Pattern Recognition Research

This report is concerned with the adaptive estimation of joint probability densities from a finite number of multi-dimensional vectors of known classification. An estimation procedure for the approximation of probability densities in the form of an n-dimensional histogram is described. The location and shape of the cells in the histogram are dependent on the data. The quality of the estimation procedure and its dependence on the order in which samples of known classification are introduced are described. Two quality measures are studied, one that estimates the probability that the decision is optimum and the other that the decision is correct. Techniques for analysis of data of unknown origin prior to the application of the adaptive pattern recognition techniques are studied. The measurement selection problem of pattern recognition is investigated and the mathematical and engineering problems are separated. Figures of merit to evaluate the usefulness of parameter sets are developed, and mathematical formulations of the parameter selection problem are given.