BDI vs FSM Agents in Social Simulations for Raising Awareness in Disasters

HAL Id: halshs-02107058

https://halshs.archives-ouvertes.fr/halshs-02107058

Submitted on 23 Apr 2019

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Awareness in Disasters

Carole Adam, Patrick Taillandier, Julie Dugdale, Benoit Gaudou

To cite this version:

Carole Adam, Patrick Taillandier, Julie Dugdale, Benoit Gaudou. BDI vs FSM Agents in Social Simulations for Raising Awareness in Disasters. International Journal of Information Systems for Crisis Response and Management, IGI-Global, 2017, 9 (1), pp.27-44. �10.4018/IJISCRAM.2017010103�.

�halshs-02107058�

DOI: 10.4018/IJISCRAM.2017010103

Copyright©2017,IGIGlobal.CopyingordistributinginprintorelectronicformswithoutwrittenpermissionofIGIGlobalisprohibited.

BDI vs FSM Agents in Social Simulations for Raising Awareness in Disasters:

A Case Study in Melbourne Bushfires

Carole Adam, Grenoble Informatics Lab (LIG), University Grenoble Alpes, Grenoble, France Patrick Taillandier, MIAT, University Toulouse, Toulouse, France

Julie Dugdale, Grenoble Informatics Lab (LIG), University Grenoble Alps, Grenoble, France Benoit Gaudou, University Toulouse 1 Capitole, Toulouse, France

ABSTRACT

EachsummerinAustralia,bushfiresburnmanyhectaresofforest,causingdeaths,injuries,and

destroyingproperty.Agent-basedsimulationisapowerfultooltotestvariousmanagementstrategies

onasimulatedpopulation,andtoraiseawarenessoftheactualpopulationbehaviour.Butvalidresults

dependonrealisticunderlyingmodels.ThisarticledescribestwosimulationsoftheAustralian

population’sbehaviourduringbushfiresdesignedinpreviouswork,onebasedonafinite-statemachine

architecture,theotherbasedonabelief-desire-intentionagentarchitecture.Itthenproposesseveral

contributionstowardsmorerealisticagent-basedmodelsofhumanbehaviour:amethodologyand

toolforeasilydesigningBDImodels;anumberofobjectiveandsubjectivecriteriaforcomparing

agent-basedmodels;acomparisonofourtwomodelsalongthesecriteria,showingthatBDIprovides

betterexplanabilityandunderstandabilityofbehaviour,makesmodelseasiertoextend,andistherefore

bestadapted;andadiscussionofpossibleextensionsofBDImodelstofurtherimprovetheirrealism.

KeywoRDS

Agent-Based Social Simulation, BDI Architecture, Bushfires, Crisis Management, Model Comparison

INTRoDUCTIoN

EachsummerinAustralia,bushfiresburnmanyhectaresofforest,causingdeaths,injuries,and

destroyingproperty.Societiescanmanagesuchcrisisandemergencysituationsinseveralways:adopt

urbanandterritoryplanningpoliciestoreducetherisks(e.g.forbidconstructioninexposedareas);

raiseawarenessandpreparethepopulationinadvance;orcreateefficientemergencymanagement

policiestodealwithcriseswhentheyhappen.Modellingandsimulationoffertoolstotesttheeffects

andcomplexinteractionsofthesedifferentstrategieswithoutwaitingforanactualcrisistohappen,

withoutputtinghumanlivesatrisk,withlimitedcost,andwithagreatdegreeofcontrolonall

conditionsandthepossibilityofreproducingexactlythesamesituationasmanytimesasneeded.

Whenmodellinghumanbehaviour,mathematical,equation-basedmodelsaretoolimited

(Parunak,Savit,&Riolo,1998)whereasagent-basedmodelsoffermanybenefits(Bonabeau,2002).

Theyallowcapturingemergentphenomenathatcharacterisesuchcomplexsystems;theyprovidean

intuitiveandrealisticdescriptionoftheirbehaviour;andtheyareflexible,offeringdifferentlevels

ofabstractionbyvaryingthecomplexityofagents.Agent-basedmodellingandsimulationplatforms

overvariousarchitecturesofdifferentcomplexityfortheagents:reflexorreactiveagentsarevery

simplistic,reactingtoenviron-mentalstimuliwithoutanylong-termreasoning;finite-statemachines

requirescriptingallofthepossiblestatesoftheagentsandthecorrespondingtransitions;cognitive

agentsofferamoreflexibledescriptionofbehavioursintermsofgoalsandplans.

InparticulartheBDI(Belief,Desire,Intention(Rao&George,1991))architectureisvery

sophisticatedandrealistic,groundedinthephilosophyofhumanrationality(Bratman,1987),and

linkedwithemotions(Adam,Herzig,&Longin,2009).Suchrealismofthehumanbehaviourmodel

isimportantforthesimulationtoproducevalidresults(vanRuijven,2011).Forthesereasonsandas

previouslydiscussedin(Adam&Gaudou,2016a),theBDIarchitectureisthereforemoreadaptedfor

crisissituationsthatinvolvecomplexindividualdecision-making,influencedbyemotions(sometimes

causingirrationalactions),andbythesocialcontext(effectofgroup,family,etc.).Accordingto

(Norling,2004),BDIalsoprovidesanadaptedlevelofabstractiontodescribehumanbehaviourin

termsoffolkpsychology,whichisthepreferredlevelofdescriptionforhumans.Itthereforeaddresses

theproblemofthescarcityof(quantitative)behaviouraldatabyallowingtheuseofqualitativedata

suchaswitnessstatementsorexpertreports.

Despitetheseadvantagesthatmakeitverysuitableforsocialsimulation,BDIhashadlimited

useinthisfieldduetothelackofadaptedtoolstoharnessitscomplexity(Adam&Gaudou,2016a).

Inpreviouswork(Adam,Danet,Thangarajah,&Dugdale,2016;Adam,Beck,&Dugdale,2015;

P.Taillandier,Bourgais,Caillou,Adam,&Gaudou,2016)wehavedescribedhowtwonewtools

couldbeusedtode-velopBDIagent-basedmodelsfrominterviews.Weillustratedhowtousethese

toolsbyturninganexistingmodelofthebehaviouroftheAustralianpopulationinbushfires(witha

finite-statemachinearchitecture(Adam&Gaudou,2016b,2017))intoaBDImodel.Inthispaperwe

nowwanttocom-parethesetwomodels,addressingthesameproblemwithdifferentarchitectures,

usingbothobjectiveandsubjectivecriteria.Webelievethatsuchmodelcomparisonisimportantto

furtherjustifytheinterestofBDImodels.

