69

BASICSOF DYNAMICPROGRAMMING FORREVENUEMANAGEMENT

* JeanMichelChapuis

Revenuemanagement is the control of inventory and pricing ofa perishable product in order to improvetheefficiencyofitsmarketing. Atthetimeacustomermakesareservationforaflight,an airlinerevenuemanagermustacceptthereservation,whichiscertainatthattime,orrejectitand wait for a possible demand for seats at a higher rate in the future. Even though this pattern of behaviourhasbeenuniversalforthelastdecade,implementationinGDSremainsproblematic.This article sets out some elements of Dynamic Programming as a new optimisation method. The propertiesoftheopportunitycostofusingaunitforagivencapacitywhichisthekeytoallRM optimisations areherepresentedindetail. Le Revenue Management,c'estdirele contrle de l'inventaire et de latarification d'un produit prissable,viseamliorerlefficiencedesonmarketing.Lorsdelenregistrementdesrservations dun vol, le revenue manager dune compagnie doit soit accepter une demande de rservation actuelle et certaine ou la rejeter et attendre une probable demande un tarif plus lev dansle futur. Alors que sa pratique est omniprsente cette dernire dcennie,la mise en uvre dans les GDS soulvent de nombreux problmes. Cet article propose quelques lments de la Programmation Dynamique comme une nouvelle mthode doptimisation. Les proprits du cot dopportunitdelutilisationduneunitpourunecapacitdonne,lacldetouteoptimisationde RM,sontprsentesdansledtail.

INTRODUCTION

Fordecades,wehavelearnedtoreverethelettersofRevenueManagement(RM).Itsspiritof pricing and of inventory controlis widely shared among variousindustries, major companies and academic researchers. Revenue maximization is usually reached by comparing a current discount

IRIDIP. Mail: University of French Polynesia, BP 6590, 98702 Faa'a, French Polynesia. Email: jean michel.chapuis@upf.pf. AssistantProfessorofManagementattheUniversityofFrenchPolynesia, PhD(CorporateFinance)Head oftheInternationalHospitalityManagementMasteroftheUniversityofLaRochelle,2002to2004Guest LecturerfortheCHNUniversityofDoha,QatarandtheESCAofCasablanca,Morocco.

70

(2007) 13 REVUEJURIDIQUEPOLYNSIENNE

1 farerequestwithanexpectedhigherfarerequestinthefuture.ThewellknownEMRSmodel gives a heuristic solution but assume that the demand is static and will not evolve during the booking period.However,newlowcostcarrierentrants,faretransparencyfromInternetfaresearchengines andchangingcustomerpurchasingpatternsarethemajorforcespushingairlinerevenuemanagersto adapttheiroptimizationmodels.Inaddition,usingtraditionalforecastingandoptimizationmodels that assume fare classindependence causes continuous erosion of yield. Especially when thereis excesscapacity,thesurplusseatsflowdownthefarestructure,allowinghighyieldcustomerstobuy down. Later, this behaviour affects forecast, ie expecting lower demand for high fare classinthe future,andleadsrevenuemanagerstoprotectlessseatsforit.

Almost RM situations can usefully be modeled as Dynamic Programming problems (DP) becausethosemodelstakeintoaccountfuturepossiblebookingdecisionsinassessingthatcurrent decision.Moreover,DPallowstorelaxtheearlybirdshypothesis(lowcontributioncustomersbook in advance of the high contribution ones). The optimal controls are then timedependent as a functionoftheremainingcapacity.Dynamicpricingisthespecialcasewherepriceisthecontrol usedbythemanagers.
2 Nonetheless, the practising of Bellman equations is relatively new in RM since none of the 3McGillandVanRyzin(1999)concludethat: pioneeringpapersuseDP.

