From: FLAIRS-01 Proceedings. Copyright © 2001, AAAI (www.aaai.org). All rights reserved.
The benefits of multi-agent systems in spatial reasoning
Christian Kray
GermanResearch Center for Artificial Intelligence (DFKI) GmbH,Stuhlsatzenhansweg
66123 Saarbrficken, Germany
kray@dfki.de
Abstract
This paperpresents argumentsfor the applicationof a multiagent approachto spatial reasoning, andit is shownhowspatial reasoningbenefits from an agent-basedimplementation.
In orderto achievethis, the basic processesin spatial reasoning are analyzedand mappedonto a multi-agentarchitecture.
Additionally, common
problemsin the context of real-world
applications are identified, and it is shownhowthese can be
addressedusing the proposedapproach.Furthermore,it is arguedthat a realization based on a multi-agentsystemoffers
advantagesin termsof extensibility andflexibility. Finally,
an exemplaryimplementationof a spatial reasoning system
withinan interactive tourist guideis presentedalongwithits
basic mechanismsand a short review of its performancein
real worldtasks.
Introduction
Spatial cognition is a central componentof humanintelligence, and it is involved in manyevery day tasks such as
wayfmding, locomotion, and vision. Spatial concepts are
also very important for communication,e.g. when describing spatial constellations to a listener so that he can identify a certain object. Manyapproaches for enabling systems to reason about space have been proposed, ranging
from strictly qualitative calculi (e.g. based on topological
relations (Egenhofer & Herring 1990)) to more quantitative
methods (e.g. using potential fields to computelinguistic
descriptions of spatial constellations (Gapp1994)).
There is a great number of possible applications (Cohn
1996) for these methodsranging from verbal interfaces for
the blind to autonomous robot navigation. Additionally,
the advent of affordable positioning systems (e. g. GPS,
Gallileo) and the widespread adoption of mobile devices
(such as mobile phones and PDAs)has sparked a great dernand for location-aware services, and consequently for spatial reasoning. While this has led to the development of
smart infrastructures for this kind of services ((Hohl et al.
1999), (Butz, Baus, & Krtiger 2000)) - facilitating the
dling of nomadicusers - there are still somemajor hurdles
that hinder the widespreadapplication of spatial reasoning.
On one hand, the computations that have to be performed to evaluate spatial constellations and to arrive at
Copyright(g) 2001, AmericanAssociationfor Artificial Intelligence(www.aaai.org).
All rights reserved.
552
FLAIRS-2O01
somesuitable solution are highly complexand time consuming. Moreover, the fundamental data upon which the reasoning takes place is frequently providedin a low-level metrical
format (e. g. vector data in a GIS), while in manycases
symbolicor qualitative format is needed. Onthe other hand,
the computational methods have to be adapted to take into
account real-world characteristics (such as the current context). These points will be morethoroughly discussed in the
following section along with a possible solution using multiagent systems (MAS).
Theremainderof this article is structured as follows. The
next section shortly reviews the current links betweenspatial cognition and agent-based systems. After this, several
reasons are highlighted for the adoption of a multi-agent approach in spatial reasoning. The following section describes
an implementation of a spatial reasoning system along the
lines of the points madein the previous section. The final
section summarizesthe ideas presented here, and provides
someideas for future research directions.
Basingspatial reasoningon multi-agent
systems
In the context of spatial reasoning, the term agent is mostly
used to describe a system that performs a specific task such
as route advice (Rogers, Fiechter, &Langeley 1999) or incremental route instructions (MaaB, Bans, & Paul 1995).
Whilethese certainly meet the definition of an agent - which
Wooldridge describes as computer systems that are capable of autonomousaction in some environment in order to
meet their design objectives (Wooldridge 1999) - they are
not multi-agent systems per se. Multi-agent systems can
be defined as systems that are designed and implemented as
several interacting agents (Jennings, Sycara, & Wooldridge
1998). Although there are a few systems that take a decompositional approach to spatial cognition and exploit the
benefits of parallel or resource-aware computations (e. g.
