Research On Task Offloading And Resource Allocation Of Edge Computing Network Based On Federated Learning

In recent years,the number of smart devices connected to the internet has increased dramatically,benefited from the booming development of wireless communication technology and artificial intelligence.Therefore,there is an increasing trend that the demand of users has begun switching from the conventional communication to the computation.In other words,the communication and intensive computation are both involved by lots of applications of the wireless networks.In practice,the computational results are the target,and the commucation is a way to establish connection for users.To support the various demands of users,researchers of the wireless communications proposed the mobile edge computing(MEC).In MEC networks,users with sufficient computational capability can execute the computational tasks locally,while others can offload tasks to the computational access points(CAPs)with powerful computational capabilities rather than the cloud server.Moreover,although the existing wireless technologies have been quite mature,in the face of the dynamic and complex wireless networks,how to design an effective offloading and resource allocation strategy has become a key issue in MEC networks.In this paper,we study the offloading and resource allocation issues of MEC networks with the non-eavesdropping and eavesdropping environment,respectively.Specifically,we first consider a MEC-assisted industrial Internet of Things(IIo T)network with the noneavesdropping environment.In the considered IIo T network,users have some computational tasks,which may need to be offloaded to CAPs to be execute.We define the system optimization target as the linear combination of the normalized system energy and the system latency.Then we design a federated optimization framework based on deep reinforcement learning(DRL)and federated learning(FL)algorithms,to adjust the task offloading ratio,the wireless bandwidth allocation ratio and the transmit power to improve the system performance.In further,we consider a priority-aware multi-user MEC network with the eavesdropping environment.In the considered MEC network,users have some priorityaware computational tasks,which may need to be offloaded to CAPs to be execute.We present three optimization criteria for different application scenarios.Then we develop a priority-aware joint optimization strategy of task offloading,wireless bandwidth allocation and transmit power allocation,by a distributed mechine learning.This optimization method ensure that users with high priority can be allocated more resources,while protecting user data privacy and reducing the learning overhead.Finally,lots of simulation results demonstrate the effectiveness of our proposed schemes for the above key issues from different perspectives.