Goal-Oriented Communications for federated learning in mobile networks

Federated Learning (FL) faces significant challenges in mobile environments, as the frequent exchange of large model parameters results in substantial communication overhead. Additionally, varying device capabilities and data heterogeneity lead to biased models and inconsistent training efficiency. This paper addresses these challenges by integrating a Goal-oriented Communications (GoC) approach into FL. Unlike traditional metrics such as Quality of Service (QoS) or throughput, GoC focuses on maximizing the impact of transmitted data for specific application goals, and is considered as an essential enabler for 6G networks. By extending GoC to FL applications, we aim to enhance FL communication efficiency through strategic user selection for model parameter transmission, considering network conditions and application-related metrics. We propose an optimization problem to derive the optimal resource allocation policy that maximizes FL model quality while adhering to network resource constraints. The solution involves a hybrid Genetic Algorithm (GA) approach that dynamically adapts policies to find a sub-optimal solution, balancing network and application demands. Our approach ensures resource efficiency, minimizes communication overhead, and maximizes the utility of transmitted data towards model performance.