Latency Minimization for TDMA-Based Wireless Federated Learning Networks

Wireless federated learning (FL) is a new distributed machine learning framework that trains a global model through user collaboration over wireless networks. However, the resource heterogeneity of different users greatly impacts on the performance of wireless FL networks. To tackle this challenge, this paper proposes a novel time division multiple access (TDMA) based scheme for wireless FL networks, where local model training and model transmission of different users can be performed in parallel. The problem of minimizing the total latency per training round by jointly optimizing user scheduling, time allocation, computing frequency, and energy allocation, subject to the maximum energy consumption constraint and the maximum computing frequency constraint of each user, is investigated. Based on the particular problem structure, the optimal computing frequency and energy allocation are derived as the functions of time allocation. On this basis, a heuristic user scheduling policy is proposed and the optimal time allocation is obtained via convex optimization. Simulation results demonstrate that the proposed algorithm is of low-complexity and close to optimal. It is shown that the proposed TDMA-based scheme can outperform the state-of-the-art schemes in existing literature.