| In recent years,the rapid development of Internet of Things(Io T)and mobile communication has enabled hundreds of millions of devices to connect to the wireless networks and generate massive data in the edge of Io T.Meanwhile,with the progress of artificial intelligence(AI)technology,it has become an important trend to train machine learning models based on the local data of devices to provide intelligent services.Current approaches to train the machine learning models are usually implemented at the cloud data center,which has a powerful computation capacity to finish the learning tasks by collecting training samples from widely distributed devices.However,uploading raw data to the cloud data center can cause extreme transmission latency and privacy concerns.Besides,the capability of wireless networks is adversely affected by the heterogeneous devices,the limited communication resources,the dynamic network environment and the diversified service requirements,thus it has brought unprecedented challenges to the optimal resource management for the intelligent services in terms of performance,latency,energy consumption,and other indicators.This dissertation aims to adapt wireless networks to the complex intelligent services in the edge of Io T by designing wireless federated learning methods and multiple resource management strategies.The main contributions of this dissertation can be summarized as follows.Firstly,due to the problem of extended processing delay of intelligent services and unbalanced data distribution,a federated learning and device scheduling method for the online latency optimization is proposed.Specifically,based on the local data of devices,the federated learning method is used to train the machine learning model,and an optimization problem is formulated to minimize the training latency for a given learning performance by the online device scheduling.Instead of assuming that the prior information such as channel state information(CSI),transmission latency and local computing power of the devices is available,a more practical scenario is considered without knowing the prior information,in which the device scheduling problem is reformulated as a multi-armed bandit(MAB)program and then an online device scheduling scheme is proposed based on the ε-greedy algorithm to achieve a tradeoff between the exploration and the exploitation.The proposed device scheduling scheme jointly considers the online training latency evaluation and the importance of local models.In addition,selecting a part of devices to participate in training may cause the deviation of global gradient,and due to the training latency budget,some devices are unable to successfully participate in training,therefore,an unbiased gradient aggregation method is used to solve the problem of gradient deviation and improve the learning performance with the analyses of unbiasedness and convergence.Simulation results show that compared with the benchmark schemes,the proposed method can achieve higher test accuracy in the same training time.Secondly,considering the limited energy resources of devices,a joint hierarchical federated learning and dynamic device association method for long-term energy control is proposed.Specifically,the hierarchical federated learning method is used to train the machine learning model,in which the models are aggregated by the cloud server and edge servers.Then a joint problem of resource allocation and device association is proposed to minimize the training latency while also achieving a targeted minimum value of the training loss function and satisfying the long-term energy consumption constraints of individual devices.Since the CSI is unavailable for all rounds,an alternative problem is formulated based on the general Lyapunov optimization framework incorporating the importance of local models,training latency and energy consumption.To solve the alternative problem,by evaluating the energy consumption queue of devices,the resource allocation problem including local computing power control and bandwidth allocation is solved under given device association strategy.Then a low-complexity heuristic algorithm with adding and removing operations is proposed to achieve a suboptimal solution of the dynamic device association problem.Numerical results show that the proposed algorithm can meet the long-term energy consumption budget after multiple training rounds and achieve better learning performance with lower latency,compared to state of the art benchmarks.Thirdly,due to the complex communication environment and the diversified service requirements,a joint hierarchical federated learning and resource allocation method for adaptive transmission is proposed.Specifically,adaptive hierarchical federated learning method is used to train the machine learning model,and a joint problem of transmission control and resource allocation is formulated to reduce the training loss and training latency.To quantify the learning performance,an upper bound of the expectation of global gradient squared,in terms of the edge aggregation interval,the training latency,and the number of successfully participating devices,is derived.Then an alternative problem is formulated and solved by an iterative optimization algorithm.Specifically,given the resource allocation strategy,a relaxation and rounding method is proposed to optimize the edge aggregation interval.Besides,the problem of resource allocation including training time allocation and bandwidth allocation is solved separately based on the convex optimization theory.Simulation results show that the proposed algorithm,compared to the benchmarks,is capable of reducing the training latency by adaptively adjusting the transmission interval and the resource allocation strategy according to the data distribution and the convergence performance.Finally,due to the diversified service types with limited communication resources,a joint clustered federated learning and device clustering method is proposed for multi-task collaboration.Specifically,clustered federated learning method is used to train the machine learning models,in which the devices with congruent data distribution can be grouped into the same cluster.By analyzing the convergence performance and the generalization ability of clustered models,the utility of the clustered model training is defined by jointly considering the cosine similarity,the number of devices per cluster,and the device participation probability.Aiming at maximizing the average utility of devices,a joint problem of resource allocation and device clustering is formulated,which can be solved by decoupling it into two sub-problems.First,given the results of device clustering,resource allocation method based on the convex optimization theory is proposed to make the bandwidth allocation and the transmit power control.Then,according to the individual stability,a coalition formation algorithm is proposed for the device clustering.Finally,the real-data experiments on the classification tasks validate the advantages of the proposed algorithm compared to the baselines in terms of the test accuracy. |
