RL-Assisted Energy-Aware User-Edge Association for IoT-based Hierarchical Federated Learning

The extremely heavy global reliance on IoT devices is causing enormous amounts of data to be gathered and shared in IoT networks. Such data need to efficiently be used in training and deploying of powerful artificially intelligent models for better future event detection and decision making. However, IoT devices suffer from many limitations regarding their energy budget, computational power, and storage space. Therefore, efficient solutions have to be studied and proposed for addressing these limitations. In this paper, we propose an energy-efficient Hierarchical Federated Learning (HFL) framework with optimized client-edge association and resource allocation. This was done by formulating and solving a communication energy minimization problem that takes into consideration the data distribution of the clients and the communication latency between the clients and edges. We also implement an alternative less complex solution leveraging Reinforcement Learning (RL) that provides a fast user-edge association and resource allocation response in highly dynamic HFL networks. The proposed two solutions are compared with several state-of-the-art client-edge association techniques, leveraging MNIST dataset. Moreover, we study the trade-off between minimizing the per-round energy consumption and Kullback-Leibler Divergence (KLD) of the data distribution, and its effect on the total energy consumption.