A blind information theoretic approach to automatic signal classification

Previous information theoretic approaches for empirical signal classification have been developed for applications where labeled training data from each of the signal sources is available. These so-called "universal" classifiers have been shown to be asymptotically optimal under very broad statistical conditions on the signals of interest and have been successfully applied to problems in communication signal modulation classification, face recognition for entry control, and as a CDMA receiver for wideband communications. Unfortunately, there are many important applications where training data may not be readily available or is unreliable, or is simply too costly to obtain. To address these limitations of trained systems, we present a new formulation of the universal classifier which does not require explicit training data. This new blind classifier extracts the necessary training information directly from the test data and uses it optimally in constructing the decision statistics. Examples from communication systems are presented. The results show that the performance of this new receiver exceeds that of standard "trained" receivers and nearly matches that of the globally optimum receiver.