OATAO is an open access repository that collects the work of Toulouse
researchers and makes it freely available over the web where possible
Any correspondence concerning this service should be sent
to the repository administrator:
tech-oatao@listes-diff.inp-toulouse.fr
This is an author’s version published in:
https://oatao.univ-toulouse.fr/27139
To cite this version:
Salas, David and Van, Kien Cao and Aussel, Didier and Montastruc,
Ludovic Optimal design of exchange networks with blind inputs and
its application to Eco-industrial parks. (2020) Computers &
Chemical Engineering, 143. 107053. ISSN 0098-1354
Open Archive Toulouse Archive Ouverte
Official URL :
Optimal
design
of
exchange
networks
with
blind
inputs
and
its
application
to
Eco-industrial
parks
David
Salas
a,b,c,∗,
Kien
Cao
Van
a,
Didier
Aussel
a,
Ludovic
Montastruc
ba Laboratoire PROMES, UPR CNRS 8521, Université de Perpignan Via Domitia, Perpignan 66100, France b Laboratoire de Génie Chimique, UMR 5503 CNRS/INP/UPS, Université de Toulouse, Toulouse 31432, France c Instituto de Ciencias de la Ingeniería, Universidad de O’Higgins, Libertador Bernardo O’Higgins 611, Rancagua, Chile
a
r
t
i
c
l
e
i
n
f
o
Keywords:
Optimization Eco Industrial Park Game theory
Single-Leader-Multi-Follower
a
b
s
t
r
a
c
t
MotivatedbythedesignandoptimizationofthewaterexchangenetworksinEco-IndustrialParks(EIP), weinvestigatetheabstractBlind-Inputmodelforgeneralexchangenetworks.Thisabstractmodelisbased onaGameTheoryapproach,formulatingitasaSingle-Leader-Multi-Follower(SLMF)game:attheupper level,thereisanauthority(leader)thataimstominimizetheconsumptionofnaturalresources,while, atthelowerlevel,agents(followers)trytominimizetheiroperatingcosts.Weintroducethenotionof Blind-Inputcontract,whichisaneconomiccontractbetweentheauthorityand theagentsinorderto ensuretheparticipationofthelatteronesintheexchangenetworks.Moreprecisely,whenparticipating intheexchangenetwork,eachagentacceptstohaveablindinputinthesensethatshecontrolsonlyher outputfluxes,and theauthoritycommits toguaranteeaminimalrelative improvementincomparison withtheagent’sstand-aloneoperation.TheSLMFgameisequivalentlytransformedintoasingle mixed-integeroptimization problem. Thankstothisreformulation,examples ofEIP ofrealistic sizearethen studiednumerically.
1. Introduction
Inthe last fewdecades,the developmentoftheindustrialized countries hasled to an increasing depletion ofnatural resources suchasfreshwaterandenergy(see,e.g.,UNEP,2000;Scientificand Organization),2009).Theconservationandsustainableuseofsuch resources play an important role in both, environmental impact andbusinesssuccess withintheindustry.Inresponse topreserve the environment while increasing the utilities of the enterprises, theconceptofindustrialecologyhasemerged(Boixetal.,2015).
Industrial ecology (IE) wasfirstintroduced in Froschand Gal-lopoulos(1989).Theywrote“theconsumption ofenergyand ma-terials isoptimized,wastegenerationis minimizedandthe efflu-entsofone process... serveastherawmaterialforanother pro-cess”.Thisisanapproachtotheindustrialdesignofproductsand processes and the implementation of sustainable manufacturing strategies. The idea is directly relatedto another concept, indus-trial symbiosis, whichinvolves “separate industriesin a collective
∗ Corresponding author at: Laboratoire de Génie Chimique, UMR 5503
CNRS/INP/UPS, Université de Toulouse, 31432 Toulouse, France.
E-mail addresses: david.salas@uoh.cl (D. Salas), kien.van@promes.cnrs.fr (K.C. Van), aussel@univ-perp.fr (D. Aussel), ludovic.montastruc@ensiacet.fr (L. Montas- truc).
approachtocompetitiveadvantage involvingphysicalexchangeof materials,energy,waterand/or by-products” (seeChertow,2000). Onekeyconcept ofindustrialsymbiosisis thentheexchange net-works.
A perfect example of an exchange network which illustrates thenotionofindustrialsymbiosisistheconceptofEco-Industrial Parks(EIP).Thisnotionhasseveraldefinitions,butonewidely ac-cepted is “an industrial system of planned materials and energy exchangesthat seeks to minimize energy andraw materials use, minimize waste, and build sustainable economic, ecological and social relationships” Alexander et al. (2000); Boix et al. (2015);
Montastrucetal.(2013).
Recently,inworksofBoixetal.(2015)andKastneretal.(2015), ithasbeenpointedoutthatthereisstillalackofsystematic meth-odsfordesigning theoptimalconfigurationofan EIP.Inprevious studies (Boix et al., 2011; 2015; Montastruc et al., 2013), water integration networks (which is a classical example of EIP) were modeled as a cooperative economy, in the framework of multi-objectiveoptimization(MOO).Thisapproach consistincreatinga vectorfunctionofn+1coordinatesgivenby
C
(
F)
=Cost1(
F)
,...,Costn(
F)
,Z(
F)
where Costi( · ) is the cost function of the enterprise i, Z( · )
isthe global consumption ofnatural resources, andF is the flux
Nomenclature
Latinsymbols
n numberofindependentagents
m numberofregulatedagents
P setofindependentagents
R setofregulatedagents
IP indexsetofindependentagents IR indexsetofregulatedagents I assemblyofindexsetsIPandIR I0 assemblyofindexsetIandsinknode0
E networktopology
Emax setofalladmissibleconnectionsofthenetwork
Ec setofconnectionsthatarenotinE Ei,act setofactivearcsofagenti
Est stand-alonetopology E setofallvalidtopologies Costi(· ) operatingcostofagenti
STCi stand-alonecostofagenti C(i,j) arcclassof(i,j)
Ci familyofallarcclassesexitingfromagenti D setofallarcclassesofactiveagents
y booleanvariable
xi,j fluxthroughtheconnection(i,j)
xi outletfluxvectorofagenti
x−i vectorofallfluxesnotexitingfromagenti xP
−i vector ofall fluxesexiting froman independent
agentotherthani
x completevectoroffluxesthroughthenetwork
zi consumptionofnaturalresourceoftheithagent Z(· ) totalconsumptionofnaturalresources
gi(· ) inputvalidationfunctionofagenti
Fi vector of fluxesexiting from enterprise i (water
exchangenetwork)
F−i vector ofall fluxes not exiting fromenterprise i
(waterexchangenetwork)
FP vector of fluxes exiting from enterprises (water
exchangenetwork)
FP
−i vector of all fluxes exiting from an enterprise
otherthani(waterexchangenetwork)
FR vector of fluxes exiting from regeneration units
(waterexchangenetwork)
F fluxvectordescribingthedistributioninthe wa-terexchangenetwork
Mi contaminantloadofenterprisei[g/h]
Ci,in,Ci,out maximumcontaminant concentration allowedin
inlet/outletofprocesses[ppm]
Cr,in minimum inlet concentration allowed of reg.
