Organizing multiple experts for efficient pattern recognition
Venugopal Govindaraju, Krassimir Ianakiev · 2000
The purpose of this research is to investigate a general theory of classifier combination under complete abstraction of classifier details. In order to completely separate the design of classifiers and the design of combinators, we will treat each classifier as a black box whose sole interface to the combinator is a set of transformed decision vectors. Since the decision vector transformation is classifier dependent, it has to be provided. Also provided is a set of labeled data whose distribution reflects the environment in which the combined system will operate. All the necessary knowledge about classifiers can be derived through observations of their responses to the observation data, The classes to be recognized in this observation data set are also predetermined. This black box treatment of classifiers can be correlated, trained on different data sets, generating mixed types of outputs, and behaving in various ways. Under these presumptions, we wish to develop a combination theory.