A dynamic programming procedure for sequential pattern classification and feature selection

Abstract

To classify an input sample to a pattern class in a statistical pattern recognition system, measurements are made on a limited number of features, which are used as a basis for classification. For situations in which the cost of taking measurements is large, sequential techniques are employed, and two problems arise: (1) How many measurements of a specified sequence are required in order to classify a particular sample? (2) In what order should the features be measured to insure that the expected cost of the classification process is minimized? In the past, procedures have been proposed for handling each problem individually under assumptions that excluded consideration of both simultaneously. A procedure based on backward programming, which allows the solution of both optimization problems simultaneously, is proposed. A nontrivial example is presented and results are compared to other sequential and nonsequential statistical schemes. A device for nonsupervised learning is proposed for incorporation into the resulting sequential system. A description of the learning algorithm and an example of its application are presented. An example of the use of this backward programing procedure to solve the problem of selecting an optimum subset of features from a given set for a nonsequential pattern recognition process is also presented.