Adaptive Client Model Update with Reinforcement Learning in Synchronous Federated Learning

Federated learning is widely applied in green wireless communication, mobile technologies and daily life. It allows multiple parties to jointly train a model on their combined data without revealing any of their local data to a centralized server. However, in practical applications, federated learning requires frequent communication between clients and servers, which brings a considerable burden. In this work, we propose a Federated Learning Deep Q-Learning (FL-DQL) method to reduce the communication frequency between clients and servers in federated learning. FL-DQL selects the local-self-update times of a client adaptively and finds the best trade-off between local update and global parameter aggregation. The performance of FL-DQL is evaluated via extensive experiments with real datasets on a networked prototype system. Results show that FL-DQL effectively reduces the communication overhead among the nodes in our experiments which conforms to the green initiative.