units[ppm]
Cr,in exact outlet contaminant concentration of reg. units[ppm]
A thelifetimeofthepark[h]
Coef Penalizationcoefficientofstand-aloneagents
Acronyms
EIP Eco-IndustrialPark
GNEP generalizedNashequilibriumproblem Eq thesetofequilibriafortheinducedGNEP KKT Karush-Kuhn-Tucker
MILP Mixed-integerlinearprogramming
MPEC mathematicalprogramswithequilibriumconstraint SLMF Single-Leader-Multi-Follower
STC stand-alonecost
Greeksymbol
α
the minimal relative gain that each agent ask for participatinginthenetworkc themarginalcostoffreshwaterconsumption[$/T]
β
i,0 thedischargecostofpollutedwaterofenterprise i[$/T]
δ
i,j thecostsendingpollutedwaterfromenterpriseito j[$/T]r themarginalcostofregeneratingwater[$/T]
ψ
powerassociatedtor
vector describingthe distributioninthe exchangenetwork. Then, the aim is to solve the problem of “minimizing” C with respect to F, satisfying the physical constraints of the model. The result of such minimization is called a Pareto front, which consists in all vectors F for which noneof the coordinates of C can be im-provedwithoutworsenanotherone(McCain,2010;Emmerichand Deutz,2018).Usuallyanauthority,representingtheEIP’sdesigner, selectsoneofthissolutionsconsideringascriteriathedistanceto anutopiapoint.
Themainproblemwithsuchan approachisthatpointsofthe Paretofrontarenotnecessarilyeconomicallystable:first,aPareto pointrequirestheenterprisestocooperateandshareinformation, whichisrarelythe caseofan EIP.Second, dueto the noncooper-ativeeconomy,thedifferententerprisesmaydeviate fromthe se-lectionoftheauthoritysincetheymayimprovetheircostfunction byunilaterallychanging theiroperation.Intermsofgametheory, asolutionoftheMOOapproachisasocialoptimizationwhichmay failtorespectincentives(seeNisanetal.,2007,Chapter1).
Tosolvethisincompatibility,againinthe contextofwater in-tegration networks, in the seminal work of Ramos et al. (2016), further developed in Ramos et al. (2018b), a novel game theory approach has been proposed, by modeling the EIP design prob-lem as a Single-Leader-Multi-Follower (SLMF) game (see Aussel and Svensson,2020; Hu and Fukushima, 2015): since the agents donotwanttoexchangeinformation,aconfidentialcentralization through an authority of the parkis introduced. Then, at the up-per level,there istheEIPauthority whichwantstominimize the consumption of natural resources Z(F), while at the lower level, each enterprise tries to minimize her cost function Costi(F), re-latedtoherprocesses,consumptionofnaturalresourcesand activ-itywithintheEIP.Theauthorityoftheparkmustchoosethe con-nectionsoftheexchangenetworkandtheoperationofthe regen-erationunits,whileeachenterprise controlstheir consumptionof naturalresourcesandtheir output fluxdistribution.Based onthe EIP authority decisions, all enterprises compete with each other ina parametric non-cooperative generalizedNash game withthe strategies of the EIP authority as exogenous parameters. Fig. 1.1
showsthegeneralschemeofsuchamodel,wheretheenterprises are considered the economic agents of the game. We refer the readertoNisanetal.(2007);Ichiishi(1983)fora primerin non-cooperativegames, toPangandFukushima (2005); Facchineiand Kanzow(2010)forasurveyofGeneralizedNashEquilibrium prob-lems,andDempeetal.(2015);DempeandZemkoho(2020)forthe theory of bilevel optimization. For Single-Leader-Multi-Follower games, we refer to Hu andFukushima (2015) andthe references therein.
The main implicitassumption done in Ramos et al. (2016) is thateachenterprisecanonlycontrolheroutletdistributionandher ownfreshwaterconsumption,buttheyare forcedtoaccept what-everissenttothem throughtheexchangenetwork.Furthermore, they haveno knowledgeabouttheparticularactionsoftheother agentsof thenetwork, excepting onlythe amountandquality of the final inlet flux. In practice, this situation corresponds to the
Fig. 1.1. General scheme of SLMF game.
Fig. 1.2. Blind-Input Schema. z i , F k,i and F r,i are freshwater consumption, wastewater sent from agent k to i , and regenerated water sent from regeneration unit r to i , respectively.
casewhen,attheentrance,eachagentofthenetworkhasamixer, and so she is only aware of the total input she is receiving, as
Fig. 1.2 illustrates. In other words, when participating in the ex-changenetwork,eachagentacceptstohaveablindinput.
While this model respects incentive consistency, it has two main drawbacks:thefirstone isthat therulethat thepark’s au-thorityimposes, that is,theblind input,istoorestrictive. Indeed, under thisparadigm, an enterprise may be forcedto receive too much pollutedwaterwhichcouldturnintohighercosts thanthe stand-aloneoperationoutsidethepark(examplesareeasyto con-structwithtwoenterprises).Thisviolatestheeconomicalprinciple (wellknownincontracttheoryandmechanismdesign)of individ-ual rationality: an enterprise will participatein the EIP only ifit isconvenienttoher(see Jackson,2014;Salanié,2005;Boltonand Dewatripont, 2005);thesecond one is thestrategy tocompute a solution.InRamosetal.(2016),theauthorsimplementedthe clas-sic general approach to solve bilevel games, that is, to reformu-late it as a mathematical programming with complementarity con-straints (MPCC): looselyspeaking,fora givennetwork,they write the Karush-Kuhn-Tucker(KKT) conditionsof each problemofthe lower level game, andput themas constraintsin the authority’s problem.ThentheyimplementedaBranch-and-Boundheuristicto obtainanapproximatedoptimalexchangenetwork,solvingateach iteration theproblemdescribedabove.However, itis knownthat the MPCC problems, which is a particular class of mathematical programmingwithequilibriumconstraints(MPEC), arehardtosolve (see, e.g., Baumruckeret al., 2008; Tseveendorj, 2013;Luo et al., 1996) andtheheuristic itself doesn’tguaranteea realsolutionof theproblem(AusselandSvensson,2019;DempeandDutta,2012). Theliterature ontheoreticalandalgorithmicaspects ofMPCC and MPEC problems is large and still an active field of research in mathematics.
