With the rapid proliferation of mobile devices and network applications, user service demands exhibit significant spatiotemporal variability and differentiation, posing considerable challenges to efficient resource allocation in Mobile Edge Computing (MEC). To address the differentiated characteristics of user demands across spatial and temporal dimensions, this paper develops an integrated model explicitly embedding spatiotemporal features into task computation offloading and service caching. Considering user-specific constraints such as delay tolerance, we construct a joint multi-dimensional resource optimization model integrating latency constraints with resource allocation conditions. To effectively manage inherent spatiotemporal variability and deep uncertainty in MEC environments while minimizing task processing delays, we propose a spatiotemporal graph-attention-enhanced Multi-Agent Deep Reinforcement Learning (MADRL) approach for joint task offloading and service caching optimization. Specifically, the proposed method employs Graph Attention networks (GAT) to effectively capture node dependencies and spatial correlations in task offloading decisions, and utilizes Long Short-Term Memory (LSTM) modules to accurately model temporal dynamics of service demands, enabling autonomous learning and adaptive decision-making for task offloading, resource allocation, and service caching strategies. Extensive simulations demonstrate that the proposed method achieves superior convergence performance, significantly reduces latency, and improves efficiency compared to existing methods.
Liang et al. (Sun,) studied this question.