(Blocher 1999)), the advantages of a true multi-agent system in spatial reasoning have not yet been thoroughly examined. In the following sections, several reasons are presented
for the application of multi-agent systems generally, as well
as arguments for modelingspatial reasoning using a multiagent approach.
¯ Frameof Reference
General qualifies of multi-agent systems
Several
relations,
especially
angular
relations
depend
on
Beforelookingat the specific advantagesin the contextof
theestablishment
ofa frame
ofreference
inorder
tobe
spatial-reasoning,somegeneralqualities of multi-agentsysnon.ambiguons.
Some
calculi
alsousetheframe
ofrefertemscan be identified. (Bond&Gasser1988)list several
ence
to scale space accordingly(Gapp1995).
rationales
for dis~’butingan AI system,such
asadaptabil¯ Route Segmentation
ity,cost-effectiveness, and improvements
in the developIn the case of reasoningin the context of navigationor
mentand management
process. In addition, they point out
route descriptions,a complexroute mighthaveto be ditl~theinherent
isolation/autonomy
oftheparts
provides
videdinto severalsegments
to allowfor properreasoning.
notonly
protection
forlocal
information,
butmayalso
be
a more’natural’
waytoaddress
certain
problems.
Further- Theseprocessesdo haveto interact heavilyin orderto solve
more,theyarguethat the distributionfacilitates spacializaspatial problems.Consider,for example,the selection of
lion, andmayincreasethe reliability, robustness,and/oreffia referenceobject. Toidentify the mostappropriateone for
~iency
oftheentire
system.
Finally,they
state
that
resource the taskat hand,static charact~’istics
are oftennotsufficient.
limitations
canbebandied
bothonanindividual,
andona
Therelation chosento go alongwiththe referenceobject as
larger
scale.
Butbefore
these
benefits
canbeexploited
in
well as whetherit has beenmentioned
beforecontribute to
spatial
reasoning,
itisnecessary
toanalyze
howthereason- its overallrating. Theprocessof selecting a frameof refingprocess
canbemapped
ontoagents.
¯ rc~cealso r~uiresinput fromother processes,as the current referenceobjector the actuallocationof the user havea
Mappingprocesses onto agents
great influenceon the outcome
of the selectionprocess.The
When
lookingat the different computational
modelsfor spasegm~tationof a complexroute is another exa~le of the
intensiveinteractionbetween
these coreprocesses:a chosen
tial reasoning((Cohn1996),(’Mukerjee
1998)),onecan dividethe reasoningprocessinto severaldistinct butinteractsubdivisiondependsnot onlyon characteristicsof the route,
ing sub-processes.Manyproblemsrelated to spatial conbut is also determined
bythe qualityof relationsthat canbe
stellations call for the evaluationof relations of somekind
establishedfor the resultingparts, andbythe context(e.g, at
betweensomeobjects, or oonv~sely,for the mapping
of rewhatspeeda user is traveling).
lations to a specific constellation.I)q~ndingon the set of
ModeLing
these interactingprocessesas agentsoff~s serrelationsin question,there mayalso be the needto establish
ial advantages
overa monolithic
architecture. Firstly,
the
a frameof reference(e.g. whenconsideringangularrelasubproc~sscan be clearly separated, whichallowsfor the
tions). Inthe context
ofnavigation,
additional
processes
can
concentrationon the correspondingfundamentalreasoning
beidentified.
Inorder
tomason
about
routes,
itoRen
makes methodswithouthavingto take into accountrequiremems
of
sense to divide theminto segments,whichcan be mapped
other subprocesses
fromthe start, Secondly,the interaction
between
tl"asn is madeexplicit, whichnot onlysimplifiesits
ontoqualitative descriptions(Tversky&Lee1999).In this
case, a different bodyof relations- whichmilesto somedeformalization,but also makesit easyto changethe interacgre~ on the shapeof the objects involved((Kray&Blocher
tion patterns (mendynamically)
andfacilitates the realiza1999),(Mathet2000))- canalso be taken into accoum.
