Research On Wireless Resource Allocation In Mobile Edge Intelligence

The recent revival of artificial intelligence(AI)is not only revolutionizing almost every branch of science and technology,but also driven the development of relevant technologies in the business sector.At the same time,a large number of emerging applications have emerged.With the continuous popularization and vigorous development of mobile terminal devices and the Internet of Things,a large number of intelligent applications are deployed on the edge of wireless network,which generates massive application data.This trend has generated strong interests in realizing an “intelligent edge”to support AI-enabled applications at various edge devices.However,the traditional cloud computing is limited by long-distance links and band-width bottlenecks,which is not enough to meet the requirements of high-speed data processing in edge network and emerging applications.In order to solve this problem,a new research area,called edge learning is proposed,which crosses and revolutionizes two disciplines: wireless communication and machine learning.Combing with wireless channel transmission,edge intelligence system trains high-performance machine learning model at the edge of networks by using edge servers and computing units of distributed edge devices.A large number of devices are deployed at the edge of the network,leading to the scarcity of wireless network resources.Therefore,how to use the limited wireless network resources to train high-performance machine learning models in the edge network to support the application of artificial intelligence on edge devices is a key research issue of edge intelligence.Aiming at the problems related to wireless network resource allocation and machine learning in edge network,this thesis respectively considers the characteristics of wireless channel and data sample to allocate channel resources in edge network.The specific research work is as follows:(1)In order to improve the channel resource utilization of edge intelligent network and the performance of trained machine learning model,a selection and resource allocation algorithm based on data diversity is proposed.The algorithm defines the ”difference” between data samples.If the difference between the data samples is larger in machine learning model,the characteristics of the model can learn more.We derived the expression of an explicit,which the expression integrates the difference between data samples and the signal-to-noise ratio of the wireless communication channel.This algorithm can effectively improve the performance of machine learning model for wireless resource allocation.(2)The automatic retransmission and resource allocation strategies in edge intelligent net-works are optimized.This thesis proposes a wireless resource allocation algorithm based on the representation of data samples,which has global cogitation and considers the representation of data samples on edge device side and edge server side.The strategy of wireless resource alloca-tion is determined according to the representativeness of data samples.Through this algorithm,We not only give priority to the data samples with high representative data on the user side of the device,but also give more communication resources to the data samples without representative data on the server side to resist the interference of channel noise.Finally,the simulation results show that the algorithm can significantly improve the machine learning performance.(3)For the centralized edge learning scenario,we also propose a retransmission algorithm based on the importance of data.The logic of the algorithm is simple and easy to understand.The misclassified data samples are considered to have more information and more importance,that more communication resources are allocated to ensure the quality of the received data samples.In addition,for the problem of uneven distribution of data samples in distributed edge learning,we also propose an algorithm to allocate wireless communication resources according to the proportion of data samples,so that machine model training can better fit parameters.