Modeling preferences, strategic reasoning and collaboration in
agent-mediated electronic markets
Citation for published version (APA):
Robu, V. (2009). Modeling preferences, strategic reasoning and collaboration in agent-mediated electronic
markets. Technische Universiteit Eindhoven. https://doi.org/10.6100/IR642817
DOI:
10.6100/IR642817
Document status and date:
Published: 01/01/2009
Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be
important differences between the submitted version and the official published version of record. People
interested in the research are advised to contact the author for the final version of the publication, or visit the
DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page
numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners
and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research.
• You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please
follow below link for the End User Agreement:
www.tue.nl/taverne
Take down policy
If you believe that this document breaches copyright please contact us at:
openaccess@tue.nl
providing details and we will investigate your claim.
Graduationcommittee:
Prof.dr.A.G.L.Romme(chair)
Prof.dr.HanLaPoutr´e(promotor)
Corereadingcommittee:
Prof.dr.NicholasR.Jennings(Univ.ofSouthampton,UK)
Prof.dr.DavidC.Parkes(HarvardUniversity,US)
Prof.dr.TonG.deKok(TechnischeUniversiteitEindhoven)
Prof.dr.EmileH.L.Aarts(TUEindhoven&PhilipsResearch)
Additionalmembers:
Prof.dr.CatholijnJonker(TechnicalUniversityofDelft,NL)
Prof.dr.CeesWitteveen(TechnicalUniversityofDelft,NL)
SIKSDissertationSeriesNo.2009-19
TheresearchreportedinthisthesishasbeencarriedoutatCentrumvoorWiskundeen
Informatica(CWI),AmsterdamundertheauspicesofSIKS,theDutchResearchSchoolfor
InformationandKnowledgeSystems.
ISBN:978-90-386-1816-6
AcataloguerecordisavailablefromtheEindhovenUniversityofTechnologyLibrary
PublishedbyTechnischeUniversiteitEindhoven,2009
PrintedondemandthroughLulu.com
Copyright c 2009byValentinRobu.Allrightsreserved.
Coverillustrations:
Front:Afractalmodelofthebrain-courtesyofSvenGeier.
Agent-Mediated Electronic Markets
PROEFSCHRIFT
terverkrijgingvandegraadvandoctoraande
TechnischeUniversiteitEindhoven,opgezagvande
rectormagnicus,prof.dr.ir.C.J.vanDuijn,voor
eencommissieaangewezendoorhetCollegevoor
Promotiesinhetopenbaarteverdedigen
opdonderdag2juli2009om16.00uur
door
ValentinRobu
Firstandforemost,I'mdeeplygratefultomyadvisor,HanLaPoutr´eforhavingmeashis
PhDstudentandforhispatientguidancealltheseyears. Ourdiscussionsprovidedmewith
lotsofinsightsandmotivation(attimeswhenitwaslacking),andhadadeningcontribution
inshapingmyresearchinterestsandcareer. IhadagreatdealtolearnfromHan,notonly
aboutwhatitmeanstogetsolidresearchresults,butalsoaboutacademiclifeingeneral.
Also,aspecialthankyoutoCatholijnJonker,who,asmy Masterthesissupervisor
attheFreeUniversityAmsterdamintroducedmetotheworldofmulti-agentsystems,and
toautomatednegotiationinparticular.Herenthusiasmanddedicatedguidancewerehighly
motivatingindeterminingmetopursueaPhDaftergraduation.
Iamalso gratefultoNickJenningsandDavid Parkes,my externalthesiscommitteee
membersforthecarefulreadingofmydissertationandtheirsuggestions,whichmadethe
presentation ofthisthesis much stronger. Furthermore, Iwouldliketothank professors
Ton de Kok, Emile Aarts andCees Witteveenforreading my thesis andtheirinsightful
comments. I alsoacknowledgeprofessorIoanAlfredLetia(TUCluj-Napoca, Romania),
whorstrecommendedmetopursueamasterdegreeintheNetherlands.
ThisthesiswouldnotbewhatitiswithouttheexcellentcollaboratorsandcolleaguesI
hadovertheyears.Iwillbeginbymyco-authors(and,attimes,co-advisers)atCWI:Koye
Somefun,PieterJan'tHoen,SanderBohte,HanNoot,PeterBosmanandEnricoGerding
(whomIhadthepleasureofmeetingagaininSouthampton).Thebreadthoftheirresearch
interestsandtheideasweexchanged(someofwhichmaterializedinpublications)werean
assettomyresearchcareer.
Also,manythankstoprof.CeesWitteveenandthegroupfromDelft:MathijsdeWeerdt,
TomasKlos(whowasalsomyroom-mateatCWIforayear),RenzeSteenhuizen,Yingqian
Zhang,SiccoVerwerandTamasMahr.Theinvitationtolecturethere(part-time)forayear
wasagreatlearningexperienceforme,sinceIgottoseersthandthechallengesinvolved
inteachingsideofacademiclife. AlsoKoenHindriksandDmytroTykhonovfromDelft
providedmanyinterestingdiscussions.
AnotherdeningmomentinmyPhDresearchwasattendanceoftheSantaFeComplex
SystemsSummerSchool.Iamindebttoallteachersandfellowparticipants,butin
partic-ulartoHarryHalpin(EdinburghandMIT)andHanaShepherd(Princeton). Theexcellent
collaborationforgedwiththemledtoaveryinterestingandsuccesfulllineofresearchon
collaborativetagging,whichresultedinachapterofthisthesis.Furthermore,Ihadthe
andhisenthusiasmandinsightsmotivatedmetolearnalotaboutwebscience.
Iwouldalsoliketothank TakayukiIto(Nagoya University andMIT),Minjie Zhang
(WallongongUniversity,Australia),ShaheenFatima(Liverpool),TokuroMatsuo(Nagoya)
andHiromitsuHattori(KyotoUniv.-whoalsovisitedusatCWI).Ihadthepleasureof
co-organisingwiththemasuccesfulworkshoponcomplexautomatednegotiations(thisyear,
forthethirdyearrunning),aswellascoordinatingonanAAMASconferencetutorial.
IamalsogratefultothepeoplefromAlmendeBV,Rotterdam,whoofferedmea
year-longinternshipduringmymasterthesisdays, ofwhichImentionhere: PeetvanTooren,
HansAbbink,TamasMahr,JPLarsenandStefanKroon. Besidesimprovingconsiderably
my programmingskills, thisinternshipgavemeagreatperspectiveonthe challenges
in-volvedindeployingagentsystemsinpractice,whichIwouldnevergetfromapurely
aca-demiccareer. ThelivelyargumentswithPeetalsotaughtmealotabouttheimportanceof
keepinganopenmindanddefendingthepracticalrelevanceofone'swork. Also,Iwould
liketothankWillem-JanvanSchijndel,fromourprojectpartnerVosLogistics,forhistime
inworkingwithustogetaconcreteapplicationfromourworkonauction-basedallocation.
DuringmytimeatCWI,Ihavebeenpriviligedtosupervisetwoexcellentmaster
stu-dents: Sandervander Putten(TechnologyManagement, Eindhoven)andLonnekeMous
(Econometrics, Univ. Rotterdam). Both hadexcellentresults(includingcumlaude
grad-uationsandprizesforbestthesisintheeldsintheNetherlands). Sanderhelpedusclose
the conceptualgapbetweenthe way AIresearchersandlogistic managerssee theworld.
TheworkwithLonnekeonusingpricedoptionstosolvetheexposureprobleminsequential
auctionsdenitelyledtooneofthemostinterestingdirectionsofworktoemergefromthis
thesis,onethatIdenitelyintendtopursefurtherinmyresearchwork.Moreover,Lonneke
isagreatpersontoworkwith,andIrememberfondlythetravelsthroughthenarrow(and
sunny)streetsofLisbonwithLonnekeandAnke.
WhichbringsmetooneofthemostimportantgroupsofpeopleImetduringmyPhD.
I'mrefering,ofcourse,totheothermembersofthegangoffourPhDstudentsofHan:
AnkeHutzenschreuter,IvanVermeulenandMengxiaoWu(ofwhich,inthatorder,therst
twowerekindenough tobemy paranimfem,while the lasttwowere myroom-matesat
CWIforthepast3years).Ourdiscussionscoveredeverythingfrombargainingmodelsand
distributedpatientschedulinginhospitalstofoodrecipesandthevirtuesofwheatbeer.Also,
ourculinaryexplorationsofthedelightsofAmsterdamwillbeveryfondlyremembered.
Also,I'dliketothankallmyfriendsinAmsterdam,startingwithmyRomanianfriends:
Tudor,Radu,Ghita,VladandLiviu,whohavebeenagreathelpandsupportwhenmoving
tothe Netherlands, andwerealwaysthereto shareabeerinCafe Uilenstede(andmove
furniture,anotherconstantinthelifeofaPhDstudent). Also,manythanks tomy many
otherDutchfriends:Benno,Abel,Richard,Leopold,Jan,Onno,RuggieroandJoost. You
showedme thatdoingaPhDinAmsterdamcanhaveitsfairshareofexcitement,besides
fromtheprofessionalexcitementofbrowsingthelatestAAMASconferenceproceedings.
