On Priing, Inentives and Congestion
Control in Heterogeneous Networks
Gergely Bizók
PhDDissertation
Advisor:
Dr. Tuan A.Trinh
High Speed NetworksLab
Departmentof Teleommuniations andMedia Informatis
Budapest University of Tehnology andEonomis
Budapest, Hungary
2010.
Heterogeneity is inherently present in multiple aspets of thewired and wireless Internet. Un-
derstanding, overomingoreven exploitingthisheterogeneityatdierentlevelsarefundamental
goals for researhersinomputernetworks. Thisdissertation presentsresults on Internet aess
priing, wireless ommunity networks and ongestion ontrol. All proposed approahes share
one thing in ommon: they intend to help a diverse set of network providers, e.g., Internet
Servie Providers, ommunity wireless operators, miro-operators (end-users themselves) and
mobile operators, to addressthe hallenges stemmingfromthe heterogeneity of theirrespetive
networks.
First, I quantify the impat of ustomer loyalty on the priing ompetition between net-
work providers. Ithen proposea priingmehanismfor Internet aess, whih enables network
providers to plantheir revenues, while users an diretlyinuene the implemented billing pol-
iy. Seond, I analyze the eonomi interations in wireless ommunity networks. I show that
properinentivedesignanduserheterogeneityfailitatestheemergeneofatrulyglobalwireless
ommunity, where both users andnetwork providersprot fromthenetwork. Third, Ishowthe
limitationsofexistingTCPversionsindynamimobileenvironments,partiularlyafterasudden
apaityinrease. Motivatedbytheselimitations, Iproposeasimple,end-to-endnon-ongestion
detetion mehanismfor TCP whih solvesthis problemeetively.
AvezetékesésvezetéknélküliInternetalapvet®tulajdonságaasokszín¶ség. Ennekasokszín¶ség-
nek a megértése, legy®zése,s®t, akár kihasználása különböz® szervezésiszinteken aszámítógép-
hálózatokterületén tevékenyked® kutatókegyiklegfontosabb élja. Eza disszertáióaz Internet
hozzáférés árazásával, vezetéknélküli közösségi hálózatokkal és torlódáskezeléssel foglalkozik. A
javasolt megoldásoknak közös a élja, segíteni a különböz® hálózati szolgáltatóknak, ideértve
az internetszolgáltatókat, vezetéknélküli közösségi operátorokat, mikrooperátorokat (magukat a
felhasználókat) és mobilszolgáltatókat hogy a saját hálózatuk heterogenitása miatt felmerül®
m¶szaki éstársadalmi-gazdasági problémákat orvosolnitudják.
Az els® részben a felhasználói h¶ség hatását számszer¶sítem a hálózati szolgáltatók ár-
versenyére vonatkozóan. Egy olyan árazási mehanizmus is kidolgozásra kerül, amely lehet®vé
teszi, hogya hálózatiszolgáltatók megtervezzék bevételüket, míg afelhasználók közvetlenülbe-
folyásolhatjákazalkalmazottszámlázásimódszerkiválasztását. Amásodikrészbenavezetéknél-
küli közösségi hálózatokban fellép® gazdasági kölsönhatásokat elemzem. Megmutatom, hogy
az ösztönz® mehanizmusok helyes tervezése és a felhasználók sokfélesége el®segíti egy valóban
globális vezetéknélküli közösség kialakulását, miközben mind a felhasználók, mind a hálózati
szolgáltatók jól járnak. A záró részben a létez® TCP verziók teljesítménykorlátait vizsgálom
dinamikus mobilhálózatokban, a hirtelen kapaitás-növekedés esetére konentrálva. Továbbá,
mivelazeredményekkedvez®tlenek,egyúj,egyszer¶,végpont-végpont jelleg¶,ahirtelenkapai-
tásváltozástjól kezel®mehanizmust(TCPkiterjesztést) dolgozokki.
Thisdissertation marks the end ofa longjourney (and thestart of another), thereforeIwould
like to pay myrespets to anumberof peoplewhohelped mereah thesummit.
First,Iwould liketo thank myadvisor,Dr. Tuan A.Trinh (HSNLab),for hisguidane and
perseverane. Without hishelpthis dissertation wouldhave neverbeen ompleted.
Myspeialthanksgoesto Dr. András Veres (Erisson Researh). Having aresearhdisus-
sion withhimhasalways been inspiring andfun.
Iowe a greatdeal ofgratitude to Prof. AleksandarKuzmanovi (Northwestern University).
He hastaught menevertoaim low.
I would also like to thank Dr. Róbert Szabó (Head of HSNLab) for being supportive with
this dissertation.
Thanks to the olleagues and fellow students I have worked/rested/suered together with.
Espeially Sándor Kardos, László Toka, Dr. András Gulyás, Balázs Sonkoly, Felíián Németh,
Dr. Zalán Heszberger, Dr. Attila Vidás and Prof. József Bíró at HSNLab; Kristóf Fodor,
Miklós Aurél Rónai, MátéCsorba, Dr. Sándor Palugyai, Ákos Kovás, Zoltán Turányi, Gábor
Németh, Feren Kubinszky, Péter Tarján, Ágoston Szabó, Dr. Balázs Kovás and Dr. András
Valkó at Erisson Researh; and Ionut Trestian, Amit Mondal, Karl Deng and Ao-Jan Su at
Northwestern University.
At last,but foremost,mydeepestthanksaredueto myfamilyfor their ontinuoussupport.
To Mother,for heraring. ToFather, forbeingarole model. ToZsuzsi,forherloveand endless
patiene. Ilove you all.