In this work, we further investigate the model proposed in
Ramosetal.(2016)forwaterexchangenetworks,brieflydescribed in Section 2 and fully exposed in Section 5, but considering its abstract formfor generalexchange networksin Section 3.2. This abstractmodel is calledBlind-Input model, since we consider the constraintoffullacceptanceforeachenterprise.Tosolvethe draw-backgivenbytheIndividualRationalityconstraint,weintroducein
Section3.3thenotionofBlind-Inputcontract,whichisan econom-icalcontractbetweentheauthorityandeachenterpriseinorderto participatein theBlind-Input model.We prove that, undersome linearstructureofthecostsfunctionsCosti( · )ofeachenterprise, theBlind-Inputmodelcanbereducedfroma Single-Leader-Multi-Followerproblemtoasinglemixed-integer optimizationproblem. This reduction, which is our main contribution, is presented in
Section4.
TheproposedreformulationoftheBlind-Inputmodelopensthe doortoalotofnewdevelopments,fromthenumericaltreatment ofhugesizeproblemsthanks toclassical MILPsolvers to exhaus-tivesearchofequilibriaforsmall/mediumsizeapplications.Thisis illustratedinthesecondpartofthearticleforwaterexchange net-worksinEco-IndustrialParks:Section6illustratesacaseofstudy and the obtained results which are then discussed in Section 7. ConclusionsandperspectivesarepresentedinSection8.
Itisworthtomentionthat,eventhoughthisworkismotivated by the design problem of water exchange networks, its abstract formulationpresentedinSection3allowstoapplyittoothertype of networks, asfor example energy networks (Boix et al., 2015; Nevesetal.,2020). InSection8,we willcommentwhichare the main elements needed to apply the Blind-Input model to other contexts.
Tosurvey our contributions, a comparison betweenthiswork andRamosetal.(2016)isgiveninTable1.Itisimportantto men-tionthat the nooncoperative approachusing SLMFgames inEIPs isvery recentand, up to ourknowledge, thereis noother refer-enceintheliteraturedifferentfromRamosetal.(2016,2018b)to compareourresultswith.
2. Motivation:EIPmodelforwaterexchange
Inthissection,webrieflydescribethemodelofwaterexchange networkused to describeEco-Industrial Parks. The model canbe found inRamos etal.(2016); Boixet al.(2015) among others.A detailedversionisfurtherexposedinSection5.
Table 1
Comparison between Ramos et al. (2016) and the present work. The first two rows are related to the numerical exam- ples used in each article.
Comparison criteria Ramos et al. (2016) . This work
Number of enterprises 3 15
Number of processes per enterprise
5 1
Regeneration units Yes Yes
Admits multiple processes per enterprise
Yes No
Tools to model the EIPs SLMF game SLMF game
Presence of Blind-Input model Implicitly used. Not formalized. Economic drawbacks.
Explicit formalization. Introduction of Blind-Input contract as economic instrument.
Solution Method MPCC reformulation + Branch-
and-Bound Heuristic
Mixed-Integer Linear programming (MILP) reduction. Properties of the solution MPCC is hard to solve and existing
algorithms are not robust. The solution of the MPCC may fail to be a solution of the SLMF game.
MILP alogirthms are robust. Commercial solvers are available. Any global solution of the MILP problem is a global solution of the SLMF game.
The operating cost of each participating enterprise in the EIP is lower than that of stand-alone.
No Yes
In an Eco-Industrial Park (EIP), several enterprises exchange wastestoreducetheglobalconsumptionofnaturalresources.Each timeanenterpriseusesthenaturalresourceinherindustrial pro-cess, it comes out degraded, but still can be used as input for otherenterprisesin thepark.Oneofthe mostclassicalexamples ofEIP (see,e.g.,Boix etal., 2015;Boixetal.,2012corresponds to themodelingof waterexchangenetworks: eachenterprise needs toconsumewaterforher industrialprocessesandtheoutcoming waterispartiallypolluted.Other examplesusingdifferentnatural resources likeenergy orheat can be found in Boix etal. (2011);
Ramosetal.(2018a).
InRamosetal.(2016),thedesignofawaterexchangenetwork istreated accordingto the following assumptions:first, the park hasafixednumberofnenterprises,eachenterpriseihastodilute anamountMiofcontaminant,andtheoutletconcentrationof
con-taminantmustbelessthanafixedconcentrationCi,out.Itisusually
assumedthateachenterpriseihasalwaysanoptimaloperation,in thesense that theoutlet concentration ofcontaminant is always equaltoCi,out.
Second, each enterprise i can accept partially polluted water, butwithamaximalconcentrationCi,in.Thisconcentrationis
mea-suredafteramixer(seeFig.1.2)insuchawaythatnoenterprise canreally knowthe operation ofthe other enterprises. However, thismeasurement,that we willdenote gi andwhichdependson
theactionsoftheotherenterprises,allowsenterpriseitoperform twofundamentalactions:(1)reportinfeasibilitiestotheauthority ofthe park, wheneverthe income water afterthe mixer doesn’t fulfilltheconstraints;and(2)computehowmuchfreshwatershe needs to complete its process attaining the outlet concentration
Ci,out.
Third, each enterprise has a cost function that depends on four factors: (1) the marginal cost of fresh water that she con-sumes, that we denote ci; (2) the marginal cost ofpolluted
wa-ter that she dischargesto the environment, that we denote
γ
i,0;(3) the cost of sending polluted water through a connection of the park; and (4) the cost of receiving water from other agents of the park (other enterprises but also regeneration units con-trolledby the authority). The authority transfers the investment costofthe EIP tothe enterprisesvia thelast two costs:the first one, via a marginal cost
γ
which depends on the connections that enterprise i uses to send water; and the second one via a costfunction Costini that will depend onthe actions ofthe other enterprises.Moreover,themainassumptionsforthepricinginstrumentsare thatthepricesoffreshwateranddischargedwaterareexogenous, andthat theauthority hasno interest of making anyprofit, and thereforeshe willfix the pricesofusing theconnections only to recoverthe investmentandmaintenance costs. Thisyields to the following scenario:each enterprise wants to minimizeits cost of theuseofwaterwhiletheauthorityisinchargeoftheecological concernsbyminimizingthefreshwaterconsumption.
Finally,aswementionedbefore,theauthoritymayhave regen-erationunits.Eachregenerationunitrreceivespollutedwaterand reducesitscontaminantconcentration uptoacertain valueCr,out.
Then,itsendsthewatertotheenterprisesforreuse.Thecosts as-sociated to the regeneration units are chargedto the enterprises throughtheinletcostfunctionCostini .
3. Blind-inputmodel
Taking inspiration from the water management model de-scribedinSection2,ouraiminthissectionistodefinetheconcept ofabstract Blind-Inputmodel forgeneralexchange networks. We divided themodel intwo parts: the physicalmodel, whichgives theconstraintsthatthenetworkmustsatisfy;andtheeconomical model,whichgivestheincentivesofeachagentofthenetwork,as well as theBlind-Input contract betweenthe agents andthe au-thority,whichwillensuretheparticipationoftheagents.
