Artificial intelligence-empowered resource management system for fog computing networks

Abstract

Nowadays, technology has become an essential part of human life and the digital world is expanding rapidly and advancing in terms of networking technologies. The rapid increase wireless services and applications is demanding faster and higher-capacity networks. Additionally, the Internet-of-Things (IoT) is fueling the need for massive device connectivity and ultra-reliable and real-time interactions. In this context, fog computing has emerged as an appealing solution to meet these requirements while closely handling growing data demands by bringing cloud computing services closer to end devices and integrating virtualized servers. On the one hand, due to volatile traffic demands and capacity-limited resources (e.g., computation, storage, and batteries), fog networks require a distributed intelligent platform that can adapt to network changes and efficiently manage the execution of complex tasks based on the application requirements. On the other hand, given the fact that computing nodes are relatively close to each other in fog networks, task offloading is particularly useful and enables load balancing by distributing the workload among different nodes throughout the network. In this thesis, three main objectives for designing distributed intelligent computing networks are considered. The three objectives are scalability, heterogeneity, and quality of service (QoS) management. Specifically, in order to improve resource utilization and network performance, heterogeneity in resources, QoS, and task characteristics are investigated. Meanwhile, hierarchical fog computing architectures are applied to distribute workloads and communications across time and space, enabling flexible management. Moreover, distributed resource management algorithms at the fog nodes, access points (AP), and mobile users (MU), are implemented to optimally manage and allocate resources. In particular, this thesis proposes promising schemes to enable efficient and scalable resource management for fog computing networks using reinforcement learning (RL) and deep learning as primary tools. In this vein, Chapter 2 studies a joint task offloading and resource allocation problem that considers heterogeneous service tasks in terms of resource characteristics and QoS requirements. We propose a scheme in which each fog node independently finds the optimal task offloading and resource allocation policies in partially-observable environments with the aim of maximizing the processing tasks successfully completed within their time limits. Hence, a deep recurrent RL-based approach is proposed to tackle the challenges associated with incomplete network information and partial observability. Chapter 3 proposes a novel partial offloading and resource scheduling algorithm for multifog networks where the amount of offloading tasks, computational speed, and CPU utilization level are jointly optimized. Hence, a novel method based on the deep recurrent Q-network is provided to minimize the total energy consumption while maximizing the number of tasks successfully executed with limited bandwidth and CPU resources. In contrast to Chapters 2 and 3, where decision-makers (fog nodes) are fully decentralized, i.e., independent learners that do not directly communicate with each other, in Chapter 4, APs as the decision-makers learn to communicate with neighboring APs over limited communication channels to coordinate their behavior. In this context, a new actor-critic RL framework is proposed to minimize the overloaded links and servers and overall bandwidth cost. We extend the actor-critic model whereby the critic network is designed for centralized learning by sharing parameters among the APs. In contrast, the individual actor networks in each AP strive to learn the optimal policy only using local information and communication messages. The proposed scheme can advance the development of communication for efficient edge learning and the application of distributed learning algorithms. Chapter 5 introduces multi-agent reinforcement learning (MARL) and major potential applications of MARL for sixth-generation (6G) networks. As wireless services and applications become more sophisticated and intelligent, it is foreseeable that future wireless networks will become AI-pervasive. Given the ubiquitous AI applications and dynamic wireless communication networks, it is crucial to build AI agents that are capable of adapting to network changes as well as cooperate with each other. This chapter includes a case study of coordinated multiagent resource management in 6G edge computing networks. The study demonstrates the importance of coordination methods to achieve distributed intelligence in 6G networks.

Date	18 May 2022
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Georges Kaddoum (Supervisor)