Finally,my familywhohavehelped methroughthisallalong, especiallymymother
1 Introduction 1
1.1 Agentsandelectronicmarkets . . . 1
1.2 Negotiation(bargaining)vs.auctionprotocols . . . 2
1.3 Designingforstrategicbehaviour:marketmechanismsvs. individualagent strategies . . . 3
1.4 Modelingpreferencesandutilitiesinagent-mediatedmarketsettings . . . . 5
1.5 Emergenceofcollaborationandstructureinmulti-agentsystems . . . 6
1.6 Positioningofthecontributionsofthisthesis. . . 7
1.7 Modelingofcombinatorialpreferences(multi-issueormulti-item)in bilat-eralnegotiations. . . 8
1.7.1 Pareto-optimaloutcomesinmulti-issuenegotiation . . . 9
1.7.2 Modelingmulti-attributenegotiationwithincompletepreference in-formation . . . 11
1.7.3 Non-linearandcombinatorialpreferencesinnegotiation . . . 11
1.7.4 Modeling multi-itemnegotiationsoverk-additiveutilityfunctions usingutilitygraphs . . . 13
1.7.5 Individualpreferencesandsocialinuence . . . 14
1.7.6 Learningthe structureofutilitygraphsusedinmulti-item negotia-tionthroughcollaborativeltering . . . 15
1.8 Preferencesunderuncertaintyandbiddinginsequentialauctions . . . 16
1.8.1 Sequentialauctionsandtheexposureproblem. . . 16
1.8.2 Designingsequentialauctionstrategiesforrisk-aversebidders . . . 17
1.8.3 Optionsmechanismsinsequentialoptions . . . 18
1.8.4 Usingpricedoptionstosolvetheexposureproblem . . . 19
1.9 Applicationstotransportationlogistics . . . 20
1.9.1 Auction-basedallocationoftransportationloadsinmulti-party trans-portationlogistics. . . 20
1.10 Preferencesinsocialwebcommunitiesandonlinemarkets . . . 21
1.10.1 Thecomplexdynamicsofcollaborativetaggingsystems . . . 22
1.10.2 Anempiricalanalysisofsponsoredsearchmarkets . . . 23
1.11 Structureofthethesis . . . 23
1.12 Publicationsrelatedtoeachchapter . . . 26
I Modelingcombinatorialpreferencesinbilateralnegotiation
(bar-gaining) 31
2 AnAgent ArchitectureforCooperativeMulti-Attribute NegotiationWith
In-completePreferenceInformation 33
2.1 Introduction . . . 33
2.2 Themulti-attributenegotiationmodel . . . 34
2.2.1 BidUtilityDeterminationandPlanningComponent . . . 35
2.2.2 TheAttributePlanningComponent . . . 36
2.2.3 TheTargetEvaluationPlanningComponent . . . 37
2.2.4 EstimationofOpponent'sParametersComponent. . . 39
2.2.5 GuessCoefcientsComponent. . . 40
2.3 Implementation&Experimentalvalidation . . . 41
2.3.1 Experimentalset-up . . . 41
2.3.2 Anexamplenegotiationtrace . . . 42
2.3.3 Comparingtracesfromthesametestset . . . 45
2.3.4 Comparingresultsfromalltestsets . . . 48
2.3.5 Human-computerexperiments . . . 49
2.4 Discussion. . . 50
2.5 Conclusions . . . 51
3 ModelingComplexMulti-IssueNegotiationsUsingUtilityGraphs 53 3.1 Introduction . . . 53
3.2 Thenegotiationsetting . . . 57
3.2.1 NetUtilityfunctionsofBuyerandSeller . . . 57
3.2.2 Usinggainsfromtradeasefciencycriteria . . . 58
3.2.3 Outlineofthenegotiationsettingandprotocol. . . 58
3.2.4 Assumptionsaboutbuyerknowledge . . . 60
3.2.5 Top-levelnegotiationalgorithmusedbytheseller . . . 60
3.3 Decomposableutilityfunctionsandtheirgraphicalrepresentation . . . 61
3.3.1 Thek-additiveutilityform . . . 62
3.3.2 Usinggraphstomodelcomplexutilityfunctions . . . 63
3.4 Negotiationheuristicsbasedonutilitygraphs . . . 64
3.4.1 Selectingthebestcounteroffer. . . 65
3.4.2 UpdatingSub-utilityFunctions. . . 69
3.5 Constructingthestructureofutilitygraphsusingaggregatenegotiationdata 72 3.5.1 Collaborativeltering:briefintroduction . . . 72
3.5.2 Overviewofourlteringandnegotiationapproach . . . 73
3.5.3 Minimalsuper-graphforaclassofbuyers . . . 74
3.5.4 Extractinginformationfromconcludednegotiationdata . . . 75
3.5.5 Computingthesimilaritymatrices . . . 76
3.5.6 Buildingthesuper-graphofbuyerutilities . . . 77
3.5.7 MinimizationofexpectedlossinGainsfromTradeascut-offcriteria 78 3.6 ExperimentalAnalysis . . . 79
3.6.3 Set-upandresultsfordifferentgraphstructures . . . 83
3.6.4 Experimentalset-upandanalysisofthecollaborativelteringmodel 87 3.6.5 Resultsfromretrievalexperimentsusingcosine-basedvs. correla-tionbasedsimilarity . . . 89
3.6.6 Experimentalresultsforselectinggraphcut-offnumberofedgesin themaximalgraph . . . 91
3.7 Discussion. . . 94
3.7.1 Comparisontootherautomatednegotiation(bargaining)approaches 94 3.7.2 Relationtographicalutilitymodelsandpreferenceelicitation . . . 97
3.8 Conclusionsandfuturework . . . 98
II Preferencesunderuncertaintyandstrategicreasoningin sequen-tialauctions 103 4 Designingbiddingstrategiesinsequentialauctionsforriskaverseagents 105 4.1 Introduction . . . 105
4.1.1 Goalsandorganisationofthischapter . . . 107
4.2 ModelingUtilityFunctionsUnderRisk . . . 107
4.2.1 Theimportanceofriskaversionindecisionmaking:anexample . . 110
4.3 Biddinginsequentialauctionswithcomplementarities . . . 111
4.3.1 Optimalbiddingpolicyforsequential2ndprice(Vickrey)auctions. 113 4.3.2 Optimalbiddingpolicyforsequential1stpriceauctions: numerical solutions . . . 115
4.3.3 Biddingstrategyformultiplecopyauctionsequences . . . 117
4.4 Experimentalanalysis. . . 118
4.4.1 Experimentalhypotheses . . . 119
4.4.2 Experimentalsetup . . . 119
4.4.3 Experimentalresultsforone-typeitemauctions . . . 120
4.4.4 Resultsforoneitemanddifferentauctionlengths . . . 121
4.4.5 Settingwithdifferentitemtypesandmorecomplexpreferences . . 123
4.4.6 Multipleitemsetting:hypotheses . . . 125
4.4.7 Resultsfortwo-itemcase. . . 127
4.5 Conclusionsandfurtherwork. . . 127
5 UsingPricedOptionstoSolvetheExposureProbleminSequentialAuctions 129 5.1 Introduction . . . 129
5.1.1 Options:basicdenition . . . 130
5.1.2 Relatedwork . . . 131
5.1.3 Outlineandcontributionofourapproach . . . 132
5.2 Expectedprotforasequenceofnauctionsand1synergybuyer . . . 133
5.2.1 Protwithnuniquegoodswithoutoptions . . . 133
5.2.2 Protwithnuniquegoodswithoptions . . . 135
5.3.1 Whenagentsarebetteroffwithoptions . . . 137
5.3.2 Synergybuyer'sprot-maximizingbid . . . 145
5.4 Simulationofamarketwithasinglesynergybuyer . . . 151
5.4.1 Synergybuyer'sbidstrategy . . . 152
5.4.2 Experimentalresults:marketentryeffectforonesynergybuyer . . 155
5.5 Settingswithlongersequencesofauctionsandeffectofauctionorder . . . 158
5.6 Multiplesynergybuyers . . . 160
5.6.1 Twosynergybuyersinteractingindirectlythroughtheexerciseprice level. . . 161
5.6.2 Directsynergybuyercompetitioninthesamemarket . . . 162
5.6.3 Largersimulationwithrandomsynergybuyers'marketentry. . . . 163
5.7 Discussionandfurtherwork . . . 164
III Emergenceofcollaborationandsocialpreferences inweb com-munitiesandonlinemarkets 167 6 EmergenceofConsensusandSharedVocabulariesinCollaborativeTagging 169 6.1 Introduction . . . 169
6.1.1 TaggingversusTaxonomiesontheWeb . . . 170
6.1.2 OverviewofRelatedWork . . . 172
6.1.3 TheTripartiteStructureofTagging . . . 175
6.1.4 Organizationofthechapter. . . 177
6.2 DetectingPowerLawsinTags . . . 177
6.2.1 PowerLawDistributions:Denition. . . 177
6.2.2 EmpiricalResultsforPowerLawRegressionforIndividualSites. . 178
6.2.3 EmpiricalResults forPowerLaw Regression UsingRelative Fre-quencies. . . 180
6.3 TheDynamicsofTagDistributions. . . 181
6.3.1 Kullback-LeiblerDivergence:Denition . . . 181
6.3.2 ApplicationtoTagDynamics . . . 182
6.3.3 EmpiricalResultsforTagDynamics . . . 182
6.3.4 Examiningthedynamicsoftheentiretagdistribution . . . 184
6.4 ConstructingTagCorrelationGraphs . . . 185
6.4.1 Methodology . . . 185
6.4.2 Constructingthetagcorrelation(folksonomy)graphs . . . 186
6.5 Identifyingtagvocabulariesinfolksonomiesusingcommunitydetection al-gorithms . . . 189
6.5.1 Usingcommunitydetectionalgorithmstopartitiontaggraphs . . . 190
6.5.2 Edgelteringstep . . . 191
6.5.3 Normalizedvs.non-normalizededgeweights . . . 191
6.5.4 Thegraphpartitioningalgorithm. . . 192
6.5.5 Graphpartitioning:experimentalresults . . . 194
fromsearchenginequerydata . . . 196
6.6.1 Datasetandmethodologyemployed. . . 198