1 Introdution 1
1.1 Researh Objetives . . . 3
1.2 Methodology . . . 3
1.2.1 Basinotions of gametheory . . . 4
1.3 Struture . . . 6
2 Priing InternetAess under User Inuene 9 2.1 Bakground . . . 10
2.1.1 Interneteonomis . . . 10
2.1.2 Brand loyalty . . . 11
2.2 PriingInternet aessinthe preseneofuser loyalty . . . 12
2.2.1 Motivation . . . 12
2.2.2 Assumptions . . . 14
2.2.3 Inentive to ooperate . . . 16
2.2.4 Dierentiated reservation pries . . . 19
2.2.5 Enhaned modelsof userloyalty . . . 22
2.2.6 Experimental evaluation . . . 24
2.2.7 Disussion onserviebundling . . . 33
2.3 User-inuenedpriing for Internet aessproviders. . . 36
2.3.1 User-inuened priing . . . 36
2.3.2 The votinggame model . . . 37
2.3.3 Equilibrium solutions. . . 40
2.3.4 Distribution of power. . . 44
2.3.5 Disussion onfeasibility . . . 46
2.4 Conlusion. . . 47
3 Inentives for Wireless Community Networks 49 3.1 Bakground . . . 51
3.1.1 Wireless ommunitynetworks . . . 51
3.2.1 Assumptions . . . 54
3.2.2 The mediator . . . 56
3.2.3 The user. . . 57
3.2.4 The Internet Servie Provider . . . 57
3.2.5 Tehnology penetration . . . 58
3.3 Equilibrium analysis . . . 60
3.3.1 One-shot usergame withhomogeneouspayments . . . 60
3.3.2 Evolutionary usergamewith homogeneouspayments . . . 63
3.3.3 One-shot usergame withheterogeneous payments. . . 65
3.3.4 The ISPgame. . . 67
3.3.5 Mediator: anoptimization problem . . . 69
3.4 Experimental evaluation . . . 70
3.4.1 Modeling usermobilityand relevane. . . 71
3.4.2 The evolutionarygame . . . 73
3.4.3 Simulationresults . . . 73
3.5 Conlusion. . . 77
4 Congestion Control in Dynami Mobile Environments 79 4.1 Bakground . . . 81
4.1.1 TCP variants for high-speed networks . . . 81
4.1.2 TCP performaneinwireless mobilenetworks . . . 82
4.2 Measuringhigh-speed TCP performaneduring mobilehandovers . . . 83
4.2.1 Emulationtestbed . . . 83
4.2.2 Handovermodel . . . 84
4.2.3 Measurement results . . . 87
4.2.4 Bueroverow probability at handovers . . . 91
4.3 SpeedDetet: handling suddenapaityinrease . . . 96
4.3.1 The SpeedDetet algorithm . . . 98
4.3.2 Parametertuning . . . 102
4.3.3 Measurement results . . . 104
4.3.4 Simulationresults . . . 107
4.4 Disussion onFAST . . . 109
4.5 Conlusion. . . 110
5.1.1 Contribution . . . 113
5.1.2 Appliation possibilities . . . 114
5.2 Inentivesfor wireless ommunitynetworks . . . 114
5.2.1 Contribution . . . 114
5.2.2 Appliation possibilities . . . 114
5.3 Congestion ontrol indynamimobile environments . . . 115
5.3.1 Contribution . . . 115
5.3.2 Appliation possibilities . . . 116
2.1 Demand funtionsfor
G 1
andG 2
. . . . . . . . . . . . . . . . . . . . . . . . . . . 202.2 Bestresponsefuntionsof
G 1
withtwo players . . . . . . . . . . . . . . . . . . . 212.3 Referene senario: pries, prots andmarket sharesfor 2,3 and10 ISPs (deter-
ministi loyalty, fullyinformed,
L = 1
,equal initial market share) . . . . . . . . . 262.4 Pries,prots andmarketsharesfor2(
L = 0.05
), 3(L = 0.05
andL = 0.15
) and10 ISPs (
L = 0.05
),deterministi loyalty . . . 27 2.5 Pries, prots and marketsharesfor 2 (L = 0.15
),3 (L = 0.15
andL = 1.0
) and10 ISPs (
L = 0.15
),stohastiloyalty . . . . . . . . . . . . . . . . . . . . . . . . . 292.6 Pries,protsandmarketsharesfor3ISPs(
L = 0.05, 0.20, 0.35, 1.0
),1inumbent,2 entrants, deterministi loyalty . . . 30
2.7 Pries, prots and market sharesfor 3 ISPs with fullyinformed (
L = 0.05, 0.35
)and partiallyinformed users (
L = 0.05, 0.35
), 1 inumbent, 2entrants, determin-istiloyalty . . . 32
2.8 Total individual prots for 2 ISPs (
L = 0.05, 0.30, 1.0
), 1 inumbent, 1 entrant,deterministi loyalty . . . 33
2.9 Overallumulative protsof 2 ISPsat dierent loyaltylevels . . . 33
2.10 User-Inuened Priingmehanism (UIP) . . . 37
3.1 Interationsamongusers,ISPsandmediator(strutureoftheStakelbergmodel).
Solidlinesdenotediretinput(ost parameters andrevenue shares),dashedlines
denoteimpliit feedbak(outome ofsubgames, for subgame perfet design) . . . 55
3.2 Expetedrevenueforusersundervariousparameterssettings(defaultparameters:
c o = 10, c i = 5, α = 0.3, ρ = 2
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633.3 Expeted revenue for users: heterogeneous vs. homogeneous payment struture
(all otherparameters xed) . . . 67
3.4 Expeted revenue for ISPs: lowand highrevenue shares(
β
) . . . . . . . . . . . . 693.5 Mediator payos: greedy vs. soial welfare (xaxis: revenue shares
(α, β)
, yaxis:ost parameters
(c i , c o )
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703.6 Geographi distribution ofFON users inBerlin . . . 71
(u = 4, c o = 3, c i = 3, α = 0.5, µ = 0.001)
3.8 Evolutionofthe userpopulationandtheaverageinsiderrelevaneunderdierent
initial onditions (default parameters:
u = 4, c o = 3, c i = 3, α = 0.5, µ = 0.001
) . . 764.1 Measured round-trip timesduring ahandover . . . 80
4.2 Throughput of various TCP variants during intra-system handover (left), inter- systemup-swith (middle) and down-swith (right).. . . 88
4.3 NewReno reeiver-side time-sequene graph of data pakets. There is a bursty lossof pakets around10.14s,whih later areretransmitted. . . 89
4.4 A simplied TCP ontrol loop as seen by the sender. Data pakets leaving the senderS rstgetqueued(andpossiblydropped)atthebottleneklink,thenarrive at thelient C. There,ACK paketsaretriggered that travelbakto thesender. Base propagationdelay(
RT T
base ) isaumulated atsome point duringtheyle. 92 4.5 Buer oupany dynamis with peaks due to link outages (AIMD ongestion ontrol). Dotted line indiates the maximum buer size above whih paketsare droppedfromthe buer. (Note: Therelative widthofthepeaksisillustration; in a typial senarioit isnegligibleompared toT p
.) . . . . . . . . . . . . . . . . . . 934.6 Buer oupation growth around thepeaks. Dashed linesmark buer growth in ase ofno linkoutages. Non-zerogrowth rates arealsonoted. . . 95
4.7 Buer overow probability of NewReno. (Note: Lowest value for
B
max isµT = 2
Mbit= 250
kbyte,belowwhihP (B p ) = 1
.) . . . . . . . . . . . . . . . . . . . . 964.8 Measured buerdynamis (NewReno) . . . 97
4.9 Blok diagramofSpeedDetet . . . 100
4.10 Congestion windows duringhandover . . . 106
4.11 SD versus PNCD:
N = 20, K = 2
. . . . . . . . . . . . . . . . . . . . . . . . . . . 1074.12 SD vs. PNCD:distribution ofdetetion time . . . 108
4.13 Throughputomparison withNewRenoasreferene . . . 109
4.14 RTT-fairness(lean run) . . . 110
4.15 Inter-protool fairnesswithNewRenoasreferene . . . 110
2.1 Payomatrix for thebasi game (inthousanddollars) . . . 13
2.2 Payomatrix for thebrand loyaltygame (inthousand dollars) . . . 13
2.3 Charateristi funtion for the user-inuened priing game (
w 1
andt 1
are the population ratio andtra ratioof heavy users,whilew 3
andt 3
arethose ofthe light users,respetively) . . . 412.4
ψ
-stable pairs inthebalanedregime . . . . . . . . . . . . . . . . . . . . . . . . . 433.1 Notationusedinthemodel . . . 54
3.2 ISPgame payos . . . 58
3.3 Simplied payomatrix for theISPgame . . . 68
4.1 Link parameters of radioaesstehnologies . . . 85
4.2 Emulated systemparameters . . . 86
Introdution
Designing the next-generation global network is a omplex task. What was one onsidered
a pure engineering problem has beome ross-disiplinary: engineers, eonomists, soiologists
and legal experts ooperate to propose eient methods for future networked systems. While
performane enhaning tehniques like ongestion ontrol are still very important as the
popularityof bandwidth-hungry andreal-time appliations andservies (lesharing, streaming
video, et.) isever-growing, soio-eonomi aspetsarebeominginreasinglysigniant [1℄.
