Optimal Information-based Classification.

Abstract

Classification is the allocation of an object to an existing category among several based on uncertain measurements. Since information is used to quantify uncertainty, it is natural to consider classification and information as complementary subjects. This dissertation touches upon several topics that relate to the problem of classification, such as information, classification, and team classification. Motivated by the U.S. Air Force Intelligence, Surveillance, and Reconnaissance missions, we investigate the aforementioned topics for classifiers that follow two models: classifiers with workload-independent and workload-dependent performance. We adopt workload-independence and dependence as "first-order" models to capture the features of machines and humans, respectively.

We first investigate the relationship between information in the sense of Shannon and classification performance, which is defined as the probability of misclassification. We show that while there is a predominant congruence between them, there are cases when such congruence is violated. We show the phenomenon for both workload-independent and workload-dependent classifiers and investigate the cause of such phenomena analytically.

One way of making classification decisions is by setting a threshold on a measured quantity. For instance, if a measurement falls on one side of the threshold, the object that provided the measurement is classified as one type, otherwise, it is of another type. Exploiting thresholding, we formalize a classifier with dichotomous decisions (i.e., with two options, such as true or false) given a single variable measurement. We further extend the formalization to classifiers with trichotomy (i.e., with three options, such as true, false or unknown) and with multivariate measurements.

When a team of classifiers is considered, issues on how to exploit redundant numbers of classifiers arise. We analyze these classifiers under different architectures, such as parallel or nested. First, we consider a team of homogeneous (identical) classifiers and provide a fusion-rule, supervisor-based strategy using a parallel architecture. Then, we consider a team of heterogeneous classifiers and provide a strategy using a nested architecture. We show results that confirm that both strategies outperform a single classifier.