tion ofmeta.l~/elreasoning,Thirdly,the agentsare l~c¢ to
tlre 1 tries to summarize
the precedentreflectionsbydepictdecidewhattheywantto donext. Thisenablesthorn, e. g. to
ing thosefundamental
processesin the center part:
reasonaboutpendingrequestsfromother agents, and cancel
obsoletejobs (suchas the evaluationof a mediocre
reference
¯ Two-PointRelations
objcg~tif there is alreadya better one)or changethe orderof
Thisincludestbe evaluationof topological,angular
distance.dependentrelations, whichshare the common the processingofth~jobs (e,g. if informationrequiredby
later job is alreadyavailable).FinaUy,
it is a straightforward
characteristicsthat they- in their mostbasicform-relate
1.
task to extend and modifysuch a multi-agent system. B~
twoobjectsto eachother
cause
the subprocesses
arestrictly separated,their internal
¯ N-PolntRelations
reasoningmechanisms
can be exchang~without disturbing
Relationsthat cannotbe reducedto a two.point-problem, the rest of the system.Addingnewagents, that modeldifsuch as path relations (Kray&Blocher1999), special
ferent processes,canalso be achievedrathereasily, since the
caseslike inbetween,or constellationalrelations (Mathet
modifications
requiredto let existingagentsbenefitfromthe
2000),are qualitativelydifferent fromtwo-pointrelations.
newfunctionalityare cle~arlyencapsulated
withinthem.In a
They
require
more
than
twoobjects,
and/or
additiortal
armonolithicarchitecture, isolating the placeswherechanges
guments
suchas shapeor outline information.
haveto be appliedor changingthe implicit interaction can
be tediousanderror prone.
¯ ReferenceObject
Common
tasks in the realmof spatial reasoningsuch as
Challenges posed by real-world applications
localizationor object identificationrequirethe selection
era suitablereferenceobject(or anchorobject)froma set
Whilethese advantagesapplyto almostany spatial reasonofpossiblecandidutes.
ing task, they do so especiallyin the contextof real-world
uses. Whena user employsa systemto solve a specific
~Two
pointscanbe seenas the mostbasic entities, between
task, such as q~ryinga geographicalinformationsystem
whicha relation canbe established(Kray&Blocher1999).hence
(OlS)to obtainsomeinformation,or askingfor navigational
thename.
Figure1: Processes,factors, andagents in spatial reasoning
instructions, the computationalcomplexityof mostspatial
reasoning calculi (Cohn1997) can becomea big obstacle.
Althoughthere are methods,whichcan be solved in polynomialtime, the great numberof possible arguments(e.g.
the numberof objects within a GIS)and the amountof time
to deducequalitative informationfromthe mostly metrical
data (e.g. performingset operations on bitmapbaseddata)
require the applicationof heuristics, or at least someformof
adaption strategy (such as anytimealgorithms(Zilberstein
1993)),in order to keepanswertimes in a tolerable range.
A multi-agent approachhelps to address these problems.
Onone hand, the freedomof decision that each agent has,
allowsfor specific adaptationstrategies and heuristics to be
applied to the specific subproblemshandledby the agent.
Onethe other hand,it is an inherentability of such systems
to allowagents to duplicate themselves,or to moveto a different site should they wantto do so. This offers newways
to addressthe needfor a timely solution: in case there is
a large amountof informationto be considered,several instances of a specific agent can concurrentlyworkon small
subsets. If the computationalpowerof the current platform
is insufficient to solve the task in time, an agent candecide
to moveto a morepowerfulsite. Finally, the messagepassing mechanism,whichis used for the inter-agent communication, makesit easy to realize a transactional paradigm:
instead of waitingfor the completionof a requestsent to another agent, a specific agentmaystart working
on a different
task until it reachesa point whereit again needsexternalinformation.After sendingoffthe correspondingrequest, the
agent cancheckwhetherthere is already a reply to the first
request.If that’s the case, it cancontinueto workonthe orig-
SS4
FLAIRS-2001
inal task, therebyexploiting the ’waiting’ time, whichwould
havebeenwastedotherwise.Thesestrategies canalso be applied concurrently,to a single basic agentor to severalones,
dependingon the task at hand.
