LAUSR

Federated learning (FL) leverages the private data and computing power of multiple clients to collaboratively train a global model. Many existing FL algorithms over wireless networks adopting synchronous model aggregation suffer from the straggler issue, due to the heterogeneity of local computing power and channel conditions. To address this issue, we in this paper advocate an asynchronous FL framework with adaptive client selection for training latency minimization, taking into account the client availability and long-term fairness. We consider a practical scenario, where the channel conditions and the locally available computing power are not known in prior. This makes the client selection problem challenging, as the training latency consists of the uplink/downlink transmission time and the local training time. To this end, we tackle the asynchronous client selection problem in an online manner by converting the latency minimization problem into a multi-armed bandit problem, and leverage the upper confidence bound policy and virtual queue technique in Lyapunov optimization to solve the problem. We theoretically show that the proposed algorithm achieves sub-linear regret performance, ensures long-term fairness, and guarantees training convergence. Results show that the proposed algorithm can reduce the training time by up to 50% when compared to the baseline algorithms.