Section: New Results

Quality of Experience

Participants : Yassine Hadjadj-Aoul, Gerardo Rubino.

QoE in mobile networks. We consider in [43] an important Quality of Experience (QoE) indicator in cellular networks that is reneging of users due to impatience. We specifically consider a cell under heavy load conditions, modeled as a multiclass Processor Sharing system, and compute the reneging probability by using a fluid limit analysis. In order to enhance the user QoE, we propose a radio resource allocation control scheme that minimizes the global reneging rates. This control scheme is based on the $\alpha$-fair scheduling framework and adapts the scheduler parameter depending on the traffic load. While the proposed scheme is simple, our results show that it achieves important performance gains. This work is extended in [42] . By solving the fixed point equation, we obtain a new QoE perturbation metric quantifying the impact of reneging on the performance of the system. This metric is then used to devise a new pricing scheme accounting of reneging. We specifically propose several flavors of this scheme around the idea of having a flat rate for accessing the network and an elastic price related to the level of QoE perturbation induced by communications.

In order to offer a high media quality and a good user satisfaction, the media streaming service requires that transport protocols can be adapted continuously to the network parameters. However, the diversity of terminals (e.g., tablets, smart phones, laptops) and their corresponding capabilities, mean that users' agnostic solutions are inefficient to cope with such diverse contexts. Indeed, the intrinsic characteristics and parameters of the terminals (i.e., devices) need to be taken into account on the video streaming adaptation process. In [17] , we propose an adaptive video streaming solution to improve the user satisfaction factor by adapting the TCP parameters according to the user's parameters on mobile networks. The user satisfaction factor is calculated according to some metrics driven from the user's quality of experience (QoE). The work is validated through our proposal based on a new mobile agent developed on a Linux script platform and tested on different kinds of devices with different scenarios.

Learning tools. Our QoE measuring techniques (see  3.2 ) are based on statistical learning methods, and we have been using Random Neural Networks as our main learning tool. These are actually open queueing networks where customers have a “sign” and behave analogously as neural spiking signals. They have been proposed by Gelenbe in the 80s, and have been used in many areas since then. In [26] , we published a survey about the tool, where we develop in some detail their use in supervised learning, not only for the case of interest in PSQA, our QoE measuring technology. We also discuss the use of powerful optimization methodology, first and second order techniques, that have proved to be very effective in the standard Neural Network area.

Recently, we started to explore new learning techniques. The first reason is not the search for more accurate tools, because ours are, we claim, as accurate as they can be, it is to improve robustness. The second reason is to extend our QoE measuring tools to richer contexts, mainly when we take into account time, that is, time series data. This comes from the observation that in many cases, the way people perceive quality has some “inertia” and depends on the quality perceived some minutes ago. In [66] we explored the capabilities of a recently proposed method called “Reservoir Computing (RC) with Random Static Projections” which combines two ideas, the now classic Reservoir Computing approach and Extreme Learning Machines (ELMs). In our paper, we replaced the ELMs by Radial Basis Functions (RBF) projections. We illustrated the good behavior of this variation of the original technique basically using known benchmarks.

In [67] , we perform a detailed analysis of one of the main instances of the Reservoir Computing idea, called Echo State Network (ESN). This type of model has several parameters to adjust, that have an impact on the performances of the learning procedure. For instance, it has been shown that the spectral radius of the reservoir matrix (the recurrent network structure that doesn't learn during the process) is related to the accuracy and the memory capabilities of the technology. The size of the reservoir is also a parameter to adjust when configuring an ESN for performing some specific task. One of the results of our work is the fact that the periodic or pseudo-period nature of data is also an important factor to be taken into account when designing an ESN, since it has an influence on the impact of parameters such as the previously mentioned spectral radius.

QoE and emergency management. As a by-product of our activities around QoE, we started to work on an application where, instead of evaluating the QoE of, say, a video or voice application, we wanted to evaluate the way users perceive a service not necessarily based on audio or video content. This was related to our participation to the European project QuEEN (see  9.2.2 ). We finished by building a platform where we test different ideas for managing an emergency situation. In our system, we include an automatic evaluator of the perceived quality of the related voice and video communications, since in the case of some catastrophes, the communications can be seriously damaged and it is critical to automatically detect the issue in order to report the problem and to take appropriate countermeasures, when possible. In [55] , we describe some of the aspects of our system and of the implemented mechanisms, and we present some design problems and their solutions, together with illustrations of the capabilities of the tool.