6.6.2 Discussionoftheresultsfromthequerylogdataandcomparison . 198 6.7 ConclusionsandFutureWork. . . 200
6.8 Acknowledgments . . . 201
7 TheComplexDynamicsofSponsoredSearchMarkets:AnEmpiricalStudy 203 7.1 Introduction . . . 203
7.1.1 Thedataset . . . 204
7.2 Complexsystemsanalysisappliedtothewebandeconomics . . . 204
7.2.1 Powerlaws:denition . . . 205
7.3 Inuenceofdisplayrankonclickingbehaviour . . . 206
7.3.1 Resultsondisplaypositionbiasandinterpretation . . . 207
7.4 Marketstructureattheadvertiserlevel . . . 209
7.4.1 Distributionofimpressionsvs. distributionofclicksforthetop ad-vertisers . . . 209
7.4.2 Distributionofmarketshareperdisplayrankposition. . . 210
7.5 Usingclickdatatoderivesearchtermrecommendations . . . 211
7.5.1 Derivingdistancesfromco-occurrenceinsponsoredclicklogs . . . 212
7.5.2 Constructingkeywordcorrelationgraphs . . . 212
7.5.3 Graphcorrelationgraphs:results. . . 213
7.5.4 Automaticidenticationofsetsofkeywords . . . 214
7.5.5 Discussionofgraphpartitioningresults . . . 215
7.6 Discussion. . . 216
7.6.1 Contributionofthechapter&relatedwork . . . 216
7.6.2 Futurework. . . 217
7.6.3 Acknowledgements. . . 217
IV Conclusions 219 8 Conclusionsandfurtherwork 221 8.1 Overviewoftheresearchcontributionsperchapter . . . 221
8.2 Furtherwork . . . 224
V Anindustrialapplicationcase 229
A APlatformforAuction-BasedAllocationofLoadsinTransportationLogistics 231
A.1 Introduction . . . 231
A.1.1 Themulti-partylogisticsdomain . . . 232
A.1.2 Companyprole . . . 232
A.1.3 Automatingmulti-partylogisticsusingagents . . . 233
A.1.4 Goalsofthiswork . . . 233
A.2 Overviewofthebusinesscaseandourplatform . . . 235
A.2.1 Generatingtransportationorders . . . 235
A.2.2 Computingpricesandcosts . . . 237
A.3 Auctionprotocolanddesignoftheauctioneeragent . . . 237
A.3.1 Auctionset-up . . . 238
A.3.2 Auctionsforloadswithashortleadtime . . . 238
A.3.3 Auctionsfororderswithalongertimehorizon . . . 239
A.3.4 Totalcapacityofloadstobegeneratedperday . . . 239
A.3.5 Auctioneeruserinterface . . . 240
A.4 Automatedbidders:descriptionanduserinterface . . . 241
A.5 Thecarrieragents:descriptionanduserinterfaces . . . 242
A.5.1 Transportationmodelandcarriercosts. . . 243
A.5.2 Penaltyforlatedeliveries. . . 244
A.5.3 Informationsuppliedaboutothercarriersduringthecompetition . . 244
A.5.4 Planningandbiddingdecisionsupportinterface. . . 245
A.6 Outlineofpreliminaryhumanbiddingresults . . . 246
A.7 Discussion. . . 247
Introduction
1.1 Agentsand electronicmarkets
Multi-agentsystemsare oneofthemost promisingnewtechnologiestoemergeinrecent
decades, atthecrossroadsbetweenseveraleldssuchas articialintelligence,distributed
systems,economicsandevensociology. Someauthors[16,231]haveoutlinedavision,in
whichmanyofthetasksperformedtodaybyhumansaredelegatedtointelligent,autonomous
andproactiveprograms,genericallycalledsoftwareagents. Asystemcomposedofseveral
suchagentsiscalledamulti-agentsystem(MAS).
Electronicmarketsrepresentkeycoordinationmechanismsinmulti-agentsystems.They
allowpartiestoefcientlyallocateresources,tasksandcapabilitiesinlargedistributed
sys-tems,composedofself-interestedagents. Therapidriseinelectroniccommerceand
mar-keting,logistics,distributednetworks(amongmanyothers)havemadethedevelopmentof
agenttechnologiescapableofautomatingsuchprocessesincreasinglyimportant.For
exam-ple,electroniccommercehaswitnessedanexponentialincreaseinthevalueofthegoodsand
servicessoldonlinejustinthepastfewyears. Itisnotjustthesaleofphysicalgoodsthat
hasgreatlyincreased,butalsothesaleofvirtualservices,suchasscreenattentionspace
fordisplayingadvertisingine-commerce,orkeywordhitsbysurfersusingsearchengines.
Suchsalesrequirefrequent,repeatedinteractions,whicharethetypeofprocessesthatare
likelytobenetmostfromautomationusingsoftwareagents.
Therearemanychallengesthat designersofagents actinginelectronicmarketsmust
face. Perhapsthemosteasilyrecognizedchallengeindesigningandusingsuchasystem,
isthelackofcentralizedcontrol. Agentsareautonomousactors,thattaketheirown
deci-sions,ratherthansimplyexecutingoperationsassignedtothembyanoutsideprocess(such
asobjectsorwebservicesdo).Furthermore,perhapsmoreimportantly,theyareoften
self-interestedactors,whosegoalsandobjectivesmaynotmatch. Forexample,inoptimizing
alloca-tionofloadsfromtheperspectiveofeachcompanymaybeverydifferentthantheoptimal
allocationfortheentiresystem. Inotherapplicationscenarios,suchasonlineadvertising,
agentsrepresentingdifferentcompaniesactivelycompeteforvirtualcommodities,suchas
consumerattentionspace.
AnimportantchallengeinMASisthepresenceofuncertainty,i.e. incompleteor
im-perfectinformation,bothregardingthemarketenvironment,thepreferences,strategiesand
behaviourofthe otheragents and,sometimes, evenuncertaintyinspecifyingtheagent's
own preferences. Furthermore, unlikeassumptions commonlymadein gametheory, the
agentsareboundedrationalactorsandoftenhavetomakedecisionsinlimitedtime,under
riskaversionorbasedonotherconstraintsimposedbytheirownersorthemarket
environ-ment.Moreover,theopposingagentsparticipatinginthesamemarketmayalsobebounded
rationalandevenact `irrationally, whichmakes modelingtheagent'sown optimalor
rationalbehaviourinsuchasettingevenharder.
Anotherimportantapproachinthestudyofagent-mediatedelectronicmarketsare the
so-calledcomplexsystemstechniques.Theaimofsuchapproachesistoexaminehoworder
and structurecanemerge ina large systemcomposed ofmanyautonomousentities(i.e.
agents),actingindependently,withoutanycentralcontrollertoprovidecoordination. The
recentsurgeofinterestinsystemssuchaswebcommunitiesandonlineelectronicmarkets,
wherestructureemergesoutofindividualagentdecisions,makessuchquestionsincreasingly
important.
1.2 Negotiation(bargaining)vs. auctionprotocols
Negotiation,verybroadlydened,istheprocessbywhichagroupofagentscommunicate
withoneanothertotrytoreachagreementonsomematterofcommoninterestc.f.[111,
189].Automatednegotiationhasbeenattheforefrontofresearchinterestsinthemulti-agent
researchcommunityeversincethebeginningoftheeld[129,189].
Oneofthemaindistinctionlinesbeingdrawninexistingliteratureisbetweenautomated
negotiation(bargaining)protocolsandauctionprotocols[111,147]. Bargainingisalways
adecentralizedprocessandistypically(thoughnotnecessarily)based onanalternating
offers-typeofprotocol. Someauthors[111,115,175](amongothers)arguethat
bargain-ingdoeshavesomeadvantagesoverauctions,especiallyinmultipleissuecases, inwhich
thereisincompleteinformationabouttheopponentpreferences(orevenuncertaintyabout
the agent'sownpreferences)andthespace ofpossibledealstobeexploredisverylarge.
Bargainingalsoallowsmoreexibilityinhowthenegotiationismodeled,aswellasa
de-greeofself-interestonthepartoftheagents. Somesources[87,115,175,179]evenargue
that,inelectroniccommerce,multi-issuenegotiationshouldbemodeled,atleastpartially,
asacooperativeprocess,becausesellershaveaninterestinmaintainingagoodrelationship
andthelong-termsatisfactionoftheirbuyers.
Auctions,ontheotherhand,followprotocolswithxedrules,thattypicallyrelyona
managementscience[190],butalsoarticialintelligenceandtheoreticalcomputerscience
[27,55,58,59,81,142,174,194].Theyhavebeenthemethodofchoiceforautomating
elec-tronicmarketplaces.
Inourwork,wehavelookedatbothmechanisms,fordifferentsettings.Ourinitialwork
inthetopicstartedondesigningefcientbilateralnegotiationmechanisms,rstforlinear
utilityfunctions(Chapter2ofthisthesis),thenforcomplex,interdependentutilityfunctions
(Chapter3). Wehavealso lookedatdesigningbiddingstrategiesforsequentialauctions,
insettingsnotpreviouslyconsideredinexistingliterature,suchasthecasewhensomeof
theagents areriskaverse(Chapter4), orwhenoptionsareauctionedinstead oftheitems
themselves(Chapter5).
Thisthesistakesanengineeringapproach,meaningthatweaimtoidentifyopen
prob-lems,andthenengineerandvalidatesolutionsforthem. Wedostudytowhatdegreethese
problemscanbeaddressedusingananalytical,mathematicalapproachinsofaraspossible.
