摘要
This work offers an effective solution to minimize the upload delay in wireless federated learning (FL). Our solution is suitable for typical FL applications where users employ the same model structure and quantization precision to train their local models. In this case, the time to complete model training using FL is dominated by the slowest uplink transmission rate among users, known as the straggler effect. We formulate an upload delay minimization problem considering the transmission power and model accuracy constraints for a frequency division multiple access (FDMA) system due to its popularity. Since the problem is difficult to solve directly, we reformulate the problem under some mild relaxations that are useful to simplify the problem without violating the goal of timely model upload in wireless FL. We develop a simple algorithm to find the optimal solution and present simulation results to demonstrate the performance of the proposed method in comparison with some benchmark schemes.