Dynamic formulations of the revenue management problem are required to properly model real world factors like cancellations, overbooking, batch bookings, and interspersed arrivals. Unfortunately, DP formulations,particularly stochastic ones,are wellknown fortheirunmanageablegrowthinsize when realworld implementations are attempted. Usually, the only hope for dynamic optimisation in these settings lies in identification and exploitation of structural properties of optimal or near optimal solutions. Knowledge that an optimal solution must be of a controllimit type or be represented by a

See P P Belobaba "Application Of A Probabilistic Decision Model To Airline Seat Inventory Control", Operations Research (1989) vol 37, n 2, 183198 which is an extension of Littlewood's model KLittlewood"ForecastingandControlofPassengersBookings"(1972)ProceedsofAGIFORS,Nathanya, Israel. SeeRBellman AppliedDynamicProgramming (PrincetonUniversity Press,1957). The basics of Revenue Management can be found in P E Pfeifer "The Airline Discount Fare Allocation Problem" (1989) Decision Sciences, vol 20, n 1, 149157 Belobaba, op cit or S E Kimes "Yield Management:AToolforCapacityConstrainedServiceFirms"(1989)JournalofOperationalManagement, Vol 8, n 4, 34864 among others. For an overview, check J I McGill and G R van Ryzin "Revenue Management:ResearchOverviewandProspects"(1999)TransportationScience,vol33,n2,233257orG BitranandRCaldentey"AnOverviewofPricingModelsforRevenueManagement"(2003)Manufacturing andServiceOperationsManagement,vol5,202229,and morerecently WCChiang,JC HChenand X Xu "An Overview of Research on Revenue Management: Current Issues and Future Research" (2007) International Journal of Revenue Management, vol 1, n 1, 97128. The experience of American Airlines whileimplementingRevenueManagementinaworldwidescaleisreportedin B C SMITH, J F LEIMKUHLER AND R M DARROW "YieldManagementatAmericanAirlines"(1992) interfaces, vol22,n1,831.

2 3

BASICSOFD YNAMICPROGRAMMINGFORREVENUEMANAGEMENT

71

monotonic threshold curve can be invaluable in development of implementable systems. The existing literature has already identified such structures in special cases of the revenue management problem however,therearedifficultareasstillrequiringwork(emphasisadded).

The goal of this research is to review some Dynamic Programming models dedicated to Revenue Management in order to provide a solid basis for future work. The structure is the following: PartII,formulatesboththeRMproblematicanditsDPmethodology.Then,theproblem of setting booking limits and prices using DP approach isaddressedinthePart III. The last Part concludesandopensapplicationsofDPinRM.

II

DYNAMICPROGRAMMINGOFREVENUEMANAGEMENT OPTIMISATION

AfterreviewingtheRMproblem,abriefreviewoftheliteratureofDPmodelsappliedtoRMis presented in A and a mathematical formulation isin B. As anexample, considerthe following case.Youaremanagingahoteloftworoomsforthenextweekend(theFridayandSaturdaynights), meaningyouhavetosell4unitsofasameresource,marketedthrough3products.Youcharge$100 foraroompernight,exceptfortheguestswhostayfor2consecutivenights.Thelattercosts$160. ThisisMondayandyoucheckonyourwebsitetheactualactivityinordertoconfirmreservation. There are 2requests for the Friday night(at $100each),1 request forthe Saturday night(still at $100). Atthe sametime, your travel agent calls you to book 2 couples for 2 nights(so theprice would be $160 each). What should you do? Revenue Management can help to choose which requeststoaccept. In order to maximise operating revenue, RM isthe process by which a manager controls the availabilityofaproductorservice,marketedwithadifferentialanddynamicpricing.Controlscan 4 beeffectivebyvaryingprices,settingbooking limitsandmanagingfences. Thisapproachisinline 5 whosupportedbothpriceandquantitybasedmodelsofRM. withTalluriandVanRyzin(2004), MostauthorsoftendefineRMastheapplicationofcontrolandpricingstrategiestoselltheright 6 capacitytotherightcustomer,intherightplace,attherighttimeandattherightprice .RMhas 7 provenitspotentialimpactonprofitabilityinthepast .

Inthe previousexample,onecan easily seethatthe firstcome first serveruleisnot optimal for thehotel. Thefenceof"stayatleasttwonights",thatcanjustifythediscountrateandhelptoincreaserevenue,isnot totallyefficient.Onlyabookinglimitofone"Twonights"packageorabidpriceforeachnightareoptimal. SeeKTTalluriandGJvanRyzin"TheTheoryand PracticeofRevenueManagement" 712. Kimes,opcit. Smith,LeimkulherandDarrow,opcit.

