An Inference Technique for Integrating Knowledge from Disparate Sources

This paper introduces a formal method for integrating knowledge derived from a variety of sources for use in "perceptual reasoning". The formalism is based on the "evidential proposltlonal calculus", a derivative of Shafer's mathematical theory of evidence [4]. It is more general than either a Boolean or Bayeslan approach, providing for Boolean and Bayeslan inferencing when the appropriate information is available. In this formalism, the likelihood of a proposition A is represented as a subinterval, [s(A), p(A)], of the unit interval, [0, 1]. The evidential support for proposition A is represented by s(A), while p(A) represents its degree of plausibility; p(A) can also be interpreted as the degree to which one fails to doubt A, p(A) being equal to one minus the evidential support for A. This paper describes how evidential information, furnished by a knowledge source in the form of a probability "mass" distribution, can be converted to this interval representation; how, through a set of inference rules for computing intervals of dependent propositions, this Information can be extrapolated from those propositions it directly bears upon, to those it indirectly bears upon; and how multiple bodies of evidential Information can be pooled. A sample application of this approach, modeling the operation of a collection of sensors (a particular type of knowledge source), illustrates these techniques.