3.1. Networkmodel
Wefirstconsidertwomainactors:asetofagentsparticipating to an exchangenetwork, andan authoritythat aims to minimize theconsumption ofnaturalresources. Among the agents,we dif-ferentiate aset P:=
{
P1,...,Pn}
ofindependent agents,andaset R={
R1,...,Rm}
ofregulatedagents(controlledbytheauthority).Regulatedagentsdon’t haveeconomicalmotivations,butthey act ontheexchange networkfollowing theindicationsofthe author-ity.InthecontextofwaterexchangeinEIP,theindependentagents aretheenterprises,andtheregulatedonesmodeltheregeneration units(Ramosetal.,2016).
We identify the independent agents with the index set IP=
{
1,...,n}
andtheregulatedagentswithIR={
n+1,...,n+m}
.WesetI=IP∪IR andI0=
{
0}
∪I,where0representsthesinknode.Wedefine an exchangenetworkasa simpledirectedgraph(I0,
part of her output to the agent j. The extra node 0 is identified asasinknode,whichrepresentthepossibilityofdischargeofthe output.Avalidnetwork(I0,E)mustsatisfythefollowingfive
con-ditions:
I. E⊆Emax,whereEmax isthesetofall admissibleconnections of
thenetwork.
II. (I0,E)isasimplegraph,thatis,thereisnomultipleedges nor
graphloopsinE.
III. Eachindependentagenti∈IPisconnectedwiththesinknode,
thatis,(i,0)∈E.
IV. Eachregulatedagentr∈IRisnotconnectedwiththesinknode,
thatis,(r,0)∈E.
V. The sinknode has not exitedges inE that is(0, i)∈Emax, for
anyi∈I.
Inwhatfollows,wewillcallEthetopology ofthenetwork(I0,
E),andwewilldenotebyE the setofallvalidtopologies. Never-theless, inorderto simplifynotations,thenetwork(I0, E) willbe
onlyrepresentedbyitstopologyE.Observethatthisrepresentation mayleadtoambiguity,sincethesetEdoesn’tallowtodistinguish possible isolated regulated agents (independent agents are never isolated,givenhypothesisIII).However,thisisnotaproblem,since anyisolatedregulatedagentwillbesimplyremovedfromthe net-work.
Foreach edge (i, j) ∈Emax, weset thevariable xi,j which
rep-resents the flux through the connection (i, j). For each i ∈ I, we setxi:=(xi,j:(i,j)∈Emax),beingthustheoutcomevectorofagent
i.Finally, weset x=
(
xi, j :(
i,j)
∈Emax)
, thecomplete vector offluxesthroughthenetwork.
To simplify the mathematical model we use, let us introduce some notation. Weput xR := (x
r:r∈IR)andxP := (xi:i ∈IP). In
whatfollows,foranagenti∈I,wewillwrite
x−i:=
xk, j :(
k,j)
∈Emax, k∈I\
{
i}
, xP−i:=xk, j :(
k,j)
∈Emax, k∈IP\
{
i}
.ForatopologysubsetA⊆Emax,wewrite
x
A:=(
xi, j :(
i,j)
∈A)
.Similarly,wedefinexi|A,x−i
A,x P|A,xP−i
AandxR|A.ItwillbeusefulalsotodenoteAc:=E
maxࢨA.
3.2. Physicalmodel
Letusfixa networktopology E∈E. IfE isimplemented,then foreachagenti∈I,thephysicalmodelofthenetworkisgivenby thefollowingsixoperationalconstraints:
1. Nullfluxesoutsidethenetwork:each agentcanuseonlythe connectionsinthetopologyE.Thus,weset
xi
Ec=0, (3.1)thatis,foreveryedge(i,j)∈E,thefluxxi,j iszero.
2. Consumptionofnaturalresource:theconsumptionofnatural resource ofthe ith agentis givenby the output fluxesofthe otherplayers,thatis,
zi=zi
(
x−i)
. (3.2)Thisassumption is derived froman optimal response hypoth-esis: we assume that, fora given value ofx−i, the agenti is capableofcomputeexactlytheminimalamountofnatural re-sourcezithatshehastoconsumeinordertoperformherinner
processes.
3. Balance constraint:the fluxes must satisfy the Kirchoff’slaw fortheagenti∈I,thatis,
zi
(
x−i)
+ (k,i)∈E xk,i= (i, j)∈E xi, j. (3.3)Since 0isthe sinknode,it isnotsubjectto thisbalance con-straint.
4. Inputconsistency:thereexistsareal-valuedfunctiongiwhich
allows the agenti ∈I to validate the input coming from the other agents.Wewritethisvalidationasan abstractinequality constraint
gi
(
x−i)
≤ 0. (3.4)Thisconstraintmayrepresentmaximalinletfluxes,maximal in-letcontaminantconcentration,minimalinlettemperature,etc. 5. Positivityoffluxes:weassumethatthefluxesonthegraph,as
well astheconsumednaturalresource areallpositive, thatis,
xi≥ 0 and zi
(
x−i)
≥ 0. (3.5)6. Extra authority constraints: the exchange network may re-quireadditionalconstraints.Wewillmodelthemherethrough anabstractinclusion
x∈X,
where X⊂ R|Emax| represents the abstract additional feasible set.
Remark3.1. Here,weassume thatthedegradationofthenatural resource is implicitin the connections of the topology E.In this generalmodel,wesupposethatagenticancomputethe degrada-tionofitsinletfluxthroughthefunctionsgiandzi.
Animportantelement ofthismodel isthe totallack ofdirect information among the agents. We suppose that agent i cannot know the actions of other agents, that is, she doesn’t have ac-cessto theexact value ofx−i.However, she countswithindirect observations: even though x−i is unknown, the values of zi
(
x−i)
,gi
(
x−i)
andthe total inlet flux(k,i)∈Exk,i are available. For wa-ter exchange, thiscould be interpreted asa measurement of the amountofwaterandcontaminantconcentrationafterthemixerof
Fig.1.2.Thisisaveryimportantfeatureofourmodel,since enter-priseswanttokeep asmuch private informationaspossible.The onlyagentthathasallinformationistheauthority,whohasaccess tothefullvectorx.
3.3.Economicalmodel
Inthissetting,thenetwork authorityhastwo vectorsof deci-sionvariables:shemustchoosethetopologyofthenetworkE∈E
andshecontrolstheoperationoftheregulatedagents,thatis,the outputvectorsxr,forevery r∈IR.Each independentagenti ∈IP
controlsheroutputvectorxi.
