Research On Intelligent Configuration And Resource Scheduling For Network Slicing

One of the most prominent features of 5G is its capability of providing differentiated services,including but not limited to e MBB,u RLLC,and m MTC.As the mobile networks continue to evolve,it is reasonable to deduce that the forthcoming 6G needs to support more diversified services.Such diverse applications have different or even contradictory requirements in terms of bandwidth,latency,energy efficiency,mobility,etc.Consequently,it is unreasonable to devise a one-size-fits-all network architecture to fulfill the diverging requirements.Benefiting from the development of Software Defined Network(SDN)and Network Function Virtualization(NFV)technologies,network slicing has been proposed as a key architectural technology to resolve this issue.With the aid of network slicing,mobile operators can build customized logical networks(i.e.,network slices)by tailoring the physical network to provide differentiated services to their subscribers.However,the introduction of network slicing has incurred several issues for the management and resource scheduling(MRS)of wireless networks.First,resource scheduling in network slicing should not only maximize the resource utilization,but also ensure slice isolation.Second,due to the dynamic nature of network slices,traditional static resource allocation schemes are no longer applicable since they will lead to low resource utilization.Third,network slices are unable to support latencycritical applications due to the inevitable backhaul delays caused by the cloud-based architecture of the underlying physical network.Recently,with the advancement of machine learning(ML),a number of researchers resort to ML for automating the MRS of network slices.Benefiting from their capability in forecasting the uncertain traffic and the dynamic environment,intelligent MRS schemes can adaptively adjust network slices in a proactive manner,thereby optimizing long-term resource utilization without service interruption.Therefore,this dissertation focuses on devising intelligent MRS schemes for network slices by exploiting ML,stochastic games,etc.The main focus of this dissertation includes the following four parts:(1)Inner slice reconfiguration for core networks at small time-scale;(2)Inter slice reconfiguration for the core networks at large time-scale;(3)Dynamic RAN slicing problem;(4)Resource scheduling for MEC-enabled RAN slices.First,the problem of inner slice reconfiguration for core network slices at small timescale is considered,which is modeled as a Markov Decision Process(MDP).Since the multi-dimensional discrete action space of the MDP is hard to explore,the problem is intractable by directly employing conventional Deep Q Network(DQN).To address the curse of dimensionality,a discrete Branching Dueling Q-network(discrete BDQ)is proposed by incorporating the action branching architecture into DQN,for drastically decreasing the number of estimated actions.Based on the discrete BDQ network,an intelligent network slice reconfiguration algorithm(INSRA)is developed.Extensive simulation experiments are conducted to evaluate the performance of INSRA and the numerical results reveal that INSRA can minimize the long-term resource consumption and achieve high resource efficiency compared with several benchmark algorithms.Second,the problem of inter-slice reconfiguration problem for core networks at large time-scale is investigated.Existing researches on network slice reconfiguration are either model-driven or data-driven methods.However,model-driven methods may cause resource over-provisioning due to a lack of prediction mechanism,while data-driven methods are unrealistic in inter-slice reconfiguration that involves costly and time-consuming operations such as VNF migration.To address these issues,this dissertation proposes a Hybrid Model-Data Driven(HMD)framework that intelligently performs inter-slice reconfiguration by leveraging prediction interval and robust optimization.Specifically,a Prediction Interval-oriented Predictor(PIP)is designed to produce a prediction interval that can bracket the future traffic demand with a prespecified probability.Based on the prediction interval,an inter-slice reconfiguration scheme(named box optimizer)is proposed to perform fast inter-slice reconfigurations.To tackle the over-conservativeness of the box optimizer,an ellipsoid optimizer with better optimality is proposed at a cost of increased complexity.Numerical results demonstrate that the proposed framework can provide high robustness with low power consumption.Third,the problem of dynamic RAN slicing under demand uncertainty is investigated.Existing solutions rely on heavy-weight machine learning methods that require numerous trial-and-error training,which is computation-intensive and may incur substantial signaling overhead.Due to the limited computing capacity of the RAN,it is unrealistic to apply them in practice.Thus,in this dissertation,a two time-scale framework that performs RAN slicing at different levels of granularity is proposed.Particularly,at the large time-scale,the problem of resource slicing at coarse granularity is formulated as a nonlinear integer programming and a dynamic programming-based algorithm is proposed to tackle it.At the small time-scale,the fine-grained resource adjustment problem is formulated as a stochastic game.Based on the potential game theory and learning automata,a fully distributed learning algorithm is devised to find the Nash Equilibrium of the game.Simulation results demonstrate that the proposed algorithm converges fast and the performance gap is within 10% compared with the centralized algorithm.Finally,the problem of resource scheduling for MEC-enabled RAN slices is studied.The problem is formulated as a bi-level Stackelberg game with multiple leaders and disjoint followers.At the upper-level game,the networks slices act as the leaders to compete for the limited computation and spectrum resources.At the lower-level game,the users that subscribed to the same slice act as the followers to decide their optimal computation offloading strategies.Based on the potential property of both the leaders' game and the followers' games,the bi-level game is proved to admit at least one pure-strategy global NE.Then a distributed algorithm is proposed to find the global NE of the bi-level game.Simulation results demonstrate that our proposed algorithm possess superior performance with a gap within 9% compared with the centralized solution.