HAL Id: hal-01900381
https://hal.archives-ouvertes.fr/hal-01900381
Submitted on 22 Oct 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.
Flexible knowledge representation and new similarity
measure: Application on case based reasoning for waste
treatment
Philippe Chazara, Stéphane Négny, Ludovic Montastruc
To cite this version:
Philippe Chazara, Stéphane Négny, Ludovic Montastruc. Flexible knowledge representation and new
similarity measure: Application on case based reasoning for waste treatment. Expert Systems with
Applications, Elsevier, 2016, 58, pp.143-154. �10.1016/j.eswa.2016.03.014�. �hal-01900381�
O
pen
A
rchive
T
oulouse
A
rchive
O
uverte
(OATAO)
OATAO is an open access repository that collects the work of some Toulouse
researchers and makes it freely available over the web where possible.
This is an author’s version published in:
http://oatao.univ-toulouse.fr/
20505
Official URL:
https://doi.org/10.1016/j.eswa.2016.03.014
To cite this version:
Chazara, Philippe and Negny, Stéphane and Montastruc, Ludovic Flexible knowledge
representation and new similarity measure: Application on case based reasoning for
waste treatment. (2016) Expert Systems with Applications, 58. 143-154. ISSN
0957-4174
Any correspondance concerning this service should be sent to the repository administrator:
Flexible
knowledge
representation
and
new
similarity
measure:
Application
on
case
based
reasoning
for
waste
treatment
Philippe
Chazara
∗,
Stéphane
Negny
,
Ludovic
Montastruc
Universite de Toulouse, Laboratoire de Genie Chimique UMR CNRS/INPT/UPS 5503, BP 34038, 4 allee Emile Monso, Toulouse 31030 Cedex 4, France
Keywords:
Case-based reasoning Similarity measures Ontology Dynamic cases
Flexible case representation Waste management
a
b
s
t
r
a
c
t
InCaseBasedReasoningtherepresentationofacaseandthesimilaritymeasuresaretwodifficultsteps intheconceptionofasystem.Often,thesestepsaredevelopedtoresolveonekindofproblem.However, insomeofthemsuchasrecoverytreatmentprocessesgeneration,itisnecessaryforthesystemtobe abletomodifyandadapttherepresentationofacaseandthesimilaritymeasureswithrespectofthe contextandalsothekindofsolutionsproposed.Inthispaper,authorsintroduceanewmethodto repre-sentcaseswithaflexibilitybasedonastructureinaconnectionistmodel.Thisflexibilityisneededdue tothecomplexityofcases,thenumberofpossibleoptionsandtoensurethedurabilityofthesystem.In asecondmaincontribution,authorsintroduceamethodfortheselectionofsourcecasesusing abstrac-tion,conceptualisationandinferencemechanisms.Finally,authorstesttheirsysteminaCBRdeveloped onSWI-Prologwithdifferentproblems. TheCBRisappliedtofindnew recoveryprocessesand try to estimatethenewupgradedproductgenerated.
1. Introduction
The problemof wasteand inparticular theproblem ofwaste management hasincreased sharply during the last decades, pro-ducing three kinds of effects. First, the problem of waste treat-ment isbecoming moreandmoreimportant duetothe quantity produced with the increase of human population size and con-sumption. Second, the pricesof some raw materials are growing sharplyduetothephenomenonofdepletion.Itbecomesmoreand moredifficulttofindnewsources andtheirexploitationcosts en-hance. Third, thetreatment ofwaste canhave astrategic dimen-sion.Actually,itcanreducetherawmaterialdependencyforsome countries, itcandevelop newindustriesandcreatenewjobs.But currently, waste is considered as a pollution source for environ-ment andas a costly burden for companies because of the loss of material and the waste treatment. Consequently, it is neces-sarytoproposenewrecovery processesandnewwaystomanage waste.However, some elements inducelimitations.First,contrary to a newproduct, a waste has not essence by definition. There-fore,thefirstquestionisto findoneormoreessencesforit.The
∗ Corresponding author. Tel.: +330534323663.
