USF Libraries
USF Digital Collections

Multi-robot task allocation using affect

MISSING IMAGE

Material Information

Title:
Multi-robot task allocation using affect
Physical Description:
Book
Language:
English
Creator:
Gage, Aaron
Publisher:
University of South Florida
Place of Publication:
Tampa, Fla.
Publication Date:

Subjects

Subjects / Keywords:
recruitment
emotions
multi-agents
robotics
Dissertations, Academic -- Computer Science and Engineering -- Doctoral -- USF   ( lcsh )
Genre:
government publication (state, provincial, terriorial, dependent)   ( marcgt )
bibliography   ( marcgt )
theses   ( marcgt )
non-fiction   ( marcgt )

Notes

Summary:
ABSTRACT: Mobile robots are being used for an increasing array of tasks, from military reconnaissance to planetary exploration to urban search and rescue. As robots are deployed in increasingly complex domains, teams are called upon to perform tasks that exceed the capabilities of any particular robot. Thus, it becomes necessary for robots to cooperate, such that one robot can recruit another to jointly perform a task. Though techniques exist to allocate robots to tasks, either the communication overhead that these techniques require prevents them from scaling up to large teams, or assumptions are made that limit them to simple domains. This dissertation presents a novel emotion-based recruitment approach to the multi-robot task allocation problem. This approach requires less communication bandwidth than comparable methods, enabling it to scale to large team sizes, and making it appropriate for low-power or stealth applications.Affective recruitment is tolerant of unreliable communications channels, and can find better solutions than simple greedy schedulers (based on experimental metrics of the time necessary to complete recruitment and the total number of messages transmitted). Experimental results in a simulated mine-detection task show that affective recruitment succeeds with network failure rates up to 25%, and requires 32% fewer transmissions compared to existing methods on average. Affective recruitment also scales better with team size, requiring up to 61% fewer transmissions than a greedy instantaneous scheduler that has an O(n) communications complexity.
Thesis:
Thesis (Ph.D.)--University of South Florida, 2004.
Bibliography:
Includes bibliographical references.
System Details:
System requirements: World Wide Web browser and PDF reader.
System Details:
Mode of access: World Wide Web.
Statement of Responsibility:
by Aaron Gage.
General Note:
Includes vita.
General Note:
Title from PDF of title page.
General Note:
Document formatted into pages; contains 127 pages.

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
aleph - 001488423
oclc - 56801129
notis - AJT5717
usfldc doi - E14-SFE0000465
usfldc handle - e14.465
System ID:
SFS0025157:00001


This item is only available as the following downloads:


Full Text

PAGE 1

Multi-RobotTaskAllocationUsingAffectbyAaronGageAdissertationsubmittedinpartialfulllmentoftherequirementsforthedegreeofDoctorofPhilosophyDepartmentofComputerScienceandEngineeringCollegeofEngineeringUniversityofSouthFloridaMajorProfessor:RobinMurphy,Ph.D.KimonValavanis,Ph.D.LarryHall,Ph.D.RajivDubey,Ph.D.DateofApproval:August18,2004Keywords:robotics,multi-agents,recruitment,emotionscCopyright2004,AaronGage

PAGE 2

DedicationThisworkisdedicatedtothefamilyandfriendswhosesupportmadeitpossible.

PAGE 3

AcknowledgmentsPortionsofthisworkweresupportedbyONRGrantN00014-03-1-0786andDOEGrantDE-FG02-01ER45904.TheauthorwouldalsoliketothankRobinMurphyforherguidanceandsupportthroughoutthedevelopmentofthisthesis;MiguelLabradorforpointingoutMarkovmodelsforwirelessnetworklosses;andMattLongformakingtheunderlyingSFXrobotarchitectureworkjustintimeforsimulationsandrealrobottests.

PAGE 4

TableofContentsListofTablesiiiListofFiguresvAbstractviChapterOneIntroduction11.1Multi-RobotTaskAllocation11.2MotivatingExample31.3ResearchQuestion61.4WhyUseAffect?61.5CommunicationsChallenge81.6TheNeedforaFitnessFunction101.7Contributions101.7.1ArticialIntelligence101.7.2Robotics111.7.3CognitivePsychology121.8OrganizationofThesis12ChapterTwoRelatedWork142.1Multi-RobotTaskAllocation152.1.1Motivation-based:ALLIANCE152.1.1.1OtherMotivation-basedAllocationResearch202.1.2Auctions:MURDOCH202.1.2.1OtherAuction-basedApproaches232.1.2.2UtilityMetrics242.1.3OtherApproaches242.2DistributedSensing262.3EmotionsandAffectiveComputing302.3.1EmotionsinRobots302.3.2OCCModelofEmotions312.4FoundationofApproach342.5Summary34ChapterThreeApproach383.1RobustCommunicationProtocol383.2FormalDescriptionofAffectiveRecruitment413.3MultivariateMetricEvaluationFunctions463.4Summary48i

PAGE 5

ChapterFourExperiments514.1ExperimentalDesign524.1.1Scenario524.1.2RecruitmentStrategies544.2ExperimentalSimulations554.2.1EffectsofTeamSize554.2.1.1StatisticalAnalysis554.2.1.2ResultsforNumberofMessagesMetric574.2.1.3ResultsforAverageWaitTimeMetric574.2.1.4SummaryofTeamSizeSimulations624.2.2EffectsofCommunicationLoss634.2.2.1StatisticalAnalysis634.2.2.2ResultsforNumberofMessagesMetric644.2.2.3ResultsforAverageWaitTimeMetric644.2.2.4SummaryofCommunicationLossSimulations644.2.3BroadcastversusUnicastMessaging694.2.4IllustrativeUseCases704.2.5FairnessofRecruitment714.3RobotImplementation734.3.1RestrictedScenario734.3.2SFXImplementation734.3.3RobotTrials754.4Summary75ChapterFiveDiscussion805.1LimitationsofExperiments805.2ComparisontoExistingResults815.3ParametersandFitnessMetrics835.3.1FitnessFunction835.3.2SHAMEAccrualFunction835.3.3SHAMEDecayFunction845.4Contributions855.4.1ValidatesApplicationofEmotions855.4.2ReducedCommunicationOverheadandBetterScaling855.4.3SuperiorSolutionQuality865.4.4DemonstratedRobustness865.4.5HandlesHeterogeneity865.4.6FairnessofAllocation875.5Summary87ChapterSixSummaryandFutureWork896.1SummaryofThesis896.1.1Contributions916.2FutureWork93References95Appendices102AppendixA:RawSimulationResults103AbouttheAuthorEndPageii

PAGE 6

ListofTablesTable1.RelatedMulti-robotTaskAllocationWorkAccordingToResults.16Table2.TestDomainsForALLIANCE.19Table3.MeanAndStandardDeviationOfTheElapsedTime,InSeconds,ForSuccessfulPushingTrialsInEachOfFourBoxPushingExperimentsForMURDOCH.21Table4.DistributedSensingLiterature.27Table5.SummaryOfLiteratureApplyingEmotionsToRobots.30Table6.Standards-basedEmotionsAlsoCalledAttributionEmotions.33Table7.Standards-basedEmotionsInWhichAnAgentHasANegativeReactionToItsOwnAc-tions.33Table8.RecruitmentProtocolMessagesAndParameters.39Table9.SummaryOfTheNotationUsedInAffectiveRecruitment.42Table10.AverageNumberOfMessagesTransmittedForEachStrategyForVaryingTeamSize.58Table11.PairwiseCondenceIntervalsForAverageNumberOfMessagesForVaryingTeamSize.59Table12.AverageTime,InSeconds,TheUAVSpentWaitingAccordingToTeamSize.59Table13.PairwiseCondenceIntervalsForAverageTimeUAVSpentWaitingAccordingToTeamSize.61Table14.AverageNumberOfMessagesTransmittedForEachRecruitmentStrategyAccordingToNetworkLossRates.65Table15.PairwiseCondenceIntervalsForAverageNumberOfMessagesForEachMessageLossRate.67Table16.AverageTime,InSeconds,TheUAVSpentWaitingAccordingToRandomMessageLossRate.68Table17.PairwiseCondenceIntervalsForAverageWaitTimeForEachMessageLossRate.69Table18.AverageNumberOfMessagesTransmittedAccordingToMessagingType.70Table19.BiasOfEachRecruitmentStrategy.72Table20.NumberOfTimesEachRobotWasRecruitedUsingAffectiveRecruitment.103iii

PAGE 7

Table21.NumberOfTimesEachRobotWasRecruitedUsingAffective1=D2Recruitment.104Table22.NumberOfTimesEachRobotWasRecruitedUsingGreedyRecruitment.104Table23.NumberOfTimesEachRobotWasRecruitedUsingRandomRecruitment.104Table24.RawDataForTimeMetric,4Robots,And0%CommunicationFailureRate.105Table25.RawDataForNumberOfMessagesMetric,4Robots,And0%CommunicationFailureRate.106Table26.RawDataForTimeMetric,8Robots,And0%CommunicationFailureRate.107Table27.RawDataForNumberOfMessagesMetric,8Robots,And0%CommunicationFailureRate.108Table28.RawDataForTimeMetric,13Robots,And0%CommunicationFailureRate.109Table29.RawDataForNumberOfMessagesMetric,13Robots,And0%CommunicationFailureRate.110Table30.RawDataForTimeMetric,23Robots,And0%CommunicationFailureRate.111Table31.RawDataForNumberOfMessagesMetric,23Robots,And0%CommunicationFailureRate.112Table32.RawDataForTimeMetric,53Robots,And0%CommunicationFailureRate.113Table33.RawDataForNumberOfMessagesMetric,53Robots,And0%CommunicationFailureRate.114Table34.RawDataForTimeMetric,13Robots,And5%CommunicationFailureRate.115Table35.RawDataForNumberOfMessagesMetric,13Robots,And5%CommunicationFailureRate.115Table36.RawDataForTimeMetric,13Robots,And10%CommunicationFailureRate.116Table37.RawDataForNumberOfMessagesMetric,13Robots,And10%CommunicationFailureRate.116Table38.RawDataForTimeMetric,13Robots,And25%CommunicationFailureRate.117Table39.RawDataForNumberOfMessagesMetric,13Robots,And25%CommunicationFailureRate.117iv

PAGE 8

ListofFiguresFigure1.UnmannedAerialVehiclesForTheDeminingTask.3Figure2.UnmannedGroundVehiclesUsedInTheDeminingTask.4Figure3.UGVAndUAVTogetherInTheDeminingTask.4Figure4.GraphOfTheCommunicationsUseByMURDOCH.22Figure5.TheOCCModel.32Figure6.RecruitmentProtocolInTermsOfTheMessagesSentBetweenRobots.40Figure7.ExampleOfAverageBestFitnessBeingUsedToGenerateReplies.45Figure8.UserInterfaceForRecruitmentSimulator.53Figure9.HistogramOfTheNumberOfMessagesTransmittedUsingTheAffectiveRecruitmentStrategyForTeamSize13.56Figure10.MessagesTransmittedAtDifferentTeamSizes.58Figure11.BoxPlotsOfTheSimulationResultsForTheCommunicationOverheadAccordingToTeamSize.60Figure12.TotalWaitTimeAtDifferentTeamSizes.65Figure13.BoxPlotsOfTheSimulationResultsForTheWaitTimeMetricAccordingToTeamSize.66Figure14.MessagesTransmittedAtDifferentNetworkFailureRates.67Figure15.WaitTimesAtDifferentMessageLossRates.68Figure16.SimpliedOverviewOfTheSFXArchitecture.74Figure17.OperatorUserInterfaceForRealRobotTests.76Figure18.OperatorUserInterfaceForRealRobotTests.76Figure19.OperatorUserInterfaceForRealRobotTests.77Figure20.UGVArrivingAtASimulatedMine.77v

PAGE 9

Multi-RobotTaskAllocationUsingAffectAaronGageABSTRACTMobilerobotsarebeingusedforanincreasingarrayoftasks,frommilitaryreconnaissancetoplanetaryexplorationtourbansearchandrescue.Asrobotsaredeployedinincreasinglycomplexdomains,teamsarecalledupontoperformtasksthatexceedthecapabilitiesofanyparticularrobot.Thus,itbecomesnecessaryforrobotstocooperate,suchthatonerobotcanrecruitanothertojointlyperformatask.Thoughtechniquesexisttoallocaterobotstotasks,eitherthecommunicationoverheadthatthesetechniquesrequirepreventsthemfromscalinguptolargeteams,orassumptionsaremadethatlimitthemtosimpledomains.Thisdissertationpresentsanovelemotion-basedrecruitmentapproachtothemulti-robottaskallocationproblem.Thisapproachrequireslesscommunicationbandwidththancomparablemethods,enablingittoscaletolargeteamsizes,andmakingitappropriateforlow-powerorstealthapplications.Affectiverecruitmentistolerantofunreliablecommunicationschannels,andcanndbettersolutionsthansimplegreedyschedulersbasedonexperimentalmetricsofthetimenecessarytocompleterecruitmentandthetotalnumberofmessagestransmitted.Experimentalresultsinasimulatedmine-detectiontaskshowthataffectiverecruitmentsucceedswithnetworkfailureratesupto25%,andrequires32%fewertransmissionscomparedtoexistingmethodsonaverage.Affectiverecruitmentalsoscalesbetterwithteamsize,requiringupto61%fewertransmissionsthanagreedyinstantaneousschedulerthathasanOncommunicationscomplexity.vi

PAGE 10

ChapterOneIntroduction1.1Multi-RobotTaskAllocationCollaborationamongmembersofamulti-robotteamismotivatedbyaneedtocompletetasksthatrequiremorecapabilitiesthanasinglerobotcanprovide.Robotteamsmaybeheterogeneous,meaningthattheindividualrobotshavedifferentcapabilities.Thesedifferencesmaybeinhardware;forinstance,robotscanbedifferentintermsoftheirsize,mobilitydrivingversusying,sensors,orcomputationalpower.Robotscanalsobedifferentinsoftware,suchthattheyhavedifferentavailablebehaviorsorperceptualalgorithms.Robotsinateamoftenhavedifferentsensors,eitherbydesignorbycircumstance:evenifallrobotsareidenticalatthestartofatask,hardwarefaultscanmakethemdifferent.Tocompleteatask,thesensorsdistributedacrossarobotteamshouldbebroughttowheretheyaremostneeded.Twodomainsinwhichrobotswithsensorsmayneedtoberecruitedtoachieveteamobjectivesareerrorhandlinganddistributedsensing.Intermsoferrorhandling,arobotmayexperienceasensorfailureorbecomestuckandrequireanotherrobot'sexternalviewpointfordiagnosisandrecovery[53].Fordistributedsensing,anautonomousaerialvehiclemaydetectasuspiciousobjectonthegroundandemployagroundvehicletoinvestigatemoreclosely.Inparticular,forthecooperativeidenticationanddisposaloflandmines,searchandrescue,mapbuilding,andforaging,itmaybenecessarytouseamultitudeofviewpointsandsensingmodalitiestoaccomplishateamgoal.Theproblemofrecruitment,whereonerobotistaskedtohelpanother,isaspecialcaseofthemulti-robottaskallocationMRTAproblem[31],whichhasreceivedmuchrecentattentionseeChapterTwo.MRTAandrecruitmenthavebeenappliedtoanumberofdomains,practicallyanywherethatmultiplerobotsoperateinthesamelocation.TheMRTAproblemhassixcharacteristicsthatmakeitchallenging.Teamsofmobilerobotsmaybeheterogeneous,suchthattherobotsintheteamhavedifferenthardware,software,orareperformingunrelatedtasks.1

PAGE 11

RobotstypicallyshareaniteandunreliablecommunicationschannelsuchaswirelessEthernetwhichdespitebest-effortnetworkprotocolsi.e.TCP/IPmayperiodicallyfail,andcansaturateifusedheavily.Therobotteamcanbecomequitelarge,androbotsmaybeaddedorremovedfromtheteamatanytime.Thisimpactsthecommunicationsoverheadofanysolution,andimpliesthatrobotscannotbeexpectedtomodelthestatesoftherestoftherobotsontheteam.Controlisgenerallydistributed,notcentralized.Distributedcontrolisessentialforrobustness,aspartialfailuresmakecentralizedapproachesbrittle.However,distributedcontrolismoredifculttomanage,astheinformationrequiredtomakeinformeddecisionsisoftenspreadacrosstheteam.Itmaynotbepossibletoreassignarobottoanewtaskonceithasatask[28].Taskpreemptioncannotbeassumed,sothebestrobotforataskmaynotbeavailable.Accuratepredictionoffuturetaskallocationrequirementsisoftennotpossible.Assignmentsmustbemadeinreactiontonewtasks,opportunities,orrobotfailuresastheyarrive.Althoughdecentralizedsolutionstothisproblemexist,mostrequirelargeamountsofcommunicationamongrobots[30]withouttoleranceforcommunicationlossesasin[40],andthequalityofthesolutionsmaybepoor[28].Todate,mostsolutionstotheMRTAproblemhavebeenbasedongame-playingstrategiesfromthearticialintelligencecommunity,orvariantsonauctionprotocolsespeciallytherst-priceauction[63][105]combinedwiththecontractnetprotocol[90][21].Intheseapproaches,anewtaskisannouncedtotheteam,robotsrespondwithanestimateoftheirsuitabilityorcostforthetask,andthetaskgoestotherobotthatsubmittedthebestbid.However,noneofthesehasappliedanaffectivesolutionthatis,usingemotions.Emotionsareusefulinrobots,astheyprovideamechanismforself-regulation,suchthatachangeinarobot'sstateorbehaviorcanbeinducediftherobot'smotivationallevelishighenough[64][62][60][61][3][6][5][25][52][98][99].Emotionsprovideacomputationallysimpleandlow-communicationmethodforlettingarobot'srecenthistorybiasitscurrentchoiceofaction,ratherthanhavingtherobotmakedecisionsbasedsolelyonitsinstantaneousstate.2

PAGE 12

Figure1.Unmannedaerialvehiclesforthedeminingtask.Fourhelicoptersareshown,eachwithdifferentcapabilities.1.2MotivatingExampleAnautomateddeminingtaskprovidedbyNAVSEACoastalSystemsStationCSSservestomotivatetheMRTArecruitmentproblem.Inthistask,alargeareaissurveyedbyarobotteamsothatlandminescanbeidentied,disabled,andremoved.TherobotteamconsistsofatleastoneunmannedaerialvehicleUAV,e.g.ahelicopter,andmultipleunmannedgroundvehiclesUGVs.Therobotsarephysicallyheterogeneous,bothinmobilityandsensorsuite.TheUAVcarriesacolorcamera,aforward-lookinginfraredFLIRcamera,GlobalPositioningSystemGPSreceiver,andInertialMeasurementUnitIMU.TheUAVisownthroughacombinationofonboardandoffboardcontrol,andtransmitssensorreadingstoagroundcontrolstation.TheUAVcanbeseeninFigure1.TheUGVsareiRobotATRVJr.groundrobots,eachequippedwithacolorcamera,FLIR,GPS,IMU,compass,laserrangender,andothertask-specicsensors.TheUGVsareautonomousandarecontrolledbyanonboardPentium-IIIclasscomputer.TheycommunicatewitheachotherandbacktoOperatorControlUnitsOCUsviawirelessEthernet.11b/g.TheUGVscanbeseeninFigure2,andtheteamworkingtogethercanbeseeninFigure3.Thecapabilitiesthateachplatformbringstothedeminingtaskareasfollows.TheUAVcanmovequicklyoverthesearchareawithouttriggeringmines,butisunabletooperatenearobstaclessuchastreesandhigh-tensionlines.TheUAVmustalsomaintainaminimumaltitudewhileunderautonomouscontrolforsafetyreasons.Further,theUAVcannotliftaheavypayload,soitcarriesaminimalsensorsuiteand3

PAGE 13

Figure2.Unmannedgroundvehiclesusedinthedeminingtask.ThreeiRobotATRVJr.groundrobotsareshown. Figure3.UGVandUAVtogetherinthedeminingtask.4

PAGE 14

delegatetheanalysisofitssensordatatoanoffboardcomputer.TheUGVsarecomplimentaryinthattheycancarryanarrayofsensorsanddoonboardprocessing,butmoveatalowervelocityandmustavoidhazardsobstaclesandminesontheground.Forthedeminingtask,theUAVsearchesanareaforinterestingtargetssuspectedmines,butitcannotcloselyinvestigatethemwithitslimitedperceptionanditsneedtomaintainaltitudeforsafety.Whenitdetectsasuspectedmine,theUAVrequeststheassistanceofaUGVthatcaninvestigatefully,thencontinuesitssearch.GiventhattherearemultipleUGVs,theUAVcanquicklysurveyanareaandsummonUGVstoeachinterestingartifact,allowingthemtoexaminethetargetsinparallel.TheUGVscanalsosearchformines,butarelimitedbytheirmobilityandgroundhazards.ThechallengesoftheCSSdeminingmissionarecharacteristicofMRTAingeneral:Therobotteamisheterogeneous.TheUAVandUGVsarephysicallydifferentplatformsthatcarrydifferentsensorsandbehavedifferently.TherobotscommunicateviawirelessEthernet,whichisknowntobeunreliable[49].Therobotteamsizeisdynamicandcangrowquitelarge.Robotsineldenvironmentsarepronetofailures[14],andcommunicationfailurescanhavetheeffectoftemporarilyremovingarobotfromtheteam.IntherobotresultsdescribedinChapterFour,onlythreegroundvehicleswereused,butthiswasalimitationofavailablehardware,notofthetaskdomain.Controlisdistributed.AlthoughOperatorControlUnitsallowahumantosuperviseandmanagetheUGVs,theUGVsareautonomousandmaynotalwaysbeincommunicationscontactwiththeOCUs.Whentaskallocationisrequired,theUGVsareexpectedtoresolveitontheirown,withoutrelyingontheOCUsoranyothersinglepointoffailure.UGVscannotabandonanassignedtask,butmaybepreemptedbyanoperator.ItisexpectedthatmultipleUGVswouldsearchformineswhileothersremainidleinanticipationofadiscoverybytheUAV.Intheeventthatnorobotsareavailabletoinvestigateanewtarget,thehumanoperatorhastheoptionofpreemptingarobottosatisfythatneed.Thereisnoknowledgeoffuturetasks.TheUAVrequestsassistancewhenitdiscoversanewtarget,whichcanhappenatanytime.5

PAGE 15

Thus,thedeminingtaskisachallengingexampleoftheMRTAproblem.Deminingisareal-worldconcern,asitaffectsbothhumanitarianandmilitaryefforts.ThisthesisprovidesanovelapproachtotheMRTAproblemthroughthedeminingtaskdomain.1.3ResearchQuestionTheresearchquestionthatthisworkaddressesisasfollows:Howcanaffectivecomputingbeusedforrecruitmentinateamofdistributed,heterogeneousmobilerobotswithunreliablecommunications?Therearethreeprimaryissuesraisedbythisquestion:Whatisaffectivecomputing,whyshoulditbeused,andwhatmodelsofemotionshouldbeapplied?Affectivecomputingreferstotheuseofanemotionalmodelinacomputationalsystem.Inthiscase,itmeansusingemotiontocontrolhowrobotsarerecruited.ThemotivationforaffectivecomputinginthisapproachisprovidedinSection1.4,andmodelsofemotionarediscussedinChapter2.3.Howcanunreliablecommunicationsbeovercometoproducerobustmulti-robottaskallocation?Coordinationofagentsinanydistributedsystemrequiresthattheagentsbeabletoperiodicallyexchangeinformation.Inreal-worldapplications,robotscommunicateviawirelessnetworks,butthesearepronetolosses[49].ThechallengesthatunreliablecommunicationsposearediscussedinSection1.5,andChapter3.1providestheprotocolthatwasused.Howshouldrobotsberecruited?Thedifferencebetweenagoodtaskallocationstrategyandastrategythatpicksatrandomishowrobotsarechosenforeachtask.Thisassumestheabilitytodiscriminatebetweenrobotsandtoselectthebestone.However,whatdoesitmeanforarobottobethebest?Theremustbesomemeasureofthetnessofarobottoataskbeforethisdecisioncanbemadeintelligently.TheneedforatnessfunctionismotivatedinSection1.6.1.4WhyUseAffect?Affectivecomputingreferstotheuseofanemotionalmodelinacomputationalsystem.Emotionsaredescribedin[52]ascrucialforsocialintelligence,especiallyforagentswithlimitedresources.Theapproachinthisthesisusesemotiontomodulaterobotbehavior.Thetaskdomainrequiresamulti-robotteam,wheretherobotsmustcooperatetoreachteamobjectives.Thiscooperationrequiresthateachrobotbe6

PAGE 16

awareoftheothersintheteamandactaccordingtosocialrules[55]thatallowtherobottoberecruitedtohelpateammate.Inotherwords,therobotsmustbeprovidedwithatleastrudimentarysocialintelligence.In[64],itisobservedthatapplyingresearchinemotionalintelligencemayleadtomoreautonomousandefcientrobotsandrobotteams,andwhileitmaybepossibletoobtainthesameresultsthroughmoretraditionalengineeringsolutions,suchasproductionsystems,thecognitiveapproachshouldnotberuledout.In[89],page198,SlomanandCroucherobserve,Incooperativecommunities,individualsshoulddevelopmotiveswhichdonotnecessarilymaximizetheirownadvantage,butwhichenablethecommunityasawholetofunctionwell.Thiscan,ofcourse,leadtoconictingmotiveswithinandbetweenindividuals...Choosingbetweenalternativeswillnotbesimple.Thenotionofanoptimalchoicewillnotnecessarilyevenbewell-dened.Achievingalongtermbalancebetweendifferentneedsoftheindividualorthecommunitycanbeamajorproblem.Decisionmakingprocesseswillhavetobecapableofcopingwithsuchconicts.[89]thenpresentsemotionsasameansofresolvingtheconictsbetweenarobot'scompulsiontomaximizeitsownutilityversustheneedsoftheteam.Borrowinganemotionalmodelfromthecognitivesciencecommunitycanresultinacognitivelyplausibleemotionalsystemthatreectsthesocialinteractionsinnaturallyoccurringsocieties.Thus,anemotion-basedapproachtosolvingcoordinationproblemsinthedomainofdistributedmulti-robotteamsisappealing,especiallysinceitisadirectionthathaslargelybeenleftunexplored.Emotionsaredesirableinrobotsbecausetheyprovideasimplemotivationalmechanismtoperformanaction,oftenasareactiontoaneventorinternaldrive[64][62][60][61][3][6][5][25][52][98][99].Forinstance,supposethatarobotisunabletoachieveagoalbutcannotdetectitslackofprogress.Iftherobotisequippedwithemotions,anongoingunfullledgoalmayincreasetheintensityoftherobot'sfrustratedorangryemotions,andmotivateittochangetoadifferentstrategyorcallforhelpwhenthesebecomeintenseenough.Notethatinthiscase,arobotneednotkeepahistoryofitsactionsormaintainacomplexstateinordertouseemotionseffectively.If,ateveryupdate,therobot'sleveloffrustrationincreasesbecauseitsobjectivehasnotbeenfullled,thentherobotonlyneedstoknowthatitisfrustratedandshouldattemptadifferentstrategy,withouthavingtoreasonaboutwhythisisthecase.Thismakesemotionsusefulforrobotsthatmaintainaminimalamountofstateinformation,suchasthoseusingthereactiveparadigm[65].7

PAGE 17

Toexpandontheaboveexample,thefrustrationemotioncouldalsobeinuencedbyotherfactors.Forexample,iftherobot'sgoalwastoreachitsrechargingstation,itsfrustrationcouldbecompoundedbyadangerouslylowenergyreserve.Inthisway,emotionsserveasasimplemeansofcombiningmultipleinuencesontherobotintoasingletypeofreaction.Havingthelow-batterystatusandlackofprogressbothfeedingintofrustrationcanmotivatetherobottochangetasks.Thisissimplerthanenumeratingrulessuchasifnotmakingprogressaftert1seconds,changestrategiesandifnotmakingprogressaftert2secondsandbatteryisbelow25%,changestrategiestoachievethesamebehavior.Inthisway,seeminglycomplexbehaviorcanbeassembledfromprimitivedrivesandreactions.Emotionscanalsoprovideanaturalisticinterfaceforhumans.Ifarobot'sstatecanberepresentedintermsthatahumancanassociatewithhappy,angry,depressed,thenthehumancanunderstandtherobot'ssituationmoreintuitively.Robotsandagentsthatexpressemotionsforthepurposeofhuman-robotinteractionhavebeenexploredin[6][52][98][99][25].Thisworkusesanemotion-basedmodel,groundedinaplausibletheoryofemotions,toleverageresearchinthisareathathasbeenmappedovertothedomainsofarticialintelligenceandagents[73][71][72][84][79][50].Thisimpartstwobenets:natureprovidesanexistenceproofthattheemotionsareeffectiveforregulatingbehavior,andmuchofthedifcultyindevisingacoherentemotionalsystemsuitableforbehavior-basedrobotshasalreadybeendone.Thus,thisworkborrowsfromanexistingformaltheoryofemotions,theOCCmodelnamedafterOrtony,Clore,andCollins[72].Inthiswork,aSHAMEemotionisgeneratedinreactiontoanevent:arequestforhelpthattherobotignored.TheintensityoftheSHAMEemotioncontrolswhetherornotarobotwillrespondtoasubsequentrequest,andintheabsenceofadditionalstimulusfurtherrejectedrequests,SHAMEwilldecayovertimebackdowntonothing.Inthiswork,SHAMEisdrivenbyasinglekindofevent,butintheory,otheraspectsoftherobot'sstatuscouldinuenceitaswell.Forinstance,ifarobotdetectedthatitsactionswereinhibitingtheprogressoftheteam,itcouldreactwithanincreaseinSHAME.1.5CommunicationsChallengeRecruitmentofrobotsinadistributedteammustberobustintermsofcommunications,suchthatrecruitmentsucceedsdespitemessagelossortheunexpectedlossofanyrobot.Inadistributedteam,theremaybepartialfailures,whereoneormorerobotsfail,areunabletocommunicateduetointerferenceorobstructions,orareotherwiseunabletorespond.Robotsmayalsobeaddedtoorremovedfromtheteamatanytime.8

PAGE 18

Communicationsshouldbeusedconservatively.Notethatthisisnotthesameasusingaminimalamountofcommunications,becauseitispossibleforrobotstocoordinatewithoutexplicitlycommunicating,asin[41][42][47][55].Robotsthatcansufcientlyperceiveormodeleachothercaninferenoughtomakeexplicitcommunicationunnecessary.Suchsystems,however,areboundbythelimitsoftheirperception,andwouldnotbeadequateforthedomainsaddressedinthiswork,becauserobotswillbetoofaraparttoperceiveeachother,androbotteamsmaybetoolargetorealisticallyallowmodeling.Itisthereforeassumedthatasharedcommunicationschannelisrequired,butshouldbeusedconservativelyforthefollowingreasons:Communicationsbandwidthisnite.Othermission-relateddemandsonasharedcommunicationschannel,suchasstreamingvideoorcontrolcommands,mayconsumeanyavailablebandwidth.Recruitmentshouldnotinterferewithsuchdemands.Teamscanvaryinsizewithoutbound.Currentresearchinmulti-robottaskallocationusuallyonlyconsidersafewrobotsatatime,aswillbediscussedinChapterTwo.However,largerrobotswarmssuchastheSRICentibotswarm,ateamof103robotsarebecomingmorecommonandinthefuture,teamsmaygrowtoincludehundredsofrobotsormore.Thecommunicationsrequirementsforrecruitmentshouldscalewellwithteamsize,sothatlargeincreasesinthenumberofrobotsdoesnottranslatetoalargeincreaseinrequiredbandwidth.Robotsmayhavepowerconstraints.Asmallmobilerobotoranodeinasensornetworkmaynothaveenoughpoweravailabletofrequentlytransmitmessages.Thefewermessagessucharobotsends,thelongeritsbatterieswilllast.Therobotsmaybedeployedinadomainrequiringstealthwheretheyarecapableofreceivingmessages,butriskrevealingtheirlocationbyresponding.Inthiscase,anytransmissionsshouldbewelljustied.Therobotswillalsocommunicateusingabroadcastmessagingscheme,meaningthatasinglemessagecanbetransmittedwithallotherrobotsasrecipients.Broadcastmessagingrequireslesscommunicationsbandwidthforcoordinatingmultiplerobotsthanunicastmessaging,whereonlyasinglerecipientreceivesamessage.ThisisdiscussedfurtherinChapterThreewithexperimentalvalidationinChapterFour.9

