Tactical communication networking faces diverse operational scenarios where network optimization is a very challenging task. Learning from the network environment, in order to optimally adapt the network settings, is an essential requirement for providing efficient communication services in such environments. Cognitive networking deals with the application of cognition to the entire protocol stack for achieving network-wide performance goals. One of the key requirements of a cognitive network is to learn the relationships between network protocol parameters spanning the entire stack in relation with the operating network environment. In this paper, we use a probabilistic graphical modeling approach, Bayesian Networks (BNs), in order to create a representation of the dependence relationships between significant parameters spanning transport and medium access control (MAC) layers in multi-hop wireless network environments. We exploit this model to face one of the problems of the TCP protocol, that does not have any mechanism to infer when congestion is about to occur in the network and therefore waits till some packets are lost for reacting to congestion in the network. Such a reactive nature of TCP leads to wastage of precious network resources like bandwidth and power. In this paper we show how to infer in advance the congestion state of the network. We constructed BNs for different network environments by sampling network parameters on-the-fly in the ns-3 simulation platform. We found that it is possible to predict the congestion state of the network with quite good accuracy given sufficient training samples and the current value of the TCP congestion window. ©2010 IEEE.