RELATION-BASED EVIDENTIAL REASONING

We argue that the difficult and ill-understood task of estimating numerical degrees of belief for the propositions to be used in evidential reasoning can be avoided by replacing estimations of absolute values with the much simpler and more defensible assignments of relations. Using a well-known example we demonstrate that it is difficult to justify decisions based on numerical degrees of belief and to answer questions such as ''How do changes in the numbers affect the answers.'' This shows that the epistemological adequacy criterion for representations is violated and leads to the proposal of a representation method for uncertain information based on representing arguments such as ''evidence e supports alternative set A'' and relative strengths of arguments such as ''e1 supports A1 better than e2 supports A2''. We prove that, in a certain sense, belief functions are equivalent to a special case of the proposed method, namely, the cases where all arguments are based on only one piece of evidence. Finally, we compare our approach to those of, for example, Cohen, Grosof, and Pearl.