198- advanced decision architectures for the warfighter.pdf
TRANSCRIPT
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
1/557
E D I T E D B Y PAT RI C I A M C D E RM O T T AN D LAU RE L ALLE N D E R
ADVANCED DECISI ON ARCHITEC TURES
FOR THE WARFI GHTE R:F O U N D AT I O N S A N D T E C H N O L O G Y
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
2/557
i
A D VA N C E D D E C I S I O N A R C H I T E C T U R E S
F O R T H E WA R F I G H T E R :
F O U N D AT I O N S A N D T E C H N O L O G Y
E D I T E D B Y PAT R I C I A M C D E R M O T T
A N D L A U R E L A L L E N D E R
This book, produced by the Partners of the
Army Research Laboratory Advanced Decision Architectures
Collaborative Technology Alliance, also serves as the
Final Report for DAAD19-01-2-0009.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
3/557
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
4/557
SECTIONI.COLLECTING,
PROCESSING,
AND
DISTRIBUTING
BATTLEFIELD
INFORMATION
SECTIONI
COLLECTING,PROCESSING,ANDDISTRIBUTINGBATTLEFIELD
INFORMATION
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
5/557
14 SECTIONI
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
6/557
15
Chapter1
1.
DYNAMICNETWORKANALYSISAPPLIEDTOEXPERIMENTS
FROMTHEDECISIONARCHITECTURESRESEARCHENVIRONMENT
KathleenM.Carley,Ph.D.
CarnegieMellonUniversity,Pittsburgh,PAMichaelK.Martin,Ph.D.
CarnegieMellonUniversity,Pittsburgh,PAJohnP.Hancock
ArtisTech,Inc.,Fairfax,VA
INTRODUCTION
ADACTAresearchisproducingexperimentalresultsfromsimulations
ofintermeshednetworksofwarfightersandbattlefieldsurveillance
assets.Thesenetworksformacomplexsystemwithbehaviorsthat
emergefrompatternsofinteractionamongconstituententities.Thesim
ulatedinteractionsarespatiallysituated,temporallydistributedcommu
nicationsamongpeople,robots,andsoftwareagents.Ingeneralterms,
thecomplexsystempartsofwhichweaddressinthispapercanbe
conceptualizedasatwolevelmetanetworkthatincludesinteractions
amonghumanagentsatonelevel,interactionsamongartificialagentsat
anotherlevel,andcrosslevelinteractionsbetweenhumanandartificial
agents.
TheissueaddressedinthispaperisDynamicNetworkAnalysis(DNA)
ofsystembehavior.Forthepurposesofthischapter,wearenotinter
estedinexaminingtheperformanceofonesystemrelativetoanother.
Althoughthis
type
of
comparative
analysis
can
be
useful
for
researchers
orsystemdesigners,itisofquestionableusetowarfighters.Instead,we
areinterestedinanalysesthatproducetacticallyrelevant,actionable
resultsthathighlightthestrengthsandweaknessesinthesystembeing
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
7/557
16 CARLEY,MARTIN,ANDHANCOCK
observed.Tobeuseful,theanalyticalresultsmustfostertacticalinsight
andstimulatebattlefielddecisionsthatprudentlyinfluencefuturesys
tembehavior.
Tothisend,wedescribetwocasestudiesthatapplyDNAtothesim
ulatedbattlefield
data
being
generated
by
experiments
in
the
Decision
ArchitecturesResearchEnvironment(DARE).Thefirstcaseinvolvesinter
ceptsofsimulatedcommunicationsamonghumanagents,whichwe
frameasanexerciseinadversarialreasoning.Thesecondcaseinvolves
simulatedcommunicationsamongsurveillanceassets(i.e.,software
agentsandrobots),whichweframeasanexerciseinunderstandingthe
automatedcontrolofaPersistentCoordinatedVideoSurveillance(PCVS)
system.Together,webelievethecasestudiesdemonstratehowDNA
(Carley,2002)
can
foster
tactical
insight
in
complex
multi
entity
scenar
ios.TheyalsodemonstratethatthecombinationofDNAtechniques
requiredfortacticalinsightmayvaryaccordingtothetypeofnetwork
beinganalyzed.Finally,theyshowhowDNAassistsinthedevelopment,
understandingandtuningofsoftwareagentsystems.
ANALYTICALAPPROACHOneapproachtoanalyzingsystembehavioremploysmainstream
statisticaltechniquesonsummarymeasuresofperformance(e.g.,percentoftargetstracked,percentofprioritytargetstrackedwithaccept
ableaccuracy,trackcorrelation,etc.).Thistypeofanalysis,however,
doesnotfullyexploittheinformationgeneratedinDAREexperiments
orbybattlefieldsurveillanceingeneral.Moretothepoint,itprovideslit
tleinsightintotheidentificationofleversinthenetworksunderlyingsys
tembehavior.Onceidentified,leverscanbeusedtodetermineactionsto
taketoinfluencefuturesystembehavior.
DNAprovides
an
alternative
analytical
approach
that
compliments
mainstreamstatistics.WithDNA,thefocusshiftsfromaggregatemea
suresofperformanceforacollectionofbattlefieldentitiestotheperfor
manceimpliedbythestructureofrelationsamongbattlefieldentities.
Thisshiftistheessenceofwhatitmeanstoviewthebattlefieldfroma
networkscienceperspective.Thatis,fromanetworkscienceperspective
wearelessconcernedwithwhatisnormal(e.g.,averages,dispersions)
aboutentitiesinthebattlefield(e.g.,people,places)andmorecon
cernedabout
detecting
substantive
patterns
in
the
observed
relations
amongentities.TheemphasisonthestructureofrelationsinDNAmakes
itparticularlywellsuitedtothedetectionofanomaliesandexceptions
(e.g.,centralities,exclusivities)theleverswithpotentiallylargeinflu
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
8/557
DynamicNetworkAnalysis 17encesonsystembehavior.DNA,therefore,fostersscrutinyofstrengths
andvulnerabilitiesintherelationsamongbattlefieldentities(i.e.,the
observedsystem).WithaDNAmodel,wecanidentify(amongother
things)implicitgroupsofentities,keypeopleandlocations,andopera
tionallysignificant
time
frames.