Thispaperisstructuredasfollows.Wefirstdescribethecontextofourcasestudy(Melbourne

bushfires)andourinitialmodelofpopulationbehaviourusingafinitestatemachine(FSM)architecture

(Adam&Gaudou2017).WethendiscussBDIagentsasamorecomplexalternative,exposesome

oftheobstaclespreventingtheiruseinsocialsimulations,proposesometechnologicalsolutionsand

illustratethembyprovidingaBDImodelofpopulationbehaviour.Thenextsectionisdedicated

tocomparingthisBDImodelwiththeinitialFSMmodel:wediscusstheliteratureaboutmodel

comparison,deriveourownlistofcomparisoncriteria,andpresenttheresultsofcomparingour

modelsfollowingthesecriteria.Thelastsectiondiscussesfutureprospectsandconcludesthepaper.

INITIAL MoDeL oF THe PoPULATIoN IN BUSHFIReS Context

Wefocusontheso-calledBlackSaturday,7thFebruary2009,whenparticularlystrongbushfires

hitthestateofVictoriainAustralia,killing173peopleanddestroyinghectaresofbushandmany

properties.Theofficialrecommendationtothepopulationwasto“prepare,stayanddefend,or

leaveearly”.However,reports(AlanRhodes,2014)writtenafterthesefiresshowedthatemergency

managementpoliciesweredesignedbasedonan(ideal)expectedbehaviourthatdifferedfromthe

residents’actualbehaviourontheday.Itisthereforeimportanttoprovidedeciderswithasimulatorto

raisetheirawarenessaboutresidents’actualdecisionmaking,andletthemtrydifferentcommunication

strategies.Currently,theavailabledataismostlyintheformofwitnessstatements(Exell,2009)

andpolicehearings.TheRoyalCommission(Teague,McLeod,&Pascoe,2009)alsogatheredthe

followingstatisticsaboutthevictims:

• Preparation:58%ofthevictimshadmadenopreparationatall;manypreparedtoleavebut

expectedawarningbeforegoing;20%intendedtostayanddefendandwerewellprepared;14%

hadmadelimitedpreparation;

• Awareness:Thefiretook30%ofthosewhodiedbysurprise;24%ofthevictimswereunaware

thattheylivedinabushfirepronearea;38%hadnobasicknowledgeaboutwhattodo;

• Causes of death:14%diedwhileescaping(4%bycarand10%byfoot);69%diedwhilepassively

shelteringinabuilding;theothersdiedwhiletryingtodefend;

• Vulnerability:44%ofthevictimswereconsideredvulnerablebecauseoftheirage(under12or

over70),illnessordisability;32%diedonpropertieswhosedefendabilitywasquestionable(it

wasnotworthtryingtodefendthembecausetheywerelikelytoburnanyway).

Tofacilitatecomparison,wehaveimplementedageneralmodelintheGAMAsimulation

platform (Grignard. et al., 2013) that allows choosing between two models of the civilians’

behaviour:thefirstoneisbasedonaFSM(finitestatemachine)andthesecondoneonBDI

(beliefdesireintention).Figure1presentstheclassdiagramoftheglobalmodel.Forthesakeof

comparison,wedefinedagenericCivilianclassthatregroupsallthecommonproperties,actions

andreflexesofbothbehaviourmodels.

Belowwedescribethemodeloftheenvironment,buildingsandfires;thenthegenericmodel

ofciviliansandfinallytheFSMarchitectureforcivilians.ThefollowingsectiondescribestheBDI

Figure 1. UML class diagram of the model

architectureandthetoolsthatwereusedinitsdesignandimplementation.Wetrytogiveenough

detailsabouteachmodeltoallowthereadertounderstandthecomparisonprovidedinthenextsection.

MoDeL oF THe eNVIRoNMeNT AND THe FIRe

Forthesakeofsimplicity,theenvironmentisrepresentedasagridcontainingthedifferenttypesof

agents(houses,shelters,fires,andresidents).Thissimplisticenvironmentisnotrealistic,butwas

provensufficienttosimulatetheresidents’decision-makingbehavioursinreactiontofires(Adam

&Gaudou,2017).

Fire

Verycomplexanddetailedmodelsofthespreadoffirealreadyexist(Du,Chong,&Tolhurst,

2013;Miller,Hilton,Sullivan,&Prakash,2015),butrealisticfirebehaviourisnotthefocus

here.Withthegoalofnotaddingunnecessarycomplexity,wehavedesignedaverysimplistic

firemodelthatissufficienttotriggerandvisualisethereactionsofthepopulationinwhich

weareinterested.Thefireiscomposedoffireagents(eachwithalocationandanintensity

representingitsradiusofaction),havingareflexarchitecture,i.e.thefollowingreflexesare

triggeredateachstepofthesimulation:

• Grow (increase or decrease intensity):Probabilitiesareparameters;

• Propagate:Toanon-burningneighbourcell,creatinganewfireagent.Theprobabilityof

propagating,andstartingintensityofnewfires,areparameters;

• Deal damage:Tobuildingsinitsradiusofaction(basedonitsintensity).Theamountofdamage

ispickedrandomlybetween0andamaximumvalue,afunctionofintensityanda“damage

factor”parameter;

• Deal injuries:Toresidentsinitsradiusofaction,alsoarandomnumberbetween0andthe

maximumvaluebasedonitsintensityandonan“injuryfactor”parameter.Ifthepersonisin

theirhousetheinjuryismoderatedbyitsresistanceweightedbya“protectionfactor”parameter;

• Disappear:Whenitsintensityisnull.

Thedifferentparametersinvolvedallowtheuserofthesimulationtomakethefiremoreorless

dangerous(growingandpropagatingmorequickly,dealingmoredamageandinjuries,etc.),inorder

toobservethedesiredbehaviours.Actionsarealsoavailabletostartnewfiresorstopallfires(and

thusthesimulation).

Houses

Theenvironmentinitiallycontainsanumber(parameter)ofhouseseachinhabitedbyexactly1

resident(infutureworkweplantoconsiderfamiliesandtheirrelationships).Eachhouseisanagent

withthefollowingattributes:

• Owner:Theresidentofthathouse;

• Resistance:Randominitialvaluebetween100and200tosimulatehouseresistance.Thisvalue

isincreasedbytheresidentpreparingthehouseforafire,ordecreasedbyfiredamage.Resistance

alsoofferssomeprotectionfromfireinjuriestoitsresident;

• Damage:Thedamagereceivedfromfire.

Thehousescollapsefromfiredamagewhentheirresistancedropsto0.Theythenceasetooffer

protectionandtheresident’smotivationtodefendthemalsodisappears.Housesstayintheenvironment

asruinsforfinalvisualisation.

Shelters

Sheltersaresafeplacesofferingtotalprotectionfromthefires(noinjuriescanbereceivedwhilein

ashelter).Oncearesidenthasreachedashelter,itstaysinsideuntiltheendofthesimulation.

