloading in two-layer MEC: integrating spatiotemporal prediction with cooperative multi-agent reinforcement learning

BMC Bioinformatics · Springer Nature · 2026-10-31 · DOI: 10.1007/s11235-026-01521-0

Аннотация

Mobile edge computing (MEC) is a key enabler for latency-sensitive 6 G services; however, task offloading in multi-base-station multi-MEC systems remains highly challenging due to inter-BS resource contention, time-varying traffic dynamics, and the limited realism of synthetic task-generation models. This paper proposes a task-offloading framework for a two-tier MEC architecture that integrates spatio-temporal traffic forecasting with multi-agent decision making to jointly reduce processing delay, task drop rate, and energy consumption. Specifically, task generation is derived directly from the Abilene backbone origin–destination traffic matrices. A ConvLSTM encoder-decoder is trained to capture the spatial-temporal correlations of network traffic and to generate predictive features that are incorporated into the local observation of each base station. The offloading problem is then formulated as a cooperative Dec-POMDP and solved by multi-agent proximal policy optimization (MAPPO) under a centralized training and decentralized execution paradigm. Experimental results under four traffic scales show that the proposed method consistently outperforms No-offloading, Random, Greedy, and DQN baselines. At the highest traffic load, the proposed framework achieves an average delay of 6.320 slots, a task drop rate of 31.749%, and an energy consumption per successful task of 0.435, corresponding to improvements of 3.8%, 10.8%, and 10.6%, respectively, over DQN. These results confirm that combining proactive traffic awareness with multi-agent coordination is effective in mitigating MEC congestion and improving offloading performance under heavy-load conditions.

Читать у издателя →

Похожие статьи