We
can
even
begin
to
infer
relations
amongentitieswherenonehavebeenobserved.
Networksciencehasbeenhamperedhistoricallybyadominantly
socialperspectivefocusingonwhointeractswithwhom.However,Carley
(2002)arguedthatthesesocialnetworksexistwithinanecologyofnet
worksthatcanusefullybecharacterizedintermsofthedynamicsofthe
relationsamongthewho,what,where,how,andwhy.Thisisknownas
themetanetworkperspective.*ORA(e.g.,Carley,Columbus,DeReno,
Reminga&
Moon,
2008)
is
adynamic
network
analysis
package
that
can
beusedtoassessmultimode,multiplexnetworks;identifykeyplayers,
groupsandvulnerabilities;enablecomparisonoftwoormorenetworks;
andfacilitatereasoningaboutspatiotemporalnetworks.
*ORAsupportsanalysisofdynamicnetworksinmanyways: (1)com
parisonsoftemporallyorderedsnapshotsofstaticnetworks,(2)statisti
calchangedetectiononsequencesofnetworks,(3)trailanalysisfortrail
dataandconversionoftraildatatonetworks,(4)simulationofchangein
networks,and(5)comparativestaticsforimmediateimpactassessment.
Herein,wemakeuseofstatisticalchangedetectionandtrailanalysis,
alongwith*ORAsvisualizationcapabilities.
CASESTUDY1:TERRORISTSINADELPHIThescenarioforcasestudy1wasframedasinterceptsofsimulated
communicationsamongagentsrepresentingterroristsandnoncomba
tants(e.g.,pizzadeliveryguys).Theagentscommunicatedviaphoneand
emailas
they
moved
about
the
Adelphi
region.
DataGenerationThedataweregeneratedbyArtisTechsAlgoLinksimulator.The
AlgoLinksimulatorwasoriginallydevelopedtotestthecapabilitiesof
messageanalysistoolstosupportintelligenceanalysisrequirementsin
battlefieldcommunications.AlgoLink(seeFigure1.1)facilitatescustom
constructionof
entity
networks
that
follow
specified
communication
structures,times,places,durations,andbehaviors.Itgeneratesboththe
foreground(networkofinterest)andthebackgroundcommunicationas
specifiedandstitchesthemtogetherintoasinglecommunicationrecord.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
9/557
18 CARLEY,MARTIN,ANDHANCOCK
Backgroundcommunicationisbothstructured(e.g.intermingledhierar
chicalandsocialorganizations)andrandomasspecifiedbythehuman
simulationoperator.AlgoLinkusesrealisticcommunicationandorganiza
tionaldata,timing,andmorphologiesbutcontainsnoinformationabout
anyreal
person
or
organization.
Figure1.1TheAlgoLinkmessagesimulatorinterfaceThedatasetcreatedforthisexperimentwasgeneratedbyArtisTech
stafflookingforwardtosystembasedIntelligentAgentcommunication
behavioranalysis.Theentitiesofinterestwereorganizedintoasmall
numberofcellsthatwereuniquelyconnectedandstitchedintoa
largerbackgroundcommunity. Thedatasetwasgeographicallycen
teredontheARLAdelphicampusbecausetheARTEMISprojectiscoordi
natedwiththeARLComputationandInformationSciencesDirectorateresearchsystems.Thesimulatedsocialstructurewasmorphologically
similartoamediumsizedcommunityofintelligentagentsactingwitha
specificpurposeinabattlefieldCommand,Control,andCommunications
(C3)system.Tosimulateaneventthatstimulatesthecommunication
network,aspikeincommunicationvolumewasinsertedataselected
time.TheAlgoLinkgenerateddatasetwasarapidwaytoassessthefea
sibilityofcollaborationbetweentheARTEMISPCVSandCASOSteams.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
10/557
DynamicNetworkAnalysis 19Analysis
TheAlgoLinkoutputwasdeliveredasanXMLfilecontaininga
sequenceofcommunicationsrecords.Eachsimulatedcommunication
recordidentified
the
sender,
the
receiver,
the
time
the
communication
occurred,itsduration,andwhetheritscontentwasoperationallyrele
vant,irrelevant,orambiguous.Italsocontainedlatitudeandlongitude
forthepositionofthemobilesendersandreceiversduringeachcommu
nication.
Ouranalysisstrategyinthiscasecanbegenerallydescribedasan
overviewandzoom.Thatis,wefirstexaminedthegeneralcontextof
communicationsactivities,andthendrilleddowntodetermineimpor
tantagents,
time
frames,
and
locations.
The
subset
of
*ORA
capabilities
thatprovedparticularlyusefulhereincludedgeospatialvisualization,key
playeridentification,changedetectionanalysis,andthecorrelationof
standardnetworkandgeospatialvisualizations.
Using*ORAsgeospatialvisualizationcapabilities(e.g.,Davis,Olson,
&Carley,2008)weexploitedthepresenceoftimevarying,geolocated
attributesoftheinterceptedcommunicationstodiscoverthescenario
involvedsuspiciousentitiesfleeingtheAdelphiarea(seeFigure1.2).
