Signal classification by probabilistic reasoning

Much of the work into developing environmental and network awareness for cognitive radios has been focused on developing new metrics to identify the modulation schemes in use by neighboring radio nodes. Unfortunately, the metrics are used to derive only hard decisions which are often threshold-based and therefore unable to assign a measure of likelihood to the candidate modulation schemes. Bayesian Networks are a graphical representation of a joint probability distribution for a set of variables, but have been investigated for the task of classification of an unknown and unobservable variable whose related variables can be observed and measured. Bayesian Network Classifiers (BNC) are an exciting approach to the problem of using a set of observable features to infer the distribution of a class variable like the modulation scheme. BNCs are especially attractive because they are agnostic to the types of features which are observed and because they give a measure of probability that the classification is correct as well as the probabilities for the alternatives. In this work, we extend a previous effort [1] to explicitly investigate the potential use of BNC for classification in cognitive radio applications. We present some preliminary results using a simple set of metrics to demonstrate signal classification using a Naive BNC. We show how performance can be improved with a more complex BNC known as a Tree-Augmented Bayesian Network Classifier (TAN). Finally, we show how easily the soft, probabilistic outputs of the Bayesian classifier can greatly improve performance.