Automated Deduction of Cross-Organizational Collaborative Business Processes

HAL Id: hal-01744194

https://hal-mines-albi.archives-ouvertes.fr/hal-01744194

Submitted on 6 Nov 2018

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

Automated Deduction of Cross-Organizational

Collaborative Business Processes

Aurelie Montarnal, Wenxin Mu, Frederick Benaben, Jacques Lamothe,

Matthieu Lauras, Nicolas Salatge

To cite this version:

Aurelie Montarnal, Wenxin Mu, Frederick Benaben, Jacques Lamothe, Matthieu Lauras, et al..

Au-tomated Deduction of Cross-Organizational Collaborative Business Processes. Information Sciences,

Elsevier, 2018, 453, p.30-49. �10.1016/j.ins.2018.03.041�. �hal-01744194�

Automate

d

de

duction

of

cross-organizational

collaborative

business

processes

Aurelie

Montarnal

a,∗

,

Wenxin

Mu

b,∗

,

Frederick

Benaben

a

,

Jacques

Lamothe

a

,

Matthieu

Lauras

a

_,

_Nicolas

_Salatge

a

a IMT Mines Albi, University of Toulouse, Albi, France

b Department of Information Management, Jiao Tong University, Beijing, China

Keywords:

Business process management Interoperability

Decision support system Model-driven architecture Ontology

Business services composition

a b s t r a c t

Being able to implement efficient cross-organizational collaborations has become a key factor for enterprises to respond to emerging market opportunities. The business pro-cess management approach is commonly used to design cross-organizational collabora-tions.Thistype ofbusinessprocessaimsat achievingspecificcollaborativeobjectivesby addressingthreemainstepsaccordingtoatop-downapproach: (i)definingthebusiness servicesthathavetobeperformedtoreachtheobjectives,(ii)findingthebestsetof part-nerstoprovidethemand(iii)orderingthebusiness servicesinan optimizedway.While theresultingbusiness processesarea cornerstonetosupporttheinteroperabilityamong thepartnersofacollaboration,theirdesignstepremainsoftenhumanly-conductedand la-borious.Moreover,seekingthe“best” setofpartnersinvolvesnon-additivecriteriasuchas thedeliverytime(i.e.businessservicescanbeperformedinsequenceorinparallelwithin theprocess).Inthis context, thispaper presentsa decisionsupport systembased onan AntColonyOptimizationalgorithmtoexploitcollaborativeknowledgegatheredfrom com-paniesonadedicatedplatform(companies’profilemodelsregisteredtotheplatformand collaborative opportunity models) and deduce quasi-optimal collaborative business pro-cesses.Aprototypethatsupportsthissystemisalsopresented.

1. Introduction

Inthecurrent economic context,surviving asanisolated companyhasbecome unrealistic. Thenecessityto quickly re-spond to the emerging market opportunities leads to two consequences: (i) the companies have to react quickly and (ii) must be ready to establish new collaborations dedicated to the specific opportunities. While the recent advances in in-formation technologies have enabled the creation of useful tools to support the coordination between the partners of a collaboration, thereisstill alackwhen itcomes tohelpcompaniesin theircollaborations(i.e. shareandachieve common goals with other heterogeneous enterprises). In most of the cases designing cross-organizational collaborations relies on empiricalabilities, informallyheld byspecific peoplein thecompany. As a result,thescope of potentialpartners remains narrowandthereisalackofflexibilitywhenmountingon-the-flycollaborationsdedicatedtoemergingopportunities.

∗ Corresponding author.

Fig. 1. Functional big picture of the collaborative process deduction service of OpenPaaS.

Inthissense,thenotionofinteroperabilityoforganizationshasemergedastheabilityofheterogeneoussystemstobuild harmonious and intentional collaborative behaviors without needing deep changes in their structure [34]. Among many methods,Muetal.[27]claimthat settingupanontology-based mediationinformation systemappearsto bethebest way toensureinteroperabilitybecause,on thecontrarytoothermethods,itdoesn’trelyon anystandardization oftenlaborious to integrate and update on a long term. Moreover, interoperability concerns the business level as well as the technical level ofa collaboration, which,in informationsystems, can betransposed asbusinessprocess and workflows respectively. A MediationInformation System (MIS) [46]is a suitable solution to bridge thegap between both businessand technical levels bydealing with processorchestration and technical serviceselection. However, before thecollaborative process can beorchestrated,ithastobedesigned.

This researchis part of a bigger project called MISE (MediationInformation System Engineering) [2] that proposes to automaticallyset upMIS tofitthespecific needs ofinter-organizationalcollaborations.It isbasedon anModel-Driven Ar-chitectureandpresentsaknowledge-basedsystemthatcanbedecomposedintothefollowingsteps:(i)gathercollaborative knowledgeviabusiness-orientedmodels,(ii)elaborateasuitablecollaborativebusinessprocesstoreachthelatter gathered collaborativeobjectives,(iii)turnthecollaborativebusinessprocessintoatechnicalworkflowand(iv)orchestratethe work-flow. The design-time of the system consists in the three first steps, while the run-time concerns the final orchestration of the workflow. In continuation with the works presented by Mu et al. [27] (Section 2.1 highlights thenovelties of our research),thispaperfocusesthenon thetwo firstpartsand aimstoaddress thequestionoftheautomated elaborationof collaborativebusinessprocessesinresponsetospecificcollaborativeobjectives.ItisalsopartoftheOpenPaaS1 projectthat aims to implement a collaborative Platform as aService that offers companies helpfulservices in order to improve their cross-organizationalcollaborations. EnterprisescandescribethemselvesthroughaProfile Modeler(cf.Fig.1). Then,an Op-portunityModeler (cf.Fig.1)allowsenterprisesto proposenewcollaborative opportunities.Limitedtothese two inputs,it isproposedtoinfer asuitablecollaborativeprocessesto setupspecificMIS dedicatedtoaddresseachopportunity.As itis expectedtohavenumerousorganizationsabletoprovidethesamecapabilitiestoreachaspecificcollaborativeopportunity, weaimtofindanoptimalprocessaccordingtonon-functionalcriteria(e.g.cost,deliverytime...).However,becauseofsome criterialiketheentireprocessexecutiontime,thewholesetofpartnersandtheirorderedbusinessservicesmustbeknown beforeassessingacandidatesolution.

Hence,thisresearchcanbedecomposedintothreemainproblemsthathavetobeaddressedsimultaneously: • Whichbusinessservicesshouldbeperformedinordertofulfilltheobjectives?(What?)

• Who couldprovide these businessservicesto achieve thebest results(in termsofcriteria suchastime ofcompletion, orcost)?(Who?)

• When shouldeachorganizationexecuteitsbusinessservices?(When?)

It should to be noted that the Business Process Modeling Notation 2.0 (BPMN 2.0) has been chosen as the expected formatofthefinalprocess,sinceitprovidesorchestrableprocesseswithoutrequiringanymodeltransformation[35].

Thispaperisarticulatedinfourmainparts.Section2proposesastateoftheartfocused onhowsuchsystemscouldbe implementedtosupportcollaborations,aswellasonthedifferentalgorithmstypicallyusedwithintheservicecomposition topic.Then,Section3explainstheknowledgerepresentationadoptedwithinoursystemandfocusesparticularlyonits spe-cificstructure,whichcouldbedepicted asaknowledgegraphwithANDand ORnodes.Based onthat,Section 4isentirely dedicatedto theexploitationofthisknowledge base,with theimplementation ofanoriginal and particularlysuitableAnt ColonyOptimization algorithm,andtheevaluationofitsperformance.Section5proposesause-casetoillustratethe bene-fitsandlimitsoftheimplementedsystem. Finallytheideasoutlinedinthispaperwillbediscussed andsomeperspectives willbehighlightedintheconclusion.

2. Literaturereviewandresearchpositioning

SincethebeginningoftheriseoftheITtechnologiesandwiththeimplementationofweb-basedplatforms,the applica-tionsfor improvinginter-enterprisescollaborations havebecome apriority. First, withthisgrowinginterest,standardslike RosettaNet [18] have been set up to providemore efficient collaborations. Butdue to theirlack of flexibility and because theyonlyfocuson atechnical level,other workshave beenled,ratheroriented towarddesign-timeautomation: in partic-ular,theservicecompositionresearchfieldconcernsalargepartoftheliteraturereview, bothinManagement Scienceand ComputerScience.InthisSection,thefirstpartprovidesanoverviewon thefewalready existingmethodstodeduce busi-ness processes in response to collaborative opportunities. Then, thesecond part focuses on theservice composition topic and bringsthe fact that these kind ofproblems can be considered as NP-hard problems. That is why themetaheuristics domainisstudiedand inparticulartheuseofACOinservicecompositioncontext.