Figure1.2AgentsfleeingAdelphioverthetimecourseofthescenario
Usingafuzzygroupclusteringtechnique,FOG(e.g.,Davis&Carley,
2008),wefoundthatthesuspiciousentitieswereorganizedintofive
groupswithsharedmembers(seeFigure1.3).Theinterstitialmembers
arelikelytocontainthecoordinatorsandleaders.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
11/557
20 CARLEY,MARTIN,ANDHANCOCK
Figure1.3Agentsorganizedin5groupsthatsharemembers
Todrilldown,wefirstused*ORAsKeyEntityReporttoidentifythe
threeagentsmostcriticaltooperations(seeFigure1.4).Becausethe
datawereaboutcommunications,twodifferentcentralitymeasures
wereuseddegreecentralityandbetweennesscentrality.Degreecen
tralitymeasureswhoisconnectedtomostothers(i.e.,theactormost
likelytobeintheknow).Betweennesscentralitymeasureswhois
mostlikelytobeonallthepathsbywhichinformationflows(i.e.,the
actormostlikelytobeinfluential).Thisenablednarrowingourfocustoa
smallgroupofleadersinsteadoffocusingonthesetofinterstitialmembers.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
12/557
DynamicNetworkAnalysis 21
Figure1.4KeyEntityReportidentifies3importantagents
Wenext
asked,
is
something
happening?
One
way
of
answering
this
istoseewhetherthereisachangeinstandardbehavior.Usingthe
ChangeDetectionReport(seeFigure1.5),weidentifiedperiod3asthe
timeframeinwhichoperationsmostlikelyoccurred.
Figure1.5ChangeDetectionReportsignalstimeperiod3isdifferent
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
13/557
22 CARLEY,MARTIN,ANDHANCOCK
Thenetworkchangedetectionanalysis(McCulloh&Carley,2008;
2009;McCulloh,Webb,Graham,Carley&Horn,2008)extendschange
detectionfromoperationsresearch,whereithasbeenusedonvariable
leveldata,
to
relational
data.
This
is
astatistical
approach
for
detecting
smallpersistentchangesinorganizationalbehaviorovertimeusingsta
tisticalprocesscontroltechniquesappliedtonetworksummarystatistics.
Period2wasthepointatwhichorganizationalbehaviorchanged,leading
toradicaldifferencebyperiod3.Thisappearstohavebeenaplanning
executionphaseshift.
Examinationofindividuallevelmetricsforthethreekeyplayersand
networklevelmetrics(e.g.,centrality,betweenness,efficiency,connect
edness)corroborated
our
interpretation
that
period
3was
operationally
significant.Individuallevelmetricsindicatedthatactor286engagedin
extensivecoordinationatperiod2,passingthereignsofcontroltoactor
652atperiod4(seeFigure1.6).
Figure1.6Individuallevelmetricsconvergeonperiod3
Examinationofnetworklevelmetrics(seeFigure1.7)showedthatthe
groupwasgenerallyaverydistributedstructurethatcoordinatedintoa
centrallycontrolled,moreefficient,unitatperiod3.Then,itwentback
toits
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
14/557
DynamicNetworkAnalysis 23
Figure1.7Networklevelmetricsconvergeonperiod3
standardformandbecamemorehiddeninthesociocommunication
environment.
Havingidentifiedthekeyplayersandtimeperiod,wefocusedour
analysesondiscoveringwhatmayhavehappened.Examinationofthe
agentxlocationnetworkforperiod3(seeFigure1.8)indicatedalarge
clusterofsuspiciousentitiesintheAdelphiarea(includingkeyplayer,
Agent286),afairlylargeclusterofsuspiciousentities(includinganother
Figure1.8Agent
xLocation
Network
for
period
3
keyplayer,Agent97)inwhatappearstobeastagingarea,anapparent
waypointbetweenthestagingareaandtheclusterofsuspiciousentities
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
15/557
24 CARLEY,MARTIN,ANDHANCOCK
inAdelphi(alsovisitedbyAgent97,whoseemstobealiaison),anda
runner(keyplayer,Agent652)whovisitsmanylocationswithfewsuspi
ciousentitiespresent.
Anotherkeyadvance,developedaspartoftheCTA,wasthecapabil
ityto
move
between
trail
data
who
was
where
when
and
networks.
Whenweexaminedthetrailsvisualizationforperiod4(i.e.,theperiod
immediatelyfollowingtheapparentoperation,perhapsaperiodofinitial
surveillance),wesawthatthethreekeyplayerswereneverinthesame
placeatthetime;Agent652wasagainrunning,whereastheactivitiesof
Agents97and286wererestrictedtooneortwoareas.Theapparent
coordinationhandofffromAgent286toAgent652isrelatedto652s
increaseinspatialmovementandcoordinationneededduetoincreased
movement.In
the
trails
visualization
(see
Figure
1.9),
time
progresses
downtheyaxis.Geographicregionsformverticalbinsalongthexaxis.
Arrowsareplottedasagentsmovefromregiontoregion(orwithin
regions),andinthiscasearecolorcodedtothethreekeyplayers.
Figure1.9*ORALoomvisualizationoftrailsfor3keyplayersduringperiod4
Finally,correlatingastandardagentxlocationnetworkvisualization
withageospatialvisualizationfortheendofthescenariowefoundthat
Agent286wasalonewithasinglemovementbetweentwolocationsin
Adelphi,Agent97washoledupwithasizablegroupofsuspiciousenti
tiesnorth
of
Adelphi,
and
Agent
652
was
alone
but
on
the
run
(see
Figure
1.10).
