421
Cost-Sharing in Bayesian Knowledge Bases
Solomon Eyal Shimony and Carmel Domshlak
Dept. ofath. andComp. ci.
Eugene Santos Jr.
Dept. ofElectrical andComp. Eng.
Ben Guionniversity of the Nege
.0.Bo653,Beer-heva84105,IAE
e-mil: shimony, carmel @cs.bgu.ac.il
Abstract
Bayesianowledgebases (BKBs) are agen
ralization of Bayes networks and weighte
proofgrahs (WAODAGs), thatallowcycles
nthecausalgraph. easoning nBKBs e
uires indin the most robable inferences
onsistent with the vidence. The cost
sharing heristic for iding least-cost ex
planationsinWADAGswaspresented an
shown to be eective by Charniak and Hu
sain. owever, the cycles in BKBs would
make te enition ocost-sharing cyclic s
well, ifapplied directly to BKB. B teat
ingthedening equations ofcost-sharingas
asystem of equations, onecanroperly de
ne an adissible cost-sharing heuistic for
BKBs. Empiricalevaluationshowsthatcost
sharng improves performance signcanty
whenappliedtoBKBs.
1 INTRODUCTION
Bayes networks 7] are a commonly used eaoning
toolwithin the uncertaint in Acomunity. ately,
graphcal causl probabiistic modes hae shown up
that generalize on the acyclc Bayes networks, in or
der to cater for causal phenomena whh cannot be
strictly partialy ordered. These models have causal
cycles 1, 8], or ndirected sections in the directed
graphs 2NJ. Clearly, one sti! needs o do eie be
liefrevisionorbeliefupdatng 7] inorder t perfor
reasoniginthesechemes.These moregeneral mod
els, being less restrictive, pause ineresting problems
ileentngeoningagorithmsfor them.
Baysian knowledge ase (BKBs) [8] ar a gneral
ization of Bays networks and weighted (AND/O,
dreted acycli) proof gaphs (acronym WAODAGs)
4,that allow cclesint causalgraph. Considerthe
prolem of inding the mst probablenference ("ex
planation") consisten with the evidene in a BKB.
Thisproblemisanalogousto(andmregeneralthan)
theN-hardproblemofbeliefreision inBaysnet-
AirForcenstitute ofechnlogy
Wright-Patterson AFB, OH
e-ail:esantos@ait.af.mil
woks,ordnginimum-costpoof onaWAODAG.
As for Bayes networks, reasoning with tre-shaped
BKBscanbedoneecently. However,i tislearthat
inactualappicationswecannotusualyforcourrep
esentation tobelongtotheeasyclassofprbems.
To-date, inding most-probable inference in general
BKBs h been implemented s bes-ist heursc
search, whee the heuristic used wascost-so-far, with
dismal results. The reason is that this lcal heuris
tic does no take into account the cost of odes (or
variables) tbeassignedateronnthesearch. Prop
agation ofcoststobeincurredismuchpreferable,but
it isnon-trivial to do s oinamanner resting in an
ssleeurstc. The latter wasrstachievedby
usingthecost-sharingpropagatedcostmethd3(see
nextsectio forabrief denition).
twasshownbyCharniakandHusain [3] thatforind
igest-ct epnationsnWAODAGs,the(admis
ible)cost-sharingheuristichas amuchbetterperfor
mance. Thecostsharingheursti wasalsofounduse
fulfor beliefrevision inBayesnetworks 9). Here, we
generalizecost-sharing to apply tocyclic graphs, and
show that the resultingheristc i asoadmissble.
Thegeneralization ofthecost-sharing heuristic, while
straightforward, causes several problems. irst, the
cyclesntheBKBmaketheproblemofproperly dein
ing the heuristic nontrivial. If weustsed the same
deiningequations,thefact thatthereaecyleswould
make the dening equations cyclic. But by looking
a theseequations as a system ofequations, we state
hat a io to th sse is o heris. An
suchsoutin to the systemofeqations is shownto
be an admssible heuristic. A secnd problem ishow
tosoletheseequations. Thestandardtop-downago
ithmused inpio wo wouldbehnderedbythecy
cles: eveniwe conert i toakindofmessage-passing
updangalgorithm, in manycases the algorithmwil
oopindeinitely. nstead,weshowthatconertingthe
ystemofsemi-linearequationstoainearprogram, we
nvalt theheristi n polynomial time.
Webegin with a motivatigBB eamp, folowed
by a formal deniton of BKBs (secton 2). We then
reae BKBs toWAODAs,andreviewthecostsar-