2.1. Existingsolutionsforbusiness processdeduction

In [26], Mu et al. presents a metamodel and the corresponding knowledge exploitation for the deduction of business processes following the Business Process Modeling Notation (BPMN) [30]. [26] is centered on the use of a collaborative ontology via first order logical and inferences rules to obtain a BPMN process from (i)a set of partner wishing to work togetherbut without knowinghowexactlyto do thisand (ii)businessobjectivesof collaboration. Theresearchconducted byMuetal.[26]setsthefoundationsofthispaper.

Totheauthors’knowledge,theresearchofKoetal.[21]facedtheclosestissues.Koetal.[21]havebasedtheirapproach on aHierarchicalTask Network (HTN) ontologycalledthe Business-OWL (BOWL).Thestrength of theirwork relieson the implementationofanontologynotonlyasaknowledgebasebutthatalsoembedstheresultingbusinesstaskdecomposition of the HTN. At the time of the beginning of these works, in 2007, Cloud Computing was just a new research topic and was not yet one of the highest concerns in Computer Science and the idea of social networks were also at a very early stage. Since then,both have revolutionized IT habits, particularlywith the growingneed formore flexible approaches.To theauthors’beliefone ofthemainkey factorofcurrent process deductionsystemfor industrialapplications istheability to offer this flexibility. For instance, a simpler decomposition of the central collaborative ontology would allow users to bringtheir own knowledge,either bycreatingtheir own already structuredontology, or byanalignment of theirexisting ontologies withthecollaborativeontologyofthesystem. Weproposetouseacollaborative ontologythat doesnotinclude theintelligenceofthedeductionbutaflatand‘accessible-to-newcomers’description,byansweringthefollowingquestions ‘Whataretheactivitiesusuallyneeded tofulfillthiscollaborativeobjective?’and ‘Isitpossibletodividethisobjectiveinto lower-levelsub-objectives?’.Moreover,theexplosionofthecloudcomputingtechnologiesnowallowsfortheuserstoaccess toaPlatform asaService (PaaS)on whichthebusinessprocessdeductionservice canbedeployed,and alsotobring their ownSoftwaresasaService(SaaS)and linkthemtotheirbusinessservices:insuchawaythetasksoftheprocessescanbe human(e.g.production tasks),automated(e.g.customizedformscreation)oralsodirectlylinkedtoaSaaS.

2.2. Researchpositioningregarding servicecomposition

Manyresearch works have been led in the field of service composition with the guiding thread of the establishment ofbusiness processesin orderto model newcross-organizational collaborations. These worksusually deal with two main topics:(i)Management Science and most ofall(ii) ComputerScience. Thus, thoseare thetwo investigated topicsfor this section.

2.2.1. Servicecompositioninmanagementscience

InManagement Science, many worksfocus on theestablishment of supply chainnetworks and moregenerally on the creation of Virtual Enterprises (VEs).Sha and Che in [40]propose a Genetic Algorithm to findan optimal set of partners providingtherequiredbusinessservicesin thecontext ofasupplychain.Based ona‘macro-process’ alsocalleda collabo-rativepattern,theydiscovercandidatepartnersandachieveaglobaloptimizationonvariousnon-functionalcriteriaascost, capacity...Thismethodisinterestingbecauseofthemulti-objectivecontextandthenecessityoffindingawholeoptimalset ofpartners,whichisnotthesameasfindingeachsinglebestpartner. Evenifinourcasethisalgorithmcannotbeapplied,

because ofthe hypothesis ofan already known ‘macro-process’,it leads to an interestinguse ofa metaheuristics. In [8], Crispimand Sousapropose awaytoselectpartnersinthecontext ofaVE.Making thehypothesisofastar-like collabora-tion(i.e.acentraldecision makerand anetwork oflinkedpartnersallaround),and project(s)whoseactivities areknown, CrispimandSousauseaTabuSearchalgorithminordertofindthebestsetofpartnerstoachievetheproject(s),according to various global objectives and constraints. One can also note that theplanning discipline hasbeen widelydiscussed in Management Science, as shown in [24]. Tang and Shen [41], for example, describe a HTN approach to provide on-the-fly andspecificallysuitablecrisisresponseprocesses.However,suchplanningtechniques assumeatleastthat thepartnersare alreadyknown.

As a result, the relative works led in Management Science on the service composition and partners discovery topics illustrate(i)themethodsusuallyapplied inthisfield(i.e.metaheuristics)and (ii)thestrongassumptions made(i.e.either themainstepsofthecollaborative processorthepartnersofthecollaboration arealready known).

2.2.2. ServicecompositioninComputerScience

ConcerningComputerScience,theliteratureisprolificonservicecomposition,sinceitisone ofthemainbreakthroughs madepossible particularlywith the emergenceof theSOA (Service Oriented Architecture) paradigm asa wayto conceive services as the fundamental elements of an application [32]. In [13], Dustdar and Schneider have proposed a survey on web services composition solutions, which are not so far from the scope of this paper, since web services composition approachescan generally be adapted and conducted on a business level (and vice versa). They divide the topic into two approaches: (i) the static composition composes services in design-time, and builds on the hypothesis that partners are fixed;(ii)thedynamiconeallowsthediscoveryofservicesduringtherun-time andleadstoevolvingprocesses.Because of obviousreasons linkedto inter-enterprisebusinessagreements, and in linewith theidea ofautomated oneshot bids, the dynamicapproachisclearlydiscarded. According tothem,theservicecomposition iscloselylinkedtobusinessworkflows, sinceeachbusinesstaskprovidesinformationforfindingcorrespondingtechnicalservices:aconditiontofittheoperational needs, thelinkageamong thewebservices withmessage flows, events,itsprovider.Theselast informationshouldalso be deducedinthecontextofbusinessprocessdeduction.However,onecanobservethatthetechnicalandbusinessapproaches areconductedontwocompletelydifferentabstractionlevelsandconsequentlywithdifferenthypothesiswithregardtothe input.For instance, in[1],Benaben etal.proposea semanticreconciliationfrom nbusinessservicesto mweb services, in ordertoobtain aBPELfilefromaBPMNprocess. Thetransformationconcernstheuseofontologiesasknowledgebasesfor thematchingbetweenbusinesstasksinformationandweb servicesannotations:whatisknownaboutabusinesstaskis(i) itsbusinessrole,(ii)itsflows(input,output)andoptionally(iii)itsprovider(i.e.onecanthinkof‘generic’businesspartners tobefound and whoserole isonlymadeclear onatechnical level). Whencompared, thedeductionofabusinessprocess ishoweveronlybased oninformationabout ‘whatthebrokerwantstodo’, whichcould beonly comparedto avery high-level ‘businessrole’ oftheprocess, that shouldbedecomposed. In [17], Garcia-Crespo et al.highlight theimportance and thecontribution of theSemantic Web and the numerous derived knowledge-drivenapproachesdedicated to the business processmanagement,inparticularwebservicescomposition.

Afterworkflow basedcomposition,RaoandSu[36]presentArtificialIntelligence(AI)planningcompositiondefinedbya five-tuple{S,S0,G,A,

!

}, withS theset ofallpossiblestateoftheworld, S0 theinitialstate,Gthegoalstateoftheworld, Athe set ofactions availableand

!

the presetsand effects of eachaction if executed.Besides, Wanget al.[44] dedicates theirsurvey on bio-inspiredalgorithm for web servicescomposition. According tothem, with thecurrent needQuality of Service(QoS)evaluation,theissuesofservicescompositionhaveevolved:andbasedontheexplanationofCanforaetal.[6], ithasbecome NP-hardproblems, and theirsolving algorithmshave evolvedconsequently.Section 2.3is thereforeentirely dedicatedtotheservicecompositionthrough currentpromisingmetaheuristicapproaches.

