Research On Collaborative Solution Of Edge Computing Resources Based On Deep Learning

One of the research priorities in the field of collaborative edge computing is to effectively analyze the information of computing resource in edge devices,the utilization of computing resource is an important research point of edge collaboration.This thesis focus on energy consumption and time delay are considered in more detail,and more significant research value is put forward.Therefore,to counter the deficiencies of the existing collaborative edge models,this thesis conducts a thoroughgoing analysis on collaborative edge in terms of edge computing node and edge computing layer respectively,specifics are as follows:1.In terms of edge computing nodes: considering that numerous data generated by mobile terminals are transmitted to edge servers,which would multiply edge server computation,lower efficiency and confine collaboration scope and so on,thus,task transfer algorithms and computing resource allocation algorithms are proposed.Considering that each mobile terminal has multiple tasks unloaded to edge computing system of the edge server,the intelligent agent mechanism of the edge server is used to design collaborative decision-making and optimize computing resource allocation.This thesis puts the state decision of the current server as state sample is iterated in the experience pool of the intelligent agent,the selection of the experience pool will result in the increase of the Q value,and the target value will increase with the optimization process.At last,the approximate solution of the optimal decision is gained to minimize the edge server total unloading cost in terms of energy consumption and time delay.Simulation results show that the proposed method can achieve better performance.2.In terms of edge layering: considering the problems including limited resources and high energy consumption are caused by the great difference of computing resource demand and output data between edge layer and terminal layer.A deep learning model is proposed in this thesis to deploy the edge layer efficiently.Based on the diversity of the depth model,this thesis divides the model into the edge server and the terminal device.The time delay of the deep learning model is effectively reduced through the cooperation between the terminal equipment and the edge server,firstly,the deep learning model is divided into neural networks of different sizes in the face of the computation demanding relatively vast resources,the best model segmentation is trained to give full play to the computing advantage of cooperation between terminal and edge layer.Secondly,the simplified "small model" is required to improve the utilization ratio of computing resource in the face of the computation demanding relatively small resources.It is advisable for any deep learning task to train the exit points of branch networks with multiple choices in off-line condition.Higher accuracy requires larger model,and later exit point causes longer relative time.The simulation results show that the accuracy of the model can be sacrificed to better performance when the time of deep learning task is tight.