5 6 7

72

(2007) 13 REVUEJURIDIQUEPOLYNSIENNE

A LiteraturereviewofDPModelstoRMOptimisation. Dynamic programming addresses how to make optimal decisions over time under uncertain 8 conditionsandtocontrolasystem .MostRMsituationscanbeanalysedassumingadiscretestate 9 andadiscretetimeoverafinitehorizonmodelling .Table1summarisestheliteratureaccordingto four criteria:(i) A paper could consider a single product (at various prices) or multiple products (depending on purchase restrictions or independent demands for example) (ii) A paper could considerastaticpolicy(assumingastrictorderofbookingarrivals)orallowforadynamicpolicy (not assuming the early birds hypothesis) (iii) A paper could consider various forms of demand process (iv) A paper could consider either a single resource for 1 to n products or multiple resources(suchasanairlinenetworkofhubsand spokes). Table 1:classificationofliterature Product Single Curry(1990)Wollmer(1992)BrumelleandMcGill(1993)Leeand Hersh(1993)GallegoandVanRyzin(1994)BitranandMondshein (1995)Robinson(1995)LautenbacherandStidam(1999)Zhaoand Zheng(2000)You(2001) Gallego and Van Ryzin (1997) Talluri and Van Ryzin (1998) Feng and Xiao (2001) Kleywegt (2001) Bertsimas and Popescu (2003) ElHaberandElTaha(2004)BertsimasandDeBoer(2005) Policy Curry(1990)Wollmer(1992)BrumelleandMcGill(1993)

Multiple

10 Static

8 9

The methodologicalreferenceisD BertsekasDynamic ProgrammingandOptimalControl(1995)Athena Scientific,vol1,BelmontMA. T Lee and M Hersh"A Model for Dynamic Seat Inventory Control with Multiple Seat Bookings" (1993) TransportationScience,vol27,pp.233247isthemostpopularreferencewiththeirmodelfordynamicseat inventory control. G Gallego and G van Ryzin "Optimal Dynamic Pricing of Inventories with Stochastic DemandoverFiniteHorizons"(1994)ManagementScience,vol40,n8,9991020arecommonlycitedas thereferenceindynamic pricing modelusingdynamicprogramming.

10 DBertismasandSdeBoer"Dynamicpricingandinventorycontrolformultipleproducts"(2005)Journal of Revenue and Pricing Management vol 3, n 4, 303319 D Bertismas and I Popescu "Revenue ManagementinaDynamicNetworkEnvironment"(2003)TransportationScience,vol37,n3,257277G BitranandSMondschein"AnApplicationofYieldManagementtotheHotelIndustryconsideringMultiple Day Stays" (1995) Operations Research, vol 43, n 3, 427443 S Brumelle and J I McGill "Airline Seat Allocation with Multiple Nested Fare Classes" (1993) Operations Research 137 R Curry "Optimal Seat Allocation with Fare Classes Nested by Origins and Destinations" (1990) Transportation Science vol 24, 194204 Y Feng and G Gallego, 2000, Perishable asset revenue management with Markovian time dependentdemandintensities, ManagementScience,vol46,n7,941957YFengandBXiao,2001,A dynamicairlineseatinventorycontrolmodelanditsoptimalpolicy OperationsResearch,vol49,n6,938 952 G Gallego and G van Ryzin "A Multiproduct Dynamic Pricing Problem and its Applications to