second question is how to transform a waste into new valuable products.Tosolvethesequestions,authorsproposetousean arti-ficialintelligencesystem,andmoreparticularlycasebased reason-ing (CBR).CBRisrelevantforthiskindofproblemsbecauseit al-lowssolvingproblemswithoutaclearlydefinedknowledgeofthe processneededfortheresolution.Thereasoningcanrelyonavast numberofcases,withtheirprecisedescriptionofprevioussolved problems and their associated solutions (Cordier, Mascret, Mille, 2009). Secondly, in the domain of waste treatment, cases may contain differentinformation:valorisationprocessesandessences for thenew createdobjects. Inthe literature, casebased reason-ing systems are used in different waste treatment problems and inprocesses research.Forexample,López-Arévalo,Bañares Alcán-tara,Aldea,Rodríguez-Martínez,andJiménez(2007)describeatool basedonCBRforthegenerationofprocessalternatives.Yang and Chen(2011)proposeaclassicalCBRretrievemethodusedfor Eco-innovationKuo(2010)givesanexampleofCBRusedtodetermine a recyclable index of some components. Liu and Yu (2009) use CBR for problems linked to environmental topic. Zeid,M. Gupta, and Bardasz (1997) propose a model dedicated to disassembling problems.
AsdetailedinSection2,CBRmethodisdecomposedindifferent steps: Retrieval,Adaptation,Memorisation or Learningasexplained byAamodtandPlaza(1994)andNapoli,Lieber,andCurien(1996), similaritymeasureisonekeycornerstoneofaCBRsystemandof
E-mail addresses: philippe.chazara@ensiacet.fr (P. Chazara), stephane.negny @ensiacet.fr (S. Negny), ludovic.montastruc@ensiacet.fr (L. Montastruc).
the Retrieval partin particular. Thismeasure allows finding close andrelevantcasesto solvethenewproblem. Therefore,withour goaltoreusetheknowledge relatedtorecovery methodsfornew waste valorisation,itisimportanttopropose newapproachesfor this step respecting the constraints imposed by this category of problems.
On this topic, authors tackle several problems related to the similarity question. The first one is how to represent a case and moreparticularlyforthedomainofapplication,how torepresent a waste.There are many kinds ofwaste andthey need different representations.Moreover, thedomainsofwasteandwaste treat-ment have an importantdynamic. Indeed, thesedomains change quicklyi.e.thecomposition ofwaste,orthewastetreatment pro-cesses evolve over time. To take into account these points, it is necessary to develop a flexible case representation to ensure a precise descriptionofproblem, knowledgereuse,CBRsystem effi-ciencyanddurability.Anotherconsequenceofthesespointsisthat the problemsofwastecannotbe consideredasroutine problems. However,CBRsystemsaredevelopedtoresolveonlyroutine prob-lems i.e.problemswhichare verysimilar.Consequently,a system used for thesekinds ofproblems need togo beyond this limita-tionbytheintroductionofflexibility.Anotherpointishowtotake intoaccountthattherearedifferentpossibilitiesofvalorisationfor a samewaste.Forexample,inthecaseofusedtyres,theycanbe burntto produceenergy,reusedastyres, transformedby crunch-ingintomaterialfordifferentkindsofnewobject,transformedby fermentation toproduce syngaz. For each solution,the same de-scription parameters are not selected: for some solutions is the chemicalcomposition;forotheronesistheformorthe functional-ity,forotheronesmechanicalproperties.Therefore,asshowedby
Lieber (2002), problemsandtheir solutions depend on their use. As a consequence,authors thinkthat problemrepresentationand similaritymeasuredependonthesolutionorthekindofsolution targeted.