PAGE 19

1.6TheNeedforaFitnessFunctionWhenataskarisesforwhicharobotmustberecruited,itisnecessarytochoosearobotthatiscapableofcompletingthetask.However,matchingthecapabilitiesofarobottotherequirementsofthetaskisstillanopenissue.Thisworkusestheestimatedtimethatarobotneedstoreachatasklocationasametricthatdescribestherobot'ssuitabilityforatask.Similarly,in[30],ametricisdescribedasafunctionoftherobot'sstate,evaluatedinthecontextofthetask.Thisdenitionisverybroad,but[30]providesanotherexampleofametric:thedistancebetweentherobotandwherethetaskwillbegin.Thesemetricscanbedenedasrequiredandaretask-specic,butarestillsomewhatadhoc.Ideally,thesuitabilityofarobottoataskwouldbeafunctionofresourcecostssuchastime,energy,ormaterialsthattherobotincurstocompletethetaskandhowwelltherobot'scapabilitiesoverlaptherequiredcapabilitiessuchassensingmodalities,sensorresolution,eldofview,orsimplytheperceptsthattherobotcangenerate.Thedistributedsensingcommunitywouldbenetfromauniedmodeloftasktness,suchthatcompletelyheterogeneousrobotscouldbecompareddirectly,eveniftheyhadnosensorsincommon.Adiscussionofthemetricideafrom[30]isprovidedinChapter3.3.1.7ContributionsThisthesisdescribesanaffectiverecruitmentprotocol,basedonthecontractnetprotocol[90][21],thatenablesrobotstorequestandreceiveassistancefromothermembersoftheteamusinganemotionalmodelbasedonworkbyOrtonyetal.[73][71][72].Thisapproachisnovel,becauseemotionshavenotpreviouslybeenappliedtothemulti-robottaskallocationproblem,andproducesbetterresultsthanthestateoftheartintermsofcommunicationoverhead[28]seeChapter4.2.Thismethodmakesatleastsixcontributionstothearticialintelligence,robotics,andcognitivepsychologycommunities,asfollows.1.7.1ArticialIntelligenceSuperiorsolutionquality:Theaffectiverecruitmentapproachbenetsthedistributedagentscommunity,asitcanreachbettersolutionsthanexistinggreedyrst-priceauctionstrategiessuchasMURDOCH[30].Suchgreedyschedulerscanbeadverselyaffectedbychangingtheorderofnewtasks,whichmayleadtogreatlyreducedsolutionquality[28].Thisapproachdependslessontheorderofnewtasks,andcanndsolutionsthatexistingmethodsmissasisshowninChapter4.2.The10

PAGE 20

behaviorofaffectiverecruitmentiscontrolledthroughparameters,andprovidesmoreexibilityindesignthanatraditionalgreedyapproach.Fairnessofallocation:Aninterestingside-effectofthisapproachisthatrobotsthatareequallysuitedforataskwilltendtotaketurnsbeingrecruited,suchthatthedisruptiontoeachrobotisdistributedacrosstheteam.1.7.2RoboticsReducedcommunicationoverheadandbetterscaling:Theuseofanemotionalmodelreducesthecommunicationsrequiredfortaskallocationcomparedtothestateoftheart,andrepresentsanimprovementoverthegreedyapproach.Thisreductionofoverheadcontributestwobenetstothedistributedsensingandroboticscommunities.First,theprotocolcanscaletolargeteamsorswarmsofrobotsmorereadilythanexistingmethodsasshownbyresultsinChapter4.2.Second,thisapproachreducesunnecessarytransmissions,whichbenetsbothlow-powerandstealthapplicationsthough,asinSection1.5,noclaimismadeaboutatheoreticalminimum,becausethetheoreticalminimumiszerogiventherightassumptions.Demonstratedrobustness:ThisapproachusesacommunicationprotocolsimilartothatinMURDOCH[30],whichprovidesrobustnessintermsofcommunicationfailures.Messagesbetweenrobotsfollowasequenceofsteps,andthelossofanymessagecanbedetectedandcompensatedfor.Experimentalresults,showninChapter4.2,indicatethattherecruitmentprotocolwillcontinuetofunctionwithupto25%randommessagelossregardlessoftherecruitmentstrategyused,butuptothatpoint,theaffectiverecruitmentstrategytransmitsfewermessages.Theseresultsbenetthedistributedsensingcommunitybydemonstratingtheperformanceofadistributedprotocolwithrealisticcommunicationlosses.HandlesHeterogeneity:Thisapproachmakesnoassumptionsaboutthecompositionoftherobotteam.Robotscanbecompletelyheterogeneousinhardwareandsoftware,anddonotneedtobeoperatingoverthesamesetoftasksorgoals.Thisbenetstheroboticscommunity,whererobotsareoftenheterogeneous,eitherbydesignorduetopartialfailures.11

PAGE 21

1.7.3CognitivePsychologyValidatesapplicationofemotions:Thisapproachaddsmotivationstoanauction-basedmulti-robotrecruitmentstrategythroughanemotionalmodel.ThisthesisvalidatestheemotionalmodelofOrtonyetal.,andbenetsthecognitivesciencecommunitybydemonstratingthattheemotionsfunctionasexpectedinarticialagents.Emotionshavebeenusedinrobotsinthepast,buttypicallyforhuman-robotinteractionandentertainmentresearch.Thereisonlyoneinstanceintheliteratureofemotionsbeingusedtocontrolateamofrobots[64],andinthatcase,therewasnotaskallocationastherobotshadxedroles.Emotionscanalsoprovidemeaningfulstateinformationtohumansupervisors.Inthisthesis,ahumanoperatorcanusetheemotionalstateofrobotsinateamtomakeinformeddecisionsabouthowtheteamisperforming.1.8OrganizationofThesisTherestofthisthesisisorganizedasfollows.ChapterTwosurveysthepriorworkinmulti-robotcoordination,distributedsensing,andemotionsinrobots.ChapterThreeformallyintroducestheaffectiverecruitmentapproachmotivatedaboveanddiscusseswhycreatingaformalmechanismfordeterminingthetnessofarobotforataskisahardproblem.ChapterFourdetailsexperimentsthatwereconductedtovalidatetheapproachinsimulationandonmobilerobots.Thetestscomparedaffectiverecruitmenttogreedyandrandomstrategies,wheregreedyisconsideredthestateoftheart.Themetricsfortheexperimentswerethetimenecessarytocompleterecruitmentandthenumberofmessagestransmitted.Theobjectivesoftheexperimentsweresix-fold:Testtheeffectsofvaryingteamsizefrom4Testtheeffectsofrandomcommunicationlossesupto25%Testtheeffectoflinearversusnon-linearSHAMEupdatefunctionswithregardstochaoticbehaviorinverylargerobotteamsJustifytheuseofbroadcastmessaginginsteadofunicastVerifythataffectiverecruitmentcanreachbettersolutionsthangreedywherebetterisintermsofthemetricsTestthedegreetowhichrobotsarerecruitedequallyoften,orfairly,accordingtothesestrategies.12

PAGE 22

Theresultsfromthesetests,providedinChapter4.2,indicatethatthisapproachispromising,sinceaffectiverecruitmentrequired32%fewertransmissionsoverallcomparedtothegreedystrategy,andsucceededwithrandomnetworkfailureratesashighas25%.ChapterFivediscussestheexperimentsandresultsintermsoftheirlimitationsandinthecontextoftheliterature,andalsoprovidesamorethoroughanalysisofthecontributionsofthisthesis.ChapterSixsummarizesthethesisandprovidesdirectionsforfuturework.13

PAGE 23

ChapterTwoRelatedWorkThefocusofthisresearchisfault-tolerantrecruitmentforateamofmobilerobots,withasmallerimpactonlimitedcommunicationsresourcesthanothermethods.Beforepresentingtheapproachforthisthesis,itisnecessarytoexaminewhathasalreadybeendoneinrelatedareas.Therearethreeareasofresearchthatdirectlyinuencethiswork:multi-robottaskallocationMRTA,theoriesofemotions,anddistributedsensing.Asthenamesuggests,multi-robottaskallocationreferstotheproblemofassigningrobotstotasksthatcontributetoalargerteamobjective.RecruitmentisaformofMRTA,wherearobotneedingassistancecreatesataskforinstance,investigatingapossiblemineatagivenlocationthatisthenallocatedtoasuitablerobot.AsurveyoftheMRTAliteratureisprovidedinSection2.1.Althoughtherecruitmentproblemaddressedinthisthesisdealswithmulti-robotteams,itismotivatedbysensing.Robotswillberecruitedforthesensorstheycarryinordertoperformasurrogatesensingtask,suchasprovidingperceptsforsensorfaultdiagnosis,orcloserinspectionofsuspectedmines.Distributedsensingdealswithasimilarproblemofcontrollingwheresensorsshouldbelocatedtoaccomplishagoal,andhowtodealwiththeunreliablecommunicationsandthelossofteammembersthatarecharacteristicofdistributedsystems.ThedistributedsensingworkthatrelatestothisapproachisprovidedinSection2.2.Theapproachtakeninthisthesisusesanemotionalmodelwithineachrobottoregulatehowtherobotwillrespondtoarequestforassistance.Emotionsareausefuladditiontorobots,astheyprovideacontextinwhichtherobotsmakedecisionsandtheyallowtherobottomonitortrendsinitsactivityi.e.thattherobothasbeenonthesametaskforalongperiodoftimewithoutprogresswithouthavingtorecorditshistory.Inthiswork,robotswillrefusetorespondtorecruitmentrequestsunlesstheiremotionalmotivationissufcientlyhigh,whichleadstoareductionincommunicationsoverheadwithoutsignicantlyalteringtheoutcomeoftherecruitment.Theuseofemotionscanalsoimprovehuman-robotinteraction,suchthatthehumancanexaminetherobot'semotionalstateandquicklyassessitsoverallstatus.EmotionsinrobotswillbediscussedinSection2.3.14

PAGE 24

2.1Multi-RobotTaskAllocationMRTAdealswiththeassignmentoftaskstomobilerobotsinacooperativeteam.Theliteraturedescribesfourbasicstrategiesforsolvingthisallocationproblem:Motivation-based,usinganinternalmotivationmechanismtocausebehaviorchanges.Parker'sALLIANCE[75]isoneexample.Auctions,whererobotsexplicitlynegotiatefortasksthroughabiddingprocess,asinGerkeyandMataric'sMURDOCH[30].SpreadingActivation,inwhichrobotsdirectlyinhibitthosearoundthemfrombeingchosenforatask,asinMataricandSukhatme'sBroadcastofLocalEligibility[56].TeamConsensus,inwhichentireteamsofrobotsagreeonateamstrategyorformation.ThishasbeenusedbyChaimowiczetaltocoordinateteamsforRoboCup[16].TheapproachtakeninthisthesisisanextensionoftheALLIANCEarchitecture[75],whichwillbedescribedindetailinSection2.1.1.ThethesistakesasimilarapproachtoMURDOCH[30]whichisintroducedinSection2.1.2.AdiscussionofothermethodscanbefoundinSection2.1.3.AsummaryofkeyarchitecturesisprovidedinTable1.2.1.1Motivation-based:ALLIANCEALLIANCE[75]isadistributedrobotarchitectureinwhichrobotschoosetasksbywayoftwomotivationalmechanisms:impatienceandacquiescence.Withacquiescence,arobotperformingataskwilldetectwhenitisnotmakingprogress,andmayeventuallyacquiesce,orabandon,thetask.Conversely,withimpatience,arobotwilldetectthatataskisnotbeingcompletedsatisfactorily,eitherbecausenorobotsareattemptingtofulllit,orbecauseadifferentrobotisonthetaskbutnotmakingprogress.Withsufcientimpatience,arobotwillbeginworkingonatask.Thesemechanismsallowateamofrobotstocompensateforfailures.Ifarobotfails,getsstuck,orisotherwiseunabletocompleteatask,thenanotherrobotwilleventuallytakeover.ThefollowingisaformaldescriptionofALLIANCE,condensedfrom[75].Notethattheformulasprovidedbelowhavebeenreproducedexactlyfrom[75],pp.227toensureaccuracy.InALLIANCE,eachrobotisequippedwithanumberofbehaviorsets,eachofwhicharecapableofcompletingsometask.Supposethatateamismadeupofrobotsfr1;r2;:::g,andeachrobotrihasbehaviorsetsfai1;ai2;:::g.Foreachbehaviorsetaij,thereisamotivationvaluemij.Whenmijexceedsathreshold,thenbehavior15

PAGE 25

Table1.Relatedmulti-robottaskallocationworkaccordingtoresults.NotethatCNPreferstoContractNetProtocol. Name Approach Task Simulated Real PreDomain Robots Robots emption ALLIANCE[75] impatience, Foraging N/A 3 Yes acquiescence MURDOCH[30] GreedyCNP Boxpushing N/A 3 No M+[4] CNP Loadtransfer 3 N/A No CEBOT[10] CNP MapBuilding 3 N/A No ACTRESS[57] CNP Boxpushing 4 N/A Yes LEMMING[70] CNP Foodserving 9 N/A No BLE[101] Port-Arbitrated Multi-target N/A 3 Yes Behavior-based observation Control First-priceauctions[105] Exploration N/A 4 Yes Dynamicrole Cooperative 20 N/A Yes assignment[16] Transport TeamMember PeriodicTeam Soccer 11 N/A Yes AgentArchitecture[94] Synchronization AntSwarms[48] Demining 4 N/A No Coordination Emergency N/A 3 Yes vs.Commitment[74] Handling Implicit Multi-robot 6 N/A Coordination[42] Construction No LMMS[41] Foraging 20 N/A No MOVER[40] Boxpushing N/A 2 No HIVEMind[45] Search N/A 2 No [88] robotcall MapBuilding 2 N/A No queue [87] Tasktrees Construction N/A 3 Yes [47] Motivational BoxPushing 10 N/A Yes 16

PAGE 26

setaijwillbecomeactiveonrobotri.Thetaskthatriisattemptingtocompletewithbehaviorsetaijisreferredtoashaij.Themotivationmijtoexecuteaijiscomputedasshownbelow:mij=0mijt=[mijt)]TJ/F8 9.963 Tf 9.51 0 Td[(1+impatienceijt]sensory feedbackijtactivity suppressionijtimpatience resetijtacquiescenceijt:Thetermsintheexpressionabovearedenedbelow,beginningwiththederivationofimpatience.Notethat slowijand fastijareratesatwhichimpatienceaccrues.Ifnorobotisworkingontaskhaij,orifarobothasbeenworkingonhaijforlongerthanijtimeunits,thenimpatienceincreasesatrate fastij.Otherwise,ifarobothasannouncedthatitisworkingonhaijwithinthepasttimeunits,thenimpatienceaccruesatrate slowij.impatienceijt=8>>>><>>>>:mink slowijk;t;ifcomm receivedi;k;j;t)]TJ/F11 9.963 Tf 13.256 0 Td[(i;t=1andcomm receivedi;k;j;0;t)]TJ/F11 9.963 Tf 9.963 0 Td[(ijk;t=0 fastijt;otherwise.comm receivedi;k;j;t1;t2=8>>>><>>>>:1;ifrobotrihasreceivedmessagefromrobotrkconcerningtaskhiaijinthetimespant1;t2,wheret1>>><>>>>:1;ifthesensoryfeedbackinrobotriattimetindicatesthatbe-haviorsetaijisapplicable0;otherwise.activity suppressionprovidesmutualexclusionsothatonlyonebehaviorsetofmanythatexistontherobotwillbeactiveatatime:activity suppressionijt=8>>>><>>>>:1;ifanotherbehaviorsetaijisactive,k6=j,onrobotriattimet0;otherwise.17

PAGE 27

impatience resetcausesthemotivationtodotaskhaijtoresettozeroifsomeotherrobotrkannouncesthatithasbegunperforminghaij:impatience resetijt=8>>>>>>><>>>>>>>:0;if9ksuchthatcomm receivedi;k;j;t)]TJ/F11 9.963 Tf 10.811 0 Td[(t;t=1andcomm receivedi;k;j;0;t)]TJ/F11 9.963 Tf 10.294 0 Td[(t=0,wheret=timesincelastcommunicationcheck1;otherwise.Finally,acquiescencecausesarobotritoacquiesceafteracertainamountoftime,eitherijtifthenootherrobothastakenovertaskhaij,orijtifanotherrobothas,whereijt>>>>>>>>><>>>>>>>>>>:0;ifbehaviorsetaijofrobotrihasbeenactiveformorethanijttimeunitsattimetand9ksuchthatcomm receivedi;x;j;t)]TJ/F11 9.963 Tf 9.036 0 Td[(i;t=1orbehaviorsetaijofrobotrihasbeenactiveformorethanijttimeunitsattimet1;otherwise.ALLIANCEhasbeenshowntoproducecorrectresultsinateamofmobilerobots.In[75],experimentsusingthreerobotsinaforagingtestdomaindemonstratedthattherobotswoulddivideuptasksandbeginexecutingthemwithoutcentralizedcontrol.Intheforagingdomain,onerobotwouldperformamonitoringtask,reportingtheprogressoftherestoftheteamatregularintervals.TheremainingrobotswouldperformtheMove-SpillleftandMove-Spillrighttasks,whichinvolvedmovingtotheleft-mostandright-mostconcentrationsofspillobjectsandmovingthemtoatargetlocation.Ifarobotwasremovedfromtheteam,theotherrobotswouldeventuallybecomeimpatientandtakeonthetaskthemselves.In[77],sevenadditionaltestdomainsaresummarizedboxpushing,janitorial,boundingoverwatch,formationkeeping,manipulation,tracking,productiondozing.TheseareshowninTable2.ALLIANCEhasalsobeenextendedtoL-ALLIANCE,orlearningALLIANCE,inwhichtheparameters slowij, fastijandijwereadaptedbasedonpreviousruns.ALLIANCEpresentsasolutiontotheproblemofallocatingtasksamongrobotsusinganinternalmotivationwithineachrobot.However,ithasthefollowingcharacteristicsthatmaybeundesirable.ALLIANCEispseudo-emotional;itusesemotion-likemotivations,butdoesnotderivethesefromaformaltheoryofemotions.18

PAGE 28

Table2.TestdomainsforALLIANCE.Notethatthistablehasbeenadaptedfrom[77],Table1.Inthesecondcolumn,Sreferstosimulatedrobots,andPreferstophysicalrobots. ApplicationDomain #Robots Metricdescription Mockhazardouswastecleanup 2P Timeoftaskcompletion,totalenergyused Boxpushing 1P Perpendiculardistancepushedperunittime Janitorialservice 3S Timeoftaskcompletion,totalenergyused Boundingoverwatch 4S Distancemovedperunittime Formation-keeping 4P&S Cumulativeformationerror Simplemulti-robotmanip-ulation 2P Numberofobjectsmovedperunittime Cooperativetracking 2P,2S Averagenumberoftargetsobservedcollectively Multi-vehicleproductiondozing 2S Quantityofearthmovedperunittime ALLIANCEassumesthatrobotsaretosomedegreeinterchangeable.Thoughtherobotsmaynotbehomogeneous,theymustatleasthavecapabilitiesincommon[76].AlloftherobotsunderALLIANCEmustbecooperatingononesetoftasks.InALLIANCE,robotsmayabandonacquiesceataskandleaveitforanotherrobottonish.Everyrobotmustbroadcastitsstatusincludingthetaskitisperformingataregularintervalsothatothermembersoftheteamcanincreasetheirimpatienceaccordingly.TheaffectiveapproachinthisworkisanextensiontoALLIANCE,withthefollowingcharacteristics.Thisapproachisbasedonaformaltheoryofemotions,theOCCmodel[72][73][71].Robotscanbecompletelyheterogeneousinhardwareandsoftware.Robotscanbeengagedinentirelydifferenttasksorobjectives.Robotsmaynotbepreemptedfromtheirtasks.Sharedcommunicationchannelsareinfrequentlyused.19

PAGE 29

2.1.1.1OtherMotivation-basedAllocationResearchAnotherapproachthatusedinternalmotivationforactionselectionisdescribedin[47].Inthatwork,ateamoftensimulatedantsperformedabox-pushingtask.Eachanthadaninternaltimerthatwouldincrementiftheboxwasnotmoving,andwhichwouldresetotherwise.Stagnationwasdetectedbyhavingthetimerexceedathreshold,atwhichpoint,therobotwouldchangeitsbehaviorinthiscase,byattemptingtopushonadifferentpartofthebox.In[47],therewasnocommunicationamongagents,andallagentswerehomogeneousandworkedonthesametask.2.1.2Auctions:MURDOCHThenextclassofsolutionsforthemulti-robottaskallocationproblemusessomeformofanauction.AcommonapproachistheContractNetProtocolCNP[90][21]witharst-priceauction[63].InCNP,anannouncementaboutanewtaskisbroadcasttoateamofrobots.Eachrobotthenreturnsabidthatspecieshowwell-suiteditisforthetask.Awinnerisselectedfromthebids;inthecaseofarst-priceauction,thebidwiththebestutilityorlowestcostischosen.CNPandrst-priceauctionsassumetwocomponentsofinteresttothisthesis.Communication:taskannouncementsmustbetransmittedtotherobotteam,androbotbidsmustbetransmittedinresponse,suchthatallofthebidsexistinoneplaceandthebestcanbechosen.ThecommunicationmechanisminthisthesisisassumedtobebroadcastseevalidationforusingbroadcastinsteadofunicastinChapter4.2,andthecommunicationmediumisassumedtobeunreliable.Fitness:thebidsthatrobotsprovideinresponsetoataskannouncementarearelativemeasureoftherobot'ssuitability,ortness,toperformthetask.Eachtaskmayrequiredifferentrobotresources,andasaresult,themeasureoftnessmayvary.Fitnesscanberepresentedintermsofcoststime,energy,orotherresourcesthatmustbeexpendedorcapabilitiesavailablepercepts,effectors,etc..MURDOCH[30][29][27]isanauction-basedtaskallocationsystemthatissimilartothiswork.TheauctionprotocolinMURDOCHfollowsthissequenceofstepstakenfrom[30],pp.761:Taskannouncement:arobot,taskplanner,human,etc.broadcastsanannouncementtoallrobots.ThemessagingsysteminMURDOCHissubject-baseddescribedbelow,sothatrobotswillonlyhearrequestsforresourcesorservicesthattheycanprovide.20

PAGE 30

Table3.Meanandstandarddeviationoftheelapsedtime,inseconds,forsuccessfulpushingtrialsineachoffourboxpushingexperimentsforMURDOCH.Eachsetwasrepeated10times.Notethatthistableisreproducedfrom[30],p.766. Set Description 1 Nofailurestraightpath 31.22 0.44 2 Pusherfailure 132.75 26.94 3 Partialpusherfailure 260.89 37.79 4 Pusherfailure&recovery 116.44 37.72 Metricevaluation:eachrobotcomputesitsabilitytoperformthenewtask.Thetnessofarobotisbasedonadhocmetrics;thedistancebetweentherobotandthenewtaskwasused,andcomputationalloadwassuggestedasanalternatemetric[27].Bidsubmission:everyrobotthatheardtheannouncementrespondswithitsmetricscore.Closeofauction:onerobotisselectedasawinner,andallrobotsarenotiedofthechoice.Thewinnerisgivenatime-limitedcontracttoperformthetask.Progressmonitoring/contractrenewal:theprogressoftheselectedrobotwillbemonitored,anditscontractwillperiodicallyberenewedifitmakesprogress.Iftherobotdoesnotmakeprogress,thenanotherauctionmaybeheldtoreplaceit.MURDOCHhasbeenimplementedonateamofthreemobilerobotsandvalidatedinabox-pushingtaskthrough40trialsacross4scenarios.Themetricsfortheseexperimentswerewhethertherobotteamsucceededinitstaskforatotalof36successfultrials,andthetimerequiredtocompletethetaskshowninTable3.TheamountofcommunicationsbandwidthrequiredduringoneofthetestsisshowninFigure4.CommunicationinMURDOCHisthroughasubject-basedpublisher/subscribersystem.Allmessageshaveasubjectattribute,androbotscansubscribetodifferentsubjectstoreceivemessagesofthattype.Messagesarethenpublishedbroadcasttotheteamofrobots.Ifamessagehasasubjectthataparticularrobotisnotsubscribedto,thentherobotsimplyignoresthemessage.MURDOCHhasthefollowingcharacteristics:MURDOCHactsasaninstantaneousgreedytaskscheduler.Thishastwoimplications.First,everyrobotsubscribedtothesubjectcontainingthetaskannouncementwillrespondwithabid.Theimpactoncommunications,therefore,increaseslinearlywiththesizeoftheteam[31].Second,itsperformancedependsontheorderinwhichtasksappear[30]andthesolutionitchoosesmayonlyprovide1 2oftheoptimalutility[28].21

PAGE 31

Figure4.GraphofthecommunicationsusebyMURDOCH.Therequiredbandwidthspikeswhenevertaskallocationoccurs.Notethatthiswastakenfrom[30],p.766,usedwithpermission.MURDOCHusesasubject-basedmessagingsystem,whichrequiresallrobotstousethesamenamespace;thatis,thesetofallpossiblesubjectsmustbeagreeduponinadvance.MURDOCHcontributesamethodfordeterminingthebestrobotforaparticulartask,whichistouseametrictodescribetherobot'ssuitabilityforinstance,itsdistancefromanobjectthatneedstobemanipulated,orthecomputationalloadontherobot.Nobettermethodhasbeenfoundintheliteraturefordeterminingthetnessofarobottoaparticulartask.Theaffectiveapproachinthisworkusesasimilarstrategy,buthasthesecharacteristics:Robotsdonotimmediatelyrespondtoeachannouncement,butinstead,graduallyincreaseaninternalmotivationSHAMEuntilitreachesathreshold.Thishastwoimplications.First,noteveryrobotwillrespondtoagivenannouncement,andinmanycases,atmostonewill.Thisreducestheimpactoncommunications,sothatbandwidthusewillincreaseslowlywithteamsize.Second,theorderinwhichtasksappearisnotasimportant,astheallocationisnotinstantaneous.ThisallowstheaffectiveapproachtoreachsolutionsthatthegreedystrategywouldmissseeChapter4.2.4.Parametersintheaffectiveapproachallowitsbehaviortobetuned,soitcanbemoreorlesslikegreedyasrequired.Theapproachinthisworkusesaclass-basedmessagingsystem,wheremessagetypescanbedistributedatrun-time.TheimplementationfortheaffectiveapproachisinJava,andthemessages22

PAGE 32

themselvesareinstancesofJavaclasses.AfeatureofJavaiscodemigration,inwhichclassesbotheldsandmethodscanbetransferredfromonevirtualmachinetoanotherdynamically.Inthisway,iftworobotsrecognizedifferentsetsofmessages,theycanexchangemessagetypesuntiltheymatch.LikeMURDOCH,thisworkusesametrictodeterminethesuitabilityofarobottoarecruitmenttask.Inthisapproach,themetricistheestimatedtimetherobotwouldneedtomoveintopositiontobeginthetask.2.1.2.1OtherAuction-basedApproachesTheM+ProtocolfromBotelhoandAlami[4]containstwointerestingcomponents:M+taskallocationandM+cooperativereaction.TheM+taskallocationmechanismisbasedontheContractNetProtocolCNP[90][21].Whenataskisavailable,arobotthatiscapableofcompletingthetaskwillestimateitscostfordoingsoandannouncearstoffer.Otherrobotsmayannouncebetteroffers,andbecomethebest-candidateinturn,untilarobotnallybeginsthetask.Next,theM+cooperativereactionmechanismisonlyusedintheeventofafailurein[4],butitmorecloselyresemblesthisapproach.ArobotRiwillsendoutarequestforhelpwhenitisunabletocompleteitstask.OtherrobotswillthendeterminewhetheritispossibletoachieveboththeirowngoalandRi'sgoal,andifso,willrespond.Riwillthenchoosethebestoffer.Inthisthesis,requestsforhelpareansweredonlywhentherobot'semotionalstatemotivatesittodoso,andthebestofferisaccepted.AccordingtoGerkey[31],thecommunicationcomplexityoftheM+ProtocolisOmnformtasksandnrobots.Zlotetal[105]describerst-priceauctionsusedbyrobotsinanexplorationtaskdomaintodeterminewhichrobotwillperformatask.Robotsexplicitlynegotiatethroughanoperatorexecutive,orOpExec,tomaximizetheirownprotbybuyingandsellingtasks.Bidsarebroadcastbyeachrobot,resultinginabehaviorandcommunicationcomplexitysimilartothatofMURDOCHandM+.DynamicRoleAssignment,developedbyChaimowiczetal[16],isalsocomparableincommunicationcomplexitytoMURDOCH,M+,andrst-priceauctions.In[16],anothervariantoftheContractNetProtocolisused,suchthatasingleleaderrobotbroadcastsrequestsforassistanceuntilasufcientnumberofrobotshavevolunteeredforacooperativetask.Robotscanbepreemptedfromtheirtasks,butwillonlychoosetodosoiftheutilityofthenewtaskishighenoughtojustifytheoverheadcostofmakingthetransition.TheContractNetProtocolalsoappearsinotherarchitectures,suchasCEBOT'sTaskAcquiringLayer[10].CEBOTdistributesrobotstateinformationamongmembersoftheteam,whereitisstoredinaWorldModel.InDistributedAutonomousRoboticSystemDARS[9][11],CEBOTisextendedtomodelreliabilityofsensinginformationfromrobotsinadistributedsystem.In[83],manipulatorsandmobile23

PAGE 33

robotscooperate,usingacontractnet,toprovidethebestsensingofatargetforaninspectiontask.In[69],taskdecompositionisperformedbyaplanner,andsubtasksareassignedusingacontractnet.TheACTRESSarchitecture[57]alsousesavariantofthecontractnetforateamofheterogeneousrobots.InACTRESS,tasksareprioritized,andpreemptionoflow-prioritytasksmayoccurifahigh-prioritytaskisfailing.TheLEMMING[70]systemusesCNP,butreducestheamountofcommunicationsamongrobotsbyhavingrobotsrememberwhorespondedtoaparticularrequestandassigningtaskstothatrobotdirectlyinthefuturethuseliminatingtheannouncementandbidstagesofCNP.However,LEMMINGdoesnotuseacknowledgmentsformessages,andmessagelosscouldcauselongdelaysontasks.Further,intheeventthatarobotsufferedafailure,itwouldstillbeassignedtasksthatitcouldnotcomplete.2.1.2.2UtilityMetricsAuction-basedmethodsrequireagentstosubmitbidsforataskthatdescribetheagent'srelativetnesstothetask.Thereisnogenerallyacceptedmeansofdeterminingtheutilityofarobottoatask,butatleastveapproacheshaveprovidedtheirownmetrics.Asabove,MURDOCH[30][27]usestheCartesiandistancebetweeneachrobotandthetasktodeterminetheirrelativetness.In[51],thecostofusingeachsensorisconsidered,andtheoverallcostofachievinganobservationisminimized.Thisimpliesthatthecostinpowerconsumedorthetimetakentoreadthesensorisalreadyavailableanddirectlycomparable.In[104],thelevelofintelligenceofanindustrialmanipulatorvariedfromonetovebasedonwhatcapabilitiesitssensorsallowed.[104]alsodiscussedtheutilityofusingasensor,whichwasmeasuredintermsofresponsetimeoruncertainty.In[103],sensorutilityisdenedintermsofthepositionuncertaintythatresultsfromusingdatafromaparticularsensor.Utilitydeferstoahuman-generatedpreferenceorderingorsensoruncertaintyin[26].Thisapproachusestheestimatedtimethatarobotwillrequiretoreachthetasklocationasitsmetric,aswillbediscussedinChapter3.3.2.1.3OtherApproachesTheBroadcastofLocalEligibilityapproach[101]usesPort-ArbitratedBehaviorPABtocontroltheowofinformationbetweenreactivebehaviorsinasubsumptionarchitecture.Robotsdeterminetheirownutilityforaccomplishingtheavailabletaskandinhibitnearbyrobotsfrombeingselectedforthetaskaccordingly.Thiscross-inhibitionresultsintheselectionofasinglerobot.Theapproachin[101]requiresagreateramountofcommunicationamongrobotsthanthisapproachOmnformtasksandnrobots,accordingto[31],comparedtoanupperboundofOnforaffectiverecruitment.In[101],theBLEapproachistestedagainstgreedyandrandommethods,whichisalsotrueofthiswork.24