Onthe otherhand, heterogeneity isinherently present inmultiple aspets of the wiredand
wireless Internet. Asaesstehnologies, network-enabled devies and end-usersthemselvesare
moreandmorediversied,heterogeneitywillonlygainevenmoremomentuminfuturenetworks.
Therefore, understanding, overoming or even exploiting this heterogeneityat dierent levels is
a fundamental goalfor researhers.
Heterogeneous networks aremanaged bya diverse olletion ofnetwork operators: Internet
ServieProviders,ommunitywirelessproviders,miro-providers(end-usersthemselves),ellular
mobile operators, et. Theseoperatorsfaevarious hallenges bothat thesoio-eonomialand
tehnial levels. This work fouses on a subset of these hallenges. First, the dissertation aims
to quantify the impat of user-behavior on Internet aess priing, and to propose a diretly
user-inuened priing mehanism. Seond, inentive shemes are studied whih enable the
global proliferationofa ommunitywireless network. Third,a novelTCP extension isproposed
and analyzed; this extension enables mobile operators and ontent providers to utilize eient
ongestion ontrolin dynamimobileenvironments.
Starting with the emergene of ommerial Internet Servie Providers (ISPs), ontinuing
in the dotom era, ulminating in today's feature-rih, appliation-driven, multimedia servies
network, prot-making is the single most-important driving fore behind the evolution of the
Internet. Understanding the eonomi proessesof theInternet is thereforeessential [2℄. There
is broad literature in the area of modeling interations between ISPs, most of these employ
game-theoretial tools [3℄ [4℄ [5℄. While these papers introdue and analyzeomplex models on
the interation of ISPs at dierent levels of the hierarhy, they mostly assume a very simple
user behavior model when investigating the market for last-mile ISPs: end-users hoose the
heapest provider assuming that the quality of the servie is the same. This assumption ould
beplausible inertainsenarios, butitouldbemisleadingifthereareloyalustomer segments
present inthemarket. Ontheotherhand, eonomistsarewell aware ofthenotion of onsumer
or brand loyalty, whih is very muh existing in realisti markets [6℄. Pratially speaking, a
ustomerisloyaltoabrand ifshepurhases theprodutofthatbrand,even ifthereareheaper
substitutions onthe market. Ifuser loyalty isgiven onsideration whenpriing Internet aess,
higher prots an be ahieved by the servie providers. Also, the ongoing network neutrality
debate [7℄ is driving a hange in priingshemes used by players of the Internet eosystem [8℄.
AessISPs arethinking abouttaxingontent providersfor arrying their ontent to end-users,
furthermore, they want to ap user's datatra inorder to operate protably. While ertainly
advantageous for aess ISPs, these shemes may not be optimal or even aeptable for users
and ontent providers. Furthermore, the global eonomi risis left providers worrying about
their revenues and users trying to utdown on their osts. Given theirumstanes, designing
a priing mehanism, whih lets ISPs plan their inome and gives users the possibilityto save
some money,is highlyrelevant.
User-provided networking hasseen its stok risingreently. While some see this onept as
aninterestingbut onlymoderatelyviablealternativetothetraditionalInternetServieProvider
entri paradigm, others believe ithasthe potential toindue aomplete shift inInternet om-
muniation patterns. The latter view an be justied by four important disruptive aspets of
user-provided networking. First, sine the end-user an share or sell her own resoures (e.g.,
onnetivity), thedistintion between end-user devie and network devie disappears. Seond,
thenatureofwirelessmedia,humanmobility[55℄andtheriseofmiro-operatorsreatetheneed
forprotoolsthatinherentlyhandleintermittent onnetivity,opportunistirelayingandsmooth
roaming. Third,user-provided serviesrequire traditionaltrustrelationships to be transformed:
soial networksoftrust shouldbeformedto ensurethewillingnessto ooperateandto maintain
network growth. And last, swift adoption of new tehnologies is possible as adopters are the
end-users themselves. Itis reasonable tobelieve thatthese novelharateristis andfuntional-
ities ould enable user-provided wireless networking to be the foundation of thefuture wireless
Internet. Existingworkon wirelessommunity networksfouses onuserpartiipation andom-
munity priing [56℄ [57℄ [58℄. Although FON [59℄ and wireless ommunity networks in general
show great promise, their ultimate suess dependson properly designed inentive mehanisms
for both users and ISPs. Thesenetworksshould be asmuh user-provided asISP-endorsed: a
dualsupport isessential to ahieve global wireless onnetivity.
With the proliferation of wireless ellular and ommunity networks, and more speially
evenmoredominant. High-speedwirelesstehnologieslikeUMTS,HSPA,LTE[80℄andWiMAX
are beoming reality, bringing broadband experiene to mobile users. The data rates of these
tehnologieswerealreadyrealizedandsurpassedinwirednetworksadeadeago,eventriggering
a urry of researh for making TCP (thedominant transport protoolof the Internet) apable
of utilizing high-bandwidth, high-delay network onnetions [81℄ [82℄ [83℄ [84℄. Also, there is a
signiantamountofworkonmakingTCPwireless-friendly[85℄[86℄[87℄[88℄[89℄. However,sine
new, broadbandwirelessnetworksaredeployedinanisland-likemanner,inter-systemhandovers
are beoming ommonplae. A handover between a GPRS and an LTE network means a link
apaity inrease of several orders of magnitude. This ould signiantly impat a number of
appliations, e.g., multimedia streaming over TCP [90℄ [91℄. Understanding the limitations of
existingTCP variants andproposing anewmehanismwhihan handlethese senarios well is
important,beauseuserexperieneansuerfromineienttransportprotools,drivingpeople
away from using mobile Internet. Mehanisms with expliit signaling introdue the need for
network support and additional overhead, and also have a low hane of deployment [92℄. In
order to ensure deployability and failitate adoption, the proposed sheme should be a simple,
easily implementable,server-side algorithm, adhering to theend-to-enddesign priniple.
1.1 Researh Objetives
The objetive ofmy researhis threefold.
(1) Quantify theimpat ofustomer loyalty onthepriingompetitionbetween Internet Ser-
vie Providers. Propose a priing mehanism for Internet aess, whih enables network
providerstoplanetheirrevenues,whileusersandiretlyinuenetheimplementedbilling
poliy.
(2) Understandthe eonomiinterations inwireless ommunitynetworks. Show thatproper
inentive design failitatestheemergene ofatrulyglobal wireless ommunity,whileboth
users and network providers protfrom the network.
(3) Understand the limitations of existing TCP versions in senarios of sudden apaity in-
rease. Propose a simple non-ongestion detetion mehanism whih solves this problem
eetively.
1.2 Methodology
TheresultspresentedinChapter2wereobtainedmainlybygame-theoretialanalysis[9℄. Quan-
tiationoftheimpatofustomerloyaltyonInternetaesspriingwasdonebyapplyingnono-
perfet equilibrium, Nash reversion). For proving the properties of the novel user-inuened
priing mehanism, I used ooperative game theory (transferable payos, ore,
ψ
-stable pairs,Shapley value).