However,manynegotiationandauctioningprocessesaretoocomplextobesolvedusinga
purelyanalyticalapproach,asisnormallythecaseforreal-worldproblems. Insuchcases,
experimentalvalidationisapromisingalternative,whichwasusedextensivelyinthisthesis.
Acommonthreadrunningthroughthe researchpresentedinthisthesisisthatwetake
theheuristicapproachtothedesignofbiddingagents. Thatis,we focusourattentionon
designingthestrategiesthatbiddingagentsusetobidornegotiateinagivenmarket
envi-ronment(usuallyonewidelyencounteredinpractice),notthemarketprotocolitself.Thisis
animportantdistinction,asexplainedinthenextsection.
1.3 Designingforstrategicbehaviour: marketmechanisms
vs. individual agent strategies
With the growth of interest in electronic markets, several research lines have emerged,
proposingdifferentapproachestomodelingstrategic,self-interestedbehaviourwhen
allo-catingresourcesortasksamongasetofagents.Oneofthemostpromisingsuchapproaches
iscomputationalmechanismdesign-ortobemoreprecise,thatpartofmechanismdesign
theorythatconcernsdesignofelectronicmarkets.
Mechanismdesigninitiallydevelopedasabranchofalgorithmicgametheory[168].
Ba-sicallydened,mechanismdesignisconcernedwithdeningtherulesofthegame(i.e.the
marketmechanism),suchthattheoutcome(i.e. nalallocationoftheitems,togetherwith
the correspondingpayments)guaranteescertaindesiderata (i.e. properties).
Commonly-citeddesideratainclude,forexample:Pareto-optimality,efciency,budgetbalanceor
indi-vidualrationality[58].
Besidesfromthesegame-theoreticdesiderata,computationalrequirements(i.e.the
com-putationtimeormemoryneededtondsuchamechanism)oftenplayanimportantrole.The
biddingagents,asthestructureofthemarketprovidesbidderswithanequilibriumbidding
strategy.Differentequilibriumconceptexist,varyingintheirstrength,e.g.dominant
strate-gies,ex-postefcient,Bayesian.Themostdesirablemarketmechanismsarestrategy-proof
mechanisms,i.e.thosemechanismsinwhichtruthfulbiddingisthedominantstrategy.
Themechanismdesignapproachhasproventobeverysuccessfulinmanyapplications.
However,thereexistsawiderangeofpracticalsettingsforwhichitisunrealistictoassume
thatonecandesignacompletelynewmarketmechanismfromscratch. Furthermore,many
mechanismsproposedbythislineofresearchofteninvolveallocationandbiddingrulesthat
carefullydesignedandmathematicallysound,butmaybecounter-intuitiveforhumanusers
ofthesystem.
Moreover,inmanyreal-lifeallocationproblems,thereismorethanonemarketanagent
can/shouldparticipatein,andthestrategicbehaviouracrossmarketbordersbecomesthe
crucialissue.Evenifanagenthasanoptimal(e.g.dominant)biddingstrategyineachofthe
marketsheparticipatesin,whencoordinatingthebiddingindifferentmarkets,hisoptimal
strategymaybeverydifferentfromthedominantstrategyforeachmarkettakeninisolation.
One suchexampleis bidding ina sequenceof secondprice(i.e. Vickrey)auctions, for
agents whichhavecomplementaryutilities overtheitems beingoffered[27,89,187,217]
1
. Althoughtheagents insuchasequentialauctionhaveadominantbidding strategyin
eachindividualVickreyauctiontakeninisolation,biddingoptimallyinasequenceofsuch
auctionsisacomplexdecisionproblem,andthebidsplacedintheoptimalsequentialbidding
policymaydifferconsiderablythanthedominantbidsintheindividualauctions
2 .
Similarly,arelateddecisionproblemisfacedbyagentsbiddinginasetofsimultaneously
ascendingascendingEnglishauctions,whenagentshavecomplementaryutilityfunctions[1,
184].Whereasasimple,dominantbiddingstrategyexistsforeachEnglishauctiontakenin
isolation,determiningtheoptimalbiddingstrategyfortheentiresetisachallengingproblem,
forwhichnodominantstrategyresultsareknown. Intuitivelyexplained,itisdifcultfor
abiddingagenttodistributethe additionalcomplementarityvalueacross asequenceora
setofsimultaneousauctions, becausean agentcanonlyknowifhe canbenetfrom the
complementarityonceallitemsinthedesiredbundlehavebeenacquired(i.e. onceallthe
auctionsclose).
Whilethecomputationalmechanismdesigncommunityhasbegantoaddresssomeof
thesechallenges,throughsuchtechniquesasonlinemechanismdesignoradaptive
mecha-nismdesign[33,74,92],theseapproachesstillimposeseveralrestrictionsonthestructureof
theproblem,andformanymarketsettingwidelyusedinpractice,nodominantequilibrium
strategiesare[yet]knowntoexist.
Theresearchperformedforthisthesismostlyfollowstheothermaindirectionofresearch
1
Whilethesenotionswillbeformallydenedlater,intuitively,acomplementaryvaluationimpliesthatanagent assignsacombinationofitemsasuper-additiveutility(i.e.autilityhigherthanthesumoftheitemstaken individ-ually),whilesubstitutabilityimpliesasub-additiveutility(seeChapter3)
2
Inasequenceofauctions,thisistruewhenevereithercomplementarityorsubstitutabilityeffectsexistbetween items,orthereareotherpreferenceconstraintstoaccountfor,suchasbudgetconstraintsoraversiontorisk(c.f. [27,89,187]andChapters4and5.
designtheoptimalagentstrategiesforbiddinginsuchmarkets?
Workondesigning biddingagents' strategiesalso hasarelativelylonghistoryin the
MAS community[1,27,81,83,89,184,188,223](amongmayothers). Severalplatforms
havebeenproposedtoenablecomparisonofdifferentauctiontradingstrategies(aswellas
learningandadaptationheuristics).Themostwell-knownistheTradingAgentCompetition
(TAC)platform,withits differentversions: TACclassic, TACsupply-chainmanagement
etc.[96,192,227,230]. ThemarketstructurefortheTACcompetitionsisbuilttoresemble
tradingscenariosthatwouldbeencounteredinpractice. Reasoningrequiredforthe
trad-ingagentsintheseplatformscombineselementsofbothefcientbiddinginsequentialand
simultaneousauctions,aswellaslearning,anticipationoffutureorders,inventory
manage-mentetc. Anotherdirectionofwork(mentionedhereforcompleteness)examinesbidding
heuristicsfordoubleauctionsettings[226],whichischaracteristictonancialmarkets[124].
Asalreadydiscussed,inthisthesiswealsotaketheheuristicapproach,andwearemostly
concernedaboutthedesignofagentstrategies,ratherthanthethemarketmechanismitself.
Inparticular,weareconcernedwithoneaspectoftheproblem,whichishowtomodeland
efcientlyusepreferenceinformationoftheagentstakingpartinsuchmarkets.
1.4 Modeling preferences and utilities in agent-mediated
marketsettings
Inbuildingefcient electronicmarkets, the methodofmodelingandreasoningabout the
preferences ofparticipatingagents is akeymodelingchoice. Some sourcescall
model-ingpreferencesofbuyersandprovidersremainstheAchilles'heelintheapplicationof
multi-agentresourceallocationtoindustrialprocurementsettings[46].Thereare,however,
considerabledifferencesastotheirmeaningofthetermspreferenceandutilityin
dif-ferentsources intheeconomicsandmulti-agentsystemliterature. Inthe broadestsense,
preferencesexpresstherelativeorabsolutesatisfactionofanindividualwhenfacedwitha
choicebetweendifferentalternatives[46]. Inthisthesis,webroadlydistinguishbetween
twobroadclassesofconceptsofpreferenceorutility:preferencesincombinatorialsettings
(i.e. toreasonregardingmultiplecriteria ormultipleitems)andpreferencesunder
uncer-tainty.
Mostexistingliteratureonmulti-agentresourceallocationandmarketmechanisms
con-siderscombinatorialpreferencesandutilities. Combinatorialpreferencesareeither
multi-item(i.e.involveexpressingpreferencesovercombinationsofitems)ormulti-issueor
multi-attribute(e.g. involvecombinationsofattributesforthesameitem, e.g. colour,priceand
mileageforausedcar,totaketheexampleusedinChapter2). Moreover,inmany(ifnot
most)realisticapplications,itisreasonabletoexpectthattherearecomplexdependencies
betweenattributesoritems,andthechoiceinonemayaffectthechoicemadeforasubsetof
non-linear,combinatorialpreferencesinmarketsituationsisacomplexproblem,whichwe
discussinSect.1.7below.
Anotherperspectiveondeningpreferencesconsidersthecomplexdecisionsanagent
faces,notsomuchwithrespecttospecifyingdesiredcombinationsovermultipleitemsor
issues,butwithrespecttouncertaintyaboutthefuture.Thisappearstobeastandard
under-standingofpreferenceinsomeelds,suchaseconometrics.Forinstance,ina2006MIT
textbookoneconometricanalysisofauctiondata[171],thechapterregardingpreferences
dealsexclusivelywithpreferencestowardsrisk. Inthisthesis,we considerboth
perspec-tivesonpreference,andbothtypesofmarket-basedinteractiondiscussedabove: bilateral
negotiationandauctions,althoughfordifferentproblems.
