Communication Resource Optimization For Wireless Edge Learning Systems

With the explosive growth of data generated by mobile devices and the remarkable breakthroughs made in artificial intelligence(AI)in recent years,AI has been widely applied in wireless communication,giving birth to many emerging intelligent applications,such as intelligent Internet of Things and intelligent medical care.To support such emerging intelligent applications,wireless edge learning has been regarded as one of the key technologies empowering the next generation communication.It can quickly access distributed data,utilize computing resources of various edge devices,and provide intelligent services for devices by deploying AI algorithms at the edge of the network.However,wireless edge learning still faces many challenges,mainly including:(1)Due to the limited wireless communication resources and computing resources of devices,AI model calculation and model information interaction will generate huge costs during the training process.To tackle the communication and computing bottlenecks of wireless edge learning,it is essential to design optimization methods of communication and computing.(2)Transmission errors caused by unreliable factors in wireless networks will also affect the model training performance.It is necessary to design wireless resource management algorithms in complex wireless scenarios to improve the learning performance.This thesis carries out in-depth research on the above issues,aiming to improve the learning performance(e.g.,learning efficiency,training accuracy,and test accuracy)by designing efficient distributed resource optimization schemes from the perspective of model training and wireless communication resource optimization.First,a user selection policy based on data importance and channel state information(CSI)and a wireless communication resource allocation scheme are proposed for federated edge learning system,which aim to tackle the communication bottleneck caused by the large amount of model parameters exchanged.To quantify the data importance of each device,the relationship between the loss decay and the squared norm of gradient is first analyzed.Then,by jointly considering user selection and communication resource allocation,a combinatorial optimization problem is formulated to maximize the learning efficiency of federated edge learning system.By problem transformation and relaxation,the optimal user selection policy and communication resource allocation are derived,and a polynomial-time optimal algorithm is developed.Finally,two common-used deep neural network(DNN)models are deployed for simulation.The simulation results validate that the proposed algorithm has strong generalization ability and can attain higher learning efficiency as compared with other traditional algorithms.Secondly,to minimize the training latency,a joint model parameter allocation and bandwidth allocation algorithm is proposed for partitioned edge learning system with time-varying channels.First,it is analyzed that minimizing the overall latency is equivalent to separately minimizing the latencies of all rounds.Based on this conclusion,the problem of one-round latency minimization is formulated.Then,to tackle the challenge of unknown future CSI,an equivalent Markov decision progress(MDP)based one-round latency minimization problem is derived.Given bandwidth allocation,the optimal model parameter allocation algorithm is proposed.Next,given model parameter allocation,the suboptimal bandwidth allocation algorithm is proposed for time-varying channels,where bandwidth is allocated sequentially in each coherence-time duration.Based on the above two algorithms,a joint model parameter allocation and bandwidth allocation algorithm is proposed.Finally,simulation results show that the proposed algorithms can reduce the training latency and improve the model training efficiency.Finally,a wireless communication resource allocation scheme is proposed for decentralized edge learning system via unreliable device-to-device(D2D)communications,which aims to mitigate the impact of transmission errors on model training performance.To speed up the model training performance,an optimization problem to minimize the overall model deviation is formulated under given latency requirement by jointly optimizing the broadcast data rate and the bandwidth allocation.In the high signal-to-noise ratio(SNR),sufficient bandwidth,and large latency scenarios,the optimal broadcast data rate is firstly derived to equalize the one-round latency of devices.Then,an optimal bandwidth allocation algorithm is proposed to further improve the learning performance,where the deviation reduction rates of devices must be equivalent.Furthermore,a protocol to realize the decentralized edge learning system is developed,and the convergence analysis of the proposed algorithm is provided.Finally,simulation results verify that the proposed algorithm can improve both the convergence rate and learning accuracy as compared with the baseline algorithm.These research results in this thesis can provide a theoretical basis for the application of the wireless edge learning system,as well as theoretical support and effective technical solutions for the further development of 6G.