Our Proposed Rate Control Protocol

It is likely to see a progressive adaptation of most of the multimedia RTP/UDP/IP based protocols to one or some of the proposed TCP-Friendly protocols. Although this is required to avoid the congestion and collapse of the global Internet, it can drastically reduce user-perceived quality, as explained before. In [41], a TFRC mechanism is proposed to control the rate of real-time applications such as audio and video conferencing. See Section 8.2.2 for more details. We propose to use the neural networks that we presented in the previous Chapters in order to decide on the best adaptation that a sender could use when it receives congestion control information from the network. We show that the usage of the neural networks helps in improving MOS results under the same network conditions. As shown in Chapter 7, the same MOS scores can be obtained by several combinations of input parameters. For example, for a given bandwidth, suggested by TFRC, one can choose the codec that gives the best quality. On the other hand, for a given MOS score one can choose the best-input parameters to reduce the required bandwidth. By changing the packetization interval (and hence changing the packet rate sent to the network), one may improve the loss ratio. Reducing the sending rate by choosing a different codec, adding or removing FEC, changing the packetization interval, are decisions that can change the MOS in the receiver under the same congestion circumstances and may reduce the bandwidth required by the application to deliver the real-time speech signals from the sender to the receiver (see bellow). The new control protocol (shaded parts in Figure 8.1) is composed of three parts:

1. The first part is a TFRC that periodically calculates the suggested sending rate. The periodicity of the calculation can depend on RTCP recommendations (5 seconds and not more than 5% of the data traffic) if long-term adaptation is needed or can behave like the TFRC implementation (every Round Trip Time) if shorter-term adaptation is needed. See Sections 8.2.1 and 8.5 for further information.
2. The second part is the trained neural network at the receiver side, to measure in real time the MOS based on the network conditions and the encoding parameters. It sends feedback information to the sender, which contains the network statistics as well as the MOS result evaluated at the receiver. See the description of the approach in Chapter 4 and the validation of the method for speech and video in Chapters 5 and 6.
3. The third part takes these results and based on internal rules, decides on the new parameters to be used. In this module, a set of controlling parameters should be defined (see bellow for a list of the possible parameters). In addition, the impact of these parameters on the quality should be known. (See Chapter 7 for a guideline.) The goal of this module is to select the best possible values of the controlling parameters to maximize the quality as if it were perceived by the end user (represented by the NN module) based on the network state as measured by the first part. A compromise between parameters stability and bandwidth utilization should be established. A too slow change frequency of the parameters would result in bad quality during congestion and unfairness toward the competing TCP connections. Similarly, too high change frequency, if possible, could result in too many fluctuations in the perceived quality and this is not suitable for the end users, as mentioned before.

Figure 8.1: Architecture of the proposed control mechanism.