Distributed Intelligence and Generative AI to Enhance the Performance of Next-Generation Wireless Networks

Abstract

In recent years, the demand for mobile communications has steadily increased, and the volume of wireless traffic data has reached an unprecedented level. The growing capabilities of the next-generation wireless networks offer the potential for revolutionary applications and use cases, including massive machine-type communications, immersive gaming and virtual reality, edge computing and preventive maintenance through Internet of Things (IoT) sensors. These emerging use cases impose stringent requirements on communication systems, demanding greater agility, intelligence, and adaptability.

To meet these rising demands while controlling capital and operational costs, artificial intelligence (AI) techniques have been increasingly integrated into mobile networks, particularly the radio access network (RAN). AI accelerates the automation of network tasks across various domains. For example, reinforcement learning (RL) has shown promising performance in dynamic control applications such as resource allocation and traffic steering. Similarly, supervised learning methods are effective for prediction tasks like traffic volume forecasting and classification tasks like modulation classification.

However, the integration of AI into mobile networks also introduces new challenges, including high computational demands, security vulnerabilities, privacy concerns, and increased system overhead. How to effectively incorporate advanced AI techniques such as transfer learning (TL), knowledge distillation (KD), and model compression to mitigate these concerns and optimize their performance in mobile networks remains an open problem. There is still a big gap between existing AI methodologies and their practical, accessible applications in mobile communication systems.

This thesis focuses on two major research directions that aim to bridge this gap. First, we investigate distributed AI frameworks for mobile networks. By decentralizing intelligence to various nodes, such as base stations (BSs) and user equipment (UE), distributed learning enhances scalability, reduces service latency, and supports user-centric network management. Nevertheless, it introduces challenges including conflicting local decisions, limited node resources, and the presence of adversarial agents, motivating the need for robust and efficient distributed solutions.

Second, we explore the application of generative AI (GAI) and foundation models (FMs) in mobile communications. GAI is an emerging AI technology that has recently achieved groundbreaking advancements in content generation. FMs refer to large-scale machine learning (ML) models trained on a broad data set that can be adapted and fine-tuned for various applications and downstream tasks. As a combination of GAI and FMs, generative FMs (GFMs), such as large language models (LLMs), are equipped with emergent abilities that can deliver unprecedented benefits. Their deployment opens new possibilities for automating network management and realizing intrinsic intelligence in mobile communication systems.

The findings of this research demonstrate the effectiveness of applying both distributed AI techniques and GFMs in mobile communication systems. Distributed learning enables adaptive, decentralized network management across various network nodes and unlocks the potential of on-device intelligence. GFMs introduce a higher level of cognitive capability, assisting in complex or unforeseen network scenarios. By integrating these AI approaches, this work envisions a future mobile communication system that is AI-enhanced, fully automated, and continuously self-evolving.