Files
Abstract
The proliferation of connected devices creates various use cases and heterogeneous services, e.g., augmented/virtual reality (AR/VR), vehicle-to-everything (V2X), and mobile artificial intelligence.These services and use cases have diverse networking and performance requirements such as throughput, delay, and reliability, which challenge the "one-fit-all" service architecture in current networks.
Mobile edge computing (MEC) allows the deployment of computation infrastructures in close proximity to mobile users and is a key technology to effectively serve these services in terms of cost-efficiency, flexibility, and scalability.In this research, an intelligent network management framework in mobile edge computing is explored.The primary challenges lie in the unique characteristics of heterogeneous services and complicated correlations between network management on multiple technical domains and high-dimension performance requirements of mobile users in complex mobile networks.
This research addresses these challenges with two different management approaches, i.e., context-aware service adaptation to network dynamics as service providers and network orchestration intelligence for heterogeneous services as infrastructure providers.From the perspective of service providers, multiple mobile systems are designed to allow service adaptation under complex network dynamics, e.g., channel variation and traffic workload, which dynamically and adaptively adjust resource allocations and system configurations by exploiting the unique characteristics of individual services.Specifically, two mobile AR/VR systems are proposed in distributed edge computing networks to strike the balance between the quality and round-trip latency (RTT) performance of AR/VR services.From the perspective of infrastructure providers, multiple network systems are proposed to enable orchestration intelligence with no need for accurate performance modelings of services, which automatically learn to orchestrate multiple domain network resources for supporting various services by exploiting advanced machine learning techniques.First, two resource orchestration systems are proposed, which learn to allocate cross-domain network resources to heterogeneous services by leveraging Gaussian process (GP) regression techniques.
Then, a network slicing system is designed to enable intelligent orchestration to network slices under high-dimension complex end-to-end networks by leveraging deep reinforcement learning (DRL) techniques.The proposed network management framework in this research unveils the promising directions in effectively and efficiently supporting heterogeneous services, and provides important insights for network management design in the next-generation mobile network.