BASICSOFD YNAMICPROGRAMMINGFORREVENUEMANAGEMENT

73

Dynamic

Lee and Hersh (1993) Gallego and Van Ryzin (1994) Robinson (1995)GallegoandVanRyzin(1997)TalluriandVanRyzin(1998) Liang(1999)You(1999)LautenbacherandStidam(1999)Fengand Gallego(2000)FengandXiao(2001)Kleywegt(2001)You(2001) BertsimasandPopescu(2003)ElHaberandElTaha(2004) BertsimasandDeBoer(2005) Demand Curry(1990)Wollmer(1992)BrumelleandMcGill(1993)Robinson (1995)Kleywegt(2001) Lee and Hersh (1993) Gallego and Van Ryzin (1994) Gallego and Van Ryzin (1997) Talluri and Van Ryzin (1998) Feng and Gallego (2000) Feng and Xiao (2001) You (2001) Bertsimas and Popescu (2003)ElHaberandElTaha(2004) ZhaoandZheng(2000)BertsimasandDeBoer(2005) Resources Wollmer(1992)BrumelleandMcGill(1993)LeeandHersh(1993) Gallego and Van Ryzin (1994) Robinson (1995) Lautenbacher and Stidam (1999) Feng and Gallego (2000) Zhao and Zheng (2000) Kleywegt(2001)You(2001)BertsimasandDeBoer(2005) Curry (1990) BitranandMondshein (1995) Gallego and Van Ryzin (1997) Talluri and Van Ryzin (1998) You (1999) Feng and Xiao (2001)BertsimasandPopescu(2003)ElHaberandElTaha(2004)

Both

Deterministic Stochastic

Nonhomogeneous

Single

Network

The way the behaviour of customers is incorporated in the optimisation process is the next 11 challenge.Thefollowingpartalmostborrowsfrom TalluriandVanRyzinbook.

Network Yield Management" (1997) Operations Research, vol 45, n 1, 2442 S E Kimes "Yield Management:AToolforCapacityConstrainedServiceFirms"(1989)JournalofOperationalManagement, Vol 8, n 4, 348364 A J Kleywegt "An Optimal Control Problem of Dynamic Pricing" (2001) Working paper, 25 C Lautenbacher and S Stidham "The Underlying Markov Decision Process in the SingleLeg Airline YieldManagementproblem" (1999)TransportationScience,vol33,n2,136146LW Robinson "Optimal and Approximate Control Policies for Airline Booking with Sequential Fare Classes" (1995) Operations Research, vol 43, n 1, 252263 K T Talluri and G J van Ryzin An aAnalysis of BidPrice ControlsforNetworkRM(1990)ManagementScience,vol44,n11,15771594RWollmer"AnAirline SeatManagementModelforaSingleLegRoutewhenLowerFareClassesBookFirst"(1992)Operations Research,vol40,2637PSYou"AirlineSeatManagementwithrejectionforpossibleupgradedecision" (2001)TransportationsciencepartB,vol35,507524,WZhaoandYSZheng"OptimalDynamicPricing forPerishableAssetswithNonhomogeneous Demand" (2000) ManagementScience,vol46,n3,375388. 11 2004, Ibidem,651

74

(2007) 13 REVUEJURIDIQUEPOLYNSIENNE

B ASimpleFormulationofRMwithaDynamicProgram ThispaperchooseatraditionalRMproblemby consideringn> 2fareclasses.Demandfora single product is then considered discrete as well as capacity (assuming a single resource to simplify). Dynamic models do not assume the traditional hypothesis of early birds. Demand can arriveinanonstrictincreasingorderofrevenuevalues.Demandisassumedindependentbetween classesandovertimeandalsoindependentofthecapacitycontrols.Thenclassesareindexedbyj 12 suchthat p1>p2>>pj>> pn.Multiplebookingisnotconsidered. Overtime,thissystemevolvesasafunctionofbothcontroldecisionsandrandomdisturbances accordingtoasystemequation.Thesystemgeneratesrewardsthatareafunctionofboththestate andthecontroldecisions.Theobjectiveistofindacontrolpolicythatmaximisesthetotalexpected revenuesfromthesellingperiod.ThereareTtimeperiods.Time,indexedby t,runsreversesothat t=1isthelastperiodandt=Tisthefirstperiod.Bytheway,tindicatesthetimeremainingbefore theproductisconsumedorlosesitsvalue.ApplyingbasicsofDPtoRMgivesthefollowing: x(t)is,asthestateofasystem,theremainingcapacityconstrainedbetween0andthefull capacity C(bothcancellationandoverbookingarenotconsidered). w(t)istherandomdisturbancerepresentingtheconsumer'sdemandtothefirmforproduct j.Theprobabilityofanarrivalofclassjintheperiodtis lt .Thehorizonisdividedinto j decision periods that are small enough so thatno morethan onecustomer arrives during n j eachperiod,thus l 1. t j= 1 u(t)isthecontroldecision,assumeddiscreteandconstrainedtoafinitesetthatmaydepend ontimetandthecurrentstatex(t).u(t)isthenthequantityuofdemandtoaccept.Inthe simpleststatement,uisconstrainedtobeeither0or1.Theamountacceptedmustbeno morethanthecapacityremaining,so ux.

