Unsupervised cluster discovery: the peakfinder algorithm

The PeakFinder Algorithm is an unsupervised meAhod for discovering significant clustm (classes) in a noisy histogram whose underlying distribution estimates the probability density function of an n dimensional feature space for one symbolic category. The histogram is filtered using a suitable kernel function, whose strength (window size) is searched for the smallest value that yields the largest number of statistically significant peaks. Each significant peak is then definod as the mode of a class of the feature space, and a gradient-descent: search is used to determine the class membership and class confiderice of each bin in the histogram. Because the method has a statistical basis rather than a geometrical basis, no parametric assumptions m made about the shape of the clusters, which may have arbitrarily complex boundaries in the feature space. Initial tests with compiler-generated noisy histograms demonstrate the PeakFinder Algorithm both correctly identifies the components of the underlying mixture density and indicates when sufficient data has been accumulated in the histogram during training with an adaptive learning process.