Eonomiinterations inglobalwirelessommunitynetworks(Chapter3)werealsomodeled
and evaluated with the help of nonooperative (one-shot game, Nash equilibrium, Stakelberg
game, subgame-perfet equilibrium,bakwardindution) andevolutionarygametheory(tness,
mutation, evolutionarily stable strategies) [60℄. Probability theory was used for modeling the
geographi relevanes of users. Extensive simulations were utilized for evaluating the extended
modelinludingusermobility,heterogeneouspaymentsandtemporalevolution. Real-worlddata
setswereusedasinputtothe simulations,e.g.,GPSoordinatesoftheusers'homes,population
densityand user mobility behavior.
ThemainresultsofChapter4werederivedfromtestbedmeasurementswithaworkingLinux
kernelimplementationoftheproposedmethods. ThelowerboundforFAST,andthemethodfor
SpeedDetet parameter tuning was derived byanalytial alulations. Larger sale throughput
and fairness evaluation was done by simulations in the well-known ns-2 framework using the
exat kernel implementation ode.
Sine we use game theoryextensively throughout the dissertation, we introdue some basi
onepts inthefollowing.
1.2.1 Basi notions of game theory
Here we present elements of lassial non-ooperative (players behave rationally and selshly,
i.e., maximize their individual prots) and evolutionary (in the biologial sense) game theory
we build upon throughout the dissertation. Note that onepts of ooperative game theory
used inSetion 2.3 aredened there to failitate understanding and the ow of reading. For a
omprehensive materialon gametheoryplease refer to [9℄and [10℄.
One-shot game
Also known as a stage game, this type of game assumes that players at at the same time
instant, therefore there is no ausality. A gamein strategi (normal) form an be desribed by
three elements:
•
the set ofplayersi ∈ I
,whih we take to be thenite set{ 1, 2, . . . , I }
;•
the pure-strategy spaes i ∈ S i
for eah playeri
,wheres i
isa possibleation of playeri
;•
and payo funtionsπ i
, whih give playeri
'sutilityπ i (s)
for eah proles = (s 1 , . . . , s I )
A mixed strategy
σ i
is a probability distribution over pure strategies. Also, we denote the opponentsofplayeri
asa wholeas− i
.AgeneralsolutiononeptforgamesofeonomiinterestistheNashequilibriumsolution. A
Nashequilibriumisaproleofstrategiessuhthateahplayer's strategyisanoptimalresponse
to theother players' strategies.
Denition 1 (Nash equilibrium) A mixed strategy prole
σ ∗
is a Nash equilibrium if, for all playersi
π i (σ i ∗ , σ ∗ − i ) ≥ π i (s i , σ ∗ − i ) ∀ s i ∈ S i .
(1.1)A pure-strategy Nash equilibrium is a pure-strategy prole that satises the same onditions.
Pratially speaking, a strategy prole is a Nash equilibrium point, when no individual player
an geta higherpayobyunilaterallydeviating from it.
Note,thatweassumegamesofompleteinformation inthiswork,meaning thateveryplayer
knows the exat payofuntionsof otherplayers.
Stakelberggame
Also known asa leader-follower game, it introdues multiple stages. The leader ommits itself
rst,hoosesitsstrategy,thenthefollowersrespondsequentially. Note,thatweassumeobserved
ationsthroughout thedissertation.
TheStakelbergmodelanbesolvedtondthesubgame perfetNashequilibrium or equilib-
ria, i.e. thestrategy prolethatserveseahplayerbest,giventhestrategiesoftheotherplayers
and thatentailseveryplayer playing ina Nashequilibrium inevery subgame.
Denition 2 (Subgame-perfet equilibrium) Astrategy prole
s
isasubgame perfetequilibrium if it represents a Nash equilibrium of every subgame of the original game.Theusual method to ompute thisequilibrium isalledbakward indution.
Repeated game
Themostwidelystudied repeatedgametypeiswhenastagegameisrepeatedapossiblyinnite
number of times. An important feature of a repeated game is the way in whih a player's
preferenesmaybemodeled. Weusedisountedpayostomodelthatthevaluationofthegame
diminishes with time depending on the disount parameter
Θ
. The solution onept for theoptimal strategysetisthesameasinDenition2,however, bakward indutionannotbeused
to derive it.
Onmany oasions, it isfound that the optimal method of playing a repeated game is not
optimum strategy. This an be interpreted as a soial norm. An essential part of innitely
repeatedgamesispunishingplayerswhodeviatefromthisooperativestrategy. Thepunishment
may be something like playing a strategy whih leads to redued payo to both players for the
rest of thegame (alleda triggerstrategy). Ifthepunishment strategy isa Nashequilibrium of
thestage game, the ation isreferredto asNash reversion.
Evolutionary game
Maynard Smith's evolutionary theoryoriginates from the appliation of the mathematial the-
oryofgamesto biologialontexts, arisingfromtherealizationthatfrequenydependenttness
introduesastrategi aspetto evolution. Reently,however, evolutionarygametheory hasbe-
omeofinreasedinteresttoeonomists,soiologists,andomputersientists,sineevolutionary
games an be adaptedto their domains. Themain dierenes inan original evolutionary game
ompared to a lassialgame are:
•
There is alarge populationofplayers.•
Players are hard-wiredto their strategies; theyplay,and either reprodue anddie, or are partially replaed.•
Payos represent the reprodutive tness ofstrategies, i.e., thehaneof survival.Pratially speaking,evolutionarygame theoryanalyzes theompetitionof strategies, with-
out emphasizingtheindividualplayers. Heneitssolution onept,evolutionarilystable strategy
(ESS).
Denition 3 (Evolutionarily stable strategy) A strategy
s i
is evolutionarily stable, if it has the property that if almost every member of the population follows it, no mutant (i.e., an individualwho adopts anotherstrategy
s j
,j 6 = i
)an suessfully invade.Alongthelinesofpureandmixedstrategies, theonept ofamixed ESS emerges. However,
inthisaseithastwodierentinterpretations. EitheritisanESSinmixedstrategies,i.e.,every
individual plays the same mixed strategy; or the population is partitioned to groups playing
pure strategies, and relative group sizes represent the probability distribution assignedto pure
strategies. Thisseond interpretationis usedinthiswork.
1.3 Struture
Thebodyofthisdissertationisdividedintothreehapters. Thetopisoveredinthisdissertation
are somewhat heterogeneous from a related work and methodology point of view, hene we
Chapter 2 presents our eorts in quantifying user-inuene with regard to priing in the
Internet. We introdue basi game-theoretial onepts and relatedwork on Internet priingin
Setion 2.1. Setion 2.2 presents the impat of ustomer loyalty on theompetition of Internet
ServieProviderssellingaessto end-users. Afterharaterizingthisimpliiteet,wepropose
andevaluateanexpliitlyuser-inuenedpriingshemeinSetion2.3. Weonludethehapter
inSetion 2.4.
Chapter3 desribesinentive design inuser-provided and ISP-endorsedwireless ommunity
networks. We briey introdue evolutionarygame theory, and surveyexisting work on wireless
ommunity networks in Setion 3.1. We onstrut a Stakelberg game inSetion 3.2 modeling
the three levels of the deision hain in building wireless ommunities. We analyze the user
andISPgames,and givenumerial solutionstothe StakelberggameinSetion3.3. Wefurther
supportouranalytiresultswithdata-drivensimulationsinSetion3.4. Weonludethehapter
inSetion 3.5.