Afewpointsmustbenoted.First,asthemanagerwaitsforthedemandtorealisetomakehisor herdecisionacceptorrejectanincomingbookingrequesttherandomdisturbanceisobservable. 13 Inotherwords,s/hecanbuildthecontrolactionwithperfectknowledgeofthedisturbance. Second,customersareallowedtoarriveinanyorder(interspersedarrivals),incontrasttostatic modelsthatassumesequentialbookingclassesorlowbeforehighfares.Acceptingthisearlybirds hypothesisleadstothinkintermsofthenumberofbookingsaskedperperiod(iethebookinglimit)

12 Usually, researchers consider that the time can be divided in small enough intervals such that only one bookingarrives.You,Op citdevelopsamodelallowingformultiplebookings. 13 This assumption allows to greatly simplify both the mathematical formulation and the optimal control policies because the optimisation problem could be spread in multiple subperiods optimisation problems that are independents between them. See Talluri and Van Ryzin (2004, ibid, 654) to understand the full consequenceofthisassumption.

BASICSOFD YNAMICPROGRAMMINGFORREVENUEMANAGEMENT

75

and only requires forecasting the total demand for each class, Dj. However, the optimal control wouldbe"static",meaningconstantovertime(howmuchtoprotectorwhentorejectbooking)as themanagernevercamebackonhisorherdecision.Anotherapproachwouldhavebeentoreason for only one booking request at a time. The incoming requests are not ordered by price but one needstomodelthedemandprocessandthentheoptimalcontrolchangeseveryperiodaccordingto the remainingcapacity so"dynamic"meansrealtimeadjustment. gt(x(t),u(t),w(t))isarealvaluedrewardfunction,specifyingtherevenueinperiodtasa functionoftheparameters.Thetotalrevenueisadditiveandterminaloneisassumedtobe zero whateverhappensnamelythattheremainingcapacityafterthelaststageisnovalue.

TheobjectiveistomaximisethetotalexpectedrevenueforTperiods E g x t u t),w t) t ( ), ( ( 1 t= by choosing T control decisions uT, , u2, u1 This is a Markovian control because the control . dependsonlyonthecurrentstateandthecurrenttime,andnootherinformationisneeded,suchas thehistoryoftheprocessuptotime.Thecollection{uT,, u2, u1}iscalledapolicy m.

Foragiveninitialstate x(T)= x,theexpected(total)revenueofapolicyis:

V ) . T( x =E g x(), t xt wt t t u( ()), ()

t 1 =

Inotherwords,thisisthetotalexpectedrevenuethatcanbegeneratedwhenthereareT decision periods(remaining)andxproductstosell.E[]meanstheexpectationisestimatedoverj,thefare classes or products. The optimal policy, denoted u*, is one thatmaximizes VTu*(x) and is simply 14 writtenVT(x).Theprincipleofoptimality liesattheheartofDP.Itsaysthatifapolicyisoptimal for the original problem,then it must be optimal for any subproblem of this original problem as well.Letsaythatif{ut*,ut1*,,u1*}isnotoptimalforthetsubproblemandanotherpolicy,, yields a greater expected reward, then the optimality of u* is contradicted. Because the policy {uT*,, ut+1*, t,, 1}wouldproduceastrictlygreaterexpectedrevenuethandoesthepolicy u*. Applying this principleleads to usearecursive procedure for finding the optimal policy. The followingvaluefunctionVt(x)istheuniquesolutiontotherecursionshowninequation1,forallt andall x. Once the value of the demand is observed, the value ofu is chosen to maximize the current periodt revenueplustherevenuetogo,orpj.u+Vt1(x u)subjecttotheconstraintu {0,1}.This sentenceleadstothinkintermsofV(x)=E[max{}]instead ofV(x)=max{E[]}asintraditionalDP, duetothepreviousassumptionabouttherandomdisturbance.Thevaluefunctionenteringperiodt, Vt(x), is then the expected value of this optimization (namely the maximum) with respect of the demandandisgiven intheformoftheBellmanequation: Equation1:
V (x) = E max pju + V -1 x - u t t u {0, }. 1