GeNeRIC CIVILIAN MoDeL

Civilianagentsareheterogeneousagents,eachhavingtheirownvaluesofattributes:

• Defend motivation:Randominitialvaluebetween0.0and1.0tosimulatethepropensityto

defendtheirhome;

• Escape motivation:Randominitialvaluebetween0.0and1.0tosimulatethepropensityto

escapeincaseofdanger;

• Awareness probability:Randominitialvaluebetween0.0and1.0tosimulatetheattention

towardsdangers;

• Perception radius:Randominitialvaluebetween0.0and20.0tosimulatethemaximaldistance

ofperceptionoffires;

• Defence radius:Randominitialvaluebetween0.0and10.0tosimulatetheareaofdefence;

• Danger radius:Randominitialvaluebetween0.0and10.0tosimulatetheareaofdanger;

• Velocity:Randominitialvaluebetween0.2and1.0tosimulatethemovingspeed;

• Location:Onthegrid;

• House ID:Eachagentisinitiallyinahouse;

• Health:Thehealthofthecivilian.Itisincreasedbypreparingforfireanddecreasedwhen

injuriesarereceived.Ahealthof0meansdeath;

• Injuries:Theinjuriesreceivedfromfire.Itdecreasesthehealth;

• Is safe:Definesifthecivilianisinashelter;

• Is dead:Definesifthecivilianisdead;

• In smoke:Definesifthecivilianisinsmoke(slowmovement).

Inaddition,eachcivilianagenthasthefollowingactions:

• Prepare for fire:Consistsinincreasingtheresistanceofthehouse(watering,removing

vegetation,etc.)andhealth(wearingappropriateclothing,etc.);

• Fight fire:Consistsindecreasingtheintensityofnearbyfiresbyavalue;

• Moving:Consistsinheadingtowardstheclosestshelter(amongsttheknownshelters);theagent

mightgetinjurediftravellingtooclosetothefire.

Finally,theyhavereflexes:

• Update status:Isactivatedateverystep;updatestheagentspeedaccordingtothehealthand

thepresenceofsmoke;italsoupdatestheagent’smotivationtodefendthepropertyaccording

totheagentcontext;

• Die:Isactivatedwhenthehealthoftheagentreaches0;itkillstheagent(theattributeisdead

issettotrue).

FINITe-STATe MACHINe MoDeL oF BeHAVIoUR

Civilian-FSMagentshaveafinite-state-machinearchitecture(Adam&Gaudou,2017)withthe

followingstatesandtransitions(c.f.Figure2),inspiredbythestatesobservedintheinterviews:

• Unaware:Initialstatewheretheagentis(rightlyorwrongly)unawareofanydanger,anddoes

nothing;agentscanbecomeawarebyspottingfiresintheirperceptionradius(seeflames,smell

smoke,etc.),withaprobabilitybasedontheirobjectiveabilities;theyup-datetheirvalueof

subjectivedangerbasedontheirperceptionsandmotivations;

• Indecisive:Theagentisawareofsomefiresbuthasnotyetmadeadecisionabouthowtoreact;

agentsstayindecisiveforavaryingamountoftime,untiltheyhaveenoughmotivationtoeither

fightorescape;initialmotivationsareindividualandthenvarybasedontheevaluationofthe

situation(subjectivedanger);

• Preparing to escape:Theagenthasdecidedtoleaveandstartspreparing.Theagentdoesthis

untilready,orsurprisedbythefirebeforebeingready(transitiontoEscaping),orblockedby

thefireandforcedtostay(transitiontoPreparingtodefend);

• Escaping:Theagentisevacuatingtowardstheclosestshelter(callthemovingaction);travel

efficiencyde-pendsonobjectiveabilities;injuriescanbereceivedfromfiresontheway.Unless

theagentdiesduringtravel,itsnextstatewillbeSafewhenreachingtheshelter;

• Preparing to defend:Theagenthasdecidedtode-fend,orwasforcedtostaybecausethefire

blocksescape;itpreparesthehouseanditself(calltheprepareforfireaction)untilthefireis

closeenough,whichtriggersthetransitiontoDefending;

• Defending:Theagentisactivelyfightingthefirearounditshouse(callthefightfireaction);

ifthatfireisextinguished,theagenttransitionsbacktoPreparingtodefenduntilanotherfire

comes;ifmotivationschange(e.g.subjectivedangerincreaseswhenactuallyseeingthefire,

orsubjectiveabilitiesdecreaseafterfailingtofight)andevacuationbecomesmoreurgent,the

agenttransitionstoEscaping;

• Safe:Theagentis(andwillstay)inashelter,andcan-notbeinjuredanymore;

• Dead:Finalstateofallagentswhosehealthdroppedto0asaresultofinjuriesreceivedfrom

thefire(fromanystateotherthanSafe);

• Survivor:Finalstateofallagentswhodidnotdieduringthefires(e.g.successfuldefenders,

luckypassive,andallshelteredagents).

Thissimplemodelissufficienttohighlighttheroleofsubjective,irrationaldeterminantsofthe

decisionsandbehavioursofeachresident,andthereforecapturesthediscrepanciesshownbythe

data.Indeed,theobjectivevalueofdangerinfluencesinjuriesanddamage,andtheobjectivevalueof

Figure 2. States of the FSM for civilian agent

capabilityinfluencesthesuccessofactions.Buttheseobjectivevaluesareinaccessibletotheagents,

whosedecisionsarebasedontheirsubjectivevaluesofdangerandabilities,andontheirmotivations.

FSM Model Discussion

ThismodelwasimplementedinGAMA,andourexperimentsshowedgoodresultsinhighlighting

theroleofsubjectivefactorsinthe“irrational”behaviourofthepopulation(Adam&Gaudou,2016b,

2017).However,wehavepreviouslyarguedfortheuseofBDIagentsinsocialsimulations(Adam&

Gaudou,2016a);theirmainadvantage,andevenmoresowhenaimingatraisingawareness,isthat

theirbehaviourisencodedintermsofconceptsfromfolkpsychology(Norling,2004)andaretherefore

moreeasilyunderstoodbyhumanusers.Tofurtherproveourpoint,weimplementedaBDIversion

ofthecivilianagentsinthisbushfiresimulation,inordertothencompareitwiththeFSMversion.

BDI MoDeL oF THe PoPULATIoN BeHAVIoUR IN BUSHFIReS

Asdiscussedabove,BDIagentsoffermanyinterestingadvantagesforsocialsimulation,inparticular

whenappliedtoraisingawareness.However,theyarelittleusedyet,duetotheirhighercomputational

complexityandthelackofdedicatedsupportingtoolsandmethodologies.Inthissection,wedescribe

amethodology(TDF)andtool(GAMA-BDI)tosupporttheintegrationofBDIagentsinsimulations,

andillustratethembydesigninganewversionofthecivilianmodelabove,withaBDIarchitecture

insteadoftheFSMarchitecture.