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
16/557
DynamicNetworkAnalysis 25
Figure1.10Correlationofagentxlocationnetworkandgeographiclocationatsce
narioend
Giventhepatternofcommunicationsandmovementduringthesce
nario,twocoursesofactionsappearreasonable: (1)scourAdelphifora
bomb,IED,etc.plantedduringoperationsinperiod3,or(2)goafterdis
persedsuspiciousentities.Withrespecttoaction2,Agent286maybean
easytarget
with
direct
knowledge
of
the
operations
that
occurred
during
period3.TargetingthelocationwhereAgent97ishiding,however,will
yieldmoresuspiciousentities.Notethatthesimulateddatadoesnot
includecommunicationscontentsospecificationoftheeventisnotpos
sible,buttheappliedDNAanalysisaccuratelyidentifiedthetime,place,
andleadentitiesinthesimulation.
CASESTUDY2:MOVEMENTINTHEPERIMETERThescenarioforthiscasecenteredonanautomatedsurveillance
system(i.e.,theARTEMISPCVSsystem)thatisresponsibleforidentifying
movingentitieswithintheperimeterofaBlueForce researchcom
pound.Thesystemdividestheperimeterintofourareasofresponsibility,
whereeachareaisassignedtoaTaskingAgentthatisresponsiblefor
surveillance(seeFigure1.11).TaskingAgentsgetsimulatedmovement
reportsfromsimplevideoanalysisalgorithms.TaskingAgentsthenprior
itize
targets
and
assign
mobile
robotic
assets
to
pursue
and
identify
the
targets.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
17/557
26 CARLEY,MARTIN,ANDHANCOCK
Figure1.11Tasking
Agent
Areas
DataGenerationThedataweregeneratedbyArtisTechsARTEMISPCVSsystemproto
type(seeFigure1.12).Thisprototypeuseshundredsofsmallreasoning
Figure1.123DViewofARTEMISPCVSscenarioontheGISARLAdelphitestbed
model
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
18/557
DynamicNetworkAnalysis 27algorithmsencapsulatedinsoftwareagents.Theagentscommunicate
withhumans,agents,andothersystemelementsandevencreateand
deleteotheragentswithfrequenciesthataredependentontheirreac
tionstothesensedenvironment(simulation).
Theparticular
experimental
run
analyzed
here
was
conducted
to
determinewhetherTaskingAgentreasoningandcommunicationabout
sharingmobileroboticassetswasfunctioningasexpected.Thescenario
includedashorttimeofquiescenceinthecompound,followedbythe
injectionofarelativelylargenumberofmovingtargetsthatmoved
aboutthecompoundusingreasonablepathsbutfreelycrossingareasof
surveillanceresponsibility.Whentargetscrossareasofresponsibilitythe
complexityofreasoningincreasesandrequiresthatTaskingAgentstrans
fer(handoff)
tracking
and
even
possibly
lend
robotic
assets
to
other
TaskingAgents.Theactoflendingtheassetinvolvescommunicationto
notifyanadjacentTaskingAgentofanincomingunidentifiedentity,
andamultimessageTaskerGlobalTaskerhandshaketotransfercon
trol.Thereweretwoofthesehandoffeventsinthescenario.Thissce
nariowasasimpleonetofacilitateearlycollaborationbetweenARTEMIS
andCASOSstaff.
Analysis
Aswithcasestudy1,thedataweredeliveredasanXMLfilecom
prisedofasequenceofcommunications.Thecommunicationswere
betweensoftwareagentsorsoftwareagentsandrobots.
Althoughthedatasetsweresuperficiallysimilar(i.e.,logsofcommu
nicationsrecords),casestudy2presentedseveralchallengesnotpresent
incasestudy1.Thesechallengesaroseprimarilybecausethedatafor
thiscasewererelativelyimpoverished.DNArequireslargedatasetsand
wedid
receive
more
data
for
case
study
2than
for
case
study
1.
However,
theextradatawereoflittlebenefitbecausetheyprovidedlittleinforma
tionregardingthestructureofthesystemwewereanalyzing.Intermsof
structuralinformation,theextradataweremostlyredundant.
Theanalysiswascomplicatedbythefactthatthescenarioincluded
onlytwoinstancesofthetargethandoffeventthesignalwewereto
detect.Instructuralterms,thismeansthatwewerelookingforachange
innetworkstructurethatinvolved(perArtisTechsdescriptionofthe
handshake)three
links
at
most.
Thus,
the
statistical
change
detection
analysisusedincasestudy1wasofnouse.Suchasmallchangeinstruc
turewouldnotbedetectedasbeingstatisticallysignificant.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
19/557
28 CARLEY,MARTIN,ANDHANCOCK
Theanalysiswasfurthercomplicatedbytheabsenceofgeocoordi
natesforthemobilerobots.Therefore,*ORAsgeospatialvisualization
capabilitiescouldnotbeemployed.Theinsightthatcanbegleanedfrom
visualizationsofagentsxlocations,asusedincasestudy1,wasalsolost.
Withoutlocation,
we
also
lacked
any
means
for
constructing
trails
data
toexaminewhowaswherewhen.
Finally,thegoalofanalysisincasestudy2differedfromcasestudy1.
Incasestudy1,weemployedDNAtechniquesdesignedtoidentify
importantentities,locations,andtimes.Thepurposeoftheanalysis,
therefore,wastoidentifycentralities.Incontrast,thepurposeofthe
analysisforcasestudy2wastoidentifyexclusivities(i.e.,thetwoTasker
GlobalTaskerhandshakes).
Giventhe
impoverished
data
set
and
our
goal
of
detecting
only
two
instancesofthehandoffamongasmallsetofagentsandrobots,the
analysisstrategyweadoptedforthiscasewasoneofconvergingopera
tions(tokickstartaDNAofricherdatainthefuture).ArtisTechper
sonnel(withtheirknowledgeofthesystem)manuallyanalyzedthedata.
CASOSpersonnelappliedDNAtechniquestothedata.Thefollowingdis
cussesonlytwooftheissuesweaddressed.
