Research On Collaborative Computing Optimization Techniques For Mobile Edge Networks

In recent years,with the rapid development of new applications,diversified mobile services that are data-intensive,computing-intensive,and delay-sensitive have placed increasingly high demands on network transmission capabilities and data distribution and processing capabilities.Mobile edge networks sink computing and storage capabilities to a location closer to users,enabling edge networks to have computing capabilities with low latency,large bandwidth and high reliability to meet the differentiated needs of emerging services and provide higher quality services to mobile users.However,the current research on mobile edge network computing techniques faces the following challenges:(1)highly dynamic changes in the mobile edge network environment lead to difficulties accurately perceiving and acquiring system states;(2)rigid resource management and poor scalability of mobile edge networks lead to low computational processing efficiency and resource utilization;(3)large-scale repetitive training and insufficient and single data samples of edge artificial intelligence lead to low training efficiency and quality of service.Meanwhile,the scale and resources of a single edge computing node are relatively limited.With the development and large-scale deployment of mobile edge networks,how to collaboratively schedule multi-layer and ubiquitous computing resources of cloud,edge,and end users to break through the scale and resource bottleneck of a single node to further improve computing performance is also a critical challenge that needs to be solved.Therefore,to address the above challenges,this thesis presents an indepth study of issues related to collaborative computing optimization techniques in mobile edge networks from three aspects:end-edge collaborative computation offloading for partially observable scenarios,end-edgecloud collaborative task scheduling for serverless computing,and Internet of intelligence-empowered edge-to-edge collaborative intelligence sharing,respectively.Corresponding innovative solutions are proposed,and the improvement of the proposed solutions on system performance is also demonstrated through theoretical analysis and simulation experiments.The main contributions of this thesis include three aspects,which are summarized as follows.1.An end-edge collaborative computation offloading mechanism for partially observable mobile edge networks is designed.It is difficult for distributed end users to obtain accurate and complete system state information when making computation offloading decisions.To solve such problems,based on decentralized partially observable Markov decision process theory and partially observable stochastic game theory,this thesis designs a partially observable and distributed end-edge collaborative computation offloading mechanism to study dynamic computation offloading in partially observable scenarios from the perspectives of end users cooperation and end users non-cooperation,respectively.With the proposed mechanism,end users can formulate optimal computation offloading strategies in a distributed manner only according to their local state information,without knowing the state and decision information of other end users,so as to maximize the performance of task processing while meeting the quality of experience requirements of users.2.An end-edge-cloud collaborative task scheduling mechanism for serverless computing-oriented mobile edge networks is designed.To solve the problems of rigid management,poor scalability and inflexible resource allocation in traditional mobile edge networks,this thesis innovatively proposes a serverless edge computing framework,using serverless computing technology to facilitate mobile edge networks to become efficient,economical,distributedly elastic and scalable.Under this framework,based on partially observable stochastic game theory and deep reinforcement learning algorithms,the distributed end-edge-cloud integrated collaborative task scheduling mechanism is studied.Considering the heterogeneous characteristics of serverless edge computing nodes,containerized invocation of serverless computing services,and limited service deployment in serverless edge networks,this mechanism enables serverless edge computing nodes to efficiently perform dynamic scheduling and resource allocation for function-grained serverless computing tasks,thus effectively improving the flexibility of task scheduling and optimizing the efficiency of edge resource utilization.3.An edge-to-edge collaborative intelligence sharing mechanism for Internet of intelligence-empowered mobile edge networks is designed.To solve the problem of low training efficiency and quality of service in edge intelligence caused by large-scale repetitive model training and limited data samples,this thesis introduces the innovative concept of Internet of intelligence in mobile edge networks and designs a distributed edge-to-edge collaborative intelligence sharing scheme that allows distributed edge computing nodes to improve learning performance quickly and economically by sharing learned intelligence.Reputation is introduced as a fairness metric to select reliable edge computing nodes for intelligence sharing to resist unreliable sharing operations in edge networks.Considering the time-varying state of mobile edge networks,a novel collective deep reinforcement learning algorithm is designed based on deep reinforcement learning theory and the idea of "collective learning",which enables each edge computing node to learn distributed intelligence sharing decisions locally based on the soft actor-critic algorithm.Then,the training efficiency is further optimized through the collective learning among different edge computing nodes.