Research On Network Slicing Resource Allocation Based On Deep Reinforcement Learning

As one of the key technologies of 5G,network slicing is an on-demand networking method.Network slicing supports service customization and isolation on the infrastructure by separating logical and physical resources.Network slicing completes the design,implementation,deployment,management and maintenance with the help of virtualization and software provided by Network Function Virtualization(NFV)and Software Defined Network(SDN),thereby supports the rapid development and wider application of 5G network.Network slicing can be divided into access network slicing and core network slicing according to different underlying network structures and functions.They form end-to-end network slicing together to provide users with end-to-end network slicing services.This thesis studies the resource allocation of access network slicing and core network slicing.First,this thesis studies the network slice resource allocation problem of ultra-dense network in the access network,and proposes to use the MA3C algorithm based on the attention mechanism and multi-agent cooperation mechanism to make base station access decisions for users.This thesis finds that by using the historical trajectories of mobile users and the historical traffic of all users,the problem of network energy consumption and the frequent handover of base stations of mobile users can be solved.The experimental results show that the MA3C outperforms the comparative algorithms in such indicators as network energy consumption,mobile user base station handover penalty,and constraint violation penalty.And MA3C also has ideal algorithm running time and robustness performance.Second,this thesis studies the network slice resource allocation problem of dynamic multi-tenant network in the core network,and proposes to use deep reinforcement learning algorithm to deploy network slices for users served by low-priority tenants.This thesis finds that it can make network slice deployment decisions with higher long-term rewards by inputting the historical information of network resources and user traffic into the deep reinforcement learning algorithm.Combining with the background of the problem that the network resources and user traffic have a great influence on the reward range,it is found that the Dueling-DQN algorithm is suitable for the problem scenario.And a solution is proposed to deal with large action space.The test results show that the proposed algorithm outperforms the comparative algorithms in terms of network system energy consumption,algorithm running time,and it also has ideal robustness.