TofindthehandoffwhereoneTaskingAgentloanedarobotto
anotherTaskingAgent,wereliedonArtisTechsidentificationofmes
sagetypesthatindicatesuchahandoffoccurred.Threemessagetypes
wererelatedtothehandoff:RemoveIdentity,ReportPositionToSelected
Tasker,andReportPositionToSelectedTaskerReturn.Findingthehandoff
wasthensimplyamatterofusing*ORAsSphereofInfluencecapabilities
tovisualizewhichagentswereinvolved.Figure1.13showsthethree
importantmessages,alongwiththeagentsthatsentandreceivedthem.
ItcanbeseenthatTaskingAgent3ispositionedinthenetworkdiffer
entlythanTaskingAgents1,2,and4.Furthermore,weseethatTasking
Agent3was
the
only
agent
to
send
the
ReportPositionToSelectedTasker
Returnmessage(asindicatedbythearrowonthelinkbetweentheagent
nodeandthemessagenode).GlobalTaskingAgent7receivedthismes
sageandsentaRemoveIdentitymessagetoTaskingAgent3.Thesphere
ofinfluencealsoindicatesthattheReportPositionToSelectedTaskermes
sagewasprobablynotuniquelyrelatedtothehandshake.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
20/557
DynamicNetworkAnalysis 29
Figure1.13SphereofInfluencevisualizationshowingrobothandoff
Inanticipationofricherdatasets,thesecondissueweexaminedwas
whetherwecoulduseDNAtechniquestopartitionagentsintofore
groundandbackgroundagents.Tothisend,theArtisTechteamproduced
ameticulous
message
trace
analysis
and
event
identification
as
ground
truthfortheCASOSteam.Evenasfullyknowledgeabledesignersofthe
communicationlogicthisanalysisanddocumentationtookmorethan4
hoursusingsimpletextsearchandanumericmessagefrequencyanalysis
providedbyCASOS.Thecomplexityofthisanalysisunderscorestheneed
fornetworkanalysistechniquesforamorecompletesystemexperiment
analysis.
PerArtisTechsanalysis,agentscouldbedividedintoforeground
agentsthatweresubstantivelyrelatedtothescenarioandbackground
agentsthatexistedsimplytomakethesimulationrun.Ourtaskat
CASOS,therefore,wastoemployoneoranotherDNAtechniquetosepa
rateforegroundfrombackgroundagents.WefoundthattheNewman
groupingalgorithmworkedwell,separatingforegroundandbackground
agentsintogroupsthatcloselyapproximatedthemanualanalysis.Seven
oftenbackgroundagentswerecorrectlyclassified.Buttherewasdis
agreementbetweenArtisTechjudgesregardingwhetheroneofthethree
misclassifiedagentswasabackgroundoraforegroundagent.Onefore
groundagent
out
of
29
undisputed
foreground
agents
was
misclassified.
ThattheNewmangroupingalgorithmcorrespondsfairlywellwiththe
judgmentsofdomainexpertsispromising,andsuggestsalineof
researchconcerningthepsychologicalvalidityofNewmangrouping.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
21/557
30 CARLEY,MARTIN,ANDHANCOCK
SYSTEMUNDERSTANDINGANDTUNINGACTIVITIESArtisTechpostulatedtheadditionalbenefitofperformingDNAon
theoutputlogsfromtheinternalsystemcommunications;theanalysis
supportedunderstanding
of
how
the
team
and
system
achieved
the
mea
suredperformance.TheARTEMISresearchteamwasintheprocessof
settingupsystemmodelexperimentstostudytheadvantagesthatauto
matedreasoningcouldaddtowidelyusednumericimageandvideopro
cessingforthepurposeofPCVS.Aswemodeledthereasoningandsetit
intothesimulationwehaveconductedmanytestruns.Wecaneasilysee
whenthemacrosystembehaviorisasdesignedandwhenitdeviates
fromintentions.However,todeterminewhythedeviationsoccurand
howthe
reactive
agent
networks
achieve
intended
system
performance
requiresmessagelevelanalysis.ArtisTechsharedmixedresulttestlogs
withCASOSspecificallytofacilitatetheidentificationofexpectedand
unexpectedresults,andhowtheyarise.
Resultsfrominitialanalysesinbothcasestudiesdivergedfrom
expectedresults.Inhindsightthisishardlysurprising.Designofacom
plex,multiagentsystemwithemergentreactivesystembehaviormay
notbepossiblewithoutthesupportofnetworkanalytics.
Inbothcaseswefoundevidenceofpragmatic,technicallycorrect,
programmingpracticesthatinterferedwithsubstantiveDNAofthecommunicationslogs.ArtisTechconfirmedthattheexperimentrunsthat
wereanalyzedwerenotconsideredfinalorexpectedtobeentirelycor
rect,merelytypicalincontentandform.Whetherweviewthisasaveri
ficationissue(i.e.,buildingthethingright)oravalidationissue(i.e.,
buildingtherightthing)dependsonperspective.Giventhatverification
isprimarilyaninternalactivity;thedegreetowhichprogrammers
achievedtheintentbehindspecificationsisamatterofinterpretation.
Fromthe
perspective
of
network
analysis
(i.e.,
users
of
the
output
logs),
however,wecannoteaminorshortfallinvalidation.Specifically,com
municationrecordsthatarenotrelevanttothescenariobeinganalyzed
shouldbefilteredfromthelogtoincreaseusability.
