Type-2 fuzzy logic control of PQoS driven adaptive VoIP scheme E. Jammeh, I. Mkwawa, L. Sun and E. Ifeachor Adaptive VoIP schemes have potentially suboptimal performance owing to imprecision in the metrics used to infer network state. An interval Type-2 fuzzy logic controlled scheme for VoIP services is pre- sented. It infers network state from average delivered perceived quality of service and its degradation due to network congestion and updates an AMR codec mode to match voice quality to available network band- width. Tests showed that the scheme maximised delivered voice quality and outperformed an existing adaptive scheme. The scheme achieves robust performance in the presence of input imprecision and can be implemented in VoIP terminals, and the fuzzy rule base is easy to understand and change by non-experts because of its similarity to the human decision-making process. Introduction: The proliferation of multimedia services over IP networks which are not optimised for real-time delivery has necessitated the development of adaptive schemes to efficiently use and share the avail- able network bandwidth in a controlled and friendly manner. To do this, traditional end-to-end network quality of service (NQoS) driven adap- tive Voice over Internet Protocol (VoIP) base adaptation decisions on network state which is inferred from measured end-to-end delay and loss. Perceived Quality of Service (PQoS) driven adaptive schemes base adaptation decisions on user perceived quality which is predicted from NQoS. These adaptive schemes usually dynamically vary the voice encoding parameters such that the encoded voice bandwidth requirements reflect the network state and available bandwidth. But accurately determining network state from inferred metrics is difficult. This makes the performance of traditional adaptive VoIP schemes sub- optimal in the presence of measurement imprecision. Future networks will comprise different network types which will scale in size and traffic volumes. This will introduce extra imprecision in network measurement. To solve this problem, cross-layer adaptive schemes propose making network state available to adaptive schemes. However, the currently deployed Internet and developed VoIP terminals do not have access to cross-layer information. Ideally, a more robust type of controller is needed that can provide acceptable performance in the presence of input imprecision and make adaptive schemes future proof. Controllers based on interval Type-2 (IT-2) fuzzy sets have been shown to be robust even in the presence of input imprecision [1] and they have been used in adaptive IPTV schemes [2, 3]. However, they are yet to be used in adaptive VoIP schemes. The scheme presented in this Letter computes PQoS (MOS) and determines PQoS degradation (dMOS) due to network congestion. These are used as inputs to an inter- val Type-2 (IT-2) fuzzy logic controller (FLC) in order to dynamically update the mode of an adaptive multi-rate (AMR) codec in real time. Simulation results showed that the scheme delivers the maximum pos- sible quality for different network states and outperformed adaptive VoIP (AVoIP) [4] that based adaptation decisions on packet loss. Methodology: Fig. 1 depicts the block diagram of the IT-2 fuzzy logic controlled adaptive VoIP system. The receiver computes network loss and delay which it uses to predict the instantaneous PQoS using a pre- diction model [5]. One way delay is measured from sent and received packet timestamps as, instDelay ¼ P r  P s ð1Þ where P r and P s are the packet received and sent times, respectively. voice source voice sink AMR encoder AMR decoder jitter buffer packetiser feedback control PQoS prediction depacketiser network adaptation commands feedback (NQoS, PQoS) fuzzy logic AMR mode controller Fig. 1 Conceptual diagram of adaptive VoIP system implemented as module in nsMiracle Measured delay is bursty and varies on a wide scale depending on network state and background traffic volume and type. An average delay avgDelay is therefore computed as follows, avgDelay n ¼ a  avgDelay n1 þð1  aÞ instDelay n ð2Þ where a is the filter gain and set to 0.9 in simulations. Delay measurement accuracy depends on clock resolution and syn- chronisation. We assume source and receiver clocks synchronised by Network Time Protocol (NTP). However, fuzzy logic control should be robust even in the presence of synchronisation problems. The packet loss rate measured within a 1 s period and the computed average delay are used to predict the PQoS (MOS). An average MOS is then computed to reflect the average network state by averaging pre- dicted MOS, avgMOS n ¼ a  avgMOS n1 þð1  aÞ MOS n ð3Þ A maximum possible PQoS (maxMOS) is then computed from minimum delay and zero loss to reflect a PQoS under no network con- gestion. avgMOS n represents the quality for the average network state, and max MOS n is the maximum possible quality of the network without any congestion. Quality degradation (dMOS n ) is computed to reflect the quality degradation due to network congestion as, dMOS n ¼ max MOS  avgMOS n ð4Þ avgMOS n and dMOS n are sent as feedback information to the sender every second, where they are used by the FLC to compute an updated AMR mode. The FLC is depicted in Fig. 2. It takes avgMOS n and dMOS n as inputs to generate an output fuzzy set according to rules that determine how the inputs are mapped to an output. The output fuzzy set is type reduced and defuzzified to obtain a crisp control output Ctrl which is used to vary the AMR codec mode as follows, Mode n ¼ Mode n1 þ Ctrl ð5Þ Results: The scheme was tested by having 50 fuzzy adaptive VoIP ses- sions share a network of different bandwidth capacities from 0.1 to 1.5 Mbit/s. The average PQoS obtained by the sessions was compared with that obtained when AVoIP was used instead. It was also compared when no adaptation was used. Fig. 3 shows that the adaptive scheme out- performed AVoIP and non-adaptive sources delivered the worst quality. The bottleneck bandwidth was fixed at 1Mbit/s and the number of sources increased from 10 to 110. The FLC scheme also outperformed AVoIP, as shown in Fig. 4. fuzzifier rule-base inference rules type reducer defuzzifier fuzzy input sets fuzzy output sets ctrl avgMOS dMOS Fig. 2 Conceptual diagram of interval Type-2 fuzzy logic controller 400 500 600 700 network bandwidth, Kbit/s PQoS (MOS) 800 900 1000 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 Fig. 3 Quality comparison for 50 sources sharing an increasing network bandwidth – V– fuzzy adaptive ... O .. non-fuzzy adaptive – B– non-adaptive ELECTRONICS LETTERS 21st January 2010 Vol. 46 No. 2