Inthispaper,authorsproposetoexplaintheirmethodsfor rep-resenting knowledge and cases, and for selecting relevant cases. Thesemethodstrytotakeintoaccountthesolutionandtherefore to adapt the similaritymeasure infunction to the important pa-rametersaccordingtoakindofsolution.Moreover,thesemethods do not produce a metric value of distance orsimilarity measure but, it determines if a case is similar to the current problem or not,i.e.ifthecasecanbeusedtogenerateanoriginalsolutionfor the problem. Contraryto Perner(2003), themethodis notbased ongraphs,anditdoesnotusethresholdorothermetricvalue,but it isbasedon logicaldeductions.Inconclusion, themajor contri-butionsofthispaperarethefollowing:
• Theintroductionofaflexiblerepresentationforknowledge. • A dynamic construction of cases, which allows going beyond
thelimitationofroutineproblems.
• Anewmethodforsimilaritymeasure,withoutcalculation and
withalimitedneedofknowledge.
In the remainder of this paper, the Section 2 explains some elements about CBR systems and develops some ideas for the realisation of each step finding in the literature. In Section 3, the proposed flexible representation of a case is described and more specificallythemanagement oftheknowledgeisexplained. Then, the core of this method is introduced with the presen-tation of the main assumptions, and the retrieve part is de-scribed step by step in Section 4. The Section 5 highlights the method capabilities through a case study, where some tests have been realised to assess the proposed method.Section 6 is-sues opinions about the positive points and the limitations of the method, and underlines some difficulties met during its implementation. Finally, Section 7 draws conclusions and
sum-Fig. 1. Steps in classical CBR.
marisesthepresentedwork,andproposesdifferentperspectivesto improveit.
2. Case-basedreasoning:differentrelatedsteps
Asexplainedintheintroduction,aCBRsystemisbasedon dif-ferentsteps(eachofthemdecomposesinsubprocessesnotdetail here)(Reyes,Negny,Robles,&LeLann,2015)(Fig.1).
However,therealisationofone stepimpactsalltheCBR’s pro-cesses.The representationoftheknowledge orcasesimpacts the sub-processesintheretrievalstep,forexamplethesimilarity mea-sureorthemappingphase.Therefore,itisnecessarytorepresent knowledgebytakingintoaccountthatretrievalstepusesit,i.e.the definitionofallthesub-processesdependsonthechoiceofakind ofrepresentation.Finally,thelastsub-processofthisretrievalstep oftheCBRistheselectionoftherelevantcaseinordertoreviseits solutiontomatchtothetargetcaserequirements.Onemechanism usedistheanalogy.Cornuéjols(1996)hasstudiedthefundamental ofthismechanism.He definedanalogicalreasoningasthewayto findtheexpressionwhichallowspassingfromapreviousproblem toits solutionandtoapplyit toanewtarget case.Here too,the representationofcasesisimportant.
IntraditionalCBR,theknowledgeisoftenrepresentedasa set of spaces. Napoli et al. (1996) explain that there is a space for the problem and another one for their solution. Mougouie and Bergmann (2002) define a query in CBR system as a point in thesespaces.Therefore,eachpoint ofthesespaceshastobe rep-resentedwithacommonmethod.Kokinov(1994) explainsthata cognitivemechanismisbasedonrepresentation, memorisation.In CBRandingeneralforallartificialintelligentsystems, representa-tionisonly apartialdescription ofthereality.As aconsequence,
MougouieandBergmann(2002)explainthat aqueryisonly par-tially described. For Peschl, it is an interpretation of the world which allows the construction of a behaviour (Peschl & Riegler, 1999).Underthisidea,AmailefandLu (2013)linkan ontology to aCBRsystemtofacilitatetheunderstandingofasituationandthe retrievestep.Thisinterpretationisveryimportantintheresolution phase as Richard highlights because a modification of the inter-pretationcanimprove theefficiencyofsolving methods(Richard, 1979). Finally,representation can be symbolic, based on connex-ions (Kokinov, 1994), defined as vector features, or complex as semanticnetwork (Branting & Aha,1995). Whatever, the manner to represent knowledge, it is a reduction of the reality. But, the choiceoftherepresentationapproachimpacts thesimilarity mea-sure step. Forexample, Branting and Aha (1995) and Garey and Johnson(2002)explainthattheutilisationofsemanticnetworkfor therepresentationofcasesinCBRcausesthatthemappingstepis
aNP-complextask.Napolietal.(1996)workwithanobjectbased representationallowingaclassificationofcases.