Intheinitialanalysis,weexperimentallyillustratedthatDNA
approachesprovideacapabilitytoanalyzecomplexmessagesetstofind
particularbehaviorpatternsofinterest.Inthesecondanalysiswebegan
forgingacollaborativeresearchapproachdesignedtounearththeana
lyticsteps
the
combinations
of
DNA
techniques
required
to
analyze
differenttypesofmessagingbehavior.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
22/557
DynamicNetworkAnalysis 31LESSONSLEARNEDANDFUTUREDIRECTIONS
Opencollaborationbetweendataprovidersandnetworkanalysts
createdabeneficialgapbetweenexpectedandobservedsystembehav
ior.As
data
provider,
ArtisTech
developed
the
multi
agent
system
and
environmentsimulations,designedthescenarios,andconductedthe
simulationbasedexperimentswithexpectedsystembehaviorsinmind.
Thenonlinearitiesinherenttocomplexsystemscomprisedofinteracting
agents,however,makeitnotoriouslydifficulttopredictemergentand
reactivebehavior,andareindeedthereasoncomputersimulationisnec
essary.Asnetworkanalyst,CASOSreceivedmilitarilyrelevantbattlefield
simulationdatawithoutpriorknowledgeofexpectedsystembehavior.
Thechallenge
to
CASOS,
therefore,
was
to
use
DNA
to
characterize
what
happenedinthemysteriousscenariosreceivedandCASOSobserva
tionsinitiallydivergedfromArtisTechdefaultexpectations.Toresolvethe
discrepancy,weusedDNAtoexaminewhythesimulationsdidnot
behaveasexpected.Thus,thepostulatedbenefitoftheArtisTechCASOS
collaborationwastheuseofnetworkanalyticstogaininsightintothe
performanceofmultiagentsystems.ArtisTech,withCASOS,nowplans
tousethisapproachandDNAtoolstobuildasetofreusablesystem
experimentanalysismethodsthatwillbeappliedtounderstandhow
variantHuman/AutomationPCVSexperimentsachieveobservedresultsusingnetworksofcommunicationandbehaviors.
WithregardtodemonstratingthetacticalrelevanceofDNA,the
availabilityofmilitaryscenariodataisinvaluable.Itprovidesopportuni
tiestocombineextantDNAtechniquesintoanalyticstrategiesthatpro
duceresultswarfighterscanuse.Generally,wefoundthatthefamilyof
DNAmetricsandvisualizationsthathavebeenimplementedin*ORA
overthepast10orsoyearsprovidesanamplebasisforconductingtacti
callyrelevant,
actionable
analyses
on
battlefield
data.
Two*ORAcapabilitieswereparticularlyhelpfulduringthiseffort:
geovisualizationandnetworkchangedetection.The(notsosimple)act
ofplacingnetworksonamapputsabstractsocialnetworksintoacon
crete,spatialcontext.Itprovidesexplicitinformationaboutwherethe
actionistakingplace.Thecapacitytodetectstructuralchangesamong
temporallyorderednetworksprovidesexplicitinformationaboutwhen
theactionistakingplace.
Thecase
studies
also
helped
to
identify
several
areas
where
future
developmentcouldimprove*ORAstacticalrelevance.Theimprove
mentswouldgenerallysupportdeliveryofoneormoreTacticalInsight
Reports.Asenvisioned,thesereportswouldcontaintheoutputsofall
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
23/557
32 CARLEY,MARTIN,ANDHANCOCK
DNAtechniquesthatcontributetoaparticularanalyticstrategy.From
thetwocasestudiesdescribedabove,itappearsthatcorrespondences
amongnetworksoftherelationsamongentities,networksofgeospa
tiallyanchoredentities,andnetworksdistributedovertimewillplaya
centralrole
in
such
reports.
ForArtisTechthegoalofDNAapplicationstosystemanalysisand
monitoringistoinvestigatethecreationofageneralizeddatastructure
andDNAmethodcombinationapproachthatwillallowtheencodingof
expectedordetectedmessageandbehaviorpatterns.Theinitialapplica
tionofthisconcernspostexperimentreview.Themorefarreaching
implicationsofsuchDNAmechanismsextendtorealtimeembedded
systemmonitoringtoincreasedistributedsystemsecurityandusertrust.
ACKNOWLEDGEMENTS
ThisworkispartoftheDynamicsNetworksprojectatthecenterfor
ComputationalAnalysisofSocialandOrganizationalSystems(CASOS)of
theSchoolofComputerScience(SCS)atCarnegieMellonUniversity
(CMU).For*ORA,forthepartsusedinthisanalysis,thegeospatialcom
ponentsweresupportedbyAROandERDCTECW911NF0710317,the
*ORALoombyONRN000140610104,theovertimechangedetection
basicanalysisbyARIW91WAW07C0063andtheORAimplementationofchangedetectionbyARLDAAD190120009,thefourieranalysisbyONR
N000140610104,thekeyentityanalysiswassupportedbyONR
N000140610104,extensionstohandlecommunicationsnetworksby
ARLthroughthe CommunicationsandNetworks(CN)Collaborative
TechnologyAlliance(CTA)20002504andtheimmediateimpactassess
mentbyONRN000140610104.Thebasicresearchconductedhereto
explorehowtousethesetechnologiesforadversarialbehaviorandto
linkto
the
ArtisTech
ARTEMIS
PCVS
system,
AlgoLink
simulator,
DARE
wassupportedbytheArmyResearchLabthroughtheAdvancedDecision
Architecture(ADA)CollaborativeTechnologyAlliance(CTA) DAAD1901
20009.Theviewsandproposalscontainedinthisdocumentarethoseof
theauthorandshouldnotbeinterpretedasrepresentingtheofficialpol
icies,eitherexpressedorimplied,oftheOfficeofNavalResearch,the
ArmyResearchOffice,theArmyResearchInstitute,theArmyResearch
LabortheU.S.government.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
24/557
DynamicNetworkAnalysis 33REFERENCES
Carley,K.M.(2002).Smartagentsandorganizationsofthefuture.InL.Lievrouw&
S.Livingstone(Eds.),Thehandbookofnewmedia(pp.206220).ThousandOaks,CA:Sage.Carley,K.M.,Columbus,D.,DeReno,M.,Reminga,J.,&Moon,I.C.(2008).ORAUsersGuide2008(Tech.Rep.No.CMUISR08125).Pittsburgh,PA:InstituteforSoftwareResearch,SchoolofComputerScience,CarnegieMellonUniversity.
