The Contributions of this Dissertation

In this Section we briefly resume the major contributions of this dissertation. We classify them according to the area they naturally belong to. In the area of multimedia quality assessment, the main contributions are: $\bullet$

• We provide a new methodology to measure real-time multimedia (speech, audio, and/or video) quality in real time (the associated publications are [90,92,95,91]). The main properties of our method that we want to emphasize are the following ones: (i) no access to the original signal is required, (ii) it is not computationally intensive since, once trained, the neural network gives its evaluations in negligible time, (iii) the obtained results correlate well with MOS (i.e. human perception), (iv) all the quality-affecting parameters can be taken into account, (v) it can operate in real time, and, for instance, it can be easily integrated with a multimedia application.
• We investigate the applicability of our method to the case of speech quality assessment. We present the results of real experiments we carried out between national-wide and international-wide sites. The purpose of these experiments is to identify the most typical ranges of the network parameters. We use three different speech encoders, namely PCM, ADPCM and GSM. Three different spoken languages (Arabic, Spanish and French) are considered. In particular, loss rate, consecutive loss duration as well as the packetization interval of the speech are also considered. (The associated publications are [92,90].)
• We explore the applicability of our method to evaluate real-time video in real time transmitted over packet-based networks. We use a H.263 encoder and we selected five important parameters that have the most dominant impact on video quality. These parameters are: bit rate, frame rate, loss rate, consecutive loss duration, and the amount of redundant information to protect against loss. (The associated publication is [95].)
• We provide a study of the impact of quality-affecting parameters on the perceived quality. Our study is for both video and audio/speech applications. To the best of our knowledge, this is the first analysis of the combined effects of several parameters of both the network and the encoder on the perceived quality. (Corresponding publications submitted for possible publication are [94,96].)

In the area of rate control, our contribution is the following: $\bullet$

• We present a new rate control protocol that integrates both network passive measurements (e.g., loss rates and delays) and user's perception of multimedia quality transmitted over that network. The objectives of our rate controller is twofold: first, a better use of bandwidth, and second, delivery of the best possible multimedia quality given the network current situation. (The associated publication is [93].) It can be seen as an application of the method of quality assessment described before. We tested the applicability of our protocol for the case of both speech and video transmission over the public Internet. We relay the basis of designing similar kind of protocols that take benefits of having a tool able to measure in real time multimedia quality.

In the area of traffic prediction, the contribution is following: $\bullet$

• We provide a new method for traffic prediction that takes into account not only short-term dependency of the traffic process, but also long-term one. We study and evaluate our model by using real traffic traces and we provide some possible applications that can make use of our new technique.

In the area of neural network training, the main contribution is the following: $\bullet$

• We provide two new training algorithms for random neural networks. These two algorithms are inspired from the fastest training algorithms for ANN, namely Levenberg-Marquardt (LM) and a recently proposed referred as LM with adaptive momentum. We evaluate the performance of these algorithms and the available gradient descent one. In addition, we provide some elements of comparison between ANN and RNN.