Weassume that the authoritydoesn’t payanycost associated totheimplementationandoperationofthenetwork.Instead,she transfersall thesecosts througha function
γ
:Emax→R+,whereγ
((
i,j))
=γ
i, jrepresentsthemarginalcostforsendingoneunitoffluxthroughtheconnection(i,j). Usingthispricing, the indepen-dentagentswillpaytheinvestmentcostofthenetworkandalso theoperationoftheregulatedagents.Thus,ifthereisaconnection (r1,r2)∈Emaxbetweentworegulatedagentsr1,r2∈IR,weassume
that
γ
r1,r2=0.Sinceall the investmentcost is transferedto the independent agents,theauthorityisonlyconcernedaboutminimizingthe con-sumptionof the naturalresources, andso she aims to minimize thefunction
Z
(
x)
:= i∈Izi
(
x−i)
. (3.6)Remark 3.2. It could be argued that the authority must be also concernedaboutefficiencyofthenetwork,byconsideringthetotal investmentcostofthepark.However, weassumethatthepricing
instrument
γ
isgivenexclusivelytopaytheinvestmentand main-tenancecostofthepark,andthatit willbeimplementedas effi-cientlyaspossible.Thediscussionoverefficiencyandrightpricing instruments,isoutofthescopeofthiswork.Ontheotherhandanyindependentagenti∈IPwantsto
mini-mizeherglobalcostCosti,whichcanbeseparatedintothree
com-ponents:theconsumptionofthenaturalresourcezi
(
x−i)
,thecostofdischarging(usingtheconnection(i,0)),andtheuseofthe ex-changenetwork.ThereforehercostfunctionCostiisgivenas: Costi
xi,xP−i,xR,E =ci· zi(
x−i)
+Costini xP −i,xR + (i, j)∈Eγi
, j· xi, j. (3.7) whereCostini xP −i,xRisthe inletoperatingcostofanagenti,and itsatisfiesthat (k,i)∈Emax xk,i=0⇒Costini
xP −i,xR =0.Observe that, the cost concerning the exit connections is linear, andso,thecostfunctionislinearinthefirstcomponentxi.
Remark3.3. Again, in termsofcosts, agent i doesn’thave direct accessto theactions ofthe otheragents.However, shemust pay an operating cost Costini
(
xP−i,xR
)
that is communicated to her bythe authority. The choice of this function as pricing instrument could be studied, but this is out of the scope of the work. For now,wewillsupposethatagentihasenoughindirectinformation (throughmeasurementsafterthemixerofFig.1.2)toconsiderthe costCostini
xP−i,xR
ascorrectandthereforetoacceptit.
Withthismodel,theminimizationproblemoftheith indepen-dentagent(parametrized by the topology E, the actions of regu-latedagents xR andthe actions ofthe other independent agents xP
−i)leadstoproblemPi
xP −i,xR,E : minxi Costi xi,xP−i,xR,E s.t.⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩
zi(
x−i)
+ (k,i)∈E xk,i= (i, j)∈E xi, j gi(
x−i)
≤ 0 zi(
x−i)
≥ 0 xi≥ 0 xiEc=0. (3.8)Wedenoteby Eq(xR,E)the setofequilibriafortheinduced
gen-eralizedNashequilibriumproblem(GNEP,forshort) givenbythe vectorxR andthetopologyE,thatis
xP∈Eq
(
xR,E)
⇐⇒∀
i∈I P,xisolvesPi xP −i,xR,E . (3.9)AswealreadydiscussedinSection1,themainproblemofthis model is that each independent agent only controls her output vectorxi,whichis not realistic.She isforcedby the authority to
fullyacceptanyinletfluxes,whichmaybeharmful.Thus,without anyextraconstraint, agent i maynot be willing toparticipate in thenetwork.
Thus, tosolve thisproblem, theauthority must“buy” the par-ticipationof agenti. Thisis modeled by the Blind-Input contract: agentiacceptstocontrolonlyheroutputfluxes,andtheauthority commitstoguaranteeaminimalrelativeimprovementofhercost, withrespecttothestand-aloneoperationofagenti.
Toformalizethisrequirementinthecontract,letusdenotethe stand-alonetopologybyEst∈E,thatis,
Est:=
{
(
i,0)
: i∈IP}
.Foreach independentagenti∈IP wedefine thestand-alonecost
STCi,astheoptimalvalueoftheproblemPi(0,0,Est),thatis,
STCi=
(
ci+γi
,0)
· zi(
0)
Inotherwords,STCi isthecost oftheithagentassumingthat all
other agents (independent andregulated) are inactive, i.e. when agentionlysend fluxesto thesinknodeanddoesn’treceiveany complementaryfluxesfromother agents.Then, foreach indepen-dentagentPi,we canformulate thecommitmentofminimal
im-provementintheBlind-Inputcontractasthefollowingconstraint:
Costi
(
xi,xP−i,xR,E)
≤α
· STCi, (3.10)where
α
∈ ]0,1[istheminimalrelativegain thateachagentask forparticipatinginthenetwork.Weassumethatα
>0since,itis impossibletoeliminateallcosts,andthatα
<1sincenoagentis indifferentconcerningher participationinthenetwork.Indeed,if Costi(
xi,xP−i,xR,E)
=STCi,then the agenti willprefer not topar-ticipate, since she has no gain, entering an exchange network is complicatedandsheknowsshemaybe“helpingthecompetition”.
Finally,wecanwritetheauthority’sproblemas
minE∈E,x∈R|Emax| Z
(
x)
s.t.
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
x∈X, zr(
x−r)
+ (k,r)∈E xk,r= (r, j)∈E xr, j,∀
r∈IR, zr(
x−r)
≥ 0,∀
r∈IR, gr(
x−r)
≤ 0,∀
r∈IR, xR≥ 0, xR Ec=0, xP∈Eq(
xR,E)
, Costi(
xi,xP−i,xR,E)
≤α
· STCi,∀
i∈IP. (3.11)Theoptimizationproblem(3.11)canbe interpretedasfollows: theauthoritywillproposetotheagentsatopologyEandan oper-ationx∈R|Emax| whichsatisfyallthephysicalconstraintsandalso, such thattheoperation xrespects:1) theincentiveconsistency,in the sense that no agentwillhave incentivesto unilaterally devi-atefromtheproposalduetotheconstraintxP∈Eq(xR,E);and2)
theindividualrationalityofeachagent,inthesensethatallagents will participateinthe networksince their participation hasbeen bought through the constraint (3.10). The first criteria solves the economicalinconsistencyofMOO approach,andthesecond crite-riasolves theparticipation problemof theSingle-Leader-Follower approach.
Remark 3.4. In this work, we do not claim novelty in the
constraint xP ∈ Eq(xR, E). This is the main contribution of
Ramosetal.(2016).However,theconstraint(3.10)isnew.Interms ofmodelingandinthiscontext,thefactto“attract” the indepen-dent agentstowards a participation inthe generalexchange net-workconstitutesoneoftheimportantnoveltiesofthiswork.
