Uncertainty ad Incompleteness: Breaking the Symmetry of
Defeaible Reasoning *
Piero P. Bonissone David A. Cyrluk James W. Goodwint Jonathan Stillma
Artifcial Intelligence Program
General Electric Corporate Research and Development
Schenectady, New York 12301
Arpanet: bonissone@crd.ge.com cyrluk@crd.ge.com stillma@crd.ge.com
Abstract
Two majo difculties in using default logics are
their intractability and the problem of selecting
among multiple extensions. We propose an ap
proach toheseproblemsbaedonintegrating non
onooniceaoingihlauible eaningbaed
on triangular norms. A previously propedsystem
frreasoningwithuncertainty (RUM) perfomsun
certain monotonic inferences on an acyclic graph.
We have eteded RUM toallownonmonotonic in
feences and cycle within nonmonotonic rles. By
restricting the sie ad compleity of the nonmon
tonic cycles e can still perform efcient inferences.
Uncertanty meaures provide a bai for deciding
between multiple defalts. Diferent algorithms and
heuristics for fnding the optimal defaults are dis
cussed.
1 Introduction
1.1 Moiaion
The management of uncertain inforation in first
generation expert systems, when addressed at all,
ha lagely been left to ad hc methods. This has
been efective onl becase oeational expert sys
tems normally asme that knowledge is complete,
ecise, ad unvaying. Thisfundamental asump
tion is a principal sorce of the limitation of any
Th k a ial supported by the Defense
Advaced Reeac Pojede Agency (DARPA) under
USAF /Rome Air Development Center contract F3002-85
C-033. Views ad conclwions contained in ths pape Me
those of the authors &d should not be interpreted a repre
senting the ofca oinon or policy of DARPA or the U.S.
Goverent.
!Cur ently with Knowledge Analysis, Belmont,
Masachuetts.
34
diagnosticsyste tosingle fault diagnoses, and the
limitationofcasifcationsystemstotime-invariant
phenomena.
The maagement of incomplete information ha
aso lacked a clear focus, a some reearchers have
attemptedto fnd its solution by defning new non
monotoniclogics,byaugmentingcasicallogicwih
default rules of inference, by seaching for minimal
models via fuctional optimiation, or by concen
tratig only on the instruments, Le. TMSs, rather
than the theory reuired to handle this problem.
In the pat, a subset of the athos have co
tributed to the developmet of individual theories
for eaoning with uncertainy and incompleteness.
Bonisone ha proposed RUM, a system for rea
soning with uncertaint whose underlying theory
i anchored on the semantics of many-valued log
ics (Bon87]. This system proides a representation
layer to capture structural and numerical informa
tion about the uncetainty, an infeence layer to
provide a selection of truth-functional tiagular
norm baed calculi [Bon87], and a control layer to
focus the reaoning on subsets of the KB, to (proce
duraly) reslveignoranceandconfict, and tomai
tai the integrity of the inference bae via a belief
revision ysem. RUM, however, does not provide
anydeclaraiverepresentation tohandle incomplete
inforation.
Goodwin (Goo87] and Brown [BGB87] hae pro
vided sch a representation by deelopig theoies
baed on nnmnnic deedenc networks and
algebraic euations over boolean lattices, respec
tiely. These approaches, however, hae negleced
the aspect of uncerain information.
Another otivation is the eistence of a new
class of robles, referred to a dnamic clasi
fcain blems [BW88], which cannot be prop-