Chapter4 introduesourresults onTCPbehaviorindynamienvironments;furthermore, it
proposes and evaluates theend-point-only SpeedDetet extension to TCP thathandles sudden
apaityinreaseinasimpleandeientmanner. Setion4.1givesanoverviewonTCPversions
for high-speed andmobile environments. We present ameasurement study on high-speed TCP
performane during mobile handovers in Setion 4.2. Based on the ndings of this study, we
proposeandevaluatetheSpeedDetetextensiontoTCPinSetion4.3. Weonludethehapter
inSetion 4.5.
Finally, we summarize the ndings of this dissertation and outline their possible pratial
appliations inChapter5.
Priing Internet Aess under User
Inuene
Advanesinnetworkingtehnologyandaordableserviepriesareontinuingtomake Internet
aess available for billions of ustomers. To provide end-to-end network onnetion, Internet
Servie Providers (ISPs) form a hierarhy that spans from loal ISPs who sell aess to end-
users, through regional ISPs who onnet loal ISPs to the Internet bakbone, to Tier-1 ISPs
who form the bakbone, and are peering with eah other. The eonomi interations among
servie providers of dierent levels and end-users have been in the fous of interest for several
years. Furthermore, these interations will ontinue to get speial attention, sine initiatives
like theNSFFIND [1℄and EURO-NF[19℄ promote eonomiinentivesasarst-order onern
in future network design. Also, deision-makers trying to work out a plausible solution for the
reentlysurfaednetneutralitydebate[7℄wouldgreatlybenetfroman in-depthunderstanding
of eonomiproessesinside theuser-ISP hierarhy.
On the other hand, the net neutrality debate has shed light on some problems of Internet
ServieProviders(ISPs). Sineat-ratebillingisdominantandusertrakeepsongrowing[17℄,
ISPsgetlowerprots perdataunitarried. Aninreasing numberofnewsandstudiesreporton
thetehniquesISParebeginning tolookat andusetokeep themselvesprotable: theseinlude
tra disrimination, introduing download aps and experimenting with alternative priing
shemes (e.g., usage-based priing, three-part taris and harging ontent providers) [18℄. In
parallel,thereisanongoingglobaleonomirisisofunseenproportionsfoldingout inthereent
months. This downturn makes people think twie about spending more than they absolutely
have to. Consequently,ISPs may have to faethefatof dereasing popularityof their servies
among users. Sine eonomianalystsannot really preditthelength of theglobal risis,ISPs
have to preparefor a userdemand-driven market resulting indiminishingprots, andsimilarly,
ustomershave to minimizetheir Internet aessostsforan extendedperiod oftime.
The above developments indiate a lear need for the better understanding of the impat
of user-inuene on Internet aess priing. In this hapter we deal with two aspets of this
issue. In Setion 2.2, we investigate the impat of ustomer loyalty on thepriing ompetition
between loal ISPs who sellInternet aess to end-users, both qualitatively and quantitatively,
using game-theoretial models and simulation. We assume a single-servie setting, where ISPs
sell Internet aessas astandalone produt; we also disussthe importane of serviebundling
and its relevane to user loyalty. In Setion 2.3 we develop an expliitly user-driven priing
mehanism,and showhowuser-inuened priingan providetheISPwithalulable revenues,
whilegivingtheusers ahaneto lowertheir ostsvia votingfor theirpreferredpriingsheme.
Furthermore,wemodelthemehanismasaooperativeweightedvotinggame,deriveequilibrium
solutions,andinvestigate thedistribution ofpower. Weshowthatuserswithmediumgenerated
travolumearepivotaltotheoutome. Wealsodisussthepratialfeasibilityoftheproposed
mehanism. Setion 2.4onludes thehapter.
Resultspresentedinthis hapter arepublishedin[C6℄, [C5℄ and[J2℄.
2.1 Bakground
Here we give a briefoverviewon Internet eonomis and brand loyalty.
2.1.1 Internet eonomis
There isasigniant bodyofworkonInternetpriingmehanisms. Someof thepaperspropose
sophistiated priing models for ISPs to extrat onsumer surplus [20℄ [21℄ [22℄. Others argue
that simplepriing plansare theonlyviable ones, sine there isa learuser preferenetowards
them[2℄[23℄.ATCP-sessionbasedpriingshemeisproposedin[24℄. Bypresentingtheuserwith
ost andprie information,theproposedsysteman be usedforost reovery andto enourage
eient use of network resoures. In another paper [25℄, authors propose a responsive priing
method based on the urrent utilization of network resoures. They show how this method
puts ontrol of network servie bak to the users. A simple solution for priing dierentiated
serviesisproposedin[26℄. Asmart-marketmehanism,whereusersbidfortheusageofnetwork
resouresisintrodued in[27℄. Authorsof[20℄ritiquethewelfare-optimizing paradigmonthree
grounds: marginalostpriesmaynotproduesuientrevenuetofullyreoverostsandsoare
perhaps of limitedrelevane, ongestion ostsare inherently inaessible to thenetwork and so
annotreliablyformthe basisfor priing,andthatthereareothermorestrutural goalsbesides
optimality. In [28℄ authors establish the Prie of Simpliity (PoS) referring to the dierene
in revenues between a simple priing sheme (at-rate) and the maximum ahievable revenue.
Furthermore, they haraterize a rangeof environments, where PoS islow, i.e., at-rate priing
ThereisalsobroadliteratureintheareaofmodelingeonomiinterationsbetweenISPswith
game-theoretial means [3℄ [4℄ [5℄. While these papers introdue and analyze omplex models
for the interation of ISPs atdierent levels of thehierarhy,they mostlyassumeaverysimple
userbehaviormodelwheninvestigatingthemarketforloalISPs: end-usershoosetheheapest
provider assuming that the quality of the ertain servies is the same. Thisassumption ould
beplausible inertainsenarios,but itouldbemisleadingifthereareloyalustomer segments
present inthe market.
2.1.2 Brand loyalty
Eonomists are well aware of the notion of onsumer or brand loyalty, whih is very muh
existing in realisti markets. Pratially speaking, a ustomer is loyal to a brand, when she
purhases the produt of that brand, even if there are heaper substitutions on the market.
Brand loyalty is rooted in both satisfation towards a given brand and ustomers' relutane
to try substituteproduts. Moreover, the often-debated endowment eet [29℄ ouldalso be in
play; losing something you have owned hurts more than it feels good to gain something new.
Thisovervaluation ofsomeone'sownproperty ouldbea fatorinbrand loyalty.
There is existing work dealing with lassiation of buyers into loyalty groups [30℄, and a
reent study develops and empiriallytests amodel ofanteedents of onsumerloyaltytowards
ISPs [31℄. In [6℄ authors use a game-theoreti framework to prove that if loyalty inreases
withmarketshareand penetration,ustomer retention strategiesseem tobeonsequently more
eient for market leaders. An other study [11℄ analyzes a duopolisti prie setting game in
whih rms have loyal onsumer segments, but annot distinguish them from prie sensitive
onsumers. Theydemonstrate thatonsumerloyaltyplaysanimportant roleinestablishingthe
existene andidentityof aprie leader.