PAGE 34

Otherapproachestomulti-robottaskallocationincludeemergencyhandling[74]inwhichrobotsrespondtoaudiblealarmsandfollowthesoundgradienttoitssource.In[74],robotsusecommunicationprimarilytopreventmultiplerobotsfromrespondingtothesamealarm.Thisisrelatedtotheproblemofrecruitment,wherethealarmstaketheplaceofarequestingagent.However,withoutasingleagentactingasanarbiter,therobotscoordinatethemselvesthroughmutualinhibition.In[74],ashareddatastructureblackboard[19]wasused,androbotsbroadcastupdatestotheblackboardatarateof10Hz,leadingtoahighimpactoncommunications.[40][81]discussestheMOVERsystem,whichisadistributedcontrolsystemforateamofrobotsperformingacooperativesearchandrescuetask.Whenonerobotndsavictim,alloftheotherrobotscomehelpitthereisnoselectiverecruitment,allrobotsmustassist.Theapproachin[40]requiresrobustcommunications:ifthereisalossofcommunications,therobotswillstalluntilcommunicationsarerestored.[94]usesrobotsoccertomotivatecoordinationofrobotsthroughashared,low-bandwidth,unreliablecommunicationchannel.Communicationrequirementsarereducedthroughlocker-roomagreements,whichareaprioristrategiesandformationsthatcanbespeciedwhenrequired.Robotsareassignedrolesthatspecifyasetofbehaviors,butarobotmayhavesomeautonomyinhowtofulllitsrole.ThroughPeriodicTeamSynchronizationPTS,robotsareallowedtoperiodicallyexchangeinformationwithoutrestrictionsthroughbroadcasts.[1]describestheMARTHAprojectwhichwasdesignedtomanageeetsofrobots.MARTHAusesaplan-mergingprotocol,whichworksasfollows:robotsaregivengoals,forwhichtheyindividuallyformplans,andtheyaregiventheplansofallotherrobotsintheteam.TheseplansarethenmergedintoadirectedacyclicgraphDAGtodetermineanorderingthatwillresolvetemporalconstraints.Experimentalresultswithtensimulatedrobotsandtestswiththreerealrobotsarepresented.Asimilarapproachistakenin[87],wherethreerobotscooperatedonaconstructiontask,usingtasktreestoresolvetemporalconstraints.[87]mentionsthattheallocationofrolestaskstotherobotsintheteamisnecessarybutusedxedrolesforexperimentsanddidnotindicatehowroleassignmentshouldbedone.[82]and[58]detailworkontheScoutrobots,discussinghowtoallocatealimitedcommunicationlinkthatissharedamongagroupofrobots.Inthiscase,therobotsarenotautonomousnotenoughonboardprocessingtoallowautonomy,sotheyneedenoughbandwidthtosendvideoandothersensorinformationandtoreceivemotorcommands.Allocationofbandwidthonthelimitedvideochannelsisdoneinaround-robinfashion,andthecontrolchannelisdividedintotimeslicesandportionedoutaccordingtothe25

PAGE 35

desiredupdaterateanduser-denedpriorities.Norecruitmenttakesplace,onlymanagementofthesharedcommunicationchannels.[88]giveseachrobotarobotcallqueuetowhichtaskscanbeaddedorremovedbyotherrobots.Thetestdomainismapping,anditisassumedthatrobotsofdifferentsizesareused.Whenarobotndsanareathatistooconnedforittoexplore,thelocationisaddedtothecallqueueofanotherrobotthatwouldbesmallenoughtot.Eachrobotcandecidewhethertocontinueexploringnearbyopenareasortobeginexploringlocationsonitscallqueue.Simulationresultsfortworobotsonelarge,onesmallwereprovided.Thetaskassignmentin[88]isnotuptonegotiation,androbotsareselectedfortheirsize.Alltransmissionsarebroadcast,anditisassumedthattherobotsintheteamareunchanging.Antbehaviorhasbeenresearchedwithregardstorecruitmentinswarms.KumarandSahinappliedrecruitmenttodeminingin[48],whichisthetaskdomainpresentedinthiswork.Kriegeretaldiscussrecruitmentamongantsin[46].Whenanantseesmorefoodthanitcancarry,itwillrecruitmoreantstofollowitbacktothesamelocation.However,inbothworks,thereisnodecisionprocessforrecruitmentandallrobotsareassumedtobehomogeneousandonthesametask.[13],[78]and[2]providesurveysoftheeldofdistributedrobotteams.JonesandMatarichaveexploredmulti-robotcoordinationwhererobotsuseonlytheirinternalstatewithnocommunication[41][42].[80]discussesarobotdesignusingtransputerstocontroldifferentaspectsofarobot,buttheimplementationofsensingstrategieswaslimited.[93]dealswiththedistributedcommunicationandcontrolissuesofspacecraft.[45]presentstheHIVEMindarchitecture,inwhichrobotshaveaccesstosensorreadingsfromtheotherrobotsintheteam.OtherapproachescomefromtheDistributedArticialIntelligenceDAIcommunity.[85]providesagoodsurveyofDAIasitrelatestothisarea,anddiscussescoalitionforminginauctions.[91]usesanargumentativenegotiationmodelandcase-basedreasoningandcoalitionsinanobjecttrackingdomain.Theuseofinformationinvariantstoautomaticallybuildteamswithparticularcapabilitieshasbeenexploredin[23][22][35][81].2.2DistributedSensingTheaffectiverecruitmentstrategyinthisthesisattemptstosolvetheproblemofbringingarobotandmoreimportantly,thatrobot'ssensorstoaparticularlocation.Distributedsensingsystemsdealwithsensornetworksstatic,distributedarraysofcommunicatingsensors,sensorcoverageandtrackingensuringthata26

PAGE 36

Table4.DistributedSensingLiterature. Approach TaskDomain Communications [95] ObjectTracking Broadcast [39] SensorCoverage Broadcast [34] PlumeSourceSearch Broadcast ByzantineGeneralsProblem[7] ObjectTracking Broadcast ASCENT[15] NetworkConnectivity Broadcast [18] ObjectTracking CentralizedControl [20] SensorDispersion [32] ObjectTracking 4-Neighbors DirectedDiffusion[36] LocalBroadcast deBruijnGraph[68][37][38] [44] MilitaryRecon [54] ObjectTracking [97] SensorCoverage Broadcast targetorareaiscompletelyobserved,andremoteperception.Theseareasoverlaptheproblemofrecruitmentforadistributedteamofrobotsasfollows:Sensornetworksrequirecommunicationamongdistributedsensornodes.Theindividualnodesmayhaveinsufcientpowertotransmitdatafrequentlyoroverlargedistances[43],andnodesmayfailatanytime.Thestrategiesforovercomingtheseconstraintsapplytobothdistributedrobotteamsandsensornetworks.Sensorcoverageandtrackingrequirethatmultiplerobotscooperatetoobserveaparticulartargetorarea,whichrequiresthattherobotsbeabletoshareacommoncoordinateframeandlocalizerelativetoeachother.Thisisrequiredforrecruitment,becauseonerobotmustbeabletondanotherrobotinordertoassist.Oneinsightthatdistributedsensingprovidesisthatthetrackedtargetitselfcanprovidethebasisforacommoncoordinateframe[39].Remotesensingappliestooneofthemotivatingdomainsforrecruitment:distributedfaultdiagnosis.Arobotthatdetectsafaultinitssensingmightdiagnosethefailureorrecalibratebasedontheexternalviewpointprovidedbyanotherrobot[53].Manyofthesesystemssimplybroadcastallstatusandsensorupdates,whichleadstopoorscalingperformance.AnoverviewoftheliteratureisshowninTable4.[95]describesthecombinationofsensorobservationsamongateamofrobotsusingBayes'ruleandKalmanlters,andassumingacommoncoordinateframeandperfectlocalizationfortherobots.Their27

PAGE 37

approachreliesonhavingeachrobotbroadcastitslocationandsensorreadingstotheotherrobots.Similarly,in[39]and[92],robotscontinuouslybroadcasttheirpositioninacommoncoordinateframetoproducecompletesensorcoverageofatarget.In[92],theteamconsistedofthreerobotsthatbroadcasttheirrelativebearingtothetargetat15Hz.[34]usesasimplebroadcastamongrobotssearchingforthesourceofanodorplume,andpresentsresultsonthreetypesofbroadcastswhentheplumeisdetected:NONEnocommunication,ATTRACT,andKILL.ResultsshowthatATTRACTaformofrecruitmentusesamuchlargeramountofteamenergytocompletethetaskthanKILL,whichsimplyshutsdownallrobotsotherthantheonethatdetectedtheplume.[34]indicatesthatasolutionthatrequiresslightlylongerrunningtimesmayprovidedramaticimprovementsinotherareasinthiscase,teamenergyexpenditure.Similarly,theaffectiverecruitmentapproachwilltakeslightlylongertocompletethangreedytechniques,suchasMURDOCH,butwithanimprovementincommunicationload.[7]usesananalysisoftheByzantineGeneralsProblemBGPtointroduceasensorfusionapproachthatreachesanagreementacrossadistributednetworkofsensors.BGPdealswithreachingaconsensusinadistributedsystemwheresomenodesareunreliableormalicious.Thealgorithmisshownintheoreticaltermsandcomparedtoothersimilaralgorithms.Inaffectiverecruitment,itisassumedthatallrobotsarecooperative.In[15],analgorithmispresentedtosupportroutingthroughanadhocnetwork,suchthatdistantnodescanbeconnectedwithoutanycentralizedcontrol.Thiswouldbeausefuladditiontotheaffectiverecruitmentapproach,asitallowsrobotstostayincontactwiththosethatwouldotherwisebeoutofcommunicationrange.Adistributedsensingsystemisdescribedin[18],whereasetofcamerasmountedinanenvironmentisusedforlocalizationofamobilerobot.Oncethecamerasarecalibrated,theycannotbemoved,andallprocessingisdonecentrally.Sincethisworkfocusesontheuseofsensorsattachedtofullydecentralizedmobilerobots,[18]doesnotdirectlyapply.[20]presentscontrollawstocoordinateagroupofrobotsusingaVoronoigraphtospatiallydistributetherobotsevenly.Simulationresultsareprovidedfor16robotsinapolygonalenvironmentassumingholonomicrobotswithisotropic360sensors,butnoresultsareprovidedforrealrobots.Themethodofcommunicationamongtherobotteamwasnotdiscussed.[32]discussesasensingnetworkforacquisitionandtrackingoftargetsprimarilyformilitarypurposes,andstudiesthedelaysinvolvedinprocessingandpropagatinginformationthroughthenetwork.28

PAGE 38

Simulationresultsareprovidedforthedelaysintransmissionfromonenodetoanother.[32]assumedthatthesensorsareplacedinaregularpatterni.e.aregulargridwithfourneighborspernode,whichmakesitlargelyincompatiblewithmobilerobots,whosearrangementisdynamic.[36]introducestheideaofdirecteddiffusion,whichisamethodfordistributinginformationinamesh-likesensornetwork.Thebasicprincipleisthatsensingtasksinterestsaresentintothenetworkandpropagatetotherequiredsensors.Sensorreadingsthenpropagateback,retracingthepathmadebytheinitialrequest.Theintermediatenodesalongthatpathmayberequiredtorelayinformationfrommultiplesensorsdownstreamfromthemselvestheauthorsrefertotheowofinformationasagradient,andthisaffordsthemtheopportunitytoperformaggregationorfusionofthedatainparallel.[36]presentssomeresultsfromsimulationandrealnetworktests,withupto250nodesand20%nodefailurerates,usingenergydissipationandnetworkdelaysasmetrics.[68]and[37]illustratethatasensornetworkcanbeorganizedasamulti-leveldeBruijngraphofbinarytrees.Theresultingnetworkstructureisarguedtobetolerantoffailureofanyofthenodeswhilemaintainingasmalldiameter.[38]expandsonthis,providingaformaldescriptionoftheproblemofintegratinginformationacrossanumberofsensorswherethereadingsfromaparticularsensormayhaveboundedfromtametowildinaccuracy.Simulationresultsfor60sensors,ofwhich23werefaulty,areprovided,buttheworkisstilllargelytheoretical.[44]discussestheissueofplacingsensorsinahostileareasuchthatcoverageisachievedwhilemaximizingtheeffortrequiredforanenemytodamageordestroythesensors.[44]islargelytheoretical,anddoesnotseemtoapplyoutsideofthegame-playingormilitaryscope.[54]providesanalgorithmforfusingrangereadingsofmultipletargetsusingrangereadings.ItemploysDempster'sruleandHiddenMarkovModels,buthasnorealbearingonothersensormodalitiesormobilerobots.[97]focusesprimarilyonusingmultiplerobotstosimultaneouslyobserveanobjectfromdifferentdirectionsinordertoreconstructtheobjectfromsensorreadingsandidentifyit.Therobotsinthatworksharedalloftheirsensoryinformation,sotherewasnorecruitmentrequired.Similarly,[17]isaone-pagebriefondeterminingthenumberofsensorsrequiredforuniquelyidentifyingtargetsinann-dimensionalspace,usinggraphandcodingtheoryforanalysis.In[102],theauthorsdescribearobotsystemthatusesmultiplesensingmodalitiestoexploreanareaandpresentahumanoperatorwithasummaryoftheinterestingplacesintheenvironment.[102]dealswithremotesensing,butonlyinthesensethattherobotisremoteandhassensors.[96]providesaformalframeworkforsensorallocation,butdoesnotgomuchintothepracticaldetails.Thispaperislargely29

PAGE 39

Table5.Summaryofliteratureapplyingemotionstorobots.NotethatHRIreferstoHuman-RobotInterac-tion. Approach Domain SimulatedRobots RealRobots [64] Resupply N/A 2 [3] Entertainment N/A 1 [59][60][61] Exploration,HRI 1 Simon[98] HRI 1 N/A Yuppy[99] Entertainment,HRI 1 1 KISMET[6][5] HRI N/A 1 Cherry[52] HRI 1 1 PETEEI[25] Entertainment 1 N/A theoretical,andprovidesonlynumericalexamplesratherthanexperimentalresults.[100]isoneofthemorerecentpapersonsimultaneouslocalizationandmappingSLAMwithdetectionandtrackingofmovingobjectsDTMO.2.3EmotionsandAffectiveComputingThisapproachappliesacognitivelyplausibleemotionalmodelfromOrtony,Clore,andCollins[72]totheproblemofmulti-robottaskallocation.AsurveyofthepreviousresearchinusingemotionstocontrolrobotsisprovidedinSection2.3.1.ThebasisfortheemotionalmodelusedinthisthesisisprovidedinSection2.3.2.2.3.1EmotionsinRobotsAnoverviewofliteratureinwhichemotionshavebeenappliedtoroboticsisprovidedinTable5.Roboticsresearchershaveappliedemotionsintwobasicapproaches.Therstapproach,whichistakenbythiswork,istouseemotionstomodulatethebehaviorofarobot,especiallywithregardstocooperationinateam.In[64],anemotionalmodelwasusedtopreventdeadlockinarobotteaminaresupplytask:onerobotservedrefreshmentsandwouldrequestassistancefromtheotherwhensuppliesranlow.Ratherthanwaitingforhelpthatmightneverarrive,theservingrobotwouldbecomeincreasinglyfrustrated,andeventuallyeitherattempttointerceptitsassistantorresupplyitselfdirectly.Meanwhile,theassistant'semotionalreactionscausedittochangetheparametersofitsbehaviorifitwasstymied,i.e.movingfasterandmoreaggressively.Otherapproachestendtouseemotionstogeneratebehavior,usuallybasedonconictinginternaldrives,foractionselectionwithinasingleagent.In[3],aSonyAIBOrobotdogwasprogrammedwithbiologicallyinspireddriveshunger/thirst,investigative/curiosity,play/boredom.In[59][60][61],arobot30

PAGE 40

wasprogrammedwithmotiveshungry,distress,bored,explore,etc.thatinuencedtherobot'spatternofbehaviorwhileitattemptedacomplextaskndingsitswayaroundaconferenceandgivingatalk.[98]and[99]discusstheCathexisarchitecture,rstthroughasimulatedtoddlernamedSimon[98]andsecondwithYuppy,anemotionalpetrobot[99].Yuppyhadfourdrivesrelatingtobatterycharge,temperature,energy,andinterestlevels,andSimonhadtheseplusthirst.Theseworkstendtogroundtheagent'semotionsinitsownphysicalneeds;forinstance,in[60]and[99],ahungermotivewaslinkedtotherobot'sneedtorechargeitself.Thisisexploredin[79],whichsuggeststhatemotionscanonlyexistwheretheyarerequiredforsurvivalinwildenvironments.[79]justiesthiswithanexaminationoftheFungusEateridea,inwhichanagentmustspendtimeandresourcesacquiringenergy,thoughthisisincidentaltocompletingitsgoalandleadstonodirectreward.Withtheexceptionof[60],however,theseworksdonottendtomovebeyondexistenceproofs.Emotionsarealsoemployedinhuman-robotinteractionstudies.Robotsthatdisplayemotionalresponsesmaybeconsideredmoreintelligentorlife-likebyhumans.KISMET[6][5]usesanexpressiverobotheadtoengageandinteractwithpeoplebasedonunderlyingemotionalmotives.Cherry[52]wasdesignedtobesociallyintelligenttotintoanexistingsocialstructureauniversityofceenvironment.Emotionsregulatedtheperformanceof,andreactionto,ofcetasks.Cherryisablerecognizefacesasitinteractswithpeopleandattemptstoactaccordinglyforinstance,addressingfullprofessorsmorerespectfullythangraduatestudents.PETEEI[25]modieditsemotionalprocessbasedonitsexperiences,andlearnedtoassociateeventswithemotionalstates,butwasonlytestedinsimulation.[12]and[84]providefurthersurveysofworkinemotionasitappliestorobotics.2.3.2OCCModelofEmotionsThisworkbuildsonaformalmodelofemotionsdevelopedbyOrtony,Clore,andCollins,andisreferredtoastheOCCmodel[73][71][72].AnoverviewofthemodelisshowninFigure5.TheOCCmodelconsidersemotionsasreactionstoeventsinanagent'senvironment,orreactionstoagents.Thereactionscanbepositivesuchasjoyandadmirationornegativesuchasdistressandreproach.Emotionsalsohaveavalence,orintensity,whichindicateshowstronglyaparticularemotionisfelt.IntheOCCmodel,emotionsaredividedupintofourcategories:goal-based,standards-basedalsoreferredtoasattributionemotionsinearlierwork,attitude-based,andcompound.Thesecategorieswillbedescribedbelow.Goal-basedemotionspertaintotheaccomplishmentofagoalortheanticipationofaneventthatmaypreventachievingthegoal.Anemotionalreactionmaybeinducedbyanevent,suchasthecompletion31

PAGE 41

Figure5.TheOCCmodel.Notethatthiswasreproducedfrom[72]p.195.32

PAGE 42

Table6.Standards-basedemotionsalsocalledAttributionemotions.Thesearebasedonwhichagentisbeingreactedtoandwhetherthereactionispositiveornegative.Notethatthiswasreproducedfrom[73]p.136. Identity AppraisalofAgent'sActions ofAgent Praiseworthy Blameworthy Self approvingofone'sownpraisewor-thyactione.g.,pride disapprovingofone'sownblame-worthyactione.g.,shame Other approvingofsomeoneelse'spraise-worthyactione.g.,admiration disapprovingofsomeoneelse'sblameworthyactione.g.,reproach Table7.Standards-basedemotionsinwhichanagenthasanegativereactiontoitsownactions.Thestrengthofthecognitiveunitwiththeactualagentreferstohowcloselyrelatedthereactingagentistotheagentbeingappraised;iftheyarenotatallrelatedi.e.completestrangers,thereactionwillbeweakerthaniftheyarecloselyrelatedi.e.bestfriends.Notethatthiswasreproducedfrom[73]p.137. SELF-REPROACHEMOTIONS TYPESPECIFICATION:disapprovingofone'sownblameworthyac-tion TOKENS:embarrassment,feelingguilty,mortied,self-blame,self-condemnation,self-reproach,shame,psychologicallyuncomfortable,uneasy,etc. VARIABLESAFFECTINGINTENSITY: thedegreeofjudgedblameworthiness thestrengthofthecognitiveunitwiththeactualagent deviationsoftheagent'sactionfromperson/role-basedexpectationsi.e.,unexpectedness EXAMPLE:Thespywasashamedofhavingbetrayedhiscountry. orfrustrationofagoalleadingtojoyordistress,respectively.Anagentcanalsoreacttotheprospectofattainingagoalhopeortheprospectoffailingtoachieveagoalfear.Perceivedthreatstotheagent'sabilitytoachieveagoalcaninducereactionsofreliefifthegoalsurvivesthethreat,anddisappointmentotherwise.Standards-basedemotionsarereactionstotheactionsofagents.Therearefourvariationsontheseemotions,dependingonwhetherthereactionispositiveornegative,andwhethertheagentexperiencingtheemotionisthesameastheagentbeingreactedto.ThisisshowngraphicallyinTable6.Standards-basedemotionsareusefulinmulti-agentteamsbecausetheyenableanagenttoweighitsownactions,andtheactionsofotheragents,inasocialcontext.TheSHAMEemotionusedinthisworkisastandards-basedemotionthatrepresentsthedegreetowhicheachrobotisnothelpingtheteamtomeetitsobjectives.SHAMEisanagent'snegativereactiontoitsownaction;inthiswork,ignoringrequestsforhelp.ThistypeofemotionisdescribedinTable7.33

PAGE 43

Compoundemotionsareacombinationofgoal-basedandstandards-basedemotions.Acompoundemotionmaybeinducedbytheco-occurrenceofagoal-relatedeventsuchascompletionofthegoalandappraisaloftheagentthatcausedtheeventsuchasgratitudetoanotheragent,orgraticationiftheagentcompletedthegoalitself.Finally,attitude-basedemotionscapturemorelong-termreactionsofoneagenttoanother.Forexample,ifanagenttendstoinducepositivereactions,otheragentsmaydevelopapositiveattitudetowardit.2.4FoundationofApproachThisthesisbuildsonpreviousresearchthatwasintroducedabove.Inparticular,thefollowingideasfromtheliteratureareincorporatedintothisapproach:ContractNetProtocol[90][21].Thecontractnetprotocolspeciesasequenceofmessagesbetweenmembersofadistributedagentsystemtoassigntasks.Therstmessageisataskannouncementthatspecieswhatcapabilitiesarerequiredtoperformthetask,adescriptionofthetask,howtasktnesswillbemeasured,andanexpirationtimeforbidding.Interestedagentscansubmitabidforthetask,andoneormoreagentsareawardedthetaskasaresult.Numerousotherapproacheshaveusedvariantsofthecontractnetprotocol[30][4][10][57][70][83][69].Computationalmodelofemotions.ThisapproachusesoneaspectoftheOCCmodel[72][73][71],thestandards-basedemotionSHAME,toprovidemotivationformulti-robottaskallocation.Broadcastcommunications.Thedominantmodeofinter-robotcommunicationintheliteratureisbroadcastmessaging,appearingin[75][30][95][39][34][7][15][36][97][56][74][92][88].Inthisapproach,robotswillcommunicateusingbroadcastcommunicationsexclusively.ThisdecisionisjustiedinChapter4.2.3.2.5SummaryThischapterhasprovidedasurveyoftheresearchinmulti-robottaskallocation,emotionsinrobots,anddistributedsensing.Asummaryoftheliteratureisprovidedbelow.Thereareavarietyofapproachestothemulti-robottaskallocationproblem,ofwhichtwoareparticularlyinterestingforthisthesis:MURDOCHandALLIANCE.TheMURDOCHsystem[30]relatestothisthesisasfollows:34

PAGE 44

ThisapproachandMURDOCHbothusetheContractNetProtocol[90][21]witharst-priceauction[63]toperformtaskallocationforadistributedteamofrobots.Thisappearstobethebeststrategyfordistributedtaskallocation,asitistolerantofcommunicationlossesanddoesnotrequirethattherobotsworktowardacommongoal.BothMURDOCHandthisapproachassumethatrobotscannotbepreemptedfromtheirtasks.[30]recognizesthatrobotswillhavedifferentcapabilitiesandcontributestheideaofusingmetricstodiscriminateamongrobotsforatask.Thisthesisborrowsthemetricideatodeterminewhichrobotwillbeassignedatask.InMURDOCH[30],themetricisthedistancebetweentherobotsandwherethetaskwilltakeplace.Inthisthesis,themetricistheestimatedtimerequiredtocoverthatdistance.ThefundamentaldifferencebetweenMURDOCHandthisapproachiswhathappensbetweenthearrivalofataskannouncementmessageandabidresponsebytherobot.InMURDOCH,eachrobotwillimmediatelyrespondwithabid.Thus,thecommunicationloadforMURDOCHisOn,increasinglinearlywitheachnewrobotintheteam.Inthisthesis,eachrobotwillrstevaluateitsemotionalstate,andonlyreturnabidifsufcientlymotivated.TheworstcasecommunicationloadforthisapproachisalsoOn,butsimulationresultsinChapter4.2indicateastatisticallysignicantreductionincommunicationsoverheadby32%onaverage.Thisdifferenceinstrategyhasthreeimplications.First,thereducedloadmakesthisapproachmoreappropriateforapplicationsinvolvingverylargeteams,low-powerdevices,orstealthrequirements.Second,MURDOCHisaninstantaneousgreedyscheduler,anditsperformancedependsontheorderinwhichtasksarrive.Itiseasytoconstructscenariosinwhichthegreedyapproachcausesapoordecision,resultinginaslittleas1 2oftheoptimalutility[28].Whilenoclaimismadehereabouttheoptimalityofthisapproach,itcanndbettersolutionsasshowninChapter4.2.Third,byusinganinternalmotivationwithineachrobot,thisapproachtendstodistributerecruitmentsevenlyamongrobotsthatareotherwiseequallysuitedtothetask.MURDOCHwillalwaysrecruittheclosestrobot,whichmaycauseexcessiveuseofasubsetoftherobotteam.MURDOCHusesasubject-basedmessagingsystemrequiringallrobotstoshareacommonnamespace.Theimplementationinthisworkusesaclass-basedmessagingsystemthatallowsnewmessagestobeaddedatanytime.35

PAGE 45

ThisworkextendsALLIANCE[75],whichisanarchitecturethatusesinternalmotivationswithineachrobottocontroltaskselection.Thisworkborrowstheideaofusingamotivationalsystemtoregulatetaskallocation.However,therearevedifferencesbetweenthisworkandALLIANCE.InALLIANCE,individualrobotsareabletochoosenewtasksforthemselves,whereinthiswork,robotsbidfortasksandthebestbidischosen.InALLIANCE,robotsmayabandontheircurrenttaskwhentheydetectalackofprogress.Inthiswork,robotscannotbepreemptedfromtasks.ALLIANCEusesregularbroadcaststoalloweachrobottomonitortheprogressoftherestoftheteam,whereasthisapproachonlyusesthecommunicationschannelwhenanewtaskisbeingallocated.ALLIANCErequiresthatrobotsoperateoverthesamesetoftasks,andthisapproachmakesnosuchassumption.Thisworkusesamodelofemotionsasabasisforrobotmotivation,butALLIANCEdoesnot.Therearetwoissuesinthisresearchthatfallintotheareaofdistributedsensingandsensornetworks:determiningthebestsensorforaparticularsensingtask,andfaulttoleranceintermsofbothunreliablecommunicationsandrobotfailures.However,theproblemofmatchingsensorstotasksisstillanopenissue,andnosolutionswerefoundintheliterature.Asaresult,anadhocbutextensiblemetric,theestimatedtimethatarobotwillneedtoreachthetasklocation,isusedtoestimatearobot'stnesstoatask.MetrictnessfunctionsarediscussedfurtherinChapter3.3.Regardingfaulttolerance,mostdistributedsensingapproachesfavortheuseoffrequentbroadcastcommunications,sothelossofanyparticularmessagehaslittleimpact.Unfortunately,mostoftheseapproachescannotyetbeappliedtothemobilerobotrecruitmentproblem,eitherbecausetheyassumeaparticularcongurationofstationarynodesorassumethatallnodesarealreadycooperatingonthesametaskandthus,thereisnoneedforrecruitment.Modelsofemotionshavebeenusedinotherroboticsresearch.Themostcommonuseappearstobehomeostaticcontrol,wheretherobotusesdrivesandemotionstomaintainaninternalstateforexample,tokeepitsbatterychargeatanacceptablelevel,ortodetectalackoftaskprogress.Applicationsfavorentertainmentandhuman-robotinteraction,withonlyonecasewhereemotionswereusedforcontrollingteaminteractions[64].Noworkhasbeenfoundthatappliesamodelofemotionstomulti-robottaskallocation.TheunderlyingemotionalmodelforthisworkistakenfromOrtony,Clore,andCollins[73][71]36

PAGE 46

[72].Inthatmodel,emotionsaregeneratedinreactiontogoal-relatedeventsandtheactionsofagents.ThisworkborrowstheSHAMEemotionfromthestandards-basedcategoryofthemodel,whichrepresentsthedegreetowhicharobothasanegativereactiontoitsownrefusaltohelpanothermemberoftheteam.Giventhisbodyofresearch,itisapparentthatalthoughtherearenumerousstrategiesforsolvingtherecruitmentproblem,thebestexistingmethodshaveahighcommunicationsoverhead,arepronetomakingpoordecisions[28],ormakeassumptionsabouttherobotteamthatmaynotbetruei.e.thatallrobotswilloperateonthesamesetoftasks,orhaveasharedworldmodel.Theapproachpresentedhereusesadifferentstrategy,whichistoapplyamodelofemotions.Emotionshavebeenusedinrobotsinthepast,buttheliteraturehasonlyidentiedoneinstancewherehavetheybeenusedforteamcoordination,andtherearenoinstancesfortaskallocationastheworkin[64]assumedstaticrolesfortherobots.Distributedsensingresearchhasnotprovidedaclearsolutiontomatchingthebestsensortoasensingtask,andsuggestsovercomingfailuresthroughredundancy.Theaffectiverecruitmentapproach,presentedinthenextchapter,appliesemotionstotherecruitmentproblemtoreducethecommunicationsoverheadbutwithoutsacricingrobustnessandwithoutputtingadditionalconstraintsontherobots.ThisapproachbuildsontheContractNetProtocol[90][21]usinganemotionalvariablefromtheOCCmodel[73][71][72]withbroadcastmessaging.37

PAGE 47

ChapterThreeApproachThisthesispresentsanapproachtomulti-robottaskallocation,focusingontheproblemofrecruitment,inwhichonerobotrequeststheassistanceofanotherrobotinordertocompleteatask.Therecruitmentstrategyisbasedonthecontract-netprotocolCNP[90][21]similartothatusedbyMURDOCH[30]butusesemotionsinsteadofactingasaninstantaneousgreedyscheduler.Thisrecruitmentapproachmatchesexactrequestsfortypesofperceptsprocessedsensoryinformationthatcanguidemotion,suchasapolarrangeplot,butfavorsthosewithahighertness,andisguaranteedtosucceedifanappropriaterobotisavailableandincommunicationscontactwiththerequestingrobot.Thisapproachdoesnotrequiretherobotstomodeleachother:therobotsneednotknowwhatotherrobotsareintheteamorwhattheircapabilitiesare.Theuseofemotionsresultsinaloweruseofcommunicationsbandwidthcomparedtothegreedyapproach,recruitmentthatislessdependentontheorderinwhichrequestsarrive[30],andthecapabilityofndingsolutionsthatagreedyapproachwouldmiss[28].ThischapterbeginswithadiscussionofthecommunicationprotocolusedinthisapproachinSection3.1,followedbyaformaldescriptionofaffectiverecruitmentinSection3.2.TheissueofdeterminingthetnessofarobotforataskisexploredinSection3.3.AsummaryisprovidedinSection3.4.3.1RobustCommunicationProtocolInthisapproach,axedrecruitmentcommunicationprotocolbeginswitharobotrequesterbroadcastingarequestforassistanceintheformofaHELPmessage,andendswhenanotherrobotresponderhasarrivedandbegunperformingataskonbehalfoftherequester.Thecommunicationprotocolisindependentoftheallocationmethod,andcouldbeusedwithothersystems.AsdiscussedinChapterOne,threecommunicationissuesguidethedesignofthisprotocol.First,therecruitmentalgorithmshoulduseaminimalamountofbandwidth.Applicationsthatrequirelow-powerorstealthybehaviorbenetfrompreventionofunnecessarytransmissions,andinanycase,thecommunicationrequirementshouldscalewellwiththenumberofagents.Second,thedeliverymethodformessagesisbroadcast.Therearetwoprimarynetworkmessagingmodesthatarerelevanttothiswork:38

PAGE 48

Table8.Recruitmentprotocolmessagesandparameters. Message Parameter Description HELP Percept Theperceptthatisrequiredforthetask.Anyrobotthatcanprovidethatperceptcanrespond,regardlessofhowtheperceptisproduced. Location Thelocationwherearecruitedrobotisneeded.Thisisrep-resentedinacommoncoordinatesystem,suchaslatitudeandlongitude. ACCEPT ETA TheestimatedtimeforthetransmittingrobottoarriveatthelocationprovidedintheHELPmessage.AsdiscussedinSec-tion3.3,thiscouldbereplacedbyamoregeneraltnessfunc-tion. RESPONDER ID Auniqueidentierrepresentingtherobotthathasbeenchosenrecruitedtoperformthetask. ARRIVAL Leaseduration Theamountoftime,inseconds,thattherobotiswillingtostayandperformthetaskitwasrecruitedfor. AGREE noparameters ACKACK noparameters unicastandbroadcast.Unicastmessagingimpliesthattransmittedmessagesarereceivedbyatmostonerecipient.Broadcastimpliesthattransmittedmessagesarereceivedandreadbyallreceiversinrange.Inourtestdomain,therealrobotsusedawirelessnetworktocommunicate.WirelessEthernetchannelsareasharedmedium,soanytransmissionsareautomaticallybroadcasts,andreceivedpacketsthatarenotintendedforaparticularrobotaresimplyignored.Asabenecialside-effect,theamountofnetworktrafcscalesslowlywiththesizeoftherobotteamasisshowninChapter4.2.Broadcastsaretypicalofothermulti-robottaskallocationmethods,includingALLIANCE[75]andMURDOCH[30],thoughsomeapproachesuseunicastmessagingi.e.LEMMING[70].SeealsoChapter2.2.Thethirddesignissueisthattheprotocolmustberobustintermsofnetworkfailure.Inafullydistributedsystem,itisassumedthatanythingcanfailatanytime,andthatnomemberoftheteamshouldwaitforeverforafailedrobottorespond.Therefore,therecruitmentprotocolisbasedona3-wayTCP/IPhandshakeandrecoversgracefullyfromlostmessagesorfailedrobots.Therecruitmentprotocolusesasetofsixmessages.EachmessagecontainstheIDnumberofthesender,theIDnumberoftherecipient,ifany,andamessagetype.Therearesixmessagetypesintherecruitmentprotocol:HELP,ACCEPT,RESPONDER,ARRIVAL,AGREE,andACKACK.ThecontentsofthemessagesaredetailedinTable8.TheprotocolisshowngraphicallyinFigure6.TheprotocolbeginswhentherequesterrobotbroadcastsaHELPmessagewithitslocationandaperceptthatarobotmusthavetobearesponder.If39

