Unsupervised Learning of Gaussian Mixture Models in the Presence of Dynamic Environments

Abstract

This paper tackles the on-line unsupervised learning problem of Gaussian mixture models in the presence of uncertain dynamic environments. In particular, we assume that the number of Gaussian components (clusters) is unknown and can change over time. We propose a multi-hypothesis adaptive algorithm that continuously updates the number of components and estimates the model parameters as the measurements (sample data) are being acquired. This is done by incrementally maximizing the likelihood probability associated to the estimated parameters and keeping/creating/removing in parallel a number of hypothesis models that are ranked according to the minimum description length (MDL), a well-known concept in information theory. The proposed algorithm has the additional feature that it relaxes “the sufficiently large data set” restriction by not requiring in fact any initial batch of data. Simulation results illustrate the performance of the proposed algorithm.