|dc.description.abstract||With the commercialization and maturity of the fifth-generation (5G) wireless networks, the next-generation wireless network (NGWN) is envisioned to provide seamless connectivity for mobile user terminals (MUTs) and to support a wide range of new applications with stringent quality of service (QoS) requirements. In the NGWN, the network architecture will be highly heterogeneous due to the integration of terrestrial networks, satellite networks, and aerial networks formed by unmanned aerial vehicles (UAVs), and the network environment becomes highly dynamic because of the mobility of MUTs and the spatiotemporal variation of service demands. In order to provide high-quality services in such dynamic and heterogeneous networks, flexible, fine-grained, and adaptive network management will be essential. Recent advancements in deep learning (DL) and digital twins (DTs) have made it possible to enable data-driven solutions to support network management in the NGWN. DL methods can solve network management problems by leveraging data instead of explicit mathematical models, and DTs can facilitate DL methods by providing extensive data based on the full digital representations created for individual MUTs. Data-driven solutions that integrates DL and DT can address complicated network management problems and explore implicit network characteristics to adapt to dynamic network environments in the NGWN. However, the design of data-driven network management solutions in the NGWN meets several technical challenges: 1) how the NGWN can be configured to support multiple services with different spatiotemporal service demands while simultaneously satisfying their different QoS requirements; 2) how the multi-dimensional network resources are proactively reserved to support MUTs with different mobility patterns in a resource-efficient manner; and 3) how the heterogeneous NGWN components, including base stations (BSs), satellites, and UAVs, jointly coordinate their network resources to support dynamic service demands, etc. In this thesis, we develop efficient data-driven network management strategies in two stages, i.e., long-term network planning and real-time network operation, to address the above challenges in the NGWN.
Firstly, we investigate planning-stage network configuration to satisfy different service requirements for communication services. We consider a two-tier network with one macro BS and multiple small BSs, which supports communication services with different spatiotemporal data traffic distributions. The objective is to maximize the energy efficiency of BSs by jointly configuring downlink transmission power and communication coverage for each BS. To achieve this objective, we first design a network planning scheme with flexible binary slice zooming, dual time-scale planning, and grid-based network planning. The scheme allows flexibility to differentiate the communication coverage and downlink transmission power of the same BS for different services while improving the temporal and spatial granularity of network planning. We formulate a combinatorial optimization problem in which communication coverage management and power control are mutually dependent. To solve the problem, we propose a data-driven method with two steps: 1) we propose an unsupervised-learning-assisted approach to determine the communication coverage of BSs; and 2) we derive a closed-form solution for power control. Secondly, we investigate planning-stage resource reservation for a compute-intensive service to support MUTs with different mobility patterns. The MUTs can offload their computing tasks to the computing servers deployed at the core networks, gateways, and BSs. Each computing server requires both computing and storage resources to execute computing tasks. The objective is to optimize long-term resource reservation by jointly minimizing the usage of computing, storage, and communication resources and the cost from re-configuring resource reservation. To this end, we develop a data-driven network planning scheme with two elements, i.e., multi-resource reservation and resource reservation re-configuration. First, DTs are designed for collecting MUT status data, based on which MUTs are grouped according to their mobility patterns. Then, an optimization algorithm is proposed to customize resource reservation for different groups to satisfy their different resource demands. Last, a meta-learning-based approach is proposed to re-configure resource reservation for balancing the network resource usage and the re-configuration cost. Thirdly, we investigate operation-stage computing resource allocation in a space-air-ground integrated network (SAGIN). A UAV is deployed to fly around MUTs and collect their computing tasks, while scheduling the collected computing tasks to be processed at the UAV locally or offloaded to the nearby BSs or the remote satellite. The energy budget of the UAV, intermittent connectivity between the UAV and BSs, and dynamic computing task arrival pose challenges in computing task scheduling. The objective is to design a real-time computing task scheduling policy for minimizing the delay of computing task offloading and processing in the SAGIN. To achieve the objective, we first formulate the on-line computing scheduling in the dynamic network environment as a constrained Markov decision process. Then, we develop a risk-sensitive reinforcement learning approach in which a risk value is used to represent energy consumption that exceeds the budget. By balancing the risk value and the reward from delay minimization, the UAV can explore the task scheduling policy to minimize task offloading and processing delay while satisfying the UAV energy constraint. Extensive simulation have been conducted to demonstrate that the proposed data-driven network management approach for the NGWN can achieve flexible BS configuration for multiple communication services, fine-grained multi-dimensional resource reservation for a compute-intensive service, and adaptive computing resource allocation in the dynamic SAGIN. The schemes developed in the thesis are valuable to the data-driven network planning and operation in the NGWN.||en