PAGE 49

Figure6.Recruitmentprotocolintermsofthemessagessentbetweenrobots.anotherrobotdecidestoassist,thenitrespondswithanACCEPTmessagethatcontainsanestimateofthetimeneededtoreachtherequesterbasedonthelocationprovidedintheHELPmessageandtherobot'srateoftravel.TheprocessbywhicharobotdecideswhethertoassistisdescribedinSection3.2.WhentherequesterrobotreceivesatleastoneACCEPTmessage,itbroadcastsaRESPONDERmessagetoallrobotswiththeIDofthechosenresponder.Fortheresponderrobot,thisservesasconrmationthatitsoffertohelpwasaccepted,anditwillbeginmovingtoassist.Forallotherrobots,thismessageisanexplicitnoticationthattheirhelpisnotneeded.ThoughtheprotocolresemblesthatusedinMURDOCH[30],theprotocolwasdevelopedindependently.TheHELP,ACCEPT,andRESPONDERmessagesinthisapproachareequivalenttothetaskannouncement,bidsubmission,andcloseofauctionmessagesinMURDOCH,respectively.ThesecondstageoftheprotocolbeginswhentheresponderrobotarrivesneartherequesterrobotandprovidesanARRIVALmessage,whichcontainsthedurationofalease.Leasesareausefultoolfordistributedsystems,astheypreventdeadlockinthecaseofapartialfailurei.e.onerobotstopsresponding.Byofferingalease,theresponderrobotindicatesthatitiswillingtostayandperformataskforthedurationofthelease,whichismeasuredinseconds.Ifnecessary,theleasecanberenewedextendedtokeeptheresponderontaskforaslongasnecessary.Whentheleasenallyexpires,eitherbecausethetaskhasbeencompletedorbecausetherequesterisnolongerresponding,thentheresponderrobothasdoneallthatitwasaskedtodoandisfreetoresumeitsowntasks.Iftherequesterrobotagreestothelease,thenitwillrespond40

PAGE 50

withanAGREEmessage.Finally,theresponderrobotwillsendanACKACKmessageandbeginthenewtask.ThesethreemessagesARRIVAL,AGREE,ACKACKcanberepeatedasnecessarytoextendthelease.Oncethetaskiscomplete,theagreedleasedurationwillexpireandtherecruitmentends.Therobustnesstocommunicationfailuresinthisprotocolresultsfromtheexpectationthatarobot'stransmissionwillproduceaparticularreply.Forexample,HELPmessagesproduceACCEPTresponses.Ifarobottransmitsamessageanddoesnotreceiveanexpectedreply,theneithernorobotswereincommunicationsrange,ornorobotschosetoreply,ortherewassomesortofcommunicationsfailure.Therobotcansimplyretryifanexpectedmessagedoesnotarrivewithinashortperiodoftime.IntheexperimentsinChapterFour,iftherequesterdidnotreceiveanexpectedACCEPTmessagewithinveseconds,oranyothermessagewithinfteenseconds,thenitwouldtimeoutandstarttheprotocolfromthebeginning.Thisprotocolcouldalsobeimplementedsuchthattherobotattemptstorecoverfromitscurrentstatewithoutstartingover.3.2FormalDescriptionofAffectiveRecruitmentTheaffectiverecruitmentstrategyusesanemotionalmodeltodetermineunderwhatconditionsarobotwillrespondtoaHELPmessage,assumingthatitisotherwiseavailablenotontask,abletoprovidetherequiredpercept.Themodelcurrentlyusesasinglestandards-basedemotion[72][73][71],SHAME,thatmodulatesresponsestoHELPmessagesanddetermineswhenarobotwillallowitselftoberecruited.Thenotationforthismodelwillbepresentedrst,followedbydetailsonhowtochoosetheparameters,andthenadiscussionoftheoperations.Thenotationusedisasfollows.Givenateamofnrobots,fr1;:::;rng,eachrobotriintheteammaintainsalevelofSHAME,si,suchthat0si1,andsiisinitializedtozero.AsarobotrefusestohelpitsteammatesbyignoringHELPmessages,itsSHAMEincreases.WhenitslevelofSHAME,si,passesathresholdintroducedbelow,therobotwillrespond.Oncetherobotdecidestorespond,itsSHAMEwillberesettozerojustasmotivationsinALLIANCEareresetwhentheycrossathreshold[75].AsummaryofthenotationisprovidedinTable9.TheprimaryparametersforSHAMEarethedecayrate,threshold,andaccrualrate.TheSHAMEdecayratewillbediscussedrst,withtheremaindertofollow.Theuseofadecayfunctionforemotionsismentionedin[71][79][98],butnoneoftheseprovideguidanceastohowtheemotionshoulddecay.[98]suggeststhatitcanbelinearoraccordingtosomeotherfunction.Otherapplicationsofemotionstorobots,41

PAGE 51

Table9.Summaryofthenotationusedinaffectiverecruitment. Symbol Description ri Theithrobotoftheteam si SHAMEoftheithrobot t SHAMEthreshold i Estimatedtimeofarrivalforrobotitothetasklocation d Functionthatdeterminestnessover[0;1]givenatimetoarriveover[0;1 mi=di Fitnessofrobotitotherequester'stask mideal Typicaltnessoftheidealresponder T TotaltimeelapsedsincelastreceivedHELPmessage kT SHAMEdecayafunctionofelapsedtime MaximumamountoftimethatrequesterwaitsforanACCEPTmessagebeforegivingupandsendinganotherHELPmessage suchas[3]and[6],appeartousealineardecayuponinspectionofmotivationovertimeplotsintheirpublications.Inthiswork,SHAMEdecayslinearly,butanyfunctioncanbeused.Fouradditionaltermsaffectthebehaviorofaffectiverecruitment.Eachrihasathreshold,t,wheret1,thatdeterminesthepointatwhichtherobotwillrespond.cisaconstantthatisaddedtosieachtimeriignoresaHELPmessage.disatnessfunctionthatincreasessibasedonhowwellsuitedriistothetask.Inthiscase,disafunctionoftheestimatedtime,i,thatriwouldneedtoreachtherequester.Lettheactualtnessofrobotriwithestimatedtimeibedenotedasmi,wheremi=di.Notethatamoresophisticatedmetricofarobot'ssuitabilityforrecruitmentcouldbeused.Suchmetricscouldincludeadditionalattributesoftherobotandtask,suchastheupdaterate,sensorresolution,orpowercost.MetricsareaddressedinmoredetailinSection3.3.kisadecayfunctionintermsofelapsedtimeTsincethepreviousreceivedHELPmessage.istheamountoftimethattherequesterwillwaitaftersendingaHELPmessagebeforegivingupandsendinganother.NotethedifferencebetweenandT,asfollows:pertainsonlytotherequesteranddetermineshowquicklyitwillretransmitaHELPmessagewhenitreceivesnoresponse.Tpertainstotheremainderoftherobotteam,andistheelapsedtimebetweenreceivedHELPmessages.isaconstant,whereTcanvarywithcommunicationfailuresandtheamountoftimebetweenrecruitmentepisodes.WhenaHELPmessagearrives,eachrobotriwillrstaccountforthedecayofitsSHAMEsincethepreviousrequest.RatherthanhaveSHAMEdecayincrementallyovertime,thetotalamountofdecaykTissubtractedfromsiatonce.Thatis,siisupdatedassi=si)]TJ/F11 9.963 Tf 9.963 0 Td[(kT.Next,ifsi>t,thenrobotrisends42

PAGE 52

anACCEPTmessagetotherequesterthatincludesitstnessi.Otherwise,ifsit,thenriignorestherequestandsiisupdatedassi=si+c+di.TherequesterrobotwillcontinuetosendHELPmessageseverysecondsuntilitreceivesanACCEPTmessageinreply.IftherequesterreceivesmorethanoneACCEPTmessageinresponsetoasingleHELPbroadcast,thenitwillexaminetheACCEPTmessagesandchoosethesenderthatspeciedthebesttnesstothetaskinthiswork,theleasttimeneededtoarrive.NotethattheremaybedelaysinthecommunicationsystemthatpreventanACCEPTmessagefromarrivinguntilafterthenextHELPmessageissent.However,therequesterwillconsideranyACCEPTmessagethatitreceives,regardlessofwhichinthesequenceofHELPmessagespromptedit.Itmaybedesirableforthisrecruitmentprocesstotakeplaceinstantaneously,especiallyifthereisanemergencyconditionforwhichanyavailablerobotshouldrespond,regardlessofitsrelativetnesstothetask.Thisapproachalsoallowsforsuchanemergencyrecruitment,inoneoftwoways.First,therequestercouldsendoutarapidsuccessionofHELPmessages,whichwouldquicklymotivateoneormorerobotstorespond.Second,theHELPmessagecouldbemodiedtoincludethethresholdvaluetsuchthattherequesterspeciedatwhatpointanotherrobotwouldrespond,andthisthresholdcouldbesettolessthanzero.Thelatterwouldcauseaffectiverecruitmenttorevertbacktoagreedyinstantaneousscheduler.However,affectiverecruitmenthasnotbeentestedforemergencyrecruitment.Theperformanceofthisapproachdependsonthechoiceofthevaluesc;tandthefunctionsd;k.Ingeneral,affectiverecruitmentwillrequirefewermessagesthangreedyifc+d>tisgenerallytrueforasmallsubsetoftherobotteamideally,onlytherequester'simmediateneighbororneighbors;thatis,ifonerobotaccruesenoughSHAMEinasinglerequestaccruingc+dtoexceedthethresholdt,thenthatsinglerobotwillrespond,thusconservingbandwidth,behaveexactlylikethegreedyapproachift<0,becausetheleastSHAMEthatarobotricanhaveiszero,sosi0>twouldalwaysbetrue;thus,riwouldalwayssendanACCEPTmessagewhenaHELPmessageisreceived,andrequiremoremessagesthangreedyifc+dnt,becausesionlyaccruesc+dperrequest,andmanyrequestswouldbenecessarytomakesi>t.Further,ifc+disconstant,thenallrobotswillexceedttogetherandrespondatonce,leadingtoalargecommunicationsoverhead.43

PAGE 53

Theeffectoftheseparametersontheoverallperformanceoftheaffectiverecruitmentapproachisasfollows.Therearetwosourcesofmessagesinthisprotocol:therequester,andanynumberofresponders.Inthegreedyapproach,onerequestcausesresponsesfromallotherrobots;iftherearenrobotsintheteam,thenoneHELPmessagewillcausen)]TJ/F8 9.963 Tf 9.962 0 Td[(1ACCEPTmessages.Intheaffectiveapproach,multiplerequestsmayberequiredbeforeanyresponseoccurs:largerincreasesinSHAMEperrequestresultinasmallernumberofrequestsbeforesomesi>t.Similarly,thenumberofresponsestoanysinglerequestwilltendtoincreasewiththeamountofSHAMEperrequest.Findingthepointatwhichaminimumnumberofrequestsgeneratesaminimumnumberofresponseswillproducethebestperformanceforaffectiverecruitment.Thetotalnumberofmessageswouldbethesumoftherequestsandresponses.Unfortunately,itisdifculttondtheidealamountofSHAMEtoassignarobotforignoringarequesttoreachthisminimum.Thus,reachingthebestpossibleperformanceisnotstraightforward.However,appropriatevaluescanbefoundheuristically.AdiscussionofhowthevalueswerechosenfortheexperimentsinChapterFourcanbefoundattheendofthissection.Itisrecommendedthattheparametersbechosenasfollows.Determineanacceptableupperlimitltothenumberofrequeststhatcanbeissuedwithoutaresponse.Determinewhatthetypicaltnessoftheidealresponderwillbe,denotedhereasmideal.Ifthetimetoarriveisusedasatnessmetric,thenmidealwouldbetheaveragetimethatarobotwouldrequiretoreachitsnearestneighbor.Forasensornetwork,midealcouldbetheexpecteddistancebetweenneighboringnodes.SpecifytheperiodoftimethatwillelapseafteraHELPmessagebeforetherequestergivesuponanACCEPTresponseandtriesagain.FortheexperimentsinChapterFour,avalueof=5secondswasused.Selectaninitialvalueforthethresholdt,suchas0.75.SelectanappropriatedecayratekTforyourdomain.Simplydeterminehowlongthemotivationfromonerecruitmentshouldpersistandinvertthatvalue.Forinstance,tohaveallSHAMEdecayafter200timeunits,letkT=0:005T.Fromthesevalues,settheparametersasfollows:c=t=l;44

PAGE 54

abFigure7.Exampleofaveragebesttnessbeingusedtogeneratereplies.Inthiscase,robot6makesarequest,shownina.Thiswillbeignored,buttheSHAMEof6'snearestneighborswillexceedthethresholdtasaresult.Thiscausesthemtorespondtothesecondrequest,showninb.NotethatwhiletherestoftheteamwouldhaveincreasedtheirSHAMEonlythosewithinaradiusofmidealwouldrespond.Inthegreedyapproach,alloftheotherrobotswouldhaveimmediatelyresponded,resultinginmorecommunicationsthannecessary.dmi=tmideal mi)]TJ/F11 9.963 Tf 9.963 0 Td[(c+kwheremiistherelativetnessofrobotri.Ifsetinthisway,arequesterwilltendtobeansweredquicklyaftertwocallsifmi
PAGE 55

parametersforanewdomain.Thethresholdtwaschosentobe0.75,andc=0:2wasselectedsothatdistantrobotswouldtendtorespondafterapproximatelyfourrequests.Twotnessfunctionsdiwereusedintheexperiments,whereirepresentstheestimatedtimearobotriwouldrequiretoreachtherequester.Therstislinear:di=0:5=iwaschosensothatarobotrirespondedwithintworequestsifitwaswithin1unitoftherequester,anddhadlittleeffectbeyond10units.Thesecondwasnon-linear:di=2:5=2i.ThepurposeofthesedifferentchoicesfordarediscussedinChapterFour.ThedecayfunctionkdetermineshowquicklyriwillloseitsSHAMEafterignoringarequest.ThefunctionkT=0:005TwasusedsothatriwouldloseSHAMEacquiredfromasinglerequestinabout40seconds,andwouldrequire200secondstogofromsi=1tosi=0.Thisrelativelylowrateofdecaykeepsriresponsivetotheneedsoftheteam;ifthedecaywerefaster,thenperiodicrequestswouldtendtobeignored.Ifthedecaywerenegative,thenriwouldtendtowanttohelpmoreovertimeuntilitwasrecruitedandsiwassubsequentlyresettozero.3.3MultivariateMetricEvaluationFunctionsThisapproachusesameasureofthesuitabilityofarobottorespondtoaparticularrequestforhelp,ortnessmetric,toincreasearobot'sSHAME.GerkeyandMataricdescribemetricsasameansofdiscriminatingamongateamofrobotstochoosetherobotbestsuitedtothetaskathand[30].However,theonlyexamplesprovidedaretheCartesiandistancebetweentherobotandtasklocation[30]whichdoesnotconsidertherobot'svelocityorrouteandthecomputationalloadofeachrobot[27].Theliteraturedoesnotprovideaconsistentorextensiblemeansofmeasuringthetnessofarobottoatask.Suchatnessmeasureisbeyondthescopeofthisthesis,butthemotivationfordevelopingthismeasureandthechallengesthatitpresentsareprovidedbelow.Forthepurposesoftesting,thetnessmetricusedforexperimentationinChapterFourwasanestimateofthetimerequiredfortherobottoreachthetasklocation.Considerthatarobotisneededforatask,andthatcompletingthetaskrequirescertaincapabilities,suchasdetectingamineandavoidingobstacles.Multiplerobotsintheteammayhavesuitablesensors,effectors,andalgorithms,buttheirrelativetnessdetermineswhichrobotisrecruitedtoperformthetask.Itisassumedthattherequirementsofthetaskareknown,andcanbedescribedasacollectionofindividualcapabilitiesdescribedbelow.Eachrobot,uponreceivingarequestthatcarriescertainrequirements,candetermineitsownsuitabilitytothetaskandupdateitsSHAMEaccordingly.AsdiscussedinChapter2.1.2.2,atleastsixmethodshavebeenusedtomeasurethetnessofarobottoatask:46

PAGE 56

Cartesiandistancebetweenrobotandtask.ThismethodhasbeenusedbyGerkeyandMataricinMURDOCH[30][27].Estimatedtimeforrobottoarriveattasklocation.ThisisrelatedtotheCartesiandistance,butalsoaccountsforheterogeneityintherobots,especiallytheirvelocity.Thisthesisusestheestimatedtimemetric.Costofperformingthetask.ThismethodhasbeenusedbyLindnerandMurphy[51]andZheng[104].Costcanbemeasuredintermsofthepowerorsomeotherniteresourceconsumed.However,accuratelyestimatingthecostofataskcanbedifcult.Reductionofuncertainty.Thedegreetowhicharobot'ssensorscanreduceuncertaintyinitsreadingshasbeenconsideredbyXuandVandorpe[103]andbyGageandMurphy[26].Updaterate.In[104],Zhengincorporatedtheresponsetimeofarobotgivenaparticularsensorintoautilitymeasure.Adhoc.Therelativeutilityofeachrobotorsensorcanbeenumeratedbyahuman,aswasdonebyGageandMurphy[26].Thefollowingattributescouldalsobeusedtodeterminethetnessofarobotortherobot'ssensorstoatask,butinstancesofthesehavenotbeenfoundintheliterature:Maximumscanangle.Arobotmayonlybeabletoprovidethedesiredperceptoveracertainangle.Sensorresolution.Twosensorsmightmeasurethesamepropertyforinstance,range,butmaydosoatdifferentresolutionsforinstance,onemaymeasureaccuratelytowithinamillimeter,andtheothermayroundtothenearestmeter.Maximumrange.Sensorscanoftenonlymeasureoveraparticularrange.Forinstance,camerasmayhaveaxedfocallengthandzoom,andthetimethatasonartransducerwaitsforanechomaylimititseffectiverange.Thisalsoreferstotheparticularfrequenciesorconcentrationsthatasensorcanmeasure;forinstance,somecamerasmaydetectvisiblelightwhereothersdetectonlyinfrared.Ideally,atnessmetricwouldbecapableofconsideringanyorallofthesemeasures.Notethattheproblemofdeterminingtherelativetnessofarobottoataskusingacombinationoftheseattributesisdifcult,forthefollowingfourreasons.47

PAGE 57

Inadistributedrobotteam,eachrobotmustdetermineitsowntnesswithoutanyinformationaboutthecapabilitiesofotherrobots.Theteamitselfisdynamic,anditcannotbeassumedthatthecapabilitiesofallrobotsareknownglobally.Thus,therobotcannotdoacomparisontootherrobotsdirectly;themetricmustbeobjective.Themetricmustallowdisparateattributestobecombinedandcompareddirectly:boundedvaluessuchasprobabilitiesboundedin[0;1]andanglesboundedin[0;2mustbecomparabletounboundedvalues,suchastimeordistanceover[0;1.Itmaybepossibletoscalethevaluesintoacommonrange,butsuchscalingwouldrequirebalancingtheattributesagainsteachother,whichcouldbeverydifcult.Forexample,whatpowercostwouldcontributeanamountofutilityequaltohaving75%accuratesensors?Themetricsmustnotassumethatcomparisonsaresymmetric.Forexample,thetimethatrobotArequirestoreachrobotBisnotnecessarilythesameasthetimerobotBrequirestoreachA,becausetherobotsmaytravelatdifferentvelocities.Itisnotclearhowpartialmatchesintermsofsetinclusion/exclusionshouldberesolved.Thatis,ifthreecapabilitiesarerequiredtogether,andarobotcanprovideonlytwo,itisnotclearhowtherobot'sutilityshouldbeadjusted.Solvingtheproblemofcreatingageneraltnessmeasureisbeyondthescopeofthisthesis,butthechosenmetricestimatedtimetoarriveisasuitableapproximationofarobot'stness.Thefundamentalcontributionofthisapproachistheuseofanaffectivevariabletoinuencerecruitment,andthisapproachcanbeadaptedtouseanytnessfunction.Theestimatedtimetoarriveisadequatefortestingtheperformanceoftheapproach,especiallywhencomparedtothemetricsfoundintheliteraturee.g.distance[30]andcost[51][104].Ingeneral,anyqualityoftherobotsforwhichamaximalvalueaccuracy,updaterateorminimalvaluedistance,cost,timeimplieshigherutilitycanbeused.3.4SummaryThischapterpresentedtheaffectiverecruitmentstrategyintermsofthesixmessagesHELP,ACCEPT,RESPONDER,ARRIVAL,AGREE,ACKACKthatarepassedbetweenrobotstoenablerecruitment.Thisinterchangeofmessagesbetweentherobotsistypicalofapproachesthatusethecontractnetprotocol[90],andcloselyresemblesthatusedinMURDOCH[30].Theaffectiverecruitmentstrategyusesthisprotocol48

PAGE 58

becauseitprovidesrobustnessincasethereisalossofcommunicationsbetweenrobots;aftereachmessage,therobotswillwaitalimitedamountoftimeforaresponsebeforetryingagain.Althoughthisworkreusesthecontractnetprotocol,itistherstknownworktoapplyanemotionalmodeltothemulti-robottaskallocationproblem.Theemotionalvariable,SHAME,wasintroducedformally,alongwithadiscussionofhowtheparametersthatcontrolSHAMEcouldbegenerated.TheSHAMEvariableiscentraltothiswork,asitdetermineswhenarobotwillrespondtoaHELPmessage,anddistinguishesthisworkfromMURDOCH.EachrobothasaSHAMEvariable,whichstartsoffwithazerovalue.AsHELPmessagesarrive,therobotwillonlyrespondifitsSHAMEisaboveathreshold;otherwise,itsSHAMEincreases,butitmakesnoreply.Thisincreaseintherobot'sSHAMEreectsitsownreactiontoitsunwillingnessorinabilitytorespondtoarequestforhelp,andservesasanindicationofthedegreetowhichtherobotisnotcontributingtotheoverallgoalsoftheteam.SHAMEwilldecayovertime;intheimplementationinChapterFouralineardecayisused,whichiscommonlyapplied[6][3].However,thereisnorestrictiononwhatSHAMEdecayfunctionscouldbeused[98].Thischapteralsoexaminedtheproblemofndingametricevaluationfunctionformultiplerobotcharacteristics.Suchafunctionmustobjectivelycomparethecapabilitiesofrobots,suchthatthebestrobotforaparticulartaskcanbedeterminedwithoutacentralizedarbiter.Attributesthatcouldcontributetotheutilityofarobottoataskincludethetimerequiredtobringtherobottothetasklocationandthemaximumscanangle,resolution,range,updaterate,andaccuracyoftherelevantsensororsensors.However,developingsuchametricevaluationfunctionisbeyondthescopeofthisthesis.Forthesakeofexperimentation,anapproximationofeachrobot'sutilityisused:theestimatedtimetherobotwouldneedtoarriveatthetasklocation.Thenextchapterwillpresentexperimentsthatwereperformedtovalidatethisapproach.Theexperimentstestedthefollowinghypotheses:Affectiverecruitmentscalesbetterwithteamsizeintermsofcommunicationsoverheadthanthegreedyapproach.Affectiverecruitmentisrobustwithrespecttorandomcommunicationlosses.Anon-lineartnessfunctiondperformsbetterthanalineartnessfunctionforlargerobotteams.Broadcastmessagingisbettersuitedthanunicastmessagingforthisrecruitmentprotocol.Affectiverecruitmentcanreachsolutionsthatthegreedyapproachcannot,andcanrecruitwithoutrequiringallrobotstorespond.49

PAGE 59

Affectiverecruitmentselectsrobotsequallyofteniftheyareapproximatelyequallywell-suitedtothetask.TheresultsoftheseexperimentsarepresentedinChapterFouranddiscussedinChapterFive.50

PAGE 60

ChapterFourExperimentsExperimentswereperformedtocomparetheaffectiverecruitmentstrategy,withlinearandnon-linearSHAMEaccrualfunctions,againstthegreedyinstantaneousschedulerusedinMURDOCH[30]andarandomscheduler.Thereweresixprimaryobjectivesforthesetests.Testtheeffectofvaryingtheteamsizeoneachstrategy.ThemetricsforcomparisonwerethetimenecessarytoperformrecruitmentandthenumberofmessagestransmittedamongtherobotsseeSection4.2.1.Testtheimpactofrandomcommunicationfailuresupto25%ontheperformanceofeachstrategy,againmeasuredusingthetimeneededtocompleterecruitmentandthenumberofmessagestransmittedseeSection4.2.2.TesttheeffectofalinearSHAMEupdatefunctionversusanon-linearfunctionwithregardstochaoticbehaviorforverylargeteamsseeSection4.2.1.JustifytheuseofbroadcastmessaginginsteadofunicastmessagingfortransmittingmessagesbetweenrobotsseeSection4.2.3.Testscenariosinwhichagreedyinstantaneousscheduler,suchasMURDOCH,choosesasub-optimalallocation,whereastheaffectiveapproachperformsbetterbydelayingthedecisionovertimeseeSection4.2.4.Testthedegreetowhichallrobotsarerecruitedequallyoftenbythefourrecruitmentstrategies.ThemetricforcomparisonwasthenumberoftimeseachrobotwasrecruitedrelativetoanexpectedmeanvalueseeSection4.2.5.ThischapterbeginswithadescriptionoftheexperimentaldomainandtherecruitmentsimulatorinSection4.1.ThesimulationsthatwereperformedtosatisfythesixobjectivesabovearedescribedinSection4.2,alongwiththeirresults.Theimplementationofaffectiverecruitmentonrealrobothardwareand51

PAGE 61

subsequenttestsarediscussedinSection4.3.AsummaryoftheexperimentalresultsisprovidedinSection4.4.4.1ExperimentalDesignFourrecruitmentstrategiesdescribedindetailinSection4.1.2weretestedinsimulationforaminedetectiontask:greedy,random,affective,andaffectivewithanon-linearmetric.Thepurposeoftheseexperimentswastomeasuretheperformanceoftherecruitmentstrategiesaccordingtothreemetrics:thenumberofmessagessentamongrobotsintheteam,thetotalamountoftimethatarobothadtowaitforassistance,andthetotalnumberoftimeseachrobotwasselectedforrecruitment.Thesizeoftherobotteam,rateofrandomcommunicationfailures,andmessagingtypeunicastorbroadcastwerevariedtotesttheimpactontherecruitmentprocessforeachstrategy.Section4.1.1describesthescenarioinwhichrecruitmentwastested,andSection4.1.2describestherecruitmentstrategiesinmoredetail.ExperimentalresultsareprovidedinSection4.2.4.1.1ScenarioThetaskdomainfortheexperimentswasamockmine-detectiontasksuppliedbyNAVSEACoastalSystemsStation.Inthisdomain,ateamofrobotsworkcooperativelytoidentifylandmines.Tolocateandidentifymines,asingleunmannedaerialvehicleUAVperformsacoarsesearchoveranarea,usingitsonboardsensorstondobjectsthatcouldbemines.Onceamine-likeobjectisdetected,anunmannedgroundvehicleUGVwithadditionalsensorsisdispatchedtoperformacloserinspectionastheUAVresumesitssearch.Forthesimulations,onerobotwasdesignatedasaUAVthatperformedarasterscanovera100100-unitgridatarateof3unitsperiteration.Attheendoftherasterscan,theUAVstoppedfor20secondsbeforeperformingthescanagainintheoppositedirection.Atvexedlocationsinthisscan,theUAVstopped,requestedassistance,andwaitedforanotherrobottoarrivebeforecontinuingon.ThelocationswheretheUAVstoppedweredeterminedbyhavingittravelforxeddurationsbetweenrecruitmentepisodes.Thedurationsbetweentheverecruitmentswere45,110,180,and70seconds,suchthattherobotSHAMEwoulddecaydifferentamountsbetweenrecruitments.Additionalsimulatedrobots,representingUGVs,werealsoplacedinthe100100-unitgrid.ThenumberofUGVsvaried:intheteamsizeexperimentsseeSection4.2.1from3to52UGVswereused;inthecommunicationfailureexperimentsseeSection4.2.2,12UGVswereused;andinthefairnessexperimentsseeSection4.2.5,5UGVswereused.Intheteamsizeandcommunicationfailure52

PAGE 62

Figure8.Userinterfaceforrecruitmentsimulator.TheUAVcenterrequestsassistance,andalleligiblerobotswithsufcientSHAMErespondsolidlines.ThosethatignoretherequestaremarkedwithanX.experiments,twooftheserobotsweretaskedwithrasterscansofhalfofthegrideach,travelingatarateofoneunitperiteration,andstoppingfor20secondsbecomingtemporarilyavailableforrecruitmentattheendofthescanbeforerestartingintheoppositedirection.Theremainingonetoftyidlerobotsweredistributedrandomlyacrossthegrid:30randomstartingcongurationswereusedfortheteam-sizeexperiments;5forthecommunicationlossexperiments;and10forthefairnessexperiments.TheseidlerobotswereavailableforrecruitmentbytheUAVatanytime.ThesimulatorwasimplementedinJava.RobotswererepresentedasobjectsthatcommunicatedthroughJINI,atechnologyforbuildingandmanagingdistributedsystems.Thesoftwarearchitectureunderwhichthesimulatorwasbuiltallowedforaseamlesstransitionfromsimulatedtorealrobots,soitwasexpectedthatthesimulationresultswouldbeindicativeofrealrobotperformance.Thesimulatorcontrolledthreeexperimentalparameters:whichrecruitmentstrategytouse,therateofrandomcommunicationfailures,andwhichmessagingmethodunicastorbroadcasttouse.Thesimulator'suserinterfaceisshowninFigure8.53

PAGE 63

4.1.2RecruitmentStrategiesFourrecruitmentstrategiesweretestedinsimulation.Thersttwowereaffectiverecruitment,inwhichtheclosestidlerobotwhoseSHAMEwasaboveathresholdwasrecruitedforeachrequestasdescribedinChapterThree.Fortherstaffectivestrategy,theamountofSHAMEthatarobotreceivedforignoringarequestdecreasedlinearlywiththedistancebetweentherobotandtherequesterforthedistanceDbetweentherobots,SHAMEwouldincreaseby0:5=D.Forthesecondstrategy,theSHAMEforignoringarequestdecreasedwiththeinversesquareofthedistancebetweentherobots.Thatis,giventhedistanceDbetweentherobotandtherequester,SHAMEincreasedby2:5=D2.Thepurposeoftestingtwovariantsofaffectiverecruitmentwastodeterminewhetheralineartnessmetric1=Dwouldbegintoexhibitchaoticbehaviorforlargeteamsofrobots,andwhetheranon-lineartnessmetric1=D2wouldpreventthisbehavior.Inthiscase,chaoticbehaviorreferstoalargevarianceinthetimerequiredforaffectiverecruitmentrequiresastheteamsizeincreases.ItwassuspectedthatalineartnessmetricwouldspreadtoomuchSHAMEthroughouttherobotteam,causingpoorly-suitedrobotstorespondtoHELPmessagesandberecruited,andgenerallymakingthechoiceofrobotsunpredictable.Thenon-lineartnessmetricwasaddedtodeterminethedegreetowhichthischaoticbehavioroccurred.TheparametersthatcontroltheperformanceofSHAMEthatwereintroducedinChapter3.2weresetasfollows.Notethatthesevalueswerechoseninanadhocmanner,andnotaccordingtotheheuristicmethodthatissuggestedinChapter3.2.Thethresholdtwaschosentobe0.75,andc=0:2wasselectedsothatdistantrobotswouldtendtorespondafterapproximatelyfourrequests.Therateofdecay,kT,wassettokT=0:005T.Thethirdrecruitmentstrategywasgreedyrecruitment,inwhichtheidlerobotwiththeminimumestimatedtimetoarrivewasrecruitedforeachrequest.ThisstrategyisrepresentedintheliteraturebytheMURDOCHsystem[30][27][28][31],whichisconsideredtobethestateoftheart.NotabledifferencesbetweenMURDOCHandthisapproachwerediscussedinChapter2.1.2.ItwasexpectedthatgreedyrecruitmentwouldproducefasterresponsetimesthanaffectiverecruitmentbecauseitdoesnotspendtimebuildingupSHAMEbeforearobotisrecruited.However,itwasalsoexpectedthatgreedyrecruitment'scommunicationoverheadwouldincreaselinearlywithteamsize:eachtimeaHELPmessagewassent,everyidlerobothadtoreplysothattherequestercouldchoosetherobotwiththeleastarrivaltime.Thus,itwasexpectedthatalargerteamwouldequatetogreatercommunicationoverhead.Thefourthrecruitmentstrategywasrandomrecruitment,inwhichanidlerobotwaschosenatrandomforrecruitment.WhentherequestertransmittedaHELPmessage,eachidlerobotrepliedtoindicate54

