Semi-Decentralized Federated Learning With Cooperative D2D Local Model Aggregations

Federated learning has emerged as a popular technique for distributing machine (ML) model training across the wireless edge. In this paper, we propose two timescale hybrid federated learning ( TT-HF ), semi-decentralized architecture that combines conventional device-to-server communication paradigm with device-to-device (D2D) communications training. , during each global aggregation interval, devices (i) perform multiple stochastic gradient descent iterations on their individual datasets, and (ii) aperiodically engage in consensus procedure of parameters through cooperative, distributed D2D within local clusters. With new general definition diversity, formally study convergence behavior resulting bounds ML. We leverage our to develop an adaptive control algorithm tunes step size, rounds, period over time target sublinear rate $\mathcal {O}(1/t)$ while minimizing network resource utilization. Our subsequent experiments demonstrate significantly outperforms current art terms accuracy and/or energy consumption different scenarios where device datasets exhibit statistical heterogeneity. Finally, numerical evaluations robustness against outages caused by fading channels, well favorable performance non-convex loss functions.