Edge Intelligence-based Radio Access Network Slicing

In order to meet the challenge of Internet of Everything and provide customized services to vertical industries and the Internet of Things,network slicing has become one of the key technologies for the development of next-generation mobile communications.Wherein,the core network and transport network slicing has become mature,and the relevant standardization has been basically perfected,while the establishment of radio access network(RAN)slicing has just started.Hence,RAN slicing has become the focus of research in the field of 5G/6G.RAN slicing is potentially capable of constructing a set of independent virtual logical sub-networks based upon the same physical network infrastructure and resources,where each logical sub-network is customized for a particular type of service.However,due to the complex wireless channel environment in RAN and the real-time service requirements,many difficulties such as the joint optimization of multiple performance metrics,the multi-dimensional resource scheduling,the isolation between RAN slices and the granularities of RAN slicing need to be broken through.In view of this,on the basis of the current theoretical research difficulties,and relying on stochastic geometry theory and multi-objective optimization,this paper focuses on the following issues.Firstly,we focus on the study of multi-dimensional resource scheduling in RAN slices.It is necessary to combine radio resources with caching and computing resources in slicing networks,consider the unified measurement of three-dimensional resources,and find the efficient joint management method.However,most of the current work focuses on the allocation of one or both of these resources,but radio,caching and computing resources are all very important available resources.Secondly,we focus on the study of multi-objective optimization in RAN slices,measure various key performance metrics from the perspective of the whole network,and establish accurate mathematical models for analysis.However,most of the current work focuses on single objective optimization,but each type of RAN slicing has its own performance metric.It is not comprehensive to analyze/optimize only one of these metrics from the perspective of the whole network.Thirdly,we focus on solving the non-scalar multi-objective optimization problem in RAN slices.The multi-agent deep reinforcement learning algorithm has excellent computational and learning abilities,which is conducive to solving high-complexity problems and obtaining Pareto optimal solutions.However,the existing multi-objective optimization for RAN slicing mainly uses the scalar method,but its objective function relies on empirical design,and the solution has a strong dependence on the weights.Therefore,the research in this paper is divided into three sub-topics: the edge caching/computing-based multi-dimensional resource management of RAN slicing,the Lyapunov optimization-based multi-objective analysis of RAN slicing,and the multi-agent deep reinforcement learning-based non-scalar multi-objective optimization of RAN slicing.The main contributions of this paper are:1.This paper strictly formulates the computing,caching and communication resources in the edge caching/computing-based RAN slicing,and designs a flexible multi-dimensional resource optimization method.Firstly,by introducing the edge caching and multi-access edge computing,we design a computing,communication and caching(3C)scheme for the video streaming slicing in order to maximize the network’s energy efficiency(EE),while satisfying the delay constraints.Then,a near-instantaneously adaptive edge caching(NAEC)decision is developed,where the most popular videos are cached without compression,while the remaining videos are compressed and then cached for properly allocating the limited local storage space.Next,based on Lyapunov optimization,an alternating resource optimization(ARO)algorithm is proposed for allocating the optimal subcarrier and power resources,video caching and computing resources,where the total network EE optimization problem is divided into the communication sub-problem and the edge caching/computing sub-problem.Our simulation results show that the cache hit rate of the proposed NAEC decision is always higher than the least frequently used(LFU)-40%decision and static decision,reaching over 80%.The proposed scheme outperforms both the traditional caching scheme as well as the LFU-40% regime,and strikes a compelling tradeoff between the EE and delay.2.This paper studies the Lyapunov optimization-based multi-objective analysis for RAN slicing.Firstly,by invoking the stochastic geometry theory,we solve the difficulty of dynamic modeling of multiple virtual base stations(v BSs),and we study the first-and second-priority performance metrics of both the high-throughput and the low-delay slices.For example,the first-priority performance metric of high-throughput slices is throughput,and the second-priority performance metric is delay;The first-and second-priority performance metrics for low-delay slices are delay and throughput,respectively.From the perspective of the mobile network operator(MNO),the first-priority performance metrics are mapped to a unified utility function by the scalar weighted sum method for optimization,subject to the basic second-priority performance metrics’ guarantee.Secondly,by using Lyapunov optimization,a joint virtual resource optimization algorithm is proposed for maximizing the utility,which dynamically deploys v BSs for each slice,and allocates both virtual spectral resources and the v BSs’ power.Finally,the simulation results can support both the high-throughput slices and the low-delay slices,which guarantees the effective isolation between slices.Moreover,compared with the traditional average allocation algorithm and random allocation algorithm,the system-wide utility of the proposed algorithm is increased by more than 30% and by 45%,respectively.Simultaneously,there is the tradeoff between delay and throughput in both the slices once users’ quality of service(Qo S)requirement is satisfied.3.This paper further constructs the non-scalar multi-objective optimization problem of space-air-ground integrated network(SAGIN)slicing.The proposed problem is solved by using the multi-agent deep reinforcement learning to explore Pareto optimal solutions.Firstly,we dynamically consider three typical classes of RAN slices,namely high-throughput slices,low-delay slices and wide-coverage slices,under the same underlying physical SAGIN.Then,we jointly optimize the throughput,the transmission delay and the coverage area of these three classes of RAN slices in a non-scalar form by considering the distinct channel features and service advantages of the terrestrial,aerial and satellite components of SAGINs.A central and a distributed multi-agent deep deterministic policy gradient(CDMADDPG)algorithm is proposed for solving the above problem by approaching the Pareto optimal solutions.The algorithm firstly makes optimal virtual unmanned aerial vehicle(v UAV)position decision and inter-slice sub-channel and power sharing by relying on a centralized unit.Then it optimizes the intra-slice sub-channel as well as power allocation and virtual base station(v BS)/v UAV/virtual low earth orbit(v LEO)satellite deployment in support of three classes of slices by three separate distributed units.Finally,the proposed method simultaneously optimizes three classes of SAGIN slices and ensures better throughput/average delay/average signal to interference plus noise ratio(SINR)than the traditional MADDPG algorithm in our built multi-user scenarios despite its lower computational complexity.The proposed scheme has also exhibited prominent performance advantages over the single resource allocation mechanism and the fixed v UAV position based scheme.In particular,compared with the single resource allocation mechanism,the throughput,average delay and average SINR of the proposed scheme are even more than double,respectively.Moreover,compared to the traditional scalar method,our scheme has the distinct benefit of operating without designing the utility function and weights in advance,while still finding numerous non-dominated near-Pareto optimal solutions,characterizing the set of tradeoff amongst the different slices.