Methodology: Tactics Development Framework (TDF)

Whendesigningaconceptualmodelforcomputationalsimulation,UMListhemostwidelyused

tool,mostlybecauseofitsgeneralityandeaseofuse,allowingonetodescribeentitiesintermsof

attributesandactions;butitisnotwellsuitedformodellinghumanbehaviourwhichiswhatisoften

requiredindisastermanagementandevacuationsimulationssuchasourcasestudy.Ontheother

hand,agent-basedsoftwaredevelopmentmethodologiesthatdevelopsystemsusingmentalattitudes

ofgoals,events,plans,beliefs,capabilitiesetc.arewellsuitedforthesesystems.Thisisparticularly

relevantwhentranscribingbehavioursdescribedbyhumanwitnesses,asisthecasehere,since

humansnaturallytendtoexplaintheirbehaviourintermsofmentalattitudes.Forinstance,consider

thisextract:“Ilookedoutthewindowandsawsomehazysmoketothenorth-west.Garysaidthathe

thoughtitwasjustdustbutwewentoutsideandstraightawaywenoticedthatwecouldsmellsmoke.

Itwasabout12.45pmwhenwesmeltthesmokeandassoonasthathappened,Garyagreedtogo

andgetthefirepump”.Wecanmakethementalattitudesinvolvedmoreexplicit:Gary(wrongly)

believedforawhilethatthesmokewasjustdust,butplannedtogetmoreinformation;aftergoing

outsidetheyperceivedsmokeandrealisedthatitwascomingfromafire(beliefupdate).Asaresult,

Garyadoptedthegoaltogetreadyforthefire,andstartedontheirplanwhosefirstactionwasto

getthefirepump.

Whilstthereareseveralagent-orientedsoftwareengineeringmethodologiessuchasPrometheus,

Tropos,O-MaSE,GAIAandothers(DeLoach,Padgham,Perini,Susi,&Thangarajah,2009),we

introduceamorerecentmethodologythathasbeenspecificallybuiltforelicitingandencoding

tactical/strategicbehavioursindynamicdomains:TDF(TacticsDevelopmentFramework)(Evertsz,

Thangarajah,Yadav,&Ly,2015).TDFisbasedonthePrometheusmethodology(Padgham&

Winiko,2002),amatureandpopularagent-orientedsoftwareengineeringmethodology.Apilot

studyhasshownthatTDFsignificantlyimprovescomprehensionofbehaviourmodels,compared

toUML(Evertszetal.,2015).AlthoughTDFwasinitiallydesignedtocaptureandmodelmilitary

behaviour,wehaveshowninpreviouswork(Adam,Beck,&Dugdale,2015)thatthisframework

canbeadaptedtomodelcivilians’descriptionsoftheirbehaviourincrisissituations.TheTDF

methodologyproceedsinfollowing3phases:Systemspecification:Identificationofsystem-level

artefacts,namelygoals,scenarios,percepts(perceivedinternalandexternalevents),actions,data,

actorsandroles;Architecturaldesign:Specificationoftheinternalsofthesystem,namelytheagents

thatplaythedifferentroles,theinteractionsbetweentheagents(viaprotocols)ifany,andmessages

betweenagents;andDetaileddesign:Definitionoftheinternalsoftheagents,namelycapabilities,

plandiagramsandinternalmessages/sub-goals.

Tool: GAMA-BDI Plugin

GAMA(Grignard.etal.,2013;GAMA,n.d.;P.Taillandier,Grignard,Gaudou,&Drogoul,2014)is

anopensourceplatformforagent-basedmodellingandsimulationofcomplexspatialisedsystems.It

providesbuilt-infunctionsforusingGeographicalInformationSystemsdatasuchasOpenStreetMap

forfastandprecisemappingoftheenvironment.SimulationsbuiltwithGAMAarescalable,since

theplatformcandealwithseveralthousandagents,dependingonthelevelofcomplexityoftheir

architecture.Furthermore,GAMAprovidesaverysimpleandhigh-levelprogramminglanguagecalled

GAML,whichallowsevennon-programmerstosimplybuildandmaintaintheirownmodels.Asa

result,itiswidelyusedbydesignersfrommanydifferentfields:urbangrowth(P.Taillandier,Banos,

etal.,2016),geo-historicalreproductionofpastcrises(Gasmietal.,2014),socio-environmental

models(Gaudouetal.,2014),impactoffloodsonlandplanning(F.Taillandier,Adam,Delay,Plattard,

&Toumi,2016),etc.Finally,itissupportedbyanactivedevelopmentteamthatisprogressively

improvingthesoftware.GAMAwasrecentlyextendedwithaBDIplugin(Caillou,Gaudou,Grignard,

Truong,&Taillandier,2015;P.Taillandier,Bourgais,etal.,2016)toallowdesignerstoeasilycreate

BDIagentmodelsintheGAMLlanguage.Theycanspecifylogicalpredicates,initialisetheiragents

withbeliefsanddesires,describetheeffectofnewperceptsontheagent’sbeliefs,andprovidethem

withaplanlibrary.TheBDIenginethenletstheagentsperceivetheirenvironment,updatetheirbeliefs

anddesires,selectanintentionbasedontherelativeprioritiesoftheirgoals,andchooseandexecute

anadaptedplantoreachthatgoal.WehavepreviouslyshownthatthisBDIpluginconnectswell

withtheTDFmethodologyasitusesmatchingconcepts,andhavesuccessfullyusedittoimplement

theconceptualBDImodeldesignedwithTDF(Adametal.,2016).Thenextparagraphdescribes

theresultingmodelinenoughdetailforthereadertounderstandthecomparisonofthe2models.

The BDI Model

InGAML(theprogramminglanguageoftheGAMAplatform),thedesignerfirstneedstodescribe

thedifferentlogicalpredicatesthatwillbemanipulatedbytheagent.Thisisbasicallytheontologyof

thedomain.Agentscanbeendowedwithaninitialknowledgebasewithdifferentbeliefsandde-sires.

Thesepredicatescanbeassociatedwithapriority.Wedefinetwopredicatesforthecivilian-bdiagents:

• Stay Alive:Desiretofleefirestostayalive.Itspriorityisbasedontheagent’sdangeraversion

(escapemotivation);

• Protect Property:Desiretoprotectproperty.Itspriorityisbasedontheagent’sdanger

determination(defendmotivation).

Agentscanthenbegivenperceptions,whichexplainhowtheyinterpretthestimulicomingin.

Wedefinetwoperceptionsforthecivilian-bdiagents:

• Perceive Fires:Perceivenewfires(thatarenotyetintheirlistofknownfires)andaddthem

totheirlist;

• Perceive Shelters:Perceivenewshelters(thatarenotyetintheirlistofknownshelters)and

addthemtotheirlist.

Agentsarealsoendowedwithanumberofrulesforupdatingtheirmentalattitudes.Eachrule

allowsinferringnewbeliefsordesiresaccordingtoaspecificcondition(thatcanbeaspecificbelief

ordesire).Forthecivilian-bdiagent,wedefinetworules:

• Infer Escape Desire:Inferthe stayalivedesireaccordingtotheawarenessprobabilityifthe

civilianknowsthatthereisatleastonefire(theagentreferstoitsknownfireslist);

• Infer Protect Desire:Infertheprotectpropertydesireaccordingtotheawarenessprobabilityif

thecivilianknowsthatthereisaleastonefire(knownfireslist).