Davis,G.B.,&Carley,K.M.(2008).ClearingtheFOG:Fuzzy,overlappinggroupsfor
socialnetworks.SocialNetworks,30(3),201 212.Davis,G.B.,Olson,J.,&Carley,K.M.(2008).OraGISandloom:Spatialandtemporal
extensionstotheORAanalysisplatform(Tech.Rep.No.CMUISR08121).Pittsburgh,PA:InstituteforSoftwareResearch,SchoolofComputerScience,CarnegieMellon
University.
McCulloh,I.,
&
Carley,
K.M.
(2008).
Detectingchangeinhumansocialbehaviorsimulation(Tech.Rep.No.CMUISR08135).Pittsburgh,PA:InstituteforSoftware
Research,SchoolofComputerScience,CarnegieMellonUniversity.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
25/557
34 CARLEY,MARTIN,ANDHANCOCK
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
26/557
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
27/557
E D I T E D B Y PAT RI C I A M C D E RM O T T AN D LAU RE L ALLE N D E R
ADVANCED DECISI ON ARCHITEC TURES
FOR THE WARFI GHTE R:F O U N D AT I O N S A N D T E C H N O L O G Y
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
28/557
i
A D VA N C E D D E C I S I O N A R C H I T E C T U R E S
F O R T H E WA R F I G H T E R :
F O U N D AT I O N S A N D T E C H N O L O G Y
E D I T E D B Y PAT R I C I A M C D E R M O T T
A N D L A U R E L A L L E N D E R
This book, produced by the Partners of the
Army Research Laboratory Advanced Decision Architectures
Collaborative Technology Alliance, also serves as the
Final Report for DAAD19-01-2-0009.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
29/557
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
30/557
SECTIONI.COLLECTING,
PROCESSING,
AND
DISTRIBUTING
BATTLEFIELD
INFORMATION
SECTIONI
COLLECTING,PROCESSING,ANDDISTRIBUTINGBATTLEFIELD
INFORMATION
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
31/557
14 SECTIONI
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
32/557
35
Chapter2
2.
FROMBACKGROUNDSUBTRACTIONTOTHREATDETECTIONIN
AUTOMATEDVIDEOSURVEILLANCE
JoshuaEckroth
TheOhio
State
University,
Columbus,
OH
DikpalReddy
UniversityofMaryland,UMIACS,CollegePark,MD
JohnR.Josephson,Ph.D.
TheOhioStateUniversity,Columbus,OH
RamaChellappa,
Ph.D.
UniversityofMaryland,UMIACS,CollegePark,MD
TimothyN.Miller
TheOhioStateUniversity,Columbus,OH
INTRODUCTION:PERSISTENTVIDEOSURVEILLANCEAsvideocamerasandothersensorsbecomecheaperandeasierto
network,sensornetworksbecomeincreasinglyattractivemeansto
acquireusefulinformationformilitaryoperations,suchasvideosurveil
lanceforfacilitiesprotection,andpersistentsurveillancetomaintain
sensorycontactwithtargetsofinterest(Pendal,2005).However,without
assistancefromautomation,humanswillbeoverloadedbyinformation,
andunabletouseiteffectively.
Whatinformationconsumesisratherobvious:itconsumesthe
attentionof
its
recipients.
Hence
awealth
of
information
creates
apov
ertyofattention,andaneedtoallocatethatattentionefficientlyamong
theoverabundanceofinformationsourcesthatmightconsumeit.
(Simon,1971)
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
33/557
36 ECKROTHETAL.
Itisdifficultforhumanstovigilantlymonitoralargenumberofvideo
feedsforextendedperiodswithoutfatigue,orcomplacency,especiallyif
theyareaskedtorecognizethesignificanceofrareeventsinacomplex
streamofevents(Warm,et.al.,1996;Grier,et.al.2003;Pattyn,et.al.,
2008).Moreover,
sometimes
events
of
interest
cannot
be
recognized
simplyfromthevideowithoutknowingwhereaspecificcameraispoint
ing,atarestrictedarea,forexample.Keepingtrackofthelocationand
significanceofacamerasfieldofviewimposesanadditionalcognitive
burden.Thecognitiveburdenisevengreaterifrecognizingeventsof
interestrequirestrackingentitiesastheybecomevisibleindifferentcam
eras,andaccessingmentalmapstounderstandthesignificanceof
motionpaths.
Automationcan
help.
It
can
potentially
provide
users
with
alerts
basedonrecognizingindicatorsofthreateningbehavior,includingbehav
iorthatcannotbedetectedwithasinglecamera.Forexample,themove
mentofanentityfromplacetoplaceinrelationtothemapmightshowa
patternindicativeofscoutingtheperimeterofafacility,wherenosingle
cameraviewshowsanythingsuspicious.Manyunsolvedtechnicalprob
lemsremain,however,includingproblemsabouthowtoextractneeded
informationfromvideoimagery,andproblemsofhowtoprocess
acquiredinformationtoclimbthelevelsofabstractionfromindividual
cameracenteredframeworkstoaworldcenteredframework,andfrom
motiondescription(ObjectO5604movedalongpathP52fromlocation
L12attimeT55toL13atT66.)tobehaviordescription(O5604proceeded
slowlynorthonroadR3 totheintersectionwithR7,turnedEastonR7.)torecognitionofindicatorsofthreat(O5604slowlycircledthefacil
ity.).