As aresult, Fig.2illustratesthe usual service compositionapproaches atafunctional level. One cannote that usually, twotypesofknowledgearealwaysknowninitially:(i)theconditionsofthecollaboration(e.g.objectiveofthecollaboration orobjective ofa businesstaskfor web servicecomposition)and (ii) arepositoryofavailable services,which areintended to becomposed. Then, supplementary information arerequired, either on (iii) thepartners whoalready knowthey want to work together but don’t know to formalize the corresponding process or (iii bis) the main sequence of activities not already assignedto any partner. Inthisworld ofresearch,OpenPaaS takesboth (i)and (ii)asinput,however thereareno informationabout the(iii)orthe(iiibis),and thisiswherethecontributionoftheprocess deductionserviceisdeveloped inthispaper.

Anoverviewoftheliteratureonservicecompositiontopicnowshowsthecurrentlimitationsandtheemergenceofnew relevantapproachbasedonmetaheuristics.

2.3. NP-completeproblems,towardsantcolonyalgorithm

As evoked in Wang et al. [44], new methods have emerged in the service composition field in order to optimize the deducedfinalresulteitherbasedonlocaloptimization,whentheperformancesofeachtaskisassessed(e.g.cost optimiza-tion, thecostsofallthe taskscanindeed beadded toobtain theglobalcost), orbased onglobal optimization(e.g. global deliverytime ofthe businessprocess, time assessment shouldtake taskparallelism orsequencing into account, and con-sequentlythe processshould beentirely deduced before assessment).In thecontext of OpenPaaS,thesecond category of optimizationisthemostinteresting,sincethenon-functionalpreferencesofthebrokercanbebasedonveryheterogeneous

Fig. 2. Contributions of the OpenPaaS project in service composition field.

criteria includingtime.Moreover, in [37], Rosenbergatal. claimedthat ametaheuristic is an“iterative generationprocess whichguides asubordinateheuristicbycombiningdifferentconceptsfor exploringandexploitingthesearchspace” .They areparticularlysuitableforlargecombinatorialproblemsand allowto findquasi-optimalsolution.

A brief overview on metaheuristics frequently applied to service composition positions our system and explains the authors’choiceofametaheuristics.

The SimulatedAnnealing (SA) and theTabu Search(TS) are describedas single-state methodsby Luke [22]. The SA is inspiredfromannealinginmetallurgy,whichconsistsincontrollingmaterialsheating andcoolingtoobtainabetterquality byincreasingthe size ofthecrystals (i.e.in termsof thermodynamics, lower theenergy ofthe material toobtain amore stablestate) [38]. According to Dréo etal.[12], theSA is known asa demanding methodin termsof adjustments(e.g. to controlthe cooling),and canbecome very time-consuming,usually leadingto parallel implementation.There isno notion ofmemory,whichmeansthatsolutionscannotbebasedonprevioussimulations.Intheliterature,theSAcanbeassociated withaGeneticAlgorithm toobtain moreefficientalgorithms inservicecompositionlike intheworksofGao etal.[16]or with a TS in the works of Ko et al. [20]. The TS is based on human memory mechanisms. Schematically, it consists in exploring theneighborhoodofaninitial candidatesolution, tofindbetter candidates.All alonganiteration,asolutioncan bechosenevenifitisworsethan theprevious,inordertoavoidlocalminima.Then,Dréoetal.[12]explainthatamemory ofthelastexplored solutionsiskept,sothat themechanismisabletoforbid them: itavoidsstudying analreadyretained solution(andlocalminimabythesametime).

TheGenetic Algorithm(GA) isconsidered aspopulation-based and isa typeofevolutionary algorithm [22].Jaeger and Mühl[19]considertheGA ascharacterized throughaloopwithinfourphasesasappliedinQoS(QualityofService)aware web service composition: thegeneration of abase population composed ofvarious solutions from random combinations, the selection of aneven number of solutions, the crossing of these chosen solutions to obtain new childrensolutions in the populationand finallythe mutationof individuals. Thisalgorithm isapplied both inComputer and Management Sci-ences,and for instance,Shaand Che [40]useittoselect partnersinasupply chainnetwork design work. Thelimitations ofthisalgorithmcomefromthefactthatthecrossingsteponlyworksifallthesolutions(i.e.chromosomes)havethesame number of individuals (i.e. genes). Consequently, the hypothesis that a macro-process exists enables to encode the chro-mosomestructure fromtheprocess pattern. However, inourcase, twobusiness processescould havedifferent number of tasks(dependingon theirgranularitylevels),and still fulfillthesamecollaborative objectives,thus theycannotbecrossed together.

The Particle Swarm Optimization (PSO), firstly implemented for continuous problems, has then been adapted to dis-crete ones. Nayak et al. [28] and Zhao et al. [47] have for example used thiskind of metaheuristics in VE paradigm, for theselection of partners. In bothcases, thePSO isused on fixed graphs,in which the stepsof the processesare already known.Inaddition,whilethePSOseemstorequirestrongadaptationforspecificdiscretegraph,itsparticularadaptationto constrainedAND/ORnodesconstrainedgraphsseemsquitelaborious.

Ant colony algorithms (ACO) are particularly suitable for combinatorial optimization in a graph structure. Dorigo and Birattari [10]explain that,in theory, this type of metaheuristics can be applied to any discrete optimization problem for whichsomesolutionconstructionmechanismcanbeconceived.Inthecaseofthecompositionofbusinessprocessservices composition,thisisideal.

2.4. AntColonyOptimization’s benefitsandneedofadaptation

Becauseoftheabilityoftheantstotraveloveragraphstructure,theACOseemsquitesuitabletoexploitourknowledge baseAND/ORnodestreestructure(cf.Section3).Theagents(i.e.ants)canthus beconstrainedintheir“paths” so thatthey

Fig. _{3. Proposition of a metamodel to structure the knowledge-based system.}

canexplorethesolutionsspaceinaproperway.Thisbringsarealstrength whendeducingcollaborative processes,sinceit canbeadaptedtoexploittheANDandOR nodesoftheimplementedcollaborativeknowledgebase.Italsoallowsobtaining candidatesolutionsbyanycompletion criteria,andnotonlyafixednumberofindividualasintheGAs.

The ACO possibilities have been exploredin various fields,suchas decisiontrees exploitation[31] oralso data-mining advanceswith,forexample,theAnt-Mineralgorithm,whose goalistodiscoverclassificationrulesfromgathereddata[33]. However,thecomplex structure(i.e.atreewithANDand ORnodes)bringsanunusualuseofACO. Wangetal.[43]detail such ACO, adapted to a directed acyclic graph. In that paper, the graph is actually a workflow and the goal of the ACO isto finda quasi-optimalservices compositionfor that workflow (i.e.finding thebest set ofservice candidatesthat fulfill eachoftheabstractservicesoftheworkflow).Forthispurpose,Wangetal.[43]definestheworkflowasanAND/ORgraph according to the AND/OR relationships between services and their successors. Ants begin their travel on the start event of the workflow, and then follow the workflow: when they meet AND nodes, they are replicated and when they meet OR nodes, they chose one of the path. Functionally speaking, such services composition makes the assumption that the workflowisalready known.Hence,theselectionismade“one-by-one” for each service of the workflow, which means that allthenodes(consideringeachnode isaservice)arealwaysstudied byanyants tofinda candidatesolution.However, in ourcontext,candidate processessizecannot beknownin advance:thetreeknowledgegraphdepicted inSection 3shows indeed that acollaborative opportunity can be answered at different levels of granularity (i.e. eitherat a high-levelwith turnkey solutions, or at a very low-level of granularity by for example dealing with each of the suppliers one-by-one), whichleadsto knowledgegraphsto exploit,withvariablecomplexity(i.e.moreorlesssequencesofadjacent ORand AND nodes).

3. Knowledgerepresentationandacquisition

3.1. Structureofthehandledknowledgeviaametamodel

In a first step, in orderto structure theproposed knowledge-based system, ametamodel isproposed, asillustratedin

Fig.3.Thismetamodelisbasicallydividedintothreemainpackages:

• TheSocialKnowledgeconcernstheknowledgeacquiredbythesystem,thankstomodelers.Itiscomposedofthe emerg-ingmarketOpportunitythat userswouldliketoanswerandthedifferentPartnerCapabilitiesthataremadeavailableby thecompaniesusingtheplatform.

• The PersistentKnowledgeisinitiallyinjected intothesystem,and composedofadecompositionofcollaborative Objec-tivesintoSub-Objectivesand,foreachofthem,theset ofCapabilitiesthat shouldperformedtoachieve them.

