Machine Learning Approach For Resource Management In Mobile Edge Computing Based Networks

With the development of Internet and mobile communications,the traffic carried by the cellular networks is increasing exponentially,which results in the service provided by the current cloud based mobile network cannot meet the Quality of Service(QoS)requirements of their users.One promising method is to deploy computational resources at the wireless base stations(BSs)for processing computational tasks offloaded from mobile devices hence reducing the delay and energy consumption casued by data transmission to the core network.However,the existing resource management strategies in cellular networks cannot meet the requirements of 5G systems due to their properties of large granularity and stationary.To solve these problems,a revolutionary resource management mechanism should be investigated to improve the energy efficiency and to reduce the request delay in mobile edge computing(MEC)network architecture.As a way to realize artificial intelligence,machine learning has been widely studied recently.Compared to the traditional optimization theories that mainly focus on explicit relationship expression,machine learning algorithms can obtain the hidden and understandable knowledge from huge amounts of data by studying the learning mechanism.In addition,the MEC network contains numerous environmental data that are worth mining.Therefore,this thesis studies the machine learning-based resource management technology in mobile edge computing network architecture.Specifically,aiming at the time-varying requirement of requested computational tasks,this thesis proposes a resource optimization theory combining machine learning and optimization methods to solve the challenges of nonlinear,large-scale,and fast changing resource management,thus achieving the customized resource allocation based on different tasks and environmental characteristics.The specific contributions and achievements of this thesis can be summarized as follows:1.Task scheduling and resource management strategy with known user status in a single MEC-based network.To tackle the challenges of high computing power,low delay and energy consumption of user services in a single MEC-based network,this thesis proposes a resource allocation scheme based on centralized reinforcement learning to optimize the performance of a single MEC-based network.The proposed scheme can extract the state characteristics using reinforcement learning so as to distinguish the time-varying network environment and dynamic users' needs.With such state characteristics,the BS can directly adjust the resource allocation strategy and quickly match the multi-dimensional resources with various computational task without iteration optimization process.2.Task scheduling and resource management strategy with unknown user status in a single MEC-based network.Consider the real-time changing user information in a single MEC-based network,this thesis proposes a user state perception method to meet the timeliness requirements of data acquisition in a single MEC-based network.The proposed method that is based on distributed reinforcement learning algorithms can analyze the physical dynamics using the historical monitored data to achieve dynamic environment estimation of the global enviorment.Under the premise that all user status information can be captured accurately and timely,the MEC server can optimize the resource allocation strategy to meet the QoS of all users.3.Cooperative resource management strategy in a multi-cell MEC-based network.Consider the delay and congestion caused by the control information interaction in current multi-cell MEC-based networks,this thesis proposes a cooperative resource management strategy based on federated learning in multi-cell MEC-based networks.Compared to the traditional collaborative resource allocation scheme that requires the BSs to exchange the control message,the proposed method enables each MEC-based network to extract the network status via training the key parameters using its local historical dataset.Different from the transmission of real-time dynamic network environment,the trained parameters are stable and refined and hence,reducing network congestion caused by information interaction control and realizing efficient cooperation of resource allocation in heterogeneous networks.