Integrated Resource Allocation and Customer Selection in Edge Networks: A Multi-Objective Model for Federated Learning

Federated Learning (FA) enables collaborative training of machine learning models while preserving data privacy. However, its implementation in power-constrained and heterogeneous grids poses challenges. To deal with the conflict between energy consumption, convergence time and performance, this work proposes a multi-objective nonlinear mixed integer (PIM) formulation. The modeling integrates two fundamental problems of AF: client selection and resource allocation (CPU frequency), acting adaptively with the evolution of machine learning training rounds. The evolutionary algorithm NSGA-II is used to solve the problem. The proposal is empirically validated through network emulation using MininetFed. The results demonstrate that the methodology reduces energy consumption to about 10% of that required in standard training, maintaining the accuracy of the global model in IID and non-IID scenarios.