Modeling And Optimization For The Energy Consumption Of Federated Edge Learning In Wireless Networks

Federated learning is a distributed,privacy-preserving machine learning paradigm.Collaborated with edge servers,federated learning gives birth to a new concept — Federated Edge Learning(FEEL),that is,federated learning at the edge of wireless networks.It is considered as an important enabling technology in future edge intelligence era.However,from the perspective of implementation,there is a bottleneck in FEEL — the issue of energy consumption.Different from the traditional centralized training paradigm,the training process of FEEL is carried out on mobile devices in wireless networks.Since the training tasks are computation-intensive and involve some communication overhead,the learning tasks are high power-consuming.This poses a major challenge for mobile devices with limited battery capacity.Mobile devices may even refuse to participate in training for the sake of saving power.Thus,how to reduce the impact of training energy consumption on mobile devices and improve the sustainability of devices is an issue worth exploring.This dissertation focuses on the modeling and optimization strategy design for the energy consumption of FEEL in three common wireless networks,which are Frequency Division Multiple Access(FDMA),Time Division Multiple Access(TDMA)and Non-Orthogonal Multiple Access(NOMA)communication networks,and then explores the tradeoff between energy consumption and model performance.The optimization strategy refers to the device scheduling and resource allocation strategy,in which devices involve the training devices in wireless networks,such as the smartphones,internet of things and wearable devices;resources involve the communication and computation resources,such as the available bandwidth,time slots,transmission powers and CPU frequencies of devices.Specifically,the following four research works are carried out:(1)Modeling and optimization for the energy consumption of FEEL in FDMA networks.The bandwidth resource is modeled as discrete channels,and the resulting inter-cell interference is taken into account.Therefore,the communication model of this dissertation is more practical.Due to the inter-cell interference,the communication model is complicated.Specifically,the transmission model of one user involves the participation decisions of other users.Besides,since the decision variables are integer,the formulated problem is a nonlinear integer programming problem.A low-complexity double-greedy strategy is developed,which can find the suboptimal solution.First,the original problem is reduced to multiple linear integer programming subproblems by local enumeration and relaxation,and then each subproblem is solved.The solution provides the upper bound for the optimal solution.Subsequently,this solution is improved to make it close to the optimal.Finally,the suboptimal solution of the original problem is obtained.Simulation results show that the inter-cell interference has a great impact on the training energy consumption,and the proposed strategy can reduce the training energy consumption up to 28.5% in average in interference environment compared to the comparison strategies.(2)Modeling and optimization for the energy consumptions of FEEL in TDMA networks.To improve the device sustainability,this dissertation studies the energy consumption optimization strategies from two perspectives: one is to reduce the training energy consumption;the other is to adopt the wireless power transfer technology to transmit energy from the base station to the devices and then minimize the transfer energy consumption.For the first perspective,because of device scheduling,the minimization of training energy consumption is formulated as a nonlinear integer programming problem.In order to improve the efficiency of the solution,a device scheduling strategy based on the hybrid branch-and-bound method is proposed.In the initial stage of node selection,depth first search is used to find feasible solutions to the original problem and quickly determine the upper bound of the optimal value;then,the best first criterion is adopted to traverse subsequent nodes.Simulation results demonstrate that the proposed strategy can reduce the training energy consumption up to 27.5% in average compared to the comparative strategies.For the second perspective,although the base station can transmit energy to the devices to achieve sustainable operations,the transfer of energy increases the power consumption of the base station,so it is necessary to minimize the transfer energy consumption of the base station.The problem of minimizing the transfer energy consumption is nonconvex.A resource allocation strategy based on the line search and subgradient methods is proposed to find the near-optimal solution of the original problem.Given the energy transfer duration,the original problem can be reduced to a convex problem,and then solved by the subgradient method.The line search method is used to find the near-optimal energy transfer duration.Simulation results show that the proposed strategy can reduce the transferring energy consumption up to 54.3% in average compared to the comparative strategy.(3)Modeling and optimization for the energy consumption of FEEL in TDMANOMA networks.In TDMA-NOMA networks,the training devices are grouped;devices in the same group share the same time slot and communicate with the base station in the way of NOMA;different groups communicate in the way of time division multiplexing.Since grouping is required,the formulated problem contains combinatorial property.Thus,it is decomposed into two subproblems: resource allocation and device pairing.For the resource allocation subproblem,the line search method and solver are adopted to find its near-optimal solution,and then its approximation is discussed with the help of sensitivity analysis.For the device pairing subproblem,this dissertation proposes two device pairing strategies based on matching theory and channel gain respectively.The pairing strategy based on matching theory uses the concept of stable matching,and thus can find the near-optimal device pairing form with a lower computational cost compared to the enumeration method.The pairing strategy based on channel gain sorts the devices according to their channel gains,and thus has low computational complexity.Simulation results demonstrate that the proposed strategy in NOMA communication networks can further reduce the training energy consumption up to 20.3% in average compared to the comparison strategy in TDMA communication networks.Then,they also verify the characteristics of the two device pairing strategies and their applicable scenarios.(4)Device scheduling and resource optimization for the tradeoff between training energy consumption and model performance of FEEL.It can be found that there is a tradeoff between training energy consumption and model performance of FEEL due to device scheduling.This work focuses on the tradeoff between the two performance metrics.The formulated problem can be decomposed into two subproblems: resource optimization and device scheduling.For the resource optimization subproblem,it is proved to be a quasi-convex problem without saddle points,so its local optimum is also its global optimum and the steepest descent method can be used to find the optimal solution.For the device scheduling subproblem,two greedy approximation algorithms are proposed.The devices with lower energy consumption in handling one sample have higher priority to be scheduled.Afterwards,the approximate ratios of the two algorithms are proven.The proofs adopt a classification discussion method.First,filter the whole device set based on the propositions and obtain the candidate optimal sets.Then,divide the candidate optimal sets into two cases: single-device sets and multi-device sets.Finally,discuss the approximate ratio in each case,and then make a summarization.Simulation results show that the proposed device scheduling strategy can always find the optimal solution,and the proposed strategy is able to achieve the tradeoff between the two metrics.Compared with the Energy-Efficient Federated edge learning(EEF)strategy,the proposed strategy can reduce the energy consumption up to 37.5% in average and help FEEL converge faster.