<p>Federated learning (FL) enables collaborative training of a global model among clients without centralized data collection. However, it is non-trivial to select clients during each communication round due to wireless communication channels between clients and the server, especially the difficulty of obtaining the accurate channel state. To deal with these challenges, we first demonstrate the effectiveness of using link reliability as a metric in client selection to achieve high model accuracy. To tackle the difficulty of obtaining accurate link reliabilities, we then propose an online client selection strategy for FL based on predicted link reliabilities, denoted as FedPRCS, by formulating the problem as a combinatorial multi-armed bandit (CMAB) problem. Our FedPRCS algorithm dynamically assesses link reliabilities and selects clients accordingly, without prior knowledge of link reliabilities. The theoretical analysis on regret bound demonstrates the asymptotic optimality of our algorithm. Simulation results on diverse datasets and different link reliability settings under both IID and Non-IID cases demonstrate the advantages of our algorithm in achieving high model accuracy and low loss.</p>