PAGE 64

itsavailability,andoneofthesewasthenchosenrandomly.Aswithgreedyrecruitment,thiswasexpectedtoresultinafasterdecisionthanwithaffectiverecruitment,butwouldnotscaleaswellforcommunication.Further,randommaychooserobotsthatarefaraway,whichislikelytoresultinlongerresponsetimes.Therandomstrategywaschosenasabaseline,asitisanuninformedmethodofrecruitment.Thatis,comparisonstoarandommethodprovideanassurancethattheothermethodsaresomewhatintelligent.4.2ExperimentalSimulationsTheresultsfromvesetsofsimulationsareprovidedbelow.Intherstexperiment,theteamsizewasvariedtomeasureitseffectonthecommunicationoverheadandresponsetimeoftherecruitmentstrategies.ThesesimulationsandtheresultsarepresentedinSection4.2.1.TheimpactofvaryingtherateofcommunicationfailuresusingthesamemetricsisaddressedinSection4.2.2.Next,Section4.2.3describesthedifferenceincommunicationoverheadwhenusingunicastinsteadofbroadcastmessaging.Instantaneousgreedyschedulersarenotoptimal[28],andaffectiverecruitmentiscapableofreachingsolutionsthatagreedyapproachcannot.Illustrativecasesweredevisedtoexplicitlydemonstratethisfact;thesesimulationsandresultsaredescribedinSection4.2.4.Finally,therelativefrequencywithwhicheachrecruitmentstrategyrecruitsidlerobotsiscomparedinSection4.2.5.4.2.1EffectsofTeamSizeTheeffectofvaryingthesizeoftherobotteamwasmeasuredusingsixhundredsimulations:foreachofthefourrecruitmentstrategies,simulationswereperformedwith1,5,10,20,and50idlerobotsforeachofthirtydifferentrandomlygeneratedstartingcongurations530=600.Themetricsforthistestwerethetotalnumberofmessagessentamongtherobots,andtheamountoftimethatpassed,inseconds,fromtheinitialUAVrequestuntilaUGVarrivedandwasacknowledged.4.2.1.1StatisticalAnalysisThesignicanceoftheresultsforthesetestswillbedeterminedasfollows.Inatypicalexperimentaldesign,aMultivariateAnalysisofVarianceMANOVAcouldbeusedtodeterminewhethertherecruitmentstrategiesweresignicantlydifferentacrossdifferentteamsizes.Similarly,at-testcouldbeusedtodeterminewhethertheresultsforeachmetricweredrawnfromdistributionswithdifferentmeans;inotherwords,tondwhetheronerecruitmentmethodhadsignicantlyhigherorlowerscoresthan55

PAGE 65

Figure9.Histogramofthenumberofmessagestransmittedusingtheaffectiverecruitmentstrategyforteamsize13.Notethattheresultsdonotfollowanormaldistributionabellcurve,whichmakesthecommont-testinapplicable.anotherforeachmetric.However,bothMANOVAandt-testassumethatthedataresultingfromtheexperimentsaredrawnfromanormaldistribution,andcannotbeusedifthisassumptionisviolated.TheLillieforstest[86]canbeusedtodeterminewhetherasetofsamplesisdrawnfromanormaldistribution.Applyingthistesttothesimulationresultsindicatedthatnotalloftheresultswerenormallydistributed.Foranillustrationofthehowtheresultsforatypicaltestweredistributed,seeFigure9.Thus,insteadofusingthet-test,theWilcoxonranksumtest[33][8]wasusedinstead.Theranksumtestissimilartothet-test,butdoesnotassumethatthesamplesaretakenfromnormallydistributedsources,butonlythatthesamplescomefromsimilarsources.Theranksumtesttendstobemoreconservativethanthet-test,reportinghigherp-valuesforthesamesamples,soingeneral,iftheranksumtestindicatessignicance,thenthet-testcanbeexpectedtodothesame.Aranksumhypothesistestwasconductedforeachpairofrecruitmentstrategiesforthesimulationresults.Inthesetests,thenullhypothesiswasthattherecruitmentstrategiesproducedsamplesfromdistributionswithequalmedians.Thatis,foreachteamsize,atestwasperformedtodeterminewhethertheaveragevalueforeachmetricnumberofmessages,totalwaittimeforeachstrategywassignicantlydifferent.56

PAGE 66

Inordertotestwhetherthisnullhypothesiscouldberejected,avaluewaschosenforp,theprobabilityofobservingtheseresultsbychanceifthenullhypothesisweretrue.Traditionally,p-valueslessthan0.05or0.01designatesignicance.However,repeatedexperimentstendtodepressthemeasuredp-value,aneffectthatcanbecompensatedforbyusingaBonferronicorrection[24].AconservativeBonferronicorrectiondividesthelevelofcondencebythenumberofsamples,suchthatsignicanceisclaimedonlyifthesumofthep-valuesfortheentiresetofsamplesfallsbelow0.05or0.01.Inthiscase,thep-valueforsignicancewaschosentobe0:00033,whichisacondencelevelof0.01afteraBonferronicorrectionfor30samples;thatis,0:01=30=0:00033.Thus,foreachpairofrecruitmentstrategiesforeachteamsize,theranksumtestresultisconsideredsignicantforvalueslessthanp=0:00033.Theresultsoftheranksumtestswillbeprovidedforeachmetricbelow.4.2.1.2ResultsforNumberofMessagesMetricTable10showstheaveragenumberofmessagesthattherobotstransmittedforeachrecruitmentstrategyandteamsize.ThesamevaluesareshowngraphicallyinFigure10.TheresultsoftheranksumhypothesistestsofwhetherthedifferencesbetweenthestrategieswerestatisticallysignicantareprovidedinTable11.AsetofboxplotsareprovidedinFigure11toprovideasummaryofthemeansandvarianceofthesimulationdata.Theseresultsindicatethattheaffectiverecruitmentstrategyrequiredsignicantlymoremessagestobesentthangreedyorrandomwhentherewereveorfewerrobotsavailableforrecruitment.However,oncethenumberofavailablerobotsincreasedtoten,affectiverecruitmentrequiredsignicantlyfewermessages:affectiveused63.8messageswheregreedyandrandomused75messagesatap-valueof3:310)]TJ/F7 6.974 Tf 6.227 0 Td[(11.Forlargerteams,affectiveconsistentlyusedfewermessagesthangreedyorrandom.Withaffectiverecruitment,theUAVonlyneedstosendoutHELPmessagesandwaitforasinglereplytobeginnegotiatingrecruitment.Asaresult,thenumberofmessagesthatmustbesentisalmostconstant.ThevariationinthenumberofmessagesforaffectiverecruitmentoccurswhenmorethanonerobotrespondstoaparticularHELPmessage,orwhenallUGVsarefarfromtheUAVandadditionalHELPmessagesmustbesenttopushtheirlevelofSHAMEoverthethreshold.Ontheotherhand,greedyandrandommustsolicitmessagesfromallothermembersoftheteaminordertomakeachoice,andthenumberofmessagesperrecruitmentincreaseslinearlywiththeteamsize.4.2.1.3ResultsforAverageWaitTimeMetricThesimulationresultsfortheamountoftimethattheUAVspentwaitingforUGVstorespondandarriveareprovidedinTable12andareshowngraphicallyinFigure12.TheresultsoftheranksumhypothesistestsofwhetherthedifferencesbetweenthestrategieswerestatisticallysignicantareprovidedinTable13.AsetofboxplotsareprovidedinFigure13.57

PAGE 67

Table10.Averagenumberofmessagestransmittedforeachstrategyforvaryingteamsize.Thetotalnumberofrobotsintheteamisshownacrossthetopofthetable. 4 8 13 23 53 Affective 52.6 53.7 63.8 81.9 109.1 Affective,1=D2distancemetric 54.2 59.7 75.1 94.3 86.7 Greedy 30 50 75 125 276.7 Random 30 50.2 75 125 275 Figure10.Messagestransmittedatdifferentteamsizes.Thedarksolidlinetowardthebottomisaffective,andthesolidlineatopthedashedlineisgreedy.Asexpected,affectiverequiressignicantlyfewermessagestobesentforteamswith10ormoreuntaskedrobotsi.e.teamsize13.Alsonotethattheaffectivestrategyperformsbetterwitha1=D2distancemetricthanwitha1=Ddistancemetricforverylargeteams,butthedifferenceisnotstatisticallysignicant.58

PAGE 68

Table11.Pairwisecondenceintervalsforaveragenumberofmessagesforvaryingteamsize.TheseintervalswereproducedwithaWilcoxonranksumtest.Signicanceisclaimedforvalueslessthanp=0:00033.Resultsthatarenotconsideredsignicantareshowninitalics. Affective,1=D2distancemetric Greedy Random Teamsize:4robots Affective 0.002 2:410)]TJ/F7 6.974 Tf 6.226 0 Td[(12 2:410)]TJ/F7 6.974 Tf 6.227 0 Td[(12 Affective,1=D2distancemetric 2:010)]TJ/F7 6.974 Tf 6.226 0 Td[(12 2:010)]TJ/F7 6.974 Tf 6.227 0 Td[(12 Teamsize:8robots Affective 4:910)]TJ/F7 6.974 Tf 6.226 0 Td[(6 1:510)]TJ/F7 6.974 Tf 6.226 0 Td[(6 1:110)]TJ/F7 6.974 Tf 6.227 0 Td[(5 Affective,1=D2distancemetric 4:310)]TJ/F7 6.974 Tf 6.226 0 Td[(12 1:410)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Teamsize:13robots Affective 2:710)]TJ/F7 6.974 Tf 6.226 0 Td[(7 3:310)]TJ/F7 6.974 Tf 6.226 0 Td[(11 4:210)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Affective,1=D2distancemetric 0.347 0.36 Greedy 0.33 Teamsize:23robots Affective 0.015 1:210)]TJ/F7 6.974 Tf 6.227 0 Td[(12 1:710)]TJ/F7 6.974 Tf 6.227 0 Td[(12 Affective,1=D2distancemetric 3:310)]TJ/F7 6.974 Tf 6.227 0 Td[(11 4:510)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Greedy 0.334 Teamsize:53robots Affective 0.001 1:710)]TJ/F7 6.974 Tf 6.227 0 Td[(12 1:710)]TJ/F7 6.974 Tf 6.227 0 Td[(12 Affective,1=D2distancemetric 1:710)]TJ/F7 6.974 Tf 6.227 0 Td[(12 1:710)]TJ/F7 6.974 Tf 6.227 0 Td[(12 Table12.Averagetime,inseconds,theUAVspentwaitingaccordingtoteamsize.EachsimulationconsistedofverecruitmentepisodesstartingwithaHELPrequestandendingwhenaUGVarrivedandbeganitstask.Thereportedvaluerepresentsthesumofthewaittimesforallverecruitmentepisodespersimulation.Thesizeoftherobotteamisshownacrossthetopofthetable. 4 8 13 23 53 Affective 409.4 256.0 210.3 195.0 152.2 Affective,1=D2distancemetric 424.5 264.9 221.7 183.9 153.1 Greedy 271.8 144.9 117.5 107.0 132.1 Random 272.6 298.2 304.6 361.5 404.0 59

PAGE 69

Figure11.Boxplotsofthesimulationresultsforthecommunicationoverheadaccordingtoteamsize.Thelengthofeachboxisafunctionofthevarianceofthedata,andthecenterlineineachboxdenotesthemeanover30samples.Notethatthegreedyandrandomstrategiesproducedalmostconstantresults,sotheirboxesarecompressedintolines.60

PAGE 70

Table13.PairwisecondenceintervalsforaveragetimeUAVspentwaitingaccordingtoteamsize.TheseintervalswereproducedwithaWilcoxonranksumtest.Signicanceisclaimedforvalueslessthanp=0:00033.Resultsthatarenotconsideredsignicantareshowninitalics. Affective,1=D2distancemetric Greedy Random Teamsize:4robots Affective 0.181 6:110)]TJ/F7 6.974 Tf 6.226 0 Td[(11 6:710)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Affective,1=D2distancemetric 4:110)]TJ/F7 6.974 Tf 6.226 0 Td[(11 6:110)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Random 0.923 Teamsize:8robots Affective 0.501 1:210)]TJ/F7 6.974 Tf 6.227 0 Td[(9 0.004 Affective,1=D2distancemetric 5:110)]TJ/F7 6.974 Tf 6.227 0 Td[(10 0.014 Greedy 8:210)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Teamsize:13robots Affective 0.35 1:310)]TJ/F7 6.974 Tf 6.227 0 Td[(10 2:810)]TJ/F7 6.974 Tf 6.227 0 Td[(8 Affective,1=D2distancemetric 3:210)]TJ/F7 6.974 Tf 6.227 0 Td[(10 8:810)]TJ/F7 6.974 Tf 6.227 0 Td[(7 Greedy 3:010)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Teamsize:23robots Affective 0.12 9:010)]TJ/F7 6.974 Tf 6.227 0 Td[(11 6:110)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Affective,1=D2distancemetric 6:710)]TJ/F7 6.974 Tf 6.227 0 Td[(11 3:710)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Greedy 3:010)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Teamsize:53robots Affective 0.626 0.045 3:310)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Affective,1=D2distancemetric 0.010 3:010)]TJ/F7 6.974 Tf 6.227 0 Td[(11 Greedy 3:010)]TJ/F7 6.974 Tf 6.227 0 Td[(11 61

PAGE 71

TheseresultsshowthattheamountoftimethattheUAVspentwaitingforhelptoarrivefavorsthegreedyrecruitmentstrategyoveraffectiverecruitment.ThiswasanexpectedresultbecauseaffectiverecruitmentrequirestimetobuildupthelevelofSHAMEintherobotsbeforetheywillrespond.Inthesesimulations,theUAVwaited3secondsbetweenrequestsforhelpbeforecallingagain,andneededtorequestuptotentimesperrecruitmentbeforeitreceivedaresponse.IftheparametersforupdatingSHAMEweretunedorlearned,thisdifferencebetweengreedyandaffectivecouldbereduced.LearningtheseparametersisadirectionforfutureworkseeChapter6.2.Ingeneral,thefasterarobot'sSHAMEexceedsthethreshold,thecloseraffectiverecruitmentwillresemblegreedy,tothepointthatiftheSHAMEexceedsthethresholdafterasinglerequest,thenthetwostrategiesareequivalent.Thiscanbeseenintheresultsforthesimulationswith50idlerobotsintheteam.Inthesecases,thedensityofidlerobotswashigh,suchthataUGVwasrelativelyneartotheUAVwhenevertheUAVmadearequestcomparedtothesimulationsforsmallerteamsizes.ThenearbyUGVwouldquicklyrespondandberecruited,resultinginawaittimethatwasnotsignicantlygreaterthanthatforthegreedystrategyp=0:045>0:00033.Ineffect,oncethedensityofrobotsbecamegreatenough,theaffectiveapproachfunctionedlikethegreedystrategywithinasmallneighborhood,butwithagreatsavingsincommunicationoverhead6.7messagesforgreedy,109.1foraffectivewitha1=Ddistancemetric,and86.7foraffectivewitha1=D2distancemetric.Inotherwords,forverylargeteamsormoreidlerobots,simulationresultssupporttheclaimthataffectiverecruitmentusessignicantlyfewermessagetransmissionsthangreedywithoutasignicantincreaseinthetimerequiredtocompleterecruitment.Finally,theseresultsshowthataffectiverecruitmentoutperformedrandomintermsofthetimerequiredtocompleterecruitmentoncetheteamsizereached13p=2:810)]TJ/F7 6.974 Tf 6.226 0 Td[(8.Althoughrandomrecruitmentmadeachoiceimmediately,theclosestresponderwastypicallynotchosen,sothetimerequiredfortherespondertoarrivewashigherthanfortheotherstrategies.Asexpected,affectiverecruitmentalsooutperformedrandomintermsoftotalcommunicationoverheadoncetheteamsizereached13atp=4:210)]TJ/F7 6.974 Tf 6.226 0 Td[(11.4.2.1.4SummaryofTeamSizeSimulationsTherstsetofsimulationstestedtheeffectoneachstrategyofvaryingtherobotteamsize.Themetricsforcomparisonwerethetotalnumberofmessagestransmittedbytherobots,andthetotaltime,inseconds,thattheUAVwaitedforarespondertoarriveaftermakingarequest.Teamsizesof4,8,13,23,and53weretestedinatotalofsixhundredsimulations.Theresultsfromthesesimulationsindicatethatforateamof13ormorerobots,theaffectiverecruitmentstrategyrequiredsignicantlyfewermessagesthangreedyorrandomtocompletearecruitment.Further,forteamsof5362

PAGE 72

robots,theamountoftimerequiredforaffectiverecruitmentwasnotsignicantlygreaterthanthatrequiredbygreedy.Twovariantsofaffectiverecruitmentweretested,therstusinga1=DincreaseinSHAMEbasedonthedistanceDbetweentheUAVandeachpotentialresponder,andtheotherusinga1=D2increase.Regardingwhetherthe1=Dmetricexhibitedchaoticbehaviorforlargeteamsofrobots,considerthesimulationresultsforateamof53robots.Theseresultsshowthatthenumberofmessagestransmittedusingthe1=Dmetricvariedlessthanforthe1=D2metricasillustratedinFigure11,butthatthevariantswerenotsignicantlydifferentineitherthenumberofmessagestransmittedp=0:001,wherep<0:00033wouldbesignicant,seeTable11orthetotalwaittimep=0:626,wherep<0:00033wouldbesignicant,seeTable13forthisteamsize.Inotherwords,theresultsforthe1=Dmetricweremoreconsistentlesschaoticthanforthe1=D2metric.However,the1=D2metricdidresultinfewermessagesthanthe1=Dmetriconaverage,anditseemslikelythatthe1=D2metricwouldusesignicantlyfewermessagesthanthe1=Dmetricforevenlargerteamsofrobots.Thishypothesiscouldnotbetestedduetolimitationsonthesimulatorandavailablehardware:beyond53robots,thesimulatorwouldconsumemorethan4gigabytesofmemory,whichwasthemaximummemoryonanyavailablecomputer.4.2.2EffectsofCommunicationLossAnadditional180simulationstestedtheeffectsofrandommessagelossforthefourrecruitmentstrategies.Eachofthefourstrategieswastestedwith5%,10%,and25%ofthemessagesbetweenrobotsbeingrandomlydroppednottransmitted,butwithnonoticationtothesender;thesetestswererepeated3timesforeachof5startingcongurations335=180.TenidlerobotsplusthetwotaskedUGVsandtheUAVwereusedineachsimulation,foratotalof13robots.Inthesesimulations,thechoiceofwhatmessagestodropwasmaderandomlybythesimulator,andtheimpactofthatchoicevaried.Thus,eachofthesetestswasrepeatedthreetimesforeachofvedifferentstartingcongurationstocapturethetypicalperformanceofeachstrategy.4.2.2.1StatisticalAnalysisAsintheprevioussetofsimulations,itwasnotassumedthatthedataweredistributednormally,soaWilcoxonranksumtest[33][8]wasusedtotestthehypothesisthattheresultsforeachstrategyweredrawnfromdistributionswithequalmedians.Startingwithacondencelevelofp=0:01,afteraBonferronicorrection[24]for15samples,itisclaimedthatresultswithap-valuelessthan0:01=15=0:00066aresignicant.63

PAGE 73

4.2.2.2ResultsforNumberofMessagesMetricTable14showstheaveragenumberofmessagestransmittedbytherobotsforeachrecruitmentstrategyandcommunicationfailurerate.ThesameresultsareshowngraphicallyinFigure14.TheresultsoftheranksumhypothesistestsareprovidedinTable15.Theseresultsshowthattherecruitmentprotocolcontinuestofunctiondespitenetworklosses,andthattherelativeperformanceofeachoftherecruitmentstrategiesremainsconsistentastherateofmessagelossincreases.Asbefore,affectiverecruitmentrequiresthefewestmessagestobesent,onaverage,followedbygreedyandrandom.Bythetimethatlossesreach25%,theparticularrecruitmentstrategyuseddoesnotmakemuchdifference,becauseatthatpoint,thereisonlyan18%likelihoodthatthesixconsecutiverecruitmentmessagesrequiredbytheprotocolwillallbesentproperly.Nostatisticalsignicancecanbeclaimedforthedifferencesbetweenthestrategiesforteamsize13withcommunicationfailures.4.2.2.3ResultsforAverageWaitTimeMetricTable16showstheaveragetimethattheUAVspentwaitingforaUGVtorespondandarriveforeachrecruitmentstrategyandcommunicationfailurerate.ThesameresultsareshowngraphicallyinFigure15.TheresultsoftheranksumhypothesistestsareprovidedinTable17.Theseresultsindicatethatasbefore,greedyrecruitmentstillresultedintheleasttimespentwaitingbytheUAV,followedbyaffectiverecruitment.However,aparticularweaknessofgreedyrecruitmentisthattheUAVmustobtainthelocationsofalleligiblerobotsbeforeitcanchoosethenearestone.Assumingadecentralizedteam,thisrequiresanexplicitcommunicationfromallotherrobotstotheUAV,whichmaybeimpactedbynetworklosses.IfthenearestrobottotheUAVfailstoreceiveaHELPmessageortosendareply,thentheUAVmaycommittorecruitingadifferent,moredistantrobot,andbeforcedtoawaititsarrival.Ontheotherhand,usingaffectiverecruitment,theUAVwilltendtomakemultiplerequests,whichreducestherelianceonanysingleHELPmessage.Supposethataneligiblerobot,r1,isnearesttotheUAVandfailstosendareplyduetonetworkproblems.ProvidedthatnootherrobotshadsufcientSHAMEtorespondtothatrequestforinstance,iftheywerefarawayandaccruedSHAMEmoreslowly,theUAVwouldquicklyrequestagainandhaveanotherchancetorecruitr1.Inotherwords,requestingovertimecanndsolutionsthatevenoutperformgreedyrecruitment,ifthetimebetweenrequestsislessthantheadditionaltimeamoredistantrobotneedstoarrive.4.2.2.4SummaryofCommunicationLossSimulationsThesecondsetofsimulationstestedtheeffectofrandommessagelossoneachoftherecruitmentstrategies.Aswiththecasewithnomessageloss,greedyrequiredtheleastamountoftimetocompleterecruitment,andaffectiverequiredthefewestnumberofmessagestobetransmitted.Therelativeperformanceofeachstrategyremainedthesameasmessagelosses64

PAGE 74

Figure12.Totalwaittimeatdifferentteamsizes.Thesolidlineatthebottomisgreedy.Asexpected,affectivetakeslongertocompletethangreedy,butthedifferenceisapproximatelyconstant,andisaccountedforbytheadditionalmessagesaffectivemustsendbeforeanyrobothasenoughSHAMEtorespond.Withateamof50idlerobots,thedifferencebetweenaffectiveandgreedyisnotsignicant.Table14.Averagenumberofmessagestransmittedforeachrecruitmentstrategyaccordingtonetworklossrates.Theprobabilityofrandommessagelossisshownatthetopofthetable.Notethatthe0%columnisthetakenfromTable10. 0% 5% 10% 25% Affective 63.8 75.4 84.53 149.73 Affective,1=D2distancemetric 75.1 88.27 110.13 170.6 Greedy 75 92.07 104.27 201 Random 75 95.87 103.67 193.13 65

PAGE 75

Figure13.Boxplotsofthesimulationresultsforthewaittimemetricaccordingtoteamsize.Thelengthofeachboxisafunctionofthevarianceofthesamples,andthelineinsideeachboxdenotesthemeanover30trials.66

PAGE 76

Figure14.Messagestransmittedatdifferentnetworkfailurerates.Affectivesolidlinethatstartsbelowtheothersrequiredthefewestmessagesatupto10%losses,butat25%,allapproacheswereoverwhelmedbylosses.Table15.Pairwisecondenceintervalsforaveragenumberofmessagesforeachmessagelossrate.TheseintervalswereproducedwithaWilcoxonranksumhypothesistest.Atacondencelevelofp=0:00066,itisnotpossibletoruleoutthenullhypothesisthatthenumberofmessagesusedbyeachstrategycomefromdistributionswithequalmedians.Inotherwords,oncemessagelossoccurs,thedifferencebetweenthestrategiesdiminishesforthisteamsize. Affective,1=D2distancemetric Greedy Random Randommessagelossrate:5% Affective 0.0511 0.0072 0.0012 Affective,1=D2distancemetric 0.7869 0.3396 Greedy 0.3146 Randommessagelossrate:10% Affective 0.0026 0.0079 0.0044 Affective,1=D2distancemetric 0.3948 0.5062 Greedy 1.0 Randommessagelossrate:25% Affective 0.0620 0.0136 0.007 Affective,1=D2distancemetric 0.0929 0.2371 Greedy 0.5614 67

PAGE 77

Table16.Averagetime,inseconds,theUAVspentwaitingaccordingtorandommessagelossrate.EachsimulationconsistedofverecruitmentepisodesstartingwithaHELPrequestandendingwhenaUGVarrivedandbeganitstask.Thereportedvaluesrepresentthesumofthewaittimesforallverecruitmentepisodespersimulation.Themessagelossrateisshownacrossthetopofthetable.Notethatthe0%columnistakenfromTable16. 0% 5% 10% 25% Affective 210.3 284.29 318.95 645.83 Affective,1=D2distancemetric 221.7 279.83 356.64 680.65 Greedy 117.5 137.87 176.61 402.93 Random 304.6 400.73 447.41 1026.62 Figure15.Waittimesatdifferentmessagelossrates.Asbefore,affectivemiddlerequiresmoretimethangreedybottom,andthisrelationshipremainsconsistentastherateofrandomnetworklossincreases.68

PAGE 78

Table17.Pairwisecondenceintervalsforaveragewaittimeforeachmessagelossrate.TheseintervalswereproducedwiththeWilcoxonranksumhypothesistest.Atacondencelevelofp=0:00066,greedystillrequiressignicantlylesstimetocompleterecruitmentthanaffective.Resultsthatarenotconsideredstatisticallysignicantareshowninitalics. Affective,1=D2distancemetric Greedy Random Randommessagelossrate:5% Affective 0.901 9:0710)]TJ/F7 6.974 Tf 6.227 0 Td[(6 0.0042 Affective,1=D2distancemetric 1:1010)]TJ/F7 6.974 Tf 6.227 0 Td[(5 0.0021 Greedy 3:4010)]TJ/F7 6.974 Tf 6.227 0 Td[(6 Randommessagelossrate:10% Affective 0.431 1:8910)]TJ/F7 6.974 Tf 6.227 0 Td[(4 0.0055 Affective,1=D2distancemetric 1:3310)]TJ/F7 6.974 Tf 6.227 0 Td[(5 0.020 Greedy 1:3310)]TJ/F7 6.974 Tf 6.227 0 Td[(5 Randommessagelossrate:25% Affective 0.431 8:1310)]TJ/F7 6.974 Tf 6.227 0 Td[(5 4:2210)]TJ/F7 6.974 Tf 6.227 0 Td[(4 Affective,1=D2distancemetric 3:0710)]TJ/F7 6.974 Tf 6.227 0 Td[(4 0.0032 Greedy 4:1410)]TJ/F7 6.974 Tf 6.227 0 Td[(6 increasedupto25%,atwhichpointthestrategiesperformedsimilarlyintermsofthenumberofmessagestransmitted.Therewasnostatisticallysignicantdifferencebetweenanytwomethods,includingbetweenthelinearandnon-linearvariantsofaffectiverecruitment,asinSection4.2.1.Giventhateachofthestrategiesbuildsonanunderlyingcontractnetprotocol[90],itwasexpectedthattheirperformanceundermessagelosswouldbesimilar.AsinSection4.2.1,therewasnosignicantdifferencebetweenthelinearandnon-linearvariantsofaffectiverecruitmentundercommunicationloss.4.2.3BroadcastversusUnicastMessagingNinesimulationswereconductedtotesttheeffectofusingunicastsingle-sender,single-receivertransmissionsinsteadofbroadcastintherecruitmentprotocol.Inthesesimulations,insteadofsendingasingleHELPmessagetoallotherrobots,itwasassumedthattheUAVknewaboutalloftheotherrobotsintheteamandwouldattempttosendHELPmessagestothemindividually.Aswiththenetworkfailuretests,13robotswereused,ofwhich10wereuntasked.Anetworkfailurerateof10%wasused.Iftherehadbeennolossesinthistest,thenthenumberofmessageswouldhavetriviallybeenafunctionofteamsize.TheresultsofthesesimulationsareshowninTable18.Notethatduetotherandomnatureofthemessagelosses,eachstrategywastestedthreetimesandtheresultswereaveraged.69

PAGE 79

Table18.Averagenumberofmessagestransmittedaccordingtomessagingtype.NotethattheBroadcastcolumnisthesameasthe10%columninTable14. Broadcast Unicast Affective 67.3 379.3 Greedy 117.7 194 Random 99.3 183.3 Theseresultsshowthataffectiverecruitmentreliesonbroadcastmessagingtominimizethetotalnumberofmessages.SinceaffectiverecruitmentsendsmultipleHELPmessagesbeforeanotherrobothashighenoughSHAMEtorespond,theserequestsaremultipliedbythenumberofidleteammembers,whichgoeswellbeyondthenumberofmessagesrequiredbygreedyorrandomforwhichasingleHELPmessageissufcient.Further,unicastmessagingassumedthattheUAVcouldknowaboutalloftheotherrobotsintheteam,whichisnotnecessarilytrue,asdescribedinChapterOne.Thus,broadcastmessagingisrequiredforaffectiverecruitmenttobeeffective.4.2.4IllustrativeUseCasesAlthoughtheresultsaboveindicatethatgreedyrecruitmentwilltendtoproducetheshortestwaittimesfortheUAV,thisisnotuniversallytrue.Therearecasesinwhichtheaffectiverecruitmentstrategyresultsinshorterwaittimesthangreedy.SupposethattherearethreeUGVs,r1;r2;r3,suchthatr1isuntasked,andr2andr3performarasterscan.Letr2movetwounitsperiteration,whiler1andr3moveoneunit.Thus,r2andr3willnishtheirtasksatdifferenttimes.Next,supposethatattimestept0,theUAVsendsaHELPmessage,andtworobots,r1andr3areidle,andalthoughr2canreachtheUAVfasterthanr1orr3,itisontaskandcannotrespond.Inthegreedyandrandomstrategies,r1orr3wouldbechosenforrecruitmentimmediately,whereaswithaffectiverecruitment,noselectionwouldbemade,butalloftheUGVswouldincreasetheirlevelsofSHAME.Ifr2nishesitstaskattimestept1,itcanthenberecruitedbytheUAVandarrivesoonerthanr1orr3.Thisparticularcasewastestedinsimulation,anditwasfoundthatusingaffectiverecruitment,r2waschosenandarrivedafter65.4seconds.Thegreedyandrandomstrategiesselectedr1whicharrivedafter95.2secondsand95.4seconds,respectively.Anothersimpleusecasedemonstratesthatthroughaffectiverecruitment,theUAVwillchoosethenearestrobotwithoutrequiringthatotherrobotsrevealtheirlocationsaswiththegreedystrategy,whichmakesaffectiverecruitmentsuitableforstealthapplications.Supposethat,asabove,threerobotsr1;r2;r3areidleataparticulartimet1whentheUAVmakesarequest,buttheUAVisnearestr2.Aftereachignored70