Furthermore,in a real-world context a great numberof
factors, that are usuallyleft out of spatial reasoning,haveto
be included in the reasoningprocess. Figure 1 showsfour
fields, whichplay an importantrole in a real-worldapplication: geometry,contextualfactors, languages,and resource
restrictions. Whilegeometrical information in someform
is present in almostany systemconcernedwith spatial cognition, there are somefactors that are especially important
in a real worldscenario, suchas the visible characteristics
of entities in the worldor the level of abstraction,at which
these entities are evaluated. Thesefactors, or the reasoning/computationto extract themfrom the geometryof the
world model,can again be encapsulated in agents, inheriting mostof the benefits listed aboveand simplifyingthe
interaction withthe core processesdescribedin the previous
section. Thesameis true for contextual factors, whichare
particularly important in real-world applications. Knowledge about, e.g. which meansof transportation is used
(MAX,
Bans, &Paul 1995), what scale the user is referring to (Montello1993),or whatfunctiona certain building
has (Lynch1960)strongly influence the outcomeof the reasoning process. Whilethe importanceof resource considerations havealready beendiscussed, the relevanceof the
languageswhichare used to informthe user of the results
from the reasoning wasnot. But the fact, e.g. whetherto
realize the outputgraphicallyor verbally, andwhatspecific
(natural) languageor symbolsare selected can changethe
2.
outcomeof the reasoning
This list of relevant factors for real-world spatial reasoning is certainly neither complete, nor has the exact way, in
which they influence the reasoning process, been empirically examined. But building a system that includes these
factors using a multi-agent architecture allows not only for
a clear separation but also for continuous improvements:
Whennew factors are identified, they can be added as an
agent, and their influence on the other processes can be incorporated easily by first analyzing it individually for each
process, and then establishing the interaction. Modifications
to existing processes are on a local base without the need to
search for all places, where changes would have to be made
in a monolithic system. Furthermore, once empirical results
becomeavailable about the wayin which certain factors influence the reasoning processes, they can be implemented
in a straightforward way. Either by simply changing the
explicit interaction patterns, or by additional local modifications to the affected processes. In both cases, unaffected
processes need not to be altered. These characteristics have
been very beneficial in the developmentof the spatial reasoning system presented in the following section.
An exemplary implementation
The ideas described in the previous section have led to the
3, which imrealization of a spatial reasoning system SPACE
plements manyof the concepts presented here. It is part of a
mobile tourist guide (DEEP MAP(Malaka & Zipf 2000)),
whereit is responsible for various tasks involving spatial
problems, such as navigational instructions or queries with
spatial constraints. Not only is SPACE
part of a multi-agent
system (the mobile tourist guide), but it also consists
agents itself 4. The internal agents communicatewith each
other over an internal messagebus, and can also send messages to external agents via the external messagebuss. They
modelthe basic processes identified in the previous section,
and the spatial reasoning tasks, whicharise in the context of
the tourist guide (see figure 2).
Thesetasks include the identification (Identify) of objects
in the environment of the user, and description (Describe)
of objects the user mentions, e.g. whenreplying to queries
like ’What’s this?’ or ’Tell me more about this building’.
SPACE
can also generate localizations tbr real-world entities using relational expression (Localize) and translate such
expressions into geometrical regions (TransToGeo).Finally,
there is an agent, that handles the incremental generation of
navigational instructions (RouteManager),and a scheduling
agent (Scheduler), which analyzes incoming jobs and distributes them amongthe internal agents.
During the two years of development, SPACEhas been
enhanced and extended frequently. As new information
2For example,somerelations might not be easily express~le
verbally or graphically.
3Spatial CognitionEngine
4Thisarchitecture is often called holonic(Fischer 1999).
5Thisdesign ensuresthat internal messagetraffic is kept from
the externalbus, andexternal agentscan addressa single agent for
all spatial problems.
