Sliding Differential Evolution Scheduling for Federated Learning in Bandwidth-Limited Networks

Abstract

Federated learning (FL) in a bandwidth-limited network with energy-limited user equipments (UEs) is underexplored. In this paper, to jointly save energy consumed by the battery-limited UEs and accelerate the convergence of the global model in FL for the bandwidth-limited network, we propose the sliding differential evolution-based scheduling (SDES) policy. To this end, we first formulate an optimization that aims to minimize a weighted sum of energy consumption and model training convergence. Then, we apply the SDES with parallel differential evolution (DE) operations in several small-scale windows, to address the above proposed problem effectively. Compared with existing scheduling policies, the proposed SDES performs well in reducing energy consumption and the model convergence with lower computational complexity.

Private Federated Learning With Misaligned Power Allocation via Over-the-Air Computation

“…We evaluate MPA-DPFL through training convolutional neural network (CNN) [14] on the popular MNIST dataset. The learning rate is set as η = 0.1 and N 0 = 1.…”

Multi-Objective Optimization for Bandwidth-Limited Federated Learning in Wireless Edge Systems

“…In order to describe the size of shared global model parameters w w w t , we introduce a constant J, which is similar to [17]. In addition, selected UE k is assigned with a subchannel where the channel gain h k,t and the bandwidth b k are obtained in the t-th training round.…”

“…Similar to [17], [28], [29], AoU method utilizes the age of information of the delivered dynamic contents as a metric to assess the system performance. In detail, AoU method leverages the model training times to describe the convergence value.…”

“…The parameters of the studied wireless edge FL system are set as follows [17], [34]. In the system, all UEs are distributed randomly in a rectangular area where the vertices are [70, 0, 0], [90, 10, 0], [90, 50, 0], and [70, 50, 0] m. In addition, we denote the locations of the AP as [0, 30, 1] m. The path loss in dB is considered as P L = P L 0 −10β log( d d0 ), where P L 0 = 30 dB represents the path loss at the reference distance d 0 = 1 m, and d is the distance from the signal transmitter to the signal receiver.…”

“…The optimization method is a proposal for multiple objective PSO [45], [46]. • FL with the sliding DE policy (FL-SDE) [17]: It has a single objective function denoted as F = C t + ξE t (∀t ∈ T ) where ξ = 5. The optimization method is a novel sliding differential evolution-based scheduling policy.…”