Theagentsareendowedwithalibraryofplanstoachievetheirgoals.EachGAMLplanisdefined

withseveraloptionalfeatures:thegoalitachieves(keyword:intention);acontextcondition(keyword:

when)describingwhenthisplanisapplicable;andasuccesscondition(keyword:finished_when).

Wedefinethreeplansforthecivilian-bdiagents:

• Prepare Property:Thisrealisesthegoalprotectproperty,byexecutingtheprepareforfireaction

itisapplicablewhenthefireisnottooclose(i.e.thedistanceishigherthandefenceradius);

• Fight Fire:Thisrealisesthegoalprotectproperty,byexecutingthefightfireaction;itisapplicable

whenthefireiscloseenough(distanceislowerthanorequaltodefenceradius);

• Go to Shelter:Thisrealisesthegoalstayalive,byexecutingthemovingaction.

BDI Model Discussion

Thismethodologyandtoolaregenericandnotlimitedtothisparticularcasestudy.TDFwasalsoused

formilitarytrainingsimulations.GAMAhasbeenusedinsuchvariousdomainssuchasagronomy,

epidemiology,civilengineering,andgeosciences.WearecurrentlyusingGAMA-BDIforsimulating

socialattachmentinearthquakeevacuations(Bangate,Dugdale,Adam,&Beck,2017),aswellas

cognitivebiasesinbushfires(Arnaud,Adam,&Dugdale,2017).Webelievethatthecombinationof

thesetoolscanhelpmoredesignerstouseBDIagentsintheirmodelsandsimulations.

CoMPARING THe SIMULATIoNS

Inthissection,wecompareourtwosimulations(FSMandBDI)ofthepopulationbehaviourin

bushfires.Westartbydiscussingtheliteratureaboutmodelcomparison,andthenproposeourown

listofcomparisoncriteria,finallyusingthemtoperformthiscomparison.

Model Comparison: State of the Art

Incrisismanagementaswellasotherapplicationfieldsofsocialsimulation,manyadhocsimulators

arecreated,withdifferentagentarchitectures,differentunderlyingmodels,anddifferenttools;

makingcomparisonshard.However,modelcomparisonisessentialtodeterminewhichmodelis

mostappropriateforwhichapplication.

Modelcomparisonhasbeenthetopicofmanyresearchworks.Someoftheseworksevaluateand

comparetheactualmodellingplatforms(e.g.Netlogo,Repast,Mason,etc.)(Railsback,Lytinen,&

Jackson,2006;Laclavíketal.,2011;Daudé&Langlois,2007;Bajracharya&Duboz,2013).Others

focusoncomparingtheperformanceoftheresultingsystems.Forinstance(Bartish&Thevathayan,

2002)havecomparedtheuseofBDIandFSMarchitecturesingames:evaluatingcomplexityinterms

ofthenumberofbehaviours,findingthatitwaslinearforBDIagentsandquadraticforFSM.Onthe

otherhand,run-timeperformance,usingasmallnumberofagents,degradedmorequicklyforBDI

agentsthanforFSMagentsinbiggersystems.

Comparisonshavealsobeenmadebyfocusingontheunderlyingmodels.Inthiscontext,most

ofworksbasetheircomparisononthesimulationresultsandusetwotypesofmetrics:afitness

function-oftencomputedbyestimatingtheerrorbetweentheobserveddataandsimulatedones-

andacomputationtime.Forexample,(Truongetal.,2016)comparethreeland-usechangemodels

basedonthreedifferentarchitecturesforthefarmeragents:probabilistic,multi-criteriaandBDI.The

comparisonsbetweenthethreemodelsareachievedbyusingthefuzzykappacoefficient(Hagen,

2003)thatallowsevaluatingthelocalsimilaritybetweentheobserveddataandthesimulatedones,

andthepercentabsolutedeviation.

Anothertechniquethathasbeenusedbytheagent-basedcommunityis“Docking”,sometimes

alsoknownas“replication”,orModeltoModelcomparison(Carley,2002).Dockingattemptsto

alignmultiplemodelsinordertoinvestigateiftheyyieldsimilarresults.Thecomparedmodelsmay

alluseanagent-basedapproachbutbeimplementedindifferentplatformsorlanguagesforexample

(Axtell,Axelrod,Epstien,&Cohen,1996;Arifin,Davis,&Zhou,2010;Xiang,Kennedy,&Madey,

2005),ortheymayusecompletelydifferentapproaches,byspecifyingtheirmodelsusing,forexample,

symbolicmathematicalexpressionsoragents(North&Macal,2002;Rank,2010).Thebenefitsof

dockingarewelldocumented(Arifinetal.,2010;Rouchier&Tanimura,2016)andincludeensuring

thevalidityofsimulationresults,increasingthequalityofthemodel,andassessingifonemodel

subsumesanother.

Someworksgofurtherandproposemeasurestocomparethecomplexityofmodels.Thus,

(Mandes&Winker,2015),proposemeasuresdividedinthreegroups:

1. Difficultyofdescription:

a. Numberofparameters;

b. Numberoflinesofcode;

c. Maximumcyclomaticcomplexity

; d. Averagecyclomaticcomplexity;

e. Maximumnestedblockdepthlevel;

f. Averagenestedblockdepthlevel;

2. Difficultyofcreation:

a. Computationaltime;

b. Memoryusage;

3. Difficultyoforganization:

a. Approximateentropy;

b. Fractaldimension.

our Comparison Metrics

Thisworkaimsatcomparingtwoagentarchitecturesthroughthecomparisonoftwomodels.Contrary

to(Mandes&Winker,2015),themodelswillbeverycomplex(integrationofcognitiveagents)and

sharemanyelements.Asaresult,someofthepreviouslyproposedmetricsarenotverywelladapted.

Inaddition,theseauthorsdonotproposeanyspecificmetricconcerningtheeaseofappropriationof

themodelsbyusers,whichisaveryimportantcriterionforthereusabilityofmodels.Wetherefore

proposethefollowingmetrics:

1. Difficultyofdescription:

a. Numberofcharactersinthecode:wechoosetousethenumberofcharactersratherthan

thenumberoflinesasthelengthoflinescanbeveryheterogeneous;

2. Difficultyofcreation:

a. Computationaltime;

b. Memoryusage;

3. Difficultyofappropriation:

a. Understandability;

b. Explainability;

c. Extensibility;

4. Modelcredibility:

a. Errorbetweenobserveddataandsimulateddata.

Notethatasstatedby(Mülleretal.,2014)thecommonuseofamodellingplatform(suchas

Netlogo,GAMAorRepast)canfacilitatemodelcomparison.Thisiswhyinthisworkwehaveused

theGAMAplatformtoimplementbothmodels.