1.5 Emergenceofcollaborationandstructureinmulti-agent
systems
Asnoted inSection1.1,another importantproblemarisinginmulti-agentsystems is the
lackofcentralizedcontrol.Nevertheless,manysystemsoccurringinreal-lifethatonewould
intuitively recognizeas multi-agent, exhibit aremarkabledegreeofstructure, although
theylackanyrecognizablecentralauthorityorcontroller.Instead,orderseemstoemerge
fromthedecentralizedactionsofmanyautonomousagents,actingindependentlytosatisfy
theirowninterest.Examplesofsuchsystemsinclude:theformationofequilibriaandpricing
structureinmarkets(aphenomenarstreferredtobyAdam'sSmithastheguidinghand),
emergenceofstablevocabulariesinhumanlanguages(butalsointaggingsystems)[40,93,
206],formationofstablegroupsinonlinesocialnetworks[131]etc.Thisraisesquestionsnot
onlyregardingtheexistenceandpropertiesofsuchstablestructures,butalsothedynamicsof
theprocess,i.e.howdotheyform,especiallyinanenvironmentwithnocentralinformation
sourceand/orself-interestedparties.
Oneoftherecentlyemergingeldsthataimstostudysuchphenomenaiscomplex
sys-temstheory[11,160,228]. TheseminalworkofRobertAxelroodontheevolutionof
co-operation[7]markedaturningpoint,sinceitshowed,throughcomputersimulations,how
cooperationcanemergeinamulti-agentsystem,evenintheabsenceofacentralauthority.
Arelated discipline that aims to examinecomplex-systems typephenomena through
large-scalesimulationsisagent-basedcomputationaleconomics[218].Therehasbeenmuch
workrecentlyinthisarea.Forexample,researchershavesimulatedthedynamicsofarticial
agentsocieties[7],stockmarkets[6],andevenentireeconomies[37,68].Thedevelopment
ofthewebhasgivenanewstimulustothiswork,andresearchershavebuiltcomplex
sim-ulationsoftheemergenceofsocialnetworks[131],onlinemarketsystems[54]orarticial
languagesandsemioticdynamics[42,210].
But,perhapsmoreimportant,theemergenceofthesocialwebprovides,fortherst
time,theopportunitytotestthesehypothesesempirically,onreal-worlddata.Infact,while
re-datasetsgeneratedbytheactionsofverymany(thousands,orinsomecasesevenmillions)
ofwebusers.
Interestingly,manyofthe effects foundresembled closely what was hypothesised in
complexsystemstheoryfromthebeginning. Inparticular,itappearsthatthereare
impor-tantnetworkeffectswhenmanyuserscollaborateonlineandmakedecisionsinan
on-linecommunityormarketplace. Whatthismeans,basically,isthattheactionsandchoices
madebyprevioususersmayconsiderablyinuencethechoicesmadebyfutureusers.This
typeofself-reinforcingfeedbackloopoftengivesrisetotheso-calledpowerlaw
distri-butions[11,41],whicharecharacteristicoflarge-scalesystemsthatcanbecharacterisedas
complex.
Thisthesismakestwoimportantcontributionstounderstandingtheemergenceofsocial
structureinsuchlarge-scale,decentralizedsystems. Oneiscollaborativetagging(results
presentedinChapter6)andsponsoredsearchmarkets(Chapter7).
1.6 Positioningof thecontributions ofthis thesis
The previousdiscussion identiedsomeimportantopenchallenges inunderstandingand
designingmulti-agentsystems:
Complexityofrepresenting(andreasoningwith)complexpreferences.Theseinclude
bothcombinatorialpreferencesandpreferencestowardsriskanduncertainty.
Strategicreasoningofagentsbasedonthesecomplexpreferences,especiallyforcases
whenagentsareself-interested.
Lackofcentralcontrol,andespecially,theemergenceofcooperationintheabsence
ofacentralauthority.
Inthisthesis,weaimtomakeseveralcontributionstothestateoftheartinunderstanding,
modelingandsolvingthesechallenges,asfollows.
PartIofthethesisismostlyconcernedwiththeissueofmodelingcombinatorial
prefer-ences(multi-issueormulti-item)inbilateralnegotiations. Chapter2considershow
prefer-enceinformationcanbeefcientlyusedinanegotiationmodelinwhichuserspreferences
areexpressedoverseveraldiscreteattributesandonecontinuousattribute(price).Chapter3
considershowcomplex,multi-issuenegotiationsovermanybinaryitemsorbundlesofitems
canbemodeledusingutilitygraphs. PartIalsodealswithsomeissuesrelatedtostrategic
reasoning,since,althoughbilateralnegotiationisoftenapartiallycooperativeprocess,there
isanimportantdegreeofself-interestinvolvedonthepartofthebargainingagents.
PartIIofthe thesiscanbeseenas mostlydealingwithpreferencesinuncertain
envi-ronmentsandstrategicreasoning,inparticularthestrategicreasoningofagentswith
pricedoptionsmechanismcanhelpsolvetheexposureproblembiddersfaceina
sequen-tial auctionssetting(andimplicitly, thestrategicreasoningitinvolvesduringthe bidding
process). Buttheissueofpreferenceisalsoimportantinthispartaswell,althoughinthe
formofagentpreferencestowardsrisk,whenfacedwithanuncertainfuture.Inthiscontext,
Chapter4examineshowanagent'spreferencestowardsriskaffectshis/heroptimalbidding
policyandresultingmarketallocation.
Finally,PartIIIofthethesisstudiestheissueoflackofcentralisedcontrol,more
particu-lartheemergenceofcollaborationandstructureinalargemulti-agentsystem,intheabsence
ofacentralcontroller.Forthispart,weuselargescale,empiricaldatafromtwoimportant
socialwebapplications:collaborativetaggingandsponsoredsearchmarkets.Whileinin
collaborativetagging,forinstance,theissueofstrategic/game-theoreticreasoningdoesnot
playadirectrole(sincethereisnocompetingallocationofsomescarceresource),stillthe
issueofhowagentstakedecisionsisacrucialonetomodel. Arguably,thereisalsoa
con-nectiontotheissueofpreference,sincethroughtheirchoiceoftagsandlinkstoclickagents
expressanimplicitopinion(whichmayormaynotbeinuencedbythatofotherusers).
Therestofthisintroductorychapterisorganisedasfollows. Inthefollowingsections
(Sect.1.7-1.10),wegivemoredetaileddescriptionsoftheproblemsinthisspacewhichwe
aimtoaddressinthisthesis,aswellasbriefabstractsofourresultsforeach.InSection1.11
wegivetheoverviewofthestructureof[therestof]thethesis.Section1.11alsosummarizes
thestructureofthethesisthroughadiagram,suchastomoreintuitivelyhighlightandexplain
therelationsthatexistbetweenthedifferentchapters.Theintroductionconcludeswithalist
ofresultingrefereedpublicationsrelatedtoeachchapter.
1.7 Modeling of combinatorial preferences (multi-issue or
multi-item)inbilateralnegotiations
Therearemanywaystoexpressachoicebetweenmultipleoutcomesdenedinmulti-agent,
economicsandAIliterature. Ataxonomyofpreferencesusedinthemulti-agentliterature
wouldinclude:
Qualitativepreferences: Nonumericalutilityvaluesareassignedtooutcomes,only
valuelabelssuchasgood,verygood,unsatisfactoryetc.
Quantitativepreferences: Preferenceoveroutcomesare expressedinthe formofa
utilityfunction(tobedenedbelow).Notethatsometimesqualitativeandquantitative
preferencestakentogetherarecalledcardinalpreferences.
Ordinalpreferences:onlyanordercanbespeciedbetweenranking(i.e. throughan
asymmetricandtransitivepreferencerelationbetweenalternatives).
ausercanassignanumericalvalue(eitherutilityormonetaryvalue)topossibleoutcomes
(orcombinationofitems),andthediscussioninthefollowingsectionsofthisintroduction
referstothequantitativecase.
However,notethat,ifonedenespreferenceasanychoicebetweenseveraloutcomesor
alternatives,theconceptofpreferencecanbeconstructedasbroader. Forexample,actions
suchas choosingatag that otherusershavealso selectedinthe past(see Chapter6), or
clickingonthelinkatthetopofalist,inordertosavereadingtime(seeChapter7)may
beseenasexpressinganimplicitpreference. WereturntothisideainSection1.10,when
describingthe chaptersinPartIII ofthethesis -the discussion inthefollowingsections
referringtothecaseofquantitative,economicpreferences.
Thebasisofquantitativepreferencemodelingisutilitytheory. Followingtheworkof
KenneyandRaiffa[179],manymulti-issuenegotiationandresourceallocationmodelsuse
autilityfunction,whichmapstheoutcomespaceoverasetofissues(attributes,criteria)to
autilityvalue,whichisfrequently-thoughnotnecessarily-scaledbetween0and1. The
crucialthingtonoteisthatRaiffa'smodelsandmuchoftheinitialresearchonmulti-issue
negotiationconsiderslinearlyadditiveutilityfunctions,i.e.eachissue/attributeunder
nego-tiationisassignedaweight,andtheutilityofeachpossibleoutcome/contractiscomputedas
aweightedsumovertheissuesundernegotiation.
Aspecialsubclassofquantitativepreferencefunctions,whichisimplicitlyusedinmost
ofexistingauctionliterature, are the so-calledquasi-linear preferences. Thisbasically
meansthattheutilityoftheagentsisexpressedinmonetaryterms(asanamountofmoney),
as opposedtoutils(i.e. conventionalunits,usuallyscaledbetween0and1). Thiscanbe
viewedasarestrictionforsomesettings,asrealutilityfunctionsovermonetaryendowments
areknowntobeconcave,i.e. humansareknowntohaveadecreasingmarginalutilityfor
money(see,e.g.[171]).
