Research On Intelligent Sensing And Slicing Resource Management In Wireless Networks

Mobile communications have evolved from the first generation of analog communications to the current fifth generation of digital communications(i.e.5G),and the service has expanded from simple voice communications to ubiquitous connections and a myriad of applications.As a key enabling technology and infrastructure in the era of digital economy,5G networks will bear more applications and massive connections,and face scenarios with different service level agreements.Different business models need to coexist under a unified 5G network architecture.Network slicing technology is seen as the key to 5G's ability to connect everything,which divides the physical network into multiple virtual network slices to adapt to the differentiated needs of different services.Resource allocation of network slicing plays a crucial role in load balancing,resource utilization and network performance.Therefore,the research on network slicing resource management has always been the focus of academic and industrial attention.At the same time,with the advent of the era of big data and the rise of deep learning and reinforcement learning technologies,it has become a trend to use data-driven model algorithms such as LSTM and DQN to intelligent sensing wireless network or solve complex optimization problems,and then design intelligent radio resource management methods for network slicing.In Chapter 2,this dissertation studies a solution of intelligent slicing resource management,that is,the wireless network intelligent sensing technologies are used to predict the time series of network slices,such as the number of requests and data traffic,and the radio resources are allocated according to the predicted results to meet the needs of users within the scheduling cycle.We analyzed the typical time series prediction model,i.e.LSTM,and designed Random Connectivity LSTM(i.e.RCLSTM)to solve the problem of large computation and easy overfitting in LSTM.In the simulation,we choose the data traffic prediction and user mobility prediction tasks which are of reference value to solve the problem of network slicing resource management.The simulation results show that the performance of the proposed model is significantly better than that of the traditional time series prediction methods,and the computational load is significantly reduced at the cost of partial performance loss.In Chapter 3,we study the wireless channel estimation in the intelligent sensing of wireless network and apply deep learning methods to the channel estimation of the massive MIMO system.The reason for this research is that in network slicing,CSI will inevitably be involved in the design of radio resource allocation strategy,and the channel estimation technology is just to obtain accurate CSI.This research is not only important for radio resource management,but also for the signal processing at the receiver and the transmitter in the massive MIMO system.We introduced a GAN-based channel estimation method for the case where the length of the pilot sequence is smaller than the number of antennas at the transmitter.In order to further improve the estimation quality,we inserted the operation of denoising to the noisy pilot in the front end of the channel estimation,and proposed a two-stage channel estimation method,i.e.N2N-GAN.Simulation results demonstrated that our method has better performance than the end-to-end estimation method,and has better adaptability to shorter pilot sequences and more base station antennas.In Chapter 4,we are devoted to solving the problem that random disturbance in network environment will affect the calculation results of slicing resource management methods.We find that distributionsl reinforcement learning has a strong ability to circumvent random factors in the environment and has a more stable convergence process than traditional reinforcement learning.Therefore,we applied the distributional reinforcement learning framework to solve the resource allocation problem,and used the generative adversarial network to learn the distribution of the cumulative return.Finally,we proposed GAN-DDQN algorithm and Dueling GAN-DDQN algorithm with higher training efficiency.In the simulation,we consider the RAN slicing scenario and use the proposed algorithms to learn the bandwidth allocation strategy according to the number of demands of the network slices.The simulation results verified that the strategy learned by our algorithms can provide higher system utilities.Moreover,our algorithms are more stable and converge faster than DQN in the training process.In general,this paper explores the application prospect and improvement direction of deep learning and reinforcement learning in wireless network intelligent sensing and network slicing resource management.In this dissertation,we proposed RCLSTM,N2N-GAN,GAN-DDQN and Dueling GAN-DDQN algorithms to provide solutions to problems of time series prediction,massive MIMO channel estimation and slicing resource allocation,respectively.Finally,the performance gains of all algorithms are verified by comprehensive simulations.