The latter two papers provide valuable insight to the impat of brand loyalty on ertain
markets, butalso inspireforfurtherinvestigations. Firstofall,howdoesustomer loyaltyaet
the dynami market of Internet aess? Seond, authors of [11℄ only onsider perfetly inelas-
ti demand and a single reservation prie for the whole ustomer population. While these two
assumptionsmayholdinertainsenarios, arethey validifonsidering theInternetaessmar-
ket in developing ountries or an eonomially dierentiated Internet user population? Third,
are there inentivesfor ooperative priingregarding loalInternet Servie Providers ina mar-
ket where user loyalty is present? And last, is the simple model, whih is ommonly used in
game-theoreti frameworks,a good representation ofreal-world brand loyalty? Can thereal-life
behaviorof ustomers(suh as sensitivity to the prie dierene between providers and uner-
taintyintheirdeisions)beinorporatedintoabetterusermodel? Wearguethatndinganswers
tothese questionsan bringuslosertotheunderstanding ofeonomiinterations amongISPs
involving ustomer loyalty intheir future investigations.
2.2 Priing Internet aess in the presene of user loyalty
Soio-eonomi aspets of future ommuniation networks suh as priing models for network
providers, network neutrality, and Quality of Experiene (QoE) are beoming more and more
important as the onvergene of networks is inprogress. All the above areas share a ommon
interest: the deeper understanding of user behavior. As a step towards a more realisti user
model, we investigate ustomer loyalty and its impat on the priing ompetition of Internet
Servie Providers(ISP) whosellInternet aess toend-users.
Themainontributionofthis setionistwofold. First,we analyzetheimpatofuserloyalty
withgame-theoretial meansmotivatedbytheBertrand game. Weshowhowloyaltyintrodues
anewequilibriuminarepeatedgamesettingresultingintheooperationofISPs. Furthermore,
we investigate thease of a dierentiated ustomer population by introduing dualreservation
values,and showhow itleads tonew, purestrategy Nashequilibria indiating thatISPsshould
makethemostoutoftheirrespetiveloyaluserbase. Seond,we onstruttwo novelmodelsfor
ustomer loyalty inorporating two important aspets of the users' purhasing deisions: prie
sensitivityandinherentunertainty. Weevaluatetheimpatofuserloyaltythroughthesemodels
byextensivesimulationsinanumberofrelevantsenarios. Inpartiular,weshowhowthehigher
levelofloyaltyinthe userpopulationleadstolarger protsforISPs. Wearguethatourndings
an motivate network researhers to inorporate a ner-grained user behavior model in their
investigations on priingmodels ofnetwork servies andother soio-eonomi issues.
2.2.1 Motivation
First, we showan example on the eet of loyalustomers througha simple game, and seond,
we justifythe existeneofa loyaluser segment inthe Internet aessmarket.
Brand loyalty
To illustrate theeet of brand loyalty we onsider the following game [32℄. Suppose there are
tworestaurantssellingpizzainapartiulargeographimarket. Supposetheyeahonsiderthree
possible priesfor pizzas: a highprie (H), amedium prie (M)anda low prie(L). The prot
per produtisknowntobe$12,$10 and$6 foreah rmregardlessofthevolume ofsales. Also
let us assume a perfetly inelasti demand funtion,
D(p) = 10000
, so ustomers buy 10000pizzas without regard to its prie. The game is similar to the Bertrand game as if the pries
of the two rms are dierent all demand goes to the lower pried rm, and if the pries are
equal, rms split the market evenly. It is easy to see that
(p ∗ 1 , p ∗ 2 ) = (L, L)
is theunique NashTable2.1: Payo matrixfor thebasi game (inthousanddollars)
H M L
H
(60, 60) (0, 100) (0, 60)
M
(100, 0) (50, 50) (0, 60)
L
(60, 0) (60, 0) (30, 30)
Table 2.2: Payomatrix for the brand loyalty game(inthousand dollars)
H M L
H
(60, 60) (36, 70) (36, 42)
M
(70, 36) (50, 50) (30, 42)
L
(42, 36) (42, 30) (30, 30)
Now, we hange the game a little bit, and introdue brand loyalty, suh as the rm with
thehigher prie losessome but not all of itsustomers to thelower pried ompetitor. Assume
that eah rm has a loyal ustomer base that buys
3000
pizzas, and the rms are ompetingfor the remaining demand of
4000
pizzas. In this ase the unique Nash equilibrium shifts to(p ∗ 1 , p ∗ 2 ) = (M, M )
(seeTable2.2). Itturnsoutthatbrandloyaltyremovestheinentivetotrytounderuttheprieoftheotherrminordertostealmarketshare. Thegameabovedemonstrates
qualitativelyhowtheexisteneofbrandloyaltyanaettheoutomeofthepriingompetition,
byhangingthe equilibriumpoint.
Loyalty in the ISP market
A reent empirial surveybythe Walkerompanydemonstrates theexistene ofuser loyaltyin
thewiredandwirelessISPmarket[13℄. First,itpointsouttheexisteneofatrulyloyalustomer
segment (around 38%) whih tolerates higher pries, and is likely to pay for new servies and
looks for along-term business relationship. Onthe otherhand, there arehigh risk ustomers
(around 30%)who arewillingto swithproviders at theearliestopportunity,andare drivenby
both lower pries and individual behavior. Seond, while 78% of the ustomers are satised
withtheserviethey get, only theabove-mentioned 38%aretruly loyal, henethere ismore to
loyaltythan beingsatised (again, behavioral patterns). Third, it is shownthat ISPs with the
most loyal ustomers (loyalty leaders) an expet signiantly larger revenues, faster growth
and higher stokprie performanethan their ompetitors. Inaddition, a numberof otherISP
surveys agree on the existene of loyal user bases in various geographi regions, suh as the
United States, the United Kingdom and Taiwan [31℄ [33℄ [34℄. They point out that the most
importantfatorsinhoosingaservieproviderisprie,pereivedvalueandserviesatisfation.
priingInternetaess. However,theexistingliteratureassessingthepriingompetitionamong
Internet aess providers (loal ISPs) does not take brand (or user) loyalty into onsideration
resulting in an overly simplied user model. This may lead to impreise statements regarding
equilibrium properties. But what is the partiular quantitative impat of user loyalty on loal
ISP priing ompetition? We apply the simple, stati loyalty modelused both in the pizza
game and [11℄ to the senario of multiple loal ISPs ompeting in prie to attrat ustomers
(users) to showhowloyalty ouldintrodue new equilibria, and how ooperation between ISPs
ould emergeinthe preseneof loyalty. Later inSetion 2.2.5, we addressseveralshortomings
ofthestatiloyaltymodelandintrodue twonovelmodels,whihenableusto inorporatemore
realisti user behavior to priing ompetition. These new models are inspired by the above
empirialsurveys: theyapturetheeetofpriedierene andongenitaluserbehavioronthe
loyaltyoftheuserpopulation. Weshowtheimpliation ofthesemodelsto aesspries,market
sharesand long-term prots bymeans of simulation inSetion 2.2.6.
2.2.2 Assumptions
Before gettinginto thedetails,we herebyjustifyour assumptionsusedinthegames throughout
this work. Notethat the loyaltymodel used inthegames below is similarin nature to theone
used inSetion 2.2.1. There isa xedloyal user base for eah ompeting servieprovider, and
there is anadditional group ofpotential users, not tied to anyrm, seekingthelowestprie on
themarket.
