A Deep Reinforcement Learning-based Collaborative Edge Caching Approach for AIoT

Edge caching for Artificial Intelligence of Things (AIoT) data is crucial for reducing cloud load and providing real-time services to AIoT users. AIoT caching faces challenges due to dynamic, context-sensitive data demands. Static methods based on popularity prediction struggle with rapid changes, while fully dynamic approaches may miss long-term trends, reducing effectiveness. To address this, we propose a novel approach that combines region segmentation with Temporal Convolutional Networks (TCN) and multi-head attention mechanisms for ac-curate content request prediction. This approach captures both spatial and temporal patterns in content demand. We optimize content decision-making through a Constrained Markov Decision Process (CMDP) and Deep Q-Network (DQN), enhancing edge hit rates and reducing access latency. Furthermore, our multi-agent learning framework enables centralized training with decentral-ized execution, facilitating efficient content caching across edge nodes. Experimental results show that our approach outperforms baseline methods in edge hit rate and content access latency, offering a more efficient and adaptive solution for AIoT data caching in dynamic environments.