Low-complexity sequential non-parametric signal classification for wideband cognitive radios

Abstract

This paper addresses the computational complexity of the Dirichlet process mixture model (DPMM)-based Bayesian non-parametric classifier in cognitive radios (CR's). The DPMM is an ideal signal classification tool for wideband CR's (W-CR's) due to its non-parametric structure. However, it can incur a high computational complexity since it usually requires a large number of Gibbs sampling iterations. To address this issue, we first propose a parameter selection policy that efficiently selects the cluster parameters at each Gibbs sampling iteration, leading to a faster convergence to the stationary distribution of the underlying Markov Chain Monte Carlo (MCMC). Next, we propose a sequential DPMM classifier based on a recursive formulation that allows real-time classification of newly detected signals. The proposed algorithms are shown to significantly reduce the computational burden of the DPMM-based classifier, making it suitable for both large-scale and real-time CR applications.