Video streaming over the best-effort networks is a challenging problem due to the time-varying and uncertain characteristics of the links. When multiple video streams are present in a network, they share and compete for the common bandwidth. In such a setting, a bandwidth allocation algorithm is required to distribute the available resources among the streams in a fair and efficient way. Specifically, it is desired to establish fairness across end-users' Quality of Experience (QoE).
In this research, we propose three novel methods to provide QoE-fair network bandwidth allocation among multiple video streaming sessions. First, we formulate the problem of bandwidth allocation for video flows in the context of Network Utility Maximization (NUM) framework, using sigmoidal utility functions, rather than conventional but unrealistic concave functions. An approximation algorithm for Sigmoidal Programming (SP) is utilized to solve the resulting nonconvex optimization problem, called NUM-SP. Simulation results indicate improvements of at least 60% in average utility/QoE and 45% in fairness, while using slightly less network resources, compared to two representative methods.
Subsequently, we take a collaborative decision-theoretic approach to the problem of rate adaptation among multiple video streaming sessions, and design a multi-objective foresighted optimization model for network resource allocation. A social welfare function is constructed to capture both fairness and efficiency objectives at the same time. Then, assuming a common altruistic goal for all network users, we use multi-agent decision processes to find the optimal policies for all players.
We propose a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) model for the conventional IP networks and a Multi-agent Markov Decision Process (MMDP) model for the SDN-enabled wireless networks. By planning these cooperative decision process models, we find the optimal network bandwidth allocation that leads to social welfare maximization. Distributed multi-agent reinforcement learning algorithms are also designed and proposed as a low-complexity model-free solution to these optimization problems.
Simulations of the proposed methods show that the resulting optimal policies of the novel Social Utility Maximization (SUM) framework outperform existing approaches in terms of both efficiency and fairness. The Dec-POMDP model applied to a server-side rate adaptation results in 25% improvement in efficiency and 13% improvement in fairness, compared to one popular protocol of congestion control for multimedia streaming. Our performance evaluations also show that the MMDP model applied to a client-side rate adaptation like DASH improves efficiency, fairness, and social welfare by as much as 18%, 24%, and 25%, respectively compared to current state-of-the-art.
| Identifer | oai:union.ndltd.org:uottawa.ca/oai:ruor.uottawa.ca:10393/36626 |
| Date | January 2017 |
| Creators | Hemmati, Mahdi |
| Contributors | Shirmohammadi, Shervin |
| Publisher | Université d'Ottawa / University of Ottawa |
| Source Sets | Université d’Ottawa |
| Language | English |
| Detected Language | English |
| Type | Thesis |
Page generated in 0.0019 seconds