Objective I

In this thesis, we take a system view of the adaptive bitrate selection module. We considered it to be a system made up of self-contained and interconnected functional components. Typically, the behaviour of an ABR module evolves because both its inputs and actions change in response to the change in the operating environment. Hence, it is safe to consider ABR a dynamical system. This assumption leads to the application of the Dynamical System Method in both modelling the interactions of the various components of the system (ABR) and describing the behaviour of the scheme as a whole.

Naturally, before modelling any system, there is a need for a clear articulation of how the system operates. In order to achieve this, Chapter 2 develops what we call Classic Framework, which describes the way and manner a typical state-of-the- art Adaptive Bitrate Selection module works. After that, we take an incremental approach to improving the classic framework. At the beginning of each chapter, between Chapter 3 to Chapter 5, we present a new framework that captures our understanding of the system at that stage.

Fundamentally, the behaviour of a dynamical system is described by the state vector. A state vector is a set of parameters that collectively define exactly the

state of a system. One import assumption of a dynamical system method is that

Si+1 =f(Si). In other words, the current state uniquely determines all future states.

To adequately describe an ABR system, this thesis argues that this particular aspect of dynamical system method is a vital requirement. Because as we have continuously emphasised throughout the thesis, the primary purpose of video quality adaptation is improving the user experience. And user experience is not an isolated event, the current state of a user’s perception of the quality of the delivered video depends on the previous actions taken by an ABR. Furthermore, implicit in this assumption is the presence of an evolution rule that correctly specifies the future states that follow from the current state. Therefore, the first goal we set in Chapter 1 is not to only carefully articulate the valid system states, which is a critical task, but also to define the pattern of state transition rule that maximises user experience.

However, before determining the state vector and the evolution rule, the chapter (Chapter 3) reformulates the classic framework upon which most of the existing state- of-the-art ABR modules are based. The proposed framework, through the Policy 1, breaks the cyclic relationship between throughput estimation module and chunk re- quest scheduling function. The Policy 1 is aimed at ensuring that regardless of the ABR algorithm used in video rate selection, TCP is allowed to reach a steady state without any interference from the scheduling function.

Based on the derived Framework 1, the state vector (S) is defined as a set of two parameters: the buffer level and the throughput. While it is relatively easy to measure and control buffer accurately this is not the case with the available bandwidth. So, to simplify the task of modelling a weaker definition of the system state is proposed in Definition 3.3, in which buffer state change becomes the sole indicator of state change. However, for this definition to be valid Policy 2 must be satisfied, that is, throughout the streaming session an ABR module must not select a video rate that is greater than the available throughput. Building on top of the Definition 3.3, a rate map that determines the state transition path is developed in Equation 3.11. The rate map takes an ABR from the lowest to the highest available video rate as the buffer (state) changes. Simply put, the function (Equation 3.11) takes the current state Si

(buffer level) and returns the video rate to be downloaded, in the process resulting in a new state Si+1 that can be the same or different from the previous state Si. After

receivingith chunk, an ABR evaluates Equation 3.11, using the current buffer level, to obtain the video rate of chunki+ 1. It is worth emphasising that Equation 3.11 is algorithm independent, this guarantees that regardless of the algorithm used, which usually depends on the QoE metric that and ABR designer may want to optimise, the video rate evaluation remains the same.

Equation 3.11 only tells us how to act, given a change in the system state, the actual video rate change depends on the rate evolution constantα. Which as seen in Equation 3.13, in turn, depends on three factors: (1) the maximum video rate that a client can display, (2) the size of the allocated buffer space and (3) the current buffer level. At this point, we can conclude that we have identified the valid system states and the pattern of transition between the defined states, or simply put objective I has been achieved.

At this point it is reasonable to ask, to what extent does achieving the first objective helps in validating the hypothesis? To answer this question, Chapter 3 takes a two-pronged approach: analytical and experimental. It turns out that to validate the hypothesis, all that has to be done is to show that the QoE metrics presented at Section 2.8 are improved by adopting the proposed scheme.

Using the analytical method in Section 3.4, we show that provided the available network capacity is equal to or greater than the maximum video rate; the proposed rate map will converge at the highest video rate. Furthermore, its is shown that the system is stable at the convergence point. In other words, the model guarantees that a client will be streaming, after the convergence time, at maximum video rate without suffering for video rate fluctuation. To verify this proof experimentally, two players: a buffer-based [14] and a throughput-based [22] are modified to work with the proposed rate map. Various experiments are conducted within both wired and wireless environments. The results show that adopting the proposed model increases the average video rate, capacity utilisation and the stability of a streaming session. At the same time, it reduces both the start-up delay and the convergence time. All these

happen without any adverse impact on the player’s fairness both to other players and background traffic. Evidently, using the model has an upward effect on QoE. Hence, the model has contributed in validating our hypothesis.