| Abstract: |
This article tests human inference rationality when dealing with default
rules. To study human rationality, psychologists currently use classical
models of logic or probability theory as normative models for evaluating
human ability to reason rationally. Our position is that this approach is
convincing, but only manages to capture a specific case of inferential
ability with little regard to conditions of everyday reasoning. We propose
that the most general case to be considered is inference with imperfect
knowledge - in the present case restricted to uncertain knowledge - and
that a natural framework for testing the rationality of plausible reasoning
is System P. This system provides rational postulates for nonmonotonic
inference.
The semantic of the nonmonotonic inference is given by a
possibilistic constraint introduced by Dubois and Prade (1991). This
constraint states that a rule p ( q is a plausible rule if "the degree to
which p ( q is possible" is greater than "the degree to which p ( (q is
possible". Given the choice of this constraint, we study two supplementary
postulates of rationality. Eighty-eight subjects participated in an
experiment whose results confirm - provided that reflexivity and left
logical equivalence would be tested in a further experiment - the
rationality of human nonmonotonic inference according to the rational
postulates of System P.
|