A Readily Computable Decision Rule with Variable Dimensionality

Optimal decision strategies such as Bayes and Neyman-Pearson require the computation of likelihood ratios that are difficult to compute in all but a few special cases. In practice, unfounded assumptions are frequently made about the nature of the pattern classes so that these strategies can be used. In this correspondence suboptimal decision strategies are explored that are attractive because they are easy to compute. These offer two rather unusual advantages. If, during the operation of the classifier a measurement is undefined or too difficult to make, it is easy to alter the dimensionality of the decision rule. Furthermore, it is possible to use different sets of features for testing different classes so that dimensionality can be minimized rather easily. Normally the features used for each class are "specialists" in discriminating that class from the mixture of remaining classes.