Research On Multi-Agent Communication Technology Based On Edge Computing

In recent years,with the rapid development of wireless communication technology and artificial intelligence(AI)technology,they inevitably converge,driving the further evolution of mobile communication systems,and giving rise to a large number of new services,such as Internet of Things(IoT),Internet of Vehicles(IoV),etc.Previously,Cloud Computing(CC)provides computing and storage services.However,the new services provided by the 5th Generation(5G)mobile communication,require lower latency,higher energy efficiency and privacy and security requirements.As a result,Edge Computing(EC)and the corresponding Edge Intelligence technologies have been proposed and received widespread attention.EC aims to deploy servers at the edge of the network to greatly reduce the latency of computing services.The edge intelligence relies on the distributed nature of edge computing platforms to embed the powerful learning and modeling capabilities of AI to achieve the convergence of learning,computing,and communication.However,there are numerous potential challenges in deploying edge computing and edge intelligence in specific scenarios,mainly including(1)multi-agent collaboration problems,(2)multivariate heterogeneity problems,and(3)engineering problems for multi-agent learning platforms.These issues will be investigated and studied in the thesis.To begin with,the thesis studies the task and resource allocation problems in collaborative Multi-Agent System(MAS).After the characteristics of MAS have been analyzed,a new multi-level implementation framework is proposed by combining sensing,computation and offloading,in which multiple Slave Agents(SAs)are the actual work units and a Master Agent(MA)acts as an edge server to receive and process the computation tasks offloaded by SAs.If the MA still fails to complete the computation task on time,the task will further be offloaded to the Base Station(BS).Based on this,this work proposes the joint sensing and computation-offloading optimization problem to minimize the total energy consumption of the SA.Then,this work gives an optimal algorithm to solve the problem,namely the optimal policy.On the other hand,taking the residual energy of the devices into account,the original joint optimization problem is decomposed to address the requirement of always-on devices and the adaptability to variable channel state.Then,the robust strategy also developed by solving the sub-problems.Finally,numerical simulations indicate that the optimal strategy proposed in this work effectively reduces the total energy consumption for SAs.Besides,the robust strategy extends the device’s operating time as long as possible and is more resilient to variable channel state.Secondly,the thesis deploys Federated Edge Learning(FEEL)into MAS and applies data cleaning techniques to design an agent scheduling algorithm jointly considering data quality,storage state and channel state,and a bandwidth allocation algorithm to reduce the local gradient uploading latency.First,this work extends a data samples evaluation method for FEEL networks and derives an upper bound on its evaluation error.Subsequently,the agent scheduling problem to minimize the weighted sum of data influence,communication latency,and storage state and its corresponding bandwidth allocation problem are proposed.Then,we derive the close-form expressions for the optimal agent scheduling and bandwidth allocation problems,revealing that the optimal agent scheduling probability increases sublinearly with the exponent of 1/2 as the data unbalanced indicator,the data influence indicator goes up and decreases sublinearly with the exponent of-1/2 as the storage unbalanced indicator,latency indicator goes up.Finally,the experiment indicates that the proposed agent scheduling and bandwidth allocation strategy can effectively reduce the storage pressure of the device,reduce the training latency and improve the training accuracy.In the end,the thesis develops a distributed Key-Value(KV)storage component based on the Rafty protocol,called FEELKV,to recover the device’s sudden offline problem and reduce the difficulty of development and operations.In addition to implementing the basic Raft protocol,such as leader election,log replication,and snapshot functions,the MultiNode Raft architecture is further developed to enable concurrent processing of read and write requests and to exempt some devices with heavy storage pressure from managing data.Finally,experimental results show that FEELKV can correctly handle multiple complex concurrency scenarios and the Multi-Node Raft architecture effectively improves read and write performance compared to the basic Raft architecture.These research results in this thesis can provided a theoretical basis for the application of the EC and edge intelligence in the MAS,as well as support for the evolution of 5G to the 6th Generation(6G).