Abundantvideoalsothreatenstooverloadcommunicationsand
storagesystems.
Typically,
the
sensed
bits
are
highly
redundant
for
the
tasktobeperformed.Forexample,whenitisimportanttotrackamov
ingobject,theimportantinformationiswhichpartoftheimageischang
ingcontinuously.Tounderstandwhichregionisrelevant,subtractionof
theunchangingbackgroundisneeded,andifitcanbeperformedatthe
sensoritself,significantlyfewerbitsneedtobetransmitted.Accordingly,
anewframeworkcanbeenvisagedinwhichthetraditionalwaysofsens
inganimage(inrectangularpixels)arereplacedbysensingtheimage
information
in
a
compact
form
(compressed
sensing),
and
then
recon
structingtherequiredimage,asneeded,fromthiscompactinformation.
Thispaperdescribessomeelementsofrecentprogressintechnology
forvideosurveillancethatspecificallyaddressmethodsforbackground
subtractiontodetectchangesfromframetoframeinvideo,trackingof
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
34/557
AutomatedVideoSurveillance 37
viewedobjectsincameracenteredimagespace,trackingofobjectsin
worldorientedobjectspaceusinginformationfrommultiplecameras,
andmethodsforclimbinglevelsofabstractionindescriptionsof
behavior.
BACKGROUNDSUBTRACTIONANDTRACKINGUSINGCOMPRESSIVECAMERAS
Introduction
Mostcamerasinstalledforsurveillanceapplicationsareusedforsim
plevisiontaskssuchasdetectionandtracking,poseestimation,and3D
reconstruction
from
silhouettes,
and
very
few
for
higher
level
tasks
such
asactivityrecognition.Inmanyofthesetasks,theamountofdatacol
lectedishugeforthepurposeoftheapplication.Forexample,atypical
backgroundsubtractionalgorithmcomparesasingleframe,fullresolu
tionimageobtainedfromacameratoanotherfullresolutionimagethat
hasbeenlabeledthebackground.Pixelsthatdifferbetweenthetwo
imagesareconsideredforeground.Theresultisblobsinthefore
groundthatrepresentobjectsthatarebelievedtohavemovedbetween
thetimethebackgroundimagewasacquiredandthetimethefore
groundimage
was
acquired
(usually,
the
background
image
is
updated
everytimeanewframeisobtainedfromthecamera,sothatmovements
fromoneframetothenextaredetectedandplacedintheforeground).
Theresultingforegroundimageisoftenrepresentedasablackandwhite
fullresolutionimage,whereblackrepresentsbackgroundandwhiterep
resentsforeground;iftheseimagesareobtainedfrom,say,parkinglot
surveillance,thentheforegroundblobswilllikelyresemblesilhouettes
ofmovingcarsandwalkingpeople.Yet,theseblobsoccupyonlysmall
regionsof
the
image,
since,
in
most
surveillance
video,
there
are
few
movingobjectsatanyonetime.Similarly,whentracking,weareinter
estedinonlyasmallregionoftheimagethatismoving,anddonotcare
abouttheotherpartsoftheimage.Neverthelesstheentireimageis
commonlysensedandtransmittedtoacentrallocationforprocessing.
Thismeansthatahugeamountofdata,whichisultimatelyuseless,is
collectedandtransmitted,thuswastingsensingandbandwidth
resources.Inthispaper,wepresentanapproachtoalleviatingthisprob
lem.We
show
our
approach
on
two
vision
applications:
background
sub
tractionandtracking.Inbothoftheseapplications,weusea
compressivecameratoobservetheimagesandthenprocessingis
doneonthesemeasurements.
-
7/27/2019 198- ADVANCED DECISION ARCHITECTURES FOR THE WARFIGHTER.pdf
35/557
38 ECKROTHETAL.
Acompressivecameraisadevicethathasbeenbuiltontheprinci
plesofcompressedsensing(Candes,2006).Suchacamerameasures,not
theimagepixelslikeaconventionalcamera,butasmallnumberofran
domlinearprojectionsoftheimage.Thenumberofrequiredmeasure
mentsis
significantly
fewer
than
the
number
of
pixels
in
aconventional
camera,andthisinherentlyreducesthespaceneededtostorethe
image.Thefullimage(intheformofpixels)canbereconstructedfrom
theselinearcombinations.
Wepresentaninnovativeapproachwhereinvisiontasksareper
formed,notonthefullimages,butonthecompressedmeasurements
directly,withouttheneedtoreconstructthefullimages.Becausethefull
imageisnotneededforthetasks,asignificantreductionisachievedin
theamount
of
data
that
needs
to
be
sensed.
In
the
next
section,
we
pro
videabriefintroductiontocompressedsensingtheoryandcompressive
cameras.Wethenpresentourapproachtobackgroundsubtractionand
trackingoncompressedmeasurements.
CompressiveCamerasandComputerVisionApplicationsCompressedSensing.Supposewehaveanimage ofsize
(i.e.,vectorized),
then
we
can
represent
the
image
in
some
basis
as
(1)
where isasparsecoefficientvector( sparse).Thevector has
veryfewlargenonzerocomponentsindicatingthattheimagecanbe
compressed.Waveletsareanexampleofsuchabasis.
IntheCSframework(Candes,2006)wedonotmeasurethe larg
estelementsof ,butweinsteadmeasure linearprojectionsof
theimage ontoanotherbasis .
(2)
where arethecompressedmeasurements.Since ,the
systemofequationsareunderdetermined,bututilizingthesparsityofwecanrecoverthesignalbysolvingthefollowing optimization
problemcalledBasisPursuit.
(3)
Compressivecameras.Basedoncompressivesensing(CS)theory,a
prototypesinglepixelcamera(SPC)wasproposedin(Wakin,2006).The
x 1N
x =
M N