Applied AI for QoE-Aware Video Service and Network Management

Abstract

With the vast growth of multimedia traffic on the internet, although revenue for the contributors considerably increases, video service management (VSM) becomes a complex and highly demanding task. Depending on the application, VSM systems have to deliver satisfactory Quality of Experience (QoE) to customers to retain a competitive market service. Furthermore, lack of sufficient QoE is the foremost reason for customer turnover. In this thesis, we investigate two QoE-aware VSM systems: server selection in the cloud gaming and fault diagnosis system in the video networks, where we apply artificial intelligence (AI) to address the challenges in these systems. Cloud gaming (CG) is a high-performance and cost-effective cloud-based video game system, a promising paradigm for game users and providers, where all the computational tasks are offloaded to the cloud and players with low-end devices can play high-end games without the need for advanced hardware. However, resource management is a challenging task on this platform. This study aims to present optimal resource management in a CG system by considering both the service provider and player’s benefits. Accordingly, we model an optimization problem and propose efficient metaheuristic methods: Boosted-GA and Boosted-PSO, for the GPU-based server selection in CG. The proposed methods simultaneously consider service providers’ profits and players’ experiences by maximizing GPU utilization. The second management system investigates video network diagnosis to help service providers perform fault detection and isolation properly. Fault diagnosis is the heart of every VSM system. Network faults degrade QoE and must therefore be detected, isolated, and fixed. However, this is difficult because multiple entities own the end-to-end path’s sections, and the video service provider (VSP) usually does not have access to the other entities’ networks, such as the internet service provider (ISP) and the client’s local network operator. In this study, we first collect a dataset of QoE and network metrics from an actual video streaming testbed. Multiple videos are streamed from a video server to a client network through a simplified ISP network, while faults are generated in the ISP and/or client networks. Second, we propose a novel approach that shows that it is feasible for the VSP to localize the fault with AI’s aid, using only QoE metrics, and without access to the faulty section. The two proposed deep learning methods of multi-layer perceptron (MLP) and long-short-term memory (LSTM) detect and localize the issues precisely. Furthermore, for the QoE-aware VSM systems, when QoE degradation is detected, the users have already experienced that degradation on their screens. To address this lateness, we propose a hybrid state of the art: deep learning method, i.e., BiLSTM-CNN, to forecast the QoE metrics in future time-steps before they appear on the client’s screen. The proposed approach allows VSM systems to fix a problem before causing a serious issue at the end-user or at least reduce overall QoE degradation. This approach can be used in other VSM systems, such as resource allocation in wireless networks.