ATTRIBUTE SELECTION FROM REFERENCE DESCRIPTIONS OF NORMAL 3-DIMENSIONAL SETS

The dispersion of three attributes within a class is represented as an ellipsoid of equal probability density. The method of determining the coefficients of the equation of the ellipsoid from experimental measurements of attributes in several probes of material of the same class yet with uncontrollable, varying parameters is given. The probability measure of the information content of the three attributes is external tangency of the ellipsoids of two neighboring classes. The uniqueness of the criterion is substantiated and its numerical calculation method is given. Important practical conclusions on configuring the attribute triplet to improve pattern recognition reliability are reported based on analysis and calculation results.