A contribution to the problem of feature selection with similarity functionals in pattern recognition

Abstract In dealing with the problem of selecting characteristic features for pattern recognition, this paper discusses the evaluation of feature quality or “goodness” by means of a similarity functional. The suggestion by P. M. Lewis 1 to use the trans-information functional is considered more closely, and an experimental method for estimating feature goodness is specified. Characterization of a pattern set by features is interpreted as a transmission system, and the requirement of high recognition reliability is shown to be incompatible with the requirement of low feature space entropy. Applicability of this method is outlined briefly at the end.