A classification architecture that uses probabilistic representation of support and conditionalization and expectation for updating belief is presented. The updating is guided by a utility function that determines the type of information to be acquired. Expected entropy is used as the utility measure. The three major components of a classification system are the representation of the domain information, the evidence, and the support updating paradigm. The representative of domain knowledge and evidence is described. A general overview of the classification architecture is given. The computations and assumptions required in this iterative method are presented. A detailed example illustrating the generation of support based on the acquisition of one item of evidence is given.