[mod.ai] Seminar - Approximate Deduction in Single Evidential Bodies

RUSPINI@SRI-AI.ARPA (02/25/86)

From: RUSPINI@SRI-AI.ARPA


AURA (Automated Uncertainty Reasoning Assembly) is about to resume its
AURAcles after some months of suspended animation. The next talk
(abstract below) is scheduled for next Friday, February 28, 10AM at
EK242. We plan to meet as regularly as possible each Friday thereafter
at the same time.



                       APPROXIMATE DEDUCTION IN
                       SINGLE EVIDENTIAL BODIES

                          Enrique H. Ruspini
                    Artificial Intelligence Center
                          SRI International

The main objective of this talk is the review of ongoing research on
the interpretation and manipulation of conditional evidence within
single evidential bodies. In the context of a single body of evidence,
conditional evidence is expressed as constraints on the possible
values of propositional truth under the assumption that a specific
proposition within the frame of discernment is known to be true. In
this context deductive inference consists of the combination of the
information about the probable truth of ground propositions (facts)
and conditional evidence (rules) to arrive at new (a posteriori)
estimates of propositional support. This process is both conceptually
and procedurally different from those undertaken when several bodies
of evidence are combined (e.g. using the Dempster Combination Rule).

The role of conditional evidence constraints (henceforth called
approximate or uncertain rules) is examined from the viewpoint of both
the theory of interval probabilities and the Dempster-Shafer Calculus
of Evidence. These approaches to the representation and analysis of
uncertain information will be briefly described together with their
theoretical underpinnings. Several possible interpretations of
approximate rules will be discussed and compared. Possible approaches
for the automation of approximate deduction (under each
interpretation) will also be presented.

Time permitting, the role of these results in the generalization of
Reynold's approach to the generation of support and elementary mass
measures will also be discussed.
-------