Comparison Results Difficulty of Description

Table1showsthecomparisonofcodelength.Wecanobservethatthecodeisfarmorecompact

withtheBDImodel(morethan24%morecompact).Thisisduetousingspecificfeatures,like

perceptions,thatsimplifywritingthemodel,andtothefactthattheFSMarchitecturerequiresexplicitly

specifyingallthepossibleexistingstatesandtheirtransitions.Thissecondexplanationalsoshows

alimitationoftheFSMarchitectureintermsofmodularity:enrichingthemodelrequiresadding

newstatesandspecifyingnewtransitions,whichbecomesincreasinglycomplicatedasthenumber

ofstatesincreases,whereasaddingnewdesiresandplansintheBDIarchitectureisstraightforward

(seeExtensibility,below).

Difficulty of Creation

Table2showsthecomparisonconcerningtheuseofcomputerresources.Forthetimecomputation,

weusethesamescenariowith10replicationsandthesameseriesofseeds.Thememoryusagewas

estimatedwith5000civilianagents(andagridof100x100cells)after2simulationsteps(thetime

fortheagentstodetectfires).Thememoryusageconcernsthememoryusedbyallaspectsofthe

simulation,notonlythecivilianagents,butalsoalloftheotheragentsandtheGAMAplatform

interface.

Thecomputationtimesarerelativelyclose.Adeeperanalysiswithaprofilingtoolshowsthat

thecomputationtimeismostlyduetothecivilianagentactionsandnottothecomputationlinked

tothearchitecture.

Inthesameway,theresultsforthememoryusageareveryclose.Notethatasourgoalwasto

defineaBDImodelascloseaspossibleastheFSMone,wedidnotusebeliefpredicatesthatcould

haveusedmorememorythanbasicvariables(forexample,fortheknownfiresandknownshelters).

Difficulty of Appropriation

We designed a questionnaire for test subjects to compare the two models on two aspects:

understandabilityofcode(cantheyunderstandhowitworksandmodifyit),andexplainabilityof

behaviour(cantheyunderstandwhattheagentsdoandwhy).Weaskedalimitednumberoftesters(a

Table 1. Code comparison (difficulty of description)

Measure FSM Model BDI Model

#charactersinthecode 1769 1310

Table 2. Comparison of computer resource usage (difficulty of creation)

Measure FSM Model BDI Model

Computationtime 35s 45s

Memoryusage 103Mo 109Mo

dozenstudentsfromacomputersciencelaboratory,withvariouslevelsofpreviousexposuretoFSM

andBDImodels)toanswerthisquestionnaire.Thesearemoresubjectivecriteriasothissurveyonly

providesqualitativefeedbackonthemodels.ThetestersfoundtheBDImodelmoreunderstandable

andhadlessdifficultymodifyingitthantheFSMmodel.Itwasalsoeasiertoexplainbehaviourin

termsofwhattheagentsdesiredinsteadofwhichstatetheywerecurrentlyin.Infutureworkwewill

conductthestudyonalargerscale.

Model Credibility

Inordertoevaluatethemodelcredibility,weusedsomeofthedataprovidedinthereportsconcerning

bushfires(seeContextsectionabove),inparticulartheonesrelatedtothecausesofdeath:

• 14%diedwhileescaping;

• 69%diedwhilepassivelyshelteringinabuilding;

• 17%diedwhiledefendingtheirproperty.

Asthemodelsarestochastic,wecarriedout10replicationsforeachmodelwiththesameseries

ofseeds.Table3showsthatthetwomodelsproducecorrectresultseveniftheycouldbeimproved

byenrichingthemtobettertakeintoaccounttheheterogeneityofthebehaviours.Thedifference

betweentheresultsofthetwomodelsisnotreallysignificant.ThiswaspredictableastheCivilian- bdiandCivilian-fsmagentssharethesameattributesandactions.

extensibility experiments

Toassesstheextensibilityofbothmodels,weperformedanexperimentwhereweaskedseven

researchersandstudentsfromvariousprofiles(sociology,agronomy,computerscience,civil

engineering)toundertakeanexercise.WegavethemtheBDIandFSMmodels,andaskedthemto

extendbothmodelswithanew“warnneighbours”behaviour.Thisbehaviourwasspecifiedasprecisely

aspossible:theagentsshouldhaveanattributedeterminingtheirmotivationtowarnneighbours;a

motivatedagentchoosesanneighbourwhoisunawareofthefire,communicatestotheneighbour

theirlistofknownfires,andmakesthemaware;thisactionshouldtakeexactlyonetimestep.The

subjectswereaskedtoimplementthisbehaviourinbotharchitectures,andthentoprovidefeedback

abouthoweasyitwastoperformthisextension.Wealsoreviewedandmarkedtheircodetomeasure

howwelltheysucceededinextendingeachmodel.

Results

Thesubjects’performancesonbothmodelswerecorrelatedtogether(i.e.thoseperformingbetter

onFSMalsoperformedbetteronBDI,andviceversa).Also,allsubjectsobtainedabettermark

fortheBDImodelthanfortheFSMmodel,whichseemstosuggestthattheBDImodeliseasierto

extend.Onthecontrary,fromtheparticipants’feedback,theyfeltthatFSMconceptswereeasierto

understand.Thiswasparticularlytrueforthoseparticipantswhodidnothaveacomputerscience

backgroundorthosenotfamiliarwithBDI.Butthisimpressionofeasinesswasnotcorrelatedwith

thequalityoftheircode(despitefindingiteasy,theymadeerrorsandtheirmodelswerenotcorrect).

Table 3. Comparison of model outputs (model credibility)

Measure Real Data FSM Model BDI Model

Diedescaping 14% 18% 13%

Diedpassively 69% 72% 68%

Dieddefending 17% 10% 19%

Manyprogrammingerrorswereindependentofthearchitecture(e.g.forgettingtoconsider

onlyunawareneighbours,ortoremovealreadywarnedneighboursfromtheunawarelist).Typical

errorsintheBDImodelweresyntaxerrorsorplan-managingerrors(howtoproperlyfinishaplan,

removethesatisfiedintention)thatwouldbeavoidedifthesubjectshadmoreknowledgeofBDI.

TypicalerrorsintheFSMmodelwererelatedtoforgettingtransitions,whichisinherenttotheFSM

architectureandincreasesasthemodelcomplexityincreases(morestatesmeanmoretransitionsto

addforeachadditionalstate).

Comparison Conclusion

TheconceptsofstatesandtransitionsintheFSMmodelappearedtobemoreintuitivetothesubjects

whowereunfamiliarwithBDI.However,despitethisimpressionofsimplicitytheymademoreerrors

(inparticularmissingtransitions).Conversely,BDIisinitiallyhardertounderstandduetocomplex

conceptssuchasplansandintentions,buteasiertouse(sincethereisnoneedtoexhaustivelylist

manytransitions).BDIwasalsofoundtobemoreflexibleandlesssequentialthanFSM.