PAGE 80

request,r2'sSHAMEwillincreasefasterthanthatofr1orr3,becauser2isclosesttotheUAV.Asaresult,r2willbethersttoexceeditsthresholdforSHAMEandwillrespondbeforer1orr3.Thus,theUAVrecruitstheclosestrobotwithoutrequiringallrobotstotransmittheirlocations.Thisbehaviorhasbeenveriedinsimulation.Inthisscenario,theUAVbroadcast5HELPmessages,atwhichpointr2responded,andtherecruitmentcompletednormally.Neitherr1orr3everbroadcastanymessages.Ifthiswerealow-powerorstealthapplication,r1andr3wouldhavebeensparedanunnecessarytransmissionbyusingaffectiverecruitmentratherthangreedy.4.2.5FairnessofRecruitmentAnaspectoftherecruitmentprocessthatwasnotaddressedintheabovesimulationsishowoftenaparticularrobotisrecruitedrelativetotherestoftheteam.Itisassumedthatarecruitedrobotmustexpendresourcestime,batterypowerinordertoperformataskonbehalfofanother.Ifarobotisrecruiteddisproportionatelyoften,thenitmayquicklyexhaustitsresourcesorbeunabletopursueitsowntasksakintoprocessstarvationinoperatingsystemscheduling.Assumethatinanidealrobotteam,eachrobotwouldberecruitedequallyoften,thusdistributingtheloadacrosstheentireteam.Letfairnessbeameasureofhowoftenarobotisrecruitedcomparedtotheotherrobotsintheteam,suchthatafairstrategyrecruitsallrobotsequally,andanunfairstrategyrecruitsasmallsubsetoftheteam.1Itwasexpectedthattheaffectivestrategywouldrecruitrobotsfairly,becausetheSHAMEderivedfromonerecruitmentmaypersistuntilthenextandcauserobotsthatarealmostequallywellsuitedtothetasktotaketurnsbeingrecruited.Ontheotherhand,thegreedyapproachmaytendtofavoronerobot,whichaftercompletingonerecruitmentmaystillbebestsuitedwhenthenextrequestarrives.Therandomapproachshouldbefairforlargenumbersofrecruitments,assumingauniformdistribution,butchoosesrobotsthatarenotinthevicinityoftherequest.Thus,eachstrategymaycontainabiastowardrecruitingaparticularrobot.Thedegreeofthisbiasisdenedasfollows.LetR=fr1;r2;:::;rkgbeateamofkrecruitablerobots.Letnbethenumberofrecruitmentrequeststhatoccur.=n=kistheexpectedmeannumberoftimesthateachrobotriwillberecruitedbyafairstrategy.Letibethenumberoftimesthatrobotriisactuallyrecruited.Letfi;=8><>:bi)]TJ/F11 9.963 Tf 9.962 0 Td[(cifi>0otherwise 1Notethatthetermsfairandunfairarenotintendedtobiasthereaderfororagainstaparticularstrategy.Aso-calledunfairstrategymaybeappropriateforagiventaskdomain;thegoalherewastoassignanintuitivelabeltoanaspectofrecruitmentperformance.71

PAGE 81

Table19.biasofeachrecruitmentstrategy.Tensimulationswereconductedforeachstrategy,andthebiasforeachrunisshownbelow. 1 2 3 4 5 6 7 8 9 10 Averagebias Affective 1 1 0 0 2 1 1 2 1 1 1.0 Affective,1=D2distancemetric 2 1 1 2 3 1 2 5 1 1 1.9 Greedy 2 6 4 5 5 7 5 6 5 6 5.1 Random 4 2 3 2 4 5 5 3 2 4 3.4 Giventheseterms,thebiasofastrategyisdenedasB=Pki=1fi;.Insimplerterms,biasisthenumberoftimesthatanyrobotisrecruitedmoreor,equivalently,lessthanaverage.Thefairnessofeachofthefourrecruitmentstrategieswastestedthrough40simulations.Inthesesimulations,veidlerobotswerearrangedrandomlyina100100gridforatotalof10startingcongurations.Ineachsimulation,25requestsweremadebyasimulatedUAVperformingarasterscanovertheareaasintheprevioussimulations.Themetricforthesesimulationswasthebiasdenedabove.Foreachrecruitmentstrategytestedaboveaffective,affectivewitha1=D2SHAMEupdatefunction,greedy,andrandom,tentrialswereperformed.Giventhattherewereverobotsand25recruitmentspertrial,perfectfairnesswouldbeachievedifeachrobotwasrecruitedexactlyvetimes.TheresultsfromthesesimulationsareshowninTable19,presentedintermsofthebiasofeachstrategy.Thesimulationsshowthataffectiverecruitmentwitha1=Ddistancemetrichadtheleastbias,averagingonerobotbeingrecruitedoncemorethanexpected.Usingthe1=D2distancemetricapproximatelydoubledthisbiasforaffectiverecruitment.Thegreedystrategy,ontheotherhand,hadanaveragebiasof5.1,wheresimplyignoringoneoftheidlerobotswouldhaveresultedinabiasof5.Randomfellinbetweenwithabiasof3.4.Notethatsincebiasisameasurementofvariance,itwasexpectedthatrandomwouldhaveabiasofzero,asthesamplesweredrawnfromauniformdistribution.However,the250totalrandomrobotrecruitmentsweretoofewtoproduceperfectlyuniformbehaviorfromtherandomnumbergenerator.Toverifytheuniformityoftherandomstrategy,onemillionpseudo-randomintegersintherange[0;4]wereproducedbythesamerandomnumbergeneratorthatwasusedforthesimulations.Eachintegerwaschosenwithin0.28%oftheexpected200,000times,sotherandomnumbergeneratorisadequatelyuniformforlargenumbers.72

PAGE 82

4.3RobotImplementationDuetoalackofavailablerobothardware,realtestswith5,10,20,and50idlerobotswerenotpossible,sothesecasesweretestedinsimulation.Onceithadbeenshownthataffectiverecruitmentworkedinsimulation,itwastestedwithoutmodicationonactualrobots.Thepurposeofthistestwastoverifythattheapproachworkedasexpectedonrealrobotsinadditiontosimulation.TheequipmentconsistedofthreeidenticaliRobotATRVJr.robotsshowninFigure2onpage4,andtheroleoftheUAVwassimulated.Thissectionbeginswiththerestrictedmine-detectionscenariousedforvalidationinSection4.3.1,followedbyadescriptionoftheimplementationinSection4.3.2.Section4.3.3discussestheactualrobottrials.4.3.1RestrictedScenarioThescenariofortestingontherealrobotswasasfollows.TwoATRVJr.robotsperformedrasterscansofanoutdoorarea,whilethethirdsatidleinaxedlocation.Atthetimeofthesetests,theUAVhadnotbeenequippedtoruntherecruitmentsoftware,soasimulatedagentthatrepresentedtheUAVwasusedinstead.Thisagentwaspositionedbythehumanoperatortocorrespondtothelocationofaminerepresentedinlatitudeandlongitude,atwhichpointthehumanoperatorwouldsignaltheagenttocallforhelp.TheidleUGVwouldthenberecruitedbythesimulatedUAVandnavigatetothemineusingaGPSreceivertotrackitsposition.Thisscenariowasrepeatedwithtwoidlerobotssoitcouldbeveriedthatthenearerrobotwouldbechosen.Statisticalsignicancewasnotexpectedforthisexperiment,asitwasonlyintendedtodemonstratethattheprotocolworkedonrealrobothardware.4.3.2SFXImplementationTheaffectiverecruitmentprotocolwasdevelopedaspartoftheSFXhybriddeliberative/reactiverobotarchitecture[65].AnoverviewoftheSFXisshowninFigure16,andacompletedescriptioncanbefoundin[66][65][67].EachrobotrunningSFXhasthreelayersshowninthelowerrightofFigure16:deliberative,managerial,andreactive.ThedeliberativelayercontainsaMissionPlannerthatformulateshigh-levelgoals,dividesthemintotasks,andimposesconstraintsonthemanageriallayer.ThemanageriallayerhasSensingandEffectormanagersthatcontrolresourceallocationfortherobot,andaTaskManagerthatgeneratesasetofreactivebehaviorstoperformthetasksspeciedbytheMissionPlanner.AffectiverecruitmentiscurrentlyconsideredtobepartoftheTaskManager,whichcanstartandstopreactivebehaviors;thefunctionality73

PAGE 83

Figure16.SimpliedoverviewoftheSFXarchitecture.Thetraditionalhybriddeliberative/reactivebaseofthearchitectureisshowntothelowerright;thisisinstantiatedoneachrobot.Theinterfacetotherobotthatisavailabletotherestoftherobotteamisshowntotheupperleft.GraphiccourtesyofMattLong.74

PAGE 84

couldalsobeplacedintheMissionPlannertopreventconictswithotherhigh-levelgoals.Thereactivelayercontainsreactivebehaviorsthatextractinformationfromsensorsthroughperceptualschemasandtransformthatinformationintomotionthroughmotorschemasthatcontroltherobot'seffectors.IntheJavaimplementationofSFX,entireclassesdataandmethodscanbedistributeddynamicallyamongtherobots.Onebenetofthisimplementationisthattherobotsneednotagreeaprioriaboutacommonnamespace;thatis,thetypesofsensors,percepts,andothercapabilitiescanbeenumeratedastheyareencountered.Thisprovidesanadditionaladvantageoverthesubject-basedmessagingsysteminMURDOCH,whereallcapabilitiesmustbeknowninadvance[30].4.3.3RobotTrialsAffectiverecruitmentwastestedontherealrobotsovertentrials:inveofthetrials,oneUGVwasidle,andintheremainingtrials,twoUGVswereidletotestthatthenearerrobotwaschosen.ThetrialswereconductedattheUniversityofSouthFloridainTampa,Florida,andatatesteldatNAVSEACoastalSystemsStationinPanamaCity,Florida.AtypicaltrialisshownfromtheperspectiveoftherobotoperatorinFigures17.InFigure17,theoperatorcausedtheagentrepresentingtheUAVtosendaHELPmessage,whichwasreceivedbytwoUGVs.OneoftheUGVswasinthemidstofataskandthusunabletoberecruited,soitmadenoreply.Theotherrobotwasidleandavailableforrecruitment,butithadinsufcientSHAMEanddidnotrespond.InFigure18,theUAVagainrequestedassistance,andtheidleUGVhadenoughSHAMEtorespondwithanACCEPTmessage.TherecruitedUGVthenmadeitswaytothelocationofthesimulatedUAVandannounceditsarrivalwithanARRIVEmessage,asshowninFigures19and20.4.4SummaryThischapterhaspresented833simulationsthatwereperformedtotesttheaffectiverecruitmentstrategyagainstothermethodsinaminedetectiontask.Thereweresixobjectivesfortheexperiments;thesearerestatedwiththeexperimentalresultsbelow.Testtheeffectsofvaryingteamsizeonthetimenecessarytocompleterecruitmentsandnumberoftransmissionsrequiredmetrics.Intherst600simulations,thecommunicationoverheadofaffectiverecruitmentwasshowntoscale35%betteroverallwithteamsizethanthegreedyapproachusedbyMURDOCHandarandomscheduler.Inparticular,forteamswith13ormorerobots,affective75

PAGE 85

Figure17.Operatoruserinterfaceforrealrobottests.ThesimulatedUAV,markedwitha0,sendsaHELPmessage.Robot1hasinsufcientSHAMEtorespond,andsilentlyignorestherequest.Robot2isontaskandalsosilentlyignorestherequest. Figure18.Operatoruserinterfaceforrealrobottests.ThesimulatedUAV,markedwitha0,sendsaHELPmessage.Robot1hassufcientSHAMEtorespondandsendsanACCEPTmessage.NotethatthemessageisshownasAGREEforthebenetoftheoperator.Robot2isstillontaskandsilentlyignorestherequest.76

PAGE 86

Figure19.Operatoruserinterfaceforrealrobottests.Robot1arrivesatthesimulatedUAV'spositionandsendsanARRIVALmessage.NotethatthemessageisshownasATGOALforthebenetoftheoperator. Figure20.UGVarrivingatasimulatedmine.77

PAGE 87

recruitmentusedsignicantlyfewermessagetransmissionsforrecruitment%betterwithp=3:310)]TJ/F7 6.974 Tf 6.227 0 Td[(11for13robots,35%betterwithp=1:210)]TJ/F7 6.974 Tf 6.227 0 Td[(12for23robots,and61%betterp=1:710)]TJ/F7 6.974 Tf 6.227 0 Td[(12for53robots,wherep<0:00033isconsideredsignicant.However,thisabilitytoscalecomesatacost:thegreedystrategycompletedin37%lesstimethanaffectiverecruitmentoverallbecausetheaffectiveapproachrequiredmultiplerequestsforhelpbeforearobotwouldhavesufcientSHAMEtorespondseeSection4.2.1.Testtheimpactofrandomcommunicationfailuresupto25%ontheperformanceofeachstrategy.Thesecondsetof180simulationsindicatedthattheaffectiveapproachworksinsituationswheretherearerandomcommunicationfailures,andthatitrequired22%fewermessagestobetransmittedthantheothermethodsoverall,upto25%messagelossatwhichpointtheaffectivestrategyused26%fewercommunicationsthangreedy,butthedifferencewasnotsignicantatp=0:0136.Asbefore,thegreedyapproachrequired43%lesstimetocompletethanaffectiverecruitmentsincemultiplerequestswererequiredbeforearobotwouldrespondseeSection4.2.2.TesttheeffectofalinearSHAMEupdatefunctionversusanon-linearfunctionwithregardstochaoticbehaviorforverylargeteams.Aspartoftherstsetofexperiments,theperformanceofthelinearandnon-linearSHAMEaccrualfunctionswerecompared.Nostatisticallysignicantdifferencebetweenthesemethodswasfoundforateamofsize53:thep-valueswere0.001and0.626forthehypothesesthatthelinearandnon-linearmethodswouldproduceresultssampledfromdistributionswithdifferentmeansforthenumberofmessagesandtotalwaittime,respectively,wherep<0:00033wasrequiredtodiscardthenullhypothesis.SeeSection4.2.1.Justifytheuseofbroadcastmessaginginsteadofunicastmessagingfortransmittingmessagesbetweentherobots.Thethirdsetofninesimulationsdemonstratedthatbroadcastmessagingwasjustiedoverunicast,especiallyforaffectiverecruitment,whichjumpedfromanaverageof67.3broadcastmessagesto379.3unicastmessages.SeeSection4.2.3.TestscenariosinwhichagreedyinstantaneousschedulersuchasMURDOCHrecruitspoorlywhiletheaffectiveapproachperformswell.Twocasesweretestedthroughfoursimulations.Intherstcase,affectiverecruitmentopportunisticallyrecruitedarobotthathadbeenpreviouslyunavailablebyusingmultipleHELPmessages.Therobotselectedwithaffectiverecruitmentarrivedin65.4seconds,whiletherobotsselectedbythegreedyandrandomstrategiesrequired95.2and95.4seconds,respectively.Inasecondcase,inateamwithtwoidlerobots,therobotwiththebesttnesswas78

PAGE 88

successfullyrecruitedwithoutrequiringtheotherrobottoevertransmitanymessage,whichmakesaffectiverecruitmentmoreappropriateforstealthorlow-powerdomainsthanthegreedyapproach.SeeSection4.2.4.Testthedegreetowhichallrobotsarerecruitedequallyoften.Thenalsetof40simulationsexploredthedegreetowhichallrobotsintheteamarecalleduponequallytoassistwhenusingeachrecruitmentstrategy.Thesetestsshowthataffectiverecruitmenttendstouseallrobotsequallyoverallbiasof1.0,wherezeroindicatesaperfectlyevendistributionofrecruitmentsacrossallrobots,whereasgreedyshowedmoreinclinationtousingasubsetoftherobotsoverallbiasof5.1.SeeSection4.2.5.Tentrialsonateamofmobilerobotswereconductedtovalidatetheapproachonrealhardware.AteamofthreeATRVJr.robotsandasimulatedUAVconductedamockmine-detectiontaskinwhichtheUGVswererecruitedbytheUAV.Theaffectiverecruitmentstrategyperformedasexpectedontherealrobothardware.Thenextchapterdiscussestheseresultsinmoregeneralterms,discussingthelimitationsoftheexperiments,howtheexperimentscomparetotheexistingresultsintheliterature,howtheparametersforSHAMEcouldbemodied,andthecontributionsofthisthesis.79

PAGE 89

ChapterFiveDiscussionThisworkcontributesanaffectiveemotion-basedrecruitmentstrategyfordistributedteamsofheterogeneousmobilerobots.ExperimentalresultsinChapterFourshowedthattheapproachisasrobustasthegreedyinstantaneousschedulerconsideredtobethestateoftheart[30]andimprovesthecommunicationload.Affectiverecruitmentrequired32%fewermessagestobetransmittedduringtherecruitmentprocessoverall,and61%fewerforateamof53robotswithoutastatisticallysignicantincreaseintheamountoftimerequiredtocompleterecruitment.ThischapterprovidesadiscussionoftheexperimentalresultsintermsoflimitationsoftheexperimentsinSection5.1,acomparisontoexistingresultsinSection5.2,areviewoftheparametersaffectingSHAMEinSection5.3,andcontributionsofthethesisinSection5.4.5.1LimitationsofExperimentsTheexperimentsinChapterFourconsistedof833simulationsaswellasimplementationandtestingonateamofthreeATRVJr.UGVsandanaerialvehicle.However,therewerelimitationstotheseexperimentsthatmeritdiscussion.Therstlimitationwasthemaximumnumberofrobotsthatcouldbesimulatedatonetime.ThesimulationenvironmentdescribedinSection4.1.1wasdevelopedaspartofadistributedarchitecture,suchthateveryadditionalrobotwasrepresentedbyacollectionofJavaobjectsandapproximately25threads.Asthesizeoftherobotteamgrewlarger,thenumberofthreadsandthememoryrequirementsbegantoexceedtheresourcesoftheavailablecomputerhardware.Thoughitwouldhavebeenillustrativetoperformrecruitmentonteamsof100,200,ormorerobots,thiswassimplynotpossible.However,trendsinthedatashowngraphicallyinChapterFour,especiallyFigures10and12arealreadyapparentforteamsof53robotsandcouldbeextrapolatedforlargerteams.Thenextconstraintwasthatrunningtestsontheteamofrealrobotsrequiredseveralpeopletoactasrobothandlers,favorableweatherconditions,andlogisticstomoverobotsandsupportequipmenttotheeld.Robotfailuresalsocomplicatedtestsontherobotteam.Asaresult,itwasnotpracticaltoperformenoughrealtrialstoduplicatestatisticallysignicantdatacollectionwiththerobots.Therobot80

PAGE 90

implementationandtestsserveasproofthattheconceptcanbeextendedfromsimulationtoactualrobots,buttherobotteststhemselveswereintendedonlyforillustration,andnotasasourceofadditionaldata.Athirdlimitationwasthatalthoughthetestsonrealrobotsindicatedthattheapproachworksasexpectedwithheterogeneousteams,thisheterogeneitywasnottestedinsimulation.ThejusticationformakingallidlerobotsequallycapableofbeingrecruitedthoughdifferentiatedbytheirSHAMEandtnesstothetaskisasfollows.SupposethateachUGVcouldhavecapabilityA,B,orCi.e.theyareheterogeneousandthatthesimulatedUAVrequestedcapabilityAatitsrststop.Thiswouldhavetheeffectofpartitioningtheteamintotwosubgroups:thosethatcouldprovideA,andthosethatcouldnot.ThesubgroupthathadcapabilityAwouldperformrecruitmentasexpected,andtherestwouldbeignored,effectivelyreducingthesizeoftheteam.Giventhattheupperboundoftheteamsizewasalreadyconstrainedbelowthedesired100ormorerobotsbyhardwarelimitationsasabove,thispartitioningwasundesirable.Finally,themannerinwhichmessagesweredroppedinthesimulationstestingcommunicationsfailureswassimplistic.Eachtimeatransmissionwastooccur,arandomnumberrwaschosen,where0r1.Ifrwasgreaterthanthecommunicationfailurerate,thenthemessagewouldbesent;otherwise,itwouldbedropped.However,realwirelessnetworkstendtosufferfromburstlosses,ratherthanindividualpacketlosses.Markovmodelsthatcanmoreaccuratelydescribethereallossbehaviorofwirelessnetworkshavebeendevelopedsee[49]asanexamplebutwerebeyondthescopeofthisthesis.Theeffectofusingasimplermodelisnotexpectedtobesignicantbecausetherecruitmentprotocolusedsinglemessagesthatwerespreadoutovertimesoanyparticularburstofinterferenceislikelytoeliminate,atmost,onemessageatatime.Theaffectiverecruitmentapproachhasalsobeendemonstratedonrealrobotsusingrealwirelesshardware.Itisalsoworthnotingthattherobotsmadenoattempttoretransmitthemessagesthatweredropped,whichmighthavereducedtheimpactofthesecommunicationlosses.However,sinceallrecruitmentstrategiesweretestedunderthesameconstraints,theresultsarefair.5.2ComparisontoExistingResultsThissectionwillcomparethescopeoftheseexperimentstorelatedresultsfoundintheliterature.Thepurposeofthiscomparisonistojustifytheextentoftheexperimentswithrespecttotherestofthecommunity.SeeChapterTwoandespeciallyTable1foramorecompletesummaryofexistingresults.TheexperimentalmethodologyusedinthisthesisismostsimilartothoseusedinALLIANCE[75]andMURDOCH[30]intermsofthenumberofrobotsusedandexplicittestsoffailurerecovery.A81

PAGE 91

discussionofrobotteamsizeswillbeprovidedrst,followedbyresultsintheliteraturethatincludefailuredata.ForALLIANCE[75],therobotexperimentsemployedthreemobilerobotsforaforagingtask.Twooftherobotsactivelygatheredspillobjectsandplacedtheminadesignatedlocation,andthethirdrobotremainedstationaryandperformedareport-progresstask.SimilarlyinMURDOCH[30],threerobotsperformedthepusher-watcherbox-pushingtask,wheretworobotspushedeitherendofaboxwhilethethirdremainedstationaryandmonitoredtheprogressoftheboxrelativetoagoal.Inbothcases,threerobotswereusedinanindoorenvironment,suchthatonlytwoweremobile.Inthereal-worldtestsperformedforthisthesis,threemobilerobotswereactiveinanoutdoorenvironment,whichisonparforsimilarresearch.Thelargestrobotteaminrelatedexperimentsfoundintheliteratureappearstobeve:in[77],areferenceismadetoasmanyasverealrobotsusedwithALLIANCE,butitisnotclearhowmanyrobotswereactiveorinwhatenvironmentthetestswereperformed.In[105],veActivMediaPioneerII-DXrobotscooperativelyexploredanindoorenvironment.Therefore,thescopeoftheexperimentsisconsistentwiththeroboticscommunity.Intermsofsimulationresultsformulti-robottaskallocationintheliterature,thelargestrobotteamsappeartohavebetween20and55robots.In[16],20holonomicrobotsweresimulatedforacollaborativetransporttaskinwhichtherobotswouldworktogethertomove30largeobjects.In[41],20simulatedrobotsperformedasimpleforagingtaskwith50pucks.In[48],55antsweresimulatedinademiningtaskjudgingfromagraphoftheirresults.Inthisthesis,themaximumsimulatedrobotteamsizewas53,againconsistentwiththeroboticscommunity.TheexperimentsinChapterFouralsotestedtheeffectofcommunicationfailuresrangingfrom5%to25%ontheoverallrecruitmentprocessandexceedsthetypicalamountoffailuresforthecommunity.Partialfailuresinadistributedteam,includinglossofcommunicationorrobots,isalsoincorporatedintoexperimentsintheliterature.ForALLIANCE[75],onerobotwasremovedfromtheteamandanothertookitsplace.ForMURDOCH[30],onerobotwasremovedandtheotheralternatedbetweenpushingtheendsofabox.Partialfailuresarealsoexplicitlytestedin[4]whereonerobotfailurewassimulated,in[105]whererobotsweresporadicallydisabled,andin[94]wherethecommunicationchannelwassharedwithadversariesthatactivelyinterferedwithcommunicationsorre-injectedmessagestocauseconfusion.Forsensornetworks,[36]simulatedtheeffectsofupto20%nodefailuresand[38]simulatedthelossof23outof60nodes.Therearemanyrelatedresultsinwhichnopartialfailuresweretested[10][57][101][48][74][42][41][45][88][87][47].Otherapproachesassumedreliablecommunications.Inparticular,MOVER[40]82

PAGE 92

assumedareliablecommunicationschannel,andwouldhaltforaslongastherewerelosses.In[16],reliablecommunicationswereexplicitlyassumed.InLEMMING[70],communicationswerereducedthroughstateinformation,butifmessageswerelostorrobotsstoppedresponding,thenthesystemwouldfailasitassumesthatarobotwillexplicitlydeclineataskthatitcannotfulll.5.3ParametersandFitnessMetricsTheperformanceoftheaffectiverecruitmentapproachhingesonthedesignofthreefunctions:thetnessfunctionformatchingrobotstotasks,theSHAMEaccrualfunction,andtheSHAMEdecayfunction.Therolesofthesefunctionsandtheirimpactontheexperimentsareexploredbelow.5.3.1FitnessFunctionAsdiscussedinChapter3.3,theremustbesomemechanismforchoosingthebestrobotforatask.Otherapproacheshavedenedmetricsformeasuringtheutilityofaresourcerobotorsensortoatask[51][104][103][26]basedonthecostoruncertaintyofusingthatresource.IntheMURDOCHsystem[30],tnessiscomputedintermsofdirectlymeasurablequantities,suchasthedistancebetweenarobotandthetasklocation.IntheexperimentsinChapterFour,tnesswasbasedontheestimatedtimerequiredforarobottoreachthetasklocation.Thisisasimplemeasure,atbestarst-orderapproximationofeachrobot'stness.However,thepurposeoftheexperimentswastotesttheaffectiverecruitmentapproachgivenapracticalandextensiblemeasureofutility.Forthis,thetimetoarrivewassufcient.Theremaybeadvantagestousingmoresophisticatedtnessfunctionsonrealrobots.Ingeneral,consideringseveralrobotattributestogethershouldleadtoamoreinformedmeasureoftness,aswasdiscussedinChapter3.3.Thisapproachiscompatiblewithanysuchtnessmeasure.5.3.2SHAMEAccrualFunctionOnceatnessfunctionischosenasdiscussedinSection5.3.1,thatfunctionmustbeappliedtotheaccrualofSHAMEfortherobots.SHAMEaccrualwasintroducedinChapter3.2,andtwofunctionswereusedintheexperimentsinChapter4.2:onelinearfunction,andoneverysimplenon-linearfunction.TheresultsinChapter4.2didnotshowanystatisticallysignicantdifferencebetweenthelinearandnon-linearfunctions,butthesemeritfurtherdiscussion.83

PAGE 93

ThelinearSHAMEaccrualfunctionwaschosenforitssimplicity,ashasbeendoneintheliterature.ALLIANCE[75]alsouseslinearfunctionsforitsimpatienceandacquiescence.Pfeifer'sFEELERsystemisdescribedasincreasingemotionalarousalthroughalinearfunction[71].Theemotionalmotivationalsoincreaseslinearlyin[3][98][99][6].Despiteitscommonusage,itwassuspectedthatalinearSHAMEaccrualfunctioncouldleadtoachaoticteambehavior,inwhichtherobotsthatrespondedtorequestswouldnotbeinthevicinityoftherequester,butwouldinsteadrespondasaresultofresidualmotivationfromapreviousrecruitmentepisode.InthecasewhereallrobotsstartwithzeroSHAMEwhenrecruitmentbegins,themosttrobotswillbethersttorespond.However,astheaverageSHAMEacrosstheteamincreases,thenumberofrespondingrobotswillincrease,anditbecomesincreasinglydifculttopredictwhichrobotwillactuallybechosen.Thenon-linearSHAMEaccrualfunctionwastestedasameansofkeepingtheaveragelevelofrobotSHAMEreducedwhilestillallowingthemosttrobotstoquicklyrespond.Itwasexpectedthatforlargeteamsofrobots,thecommunicationsoverheadrequiredusingthenon-linearfunctionwouldbelessthanforthelinearfunctionasrobotswouldhavelessaverageSHAMEandfewerwouldrespond,andthetimetocompleterecruitmentswouldbelessforthenon-linearfunctionbecausethechoiceofrobotswouldbebetter.However,forteamsupto53robots,thisdidnotappeartohappen.ThedistributionofresultsshowninFigures11and13inChapter4.2indicatesthatthelinearaccrualfunctionperformedmoreconsistentlythanthenon-linearfunction,andthatthesechaoticeffectswereabsent.5.3.3SHAMEDecayFunctionTherateatwhichSHAMEdecayswaschosentobealinearfunctionfortheexperimentsinChapter4.2.Ortony[71]andPfeifer[79]suggestthatemotionsshoulddecayovertime,butprovidenofurtherguidance.LinearemotionaldecayfunctionswereusedbyArkinetal[3]forhomeostaticcontrolinaSonyAIBO,andfornaturalisticinterfacesbyBreazealforKISMET[6]andbyVelasquezforSimon[98]andYuppy[99]undertheCathexisarchitecture.However,thereisnohardrequirementthattheSHAMEdecayfunctionbelinear,oreventhatitbedecreasing.AnyfunctioncouldhavebeenusedforSHAMEdecay,includinganaccrualfunctionthatcausedrobotstoincreasetheirSHAMEovertimeintheabsenceofrequests.Thechoiceofadecayfunctionwilldependonthetaskdomain,theexpectedperiodoftimebetweentasksi.e.tocontrolwhethereachrobot'sSHAMEwillpersistordecaytozerobetweenrequests.84

PAGE 94

5.4ContributionsThisapproachandvalidatingexperimentsmakeatleastsixcontributionstothearticialintelligence,robotics,andcognitivesciencecommunities.Thesecontributionsareexploredbelow.5.4.1ValidatesApplicationofEmotionsTheaffectiveapproachbuildsonamodelofemotions,makingitamixofthemotivation-basedandauction-basedstrategies.Unlikeothermotivation-basedapproachestorobottaskselection[75][47],however,affectiverecruitmentwasdevelopedfromtheliteratureinthetheoryofemotions.Thisistherstknownapplicationofanemotionalmodeltotaskallocationinadistributedteamofrobots.ItprovidesthecognitivesciencecommunitywithvalidationoftheOCCmodel[72]inateamofarticialagents,demonstratingthattheemotionsfunctionasexpected.Theemotionalstatecanprovidemeaningfulinformationtoahumansupervisoroftherobotteam.Inasituationwhererobotsarenotallowedtopreempttasks,theemotionalstatecanbeusedbyahumantomakeamanagerialdecisionaboutwhatrobotshouldbepreemptedtomakeitavailableforrecruitment.Theemotionsarealsoaneffectivetoolfordevelopingcomplexbehaviorfromreactiverobots,becausetheyprovideacomputationallysimplemethodforusingarobot'srecenthistorytobiasitscurrentactions.5.4.2ReducedCommunicationOverheadandBetterScalingAffectiverecruitmentreducesthecommunicationoverheadformulti-robottaskallocation.ThebestknownmethodsforrecruitmenttodatehaveacommunicationcomplexityofOn[31].Althoughtheworst-casecomplexityforaffectiverecruitmentisalsoOn,simulationresultsinChapter4.2indicateastatisticallysignicantreductionincommunicationsoverheadby32%onaverage.Thisreductionispossiblebecauserobotsdonotreplytoeveryrequest.Thisreductionofcommunicationoverheadprovidesthedistributedsensingandroboticscommunitieswiththefollowingfourbenets.First,theapproachwillscalebetterwithteamsizecomparedtoothermethods,becausetheimpactonthesharedcommunicationchannelsislessasshowninChapterFour.Thismakesaffectiverecruitmentsuitableforverylargeteamsandswarmsofrobots,aswellasforsensornetworks.Second,anyreductionincommunicationsforrecruitmentwillfreeupbandwidthforotherdemandsthatmaybevitaltotheteam'smission,suchasstreamingvideo.Third,fewertransmissionstranslatestoalowerenergycostforlow-powerdevices,whichmaynothavetheresourcestobroadcastabid85

PAGE 95

foreveryrecruitment.Thisbenetssensornetworks,whichmayhavelimitedcommunicationpower[43].Finally,affectiverecruitmentissuitableforstealthapplications,whereagentsonlytransmitmessageswhensufcientlymotivated,ratherthaninresponsetoeverynewtask.Thisbenetsthosedoingresearchinphysicalsecurityandmilitaryreconnaissance.5.4.3SuperiorSolutionQualityTheaffectiverecruitmentapproachallowsforateamofrobotstobalancethetimeneededtorespondtoarequestwiththecommunicationoverheadbymeansofadjustingparameters.Byallowingtherecruitmentprocesstotakeplaceovertime,theaffectiveapproachcanndbettersolutionsthaninstantaneousschedulerssuchasMURDOCH[30]whoseperformancedependsontheorderinwhichtasksarrive[28].TheaffectiveapproachisaexiblegeneralizationofthegreedyapproachseeninMURDOCH,andapartfromthetrade-offbetweenresponsetimeandmessagessent,affectiverecruitmentwillperformnoworsethanMURDOCH.ExperimentalresultsinChapterFourindicatethatforteamsof53robots,affectiverecruitmentrequired61%fewermessagesthanthegreedyapproachusedbyMURDOCH,butwithoutastatisticallysignicantincreaseinthetimeneededtoperformrecruitment.Thisimprovementongreedyallocation,whichrepresentsthestateoftheart,benetsthedistributedagentscommunity.5.4.4DemonstratedRobustnessTheContractNetProtocol[90][21]throughwhichrecruitmentisnegotiatedprovidesadegreeofrobustnessthatallowsrecruitmenttosucceed,evenatrandommessagelossratesof25%.ExperimentalresultsinChapter4.2.2indicatethataffectiverecruitmentperformscorrectlyinthepresenceofthesecommunicationfailures.Theseresultsprovidethedistributedsensingcommunitywithabasisforcomparingtheperformanceofothercommunicationprotocolsinthepresenceofhighlosses.5.4.5HandlesHeterogeneityTherewerenoassumptionsinthisthesisaboutthecompositionoftherobotteamintermsofhardware,software,ortasks.ThisapproachappliestoanyteaminwhichametrictnessfunctionisavailabletodeterminethesuitabilityofeachrobottoataskasdiscussedinChapter3.3.AlthoughthesimulationswereallperformedwithhomogeneousagentsasdiscussedinSection5.1,realrobottestsincludedthreeUGVsandasimulatedUAVthataredifferentinhardwareandsoftware.Theroboticscommunity,whichmustdeal86

