A Rate Splitting Multiple Access Interface for Clustered Wireless Federated Learning

Abstract

Consider a wireless federated learning (WFL) system where the edge devices (EDs) performing local training are closely located in a cluster, such that in addition to their private model updates, a locally common (a.k.a. consensus) model can be computed or selected at each cycle of the learning process. For such a clustered WFL (CWFL) paradigm, we design an uplink (UL) radio access scheme based on the rate splitting multiple access (RSMA) architecture, which is shown not only to significantly reduce the latency of WFL in comparison to systems employing time-domain multiple access (TDMA) and nonorthogonal multiple access (NOMA), but also to be more energy efficient, reaching a lower latency with less power than the latter alternatives. To that end, we exploit the cluster-wide consensus model as the common message used to construct the common component of the RSMA scheme and build an optimization problem aimed at minimizing the latency of each CWFL round by means of optimally allocating the corresponding computation times and uplink transmission durations of each ED, taking into account constraints such as energy consumption, data rates and computational capabilities of each ED. By adequately setting a system parameter referred to as the rate-splitting factor, the formulation also applies to systems employing conventional TDMA or NOMA methods, such that the proposed technique can be seen as a generalization of those approaches, in the context of CWFL schemes. Simulation results on the proposed RSMA-interfaced UL scheme for CWFL are given, both with and without the incorporation of optimal NOMA decoding at the base station (BS), the first of which is found to yield only a mild improvement over the latter, indicating that the proposed approach is in fact the key factor in the overall latency reduction achieved.

Keywords

• Wireless federated learning
• rate splitting multiple access
• clustered networks
• delay minimization
• convex optimization