Section: New Results

Quality of Experience

Participants : Yassine Hadjadj-Aoul, Adlen Ksentini, Gerardo Rubino, César Viho, Pantelis Frangoudis, Hyunhee Park, Kandaraj Piamrat.

We continue the development of the PSQA technology (Pseudo-Subjective Quality Assessment) in the area of Quality of Experience (QoE). PSQA is today a stable technology allowing to build measuring modules capable of quantifying the quality of a video or an audio sequence, as perceived by the user, when received through an IP network. It provides an accurate and efficiently computed evaluation of quality. Accuracy means that PSQA gives values close to those that can be obtained from a panel of human observers, under a controlled subjective testing experiment, following an appropriate standard (which depends on the type of sequence or application). Efficiency means that our measuring tool can work in real time, if necessary. Observe that perceived quality is, in general, the main component of QoE when the application or service involves video and audio, or voice. PSQA works by analyzing the networking environment of the communication and some the technical characteristics of the latter. It works without any need to the original sequence (as such, it belongs to the family of no-reference techniques).

It must be pointed out that a PSQA measuring or monitoring module is network-dependent and application-dependent. Basically, for each specific networking technology, application, service, the module must be built from scratch. But once built, it works automatically and efficiently, allowing if necessary its use in real time, typically for controlling purposes.

Learning tools. At the heart of the PSQA approach there is the statistical learning process necessary to develop measuring modules. So far we have been using Random Neural Networks (RNNs) for that purpose (see  [74] for a general description), but recently, we started to explore other approaches. For instance, in the last ten years a new computational paradigm was presented under the name of Reservoir Computing (RC)  [71] with the goal of attacking the main limitations in training time for recurrent neural networks while introducing no significant disadvantages. Two RC models have been proposed independently and simultaneously under the name of Liquid State Machine (LSM)  [73] and Echo State Networks (ESN)  [71] . They constitute today one of the basic paradigms for Recurrent Neural Networks modeling  [72] . The main characteristic of the RC model is that it separates two parts: a static sub-structure called reservoir which involves the use of cycles in order to provide dynamic memory in the network, and a parametric part composed of a function such as a multiple linear regression or a classical single layer network. The reservoir can be seen as a high-dimensional dynamical system that expand the input stream in a space of states. The learning part of the model is the parametric one. In [41] we propose a new learning tool which merges the capabilities of Random Neural Networks (RNNs) with those of RC models. We keep some of the nice features of RNNs with the ability of RC models in predicting time series values. Our tool is called Echo State Queueing Network. In the paper, we illustrate its performances in predicting, in particular, Internet traffic. In [63] , more results about the good behavior of our new tool are presented.

QoE for SVC. A recent video encoding scheme called Scalable Video Coding (SVC) provides the flexibility and the capability to adapt the video quality to varying network conditions and heterogeneous users. Last year, we started to look at the relations between the way SVC is used and the obtained perceived quality. This year we continued these efforts, together with exploring the use of QoE estimation tools for SVC video coding in network control. In [46] we evaluate different configurations for SVC-based adaptive streaming in terms of user QoE. The aim is to provide recommendations about the different rates to be used in order to create the video representation configuration. These results are part of the PhD [11] . In [25] , we extended our previous work on SVC in DVB-T2, by proposing an analytical model to evaluate the performance of associating SVC with DVB-T2 and QoE. To do this, we developed a discrete time Markov Chain model which captures the system evolution in terms of number of SVC layers that need to be decoded in order to increase user QoE. In [45] , we introduced a new solution to be used by a DASH client for selecting the video representation. Our proposal relies on using the PTP synchronization protocol in order to estimate the end-to-end delays between the client and the server, and hence to correlate this information with network load. The correlation between delays and load was based on a fitting function.

In [54] , we focus on SVC multicast over IEEE 802.11 networks. Traditionally, multicast uses the lowest modulation resulting in a video with only base quality even for users with good channel conditions. To optimize QoE, we propose to use multiple multicast sessions with different transmission rates for different SVC layers. The goal is to provide at least the multicast session with acceptable quality to users with bad channel conditions and to provide additional multicast sessions having SVC enhancement layers to users with better channel conditions. The selection of modulation rate for each SVC layer and for each multicast session is achieved with binary integer linear programming depending on network conditions with a goal to maximize global QoE. The results show that our algorithm maximizes global QoE by providing highest quality videos to users with good channel conditions and by guaranteeing at least acceptable QoE for all users.

VoIP. We continued to work on the perceptual quality of voice-based applications and services. In [17] , we consider a well-known and widely used full-reference technique for measuring speech quality called PESQ, and we propose a learning-based tool for approximating PESQ output without any need for the original signal, following the same black-box parametric PSQA approach. The procedure uses the Echo State Networks previously mentioned.

In [48] , we propose a new packet loss model that differentiates loss instances depending on their perceptual impact. In particular, the model captures the differences between short and long interruptions from the perceptual quality viewpoint. In some cases, the delays and their variation have a strong impact on the perceived quality. In [49] we explore the variability of packet delays on MANETs. For that purpose, a wide range of representative scenarios are defined and simulated. The gathered traces are then inspected from qualitative and quantitative perspectives. In [50] , a Markovian model is proposed to capture these and other features of delays in the same class of mobile networks.