Proactive Rate and Error Control for packet multimedia transmissions based on loss prediction

This thesis presents a user quality-centric joint Proactive Rate and Error Control Framework for packet multimedia that takes proactive control actions maintaining an optimal receiver session quality while preserving network bandwidth, based on packet loss prediction and on-line quality assessment. This is superior to the reactive feedback used in current sender based rate control mechanisms that depend on RTCP (Real-time Transport Control Protocol [101]) feedback of packet loss. Though initially designed with audio in mind, the mechanisms we propose are applicable to multimedia communication in general. We show the viability of our techniques applied to service-convergent IP multimedia applications, such as VoIP and emerging IPTV. We present a Loss Predictor, an end-to-end passive monitor that tracks the one-way delay, inter-packet gap, as well as short- and long-term trends, and indicates the current degree and severity of congestion and lack of available bandwidth, hence the probability of packet loss in the next window of packets. We also present a passive statistical measurement framework, Audio Genome, which tracks packet loss of a real-time audio stream at the receiver-end and assesses the audio quality objectively in terms of the Mean Opinion Score (MOS), the ITU standard of voice quality assessment [10]. The Error Control recovers individual path loss and maintains receiver quality in the short term using optimal Forward Error Correction (FEC)[60],[83]. The FEC degree and duration are chosen dynamically using a Markov Decision Process and stochastic inventory control, a novel and effective approach in the area of multimedia error control. The Rate Control detects changing bandwidth and uses a rate-quality optimization model to proactively diversify optimal codec/bitrate combination. The combined end-to-end mechanism thus ensures a bandwidth-friendly transmission maintaining an optimal session quality with an effective degree of loss recovery. We present simulation and experiment results showing the superiority of our mechanisms over other similar techniques.