Discussion

ExtendingtheFSMmodelrequiredspecifyinganewstate,theactiontobeperformedinthatstate,and

moreimportantlyallofthetransitionstoandfromthatnewstate,linkingittothepreviouslyexisting

states.ExtendingtheBDImodelrequiredwritinganewplan,specifyingthedesiretobesatisfied

bythatplan,andtheperceptiontriggeringthatdesire;thisdoesnotchangeifthecomplexityofthe

agentincreases.Asaresult,thedifferenceinextensibilityofbotharchitecturesisnotverysignificant

foranagentoflimitedcomplexitysuchastheonepresentedhere,butwouldgrowexponentiallyas

theagent’scomplexityincreases.Indeed,asthecomplexitygrows,anincreasingnumberofstates

wouldneedtobeaddedtotheFSMmodel,alsorequiringevermoretransitionstoandfromthese

states.Conversely,thecostofextendingtheBDImodelwouldstayratherstable.

CoNCLUSIoN

Inthispaper,wediscussedtheneedtocompareagent-basedmodelsforsocialsimulationsusingboth

objectiveandsubjectivecriteria,inordertohelpdesignersdeterminewhichagentarchitectureisthe

mostadaptedfortheirneeds.Concretely,wefocusedonmodellingthebehaviouroftheAustralian

populationinbushfires,withtwoverydifferentagentarchitectures:finite-statemachines,andbelief- desire-intentionagents.Wethencomparedthesetwomodelsonanumberofobjectivecriteria,andalso

askedsubjectstosubjectivelycomparetheirunderstandabilityandexplainability,andtotrytoextend

bothmodels.OurresultsshowthatBDImodels,despitebeinginitiallymorediculttounderstand,

offeragaininmodularity,flexibility,understandabilityandextensibility.Thisisessentialincrisis

managementwherethegoalofsuchmodelsispreciselytoexplainbehaviour,raiseawareness,and

explorenewstrategies.Extensibilityisalsokeyinfacilitatingthereusabilityofexistingmodels.

Theoriginalityofourworkisthatwehavedevelopedandcomparedtwomodelsintheexactsame

context,withtwodifferentarchitecturesforthesameagents.Havingalreadyarguedpreviouslyfor

theuseoftheBDIarchitectureinsocialsimulation(Adam&Gaudou,2016a),thisworkgoesfurther

bydescribingusefultoolsforcreatingBDIagents(TDFandGAMA).Italsoconcretelycompares

BDIwithanotherfrequentlyusedandseeminglysimplerarchitecture,finite-statemachines.Our

experimentsshowthatalthoughBDImightinitiallyseemmorecomplextohandle,itisthenmore

compacttoimplement,andmakesiteasiertoexplainbehavioursandtoextendcomplexmodels.This

paperthereforeaddressestwofrequentlyraisedproblemsofBDIagents:thelackoftoolsforusing

them,andtheircomplexitylimitingtheiraccessibilitytonon-specialists.

Finally,wewouldliketomentionthattheBDIarchitecturecanbeextendedwithemotions

(Adametal.,2009),andthisextensionhasnowbeenimplementedintheGAMAsimulationplatform

(Bourgais,Taillandier,&Vercouter,2016).Otherpsychologicalfactorscanalsobetakenintoaccount

suchassocialattachment(Bangateetal.,2017)andcognitivebiases(Arnaudetal.,2017).Such

extensionswillprovideagentswithevenmorerealism,whichiscrucialwhenmodellinghuman

behaviourinsocialsimulations(andevenmoresoincrisissituations),inordertodeducevalid

results.Infuturework,wewillthereforeimprovetheBDIversionofthehumanbehaviourmodel

byintegratingvariousemotionsandpsychologicalfactors,startingagainfromtheinterviewswhere

manysurvivorsdescribehowtheyfeltbefore,during,andafterthebushfires.Therandomfiremodel

hasalsobecomealimitingfactor.Wethereforeplantoincreaserealismbyusingamorecrediblefire

simulator(e.g.(Milleretal.,2015;Duetal.,2013)).Indeed,ourlong-termgoalistofurtherdevelop

thissimulatorandturnitintoaseriousgame,whichrequiresallaspectsofthesimulationtobeas

closeaspossibletoreality.

ReFeReNCeS

Adam,C.,Beck,E.,&Dugdale,J.(2015).ModellingthetacticalbehaviouroftheAustralianpopulationina

bushfire.InProceedings of ISCRAM-Med 2015,LNBIP(Vol.233).Springer.

Adam,C.,Danet,G.,Thangarajah,J.,&Dugdale,J.(2016).BDImodellingandsimulationofhumanbehaviours

inbushfires.InProceedings of ISCRAM-Med 2016,LNBIP(Vol.265).Springer.

Adam,C.,&Gaudou,B.(2016a).BDIagentsinsocialsimulations:Asurvey.The Knowledge Engineering Review,31(3),207–238.doi:10.1017/S0269888916000096

Adam,C.,&Gaudou,B.(2016b,September).Modellinghumanbehavioursindisastersfrominterviews:

applicationtoMelbournebushfires.InProceedings of Social simulation Conference (SSC).Rome,Italy.

Adam,C.,&Gaudou,B.(2017).Modellinghumanbehavioursindisastersfrominterviews:Applicationto

Melbournebushfires.Journal of Artificial Societies and Social Simulation,20(3),12.doi:10.18564/jasss.3395 Adam,C.,Taillandier,P.,&Dugdale,J.(2017).Comparingagentarchitecturesinsocialsimulation:BDIagents

vsfinite-statemachines.InProc. of HICSS-50,Hawaii,USA.doi:10.24251/HICSS.2017.032

Adam,C.,Herzig,A.,&Longin,D.(2009).AlogicalformalizationoftheOCCtheoryofemotions.Synthese,

168(2),201-248.

Rhodes,A.(2014).Whydon’ttheydowhatwethinktheyshould?

Arifin,S.N.,Davis,G.,&Zhou,Y.(2010).Verificationandvalidationbydocking:acasestudyofagent-based

modelsofAnophelesGambiae.InProceedings of theSummer Computer Simulation Conference(pp.236–243).

Arnaud,M.,Adam,C.,&Dugdale,J.(2017,May).Theroleofcognitivebiasesinreactionstobushfires.In

Proceedings of ISCRAM 2017.

Axtell,R.,Axelrod,R.,Epstein,J.M.,&Cohen,M.D.(1996).Aligningsimulationmodels:Acasestudyand

results.Computational & Mathematical Organization Theory,1(2),123–141.doi:10.1007/BF01299065 Bajracharya,K.,&Duboz,R.(2013).Comparisonofthreeagent-basedplatformsonthebasisofasimple

epidemiologicalmodel.InProceedings of the symposium on theory of modeling & simulation devs integrative M & S symposium(p.7).