PAGE 96

withchangesinrobotcapabilitiesovertimeduetodamageortheadditionofnewrobotstoanexistingteam,benetsfromthislackofconstraints.5.4.6FairnessofAllocationTheaffectiveapproachalsotendstodistributetheburdenofrespondingtorequestsforhelpacrossmoreoftherobotteam.GiventhatarobotwillonlyrespondwhenithassufcientSHAME,asinglerobotislesslikelytoberecruitedforsuccessivetasksinthesamelocationifthereisanotherrobotthatisalmostequallywell-suitedtothetask.Instead,therobotswilltaketurnsbeingrecruited,asaside-effectofretainingSHAMEfromonerequestforhelptothenext.ResultsinChapter4.2.5indicatethataffectiverecruitmenttendstorecruitmoreuniformlyfromtherobotteam.Deningthebiasofarecruitmentstrategyasthenumberoftimesthatarobotischosenmoreorlessthananyother,onaverage,affectiverecruitmenthadabiasof1.0.9withanon-lineartnessmetricwherethegreedyapproachhadabiasof5.1.5.5SummaryThischapterhasdiscussedtheresultsoftheexperimentsinChapterFour,startingwiththelimitationsofthoseexperiments.Whiletheexperimentswereequaltoorexceededstandardtestingpracticesfortheroboticscommunity,itwasnotpossibletoperformsimulationswithrobotteamsaslargeasdesired:theupperlimitforthesimulatorontheavailablecomputerhardwarewaslessthan100robots.Statisticallysignicantperformancedatawerenotcollectedfortherealrobottestsduetothemanpowerrequiredtorunmultiplerobotsoutdoors,thoughtherobotswereusedtovalidatetheapproach.Heterogeneousrobotswerenotsimulated,asthiswouldhavehadtheeffectofpartitioningtherobotteamintosmallersubgroups,whichwouldhaveoffsettheeffectofincreasingtheteamsizeandtrivializedtherecruitmentproblem.Finally,themethodforchoosingwhichmessagestodropwhilesimulatingcommunicationfailureswassimplisticcomparedtoMarkovmodelsthathavebeenproducedbythenetworkscommunity[49],buttheeffectonthesimulationresultswasexpectedtobeinsignicant.Robotsdidnotattempttoretransmitmessagesthatweredropped,whichwouldhaveimprovedtheirperformance,butthislimitationaffectedallrecruitmentstrategiesequally.Comparedtorelatedresearchinmulti-robottaskallocation,theexperimentsinChapterFourareofasimilarscopetothoseperformedelsewhere.Inparticular,theexperimentsforthisthesisinvolved53simulatedrobots,3realrobots,andpartialfailuresintheformofcommunicationlosses.Intheliterature,the87

PAGE 97

largestsimulatedteamshavebetween20and55robots[16][41][48],thelargestrealteamshave5robots[77][105],andfewerthanhalfexplicitlytestforpartialfailures.Asurveyoftheexistingliteratureindicatesthataffectiverecruitmentisanovelcreation.Experimentsshowanimprovementoverknownmethods,includingothervariantsofthecontractnetprotocol[90][21],suchasMURDOCH[30],whichistheclosestknownwork.Affectiverecruitmentprovidesatleastsixcontributionstothearticialintelligence,robotics,andcognitivesciencecommunities:Validatesapplicationofemotions.ThisworksupportstheOCCmodelofemotions[73][72]anddemonstratesthatthemodelfunctionsasexpectedinarticialagents.Reducedcommunicationoverheadandbetterscaling.ExperimentsinChapter4.2showthataffectiverecruitmentreducesthecommunicationsoverheadforrecruitmentcomparedtothestateoftheart[30].Superiorsolutionquality.Affectiverecruitmentcanreachsolutionsthatagreedyinstantaneousschedulerwouldmiss.Varyingtheparametersinthisapproachprovidesexibilityinthedesignofarobotteam.ExperimentalresultsinChapter4.2indicatethatforteamsof53robots,affectiverecruitmentrequired61%fewermessagesthanthegreedyapproachwithoutastatisticallysignicantincreaseinthetimeneededtoperformrecruitment.Demonstratedrobustness.ExperimentsinChapter4.2haveshownthataffectiverecruitmentcontinuestofunctionwithupto25%randommessagelosses.Handlesheterogeneity.Noassumptionsweremadeinthisapproachthatwouldpreventheterogeneousrobotsfromworkingtogetheronatask.Fairnessofallocation.Ingeneral,ifmultiplerobotsareapproximatelyequalintheirtnesstoatask,affectiverecruitmentwilltendtoutilizeeachofthemanequalnumberoftimes.Thefollowingchaptersummarizesthethesisandprovidesdirectionsforfuturework.88

PAGE 98

ChapterSixSummaryandFutureWorkThemulti-robottaskallocationMRTAproblemdealswiththeassignmentofnewtaskstorobotsorotheragents.Solutionstothisproblemexistintheliterature,andthecurrentstateoftheartisMURDOCH,aContractNetProtocol[90][21]withagreedyinstantaneousscheduler[30].However,existingtechniqueshaveacommunicationsoverheadthatincreasesatarateofatleastOnforateamofnrobots[28][31]ortendtobroadcaststateoreligibilityinformationatahighrate[75][74][92][39]Hzin[74],15Hzin[92].ThisthesishaspresentedanovelapproachtoMRTAthatexperimentalresultsshowrequiressignicantlyfewermessagestobetransmittedanaverageof32%andupto61%fewerthanthegreedyschedulerusedinMURDOCH[30].Inthisapproach,anaffectivevariable,SHAME,modulatesarobot'sresponsetoarequestforhelp.SHAMEhastheeffectofcausingonlythebest-suitedrobotstoreplytoarequest,thusreducingoverheadintherecruitmentprocess.Thisapproachisalsorobustwithrespecttocommunicationlosses;experimentalresultsshowthatrecruitmentsucceedsatupto25%randommessageloss.Theapproachhasbeenvalidatedin833simulationsandhasbeenimplementedandtestedonateamofATRVJr.mobilerobots.Thischapterexamineshowaffectiverecruitmentaddressedtherecruitmentproblemandpresentsdirectionsforfurtherresearch.6.1SummaryofThesisTheresearchquestionstatedinChapterOnewas:Howcanaffectivecomputingbeusedforrecruitmentinateamofdistributed,heterogeneousmobilerobotswithunreliablecommunications?AffectivecomputingcanbeappliedtotheContractNetProtocol[90][21]suchthatagentsrespondtobidrequestsonlywhentheyhaveadequatemotivation.Therateofincreaseoftheagent'smotivationisrelatedtoitstnesstothetask,suchthatthebest-suitedcandidatesrespondrst.Inthisway,theproblemofrecruitingarobotiscompletelydistributedamongtheteam,allowingittoscaleslowlywithrespecttoteamsize.Theheterogeneityoftherobotteamisaccommodatedthroughthetnessfunctionthatensuresthat89

PAGE 99

onlythemostappropriateagentswillrespond.TheContractNetProtocolitselfprovidesamechanismforovercomingcommunicationfailures,providedthattherobotsdonotwaitindenitelyforanincomingmessage.Experimentalresultshaveshownthatthisapproachrequireslesscommunicationsoverheadthanthestateoftheart,asinMURDOCH[30],aninstantaneousgreedyscheduler.Experimentsonrealrobotsshowthatthisapproachcanbeappliedtoreal-worldsystems.Thefollowingissuesrelatingtothedomainofmulti-robottaskallocationwereintroducedinChapterOne.Themannerinwhichtheapproachtakeninthisthesisaddressestheseissuesappearsbelow.Communicationsbandwidthisnite.Othermission-relateddemandsonasharedcommunicationschannel,suchasstreamingvideoorcontrolcommands,mayconsumeanyavailablebandwidth.Recruitmentshouldnotinterferewithsuchdemands.ExperimentalresultsinChapterFourshowthataffectiverecruitmenthas32%lessdemandforcommunicationsbandwidththanthegreedyapproachfoundinotherapproachesmostnotably,MURDOCH[30].Thekeytothisreductionincommunicationsoverheadistheuseofanaffectivevariable,SHAME,thatpreventsrobotsfromrespondingunnecessarilytorequestsforhelp.Teamscanvaryinsizewithoutbound.Thecommunicationsrequirementsforrecruitmentshouldscalewellwithteamsize,sothatlargeincreasesinthenumberofrobotsdoesnottranslatetoalargeincreaseinrequiredbandwidth.ExperimentalresultsinChapterFourshowthataffectiverecruitmentscalesslowlyOnworstcase,lessonaveragewiththesizeoftherobotteam.Forthelargestsimulatedteamof53robots,affectiverecruitmentrequired61%fewermessagesthanthegreedyapproach.Increasingtheteamsizefrom4to53robotsincreasedthenumberofmessagesthataffectiverecruitmentsentfrom52.6to109.1,wherethesamechangeincreasedgreedy'soverheadfrom30to276.7messages.Robotsmayhavepowerconstraints.Asmallmobilerobotoranodeinasensornetworkmaynothaveenoughpoweravailabletofrequentlytransmitmessages.Thefewermessagessucharobotsends,thelongeritsbatterieswilllast.Asabove,theexperimentalresultsinChapterFourshowthataffectiverecruitmentreducesthenumberofmessagesthateachrobotmustsend,whichtranslatestoasavingsinpower.Therobotsmaybedeployedinadomainrequiringstealthwheretheyarecapableofreceivingmessages,butriskrevealingtheirlocationbyresponding.Inthiscase,anytransmissionsshouldbewelljustied.ExperimentalresultsinSection4.2.4showthatrobotsthatareunlikelytoberecruited90

PAGE 100

duetoasmallincreaseintheirSHAMEperrequestcanremaincompletelysilent.Recruitmentcantakeplacewithasfewastworobotstransmittingmessagesforanysizeofrobotteam.Theremustbeamechanismtodeterminethetnessofarobottoatask.Thisapproachusedasinglemetrictodeterminethetnessofarobottoatask:theestimatedtimeofarrivalatthetargetlocation.Thismetricwasappliedbothasalinearandanon-lineartermwhencomputingthetnessofarobot.Itwasexpectedthatforlargerobotteams,thenon-linearmethodwouldresultinfewermessagestransmittedthanthelinearmethod.However,forthelargestrobotteamtestedrobots,thedifferencewasnotstatisticallysignicant.Thetimetoarrivewasanextensible,rst-orderapproximationoftherobot'stness,butmoresophisticatedtnessfunctionscouldbeused.ThedifcultyincreatingageneraltnessfunctionwasexplainedinChapter3.3andisdiscussedinSection6.2asanopentopicofresearch.6.1.1ContributionsThisthesisprovidesthefollowingcontributionstothearticialintelligence,robotics,andcognitivesciencecommunities.Thesecontributionsareoutlinedbelow.Validatesapplicationofemotions:Thisapproachaddsmotivationstoanauction-basedmulti-robotrecruitmentstrategythroughanemotionalmodel.ThisthesissupportstheemotionalmodelofOrtonyetal.[73][71][72],andbenetsthecognitivesciencecommunitybydemonstratingthattheemotionsfunctionasexpectedinarticialagents.Emotionsareausefultoolfordevelopingintelligentagents,astheyprovideacomputationallysimplemeansofself-regulation.Anagentequippedwithemotionscandetectalackofprogresstowardagoal,orcombinemultipleinternaldrivesintoacoherentchoiceofaction.Emotionscanalsoprovidenaturalisticinterfaces:agentscanuseemotionstoexpressmeaningfulstateinformationtoahumansupervisor.Forexample,ahumanoperatorcanusetheemotionalstateoftherobotstodeterminewhichrobotshouldbepreemptedforanewtask,forinstance,ifnorobotsareavailableforrecruitment.Agentscanalsoperceivetheemotionalstateofthehumanoperatorandaltertheirbehavioraccordingly.Reducedcommunicationoverheadandbetterscaling:Theuseofanemotionalmodelreducesthecommunicationsrequiredfortaskallocationcomparedtothestateoftheart,andrepresentsanimprovementoverthegreedyapproach.Thisreductionofoverheadcontributestwobenetstothedistributedsensingandroboticscommunities.First,theprotocolcanscaletolargeteamsorswarmsof91

PAGE 101

robotsmorereadilythanexistingmethodsasshownbyresultsinChapter4.2.Second,thisapproachreducesunnecessarytransmissions,whichbenetslow-powerandstealthapplications,andkeepsbandwidthavailableformission-specicdata,suchasstreamingvideo.Superiorsolutionquality:Theaffectiverecruitmentapproachbenetsthedistributedagentscommunity,asitcanreachbettersolutionsthanexistinggreedyrst-priceauctionstrategiessuchasMURDOCH[30].Suchgreedyschedulerscanbeadverselyaffectedbychangingtheorderofnewtasks,whichmayleadtogreatlyreducedsolutionquality[28].Thisapproachdependslessontheorderofnewtasks,andcanndsolutionsthatexistingmethodsmissasisshowninChapter4.2.Thebehaviorofaffectiverecruitmentiscontrolledthroughparameters,andprovidesmoreexibilityindesignthanatraditionalgreedyapproach.Demonstratedrobustness:Thisapproachbuildsonthecontractnetprotocol[90],whichprovidesrobustnessintermsofcommunicationfailures.Messagesbetweenrobotsfollowasequenceofsteps,andthelossofanymessagecanbedetectedandcompensatedfor.Experimentalresults,showninChapter4.2,indicatethattherecruitmentprotocolwillcontinuetofunctionwithupto25%randommessagelossregardlessoftherecruitmentstrategyused,butuptothatpoint,affectiverecruitmenttransmitsfewermessages.Theseresultsbenetthedistributedsensingcommunitybydemonstratingtheperformanceofadistributedprotocolwithrealisticcommunicationlosses.Handlesheterogeneity:Thisapproachmakesnoassumptionsaboutthecompositionoftherobotteam.Robotscanbecompletelyheterogeneousinhardwareandsoftware,anddonotneedtobeoperatingoverthesamesetoftasksorgoals.ThetargetrobotteamconsistsofthreeATRVJr.mobilerobotsandasimulatedhelicopter,whicharevastlydifferentintheirhardwareandsoftware.Thisresultinaheterogeneousteambenetstheroboticscommunity,whererobotsareoftenheterogeneous,eitherbydesignorduetopartialfailures.Fairnessofallocation:Aninterestingside-effectofthisapproachisthatrobotsthatareequallysuitedforataskwilltendtotaketurnsbeingrecruited,suchthatthedisruptiontoeachrobotisdistributedacrosstheteam.ResultsinSection4.2.5indicatethataffectiverecruitmenttendstorecruitmoreuniformlyfromtherobotteam.Deningthebiasofarecruitmentstrategyasthenumberoftimesthatarobotischosenmoreorlessthananyother,onaverage,affectiverecruitmenthadabiasof1.0.9withanon-lineartnessmetricwherethegreedyapproachhadabiasof5.1.92

PAGE 102

6.2FutureWorkTheresearchpresentedinthisthesishasbuiltuponpreviousworkinMRTA,butsomequestionswerenotadequatelyansweredintheliteratureandwerebeyondthescopeofthiswork.Thissectionidentiestwotopicsthatmeritfurtherexploration.Thisworkassumedthatthetnessofarobottoataskcouldbemeasured.TheapproachtakeninthisthesisusedtheestimatedtimeofarrivalasthesinglemetrictogenerateSHAME,andsuggestedinSection3.3thatcombiningtogethermultiplerobotandsensorattributescouldproduceamoresophisticatedmeasureoftness.However,thetopicofsensorutilitywhichistouchedonin[51][104][103][26]isstillanopenareaofresearch.Thedifcultyindeterminingthetnessofarobottoataskinthisdomaincomesfromthefollowingissues,rstoutlinedinChapterThree:Therobotteamisdynamic.Thecapabilitiesofallrobotscannotnecessarilybeknownatonetimeandplace.Evenifthecapabilitiesoftheteamarespeciedinadvance,theywillchangeovertime:sensorsmayfail,robotsmaybedisabledordropoutofcontact,oradditionalrobotsmaybeaddedatanytime.Thus,itisnotsufcienttosimplycompareaparticularrobot'scapabilitiestothoseoftherestoftheteam.Thecomparisonmustbeperformeddirectlyonwhatisknownaboutthesensor,withoutanoutsidereferencebeyond,perhaps,thetaskforwhichtnessisbeingmeasured.Sensorcharacteristicsdonotshareacommonrepresentation.Thedifferentcapabilitiesofadigitalcameracanbemeasuredinradiansviewingangle,Hertzframerate,metersfocallength,wavelengthvisibleversusinfrared,wattspowerconsumption,pixelsandcolordepthresolution,nottomentionaccuracyandprobabilityoffailure.Ideally,atnessmetricmustconsiderallofthesedisparateattributes,butitisnottrivialtocombinetheseintoasinglemeaningfulmeasure.Comparisonsarenotsymmetric.Themetricusedinthisthesiswastheestimatedtimeforarobottoreachthetargetlocation,whichdependsonthemobilityoftherobot.Forinstance,anaerialvehicleUAVmaytravelatahighervelocitythanagroundvehicleUGVandmaytakeamoredirectroute.SupposethattheUGVexperiencesasensorfault,andrequiresanothersimilarsensortobenearbyforrecalibration.Thisleavestwopossibilities:theUAVcouldgototheUGV,orvice-versa.Ifthemetricweretoconsiderthetimeneededforeachrobottoreachtheother,thenitwouldbenecessarytomeasurethetnessforbothcases.Inotherwords,forthetnessfunctionM,itisnotnecessarilytruethatMa;b=Mb;a.93

PAGE 103

Treatmentofsubsetsofcapabilitiesisunclear.Inthisthesis,itwasassumedthattheabilityofarobottoproduceaparticularperceptwasboolean:eitheritcouldgeneratethepercept,andthuswouldparticipateintherecruitmentprocess,oritcouldnot.However,itispossibleforatasktorequiremultiplecapabilities,butitisnotclearhowarobotthathasonlyapartialsubsetofthosecapabilitiesshouldrespond.InSection3.2,theparametersthatgovernaffectiverecruitmentwereintroduced,andaheuristicmethodforselectingtheirvalueswasprovided.Theperformanceofthisapproachistiedtohowtheseparametersarechosen.Allowingtherobotstoadapttheseparametersbasedontheirownactivitywiththegoalofminimizingoverallcommunicationsmayovercometherelianceontheprogrammer'schoiceofvalues.ThereisalsothequestionofhowSHAMEshoulddecayovertime.Therateofdecayforaffectivevariableshasbeenmentionedin[71][79][98],andinspectionoftheresultsin[3]and[6]indicatestheuseofalineardecayrate,whichwasalsousedinthisapproach,butthereisnoauthoritativesourcethatdescribeshowemotionalmotivationshoulddecay.Amoregeneralstudyofhowaffective/motivationalvariablesshouldbehavemaybeuseful.Toconclude,thisthesishasshownthataffectivecomputingpresentsconsiderableadvantagesforthemulti-agentrecruitmentproblem.Basedonthisandearlierworkintheuseofaffectforinternalrobotcontrol,itisexpectedthataffectivecomputingwillbecomeanintegralpartofrobotcontrol.94

PAGE 104

References[1]R.Alami,S.Fleury,M.Herrb,F.Ingrand,andF.Robert.Multi-robotcooperationinthemarthaproject.IEEERobotics&AutomationMagazine,5:36,March1998.[2]T.Arai,E.Pagello,andL.E.Parker.Guesteditorialadvancesinmultirobotsystems.IEEETransactionsonRoboticsandAutomation,18:655,October2002.[3]RonaldC.Arkin,MasahiroFujita,TsuyoshiTakagi,andRikaHasegawa.Ethologicalmodelingandarchitectureforanentertainmentrobot.InProceedingsofthe2001IEEEInternationalConferenceonRoboticsandAutomationICRA,pages453,May2001.[4]SylviaBotelhoandRachidAlami.M+:Aschemeformulti-robotcooperationthroughnegotiatedtaskallocationandachievement.InProceedingsoftheIEEEInternationalConferenceonRoboticsandAutomationICRA,pages1234,Detroit,MI,May1999.[5]CynthiaBreazealandBrianScassellati.Acontext-dependentattentionsystemforasocialrobot.InProceedingsoftheSixteenthInternationalJointConferenceonArticialIntelligenceIJCAI99,pages1146,Stockholm,Sweden,1999.[6]CynthiaBreazealandBrianScassellati.Infant-likesocialinteractionsbetweenarobotandahumancaretaker.AdaptiveBehavior,8,2000.[7]R.R.BrooksandS.S.Iyengar.Robustdistributedcomputingandsensingalgorithm.IEEEComputer,29:53,June1996.[8]KennethAlexanderBrownlee.Statisticaltheoryandmethodologyinscienceandengineering.Wiley,1965.[9]A.Cai,T.Fukuda,A.Arai,K.Yamada,andS.Matsumura.Pathplanningandenvironmentunder-standingbasedondistributedsensingindistributedautonomousroboticsystem.In4thInternationalWorkshoponAdvancedMotionControlProceedingsAMC'96-MIE,volume2,pages699,1996.[10]A.Cai,T.Fukuda,andF.Arai.Informationsharingamongmultiplerobotsforcooperationincellularroboticsystem.InProceedingsofthe1997IEEE/RSJInternationalConferenceonIntelligentRobotSystemsIROS'97,volume2,pages1768,1997.[11]A.Cai,T.Fukuda,F.Arai,andH.Ishihara.Cooperativepathplanningandnavigationbasedondistributedsensing.InProceedings1996IEEEInternationalConferenceonRoboticsandAutomationICRA'96,volume3,pages2079,1996.[12]DoloresCanamero.Issuesinthedesignofemotionalagents.InDoloresCanamero,editor,EmotionalandIntelligent:TheTangledKnotofCognition:Papersfromthe1998FallSymposium,pages23,1998.[13]Y.U.Cao,A.S.Fukunaga,A.B.Kahng,andF.Meng.Cooperativemobilerobotics:Antecedentsanddirections.InIntelligentRobotsandSystems95.'HumanRobotInteractionandCooperativeRobots',Proceedings.1995IEEE/RSJInternationalConferenceon,pages226,1995.95

PAGE 105

[14]JenniferCarlsonandRobinMurphy.Reliabilityanalysisofmobilerobots.InProceedingsInterna-tionalConferenceonRoboticsandAutomationICRA'03,2003.[15]A.CerpaandD.Estrin.Ascent:Adaptiveself-conguringsensornetworktopologies.InProceed-ingsIEEETwenty-FirstAnnualJointConferenceoftheIEEEComputerandCommunicationSocietiesINFOCOM2002,volume3,pages1278,2002.[16]LuizChaimowicz,MarioF.M.Campos,andVijayKumar.Dynamicroleassignmentforcooperativerobots.InProceedingsoftheIEEEInternationalConferenceonRoboticsandAutomationICRA,pages293,Washington,DC,May2002.[17]K.ChakrabartyandS.S.Iyengar.Sensorplacementindistributedsensornetworksusingacodingtheoryframework.InISIT2001,page157,2001.[18]T.Christensen,M.Noergaard,C.Madsen,andA.Hoover.Sensornetworkedmobilerobotics.InProceedingsIEEEConferenceonComputerVisionandPatternRecognitionCVPR2000,volume2,pages782,2000.[19]DanielD.Corkill.Blackboardsystems.AIExpert,6:40,1991.[20]J.Cortes,S.Martinez,T.Karatas,andF.Bullo.Coveragecontrolformobilesensingnetworks.InIEEEInternationalConferenceonRoboticsandAutomationICRA2002,volume2,pages13271332,2002.[21]RandallDavisandReidG.Smith.Negotiationasametaphorfordistributedproblemsolving.ArticialIntelligence,20:63,1983.[22]B.R.Donald.Informationinvariantsinrobotics,partsiandii.InProceedingsIEEEInternationalConferenceonRoboticsandAutomation,pages276,1993.[23]B.R.Donald,J.S.Jennings,andD.Rus.Analyzingteamsofcooperatingmobilerobots.InProceedingsIEEEInternationalConferenceonRoboticsandAutomation,pages1896,May1994.[24]ShirlyDowdy,StanleyWeardon,andDanielChilko.Statisticsforresearch.Wiley-Interscience,2004.[25]MagySeifEl-Nasr,ThomasR.Ioerger,andJohnYen.Peteei:Apetwithevolvingemotionalintelli-gence.InAutonomousAgents99,pages9,1999.[26]AaronGageandRobinR.Murphy.Sensorschedulinginmobilerobotsusingincompleteinforma-tionviamin-conictwithhappiness.IEEETransactionsonSystems,ManandCybernetics,PartB,34:454,February2004.[27]BrianP.Gerkey.Taskallocationforheterogeneousrobots.Master'sthesis,TulaneUniversity,April1998.[28]BrianP.Gerkey.OnMulti-RobotTaskAllocation.PhDthesis,UniversityofSouthernCalifornia,August2003.[29]BrianP.GerkeyandMajaJ.Mataric.Pusher-watcher:anapproachtofault-toleranttightly-coupledrobotcoordination.InProceedingsIEEEInternationalConferenceonRoboticsandAutomationICRA'02,pages464,2002.[30]BrianP.GerkeyandMajaJ.Mataric.Sold!:Auctionmethodsformultirobotcoordination.IEEETransactionsonRoboticsandAutomation,18:758,October2002.[31]BrianP.GerkeyandMajaJ.Mataric.Multi-robottaskallocation:analyzingthecomplexityandoptimalityofkeyarchitectures.InProc.IEEEInt.Conf.RoboticsandAutomationICRA'03,pages3862,September2003.96

PAGE 106

[32]D.E.Gossink,J.B.Scholz,andM.C.Gill.Communicationarchitecturetosupportdistributedsensors.InRecordoftheThirty-SecondAsilomarConferenceonSignals,Systems,&Computers,volume1,pages588,November1998.[33]JaroslavHajek,ZbynekSidak,andPranabK.Sen.TheoryofRankTests.AcademicPress,1999.[34]A.T.Hayes,A.Martinoli,andR.M.Goodman.Distributedodorsourcelocalization.IEEESensorsJournal,2:260,2002.[35]IanHorswill.Analysisofadaptationandenvironment.ArticialIntelligence,73:1,1995.[36]C.Intanagonwiwat,R.Govindan,D.Estrin,J.Heidemann,andF.Silva.Directeddiffusionforwirelesssensornetworking.IEEE/ACMTransactionsonNetworking,11:2,February2003.[37]S.S.Iyengar,D.N.Jayasimha,andD.Nadig.Aversatilearchitectureforthedistributedsensorinte-grationproblem.IEEETransactionsonComputers,43:175,February1994.[38]S.S.IyengarandL.Prasad.Ageneralcomputationalframeworkfordistributedsensingandfault-tolerantsensorintegration.IEEETransactionsonSystems,ManandCybernetics,25:643,1995.[39]M.JenkinandG.Dudek.Thepaparazziproblem.InProceedings2000IEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsIROS2000,volume3,pages2042,2000.[40]J.S.Jennings,G.Whelan,andW.F.Evans.Cooperativesearchandrescuewithateamofmobilerobots.InProc.IEEEInt.Conf.AdvancedRobotics,pages193,Monterey,CA,1997.[41]ChrisJonesandMajaJ.Mataric.Adaptivedivisionoflaborinlarge-scaleminimalistmulti-robotsystems.InProceedingsIEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsIROS-03,LasVegas,NV,October27-312003.[42]ChrisV.JonesandMajaJ.Mataric.Theuseofinternalstateinmulti-robotcoordination.InHawaiiInternationalConferenceonComputerSciences,Waikiki,Hawaii,January2004.[43]E.D.Jones,T.C.SteveHsia,andRandyS.Roberts.Stomp:Simulation,tacticaloperationsandmissionplanning.InProceedings2003InternationalConferenceonRoboticsandAutomationICRA'03,2003.[44]R.Kannan,S.Sarangi,S.Ray,andS.S.Iyengar.Minimalsensorintegrityinsensorgrids.InISIF2002,pages959,2002.[45]A.KhooandI.Horswill.Anefcientcoordinationarchitectureforautonomousrobotteams.InProceedingsIEEEInternationalConferenceonRoboticsandAutomationICRA,pages287,May2002.[46]M.J.B.Krieger,J.-B.Billeter,andL.Keller.Ant-liketaskallocationandrecruitmentincooperativerobots.Nature,406:992,August2000.[47]C.R.KubeandHongZhang.Stagnationrecoverybehavioursforcollectiverobots.InProceedingsoftheIEEE/RSJ/GIInternationalConferenceonIntelligentRobotsandSystemsIROS'94,pages1883vol.3,1994.[48]M.VigneshKumarandFeratSahin.Aswarmintelligencebasedapproachtotheminedetectionproblem.InProceedingsSystems,Man,andCybernetics,2002IEEEInternationalConferenceOn,2002.[49]KelvinK.LeeandSamuelT.Chanson.Packetlossprobabilityforreal-timewirelesscommunications.IEEETransactionsonVehicularTechnology,51:1569,November2002.97

PAGE 107

[50]H.LeventhalandK.Scherer.Therelationshipofemotiontocognition:Afunctionalapproachtoasemanticcontroversy.CognitionandEmotion,1:3,1984.[51]J.LindnerandR.R.Murphy.Learningtheexpectedutilityofsensorsandalgorithms.InProceedingsof1994IEEEInternationalConferenceonMFI'94.MultisensorFusionandIntegrationforIntelligentSystems,pages583,1994.[52]ChristineLisetti,S.Brown,K.Alvarez,andA.Marpaung.Asocialinformaticsapproachtohuman-robotinteractionwithanofceservicerobot.IEEESystems,Man,andCyberneticsSpecialIssueonHumanRobotInteraction,2004.Inpress.[53]MattT.Long,RobinR.Murphy,andLynneE.Parker.Multi-agentdiagnosisandrecoveryfromsensorfailures.InProceedingsIEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsIROS,pages2506,October2003.[54]F.Martinerie.Newdatafusionandtrackingapproachesinmultipletargets/distributedsensorsnet-workcontexts.In1993IEEEInternationalConferenceonAcoustics,Speech,andSignalProcessingICASSP-93,volume1,pages249,April1993.[55]MajaMataric.Minimizingcomplexityincontrollingamobilerobotpopulation.InProceedingsoftheIEEEInternationalConferenceonRoboticsandAutomation,pages830,May1992.[56]MajaJ.MataricandGauravSukhatme.Task-allocationandcoordinationofmultiplerobotsforplan-etaryexploration.In10thInternationalConferenceonAdvancedRobotics,pages61,Buda,Hun-gary,August2001.[57]AkihiroMatsumoto,HajimeAsama,YoshikiIshida,KoichiOzaki,andIsaoEndo.Communicationintheautonomousanddecentralizedrobotsystemactress.InIEEEInternationalWorkshoponIntelligentRobotsandSystemsIROS'90,pages835,1990.[58]ColinMcMillen,KristenN.Stubbs,PaulE.Rybski,SaschaA.Stoeter,MariaGini,andNikolaosP.Papanikolopoulos.Resourceschedulingandloadbalancingindistributedroboticcontrolsystems.RoboticsandAutonomousSystems,44:251,2003.[59]F.Michaud,P.Prijanian,J.Audet,andD.Letourneau.Articialemotionsandsocialrobotics.InProceedingsoftheDistributedAutonomousRoboticSystems,2000.[60]F.Michaud,P.Prijanian,J.Audet,D.Letourneau,L.Lussier,C.Theberge-Turmel,andS.Caron.Experienceswithanautonomousrobotattendingaaai.IEEEIntelligentSystems,September-October2001.[61]FrancoisMichaud.Emibcomputationalarchitecturebasedonemotionandmotivationforintentionalselectionandcongurationofbehavior-producingmodules.CognitiveScienceQuarterly,3:340361,2002.[62]FrancoisMichaud,E.Robichaud,andJ.Audet.Usingmotivesandarticialemotionforprolongedactivityforagroupofautonomousrobots.InAAAIFallSymposiumSeries,2001.[63]DovMondererandMosheTennenholtz.Optimalauctionsrevisited.InProceedingsFifteenthNationalConferenceonArticialIntelligence,pages32,1998.[64]RobinMurphy,ChristineLisetti,RussTardif,LiamIrish,andAaronGage.Emotion-basedcontrolofcooperatingheterogeneousmobilerobots.IEEETransactionsonRoboticsandAutomation,18:744757,October2002.[65]RobinR.Murphy.IntrotoAIRobotics.MITPress,2000.98

