Adaptive density estimators

Adaptive network systems (ANS) have been applied to many pattern recognition problems, but often they give a boolean answer, or one with little or no relation to probability. This work describes several similar ANS approaches to the problem of density estimation and compares their performance on stationary and drifting distributions. The approaches are variations of a proven technique in density estimation: the kernel estimator. The kernel estimator can be though of as a weighted sum of simple density functions, in this case gaussians, which approximate the distribution. This technique has three parameters that can be varied: the number of kernels in the sum, which corresponds to the number of hidden nodes in the network, the means of the kernels, and the variances, or window widths, of the kernels. Different approaches to choosing these parameters, both initially and adaptively from the data are examined.