A universal methodology for signal classification in non-Gaussian environments

Abstract

The signal classification problem is posed as an M-ary hypothesis testing problem. We develop an asymptotically optimal universal classifier which does not depend on the true statistical model of the environment. We show that the relevant error probabilities decay at least exponentially in the length of the data vector. To support these results we present simulation results comparing the performance of the proposed universal detector with that of a matched filter receiver for finite test sequences.<>