A learning neural network algorithm that learns time-varying classes
René–Vinicio Sánchez, A. Edger, Claudia I. González · 1996
A new prototype neural net is presented to classify patterns from non-linearly separable and non-stationary datasets. Local density of class membership distribution function (cmd) is estimated with an hyperspherical semi-cover (prototypes) of input data, reflecting as well local homogeneity. The learning algorithm is able to learn time variations of cmd function using adaptive prototypes with adaptive radii. The algorithm uses a modified nearest neighbor rule. Several examples, syntethic and real problem are presented. Error classirication rates were approximately of between 1.0-4.5% and the total of required memory (i.e. total of prototypes) was near optimal.