Content area
Next-generation networks, especially wireless and cellular networks, will feature massive IoT and embedded device deployments alongside various novel services, most of which may be constrained by energy and computational capabilities. As a result, these networks must address challenges such as device limitations, scalability, resource management, and energy efficiency for both computation and communication. The rise of deep learning has unlocked significant potential for various new AI-driven applications, meaning next-generation communication networks could benefit tremendously from these approaches, particularly for efficient resource allocation. Simultaneously, the growing prevalence of AI-driven applications on intelligent devices at the network edge implies that numerous applications, such as AR/VR, will significantly amplify the demand for computational resources. Therefore, it is important to study and develop efficient methods of using resources across the network to effectively implement such AI applications. Addressing these challenges requires two important research approaches: developing resource-efficient AI techniques to enhance network operations, and designing learning and inference approaches utilizing next-generation networks to effectively support demanding AI applications.
This dissertation focuses on resource efficiency in next-generation networks and deep learning methods through four key problems.
First, we deploy a cloud-native cellular network (including base station, UE, and core) as softwarized virtual containers, alongside a metrics collection approach to monitor the network. We generate user plan data based on Telecom Italia service request data and explore different machine learning approaches to investigate the most effective policy for predictively scaling various VNF resources in 5G and beyond cellular networks. In the second project, we investigate efficient deployment policies for network-efficient distributed SDN (Software Defined Networking) controllers in mobile or wireless networks. Different objectives in such an environment rely on different properties of the communication network, so the problem is formulated as an MDP (Markov Decision Process). Owing to the large state-action space of this problem, deep reinforcement and transfer learning-based strategies are applied to develop an efficient synchronization policy. The method is further extended to solve a joint controller synchronization and placement problem.
Beyond leveraging deep learning methods to enhance communication efficiency in next-generation cellular and wireless networks, such networks also improve the learning and inference processes of these methods in various ways. We develop adaptive, compression-aware split learning methods that improve resource efficiency and privacy preservation for deep learning models. This is achieved by designing an approach where a communication budget dictates network availability during learning and inference, supported by a fast and efficient technique that adapts to varying communication budgets. Finally, we develop a method for throughput optimization of LLM models via split learning during inference, resulting in improved throughput for LLM demands in an edge-cloud environment.