On the Convergence of Multi-Server Federated Learning with Overlapping Area

Multi-server Federated learning (FL) has been considered as a promising solution to address the limited communication resource problem of single-server FL. We consider typical multi-server FL architecture, where coverage areas regional servers may overlap. The key point this architecture is that clients located in overlapping update their local models based on average model all accessible models, which enables indirect sharing among different servers. Due complicated network topology, convergence analysis much more challenging than In paper, we firstly propose novel MS-FedAvg algorithm for and analyze its non-iid datasets general non-convex settings. Since number each server less FL, bandwidth client should be large enough successfully communicate training with server, indicates full participation can work Also, provide partial scheme develop new biased strategy further accelerate convergence. Our results indicate highly depend ratio area type total three strategies. extensive experiments show remarkable performance support our theoretical results.