1.7.1 Pareto-optimaloutcomesinmulti-issuenegotiation
As shown in[179],multi-issueandmulti-attribute negotiation models arefundamentally
differentfromsingle-issuenegotiation (suchas bargainingoveraprice). Multi-issue
ne-gotiationsrepresent non-zerosumgames, inthesense thatitis possibletond mutually
benecialtrade-offsbetweentheissuesundernegotiationsuchas toincreasethegainsfor
bothparties. Raiffaalsoshowsthatthemoreasymmetricpreferencesbetweenthe
negotia-torsare,thehigherthepotentialformutuallybenecialtrade-offsbetweentheissues.
Themaincriteriatomeasurehowefcientanagreement(orcontract)isthe socalled
Paretoefciency. AnoutcomeissaidtobeParetooptimalifitisnotstrictlydominatedby
anyotheroutcomeinthepreferencesofboth(orall)sides(agents)inanegotiation. That
means,therearenotrade-offspossiblethatwouldincreasetheutilityofoneagent,without
making anotheragentworse off. The setofall Pareto-optimalpointsform theso called
thatoutcome/contractis.
Animportantconceptindesigningmulti-issuenegotiationmodelsistheuncertainty
re-gardingopponentpreferences,denedhereastheamountofinformationregardingthe
op-ponentpreferencesavailablewhenmakingnegotiationoffers.
Directvs.indirectrevelationmechanisms
Theliteratureonagent-mediatedelectronicmarketsidentiestwomainapproachesbywhich
agentscansharetheirprivatepreferenceinformation:
Directrevelationmechanisms.Directrevelationmechanismsarebasedonthe
revela-tionprinciple[58]. Basicallyexplained,therevelationprinciplestatesthatany
allo-cationmechanismwithacertainequilibriumcanbetransformedintoanother
mecha-nism,inwhichatrustedcenteraskstheagentstotruthfullyrevealtheirpreferencesand
implementstheoriginalequilibrium(andallocation)ontheirbehalf.Thismeansthat
typically(thoughnot exclusively),directrevelationleadstoacentralizedallocation
mechanism,suchasacombinatorialauction.
Indirectrevelationmechanisms:Inthistypeofmechanism,theagentsarenotassumed
todirectlyrevealtheirpreferencestotheotheragents,butcommunicatetheir
prefer-encesthroughouttheircounter-offers(ortheirbids,foranauction). Forinstance,in
abilateral,multi-issuenegotiationoverthe saleofacar(see [115,116]and
Chap-ter2ofthisthesis), theagents donotdirectlyreveal toeachotherhowmuchthey
arewillingtopaytogettheirfavouritecolourortheirfavouriteaccessories(e.g. CD
player,airconditioning)installed,butinpractice,thiscanbededucedindirectlyfrom
theoffers/counter-offerstheymake. Similarly,anagentrepresentingacustomerona
largeelectroniccommercewebsite(see[185,186,220]andChapter3)doesnothave
torevealallhispreferencestothemerchant,butthemerchant(whomayormaynot
alsoactastheauctioneer)canlearnpartoftheirpreferencesfromprevious
counter-offers.
Basically,inthisthesiswetaketheindirectrevelationapproach,aswearguethisismore
realisticinmanyreal-lifeapplications,inwhichonlyalimiteddegreeoftrustexistsbetween
partiesinsharinginformationandnofullytrustedthirdpartycanbeestablished. The
rea-sonsforthismaybeendogenoustothenegotiationmechanism(e.g. thereisnooptimal
incentivecompatiblemechanismandtheopponentmayuseanyinformationsuppliedtoget
abetterdealforhimself)orexogenous(e.g.itmaybeundesirabletohavetospecify
prefer-encesoverthewholesetofalternatives,duetoprivacyconcernsorfuturebusinessinterests).
Furthermore,forcomplexnon-linearpreferences,therearealsopreferenceformulation
andcommunicationcosts. Aswe showinChapter3, becauseofboundedrationalityand
communicationability,itisoftendifcultforanagentherselftoformulateandcommunicate
searchintomodelingpreferencesinagent-mediatedmarkets,followedbybriefdescriptions
ofthecontributionsmadeinthisthesistoopenproblemsintheeld. Therefore,some
sec-tionsoftheintroductionpresentimportantconceptsfromageneralpointofview,whileother
sectionsdescribehowthesegeneralconceptswereextendedbyourownresearch,described
inthechaptersofthisthesis. Thegoalistoallowthereadertogetabetterunderstandthe
positioningandcontributionofourworkwithrespecttothestateoftheartintheeld.
1.7.2 Modelingmulti-attributenegotiationwithincompletepreference
information
Animportantdirectionofworkintheliteratureonmulti-issuenegotiationishowtodesign
efcientbargainingstrategiesinsettingswhenagentsdonothaveanyinformationaboutthe
opponent's(i.e. negotiation partner's)preferences. Theymayhave, however,someprior
knowledgeaboutthedomaintheyarenegotiatingabout. Thispriordomainknowledgecan
be, forexample,fuzzylogicdistancesbetweenattributes,suchas theperceptualdistance
betweendifferentcolours(suchasin[71,163]),oranorderingbetweenasetofqualitative
attributelabels(suchasgood,standard,meageretc),inourresearch(seeChapter2).
TheworkpresentedinChapter2and[115,116]considerssuchanincomplete
informa-tionnegotiationmodel.Theaimofthismodelistoinvestigatetherolethatpartiallyrevealing
preferenceinformationcanimprovetheoutcomeofamulti-attributenegotiation.Asa
prac-ticaldomaincase,we consideredabilateralnegotiationbetweenabuyer(customer)anda
seller(cardealer)overthesaleofacar.Thenegotiationisnotexclusivelyonprice,butalso
on thequalityoftheaccessories whichthe dealerhastoinstallinthecar togetthe deal
done(suchasaCDplayer,extraspeakers,airconditioningandtowhedge). Inthissetting,
weshowthatitispossibleforbothpartiestoreachclosetoPareto-efcientagreements,by
revealingonlypartial(i.e.incomplete)informationabouttheirpreferencesofthenegotiation
partner. Furthermore,weproposedanovelguessingheuristic,bywhichanagentusesthe
historyofopponent'sbidstopredicthis/herpreferencesinordertoproposebetterdeals.
1.7.3 Non-linearandcombinatorialpreferencesinnegotiation
Acrucialprobleminapplyingmulti-issueormulti-itemnegotiationmodelsinmanyrealistic
settingsisthefactthattheremaybecomplexinter-dependenciesbetweendifferentissues,
leadingtonon-linearpreferencesorutilityfunctions[108,126,138,186].Theproblem
ap-pearsbothwhenconsideringintegrative, multi-issuenegotiations,as wellas negotiations
overbundlesofitems [46,186]. Inboththesecases, it isimportanttoallowforconcise
representationsoftheutilitiesoverpossibleoutcomes
3 .
3
Theproblemisinfact,two-fold.Firstthereisthecomplexityrelatedtopreferenceformulationofcombinatorial preferences,aswellasoneofpreferencecommunicationcomplexity.
Theeasiestwaytorepresentpreferencesistoenumerateallpossibleoutcomes(or
com-binationsofgoods),togetherwiththeirutilityvalueforthosegoods(monetaryorotherwise).
Thisiscalledtheexplicitformofpreferencerepresentation(orbundleform).Theexplicit
formisfullyexpressive, inthesense thatanyutilityfunctionmaybedescribedbylisting
allpossiblecombinationsandtheirvalues. Itis,however,impracticalformostnon-trivial
settings,asthenumberofdescriptionswouldbeexponentialinthenumberofresources(e.g.
foronly50binaryissuesoritems,2
50 >10
15
valueswouldneedtobeassigned-see
Chap-ter3).Thishaspromptedanotherimportantdirectionofresearchinelectronicmarkets,that
ofdesigningmoreconciseutilityrepresentation(orpreference)languages.Thereareseveral
classesofsuchpreferencelanguages:
Biddinglanguages,whicharetypicallyusedincombinatorialauctionstoallowagents
toformulatetheirbids(and,implicitly,communicatetheirpreferencestothe
auction-eer).Somespecicbiddinglanguagesinclude:
The ORlanguage: The agentcan specify an array ofvaluations over
differ-entsubsetsofitems inagivenbundleofitems. Thevalueofanycombination
canthenbecomputedasthemaximalvaluethatcanbeobtainedasasumover
disjointsubsets specied[46]. Forexample,inthebid: < fI
1 g;3 > OR < fI 2 g;3> OR <fI 3 g;3 > OR < fI 1 ;I 2
g;8 >expressesthatthebidderis
willingtopay3foreitherI 1
;I 2
;I 3
or11forall3items(inthiscase,itisbetter
totakethevalueofthesubset<fI 1
;I 2
g;8>thanthevaluesofeachindividual
itemseparately).BecausetheORdependencyisnotexclusive,theORlanguage
cannotexpresssubstitutabilitydependencies,i.e. itcannotexpressthefactthat
gettingacombinationofitemshaslesserutilitythanthesumofindividualitems.
Intheaboveexample,itisnotpossiblefortheagenttoexpressthatheisonly
willing topay4ifhegetsbothI
1 ;I 2 . Ifthebid< fI 1 ;I 2 g;4 >wereadded
tothesetofbidsplaced,thentheauctioneerwouldsimplymatchthebidsover
the individualitems < fI 1
g;3 >and< fI 2
g;3 >(as anyterms ofthe OR
dependencymaybechosen).