)}

14 DuetoBellman(1957),op cit.

76

(2007) 13 REVUEJURIDIQUEPOLYNSIENNE

WiththeboundaryV0(x)=0,whateverisx,andVt(0)=0,whateverist.Thismeansthatthe unsoldinventoryleftthelastdayisuselessandasunkcost. ThemotivationofEquation1usestheBellman'sprincipleofoptimality.SinceVt1(xu)isthe optimal expected value of the future revenue given the state (x u) in the next period, t 1, the optimal value of the tsubproblem should be the result of maximizing the sum of the current expectedrewardE[pju]andtheexpectedrewardforthet1subproblem,E[Vt1(xu)].Thisisthe optimalsolutionfortheperiod t andtheprocesscanberestarteduntilreaching T.

III

APPLICATIONOFDPTO FINDOUTBOOKINGLIMITSANDPRICING ISSUES

The RM problemisbasically formulated as a marginal one,such thatthe opportunity cost or expected marginal value (hereafter EMV) becomes the cornerstone of any model. The properties that any control must hold to be optimal are shown in A. Nevertheless, there exists various controls,presentedinB,asmanagersinpracticefigureoutfewcontroltypessuchasathreshold. ThereisalsoaclosecorrespondencebetweenDPmodelandBelobaba'sEMSRmodel. A Theexpectedmarginalvalueasthekeysolution Thevaluesu*thatmaximisetherighthandsideofequation1foreachtandxformanoptimal th controlpolicyforthismodel.ThesolutioncanbefoundbyevaluatingtheEMVatperiodtofthe x unitofcapacity. Equation2.:
D t( x) = V (x)- V (x - 1 V ) t t

Themostimportantconcernsarehowthismarginalvaluebehaveswithchangesinthecapacity 15 left x andthe t remainingperiods.Theexpectedmarginalvalueof Vt(x)satisfies "x,t: (i) DVt(x+1) DVt(x)

EMV is decreasing in x, meaning that the opportunity cost of using one more unit of the capacity increases when there are less products available. In other words, the fewer products on handthemorethechancetolooseasale. (ii) DVt(x) DVt1(x)

EMV is increasing in time period left or is decreasing as time elapses, meaning that the opportunitycostofusingonemoreunitofthecapacityishigherwhenthetimetogoislongthan

15 Therearedifferentproofsofthesestatements.However,theTalluriandVanRyzin(2004,ibid,38)proofis one of themost clever. Hence, the seat allocation problem and the dynamic pricing problem satisfy well known sufficient conditions for an optimal policy to be monotonic. These problems translate to the existence of timedependent controls. In other words, the optimality value function is concave and non increasing,fromwhichitfollowsthatanoptimaladmissionpolicyismonotonicinthestate.

BASICSOFD YNAMICPROGRAMMINGFORREVENUEMANAGEMENT

77

when it's short. In other words, the longer the remaining time the more the chance to use the capacityefficiently(theprobabilityofacustomerrequestingafullfareishigh). These two properties are intuitive and greatly simplify the control decisions because the equation1canberewrittenwithequation2.
V (x) =V -1( x)+ E max pju - DV -1( x - u) 0 t t t u {01 , } The meaning of this equation is the following. The value of therevenueto go today (in t)is equaltotherevenuetogotomorrow(thatisalreadyoptimisedusingarecursivemethod)plusthe expectedmaximumrevenueofthedecisiontomake(thatisthefarenetoftheopportunitycostif positive).Theproblemisformulatedasabackwardrecursiondynamicprogram.Inotherwords,an optimalbookingpolicyisreachedbytheassessmentofacceptingabookingrequestrelativetothe decrease in expected total revenue associated with removing one product from the available inventory.