• The Inferred Knowledge is about the knowledge that has to be ultimately deduced from the two previous kinds of knowledge. It thus focuses on the concepts related to the building of the finalcollaborative process. In this sense, on the one hand, it describes the collaborative network - composedof the partnersof the collaboration - that will fulfill theObjectives,andontheotherhand,how thesepartners’capabilitiesareturnedintoactivities,thattogethershapethe collaborative processthatwillbeorchestratedbytheMediator.

As aconsequence,itcanbenotedthatthethreetypesofknowledgerelyonone anotherfollowingthisrule:theSocial Knowledgeishumanly broughtinsidethesystem,andbuildsonwhatisalreadyknownbythesystemthankstothe Persis-tentKnowledge andtheInferredKnowledgeonlycanbededucedfrom boththeSocialand PersistentKnowledges.For this specificreason,thefollowingpartsarearticulated asfollows: (i)descriptionofthePersistentKnowledge,(ii)descriptionof theacquisition ofSocialKnowledgeand (iii)descriptionofthedeductionoftheInferredKnowledge.

Fig. _{4. MIT Process Handbook schema,} [3].

Fig. 5. OpenPaaSs CO structure.

3.2. PersistentKnowledge

The choice of the structure of the Persistent Knowledge is based the recommendations found in theliterature. In the field of cross-organizational collaborations, top-down and bottom-up approaches are both commonly used: they provide two different ways ofachieving acollaborative workflow. In thefirst casebusiness objectives shouldbedecomposed into sub-objectives and/or business services, which are decomposed into technical services that could then be ordered into a sequence.In[39],Schulz andOrlowskarecommendtousethetwomethodsasfollows:thetop-down approachallowsthe coalitionoforganizationstobuilditscommonworkflowsbydescribingtheirinteractions,whilethebottom-upapproachis rathersuitableforeachorganizationtospecializeitstaskswithitsown privateworkflows.According tothis recommenda-tionandaspreconizedalsobyKoetal.[21]thetop-downapproachseemshighlysuitableinthecaseofbusinessobjectives andcapabilities decompositionasrequiredforthebusinessprocessdeduction,thatiswhyithasbeenchosen.

3.2.1. CollaborativeOntology

TheCOconsiststhus inObjectives ofCollaborationthat aredecomposedinto asetofrequired capabilities orintoaset ofsub-objectives,whichisitselfdecomposedthesameway.NowthatthestructureoftheOpenPaaSCollaborativeOntology (CO)isdefinedithasbeen populatedwithadequateindividuals.TheCO hasbeen entirelypopulatedwiththeMITProcess Handbook[23].InFig.4,itisinterestingtonotethatthehandbookincorporatesGoalsachievedbyProcessesthatrelyon Re-sourcesand alsoadecompositionoftheGoalswiththehas-partreflectiverelationship.Averysimple 1-to-1transformation hasbeenmadefrom theGoalsof theMITProcess HandbooktotheObjectives oftheCO,and from theProcessesto the Ca-pabilities.TheassociationsbetweentheseclasseshavealsobeentransformedasillustratedinFig.5:is-achieved-bybecomes

isFulfilledByandhas-partbecomeshasSubObjective.Thetransformationallowedobtainingalltheindividuals oftheOWL ver-sionoftheMITProcessHandbookintoindividualsthatfittheCOsneeds. Asalast step,thesimpleinverse logicaxioms(1) and (2) werepropagated intotheCO inorderto obtain practicalrelationships forthe furtherknowledgeexploitationstep (cf.Fig.5).

isSubob jecti

v

eO f ₌hasSubOb jecti

v

e− ₍₁₎

contributesTo₌isFulfilledBy− ₍₂₎

3.2.2. BusinessFieldOntology

ABusinessField Ontology (BFO) hasbeen implemented inorderto specify theCOâsknowledge. The researchwas ori-ented towardlarge sources ofinformation about business activities.Several classifications exist in thisarea,including the

Fig. 6. Class diagrams of both the CO and the BFO and their use by the concept PartnerCapability.

NomenclatureofEconomicalActivities[14],theNorthAmericanIndustryClassificationSystems[5]ortheInternational Stan-dardIndustrialClassificationofAllEconomicActivities(ISIC)establishedbytheStatisticsDivisionoftheUnitedNations[42]. TheISIC decomposesbusinessactivities through 21top-levelactivities, onfour hierarchicallevels and providesthelargest scopeofbusinessactivitiessinceitprovidesaworldwidepointofview,and notonly aconstrainedgeographicalvision.

3.3. SocialKnowledge

As anESN, thefirst step forthe userof OpenPaaS,is todescribe himselfvia aProfile Modeler.The profile ofan orga-nization can beconsideredas arepository ofall thecapabilities it wants toshare duringcollaborations supportedby the PaaS.

It consiststhus of a graphical interface called ‘Profile Modeler’ that allows the user to create instances of capabilities and link them to the IT knowledge (i.e. the ontologies). The description of each capabilities is based on the IDEF0 stan-dard[15]withthecentralcapabilityandthedefinitionofitsinputsandoutputs.Concretely, theusercreatessemanticlinks betweeneachcapabilityand thecapabilitiesavailableintheCO,andannotatesitsinputsandoutputsbylinkingthemwith theBFO.

Then,an‘OpportunityModeler’comesasasecondgraphicalinterfacethathelpstheuserscharacterizingtheir collabora-tiveopportunities.Itisbasedonthesamesemanticprinciple:theusercreatesaninstanceofobjectiveofcollaborationand providesa semanticlink betweenit and an objective ofthe CO.Then theuser specifiesthe opportunity byproviding the businessdomainsconcernedbythecollaborationwithsemanticlinkswiththeBFO.

Finally, Fig.6describesboththeClass DiagramsfortheCOand theBFO,and howtheyareused bywhenauserbrings anewPartnerCapability thankstotheProfile Modeler.One cannotethat theBusinessFields includedin theBFOareonly used as properties for theinput and output of thePartnerCapability (and more broadlyas aproperty ofthe Opportunity describedintheOpportunityModeler).

3.4. Logicalstructureoftheknowledge,towardthedeductionofcollaborativeprocess

At thispoint,theusershavedescribedthecollaborative contextbybringing(i)arepositoryofcapabilities theyallwant to bring within collaborations and (ii) collaborative opportunities. This acquired knowledge is directly linked to the one injected initiallyinthe COand theBFO.As aresult,aconstrainedknowledge graphisobtained. Theseconstraintsaredue tothedecompositionoftheObjectives andCapabilitiesandcanbeexpressedbysimple logicrules,asfollows:

• The decomposition fromanObjective toitscomplementary Capabilitiesismodeledwith anAND-node,sinceallthe Ca-pabilities shouldbeperformedtofulfilltheObjectiveofthecollaboration.

• The decomposition from an Objective to itscomplementary SubObjectives is modeled with an AND-node, since all the

Fig. 7. Transformation from CO to a logical graph.

• The decomposition fromaCapability ofCOto itslinkedPartnerCapabilitiesismodeledwith anOR-node,since onlyone organizationshould bechosen toexecute theCapability intheprocess (inFig.7, theCapabilitieshave beendepicted as OR-nodes, thePartnerCapabilitieshavenotbeenrepresentedinthelogical graphtokeepitreadable).

• Each Objective that hasa SubObjective ismodeled asanOR-node, since thechoice shouldbemadebetween the set of

PartnerCapabilitiestodirectlyfulfillit,orthesetofSubObjectivesthat decomposeit.

• Each Objective that has no SubObjective is modeled asan AND-node, since the only children consist in the set of the correspondingCapabilities.

Fig.7illustratesthetransformationfromCOtoalogicalgraph,whichisinfactthegraphusedbytheants oftheACO. Forinstance,theObjectiveAistransformedintoanOR-nodethat leadseithertoafinalAND-nodeforalltheCapabilities, eithertoanAND-nodeoftwoSubObjectives Band C.SinceD andEaretwofinalSubObjectivesoftheedgetheyaremerged togetherintoanAND-nodeofalltheircorrespondingCapabilities.

Intheremainingofthispaper,thetermof‘extendedCO’referstotheCOplusthecapabilitiesoforganizationsthathave beenlinkedtotheCO.Therefore,theextendedCOcontainsbothpermanentknowledge(theCO)andthe‘living’knowledge arisingfromenterprises.

