A comparison of probabilistic methods for classification

The authors study a class of problems in which the characteristics of the objects in the frame of discernment U={u₁ ,. . ., u_n} are represented probabilistically. A hypothesis is defined by attributes A₁,. . .,A _s which takes values from the sets V₁,. . .,V_s, respectively. Domain information describing a hypothesis specifies the probability of each attribute A_i assuming the values from V_i. The domain information concerning attribute A_i is given by a matrix. The generation of support is driven by the acquisition of evidence concerning attribute values. To compare evidential support generation a simple urn model is constructed to provide the probabilistic domain information. An attribute-value domain is constructed to provide a baseline by which to compare the support generated by an iterative updating architecture, a belief network, and the Dempster-Shafer theory of evidential reasoning