Equation3:

B TypesofControls Threetypesofcontrolscanbederived,studyingthesignofvalueofthemaxfunction.Needless tosaythatrelaxingthehypothesisofearlybirdsimpliesthattheprotectionlevels,thebooking limits and the bid prices are timedependent because the random disturbance associated with the value functiondependsontheprobabilityofanarrivalofclassjintheperiodt.Inordertoanefficient optimization,themanagermustknowtheprocessofdemandfor eachfareclassovertime(iethe bookingcurve). 1 Protect levelsand nestedbooking limits

First consider controlling therevenuethrough u,the accept(u =1) or reject (u = 0) decision alternativestothecurrentrequest.Since DVt(x)isdecreasinginx,itfollowsthatpju DVt(xu)is increasinginx.Foragivent,itisoptimaltokeepacceptingincomingrequestsuntiltheprevious termbecomesnegativeortheupperboundmin{Dj x}isreached,whichevercomesfirst. Theprotection level yj*isthenumberofunitstosaveforcustomerswhorequest,orwillrequest inthefuture,classj orhigher(ie j,j 1,,1).Itisthemaximumquantity xsuchthatthefarelevel istoolowtocompensatetheopportunitycost,justifyingtherejectdecision(seeequation4). Equation4: yj*=max{x: pj+1 < DVt(x)},j=1,, n1.

Theoptimalcontrolattheperiod t+1isgivenbyequation5: Equation5:

+ u*(t+1 x Dj+1)=min{(x yj*) Dj+1 }.

Thebookinglimit b*j+1isthequantityisexcessoftheprotectionlevel(x yj*)ifpositiveorzero otherwise. This is the maximum space available for class j + 1 booking request. Those booking limitsarenested,meaningthataclassj requestcanbewithdrawninanyofthej classandabove(to n). By the way, thisset of protected classes(j, j1,,1)areproposed or "open" to customers

78

(2007) 13 REVUEJURIDIQUEPOLYNSIENNE

beingdeniedarequestfareclassbelowj.Thepolicythenis"simplyacceptrequestsfirstcome,first serveuntilthecapacitythreshold b*j isreachedorthetimeends,whichevercomefirst.Thefigure1 represents the accept/reject alternatives of a decision. For anillustration, considerin the figure 2 where n =3andp1 >p2 >p3. Figure 1:criticalbookingcapacity pju DVt1(x u)

pj+1 DVt(x)

x C

Rejectzone

Acceptzone

yj*

b*j+1

Legend:foragivent,thedecisiontoacceptorrejectarequestwithfarej+1dependsonxthe levelofcapacityleft.Thecriticalbookingcapacityseparatestheacceptzoneandtherejectzone. 2 Thecontrolbytimethresholds Second,considercontrollingtherevenuethrought,thetimeleft,toaccept(u=1)orreject(u= 0) an incoming request. Based on previous properties of equation 2, one can show that time thresholds characterize the optimal policy. During the booking horizon, they are points in time beforewhichrequestsarerejectedandafterwhichrequestsareaccepted.Since DVt(x) isincreasing int,itfollowsthatpju DVt1(xu)isdecreasingint.Thus,itisoptimaltokeeprefusingincoming requestsuntiltheprevioustermbecomespositive(seefigure3). This proposition isalso easy to understand with figure 1. Since DVt(x) isincreasing in t, one should accept more booking for a given j when remaining time lessens. In other words, as time elapses,theEMVdecreasesandthelineliftsup.Thentheoptimalprotectionlevely*j willshiftto theleft(decrease).The nestedprotectionstructureiskindof ... < y *tj 1 < y *tj < y *tj+ 1 < ...