Flat-rate subsriptions
There are repeating patterns in the history of ommuniation tehnologies, inluding ordinary
mail, the telegraph, the telephone, and the Internet. In partiular, the typial story for eah
servieis thatqualityrises, pries derease, andusage inreases to produe inreased total rev-
enues. At thesame time,priingbeomessimpler [2℄. Solutions aimingtoprovide dierentiated
servielevelsand sophistiatedpriingshemes areunlikelyto be widelyadopted. Ontheother
hand, prie and quality dierentiation are valuable tools that an provide higher revenues and
inrease utilization eieny of a network, and thus in general inrease soial welfare. It is
also shown that at-rate priingwastesresoures, requires light users to subsidize heavy users,
and hinders deployment of broadband aess [35℄. However, it appears that as ommuniation
servies beome less expensive and are used more frequently, those arguments lose out to us-
tomers' desire for simpliity, so ISPs oer at rates, even if it is not revenue-maximizing [28℄.
Until there is widespread network-level arhitetural support for dierentiated priing in the
Internet, a signiant hange in these patterns is not foreseen. Additionally, mobile network
KDDI'sat-rateplanfortheirCDMA2000(CodeDivisionMultipleAess)system[36℄andNTT
DoCoMo'sfortheirhigh-speedFOMA-based(FreedomofMobileMultimediaAess)i-modeser-
vie [37℄. Nowadays, thanks in a large part for multimedia ontent, mobile at-rate plans are
availableworldwide. Furthermore,non-at ratebillingisalsoresoureonsumingfromaservie
provider'sviewpoint [38℄. Alloftheabove,andthefatthatmostInternetaessprovidersoer
at-rate subsriptions for end-users today, motivates us to assume a at-ratepriing sheme in
our models.
Consumer demand for Internet aess
Theprie elastiity ofdemandfor apartiular demandurveisgreatly inuenedbythedegree
ofneessityorluxury: luxuryprodutstendto have greater elastiitythanneessities. Thepro-
portionofinomerequiredtopurhaseaserviealsoplaysakeyrole: produtsrequiringalarger
portion of the onsumer's inome tend to have greater elastiity [39℄. These two observations
suggest that ina developed ountry, where inomes are high, Internet aess is ubiquitousand
peopletendto lean onthe Internet bya great degree(in their work andalso duringtheir spare
time), almost every household has Internet aess, so the demand an be modeled asonstant
(perfetly inelasti)[40℄. Onthe otherhand,marketsindeveloping regionsarehighlypriesen-
sitive, sine people have lower inomes, and thenumber of Internet subsriptions wouldgreatly
benetfromlowerpries. Therefore,the demandfor Internet aessinsuhregionsan bebest
modeled aselasti. We usetheinelasti modelinSetion 2.2.3 to omplywiththeassumptions
of [11℄,while weinvestigate both ofthem inthegames ofSetion 2.2.4.
Reservation priesof ustomers
Consumer population is heterogeneous in the sense that ertain groups are willing to pay dif-
ferent amounts of money for the same servie. In thedream world ofISPs, in whih they were
able to perfetly identify thereservation prie of eah ustomer in themarket, they ouldoer
individually dierentiated pries, thus squeezing o every ent from the users. Suh a perfet
identiation of reservation priesis not likely inthereal world. However, thereservation prie
ofexistingustomersisgenerallyhigherthan thatofnewustomers, beauseexistingustomers
tend to exhibit higher swithing osts and also higher brand preferene for that produt [41℄.
Furthermore, most oftheanalytial literature on prie disrimination hasfound that itis opti-
mal to penalize loyals withhigher priesthan swithers [12℄ [42℄. While we do not introdue
targetedpriingtoour models,westillassumethatloyalusersinherently tolerateahigherprie
than swithers, who are only interested in disount pries. This way, we usedual reservation
values inSetion 2.2.4 to represent the heterogeneity of the user population. In Setion 2.2.3,
In allases, reservationpries areassumedto be ommon knowledge.
Quality of servie and apaity expansion
Network apaity and the user's demand for bandwidth are related to the qualityof servie in
omputer networks. A signiant inrease of a singleISP's userbase might auseongestion in
the ore network thus reduing the quality of experiene for other users. Suh a degradation
in servie quality might results in theloss of ustomer satisfation and so it an aet loyalty.
Throughout this work, we assume homogeneous quality of servie levels among all ISPs. For
analytial tratability,we assumethat theabove-dened ongestion isnegligible, and swithing
usersexperienethesamepereivedserviequalityaftertheswith. Furthermore,inthissetion
we make theassumptionthat every singleISPisprepared to handlethewholeuser population,
i.e., they have the infrastruture and apability to servie the entire market without loss of
quality. In Setion 2.2.6 we relax this assumption and integrate the need for network apaity
expansion into themodel.
Additionally, we assume that the marginal ost of apaity is onstant (zero) for all ISPs.
This roots inthe previous assumption: ISPs have built up enough apaityto handle all users
at one. Setting the marginal ost to a onstant is reasonable for alarge rangeof demand, but
growsinstepsafterreahingsomeritiallimit[39℄. Basially,weassumethatthesystemalways
operatesintherangewhere the marginal ost isonstant.
Although the payo funtions of ISPs are pretty simple in this work, they are inline with
at-rate priing, onsumer demand elastiity and reservation values disussed above, and thus
they suit ourneeds.
2.2.3 Inentive to ooperate
Herewepresentasingle-shotgameofuserloyaltywhihwasintroduedin[11℄. Later,weextend
thisgametoaninnitelyrepeatedgame,andshowhowaooperativemaximumanbeenfored,
where the long-term prot of ISPs are higher than that of playing the equilibrium strategy of
thestage gamein eah round.
The stage game
Consider a market with two loal ISPs ompeting in pries for a xed number of ustomers.
Customers aresplit into three partitionsupon their brand loyalty: the rst group onsistsof
l 1
ustomerswho areallloyal to ISP
1
inthe sense thatifISP1
'spriep 1
islessthan or equal toareservationvalue
α
, they hoose ISP1
astheir servieprovider, otherwisethey do not purhaseInternet aess. The seond grouponsists of
l 2
loyalustomers of ISP2
,while the third groupontains
n
swithers,whobuyserviefromtheheapestprovider,ifitsprieisnotgreaterthanα
. Ifthe providersannoune thesame prie(p 1 = p 2 < α)
,thenhalfof theswithers hoosesISP
1
and the other half hooses ISP2
. The ow of thegame is that ISPs announe their priessimultaneously,thenustomersmaketheir hoies. Thisgameisreferred toas
G 0
.Note, that though values
l 1 > 0
,l 2 > 0
andα > 0
are ommon knowledge, group mem-bership of a given ustomer annot bedetermined, so there is no prie disrimination possible.
Furthermore, forsimpliitywe assumeaonstant unitostofzeroforbothrms,and thatISP
1
hasthelargerloyal userbase,
l 1 > l 2
.Given the above andthat
p 1 ≤ α
andp 2 ≤ α
,ISP1
's payoan beexpressedasπ 1 (p 1 , p 2 ) =
(l 1 + n)p 1 p 1 < p 2 (l 1 + 0.5n)p 1 p 1 = p 2
l 1 p 1 p 1 > p 2
(2.1)
Itanbeshown(see[11℄and[12℄)thatthisgamehasauniqueNashequilibriuminmixedstrate-
gies. In this ase, equilibrium prots are
π 1 = l 1 α
andπ 2 = l l 2 1 +n +n l 1 α.