Remark 3.5. After reading the forthcoming Section 4, thereader will observe that all proofs and reductions could be made con-sideringdifferentvaluesof
α
foreach independentagent, thatis, puttingavalueα
i∈ ]0,1[ foreach i ∈IP.Thevalue ofα
irepre-sentsthe“cost” ofbuyingtheparticipationoftheithindependent agent,whichisexactly
(
1−α
i)
STCi.However,allowingtohavedif-ferentcostsdependingontheenterpriserisesthenaturalquestion ofhow todecide thesevalues.This problemliesincontract the-ory(foranintroductiontothefield,werefertoBoltonand Dewa-tripont(2005); Salanié (2005))anditisout thescope ofthe arti-cle.Thus,wewillconsideronlyuniformvaluesof
α
,whichcanbe interpreted asa publiccall forparticipationin thenetwork. Uni-formvaluesofα
,however,implythat thecostofbuyingthe par-ticipationofanagentisproportionaltohersize,duetothefactor STCi.Remark3.6. Animportantfactorwedonotconsiderinthiswork isthereboundeffectthatcostsreductionsmayhaveonthe opera-tionofagents.Forexample,ittermsofwaterexchange,a
diminu-tionofcosts ofagentiwithrespecttoSTCimayinducean
incre-mentofwastesproduction,thatis,avariationinMi.Thus,this
re-boundeffectmaychangethevalueofzi
(
x−i)
.Eventhoughthisisavery interestingproblem, wesuppose thatthedemand ofnatural resourceisgivenbyafixedprocess,onwhichthecostswithinthe networkhavenoeffect.Inotherwords,theconsumptionofnatural resourceofeachagentisinelastic.
4. Mixed-integerprogrammingreduction
Theformulationoftheauthority’sproblem(3.11)hastheform of a general MPEC problem (see, e.g., Baumrucker et al., 2008; Tseveendorj, 2013; Luo et al., 1996). This section is devoted to prove that thisMPECformulation,which isknown tobe hard to solve,canbereformulatedasasingleMixed-Integerprogramming problem.
Thisreductioncanbe interpretedasfollows:Blind-Input mod-els are a social optimizationproblemwhere,through Blind-Input contracts, the cooperation of each independent agent has been bought.This socialoptimizationis alsoeconomicallystable, since implicitlyitrespectanequilibriumconstraint(xP∈Eq(xR,E)).This
reduction/reformulationwillbepresentedinthreesteps.
4.1. Characterizationofequilibria
The followingtheoremcharacterizestheequilibriumsetEq(xR, E) as a system of equations.This allows to reduce the MPEC of problem (3.11) to a single optimization problem. The reduction we do here is based on the observation that, once every agent hascommittedtoaBlind-Inputcontract,her actionsbecome pre-dictablethroughthecostfunctions.Thus,theauthoritycanchoose the network E such that each action ofan independent agentis inducedtoreachthesocialoptimum.
Toformalizethisidea,letusintroducethenotionofactivearcs. Given a topology E, foreach independent agenti ∈IP we define
the set ofactive arcs ofi,denoted by Ei,act, asall the arcse ∈E
havingminimumcost,thatis,
Ei,act:=
(
i,j)
∈E :γ
i, j=γ
i∗:=(mini ,k)∈Eγi,k
. (4.1)As convention, for any regulated agent r ∈ IR, we set Er,act=
{
(
r,j)
:(
r,j)
∈E}
.Theorem4.1. ForE∈E andxR ≥ 0fixed,theequilibriumsetEq(xR, E)isgivenby Eq
(
xR,E)
=⎧
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎩
xP :∀
i∈I P, zi(
x−i)
+ (k,i)∈E xk,i= (i, j)∈E xi, j gi(
x−i)
≤ 0 zi(
x−i)
≥ 0 xiEc i,act =0 xi≥ 0⎫
⎪
⎪
⎪
⎪
⎪
⎬
⎪
⎪
⎪
⎪
⎪
⎭
(4.2)Thus, the authority’s problem (3.11) is equivalent to the following Mixted-IntegerProgrammingproblem:
minx∈R|Emax|,E∈E Z
(
x)
s.t.
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
x∈X, zi(
x−i)
+ (k,i)∈E xk,i= (i, j)∈E xi, j,∀
i∈I xiEc i,act =0,∀
i∈I gi(
x−i)
≤ 0,∀
i∈I zi(
x−i)
≥ 0,∀
i∈ICosti
(
xi,xP−i,xR,E)
≤αi
· STCi,∀
i∈IPx≥ 0.
(4.3)
Proof. Thesecond partoftheproof iseasilyverifiedbyreplacing theconstraintxP∈Eq(xR,E)bythesystemofequationsintheright
handof equality(4.2),and thenjust reorganizing. Thus, we only needtoprove(4.2).
Tosimplify notation, let us denote by S(xR, E) the right-hand
setof(4.2).First,letusprovethatS(xR,E)⊆Eq(xR,E).FixxP∈S(xR, E). SinceEi,act⊂ Eforeach i ∈IP,itis nothard toseethat xi isa
feasiblesetofPi
(
xP−i,xR,E)
.Now, fix i ∈ IP and let xi be another feasible point of Pi
(
xP−i,xR,E)
. Then, xi≥ 0 and it satisfies the balance constraint(3.3),whichyieldsthat
Costi= (i, j)∈E
γi, j
xi, j−γ
∗ i (i, j)∈Ei,act xi, j ≥γ
∗ i (i, j)∈E xi, j− (i, j)∈Ei,act xi, j ≥ 0,where
Costi:=Costi
(
xi,xP−i,xR)
− Costi(
xi,x−iP,xR)
andthelastin-equalityisduetothefactthat
(i, j)∈E xi, j=zi
(
x−i)
+ (k,i)∈E xk,i = (i, j)∈E xi, j= (i, j)∈Ei,act xi, j.Thus,xi solves Pi
(
xP−i,xR,E)
, andsincethis holdsforevery i ∈IP,wededucethatxP∈Eq(xR,E).
Now,letusprovethatEq(xR,E)⊆S(xR,E).LetxP∈Eq(xR,E),and
suppose that xP∈S(xR, E). Since for each i ∈I
P the vector xi is a
feasiblepointofP
(
xP−i,xR,E
)
,theonlywayforxPnottobelongtoS(xR,E)isthatthereexisti
0∈IPsuchthatxi0
Ec i0,act=0.Thus,there is
(
i0,j0)
∈E\
Ei0,act such thatxi0, j0 >0.Let(i0,j1)∈Ei,act (which isnonemptybydefinition)andletusconsiderthevectorxi0 given by xi0,k=
x i0,k ifk∈I\
{
j0,j1}
, 0 ifk=j0, xi0, j1+xi0, j0 ifk=j1. Wehavethatxi0≥ 0(sincexi0≥ 0)andalso
zi
(
x−i0)
+ (k,i0)∈E xk,i0= (i0, j)∈E xi0, j= (i0, j)∈E xi 0, j. Thus, since x−i0 remains the same, xi0 is a feasible point of Pi
(
xP−i0,x
R,E
)
. Furthermore, denoting byCost i0= Costi0
(
xi0,xP−i0,x R,E)
− Cost i0(
xi0,xP−i 0,x R,E)
,wehavethatCosti0= (i0, j)∈E
γ
i0, jxi0, j− (i0, j)∈Eγ
i0, jxi0, j =γi
0, j1−γi
0, j0 xi0, j0 =γ
∗−γ
i0, j0 xi0, j0<0, since,byconstruction,γ
i0, j0>
γ
∗.Thisyieldsthatxi0 doesn’tsolvePi
(
xP−i0,x
R,E
)
,whichisacontradiction.Thus,xP∈S(xR,E),finishingtheproof.