BASICSOFD YNAMICPROGRAMMINGFORREVENUEMANAGEMENT

79

Figure 2:Casewith3fareclasses

Pu DV1(x u) j t p DV(x) 1 t p DV(x) 2 t p DV(x) 3 t x C

y* 2 y*1

b* 3 b* 2 b*1

3 Bidprices Finally,theoptimalcontrolcanalsobeimplementedthroughatableofbidprices,definedas pj+1(x)= DVt(x).Abidpriceistheminimumamountofmoneytoacceptinexchangeofaunitof capacity.Undermildassumptions(GallegoandVanRyzin,1994TalluriandVanRyzin,1998),the optimalpricepath pj(x)isdecreasinginxandincreasingint.Dynamicpricingisbecomingareal challenge for airlines since low cost carriers are introducing less restricted fares. For example, RyanairorEasyjetarenotabletobesurethatthemorepricesensitivepassengers(lowfare)book first.Thentheyapplyadynamicpricingstrategyforasingleproductwithoutpricediscrimination betweentheircustomersatagiventime.

80

(2007) 13 REVUEJURIDIQUEPOLYNSIENNE

Figure 3:criticaltimethreshold Pju DVt1(x u)

t T

Acceptzone

Rejectzone

Yj*

B*j+1

Legend:Foragivenbookingcapacity x,arequestforaseatoffareclassj indecisionperiodt is acceptediftimeremainingisshort(lessthanYj*,thecriticaltimethreshold)andrejectedotherwise. Thecapsdenotethatthecontrolisexpressedonatimeaxisinsteadofaquantityaxis.

IV

DISCUSSIONANDCONCLUSION

TheuseofDynamicProgramminginRevenueManagementhelpstodecidewhethertoaccept or reject an incoming booking reservation with more realism than older methods. There are two mainpoints.OneisthepropositionthatDPinRMallowstorelaxthelowbeforehighfareorderof arrivalbookings.Inpractice,theDPprovidestheoptimalpolicy fortheRMproblem,byevaluating thewholetreeofpossibilitiesandmakingateachpointintimethedecisionthatwouldimplyhigher future expected revenues, processing backward recursion. The dark side is the increase in the computation difficulties according to the dimension of the problem. This means that for a single productwith100unitstosellovera200timeperiods,thenumberofiterationsis100x200=20 3 000. But for three products using the same resource, this number becomes 100 x 200 = 800 millions.UnitedAirlinesexpectsa1to2%increaseinrevenue($158billionsof2005turnover) 16 duetotheimplementationofDP .

16 KSaranathanandWZhao "RevenueManagementintheNewManagementoftheNewFareEnvironment" (2005) ProceedsofAGIFORSReservationandYieldManagementStudyGroup,CapeTown.

BASICSOFD YNAMICPROGRAMMINGFORREVENUEMANAGEMENT

81

ThesecondinterestofDPapproach,andthemainavenueforfutureresearch,isthatitallows 17 RM to incorporate consumer choice within the optimisation process. ElHaber and ElTaha formulateadynamicprogrammingmodeltosolvetheseatinventorycontrolproblemforatwoleg airline with realistic elements of consumer behaviour. Ahead of the Origin and Destination formulation, they consider cancellation, noshows and overbooking. Following Talluri and Van 18 19 Ryzin's work, Van Ryzin and Vulcano consider a revenue management, network capacity control problem in a setting where heterogeneous customers choose among the various products offeredbyafirm(forexample,differentfighttimes,fareclassesand/orroutings).Customersmay thereforesubstituteiftheirpreferredproductsarenotoffered,evenbuyup.Theirchoicemodelis verygeneral,simplyspecifyingtheprobabilityofpurchaseforeachfareproductasafunctionofthe setoffareproductsoffered.Overall,thevalueof this paperistofacilitatetheunderstandingofmore complex,andprobablymorerealistic,modelsof revenue management.

17 SElHaberandMElTaha"Dynamictwolegairlineseatinventorycontrolwithoverbooking,cancellations andnoshows" (2004) JournalofRevenueandPricingManagement,vol3,n2,143170. 18 KTTalluriandGJvanRyzin "RevenueManagementunderaGeneralDiscreteChoiceModelofConsumer Behavior" (2003) ManagementScience,vol50,n1,1533. 19 G J van Ryzin and G Vulcano "Computing Virtual Nesting Controls for Network Revenue Management UnderCustomerChoiceBehavior" (2006) ColumbiaUniversityworkingpaper.

82

(2007) 13 REVUEJURIDIQUEPOLYNSIENNE