Asit an be notied, theequilibriumhasshiftedompared tothe simpleBertrandgamewithoutonsumerloyalty(whih
is (0,0)inase ofzero produtionosts), both partieshaving apositivepayoinequilibrium.
The repeated game
Now, we extend the previous model, and show that the innitely repeated
G 0
hasa sub-gameperfet equilibrium, whih an be enfored by a threat strategy, namely the Nash equilibrium
strategy of the stagegame
G 0
.In the following we onstrut
G r
as the innitely repeated extension ofG 0
. Payo is dis-ounted at step
k
with a disount fatorΘ < 1
. The game is ontinuous at innity sine thedisountedpayoinanystepisboundedby
α(l 1 +n)
. Thiswayweanusetheone-stepdeviationpriniple to prove sub-gameperfetionof agiven strategy set.
Now,ifthetwoprovidersooperateandsettheirpriesequaltothereservationvalue
α
,theywillshare"swithers"equally,inadditiontokeepingtheirownloyalusers. Thiswaytheirpayos
(
π
oop)wouldbehigherthanintheequilibriumase(π
eq),sineπ 1
oop= (l 1 +0.5n)α > π 1
eq= l 1 α
,and
π
oop2 = (l 2 + 0.5n)α > π 2
eq= l l 2 +n
1 +n l 1 α
ifn > l 1 − 2l 2
. Intheooperativeasethejoint protof thetwo ISPsis themaximumahievable
(n + l 1 + l 2 )α
. This ooperation is highlybeneialfor both parties. Ifsomehow one ISPtries to grab thewholefree marketina singlestep
k
,theotherISPan ounterat fromstep
k + 1
byhargingtheNashequilibriumpriefromG 0
furtheron,whihresults ina dereasedpayofor thetraitor. We showthatthisNashreversionassures
sub-game perfetion forthe following strategy prole under thestated onditions.
Proposition 1 The strategy prole Cooperate until the other player deviates and then play
aordingto theequilibrium in
G 0
isa sub-gameperfet Nash equilibrium for therepeated gameG r
, ifn > l 1 − 2l 2
andΘ > 1 2 + l 1 −
n+l 1 n+l 2 l 2
2n
.Proof: A strategy prole issub-gameperfet,ifthefollowing holds:
π i
nodev(k, ∞ ) > π i
dev(k) + π i
dev(k + 1, ∞ ),
(2.2)meaning that the sum prot is greater if there is no deviation from theagreed strategy. If we
assumethat ISP
2
deviates, thistranslates toX ∞
i=k
Θ i π
oop2 > Θ k π 2
dev+ X ∞
i=k+1
Θ i π 2
eq.
(2.3)After solving (2.3)for
Θ
wegetΘ > π
dev2 − π
oop2
π 2
dev− π 2
eq.
(2.4)The one-stepdeviation at step
k
isrealized by underuttingtheother ISPbya marginalǫ > 0
,and harging a prie of
α − ǫ
to the users. This wayπ
dev2 = (l 2 + n)(α − ǫ)
. Further on, wesubstitutedierent payos for ISP
2
,andsimplify the expression:Θ 2 > 1
2 + l 1 − n+l n+l 1 2 l 2
2n .
(2.5)If ISP
1
deviates, by following the same steps we getΘ 1 > 1 2
. Sinel 1 > l 2
by denition, alsoΘ 2 > Θ 1
,soiftheatualΘ > 1 2 + l 1 −
n+l 1 n+l 2 l 2
2n
,the proposition holds.The optimal deision of ISPs lies in evaluating the assumption inequalities. For Player 1
(larger loyal userbase) it isalways better to ooperateand try to ahieve sub-game perfetion.
For Player 2it is a matterof loyal user basesize: if
n > l 1 − 2l 2
, Player 2will alsoooperate.If not, thenshe will play the
G 0
equilibrium strategy,while Player 1 will play theooperative strategy for one round. Then from round 2(beause of Nash reversion), player 1 will also playthe
G 0
equilibriumstrategy.While expliit ooperation may be illegal, this inentive may lead to disussions between
servie providers. Note, that a two-ISP setting may seem artiial, it is ertainly not, e.g., a
large fration of Internet users in the US an only hoose between the loal able and phone
ompany. Moreover, there is some speulation about a artel-like ooperation among large
players in the US Internet market [43℄. Wu mentions that in the United States and in most
of the world, a monopoly or duopoly ontrols the pipes that supply homes with information.
These ompanies, primarily phone and able ompanies, have a natural interest in ontrolling
supply to maintain prie levels and extrat maximum prot from their investments, similar to
how OPEC sets prodution quotas to guarantee high pries. While this phenomena is rooted
innetneutralityandprotetinginfrastrutural investment,sine itinvolvessettingmanipulated
2.2.4 Dierentiated reservation pries
Here we onstrut and analyze single-shot games modeling the priing ompetition between
loal ISPs ghting for ustomers with dierent reservation pries. First, we deal with thease
of inelastidemand, and later,we introdue elasti demand.
Inelasti demand
Weneedtohangethesingle-shotgame
G 0
alittlebittoreetthedualityinreservationpries.We onstrut
G 1
byintroduingα 1
,thereservationvalueforswithers,andα 2
,thereservation value for loyal users (α 1 < α 2
), instead of the single reservation valueα
. Thus the demandfuntion for
G 1
is thefollowing:D(p) =
n + P N
j=1 l j 0 ≤ p ≤ α 1
P N
j=1 l j α 1 < p ≤ α 2
0 p > α 2
(2.6)
where
p
is the prie harged to users andN
is the number of ompeting ISPs. The demandfuntion an be seen in Figure 2.1(a). From that, we an dene the payo funtion
Π i (p i )
ofISP
i
,whih hasaform ofΠ i (p) =
p i l i + m n
p i = min j p j ≤ α 1 p i l i min j p j < p i ≤ α 2 0 p i > α 2
(2.7)
where
m
isthe numberof ISPs harging thesame minimumprie,thereforesharing swithersequally.
Proposition 2 Consider
G 1
with two players (N = 2
). Let us deneA = nα 1
andB i = (α 2 − α 1 )l i
fori = 1, 2
.1.
(p 1 , p 2 ) = (α 2 , α 2 )
isa pure strategy Nash equilibrium,ifA < B 1
andA < B 2
;2.
(p 1 , p 2 ) = (α 2 , α 1 )
isa pure strategy Nash equilibrium,ifA < B 1
andA > B 2
;3.
(p 1 , p 2 ) = (α 1 , α 2 )
isa pure strategy Nash equilibrium,ifA > B 1
andA < B 2
;4. There isno pure strategy Nash equilibrium if
A > B 1
andA > B 2
.Proof:
1)If
A < B 1
andA < B 2
: supposethatbothISPssettheirpriesindependentlytoα 2
. Thereis noinentive to inreasetheir priessine anyinrease will resultinzeroprot. Furthermore,
there is no inentive to derease their pries, sine their prot at