On the use of undirected probabilistic graphical modeling for cognitive wireless networks

Despite remarkable advancements in the area of computer networks in recent past, the behavior of the network protocol stack is not completely understood. Our knowledge of interaction of a given protocol in one layer of the stack with protocols in other layers is limited. There is no well known model of the protocol stack that can reveal the protocol parameter interrelationships. We propose a framework built on ideas from the fields of advanced machine learning and computer vision to tackle this problem. Our framework makes use of undirected Probabilistic Graphical Models to determine relationships among protocol parameters of different layers. We sample protocol parameters under predefined network scenarios and perform structure learning of their dependency relationships. Subsequently, we obtain graphical model of the protocol stack which is the most important contribution of this paper. We then estimate the node and edge potentials of the graphical model in the parameter estimation step based on collected data and learnt the structure of the protocol stack. We believe that many of the challenges posed by today´s cognitive networks can be properly addressed by making useful inferences from our modeling. As an example, we show that it is possible to predict network latency in terms of Round Trip Time (RTT) fairly accurately given knowledge of only its neighbor protocol parameters in the graphical model.