PAGE 108

[66]RobinR.MurphyandRonaldC.Arkin.Sfx:Anarchitectureforaction-orientedsensorfusion.InIEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsIROS,volume2,pages10791086,July1992.[67]RobinRobersonMurphy.AnArchitectureForIntelligentRoboticSensorFusion.PhDthesis,GeorgiaInstituteofTechnology,June1992.[68]D.Nadig,S.S.Iyengar,andD.N.Jayasimha.Anewarchitecturefordistributedsensorintegration.InProceedingsIEEESoutheastcon'93,page8,April1993.[69]F.R.Noreils.Anarchitectureforcooperativeandautonomousmobilerobots.InProceedings1992IEEEInternationalConferenceonRoboticsandAutomation,pages2703,Nice,France,May1992.[70]TakuyaOhko,KazuoHiraki,andYuichiroAnzai.Reducingcommunicationloadoncontractnetbycase-basedreasoningextensionwithdirectedcontractandforgetting.InProceedingsoftheSecondInternationalConferenceonMultiagentSystems,pages244,1996.[71]AndrewOrtony.Subjectiveimportanceandcomputationalmodelsofemotions.InV.Hamilton,G.H.Bower,andN.H.Frijda,editors,CognitivePerspectivesonEmotionandMotivation,chapter13,pages321.KluwerAcademicPublishers,1988.[72]AndrewOrtony.Onmakingbelievableemotionalagentsbelievable.InRobertTrappl,PaoloPetta,andSabinePayr,editors,EmotionsinHumansandArtifacts,chapter6,pages189.TheMITPress,2002.[73]AndrewOrtony,GeraldL.Clore,andAllanCollins.TheCognitiveStructureofEmotions.CambridgeUniversityPress,1988.[74]EsbenH.Ostergaard,MajaJ.Mataric,andGauravS.Sukhatme.Distributedmulti-robottaskallocationforemergencyhandling.InProceedingsofIEEE/RSJInternationalConferenceonRobotsandSystemsIROS,pages821,Maui,Hawaii,October2001.[75]L.E.Parker.Alliance:Anarchitectureforfaulttolerantmultirobotcooperation.IEEETransactionsonRoboticsandAutomation,14:220,April1998.[76]LynneE.Parker.Alliance:Anarchitectureforfaulttolerant,cooperativecontrolofheterogeneousmo-bilerobots.InProceedingsofthe1994IEEE/RSJ/GIInternationalConferenceonIntelligentRobotsandSystemsIROS'94,pages776,1994.[77]LynneE.Parker.Evaluatingsuccessinautonomousmulti-robotteams:Experiencesfromalliancearchitectureimplementations.JournalofExperimentalandTheoreticalArticialIntelligence,13:9598,2001.[78]LynneE.Parker.Currentresearchinmulti-robotsystems.JournalofArticialLife,7,2004.Toappear.[79]RolfPfeifer.Articialintelligencemodelsofemotion.InV.Hamilton,G.H.Bower,andN.H.Fri-jda,editors,CognitivePerspectivesonEmotionandMotivation,chapter12,pages287.KluwerAcademicPublishers,1988.[80]P.J.ProbertandH.Hu.Transputercontrolandsensingarchitecturesformobilerobots.InIEECollo-quiumonRobotSystemsArchitectures,pages6/1/3,October1990.[81]D.Rus,B.Donald,andJ.S.Jennings.Movingfurniturewithteamsofautonomousrobots.InProceed-ingsIEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsIROS,pages235,August1995.99

PAGE 109

[82]P.E.Rybski,S.A.Stoeter,M.Gini,D.F.Hougen,andN.P.Papanikolopolous.Performanceofadis-tributedroboticsystemusingsharedcommunicationschannels.IEEETransactionsonRoboticsandAutomation,18:713,October2002.[83]S.Sakane,H.Ohosi,T.Sato,andM.Kakikura.Distributedsensingsystemwith3dmodel-basedagents.InProceedingsofthe1993IEEE/RSJInternationalConferenceonIntelligentRobotsandSystemsIROS'93,volume2,pages1157,1993.[84]KlausR.Scherer.Criteriaforemotion-antecedentappraisal:Areview.InV.Hamilton,G.H.Bower,andN.H.Frijda,editors,CognitivePerspectivesonEmotionandMotivation,chapter4,pages89.KluwerAcademicPublishers,1988.[85]OnnShehoryandSaritKraus.Taskallocationviacoalitionformationamongautonomousagents.InProceedingsFourteenthInternationalJointConferenceonArticialIntelligence,pages65561,1995.[86]DavidSheskin.Handbookofparametricandnonparametricstatisticalprocedures.CRCPress,1997.[87]ReidSimmons,SanjivSingh,DaveHershberger,J.Ramos,andT.Smith.Firstresultsinthecoordi-nationofheterogeneousrobotsforlarge-scaleassembly.InProceedingsInternationalSymposiumonExperimentalRoboticsISER2000,2000.[88]KaransherSinghandKikuoFujimura.Mapmakingbycooperatingmobilerobots.InIEEEInterna-tionalConferenceonRoboticsandAutomation,Proceedings,pages254vol.2,May1993.[89]AaronSlomanandMonicaCroucher.Whyrobotswillhaveemotions.InProceedingsInternationalJointConferenceonArticialIntelligenceIJCAI,pages197,1981.[90]ReidG.Smith.Thecontractnetprotocol:High-levelcommunicationandcontrolinadistributedproblemsolver.IEEETransactionsonComputers,c-29:1104,1980.[91]Leen-KiatSohandCostasTsatsoulis.Reectivenegotiatingagentsforreal-timemultisensortargettracking.InProceedingsSeventeenthInternationalJointConferenceonArticialIntelligence,pages1121,2001.[92]J.SpletzerandC.Taylor.Dynamicsensorplanningandcontrolforoptimallytrackingtargets.Inter-nationalJournalofRoboticsResearch,2003.[93]P.A.Stadter,A.A.Chacos,R.J.Heins,G.T.Moore,E.A.Olsen,M.S.Asher,andJ.O.Bristow.Con-uenceofnavigation,communication,andcontrolindistributedspacecraftsystems.IEEEAerospaceandElectronicSystemsMagazine,17:26,2002.[94]PeterStoneandManuelaVeloso.Taskdecomposition,dynamicroleassignment,andlow-bandwidthcommunicationforreal-timestrategicteamwork.ArticialIntelligence,110:241,1999.[95]A.W.Stroup,M.C.Martin,andT.Balch.Distributedsensorfusionforobjectpositionestimationbymulti-robotsystems.InProceedings2001InternationalConferenceonRoboticsandAutomationICRA2001,volume2,pages1092,2001.[96]Y.Takeuchi,M.Sowa,andK.Horikawa.Aninformationtheoreticschemeforsensorallocationoflinearleast-squaresestimation.InProceedingsofthe41stSICEAnnualConferenceSICE2002,volume1,pages539,2002.[97]K.Umeda,K.Ikushima,andT.Arai.3dshaperecognitionbydistributedsensingofrangeimagesandintensityimages.InProceedings1997IEEEInternationalConferenceonRoboticsandAutomationICRA'97,volume1,pages20,April1997.100

PAGE 110

[98]JuanD.Velasquez.Modelingemotionsandothermotivationsinsyntheticagents.InProceedingsoftheFourteenthNationalConferenceonArticialIntelligence,pages10,1997.[99]JuanD.Velasquez.Whenrobotsweep:Emotionalmemoriesanddecision-making.InProceedingsoftheFifteenthNationalConferenceonArticialIntelligence,pages70,1998.[100]Chieh-ChihWangandC.Thorpe.Simultaneouslocalizationandmappingwithdetectionandtrackingofmovingobjects.InProceedingsIEEEInternationalConferenceonRoboticsandAutomationICRA'02,volume3,pages2918,2002.[101]BarryBrianWergerandMajaMataric.Broadcastoflocaleligibilityformulti-targetobservation.InLynneE.Parker,GeorgeBekey,andJacobBarhen,editors,DistributedAutonomousRoboticSystems4,pages347.Springer-Verlag,2000.[102]P.Wide,A.Safotti,andH.-H.Bothe.Environmentalexploration:anautonomoussensorysystemsapproach.IEEEInstrumentation&MeasurementMagazine,2:28,1999.[103]H.XuandJ.Vandorpe.Perceptionplanninginmobilerobotnavigation.InProceedingson1994IEEEInternationalConferenceonMFI'94.MultisensorFusionandIntegrationforIntelligentSystems,pages723,1994.[104]Y.F.Zheng.Integrationofmultiplesensorsintoaroboticssystemanditsperformanceevaluation.IEEETransactionsonRoboticsandAutomation,5:658,1989.[105]RobertZlot,AnthonyStentz,M.BernardineDias,andScottThayer.Multi-robotexplorationcon-trolledbyamarketeconomy.InProceedingsoftheIEEEInternationalConferenceonRoboticsandAutomationICRA,pages3016,Washington,DC,May2002.101

PAGE 111

Appendices102

PAGE 112

AppendixARawSimulationResultsTable20.Numberoftimeseachrobotwasrecruitedusingaffectiverecruitment.ThesevalueswereusedtocomputethefairnessofeachstrategyinChapter4.2.5. Run 1Recruited 4Recruited 5Recruited 6Recruited 7Recruited Total Imbalance 4-5-1 6 4 5 5 5 25 1 4-5-2 5 5 6 5 4 25 1 4-5-3 5 5 5 5 5 25 0 4-5-4 5 5 5 5 5 25 0 4-5-5 5 7 5 4 4 25 2 4-5-6 4 6 5 5 5 25 1 4-5-7 5 5 6 4 5 25 1 4-5-8 7 5 5 3 5 25 2 4-5-9 5 5 4 5 6 25 1 4-5-10 5 4 5 5 6 25 1 103

PAGE 113

AppendixAContinuedTable21.Numberoftimeseachrobotwasrecruitedusingaffective1=D2recruitment.ThesevalueswereusedtocomputethefairnessofeachstrategyinChapter4.2.5. Run 1Recruited 4Recruited 5Recruited 6Recruited 7Recruited Total Imbalance 4-5-1 6 5 4 6 4 25 2 4-5-2 4 5 6 5 5 25 1 4-5-3 5 6 5 5 4 25 1 4-5-4 5 3 5 7 5 25 2 4-5-5 6 7 3 5 4 25 3 4-5-6 4 6 5 5 5 25 1 4-5-7 7 5 5 5 3 25 2 4-5-8 0 6 7 6 6 25 5 4-5-9 5 5 4 5 6 25 1 4-5-10 4 5 5 5 6 25 1 Table22.Numberoftimeseachrobotwasrecruitedusinggreedyrecruitment.ThesevalueswereusedtocomputethefairnessofeachstrategyinChapter4.2.5. Run 1Recruited 4Recruited 5Recruited 6Recruited 7Recruited Total Imbalance 4-5-1 4 5 7 5 4 25 2 4-5-2 3 3 3 9 7 25 6 4-5-3 5 7 3 7 3 25 4 4-5-4 2 3 6 6 8 25 5 4-5-5 6 9 3 4 3 25 5 4-5-6 2 6 4 2 11 25 7 4-5-7 8 3 4 7 3 25 5 4-5-8 6 7 8 2 2 25 6 4-5-9 6 3 2 7 7 25 5 4-5-10 2 3 4 8 8 25 6 Table23.Numberoftimeseachrobotwasrecruitedusingrandomrecruitment.ThesevalueswereusedtocomputethefairnessofeachstrategyinChapter4.2.5. Run 1Recruited 4Recruited 5Recruited 6Recruited 7Recruited Total Imbalance 4-5-1 4 8 5 2 6 25 4 4-5-2 4 6 4 6 5 25 2 4-5-3 3 8 5 5 4 25 3 4-5-4 5 4 5 7 4 25 2 4-5-5 7 6 4 6 2 25 4 4-5-6 2 4 7 4 8 25 5 4-5-7 4 4 2 8 7 25 5 4-5-8 4 3 5 8 5 25 3 4-5-9 4 5 4 6 6 25 2 4-5-10 6 4 2 8 5 25 4 104

PAGE 114

AppendixAContinuedTable24.Rawdatafortimemetric,4robots,and0%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-1-1 328.4 441 254.3 255.2 1-1-2 392.7 398.2 258.5 259.6 1-1-3 411 416.3 267.8 268.2 1-1-4 440.2 456.6 287 286.9 1-1-5 393.1 408.5 255.2 257.4 1-1-6 442.8 456 285.3 286 1-1-7 412.9 456.4 270.9 271 1-1-8 379.8 372.6 234 233.8 1-1-9 338.7 354.3 247.6 247.2 1-1-10 448.2 456.4 290.4 290.4 1-1-11 360.2 371.2 230.4 229.6 1-1-12 389.3 412.2 258.7 261.2 1-1-13 345.1 359.6 234.4 234.1 1-1-14 404.8 418.2 265.9 265.9 1-1-15 468.6 476.6 314.2 314.4 1-1-16 391.4 406 259.5 259.3 1-1-17 489.1 496 360.4 380.4 1-1-18 385.5 360.3 227.9 228.6 1-1-19 443.4 460.8 283.3 283.8 1-1-20 325.2 375.7 215.6 212.1 1-1-21 390.6 397.5 254.5 254.3 1-1-22 350.3 364 223.7 224 1-1-23 437.4 438.3 268.6 268.7 1-1-24 468.1 472.3 318.5 317.7 1-1-25 474.8 478.7 327.5 327 1-1-26 423.6 434.8 280.4 281.4 1-1-27 440 482 290.9 291.3 1-1-28 452.1 458.2 292.3 292.2 1-1-29 410.4 409.5 280.7 280.5 1-1-30 442.8 448.1 314.8 315.2 105

PAGE 115

AppendixAContinuedTable25.Rawdatafornumberofmessagesmetric,4robots,and0%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-1-1 38 51 27 27 1-1-2 53 55 30 30 1-1-3 53 55 30 30 1-1-4 56 53 30 30 1-1-5 52 55 30 30 1-1-6 53 52 30 30 1-1-7 53 52 30 30 1-1-8 52 55 30 30 1-1-9 51 54 30 30 1-1-10 53 52 30 30 1-1-11 55 55 30 30 1-1-12 52 55 30 30 1-1-13 51 53 30 30 1-1-14 53 55 30 30 1-1-15 53 54 32 32 1-1-16 52 54 30 30 1-1-17 54 55 32 32 1-1-18 51 55 30 30 1-1-19 53 55 30 30 1-1-20 53 51 30 30 1-1-21 52 53 30 30 1-1-22 53 57 30 30 1-1-23 53 53 30 30 1-1-24 54 57 30 30 1-1-25 54 55 30 30 1-1-26 53 55 30 30 1-1-27 56 55 30 30 1-1-28 54 55 30 30 1-1-29 54 54 30 30 1-1-30 54 55 30 30 106

PAGE 116

AppendixAContinuedTable26.Rawdatafortimemetric,8robots,and0%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-5-1 178.4 173.1 77.7 255.1 1-5-2 218.2 227.6 146.8 191.7 1-5-3 246 246.6 135.4 192.5 1-5-4 204.2 212.8 106.5 300.6 1-5-5 219.2 194.6 113.7 376.8 1-5-6 228 230.5 128.6 341.1 1-5-7 302.3 315 155.7 309.5 1-5-8 175.9 196.2 104.1 316.6 1-5-9 235.1 234.7 109.1 319.7 1-5-10 165.1 173 65.5 295.7 1-5-11 177.6 191.8 118.2 198 1-5-12 194.7 208.5 115.4 227.5 1-5-13 263 288 136.9 338 1-5-14 391.7 402.5 210.9 292.8 1-5-15 297.2 328.4 151.8 281 1-5-16 259.5 295.2 159 289 1-5-17 376.7 365.7 251.6 340.5 1-5-18 242.6 269 108.3 287.8 1-5-19 316.2 294.1 170.2 349.2 1-5-20 238.1 253.5 145.8 318.6 1-5-21 232.7 271.6 122.9 239.6 1-5-22 230.3 249 162.8 223.2 1-5-23 334.3 299.5 177.7 367.2 1-5-24 279.3 287.9 187 309.8 1-5-25 272.1 301.2 165 264.7 1-5-26 256.9 290.2 121.5 276 1-5-27 319.6 286.7 190.8 334.5 1-5-28 270.3 287.1 174.2 405.3 1-5-29 298.2 289.1 192.4 381.4 1-5-30 257.2 284.8 141.5 324.2 107

PAGE 117

AppendixAContinuedTable27.Rawdatafornumberofmessagesmetric,8robots,and0%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-5-1 46 50 50 50 1-5-2 51 58 50 50 1-5-3 58 60 50 50 1-5-4 52 61 50 50 1-5-5 50 58 50 50 1-5-6 50 60 50 50 1-5-7 53 64 50 50 1-5-8 47 61 50 50 1-5-9 60 65 50 50 1-5-10 47 53 50 50 1-5-11 51 55 50 50 1-5-12 51 58 50 50 1-5-13 51 65 50 52 1-5-14 56 58 50 50 1-5-15 53 61 50 50 1-5-16 54 55 50 50 1-5-17 56 68 50 50 1-5-18 45 59 50 50 1-5-19 59 64 50 50 1-5-20 56 60 50 50 1-5-21 54 59 50 50 1-5-22 51 58 50 50 1-5-23 59 60 50 52 1-5-24 58 57 50 50 1-5-25 53 59 50 50 1-5-26 54 61 50 50 1-5-27 60 61 50 50 1-5-28 60 67 50 50 1-5-29 60 60 50 52 1-5-30 56 57 50 50 108

PAGE 118

AppendixAContinuedTable28.Rawdatafortimemetric,13robots,and0%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-10-1 151.4 143.7 67 304.6 1-10-2 197.9 247.6 114.9 222 1-10-3 179.4 186.3 97.2 258.2 1-10-4 212.9 203.1 99.9 381.8 1-10-5 205.7 190.5 105.9 314.3 1-10-6 199.1 331.8 123.7 294.2 1-10-7 240.5 272.8 132.4 304.4 1-10-8 167.2 176.4 85.4 252 1-10-9 155.6 154.6 75.1 293.9 1-10-10 162 170 80.1 389.5 1-10-11 160.3 174.3 103.8 292.4 1-10-12 199 271.9 122.9 462.5 1-10-13 168.3 152.6 70 319.9 1-10-14 278.3 293.2 182.3 255.1 1-10-15 229.5 247.7 132.4 283.7 1-10-16 229.7 244 154.3 308 1-10-17 250 260.6 134.1 281.5 1-10-18 171.7 153.5 94.4 349 1-10-19 230.5 238.1 128.5 251.9 1-10-20 295.5 310.5 147.2 261.7 1-10-21 167.1 182.5 91.6 256.8 1-10-22 155.6 175.6 91.6 402 1-10-23 273.6 224.1 124.4 334.2 1-10-24 234.3 244.1 155 295.1 1-10-25 240.6 262.1 153.8 277.3 1-10-26 170.8 182.9 95.3 373.8 1-10-27 270.3 252.8 155.4 223.7 1-10-28 202.5 220.1 120.2 231.1 1-10-29 273.9 248.9 140.9 315.4 1-10-30 234.6 235.7 145.3 349.4 109

PAGE 119

AppendixAContinuedTable29.Rawdatafornumberofmessagesmetric,13robots,and0%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-10-1 59 59 75 75 1-10-2 67 70 75 75 1-10-3 57 72 75 75 1-10-4 60 81 75 75 1-10-5 55 78 75 75 1-10-6 60 76 75 75 1-10-7 68 81 75 75 1-10-8 60 88 75 75 1-10-9 64 70 75 75 1-10-10 54 68 75 77 1-10-11 54 61 75 75 1-10-12 65 79 75 75 1-10-13 62 60 75 75 1-10-14 77 80 75 75 1-10-15 65 79 75 75 1-10-16 62 68 75 75 1-10-17 69 77 75 75 1-10-18 63 58 75 75 1-10-19 61 89 75 75 1-10-20 66 82 75 75 1-10-21 64 78 75 75 1-10-22 60 70 75 75 1-10-23 64 89 75 75 1-10-24 65 78 75 75 1-10-25 54 70 75 75 1-10-26 68 79 75 75 1-10-27 74 80 75 75 1-10-28 69 89 75 75 1-10-29 74 74 75 75 1-10-30 73 71 75 75 110

PAGE 120

AppendixAContinuedTable30.Rawdatafortimemetric,23robots,and0%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-20-1 158.2 152.6 86.6 410.4 1-20-2 166.1 157 84.1 261.8 1-20-3 201.7 195.3 105.3 305.8 1-20-4 187.9 155.2 80.4 297.4 1-20-5 174.7 202.7 106.1 294.8 1-20-6 174.2 163.1 91.2 389.2 1-20-7 230.8 208 119.1 483 1-20-8 270.7 177.1 101.6 398.6 1-20-9 180.7 166 92 427.5 1-20-10 196.5 157 85.9 399.3 1-20-11 171 169.3 96.7 371.5 1-20-12 217.9 188.2 124.2 224.2 1-20-13 175 163.7 95.8 476.4 1-20-14 237 226.5 114.8 399 1-20-15 203.7 208 122.4 342.1 1-20-16 228.9 186.4 125.4 372.6 1-20-17 212 215.8 129 346.1 1-20-18 192.9 169.6 102.5 384.1 1-20-19 220.3 191.4 117.2 362.4 1-20-20 123.4 180.3 77.4 309.1 1-20-21 184.8 176.4 100.2 400.6 1-20-22 159.7 139.5 73.7 350.5 1-20-23 172.2 171.9 96.5 338.5 1-20-24 143.3 174.4 81.3 371.9 1-20-25 203.2 228.7 143.9 330.1 1-20-26 242.3 212.2 115.5 344 1-20-27 236.9 217.6 129.7 363.3 1-20-28 194.4 203.3 117 334.9 1-20-29 195.9 179.7 164.7 407.7 1-20-30 193.7 179.8 131.3 348 111

PAGE 121

AppendixAContinuedTable31.Rawdatafornumberofmessagesmetric,23robots,and0%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-20-1 65 61 125 125 1-20-2 95 83 125 125 1-20-3 79 114 125 125 1-20-4 86 99 125 125 1-20-5 81 79 125 125 1-20-6 63 99 125 127 1-20-7 65 99 125 125 1-20-8 82 105 125 125 1-20-9 90 100 125 125 1-20-10 63 78 125 125 1-20-11 80 70 125 125 1-20-12 77 74 125 125 1-20-13 91 86 125 125 1-20-14 84 108 125 125 1-20-15 96 103 125 125 1-20-16 67 80 125 125 1-20-17 101 118 125 125 1-20-18 87 78 125 125 1-20-19 82 81 125 125 1-20-20 72 61 125 125 1-20-21 98 117 125 125 1-20-22 99 78 125 125 1-20-23 84 91 125 125 1-20-24 64 117 125 125 1-20-25 80 84 125 125 1-20-26 82 138 125 125 1-20-27 91 119 125 125 1-20-28 95 118 125 125 1-20-29 71 93 125 125 1-20-30 87 99 125 125 112

PAGE 122

AppendixAContinuedTable32.Rawdatafortimemetric,53robots,and0%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-50-1 137.7 120.4 115.2 395.7 1-50-2 134 134.4 124.1 389 1-50-3 136 105.4 124.7 423.4 1-50-4 105.1 149.3 117.8 492.1 1-50-5 119.7 159.7 124.3 344.9 1-50-6 181 87.4 114.1 486 1-50-7 93.3 140.8 117.2 415.4 1-50-8 112.8 96.3 119.2 483.2 1-50-9 118.3 129.6 120.1 362 1-50-10 149.8 236.1 115.5 293.8 1-50-11 134.1 153.2 130.8 380.5 1-50-12 156.1 166.3 137.8 519.2 1-50-13 247.7 178.9 157.2 519.2 1-50-14 152.5 170 138.8 437 1-50-15 127.4 156 122.6 247.8 1-50-16 225 201.4 168.1 339.4 1-50-17 157.1 169.5 135.4 479.2 1-50-18 126.9 133.2 145.5 348.2 1-50-19 167.8 197.7 142.6 360.1 1-50-20 149.4 118.6 133 360.9 1-50-21 172.3 193.5 132.2 405.6 1-50-22 129.6 104.5 120 354.7 1-50-23 120.9 128.5 121.8 414.1 1-50-24 99 76.3 122.2 461.6 1-50-25 174.5 164 134.4 406.7 1-50-26 213.9 174.3 149.5 393.2 1-50-27 174.8 158.2 146.1 389.2 1-50-28 263.6 198.7 143.2 391 1-50-29 131.1 220.4 123.4 368.3 1-50-30 154.8 170.2 166.1 459.4 113

PAGE 123

AppendixAContinuedTable33.Rawdatafornumberofmessagesmetric,53robots,and0%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-50-1 98 51 275 275 1-50-2 124 43 275 275 1-50-3 66 46 275 275 1-50-4 78 45 275 275 1-50-5 103 68 275 275 1-50-6 107 48 275 275 1-50-7 57 48 275 275 1-50-8 102 48 275 275 1-50-9 96 54 275 275 1-50-10 76 44 275 275 1-50-11 119 88 275 274 1-50-12 154 141 275 275 1-50-13 149 138 275 275 1-50-14 121 48 275 275 1-50-15 128 141 275 275 1-50-16 107 54 275 275 1-50-17 156 140 275 275 1-50-18 67 88 275 275 1-50-19 162 187 275 275 1-50-20 123 47 275 275 1-50-21 109 138 275 275 1-50-22 102 90 275 275 1-50-23 117 95 275 275 1-50-24 76 41 275 275 1-50-25 112 92 275 275 1-50-26 124 140 275 275 1-50-27 108 146 275 275 1-50-28 101 187 275 275 1-50-29 122 59 275 275 1-50-30 108 46 327 275 114

PAGE 124

AppendixAContinuedTable34.Rawdatafortimemetric,13robots,and5%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-10-1a 241.2 279.5 182.1 433.4 1-10-1b 217.7 162.3 107.6 389.8 1-10-1c 476.3 215.6 155.6 334.3 1-10-2a 339.6 244.9 117.1 412.9 1-10-2b 272.8 268.1 124.5 454.7 1-10-2c 296.9 313.1 128.2 469.4 1-10-3a 322.4 286.2 121.8 240 1-10-3b 236.4 262.9 155.2 466.5 1-10-3c 277.3 194.9 104.8 456.3 1-10-4a 291.3 236.6 130.8 335.7 1-10-4b 215.3 380.1 112.8 274.2 1-10-4c 417.3 230.8 239.7 245.4 1-10-5a 214 243.2 129.8 399.3 1-10-5b 205 335.3 105.4 646.6 1-10-5c 240.9 543.9 152.6 452.5 Table35.Rawdatafornumberofmessagesmetric,13robots,and5%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-10-1a 72 123 127 99 1-10-1b 70 63 104 91 1-10-1c 102 54 91 89 1-10-2a 82 71 75 90 1-10-2b 82 101 75 135 1-10-2c 85 93 75 103 1-10-3a 75 104 89 75 1-10-3b 69 79 106 117 1-10-3c 90 81 75 100 1-10-4a 73 84 88 75 1-10-4b 79 99 76 90 1-10-4c 71 83 149 75 1-10-5a 54 74 89 77 1-10-5b 64 99 75 102 1-10-5c 63 116 87 120 115

PAGE 125

AppendixAContinuedTable36.Rawdatafortimemetric,13robots,and10%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective,1=D Affective,1=D2 Greedy Random 1-10-1a 175.5 363.5 160.9 528.5 1-10-1b 176 341.8 175.8 304 1-10-1c 218.7 284.2 166.6 188.5 1-10-2a 197.4 215.6 167 422.5 1-10-2b 462.2 321.5 164.9 451.5 1-10-2c 405.9 508.3 158.7 524.3 1-10-3a 524.7 228.3 205.1 509.1 1-10-3b 299.8 368.3 147.1 596.1 1-10-3c 444.6 419.2 190.2 482.8 1-10-4a 289.6 260.5 111.4 414.6 1-10-4b 292.2 282.4 236.3 499.9 1-10-4c 418.7 416.2 156.7 531.1 1-10-5a 208.1 322 304.9 552.7 1-10-5b 393.9 391.5 111.3 288.2 1-10-5c 277 626.3 192.2 417.4 Table37.Rawdatafornumberofmessagesmetric,13robots,and10%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-10-1a 68 107 119 116 1-10-1b 67 96 130 107 1-10-1c 67 91 104 75 1-10-2a 67 89 90 118 1-10-2b 101 106 99 104 1-10-2c 102 125 90 119 1-10-3a 113 107 106 104 1-10-3b 75 139 89 104 1-10-3c 125 145 114 104 1-10-4a 90 99 77 101 1-10-4b 72 96 118 100 1-10-4c 96 123 105 101 1-10-5a 63 87 143 105 1-10-5b 83 110 76 76 1-10-5c 79 132 104 121 116

PAGE 126

AppendixAContinuedTable38.Rawdatafortimemetric,13robots,and25%communicationfailurerate.Valuesareinseconds.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-10-1a 634.9 674.2 297.3 857.6 1-10-1b 601.7 413.3 451.5 1086.9 1-10-1c 525.5 482.4 243 748.5 1-10-2a 618 691.5 538.4 878.5 1-10-2b 878.4 795.1 555.3 1820.4 1-10-2c 557.5 795.1 268.1 738.6 1-10-3a 596.1 703.1 195.5 591.6 1-10-3b 482.8 384.2 452.9 1391 1-10-3c 856.5 660.8 204.4 1347.7 1-10-4a 865 795.2 619.7 1479 1-10-4b 669.1 798.1 433.2 1201.7 1-10-4c 499.8 795.5 428.7 907.5 1-10-5a 664.6 552.7 395.3 778.3 1-10-5b 737.6 873.7 468.1 799.4 1-10-5c 499.9 794.8 492.6 772.6 Table39.Rawdatafornumberofmessagesmetric,13robots,and25%communicationfailurerate.Smallervaluesarebetter. Run Affective Affective,1=D2 Greedy Random 1-10-1a 154 178 144 175 1-10-1b 169 110 193 242 1-10-1c 119 155 148 152 1-10-2a 164 201 249 212 1-10-2b 156 192 243 288 1-10-2c 108 194 129 158 1-10-3a 139 162 133 130 1-10-3b 122 95 219 224 1-10-3c 159 153 139 218 1-10-4a 161 181 256 227 1-10-4b 190 210 230 185 1-10-4c 149 193 221 187 1-10-5a 158 130 202 195 1-10-5b 189 239 231 161 1-10-5c 109 166 278 143 117

PAGE 127

AbouttheAuthorAaronGageisaPh.D.candidateinComputerScienceandEngineeringattheUniversityofSouthFlorida.HeholdsaBSinMathematicalandComputerSciencesfromtheColoradoSchoolofMinesandaMSinComputerScienceandEngineeringfromtheUniversityofSouthFlorida.Mr.Gage'sresearchinterestsincludemobilerobots,sensing,articialintelligence,faulttolerance,computersecurity,andcomputergraphics.Hehastaughtcoursesinrobotics,computersecurity,andprogrammingattheUniversityofSouthFloridaandReykjavkUniversity.


xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 001488423
003 fts
006 m||||e|||d||||||||
007 cr mnu|||uuuuu
008 041012s2004 flua sbm s000|0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0000465
035
(OCoLC)56801129
9
AJT5717
b SE
SFE0000465
040
FHM
c FHM
090
TK7885
1 100
Gage, Aaron.
0 245
Multi-robot task allocation using affect
h [electronic resource] /
by Aaron Gage.
260
[Tampa, Fla.] :
University of South Florida,
2004.
502
Thesis (Ph.D.)--University of South Florida, 2004.
504
Includes bibliographical references.
500
Includes vita.
516
Text (Electronic thesis) in PDF format.
538
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
Title from PDF of title page.
Document formatted into pages; contains 127 pages.
520
ABSTRACT: Mobile robots are being used for an increasing array of tasks, from military reconnaissance to planetary exploration to urban search and rescue. As robots are deployed in increasingly complex domains, teams are called upon to perform tasks that exceed the capabilities of any particular robot. Thus, it becomes necessary for robots to cooperate, such that one robot can recruit another to jointly perform a task. Though techniques exist to allocate robots to tasks, either the communication overhead that these techniques require prevents them from scaling up to large teams, or assumptions are made that limit them to simple domains. This dissertation presents a novel emotion-based recruitment approach to the multi-robot task allocation problem. This approach requires less communication bandwidth than comparable methods, enabling it to scale to large team sizes, and making it appropriate for low-power or stealth applications.Affective recruitment is tolerant of unreliable communications channels, and can find better solutions than simple greedy schedulers (based on experimental metrics of the time necessary to complete recruitment and the total number of messages transmitted). Experimental results in a simulated mine-detection task show that affective recruitment succeeds with network failure rates up to 25%, and requires 32% fewer transmissions compared to existing methods on average. Affective recruitment also scales better with team size, requiring up to 61% fewer transmissions than a greedy instantaneous scheduler that has an O(n) communications complexity.
590
Adviser: Murphy, Robin.
653
recruitment.
emotions.
multi-agents.
robotics.
690
Dissertations, Academic
z USF
x Computer Science and Engineering
Doctoral.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.465