4. ExploitationoftheknowledgegraphwithanACO

Thededuction of aninter-organizational collaborative process remains thecritical part ofthe whole system. The chal-lenge ofthedeductionalgorithmdescribedinthissub-sectionistoprovidean‘optimal’cross-organizational business pro-cess.Hereoptimalisused tomention abusinessprocessthat offers goodcompromise accordingtonon-functionalcriteria as cost, delivery time quality, etc. The selection of all the cheapest complementary capabilities for an objective will def-initely provide the cheapest process. However, it isobviously not the casefor the time criterion for example because of theparallelismorthesequencingpossibilitiesduringtheservice composition(i.e.whenthecapabilities areorderedinto a collaborativeprocess).

Thatiswhytheambitionofthisdeductionalgorithmistoexploittheminimalknowledgeitholdsfromthetwoprevious modelers,inordertosimultaneouslyprovide(i)arelevantset ofpartners,(ii)theirquasi-optimalset ofcapabilitiesand(iii) theappropriatesequenceofcapabilities.

Accordingtothefamousmaxim‘divideandconquer’,thefollowingpartsallowtofirstdescribetheoptimizationproblem and provide amain ACO algorithm, and then detailits threesub-algorithms (i.e.three steps ofthe ACO):the Exploration phase,theConstructionphaseand theEvaluationphase.Finallyanimprovedversion oftheExplorationisproposed.

4.1. Problemdescription

A process can be described as a sequenced set of capabilities of organizations. Thus, thisoptimization has two main goals:(i)find a‘good’set ofcapabilities and(ii)deducethesequencesofcapabilitiestobeabletoassessthecorresponding process.

Inthefollowingsection,thevariablecapaisassociatedwiththegenericcapabilities containedintheCO,whereasoCapa

1: procedure main-ACO(objective)

2: Initialize N the number of cycles, A the number of ants per cycle 3: Initialize pheromone on all the nodes of the extended CO with τ0

4: repeat

5: Initialize S_pk

6: for Ant=1 to A do

7: oCapaSet← Exploration(objective)

8: Update local pheromone on each oCapa in oCapaSet, according to Equation 3

9: process_{← Construction(capaSet, objectiveDomains)}

10: if process is feasible then

11: for all non-functional objectives k do

12: Sprocess ← Evaluation(process,nonFunctionalObjectives

,nonFunctionalObjectivesWeights )

13: end for

14: end if

15: end for

16: Determine the best and second-best solutions

17: Update global pheromone on the corresponding capaSet according to Equation4

18: until non-functional objectives are fulfilled or N is reached 19: end procedure

Fig. 8. OpenPaaS’s ACO main algorithm.

A set of capabilities is modeled as a vector of complementary capabilities of organizations oCapaSet₌

(

oCapa₁ ,. . .,oCapan

)

.Thevariableobjectivealsocorrespondstothecollaborativeobjectiveproposedbyauseronthe

Oppor-tunityModeler.Thevariableprocessisusedtorefertothededuced processcorrespondingtothesequencingofoCapaSet. Initially,thenumberofcyclesNand thenumberofants percycleAaregiven.Eachantgeneratedbeginsthealgorithm withanemptyoCapaSet₌null.All thepaths oftheextendedCO arealsoinitialized with thesameamountofpheromone

τ

0.Oneantgoesthroughthreeparts:

• Exploration:theantexplorestheextendedCO,from thecollaborativeobjective objectiveto thecapabilitiesof organiza-tions, accordingto specificconstraintsduetothestructureoftheextendedCO. AttheendoftheExploration,thelocal pheromone is updated (i.e. decreased for the diversification of further candidate solutions) for each parent branch of eachoCapaofthechosenoCapaSet,accordingtotheEq.(3).

• Constructionandfeasibility:theprevioussetofcapabilitiesoforganizationsisordered.Accordingtotheinputandoutput ofeachcapability (i.e.thebusinessdomains),theyarelinkedto eachotherin ordertoobtain aprocess. Ifthisprocess iseffectivelydeduced,thenitshowsthat thisprocessisfeasible:itisacandidatesolution.

• Evaluation:thispotentialsolutionisevaluatedaccordingtoknon-functionalobjectivesgivenintheOpportunityModeler. For eachpotential process previously deduced,a vector Sk

process isgenerated. Thenthe k evaluation are aggregated and

finallySprocessrepresentstheglobalevaluationforthispotentialsolution. Thepheromoneonthevisitedbranches ofthe

two best solutions of each cycle evolves according to the Eq. (4) ((i.e. increases for the intensification of the visited paths).

Fig.8presentsthecorrespondingmainalgorithm. The pheromoneevolvesaccordingtotwosteps:

• Localpheromone:eachtimeapath ofthegraphisvisited, itspheromonedecreasesaccordingto:

τ

branch=

(

1−

ρ

)

·

τ

branch+

ρ

·

τ

0 (3)

• Globalpheromone:attheendofeachcycle,thetwobestsolutionsareintensified,withanincreasingofthepheromone ofallitsparentbranches accordingto:

τ

branch=

(

1−

ρ

)

·

τ

branch+

ρ

·

$

τ

branch (4)

with

τ

branchtheamountofpheromoneonthebranch,

ρ

thepheromoneevaporationrate(

ρ

∈[0;1])and

$τ

branchasfollows:

$

τ

branch=

!

_{10 if branch}

∈bestsolution

5 if branch_∈secondbestsolution

0 if otherwise

(5)

The threemodules Exploration, Construction and Evaluation are thefundamental cornerstone of thisACO and eachof thefollowingsubsectionsarededicatedtothem.

1: procedure Exploration(objective) 2: Initialize objectiveList with objective 3: Initialize listpCapa null

4: while objectiveList not empy do

5: for all currentObj in objectiveList do

6: Initialize listCapa null

7: children← children nodes of currentObj

8: child← chosen node in children, according to Equation6 9: if Type of child = ‘Capability’ then

10: Add listCapa ← child and its complementary capabilities 11: else if Type of child = ‘Objective’ then

12: Add objectiveList ← child and its complementary objectives

13: end if

14: Remove currentObj from objectiveList

15: for all capa in listCapa do

16: pCapa_{← chosen node in children, according to Equation6}

17: Add pCapa to listpCapa

18: end for

19: end for

20: end while

21: end procedure

Fig. 9. OpenPaaS’s ACO Exploration sub-algorithm.

4.2. Exploration

TheExploration algorithm (cf. Fig.9)implementation is based on the logical graphof theextended CO (cf Fig. 7). ob-jectiveList is initiated with objective, the Objective of the collaboration. All along the algorithm this list evolves with the decompositionintoSubObjectives.ThelistlistpCapa isnullatthebeginning andgrowsevery timeanantchoosesa Partner-Capability. First, theant isdroppedon objective, thenitchooses a childnode inCO, accordingto theEq. (6):itcan bean

ObjectiveoraCapability.

• If it isan Objective, this SubObjective and its complementary Objectives arekept and added to objectiveList, and as the parent ObjectivehasbeenprocesseditisremovedfromobjectiveList.Theantre-dothisloop.

• In the caseof aCapability, the childand its complementary Capabilitiesare keptin listCapa forfurther decomposition, and astheparentObjectivehasbeenprocesseditisremovedfromtheobjectiveList.OncelistCapacontainsCapabilitiesto execute, theantneedsto findpartnersabletoprovide eachCapability.For oneCapability ofCO,manychildren Partner-Capabilities areable toprovide it,but withdifferent non-functionalcriteria. The antchooses oneof them,accordingto theusualEq.(6).

Thechoice ofthechildrennodedependsusuallyon twoparameters

η

edge theattractiveness ofeachedgegoingto each

children,and

τ

edge itsamount ofpheromone. Thevariable edgesrepresents the whole set ofchildrenedges available, and edgeone ofthem.

edge₌

"

argmaxu∈ edges[

(

τ

u

)

α·

(

η

u

)

β] if q≤ q0

J if q > q₀ (6)

qisarandomvariablechosenbetween0and 1,and Jischosenaccordingtothefollowingprobabilitydistribution:

pedge=

(

τ

edge

)

α·

(

η

edge

)

β

#

h∈ edges[

(

τ

h

)

α·

(

η

h

)

β]

(7)

The attractiveness of an edge

η

edge is found by applying the further Evaluation assessment explained in Section 4.4,

appliedtoonlyoneCapability.Giventhenon-functionalObjectivesP₌p1,...,pnexpressedandweightedbyW=w1,...,wn

thebroker,theattractivenessofoneedgeis

η

_edge=#n_i=1

(

wi· pi

)

.

