Age-of-Information Minimization in Federated Learning Based Networks With Non-IID Dataset

In this paper, a federated learning (FL) based system is investigated with non-independent and identically distributed (non-IID) dataset, where multiple devices participate in the global model aggregation through a limited number of sub-channels. By analyzing weight divergence and convergence rate, a new metric is proposed based on age-of-information (AoI), which incorporates latency and can provide an advanced device selection standard. After that, device selection, sub-channel assignment and resource allocation are jointly designed in an overall AoI minimization problem under the maximum energy consumption constraint. The formulated problem is decoupled into two sub-problems. After analyzing the feasibility, the resource allocation problem is transformed to a convex problem, and the closed-from solution is obtained based on KKT conditions. By introducing virtual sub-channels, device selection and sub-channel assignment are jointly solved by a matching based algorithm. Simulation results indicate that the proposed scheme is able to outperform all baselines in terms of both test accuracy and sum AoI, and the developed strategies can achieve significant improvements for all schemes.