Bangate,J.,Dugdale,J.,Adam,C.,&Beck,E.(2017,May).Areviewontheinfluenceofsocialattachmenton

humanmobilityduringcrises.InProceedings of ISCRAM 2017.

Bartish,A.,&Thevathayan,C.(2002).BDIagentsforgamedevelopment.In Proceedings of Autonomous Agents and Multi-Agent Systems conference.

Bonabeau,E.(2002).Agent-basedmodeling:Methodsandtechniquesforsimulatinghumansystems.Proceedings of the National Academy of Sciences of the United States of America,99(3Suppl.3),7280–7287.doi:10.1073/

pnas.082080899PMID:12011407

Bourgais,M.,Taillandier,P.,&Vercouter,L.(2016).AnAgentArchitectureCouplingCognitionandEmotions

forSimulationofComplexSystems.InProc. of Social Simulation Conference.

Bratman,M.E.(1987).Intention, plans, and practical reason.CSLIPublications.

Caillou,P.,Gaudou,B.,Grignard,A.,Truong,C.Q.,&Taillandier,P.(2015,September).ASimple-to-use

BDIarchitectureforAgent-basedModelingandSimulation.InProceedings of the Eleventh Conference of the European Social Simulation Association (ESSA 2015).

Carley,K.M.(2002).Computationalorganizationalscienceandorganizationalengineering.Simulation Modelling Practice and Theory,10(5-7),253–269.doi:10.1016/S1569-190X(02)00119-3

Daudé,E.,&Langlois,P.(2007).Comparison of three implementations of Schelling’s spatial segregation model.

InAgent-Based Modelling and Simulation in the Social and Human Sciences.

DeLoach,S.A.,Padgham,L.,Perini,A.,Susi,A.,&Thangarajah,J.(2009).UsingthreeAOSEtoolkitstodevelop

asampledesign.International Journal of Agent-Oriented Software Engineering,3(4),416–476.doi:10.1504/

IJAOSE.2009.025321

Du,T.J.,Chong,D.M.,&Tolhurst,K.G.(2013).Quantifyingspatio-temporaldifferencesbetweenfireshapes:

Estimatingfiretravelpathsfortheimprovementofdynamicspreadmodels.Environmental Modelling & Software,

46,33–43.doi:10.1016/j.envsoft.2013.02.005

Evertsz,R.,Thangarajah,J.,Yadav,N.,&Ly,T.(2015).Aframeworkformodellingtacticaldecision-making

inautonomoussystems.Journal of Systems and Software,110,222–238.doi:10.1016/j.jss.2015.08.046 Exell,S.(2009).Witnessstatements.Retrievedfromhttp://vol4.royalcommission.vic.gov.au/index03a1.

html?pid=111

GAMA.(n.d.).GIS&agent-basedmodellingarchitecture.http://gama-platform.org

Gasmi,N.,Grignard,A.,Drogoul,A.,Gaudou,B.,Taillandier,P.,Tessier,O.,&Vo,D.A.(2014).Reproducing

andexploringpasteventsusingagent-basedgeo-historicalmodels(regularpaper).InE.Norling&F.Grimaldo

(Eds.),International Workshop on Multi-Agent-Based Simulation (MABS),Paris,France(Vol.9002,pp.151–163).

Springer-Verlag.

Gaudou,B.,Sibertin-Blanc,C.,Thalrond,O.,Amblard,F.,Auda,Y.,Arcangeli,J.-P.,&Mazzega,P.(2014).

TheMAELIAmulti-agentplatformforintegratedassessmentoflow-watermanagementissues(regularpaper).

InS.J.Alam&H.V.D.Parunak(Eds.),InternationalWorkshoponMulti-Agent-BasedSimulation(MABS),

Saint-Paul,MN,USA(Vol.8235,pp.85–110).Springer.Retrievedfromwww.springerlink.com

Grignard,A.,Taillandier,P.,Gaudou,B.,Huynh,N.,Vo,D.-A.,&Drogoul,A.(2013).Gamav.1.6:Advancing

theartofcomplexagent-basedmodelingandsimulation.InPRIMA.

Hagen,A.(2003).Fuzzysetapproachtoassessingsimilarityofcategoricalmaps.International Journal of Geographical Information Science,17(3),235–249.doi:10.1080/13658810210157822

Laclavík,M.,Dlugolinsky,Š.,Šeleng,M.,Kvassay,M.,Schneider,B.,Bracker,H.,&Hluchy,L.(2011).Agent- basedsimulationplatformevaluationinthecontextofhumanbehaviormodeling.InAdvanced agent technology

(pp.396–410).Springer.

Mandes,A.,&Winker,P.(2015).Complexityandmodelcomparisoninagentbasedmodelingoffinancial

markets.Journal of Economic Interaction and Coordination38.

Miller,C.,Hilton,J.,Sullivan,A.,&Prakash,M.(2015).Spark:Abushfirespreadpredictiontool.Environmental

SoftwareSystems.Infrastructures.Services and Applications,448,262–271.

Müller,B.,Balbi,S.,Buchmann,C.M.,DeSousa,L.,Dressler,G.,Groeneveld,J.,&Weise,H.et al.(2014).

Standardisedandtransparentmodeldescriptionsforagent-basedmodels:Currentstatusandprospects.

Environmental Modelling & Software,55,156–163.doi:10.1016/j.envsoft.2014.01.029

Norling,E.(2004).Folkpsychologyforhumanmodeling:extendingtheBDIparadigm.InProceedingsof

AAMAS,NewYork.

North,M.,&Macal,C.(2002).Thebeerdock:Threeandahalfimplementationsofthebeerdistributiongame.

InSwarmfest.SwarmDevelopmentGroup.

Padgham,L.,&Winiko,M.(2002).Prometheus:Amethodologyfordevelopingintelligentagents.InAgent- orientedsoftwareengineeringiii(pp.174–185).Springer.doi:10.1145/544741.544749

Parunak,H.,Savit,R.,&Riolo,R.(1998).Agent-basedmodelingvsequation-basedmodeling:Acasestudyand

users’guide.InMulti-agent systems and agent-based simulation(pp.10–25).Springer.doi:10.1007/10692956_2 Railsback,S.F.,Lytinen,S.L.,&Jackson,S.K.(2006).Agent-basedsimulationplatforms:Reviewand

developmentrecommendations.Simulation,82(9),609–623.doi:10.1177/0037549706073695

Rank, S. (2010). Docking agent-based simulation of collective emotion to equation-based models and

interactiveagents.InSpring simulation multiconference(p.6).SocietyforComputerSimulationInternational.

doi:10.1145/1878537.1878544

Rao,A.,&George,M.(1991).ModelingrationalagentswithinaBDI-architecture.InProceedings of the 2nd

international conference on principles of knowledge representation and reasoning(pp.473–484).