Deep Map
~
.
.....................
Scheduler
Sp;CE
t
BestRO
]
PathRelations
]
Segmentation
]
Relations
]
RefSys
]
Figure 2: Architecture of SPACE
sources/agents became available, the internal agents were
adjusted to include them in their reasoning. For example,
when an agent for the managementof the dialog history
was introduced, the BestROagent was modified to take into
account whether an object had been previously mentioned
when rating objects. The other agents did not have to be
adjusted in order to benefit from this additional information. Adding an internal agent was also a straightforward
task: Whenan agent for a hotel reservation system was introduced, the need arose for the translation of relational expressions (e. g. ’close to the station’) to geometricalregions.
TransToGeowas added, and did make use of the other existing agents without altering the system. The interaction
between the agents is very intense, and would be very hard
to handle if it were implicit. For example, when the BestROagent evaluates an object, it contacts the GIS, the Dialog History, the Context Manager, and all other internal
agents except Segmentation, and may engage in a message
exchangewith several of them. Thesame is true for the other
agents of the system.
SPACEas it stands now can handle the tasks mentioned
above within the city of Heidelberg (Germany) in a timely
fashion. It relies on a GIS, whichcontains a detailed model
of the entire city and its surroundings. The reasoning process takes into account information from a user model(e. g.
age, knowledge about environment, physical constitution)
and a context model (e. g. weather, means of transportation). The system also considers the user’s current position
(obtained from a GPSreceiver) and facts from the dialog
history. SPACE
has been successfully used in a mobile prototype around the castle of Heidelberg. Currently, the system
is being ported to a newagent platform, that offers enhanced
resource monitoring and adaptation facilities, and a newversion of the mobileprototype for systematic tests with casual
users is prepared.
SPATIOTEMPORAL
REASONING555
Conclusion and Future Directions
Thispaper presenteda newapproachto the modelingo f spatial reasoningprocessesusing multi-agentsystems. It was
shown,that using the agent paradigmcan be very beneficial
in several aspects. Onone hand,several basic processescan
be identified and mapped
directly ontoagents, not only simplifying the modelby encapsulatingfunctionality, but also
makingthe interaction betweenthose processesexplicit. On
the other hand,the multi-agentapproachfacilitates the modification andextensionof a spatial reasoningsystem,and is
particularly well suited in the contextof a real-worldapplication since manyadditional factors haveto be taken into
account.
In the future, severalresearchdirectionswill be followed.
Firstly, the interaction betweenthe agents will be analyzed
morethoroughlyin order to formalizeand optimizeit. Secondly,a morerigid evaluationof the entire systemis planned
throughsystematic field tests with casual users. Thirdly,
a componentfor user interaction about uncertain spatial
knowledge
(e.g. to determinethe user’s current location in
case of sensor failure by meansof a dialogue)is currently
under development.Finally, additional resourceadaptation
strategies will be examinedto further improvethe average
responsetime.
Acknowledgements
DeepMap,SpaCE,and the results presented in this paper
are fundedby the Klaus Tschira Foundation, Heidelberg.
SpaCEis co-fundedby the DFKI.
References
Blocher, A. 1999. RessourcenadaptierendeRaumbeschreibung: Ein beschrdnla-optimaler Lokalisationsagent
(resource-adaptingspatial descriptions: a boundedoptimal localization agenO.Ph.D. Dissertation, Computer
ScienceDivision, University of Saarland, Germany.
Bond,A. H., and Gasser, H. 1988. Ananalysis of problems
and research in dai. In Bond,A. H., and Gasser,H., eds.,
Readingsin distributed artificial intelligence. SanManteo,
CA: Morgan Kanfmann.
Butz, A.; Baus, J., and Krtiger, A. 2000. Augmenting
buildings with infrared information. In Proceedingsof
the International Symposiumon AugmentedReality ISAR
2000. Los Alamitos, CA: IEEEComputerSociety Press.
Colin,A. G.1996.Calculifor qualitative spatial reasoning.