Intuitively, the above theorem says that, given a topology E, each independent agent i ∈ IP will only use the connections of
minimalcost tosendtheexcessofflux,that is,shewilluseonly heractivearcs.Furthermore,eachindependentagentisindifferent tothedistributionoffluxesamongtheactivearcs,soanyfeasible vectorxPsatisfyingtheconstraintx
i
Ec i,act=0foreveryi ∈IPmust
beanequilibrium.Thissimplificationisstronglybasedonthe lin-earity ofthe costs functions withrespect to the agent’s variable
4.2.Mixed-integerformulation
Theorem4.1establishestheremarkablefactthattheMPEC for-mulationoftheauthority’sproblemcanbereformulatedasa “clas-sical” programmingproblem.Butactually,apartofthevariablesof thisprogrammingproblemliesinthesetoftopologiesofthe ex-changenetworkandso,itcanbe consideredasdifficultto imple-mentnumerically.Thisiswhy,in thissection, we willshowhow onecanfinallyworkwithamoreclassicalmixed-integer program-mingproblem.
Let us first introduce the key notion that we will use to ar-riveto thefinal formulation,that is,what we callarc classes:let (i,j)∈Emax.Wedefinethearcclassof(i,j)astheset
C
(
i,j)
:={
(
i,k)
∈Emaxγi,k
=γi, j
}
ifi∈IP{
(
i,k)
∈Emax}
ifi∈IR. (4.4)Wedenote by Ci the familyof all arcclassesexiting fromi,that
is,Ci=
{
C(
i,j)
:(
i,j)
∈Emax}
.Finally,forC∈Ciwedefinetheuti-lizationcostoftheclassby
γ
(
C)
:=γ
i, j,where(i,j)isanyrepresentativeofC.
Observethat,fortwoarcs(i,j),(i,k)∈Emaxsuchthat
γ
i, j=γ
i,k,one hasthat C
(
i,j)
=C(
i,k)
. Thus, a classC∈Ci may havemanyrepresentationsof theform C(i, j). Furthermore, thefamily Ci
in-ducesapartitionofthesetofarcs“exitingfrom” agenti,thatis
• C∈CiC=
{
e∈Emax : e=(
i,j)
forsome j∈I0}
.• ForanytwoclassesC,C∈Ci,eitherC=C orC∩C=∅.
Moreover, it isnot hard toverifythat foreach topology E∈E
andforeachagenti∈IP,thereexistsoneclassC∈Cisuchthat
Ei,act⊆ C, (4.5)
andthisclassmustsatisfythat
γ
(
C)
≤γ
(
C(
i,0))
. (4.6)ThisclassisthengivenbyC=C
(
i,j)
where(i,j)isanyelementofEi,act.WewillcallittheactiveclassofEoftheagenti,andwewill
denoteitbyCi(E).
Without loss of generality, we will assume that every class
C∈Ci satisfies (4.6). If not, any connection in a class violating
(4.6)would neverbeen used,andtherefore,inpractice,they can beerasedfromEmaxwithoutchangingtheproblem.
Now, let D=i∈IPCi, the setof allarc classesof independent
agents.Weintroducethebooleanvariabley=
(
yC)
C∈D∈{
0,1}
|D|inthefollowingway:foreachindependentagenti ∈IP andeacharc
classC∈Ci,weset
yC=
1 ifC istheactiveclassofi,
0 otherwise.
Fromy∈{0,1}|D|,wewillbuildthegraphassociatedtoyas
E
(
y)
=(
{
: yC=1}
)
∪
{
,0)
: i∈IP}
∪{
(
r,j)
∈Emax : r∈IR}
. (4.7)Weconsider thenthefollowing Mixed-Integeroptimization prob-lem: minx∈RN,y∈{0,1}|D| Z
(
x)
s.t.⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
x∈X, zi(
x−i)
+ (k,i)∈Emax xk,i= (i, j)∈Emax xi, j,∀
i∈I, C∈Ci yC=1,∀
i∈IP, (i, j)∈C xi, j≤ B· yC,∀
C∈D, gi(
x−i)
≤ 0,∀
i∈IP, zi(
x−i)
≥ 0,∀
i∈I, Costi(
xi,xP−i,xR,E(
y))
≤α
i· STCi,∀
i∈IP, x≥ 0, (4.8)where Bis a realnumber chosen arbitrarily, butbiggerthan the maximumofthetotalentering fluxover allenterprises. Asimple option to set Bis the value Z(0),which corresponds to the total consumption of the natural resource when there is no exchange network.
Here, the constraint C∈C
iyC=1 says that, for the ith agent,
onlyoneclassisactive.Also,theconstraint
(i, j)∈C
xi, j≤ B· yC,
∀
C∈Densures that, whenever(i, j) doesn’t belong tothe active classof theithagent,thenxi, j=0.
Theorem4.2. Foreveryfeasiblepoint(x,y)of(4.8),thepair(x,E(y))
isafeasiblepointof(4.3).Conversely,foreveryfeasiblepoint(x,E)of
(4.3),thepair(x,yE)isa feasiblepointof(4.8),whereyE ∈{0,1}|D|
isgivenby yE C=
1 ifC=Ci(
E)
forsomei∈IP, 0 otherwise.Finally,onehasthat
1. if(x,E) isanoptimalsolutionof(4.3),then(x,yE)isan optimal solutionof(4.8).
2. if(x,y)isanoptimalsolutionof(4.8),then(x,E(y))isanoptimal solutionof(4.3).
Proof. Let(x, y) be a feasible point of(4.8). Letus fix an agent
i ∈IP andlet Cibe theuniqueclass inCi suchthat yCi=1.Then,
byconstruction,weknowthat
E
(
y)
i,act=Ciand (i, j)∈Emax\Ci xi, j≤ B· C∈Ci\{Ci} yC=0.Wededucethenthat
xi
E(y)c i,act=0.
Sincethisconstraintisvalidforeveryactiveagenti∈IP,andsince E(y) containsall exitingarcs forevery regulated agentr∈IR,we
canrewritethebalanceconstraintinproblem(4.8)as
z
(
x−i)
+ (k,i)∈E(y) xk,i= (i, j)∈E(y) xi, j,∀
i∈I.Wededucethenthat(x,E(y))isafeasiblepointofproblem(4.3). Now,let(x,E)beafeasiblepointofproblem(4.3).Byinclusion
(4.5),foreachindependentagenti∈IP,thereexistsauniqueactive
class Ci(E). LetusdefineyE ∈{0,1}|D|asin thestatementofthe
theorem.
