Trading Agent Compe00on The Ad Exchange Game (TAC AdX)

TradingAgentCompe00on
TheAdExchangeGame(TACAdX)
EnriqueAreyanViqueira
March15,2017
AdExchange
•  AnadexchangeisatechnologyplaFormthat
facilitates the buying and selling of media
adver0singfrommul0pleadnetworks
•  Anonlineadver5singnetworkoradnetwork
is a company that connects adver5sers to
websitesthatwanttohostadver0sements
AdExchange-Schema0c
AdNetwork
Publishers
Bid
AdExchange
Impression
Opportunity
Bid
AdNetwork
.
.
.
Bid
AdNetwork
AdNetworkKeyChallenge
Fulfilladver0singcontractsatminimumcost
whilesustainingandaWrac0ngadver0sersby
achievinghighqualitytarge0ng.
EfficientMarke-ngEffec-veness
TACAdX
OneofmanyTradingAgentCompe00onGames
In a nutshell: the game reflects the challenges
facedbyanAdNetworkasitbidsfordisplayads
opportuni0es.
hWps://sites.google.com/site/gameadx/
TACAdX-Overview
Agents:
Implement Ad Network challenge, i.e., perform a
biddingstrategy.
GameServer:
Simulates the behavior of users, web sites,
adver0sers, and an Ad Exchange for a number of
days.
TACAdX-Markets
Agentspar0cipatein3typesofMarkets:
1.  CampaignsMarket
2.  ImpressionsMarket
3.  UserClassifica0onServerMarket
CampaignsMarket
Adver5sing campaigns are created by
adver0sers to promote their brands, and the
AdNet’sroleistocarrythemout.
Acampaign:
•  Targetsasubsetofusers(MarketSegment).
•  Predefinednumberofimpressions(Reach)
•  Lastsanumberofsimulateddays(Start/EndDay)
CampaignsMarket(con0nued)
•  A new campaign is announced and auc0oned
everyday.
•  AcampaignisallocatedtotheAgentthatbidsto
executethecampaignatthelowestcost(reverse
auc0on!)totheadver0ser(Budget).
•  Upontermina0onofacampaign,theagentsgets
paidthedifferencebetweentheactualreachand
thecostoftheimpressions.
ImpressionsMarket
•  Each day, a number of users from different
MarketSegmentsissimulated.
•  Eachusergeneratesanimpressionopportunity.
•  Each impression opportunity is allocated via a
secondpriceauc5on.
UCSMarket
•  Well, actually, an agent might not know the market
segmentofauser.
–  The quality of the matching depends on the UCS level
purchased.
•  UCSlevelisaprobabilityofactuallyknowingtheusers’
marketsegment
•  GSP: the highest bidder get 1.0 level, the second gets
0.9,thethird0.9^2,andsoon.(ithgets0.9^i).
MechanicsoftheGame
Foragivendayn
Afewmoredetails
QualityScore:ameasureoftheeffec0venessof
anagenttodeliveronthepromisesmade.
ü  Higherqualityscoremeanstheagentfulfills
onitspromises(campaigns)
Thequalityscoreimposesalower/upperbound
ontheadmissiblecampaignbids.
Caneffec0velythrowanAgentoutofthegame!
NextLab:OneDayGame
Wewillplaya(very!)simplifiedversionofgame.
1.  Onlyoneday.
2.  NoCampaignBidding–Agentsneedtofulfill
theirini0alcampaignonly.
3.  NoUCS.
4.  NoQualityScore.
NextLab:OneDayGame(con0nued)
YourAgent’sjobistodetermine:
1.  Wheretobid?(whichmarketsegment(s))
2.  Whattobid?
Knowing:
1.  Distribu0onofimpressions
2.  Distribu0onofcampaigns
(butnotexactlyotheragents’campaigns!)