An Algorithm for Non-Parametric Pattern Recognition

The probability densities of each of K classes must be known for a statistically optimum classification of an input into one of K categories. This article describes an economical technique for the approximation of probability densities as generalized N-dimensional histograms constructed from a limited number of samples of each class. The histogram cell locations, shapes, and sizes are determined adaptively from sequentially introduced samples of known classification. A method of storing and evaluating densities at an arbitrary point in N-space is described. A computer flow chart is given, and the method is illustrated with an example. Some computational techniques facilitating the rapid evaluation of N-dimensional histograms are discussed.