In Calmet,J.; Campbell,
L A.; and Pfalzgraf, L, eds., Artifical Intelligence and SymbolicMathematicalComputation
(LNCS1138). Berlin: Springer. 124-143.
Cohn,A. G. 1997. Qualitative spatial representation and
reasoningtechniques. In Brewka,G., Habel,C.; and Nebel,
B., eds., 1(,1-97(LNA11303).
Berlin: Springer.
Egenhofer,M. J., and Herring, J. R. 1990. Amathematical
frameworkfor the definition of topological relations. In
Brassel, K., and Kishimoto,H., eds., FourthInternational
Symposiumon Spatial Data Hadling.
556
FLAIRS-2001
Fischer, K. 1999. Agent-BasedDesign of Holonic Manufacturing Systems. Journal of Robotics and Autonomous
Systems.
Gapp,K.-P. 1994. Basic meaningsof spatial relations:
Computation
and evaluation in 3d space. In Proceedingsof
AAAI-94.
Gapp,K.-P. 1995. Processingspatial relations in object
localization tasks. In AAAIFall Symposium
on "Computational Modelsfor In tegratingLanguage
and~qsion"’.
Hohl, F.; Kubach,U.; Leonhardi, A.; Rothermel,K.; and
Schwehrn,M. 1999. Next century challenges: Nexus- an
openglobalinfrastructurefor spatial-awareapplications.In
Proc. of Mobicom99. Seattle, Washington:ACM.
Jennings, N.; Sycara, K.; and Wooldridge,M. 1998. A
roadmapof agent research and development.Journal of
AutonomousAgents and Multi-Agent Systems 1:275-306.
Kray, C., and Blocher, A. 1999. Modelingthe basic meanings of path relations. In Proceedingsof the 16th IJCAL
MorganKaufmann,San Francisco, CA, 384-389.
Lynch, K. 1960. The image of the city. Cambridge,MA:
MITPress.
MaaB,W.; Bans, J.; and Paul, J. 1995. Visual grounding
of route descriptions in dynamicenvironments.In AAAI
Fall Symposiumon ComputationalModelsfor Integrating
Languageand lqsion. MIT, Cambridge,MA:AAAI.
Malaka, R., and Zipf, A. 2000. Deep Map- challenging IT research in the framework
of a tourist information
system. In Fesenmaier, D. R.; Klein, S.; and Buhalis,
D., eds., Informationand communication
technologies in
tourism 2000. Wien:Springer. 15-27.
Mathet, Y. 2000. Newparadigmsin space and motion: A
modeland an experiment.In Rodriguez,R. V., ed., Workshop on Current lssues in Spatio-TemporalReasoningat
ECA12000.
Montello, D. 1993. Scale and multiple psychologies of
space. In Frank, A. U., and Campari,I., eds., Spatial Information Theory: A Theoretical Basis for GIS. Berlin:
Springer-Verlag. 312-321.
Mukerjee,A. 1998. Neat vs scruffy: A survey of computational modelsfor spatial expressions.In Olivier, P., and
Gapp,K.-P., eds., Computational
RepresentationandProcessing of Spatial Expressions.KluwerAcademic
Press.
Rogers, S.; Fiechter, C.-N.; and Langeley, P. 1999. An
adaptive interactive agent for route advice. In Proceedings of AutonomousAgents ’99, 198-205. NewYork, New
York: ACM
Press.
Tversky, B., and Lee, P. U. 1999. Pictorial and verbal
tools for conveyingroutes. In Spatial InformationTheory
(Proceedingsof COSIT99). Springer, Berlin, 51-64.
Wooldridge,
M.1999.Intelligent agents. In Weiss,G., ed.,
Multiagent Systems - A ModernApproachto Distributed
Artificial Intelligence. TheMITPress. 21-35.
Zilberstein, S. 1993. Operational Rationality through
Compilationof AnytimeAlgorithms. Ph.D. Dissertation,
ComputerScience Division, University of California at
Berkeley.