ExclusiveOR(i.e. XOR)biddinglanguage[194]-isanalternativetoOR,in
whichallcombinationbidsareassumedtobemutuallyexclusive.Forexample,
intheaboveexample,abidsuchas:<fI 1 g;3>XOR<fI 2 g;3>XOR< fI 3 g;3>XOR < fI 1 ;I 2
g;4>meansthattheagent(bidder)caneitheruse
onlyoneitemfromI
1 ;I
2 ;I
3
withautilityof3,orthecombinationoffI
1 ;I
2 g
witha utilityof4, butnoother combination(so, e.g. gettingbothI
1
andI
3
wouldstillonlyhavetheutilityof3).XORisfullyexpressive,inthesensethat
itcanrepresentanymonotonicutilityfunction.However,XORmayhaveahigh
communication/elicitationcost,evenforsimplesettings.Anexampleisthe
util-ityfunctionthat,foranysetRofitemsI 1
;:::I n
2R ,simplycountsthenumber
ofitemstheagentowns-i.e. u(R ) = jR j. Suchafunctionwouldrequirean
exponentialnumberofbidsintheXORlanguage,butonlyalinearnumberin
ORlanguage.Thisisbecause,usingXOR,allcombinationsspeciedare
consideredmorenaturalwaytorepresentpreferences,thereexistsalineofwork
thataimstoextenditsexpressiveness,withoutrequiringanexhaustivelistingof
XORbids[169].
Weightedpropositionalformulasandstraightlineprogramsareotheralternativesto
representingcomplexpreferences,whichmakeuseoflogicalformalisms. For
exam-ple,weightedpropositionalformulasarederivedfromaqualitativeformofpreference
representation,inwhichthepreferencesoftheagentareexpressedasgoals. Inthe
weightedcase(unlikeinpurelyformallogicapproaches),goalscanbeassigneda
util-ityweightifsatised. Wedonotdealwiththiskindofpreferencelanguagesinthis
paper,butthereadercanconsult[139]forthefulldetailsofthisapproach.
Thek-additiveform[46,55,186](alsocalledthepolynomialform[133])isanother
naturalandconcisemethodtorepresentcombinatorialpreferences. K-additive
func-tionscanencodesynergy(complementarityorsubstitutability)effectsbetweensubsets
ofuptokitems. Forinstance,ifwedenotebyx
1 ;:::x
n
theinstantiationofthesetof
nitems,theexpressionfora3-additiveutilityform(i.e.takingamaximumk=s)is:
U(x 1 ;:::;x n )= X 1in i x i + X 1i;jn i;j x i x j + X 1i;j;k n i;j;k x i x j x k (1.1) Wherex 1 ;::;x n
representsavectorof1and0,denotingwhetheranitemis(orisnot)
consideredinthecombinationbeingevaluated,thereals
1 :::
n;n;n;n
arethe
param-etersofthefunction,whilek(samekasink-additivity)isthemaximumrankofthe
polynomial,i.e. allthepolynomialtermshavingarankabovek
max
havethe
coef-cients= 0. Linearlyadditivefunctionsformasubclassofthek-additiveclass,as
denedabove,fork
max
=1. Thek-additiveformisfullyexpressive,forunbounded
k.Thismeansthat,ifkissufcientlylarge,itcanbeusedtoexpressanyutility
func-tionovera given,nite, binarysetofitems. Inpractice,although(as discussedin
Chapter3thisthesis),inorderforthisrepresentationtobecomputationallyuseful,the
maximumrankofthepolynomialkisgenerallyassumedtobeboundedtoalimited
value(e.g.2-4,asdiscussedinChapter3).
1.7.4 Modelingmulti-itemnegotiationsoverk-additiveutilityfunctions
usingutilitygraphs
InChapter3ofthethesis,weconsiderthecaseofmodelingcomplexbilateralnegotiations
overasetofmultiple,binaryissues(whichcanalsorepresentabundleofitems). Fromthe
conciserepresentationformsdiscussedinthetaxonomyfromSect. 1.7.3,theoneweused,
as wefounditmostnaturalinthe contextofthemulti-issuenegotiation,isthek-additive
form.Thisrepresentationisanaturalextensionoflinearutilitymodelsalreadyusedinmuch
nonlinearityinpreferencemakesthebargainingproblemconsiderablyharder.Forexample,
inEq.1.1above,thecaseofk
max
=2isalreadymuchharderthank
max
=1.
Multi-issuenegotiationwithnon-linearutilityfunctionsisknowntobeacomplex
prob-lem, even forthe caseof binaryissues[108,126,207]. The state ofthe art inthiseld
proposescomplexsolutionsthatinvolveamediator,aswellastechniquessuchassimulated
annealing[126]oreconometricmethods[207]thatareeithercomputationallyexpensiveor
do notscalewellforsettingswithmanyissues. In[186](correspondingtothe rstpart
inChapter3)weintroduceanovelutilitygraphformalismformodelingnonlinear(i.e.
k-additive)preferences,andweshowhowsuchgraphscanbeusedtomodelandlearn
oppo-nentpreferencesincomplex,multi-issuenegotiations.Utilitygraphsareoriginallyinspired
fromprobabilisticgraphicalmodels,buttheyencodeutilities,ratherthanprobabilities.The
mainideabehindourapproachistousethestructureofthegraphstorestricttheopponent
modelingandsearchtothemostpromisingregionoftheutilityspace. Aselleragentcan
startanegotiationwithanapproximationoftheutilityfunctionofatypicalrandombuyerin
theformofamaximalutilitygraph,andthenrenethismodelbasedonthecounter-offers
he observesduringthe negotiation. Inourcase,the initialutilitygraph reectsthe prior
informationthatthesellerhasabouthowtheutilityfunctionofarandombuyerisstructured,
inordertohelpinthesearch.
Animportantquestionis,ofcourse,howdoestheselleracquirethisinitialbuyerutility
graph approximation. Onesolution istoassume somepriordomain knowledge, suchas
plausibleconstraintsonthe shapeautilityfunctioncouldtake(whichmaybereasonable
forsomesettings). Fore-commercedomains,wehaveproposedanotheralternative:using
collaborativelteringonprevioussalesdata,thatwillbepresentedafterthediscussionin
thenextsubsection.
1.7.5 Individualpreferencesandsocialinuence
In the previousdiscussion oncombinatorial preferences, preferences are dened from a
single-agentperspective,meaningthattheutilityofanyagentisassumedtobeprivateand
independentofwhatotheragentsmaydesire. Otherwiseput,ifasellerencountersabuyer
andnegotiateswithhim thecongurationofaproductorthecompositionofabundleof
items,hewillassumethatthepreferredcombinationsofthisparticularbuyerarecompletely
independentofwhatotherbuyersencounteredwantedinthepast.Thisis,infact,astandard
assumptioninmuchofnegotiationandauctiontheory.
However,existingpracticeinelectroniccommercesuggested,fortheChapter3ofthis
thesis,analternativeapproach.Thesuccessofsocialsearchinprovidingonlinebuying
rec-ommendationsprovidesconsiderableevidencethatpreferencesarenotstrictlyindependent,
butareinsomewayclustered. Considerforexample,the caseofAmazon.com,whohas
severalmillionbooktitlesinitscollection. Eliciting,foreachindividualcustomer,his/her
preferencesoverthesebookstoproposeacceptablebundlesforthebuyerswouldbeanearly
impossibletask.However,Amazonimplicitlyassumesthatifalargenumbersofcustomers
aswell. Forinstance,ifacustomerbuysabookontravellingtoPortugal,theAmazon
en-gineassumeshemayalsobeinterestedinabookontraveltoSpain,sincemanycustomers
encounteredinthepastshowedinterestedinboth.Thereforeproposingadeal(e.g.postage
reduction,orasmalldiscount)maybeagoodwaytoincentivisethecustomertobuyboth
booksfromthesite.
Notethatthisdoesnotalwayshavetobeacorrectprediction: infact manycustomers
maynotbeinterestedintheexactcombinationproposed.However,itdoesprovideagood
approximationinsearchingthespaceofacustomer'spreferences,evenifthecustomerwas
neverencounteredbefore.
Traditionally,researchinmulti-issuenegotiationdoesnot explicitlymodelthissocial
dimensionofcustomerpreferences,orconsidertherolethatsocialinuenceplaysonthe
structureofutilityfunctions.Weshowthathavinganexplicitrepresentationthatrelatesthe
twoelds(inourcaseintheformofutilitygraphs)allowsustoconsiderablyimprovesearch
inanonlinenegotiationsetting. Furthermore,theinteractionbetweentheseeldsdoesnot
have tobe one-way: negotiation also hasa lotto addto web-based recommendationin
electroniccommerce. Throughaniterativenegotiationprocess,theinitialproposals(based
onanonymous,aggregatepreferences)canbecustomizedtothepreferencesofaparticular
customer,basedontheindirectrevelationmadethroughhis/hercounter-offersinnegotiation.
1.7.6 Learningthestructure ofutilitygraphsusedinmulti-item
nego-tiationthroughcollaborativeltering
Ourapproachtomodelingopponentpreferencesinnegotiationmakesuseoftheabove
in-tuition. Chapter3ofthisthesisproposesanovelcollaborativelteringmethodbywhich
previouslyconcludednegotiationdatacanbeusedtoconstructtheinitialapproximationof
theutilitygraphofarandombuyerthatthesellercanuseinlaternegotiations. Theseller
willthenadjust(learn)thevaluesinthegraph,foreachspecicnegotiation,basedonthe
counter-offersthebuyermakes,untilanagreementisreachedoverthebundlecombination.
Therefore,wetakewhatcanbedescribedasatwo-stepcustomizationapproach: initially,
anapproximationofthemaximalstructureofautilitygraphforarandombuyerisobtained
usingcollaborativelteringonallconcludednegotiation data(whichdoesnothavetobe
buyer-specic).Then,thisdealisrenedthroughoffersandcounteroffersduringthe
nego-tiationwithaspeciccustomer.