Obviously,thisFormulacanonly workwhentheedgeispositionedbetweenaCapability andaPartnerCapability(case1 inEq.(8)).Forthesereasons,

α

and

β

havebeenfixedasfollows:

edge₌

"

α

₌1and

β

₌2 if case1

α

₌1and

β

₌0 if otherwise (8)

4.3. Construction

The process to bededuced here must respect the BPMN 2.0 specification [30]. As stated in the literature review, the benefitofthesecondversionoftheBPMNStandardreliesonitsabilitytobeorchestrated.Insuchaperspective,thededuced processmustcontainapoolforeachpartnerofthecollaborationbutalsoaMIS(MediationInformationSystem)poolwhose goalistointeractviamessageflowswith allthepartnersduringtheorchestration. Thecreationofthepoolisindeedvery easyand thisis thevery first step ofthe construction.Furthermore, theassumption is madethat the parallel gateway is theunique gatewaydeduced,since thededuced process is‘the potentialsolution’, meaningthat the inclusivegateways is understoodhereasanotherprocessi.e.anotherpotentialsolution.

Thus, the main idea here isto deduce thesequence of activities in theMIS pool, since thetasks in partner thepools only appear as atomic activities invoked all along theorchestration. In theOpportunity Modeler,when a user provides a newobjective ofcollaboration,he/shealways linksittoone ormorebusinessdomains(i.e. individualsoftheBFO).As the first step, the Exploration ensures that the set of organizations capabilities can fulfill the objective, the remaining doubt relieson thefactthat thesecapabilities caneffectivelyworktogethertofitthedomain(s)ofthecollaboration.

Theauthorspartlyaddressedtheconstructionstepin[25]byanintuitive‘right-to-left’and‘left-to-right’processing.Itis easilyexplainedbytheinputsofthealgorithmtheobjectiveDomainsandtheset oforganizationscapabilitiescapaSet,which have to beordered into the process: only the result of the process, which matches the objective of the collaboration, is knownhere.Firstofall,anEndEventiscreatedandaddedtoprocess,thefinalprocess.ForeachdomaininobjectiveDomains, allthecapa ofcapaSet areexamined. As soon asanoutputofcapa hasthesame domainasdomain,capa is(i)added asa MIStaskin theMISpoolofprocess, (ii)addedasapartnerTask inthecorresponding partnerpool,(iii) messageflows are createdbetweentheMIStaskandthepartnertaskand(iv)theMIStaskislinkedwiththeEndEvent.capaisremovedfrom

capaSet,sinceithasnow beenused intheprocess.Then,thismatchingismadeagaininaloop,betweentheinputofeach newtasksinprocessand theoutputofremainingcapaincapaSet.

Everytimeataskislinkedwithseveraltasks,agatewayiscreatedandaddedtoprocess,andthecorrespondingtasksare consequentlylinkedtoit,insteadofbeenlinkedtooneanother:

• Either ataskcan belinkedwithseveral previoustasks,thena ‘closing’parallel gatewayiscreatedand linkedwith the subsequenttaskandalltheantecedenttasks.

• Oraseveraltaskscanbelinkedwithseveralsubsequenttaskwhichwouldprovideoutput(s)foreachofthem.Thenan ‘opening’parallelgatewayiscreatedand linkedwithallthesubsequenttasksandtheantecedenttask.

The process finishes when capaSet is empty: all the organizations capabilities have been used and ordered. However if capaSet remains non-empty, this means that the process is not feasible, and according to the main ACO algorithm (cf

Section4.1)itwon’tbepartofthefurtherEvaluation.

4.4. Evaluation

Nowthat theprocesshasbeenbuilt,andconsequentlyfitsthefunctionalneedsofthebroker,itmustnow beevaluated according to the non-functional objectives P₌p1,...,pn expressed and weighted by W=w1,...,wn in the Opportunity

Modeler,andassociatedwithobjectivetheobjectiveofthecollaboration.

The method applied here is synthesized by Berrah et al. [4]. The expression of the performance of a system can be expressed in two basic ways, directly when the performance can be defined through a single expression and indirectly whentheperformanceisthecombinationofseveral elementaryexpressionsandespecially inthecaseifthese expressions do nothave thesame dimensions.The indirectexpressionof theglobalperformance of asystem isconcretely applied by Clivilleetal.[7],forinstance.

Herethevariableprocessrepresentsthededuced processintheConstructionstep.Theperformancesofprocessare eval-uatedaccordingtothennon-functionaldimensionsand arestoredinthevector

v

₌

v

1,...,

v

n. Then,these scoresare

nor-malized accordingto their correspondingnon-functional objective and are calculated as follows:s₌ v 1

p1,...,

v n

pn. Finally all

theseunitaryelementaryperformancesareaggregatedaccordingtothecorrespondingweightsand thefinalglobalscoreis

Sprocess=#ni=1

(

si· pi

)

.

ThisapproachcorrespondstoAlgorithminFig.10.

4.5. ImprovedExploration

It is to be noticed that asproposed inits first shape, the Exploration algorithm tends to make theants exploring the ‘high-level’ paths: theprobability for an antis higher to choosea Capability ofa high-level Objective than of a low-level

SubObjective. For instance, on Fig. 11 the probability for an ant to choose a Capability of A is p, and the probability to chooseaCapability of(H+ I)isw· t· q.Inthesimplealgorithm ofFig.9,theweights p,...,wareallequalto0.5.Thusthe comparativechancetochooseaCapability ofAis0.5against0.125for(H+I).

We propose to balance the weights of all the edges established from OR-nodes (i.e. between Objectives and their Ca-pabilities or their SubObjectives), except if the Objective is at the lowest-level decomposition in order to ensure a better

1: procedure Evaluation(process, P, W ) 2: Initialize Sprocess= 0

3: n the size of P

4: Initialize v and s

5: for i = 1 to n do

6: vi← Performance of process according to Pi

7: si← v_pi_i

8: Sprocess← Sprocess+ (si· pi)

9: end for

return Sprocess

10: end procedure

Fig. 10. OpenPaaS’s ACO Evaluation algorithm.

Fig. _{11. Weighted CO’s logical graph.}

equi-probability.AsforeachOR-node,therearealwaystwoedgespossible:(i)theAND-nodeofadecomposition into Capa-bilitiesor(ii)theAND-nodeofadecomposition intoSubObjectives, then,forany OR-node,exactlytwochildrenedges exist. The probability to takeone or theother is 1, which meansthat the two childrenedges arecomplementary (for instance,

p₊q₌1).

Asit comesto therestoftheequationsto state,theapproachbegins withthelowest-level ofObjectivei.e. (H+I). The closestparent OR-nodeis G,fromwhich thetwochildrencan besimilarlychosen:

v

+w=1.Then, inordertoensurethe equi-probability, thetwo longestpaths starting fromeach ofthe childrenOR-node edges should bealso similarlychosen. Here, the second closest parent OR-node isC. Each of thechildren longest edges of C should besimilarly chosen, which meansthat u₌t_·

v

.In thesameway,for theAOR-node, itcan bestatedthat p₌q_{· t·}

v

. Finally, forthe BOR-node, only twosingle-leveledges exist,thereforer₌s

Thisleads tothefollowingequationsystem:

⎧

⎪

⎪

⎪

⎪

⎪

⎪

⎪

⎨

⎪

⎪

⎪

⎪

⎪

⎪

⎪

⎩

p₊q₌1 r₊s₌1 t+u=1

v

+w₌1 r₌s

v

=w u₌t_·

v

p₌q_{· t·}

v

(9)

Theresult,forthisexampleis: p₌1/4,q₌3/4,r₌1/2,s₌1/2,t₌2/3,u₌1/3,

v

=1/2,w₌1/2.

Thisapproachhasbeen generally appliedto thewhole ontology. Assuming thatthe rankof anObjectivedetermines its maximumnumber ofdecomposition (forinstance, rank_G=1sinceitcan onlybedecomposed onetime intoSubObjectives,

rankC=2sinceitcanbeatmost decomposedtwotimesintosubObjectives andrankB=1),thenitiseasilyverifiable that:

• The probability coefficientofan edgetobe chosen,between anObjective OR-node and itsSubObjectives decomposition iscoe fficient_edge= rankOb jective

1+ rankOb jective.

