Research On Resource Allocation Optimization For Distributed Edge Intelligence

Distributed edge intelligence is a key technology for B5G/6G networks,and it can utilize the computing capacities and data of devices at the network edge,to achieve the model training in a distributed manner.Due to the heterogenous resources in devices,the mismatch between computing capacities and training loads of devices,and the non-independent and identically distributed data,the major problems faced by distributed edge intelligence system are low efficiency of model training and low model accuracy.To solve these problems,this dissertation focuses on resource allocation optimization for distributed edge intelligence,to optimize the resource allocation for devices.The main results of this dissertation are as follows:1)A device-to-device-enabled data sharing scheme is proposed for distributed edge intelligence system in wireless networks,to optimize the resource allocation of mobile devices and the efficiency of model training.In distributed edge intelligence system,mobile devices with heterogeneous computing resources and local data can utilize device-to-device communication technologies,to balance their computing capacities and training loads,and to change their stored data,by exchanging their local data.Meanwhile,a data sharing based optimization problem is formulated,to minimize the total time required by the system for data sharing and model training,by jointly optimizing the allocation of local data,transmission power and bandwidth resource for mobile devices.Due to the discrete variables and coupled variables in the formulated problem,this dissertation transforms the formulated problem by relaxing the discrete variables and introducing slack variables.Then,an efficient solution is proposed for solving the transformed problem,by using the idea of one-dimensional search.Numerical experiment results show that the proposed device-to-device-enabled data sharing scheme can effectively improve the efficiency of model training,and reduce the impacts of non-independent and identically distributed data on model accuracy,as compared with the baseline schemes.2)A Wi-Fi network based platform is built for distributed edge intelligence,and a resource allocation problem,together with an efficient communication rate allocation scheme,is studied for the platform.Specifically,we select the computing devices including Nvidia Jetson AGX Xavier,Nvidia Jetson Xavier NX and Huawei Mate Book X Pro,and the access points including Redmi AC3000 and TP-Link AC1200,to build the platform in Wi-Fi networks.Meanwhile,the performance curve for each computing device is fitted,to characterize the relationship between the actual computing time for a single parameter update and the number of data samples.Besides,an optimization problem is formulated,to minimize the training time in the platform,by optimizing the allocation of communication rates for computing devices.Then,convex optimization technologies are utilized to solve the formulated problem.Experimental results show that the proposed communication rate allocation scheme can effectively reduce the training time in the platform,as compared with the baseline scheme.Moreover,the proposed communication rate allocation scheme can achieve the similar performance both in the platform and in the simulation software.In conclusion,the proposed resource allocation schemes can effectively improve the efficiency of model training,and reduce the impacts of non-independent and identically distributed data on model accuracy,and thus they possess the important research value.Meanwhile,the built platform can provide a powerful support for future works to evaluate their performance.