Weshowthat thecombined approachcanenablebuyersandsellers toreachefcient
agreementsevenincomplexnon-linearsettings,involvingonlyindirectrevelation(although
therearesomeassumptionsregardingthemaximalcomplexityoftheutilitygraphsthata
buyercanhave).Oneofthecontributionsofthisapproachtothestateoftheartinautomated
negotiationisthatitprovidesalinkbetweenthecustomizationtechniquesusedinmulti-issue
ormulti-itemnegotiationandthoseusedincollaborativelteringandsocialcomputing.In
andshorternegotiationsforcomplex,non-linearutilitysettingsthanwasreportedinother
research[126].
1.8 Preferences under uncertainty and bidding in
sequen-tialauctions
Inthepreviousdiscussion,wehavemainlydiscussedtheconceptofpreference(orutility)
inthecontextofintegrativenegotiation,inwhichtheallocationsforallitems(orissues)is
agreedatthesametime. Thus,whenanagentspeciesapreferencebyassigninga
mone-taryvaluetoacombinationofitems, heisbiddingforanentirecombination,andthereis
nouncertaintythathewillnotgetsomeofitemsintheagreedconguration,iftheseller
ac-ceptstheoffer.Thisisareasonableassumptionforintegrativenegotiationandcombinatorial
auctions(wheretheallocationforallitemsisnegotiatedsimultaneously).However,itdoes
not holdforotherwidelyusedallocationmechanisms, suchas sequential/simultaneously
ascendingauctions[27,89,184,217]orone-by-oneissuenegotiations[72].
Inthissection(correspondingtoChapters4and5ofthethesis),weconsiderthecase
whenagentshavetobidsequentiallyitemssoldindifferentauctions,withoutknowingwith
certaintythattheywillgettheentirecombinationofitemstheydesire. Insuchcases,
eco-nomictheory identiesanother important class ofpreferences, preferences towards risk.
Riskaversionisaveryimportantpartofeconomictheory-infact,a2006MITtextbookon
theeconometricsofauctiondata[171],thechapteronpreferencesisbasicallyconcerned
withpreferencestowardsrisk.
Thewayeconometrictheorymodels riskaversionisthroughtheso-called
Neumann-Morgensternpreferencefunctions,inwhichthe utilityderivedbyanagentfrom acertain
amountofmoneyisnotalinearfunction,butaconcaveone.Otherwisestated,utility
func-tionsarenotquasi-linear,inthesensethattheutilitythateachagentderivesfromanamount
ofmoneyisnotdirectlyproportionaltotheamountpaid/received.
Inthefollowing,webrieydenetheexposureprobleminsequentialauctions,therole
thatriskaversionplaysinthebiddingdecision,aswellasanoverviewofthecontributions
inChapters4,5andAppendixA.
1.8.1 Sequentialauctionsandtheexposureproblem
AsshowninSect.1.3above,therearetwomaindirectionsofresearchintheapplicationof
agentsystemstoauctionmarkets.Oneconcernsthedesignoftheauctionmechanismitself,
suchthatparticipantagentshaveadominantbiddingstrategy(usually,todeclaretruthfully
theirvalues),aswellascertainproperties,suchasefciency,individualrationalityorbudget
balance. However,formanymarketdesignsthatare necessarilyencounteredinpractice,
tions.
Asshownin[27,89,187]andChapters4and5ofthisthesis,themainproblemthata
bidderfacesinasequentialauctionistheexposureproblem. Informallydened,the
expo-sureproblemmeansthatanagenthastocommittobuyinganitem,beforehe/shecanbesure
thathewillabletosecureotheritemsinhisusefulsetorbundle(denedasthesetofitems
thatgiveshimapositiveutility). Ifshefailstoacquirethisbundle,thenhemakesaloss.
Hence,wesaythattheagentisexposedtotheriskofaloss.
Mostofthe modelsthatstudybiddingauctionbiddingstartfrom theassumptionthat
agents havequasi-linearutilityfunctions. Basicallydened,quasi-linearityassumeseach
agenthasasetofpayoffsthathe/sheassignstoanycombinationsofitems. Thesepayoffs
are,formanyofthemodelsstudied,private: theyarenotknowntotheotherparties. The
utility thatan agentget from participating inthe auction is assumedproportionalto the
differencebetweenhis/herprivate payoffandtheamount hepaystoacquire theitems in
question,inotherwords,itisdenedstrictlyinmonetaryprot/lossterms.
Thisquasi-linearityofpreferences assumption,whilewidelyusedandvalidformany
business models andsettings, doesnot universallyhold. Inmanyreal-lifesettings,even
assumingitistruethatagentshaveprivatevaluesfordifferentsubsetsofitemsunder
nego-tiation,protandlossarenotjudgedinthesameterms. Makingalossfromaninteraction
(i.e. payingmorethanhis/herprivatepayoffvalue)isnotproportionalasgainingthesame
amountasprot.Inotherwords,agentsarerisk-aversetomakingaloss,evenifthepotential
forgainisconsiderablylarger.
1.8.2 Designingsequentialauctionstrategiesforrisk-aversebidders
Prior to the publicationofour research, there hadbeen quite alot ofpreviouswork on
designingefcientbiddingstrategiesforagentsparticipatinginsequential[27,89,217]and
simultaneouslyascending[1,184]auctions.
Whilethisworkreportedsomepositiveresults,animportantlimitationofexisting
litera-tureweexaminedwasthatitdoesnotexplicitlymodeltherisk-takingattitudeofthebidding
agents. Byexplicitlymodel wemeanbuildingaproleofthe agent'sriskpreferences
towardsuncertain,futureoutcomes(suchasthenalallocationofasequentialauction).In
standardeconomictheory,sincetheseminalworkofK.ArrowandJ.Pratt,preferences
to-wardsriskhavebeenconsideredessentialinunderstandingandmodelingdecisionmaking
underuncertainty[5,88,153,171]. Auctionliteraturefromstandardeconomics[158,171]
considersriskaversionan importantprobleminmodelingrealbidderpreferences.
How-ever,theeconomicliteraturethatweareawareofdoesnotconsidersequentialauctionswith
complementarybiddervaluations,exceptperhapsinthesimplestofsettings(becausesuch
auctionsdonothavewell-denedequilibria). Morespecically,unliketheAIcommunity,
researchersineconomicsarenotconcernedwithdesigningautomatedbiddingheuristicsfor
ThemaincontributionofChapter4ofthisthesisismakingalinkbetweenrisk-aversion
models,andthestrategiesthatrisk-aversebidderscanuseinsequentialauctions. First,we
introducetheArrow-Prattriskmodelsfrom economicstotheproblemofmodelingagent
biddingstrategies. Wethenstudythewayinwhichtheperceivedoptimalbiddingstrategy
computedbyariskaverseagent,givenherprobabilisticmodelofthefuture,differsfrom
theoptimalstrategyofariskneutralagent.Wendthatagentsmoreaversetoriskbidmore
aggressively,inordertocovertheirsunkcostsfortheinitialitemsinthesequence.However,
ifthefuturesequenceofauctionsisinitiallyperceivedastoorisky(giventheagent'sinitial
estimationoffutureclosingprices),thebeststrategyavailabletoariskaverseagentissimply
nottoparticipateatall.
Ourexperimentalresultsshowthatariskaversebidderhas,asexpected,alowerchance
toendupwithanincompletebundleofgoods,thusmakealoss.However,whenconsidering
long-termandrepeatedinteractions,suchagentsmake,onaverage,alowerexpectedprot,
becausetheyparticipateinless auctions. Forsomemarketsettings, thisalsoaffects,ina
negativeway,theauctioneerrevenuesfromtheauctions.
Inthefollowingsection,welookatadifferentsideoftheproblemofexposuretoriskof
lossinsequentialauctions,namelywhatcanbedonetoreduceit.
1.8.3 Optionsmechanismsinsequentialoptions
Asdiscussedabove,sequentialauctionsdonotguaranteeadominantbiddingstrategyforthe
agents(unlikethecombinatorialcase). However,theproblemremains,asmanyallocation
problemsoccurringinpracticeareinherentlydecentralizedandsequential.Differentsellers
mayprefer,foravarietyofreasons,toselltheiritemsseparately-oreventhroughdifferent
markets,as thenumberofelectronicauctionsites onlineindicates. Furthermore,inmany
applicationsettings,notallresourcesthataretobeallocatedareknowninadvance,butthey
appeardynamicallyovertime. InChapter5ofthisthesis,westudyanalternativetothis
verydifcultproblemthat,althoughitcannotcompletelyeliminatebidder'sexposure,itcan
signicantlyreduceit:theuseofpricedoptions.
Intuitivelydened, anoptionisacontractbetweenthebuyerandthe sellerofaitem,
wherethebuyerhastheright tochooseinthefuturewhetherornothewillpurchase the
itemagainstthepre-agreedexerciseprice. Theselleristhenboundtosellthe itematthe
demandofthebuyer.Sincethebuyergainsaright,hehastopaytheoptionpriceregardless
ofwhetherhewillexercisetheoptionornot.
Optionsreducetheexposureproblemasynergybuyerfaces.Hestillhastopaytheoption
price,butifhefailstocompletehisdesiredbundle,thenhedoesnotpaytheexerciseprice
as well,andthushelimitshisloss. Theriskofnotwinningsubsequentauctionsispartly
transferredtotheseller,whomaymiss outonthe exerciseprice. However,thesellercan
benetindirectlyfrom theparticipationinthismarketofadditionalcomplementary-value
buyers(alsocalledsynergybuyers),whowouldhaveotherwisestayedout.