Then,foreveryi∈IP,C∈Ciy
E
C=1.Now,fixaclassC∈D,and
leti∈IPsuchthatC∈Ci.Wehavethat (i, j)∈C xi, j≤
B=B· yE C ifC=Ci(
E)
, 0=B· yE C ifC=Ci(
E)
,where thesecond inequality comes fromthe fact that,whenever
C=Ci(E),thenC⊆ Eic,act andsoxi
C=0.Foranagenti ∈IP,thefactthatEi,act⊆E(yE)leadustothefact
that
Costi
(
xi,x−iP,xR,E(
yE))
=Costi(
xi,xP−i,xR,E)
,andso,theconstraint(3.10)issatisfied.We deducethat (x,yE)is
afeasiblepointof(4.8),sinceallotherconstraintsaredirectly sat-isfiedgiventhat(x,E)isfeasibleforproblem(4.3).
Now,letusassumethat(x,E)isalsooptimalforproblem(4.8). Fromthedevelopmentabove,foreveryotherfeasiblepoint(x,y) of(4.8),weknowthat(x,E(y))isalsoafeasiblepointofproblem
(4.3),andso,Z(x)≤ Z(x).Thus,(x,yE) isoptimalfortheproblem
(4.8).
Let nowassume that (x, y) is an optimal solutionof problem
(4.8)andsuppose,byabsurd,that(x,E(y))isnotoptimalfor prob-lem (4.3). Then, there exists a feasible point (x, E) of problem
(4.3) such that Z(x) < Z(x). But,asproved above,
(
x,yE)
is alsofeasible for problem (4.8), showing that (x, y) is not optimal for
(4.8),whichisacontradiction.Theproofisthencompleted. Thereadercouldobservethat,apriori,themixed-integer prob-lem(4.8)issmallerthanproblem(4.3)insomesense,sinceit ad-mitsonlycertaintopologies(thoseonesoftheformE(y)forsome feasiblepointy∈{0,1}|D|).Howevertheabovetheoremshowsthat
thesetoffluxdistributionsxforwhich(x,E)isanoptimalsolution of(4.3) forat leastonetopology E coincides withtheset offlux distributionsxforwhich(x,y)isanoptimalsolutionof(4.8)forat leastoney.
4.3. Nullclassasexitoption
Physically, we know that the network has always a feasible point, which is the stand-alone configuration, that is, the topol-ogy Est and the fluxesgiven by the individual operationsof the
independentagentsandinactivityoftheregulatedones.However, when we include the individual rationality constraint (3.10), the problemmaybecomeinfeasible.
Infeasibilityofproblem(4.3)meansthattheauthorityisnot ca-pabletofindasolutionthatrespecttheBlind-Inputcontractswith alltheagents.Thus,weneedtoincludethepossibilityofexcluding someagentsfromthenetwork.
Formally, for each independent agent i ∈ IP, we include a
booleanvariableyi,null∈{0,1}suchthat
yi,null=
1 ifibreaks theBlind-Inputcontract,
0 otherwise.
Withthisnewvariable,wemodifyproblem(4.8)asfollows: 1. Foreachagenti∈IP,weput
yi,null+
C∈Ci
yC=1,
meaningthat,eitheronearcclassisactiveortheagentis out-sidethenetwork.
2. Foreachagenti∈IP,weput (i, j)∈C(i,0) xi, j≤ B·
(
yC(i,0)+yi,null)
(i, j)∈Emax, j=0 xi, j≤ B·(
1− yi,null)
Thisistoensurethat,iftheagentbreakstheBlind-Input con-tract,thenshewillusethedischargearc(i,0).
3. Foreachagenti∈IP,weput
(k,i)∈Emax
xk,i≤ B·
(
1− yi,null)
.This constraintestablishes that, iftheagent breaksthe Blind-Inputcontract,thennobodycansendheranyflux.
4. Foreachagenti∈IP,weput
Costi
(
xi,xP−i,xR,E(
y))
≤αi
STCi·(
1− yi,null)
+STCi· yi,null.(4.9)
Here, the individual rationality constraint is active only when
yi,null=0. Otherwise, since the agent isnot connected to the
network,hercostwillcoincidewithSTCi.
We set D=D∪
{
Nulli : i∈IP}
, where Nulli is the null class,associatedtoyi,null,andD0=D
\
{
C(
i,0)
: i∈IP}
.Denoting STCi(
yi,null)
:=αi
STCi·(
1− yi,null)
+STCi· yi,null,thenewoptimizationproblembecomes
minx∈RN,y∈{0,1}|D|Z
(
x)
s.t.⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
x∈X, zi(
x−i)
+ (k,i)∈Emax xk,i= (i, j)∈Emax xi, j,∀
i∈I, yi,null+ C∈Ci yC=1,∀
i∈IP, (i, j)∈C xi, j≤ B· yC,∀
C∈D0, (i, j)∈C(i,0) xi, j≤ B·(
yC(i,0)+yi,null)
,∀
i∈IP, (i, j)∈Emax, j=0 xi, j≤ B·(
1− yi,null)
,∀
i∈IP, (k,i)∈Emax xk,i≤ B·(
1− yi,null)
,∀
i∈IP, gi(
x−i)
≤ 0,∀
i∈IP, zi(
x−i)
≥ 0,∀
i∈I,Costi
(
xi,xP−i,xR,E(
y))
≤ STCi(
yi,null)
,∀
i∈IP,x≥ 0.
(4.10)
Observethat, wheneveryi,null=0, then all constraintsfor the
ith agent are the same that those established in problem (4.8). Also,ifyi,null=1,theonlyfeasiblesolutionforiisthestand-alone
operation.Thus,inthisnewproblem,theauthorityfirstchooseall theagentsthatwillparticipateinthenetwork,representedbythe set
IP =
{
i∈IP : yi,null=0}
,and then it solves problem (4.8) replacing I by I=IP∪IR. Of
course,as itis formulated, the authority takesboth decisions si-multaneously, by solving problem (4.10). It is not hard to verify thatanyoptimalsolutionofproblem(4.10)isanoptimalsolution ofProblem(4.8)forthereducedsetofagentsI.Weleavethis ver-ificationtothereader.
5. Blind-Inputmodelforwaterexchangenetworks
In thissection we come back to our original motivation pre-sentedinSection2,thewaterexchangenetworksinEco-Industrial Parks.Wearenowreadytodescribeindetailthemodel,andhow itfitsintotheBlind-Inputmodeldevelopedsofar.
First, an EIP consists in a set of enterprises P:=
{
P1,...,Pn}
,thatareconnectedinanexchangenetwork.EachenterprisePican beconnectedeithertoother enterprises, ortosome regeneration units,which we denoteR:=
{
R1,...,Rm}
.The regeneration unitsarecontrolledby acentralauthority. Thisauthorityplaystherole ofdesigner (when deciding the connections within the network) and regulator of the park’s operation. Finally, we include a sink node0,thatrepresentsawastes’pittodischargeuseless polluted water.WeidentifyP withtheindexsetIP:=