• The probability of an edge to be chosen, between an Objective OR-node and its Capabilities decomposition is

coe fficient_{edge= 1+ rank}1

Finally, these equi-probabilitycoefficients areappliedtotheamountofpheromoneofthecorrespondingedge.Thus the newpheromoneamountofanedgenew

τ

edgeusedinEq.(6)becomesnew

τ

edge=

τ

edge· coe fficientedge,with

τ

edgethe“basic”

amountofpheromoneandcoefficient_edgetheequi-probabilitycoefficientoftheedge,inEq.(6).

4.6. Performancesanalysis

Inordertoevaluate theperformancesofthisACO,ithasbeendecidedto assesstheglobalscoreallalongtheiterations (i.e.antscycles).Forthispurpose,differentscenariosweresetupbyvarying someoftheparameters.First, thefourchosen scenariosaredescribed,and thentheperformanceoftheACOon eachofthemaredetailed.

4.6.1. Testscenarios

Thescenarioshave beenchosensothatthebehavioroftheants couldbeassessed,dependingontwomainparameters: thenumberofdecompositionofObjectivesintoSubObjectives,whichwillbecallednbObjLevelsandthenumberof PartnerCa-pabilitiesavailablefor eachCapability oftheextended CO(ECO),whichcorrespondingvariableisnbPartnerCap.Thenumber ofCapabilitiesperObjectivehasbeenset to4,and thenumberofSubObjectivesperObjectiveisalso3.

Hence,fourECOshavebeen generatedalongtheseparameters: • ECO#1:nbOb jLe

v

els₌3and nbPartnerCap₌10.

• ECO#2:nbOb jLe

v

els₌3and nbPartnerCap₌20. • ECO#3:nbOb jLe

v

els₌5and nbPartnerCap₌10. • ECO#4:nbOb jLe

v

els₌5and nbPartnerCap₌20.

As a non-functional Objective, the top-level Objective of each ECO was always chosen - which is consistent with the choice ofnbObjLevels.As it comes to thenon-functionalobjectives, theydepend on two non-functionalcriteria, which are theglobal costand theglobal timeof deliveryof thesolution. Both objective values have been selected so that theycould notbereached,inordertoevaluate theconvergence oftheACO,withaweightof0.8forthecostand 0.2forthetime.

All thePartnerCapabilitieshavebeen generatedwiththesameinputand outputflows,assuch,even ifthecalculationof theglobal timeofdelivery ofaprocess corresponds tothe maximumtime ofthe PartnerCapabilities, thetime assessment algorithm is still the same: it has no impact on the evaluation of each candidate solution, but it is quite convenient to generateviable(andconsequentlyassessable)processes.

Based on thesimulations on different cases and also on the observations from ACO precursors Dorigo et al. [11], the relatedworksofDréoetal.in[12]and theresults ofDoerner etal.[9],thevariousparameters havebeenfixed:

⎧

⎪

⎪

⎨

⎪

⎪

⎩

numbero f cyclesN₌100 numbero f antspercycleA₌10

ρ

=0.1

τ

0=0.1 q0=0.7 (10) 4.6.2. Results

As a result,foreach ofthefour previouslystatedscenarios, 10 simulations havebeen done onthe samenative collab-orativeontology, and the average fitness values of thebest process deduced allalong the cycles have been calculated. In parallel,ithasbeencompared tothebestknownsolution, foundalongthecorrespondingmathematicalmodel, inorderto assessthequalityoftheprocessfoundbytheACO.

It should benoted that if all possibilities had been evaluated, the evaluation and construction algorithms of the ACO shouldhavebeeninvokedmorethan 1036_,2036 _,1060_,2060 _{timesrespectivelyforECO#1,ECO#2,ECO#3andECO#4.}

The whole system has been implemented using Java, whilst the ontologies have been developed as graph databases, usingNeo4j [29]and exploited with itscorresponding Java APIs. The evaluation tests have been executed on a computer embeddingai5processor,8Gb ofmemory.

Fig. 12 depicts the average evolution of thebest candidate solutions found for eachcycle, on 100 cycles along the 10 simulationsthat havebeenledonthesameECO.As aresult,adecentconvergenceisobservedfromthe80thcycleineach oftheevaluatedcases.Nevertheless,duetotheexplodingcombinatoryofECO#3andECO#4,itcanbenotedthatlogically theconvergence occursmoreslowlythanforECO#1,ECO#2,alongthe100cycles.

Table1summarizes theresults observedallalongthesimulation.Inaverage,thetimetoperform100 cyclestoexplore theECOsarethefollowing:17minforECO#1,18minforECO#2,23minforECO#3and26minforECO#4withrespectively averagereachedaccuraciesof107%,103%,114%and109%withregardtothebestrespectivesolutions.

4.6.3. Complexity analysis

More interesting,becauseindependentof thehardwareand software settings,a shortanalysis ofthecomplexityof the whole algorithmcan beperformed. Overall,theEq. (11) correspondstothemainalgorithm complexity accordingtoN the numberofcycles,Athenumberofants percycleandthecomplexityofthesub-algorithms.

Fig. 12. Average evolution of the best candidate solutions found at each cycles, on 100 cycles, 10 simulations.

Table ₁

Summary _{of the results observed for the three use cases.}

Accuracy (%) Time at 100 cycles (min) 25 _cycles 50 _cycles 75 _cycles 100 _cycles

ECO#1 193 133 119 107 17

ECO#2 135 114 110 103 18

ECO#3 158 129 118 114 23

ECO#4 150 127 121 109 26

The complexity ofthe Exploration sub-algorithmis thehighest among thethree sub-algorithm, and thus isthe single oneanalyzedfurther.Intheworstcasepossible(i.e.theExplorationreachesthe“bottom” ofthe ECO),Eq. (12)corresponds to the estimation of the complexity for this sub-algorithm. The following variables are used: nbObjLevels the number of levelsof decompositionof objectives,nbObjPerLevel thenumberofobjectives perlevel ofdecomposition and nbCapaPerObj

thenumberofCapabilityperObjective.

CExploration=nbOb jPerLe

v

elnbOb jLevels∗ nbCapaPerOb j (12)

Finallyagoodestimationofthecomplexity(intheworstcase)is:

Complexity=N∗ A∗ nbOb jPerLe

v

elnbOb jLevels∗ nbCapaPerOb j (13) Asanorderofcomparison,thealgorithmtofindtheexactbestsolutionshasamuchhighercomplexity,whichestimation is:

Complexity=nbPartnerCapnbOb jPerLevelnbOb jLevels∗nbCapaPerOb j (14) As explainedinSection 2,very few works haveaddressed thequestionof simultaneouspartners selectionand process composition.TheproblembeingbuiltonanunusualstructureofECO,noothermetaheuristicseemstobeapplicabledirectly, withoutthinkingaboutsuchadaptation(justasproposedfortheACO).Itwillthenbeinterestinginfurtherworkstobuild other suitablealgorithms and compare their complexity and ultimately findthe best alternatives depending on theinitial complexityoftheECO tobeexploited.

5. Illustrativecase

TheOpenPaaS’ ACO hasbeen implemented asdescribedin thesolution proposal ofthispaper. This Sectionintroduces the humanexperience on OpenPaaS’ modelers and shows a concrete resultof the businessservice deduction, all along a

Fig. _{13. Profile Modeler, modeling of a capability.}

Fig. 14. Profile Modeler, details of the capability description.

short use case.Thus as afirst part, thecollaboration isdefined viathe useof thepreviously evoked Profile Modeler and Opportunity Modeler.Theuser-transparentACO algorithmthendeduces theprocess, whichis writteninafile thatcan be openedinaCollaborativeProcess.

Brake France is a French food wholesaler company for enterprises and this use case focuses on a particular bid for chocolateproducts.