The Similarity step in CBR triesto findthe mostsimilar case to a new problem. Similarity is a fundamental part of the CBR (Rifqi, 2010), and it measures if two things shared some com-mon elements (Nesme & Hidalgo, 2013). In the literature, it is possible tofindthat similar casesresearch stepstarts witha de-scriptionofthecaseandsometimes byamappingstep.This pro-cess isdefined asthe identification ofthe relationships between the elements describing two cases, as suggested by Markman and Gentner (1993) which are NP-hard or NP-complete prob-lems (Sorlin & Solnon, 2005). For example, Falkenhainer, For-bus, and Gentner (1989) describe the structure-mapping engi-neering and explain the mapping result as the correspondence between the source case and the target case which can be im-proved by a set ofanalogical inferences. In addition, McFee and Lanckriet (2011)highlightthe questionof similaritybetween dif-ferent kinds of item. The authors propose a method to inte-grate heterogeneous data into a single unified similarity space and to consider some similarity comparison as a direct graph. Usually, similarity measure evaluates the distance between the target case and a source one (Richter, 1993). More generally, it can be defined as the task to find the closed point to a target one (Mougouie & Bergmann, 2002). In the literature for similar-ity calculation between two cases, it is possible to find plenty of methods. For example, Bisson (2000) proposes to estimate the similarity by the effort required to transform one case into the other one. Other methods compare each elements one by one once the mapping is realised. Avramenko and Kraslawski (2006) givethree kindsofsimilaritymeasures forCBR inprocess engineering.ThesekindsareQuantitativedistance,Hierarchical tree
andQualitativecomparisonwhichallowgivingadistancemeasure orasimilaritymeasureasanumber.SimilarityinCBRcanbe ap-pliedtoconcepts studyingthepositionofone concepttoanother one in a taxonomy structure (Wu & Palmer, 1994). Then, simi-larity could be the inverse of distance, however there are many definitions of distance (Bisson, 2000). Mougouie and Bergmann (2002) propose two methods based on optimisation. Armaghan andRenaud (2012) suggest a retrievestepbased on theuse ofa multi-criteriaselection toimprovethisstep.Therefore,some con-clusionsofpreviousresearchesleadtotheideathatsimilarity de-pends onthe studycase. Rifqi(2010)explains thatsimilarity de-pends on the general context of the domain and Goldstoneand Barsalou(1998)highlightthatitcanalsodependontheconditions of the study. Montani (2011) shows the importance of the con-text inCBRsystemwhichcan helptoreduce theretrievalsearch, to revise conclusions, and to adapt knowledge and strategies. In the samelogic,Leake studiesthepossibility toadaptthe similar-itymeasuretothecontext(Leake,Kinley,&Wilson,1996).Indeed, somedistancemeasuresarebasedonknowledgeintegratedinthe systemduringthedevelopmentstep.Itshowsthatsimilarity mea-sureneedsanadditionalknowledgewhichcomesfromthekindof problemssolved by theCBR.In thesameidea, Xiong(2011) pro-posesasystembasedonfuzzyruleswhicharelearnedbythe sys-temusinggeneticalgorithmonacasedatabase.Thissystemallows theadaptationoftheselectionandtheintegration.
