Efficient Federated Learning in Wireless Communication Systems: A Multi-Objective Optimization Perspective

This paper focuses on a federated learning (FL) system that employs a base station as a central server while clients with limited computation capabilities perform local training. The limited bandwidth leads to that only a portion of clients can participate in each FL training round. The different participated clients can largely affect the performance of FL systems, thus requiring efficient computing-communication resource allocation. In FL systems, both model convergence and energy consumption are important metrics. To this end, we formulate a multi-objective optimization problem (MOP) to simultaneously accelerate convergence and reduce energy consumption. To address the MOP, we propose a multi-objective algorithm (MOA) for FL systems to obtain a Pareto optimal solution set, where Tchebycheff approach is adopted to divide MOP into multiple single-objective problems and optimize them by differential evolution. The extensive experiments on Fashion-MNIST dataset in both i.i.d and non-i.i.d data settings illustrates that MOA outperforms other algorithms.