5.1. Profilesoforganizationsmodeling

As a first step, eachof the organizationsof theplatform uses the Profile Modeler in orderto create their own profile bydetailing the capabilities theyare able to provide and want to sharein collaborative contexts. Fig.13 showsthe main interface oftheProfile Modeler. Sinceit isgraphical, theProfile Modeler isvery intuitive for theuser. Here,Brake France enterprisecreates a new capability that is ‘Place Order’, since itis one of its maincapabilities as awholesaler. Then the inputand outputaredetailed.Theinputconcerns the‘8211 - Combinedoffice administrativeserviceactivities’ classofthe ISICclassification,whereastwodomainshavebeenchosenfortheoutput:‘1073-Manufactureofcocoa,chocolateandsugar confectionery’and‘4630-Wholesale offood,beveragesandtobacco’.Thusthefirstoutputspecifiesthefactthatthe‘Place Order’ capability concerns the ability of buying from factories of chocolate, and the second one explains that the orders actuallycorrespondtowholesale.

TheselinksbetweenthemodelandtheCOandBFOaredonebydouble-clickingoneachinstance.Thenboxopenswith allthepossibledetails,asinFig.14.Herethe‘PlaceOrder’capabilityislinkedwiththecorrespondingindividualintheCO, thenthepriceanddeliverytimearedefined(severalothernon-functionalcriteriaarealsoavailablebutcannotappearhere becauseofalackofspace).

Fig. _{15. Opportunity Modeler, modeling of a collaborative objective.}

Fig. 16. Opportunity Modeler, details of the objective description.

5.2. Collaborativeobjectivedescription

Similarly,theOpportunity Modeler providesa graphicalinterfacefor theusersto express theiropportunities of collab-oration.Here, inFig. 15, BrakeFrance wouldlike to obtain acollaborative processthat fulfillsa ‘Buy’ objective.The users canlink asmanydomainsasneeded to specifythebusinessfields ofthecollaboration. Forinstance, the‘Buy’ objective is linkedwiththreedomains: ‘4630 -Wholesale offood,beveragesand tobacco’,’1073 -Manufacture ofcocoa,chocolate and sugarconfectionery’and ‘4923-Freight transportbyroad’. Thefirst twodomainsspecify thecontextofthebuying objec-tive, whereas thethird domain concerning transport expresses another domain that Brake France would like to takeinto account.Thismeansthattheyalsowantadelivery‘flow’intheprocess.

AsintheProfileModeler,whendouble-clicking ontheinstancestheycan bedetailed.In Fig.16,BrakeFrance linksthe ‘Buy’objectivewiththecorrespondingindividualoftheCO,and expressitsnon-functionalobjectives.

5.3. Deductionofthecollaborativeprocess

AsitcanbeseeninFig.15,amainmenuatthetopoftheModelerprovidesa‘Action’list.Inthislist,thereisa‘Deduce process’button.Whenclicked,thislaunchestheACO:antsaredroppedonthecollaborativeobjectiveoftheobjectivemodel justdesigned.Withinless thantenminutes,theyareabletoprovideacorrespondingcross-organizational businessprocess. AthirdCollaborativeProcessModelerhasbeenimplementedtoeasilyobservethisprocess,asillustratedinFig.17.

F ig. 17. Collabor ati v e Pr ocess Modeler , re su lt of the pr ocess de duction service.

As explained inSection 4.3, a centralpool iscreated: the MIS Pool.It corresponds to thepoolof theMediation Infor-mation System. This poolholds the sequenced part ofthe process, corresponding to theresult of the orderingof all the organizationscapabilities that have to beexecuted. Theyare linkedwith sequenceflow. Thenmessage flows between the MISpoolandthepartners’pools illustratetheinvocationofeachpartnertask.

Thisprocessshowsthat thefinal‘best’ processfoundbytheACO concernsacollaboration ofBrakeFrance withthetwo companies‘Chocoblast’, which isinvolved inall thetasksconcerning chocolate manufacturingand ‘Spacetruck’ that offers transportservices.

6. Discussionsandconclusion

Thispaperpresents adecisionsupportsystem basedonanACOalgorithm toautomaticallyelaboratecollaborative busi-ness processesbased ononly two inputs: (i)arepositoryofenterprises’ profiles and theirbusinessservicesand (ii)a col-laborative opportunity. This system is based on two parts: (i) ‘flat’ knowledge bases that gather the information about collaborations and (ii)anACO algorithm toexploit thisknowledge.The knowledgebases haveactually been implemented asontologies:aCollaborativeOntologybringsthegenericknowledgeaboutcollaborationsandaBusinessFieldOntology al-lowstospecifythem.Theseontologiesarefirstusedfordescribingthecollaborativecontextsthroughtwomodelers:(i)the ProfileModelerhelps thecompaniestodefinetheircapabilitiesand (ii)theOpportunityModeler isaninterfacetopropose newopportunitiesofcollaboration.

Fromprofiles and objectivesof collaboration, aninnovativeACO canexploit theknowledgecontainedin theontologies and deduceanon-functionally(e.g.accordingto cost,deliverytime...)optimizedcollaborative businessprocess. Threegaps arecrossedinonetime:(i)findingtherequiredcapabilitiestofulfilltheobjectiveofcollaboration,(ii)findingthe‘best’set ofpartnersableto providethemand(iii) orderthese capabilitiesintoacollaborativebusinessprocess.

This process deduction service addresses the current lack of flexibility of such systems through the establishment of differentinnovativepartsofthesolutionsproposedinthispaper:

• Knowledgebaseshavebeen structuredandimplementedasdatagraphusingNeo4J technologies[29]inthe system.The Business Field Ontology presents a quite large knowledgebase that is intendedto fit any business activities. The Col-laborativeOntologycanbenoticedforitstop-down structurethat allowsaveryintuitivedecompositionofcollaborative objectives into sub-objectives and capabilities. These ontologies are used as knowledge bases and provide individuals linked togetheraccordingto theirmeta-models. The knowledgebasesdo not embed any ‘intelligence’, thismeansthat theexploitationalgorithmisonlybasedon thestructureoftheontologybuthasbeenimplementedindependently.As a consequence,anyontologiesrespectingtherightmeta-modelcanbeusedinthecollaborativeprocessdeductionsystem. • An ACO has been applied to exploit these ontologies and deduce collaborative process. The strength of thisalgorithm is to beable to providein the sametime (i)the required capabilities to answer thecollaborative objectives, (ii)a set ofcorrespondingpartnersableto providethesecapabilities accordingto expectednon-functionalcriteriaand (iii) order these capabilities into a collaborative process. This ability is what distinguishes this ACO from most of the deduction of collaborations in the literature. Moreover the proposed ACO proposes the originality to fit ‘hierarchical tree graph’ structureswithANDand ORnodes.

Ontheonehand,someimprovementscouldbebroughttotheACOitself.Infurtherresearch,ideastoimprovethe qual-ity ofthe outputsolutions could bebased on workssuch asthose of Zhongetal. [48] whopropose anhybrid algorithm integratingGA andACO inpartnerselectionproblemsinVEscontextsand claimtoobtainbetter solutionsthanaGA. Nev-ertheless,theresultsbroughtinourapproachseempromisingbecauseoftheoriginal adaptationoftheACOtoaparticular problem that hasnot been addressed a lot inthe literature (i.e. businessprocess deduction without knowingneither the partnersnorthestepstobeexecuted).Suchhybridizationshouldbecarriedoutwithcarebecauseoftheunusualstructure ofthisproblem.

On theother hand, a work could bedone on theuser experiencewith such platform.In thisperspective, two aspects could offer abetter experienceand fita largerscope ofcollaborations:(i)the linkagebetweentheuser’s informationand (ii)thetransformationofany structuredcollaborativeontologies totheCO’smeta-model.TheresearchworksofWangetal. in[45]focusonthecreationofasyntactico-semantic reconciliationthat allowsfirsttodeducecommonmeaningsbetween severalwords,andbasedonthattoalignontologies.Atahigherlevel,withthisontology-alignmentpromise,anyontologies couldbeuploaded andused forthedeductionofcollaborative processes.Evenmoreinteresting, thenew addedontologies could beeasily linked and merged to thealready existing so that thesystem would finally holdsa growingCollaborative Ontology.Withsuchanagileandincreasingknowledgebase,theplatformcouldsupportanalwayswiderscope oftypesof collaborations,whichisbasicallyimpossibleduetothecomplexityofgatheringallthisknowledgefromscratch.

Acknowledgments

Thiswork has been funded bythe OpenPaaS project under aDGCIS/Investissement d’Avenir grant.The authors would liketothanktheprojectpartnersfortheiradviceandcommentsregardingthiswork.