Furthermore, anotherquestion is to knowhow the data have to be saved in the system. In other words, how the information is structured in the CBR system. To reuse cases, it is important to organise them under a structure facilitating the research and therefore the applicationof the similarity measure. Different ap-proaches are detailed in the literature, for example, Díaz-Agudo andGonzález-Calero (2001b)use GaloisLatticeforaCBR system. Fortheseauthors,thismethodoffersthepossibilityforthesystem toanswertodifferentdemands.BrantingandAha(1995)propose tousestratifiedcasebasedreasoning,usingabstractionofcasein
ahierarchicalstructure.Napolietal.(1996)studytheretrievaland adaptationstepsofCBRunderthesamedataorganisation.Aswell asforcasesstructuring,theorganisationofcasesinthedatabase isimportant.Usually,theorganisationisbasedonaconcept hier-archywhichcontainsnodesorderedbyrelationas“is_a” (Gennari, Langley,&Fisher,1989).Inlatticetheory,theorganisationincludes binary relations as “is a part of” or “is contained in” (Birkhoff, 1940).Theresultofthiskindoforganisationofconceptsiscalleda taxonomy.However,forDíaz-AgudoandGonzáles-Calero,the clas-sificationprocessbasedontaxonomystructureneedstoanticipate the questions submitted to the system (Díaz-Agudo& González-Calero,2001a).
Toorganisethedatastructurewiththeaimtosimplifythe re-trieval step, an approach is to generalise cases. This idea is not newandsome authorsexplain that itispresent sincethe begin-ningofCBR(Bareiss,2014;Kolodner,2014).Theuseofgeneralised cases gives manyadvantages such as thepossibility to usethem whentheproblemsdonotrepresentstructureallowingtobe par-tiallyordered(Napolietal.,1996).Generalcasescanbedefinedas a caseglobally described and thereforeit can incorporate differ-entcases.ForMaugouieandBergmann,ageneralisedcasecovera part oftheCBRknowledge andthey define itasa subsetof rep-resentationcase(Mougouie&Bergmann, 2002).Inthesameidea, Díaz-AgudoandGonzales-Calero group caseswithshared proper-ties (Díaz-Agudo& González-Calero, 2001b). It isalso possibleto findananalogywiththeclusteringmethod.Gennarietal.explain that conceptual clustering permits understanding the world and makingpredictions(Gennarietal.,1989).
Therefore,the nextpoints to takeinto accountare howto re-alise thesegeneralised casesandhowtoorganisethem.Different approaches can be found in literature. For example, Díaz-Agudo andGonzález-Calero(2001a)thinkthatontologiescanbeusefulto designknowledgeintensiveCBR,toreducetheknowledge acquisi-tionandtheyuseFormalConceptAnalysistoproducetheconcept lattice. Amethod fortheresolutionbased onknown casesisthe analogy.Itisdefinedasamappingofknowledgefromabaseand a target andcanbe usedin reasoning(Falkenhaineretal., 1989). According them, analogy allows generalizing casesin an abstract one. The implementation of an analogical reasoning depends on the knowledgerepresentation.Forexample,Cornuéjols(1996) ex-plains that thisprocess isbased onthecomparisonbetweentwo graphswhenacaseisdescribed asanetworkstructure.However,
Bunke andRiesen (2011)observe a lackof methodin the recog-nitionpaternswithgraphswhichhighlightsthecomplexityofthe task.
3. Flexiblecaserepresentation
Asexplainedintheprevioussection,therearetwomajorkinds of representation: the classic feature values description and an-other one based on connections as graph or semantic network. The firstone allows simplifying the similarityprocess because it avoids a random mapping phase. Each feature value is fixed for each case.Thisdescriptiondefinesaprioritherepresentationand therefore a partof the interpretation ofthe reality and it limits the kind of elements which can be described. However, for the aim application domain,it is important to havethe flexibility to represent different elements and to enable a most complete de-scription of the cases. Indeed, a waste can take several ways of description which are not common. This last point is important because the description has to adapt itself to the kind of solu-tions. Therefore,authorschose tousea kindofnetworkstructure to describe the knowledge. Authors define two levels of descrip-tionfortheknowledge.Thehigheriscomposedby twoelements:
statesandrelations.Inthislevel,itispossibletocomparestatesto nodes andrelations to edges ofa graph. Ourmodel isnot based
on graphtheory,butitcanbe representedbygraphs.Ina higher representation, thereisthedescriptionofthelinksbetweenthese states. Therefore, in our CBR system, authors do not define case in the knowledge structure. The second level is the detailed de-scriptionofastate.Thesepointswillbeexplainedinthefollowing parts.
3.1. Representationofthestate
The stateisadescriptionofanelement.Inoursystem, astate isrepresentedwithanetworkstructure.Thisstructureisbasedon connectionslinkedtoobjects(forexamplerubber)orconcepts(for example metals) and some parameters. These parameters permit including quantities as for example the number of objects con-tained in anotherone, values associatedto units orvalue ranges allowing the introduction of a kind of fuzzy logic. Therefore, it is not a binary relation but a predicate which represents a fact (Falkenhaineretal.,1989).
Definition1. Aconnectionbetweenconceptsorobjectsisdefined as:def(State,Relation,Object1,Numeric value1,Numericvalue 2, Unit,Object2)
A state represents a situation and therefore, it can represent different things.For example,inthe casestudy,a state definesa wasteorasetofwastes.However, itispossibletodescribeother thingssuchashumansituationorconceptualsituation.For exam-ple,describingthesituationbetweenateamwithitsmembersand otherelementsneeded.
An object is defined only by its relations with concepts as in an ontology. The name of an object is important only to en-sure the cohesion in a state description. Therefore, in a same state, it is important to ensure that a name of an object is al-ways used for the same thing. Consequently, the definition of an object is its relations with other objects or concepts as in ontologies.
Finally, all the concepts are linked in a taxonomy, which is a limited ontology. Therefore, our modelof representation of state is based on connections. In a globalview, it is possible to con-sider that each state is linked ina huge network and, therefore, thateachstateislinkedtootherstates.
Anotherelementistheintroductionofglobaldefinition.Aglobal definitionisasetofpropertiesconstitutingthestructurewhichare sharingbyallobjects.Forexampleiftableisdefinedwithaglobal definitioncontainingthesefollowingelements,isinwood,hasfour feet, each object respecting these properties is, by definition, a
table.
Definition2. Aglobaldefinitionisastructurecontainingaminimal setofpropertiesdefiningatypeofobject.
It is a majorpoint of ourmethodology becauseit definesthe similarity. In other words, if a description of a state (state_1) satisfied a global definition of another state (state_2) then this state (state_1) can be considered as equal to the second one (state_2). Example1. def(state1,is_composed_,tyre,_,_,_,rubber). def(state1,is_composed_,tyre,_,_,_,metal). def(state1,is_composed_,tyre,_,_,_,fiber). def(state1,has_the_form_of,tyre,_,_,_,torus). def(state1,has_the_color,tyre,_,_,_,black). def(state2,is_composed_,tyre_granule,_,_,_,rubber). def(state2,is_composed_,tyre_granule,_,_,_,metal). def(state2,is_composed_,tyre_granule,_,_,_,fiber). def(state2,has_the_form_of,tyre_granule,_,_,_,granule). def(state2,has_the_color,tyre_granule,_,_,_,black). def(state3,is_composed_,tyre_powder,_,_,_,rubber).
Fig. 2. Example of connections between states : following example 1 . def(state3,is_composed_,tyre_powder,_,_,_,metal).
def(state3,is_composed_,tyre_powder,_,_,_,fiber). def(state3,has_the_form_of,tyre_powder,_,_,_,powder). def(state3,has_the_color,tyre_powder,_,_,_,black).
where tyre, tyre_granule and tyre_powder are objects, rubber, metal, fiber, black, torus, granule and powder are concepts and
is_composed, has_the_form_of and has_the_color are relations be-tween objects and concepts. This is a simple example show-ing different states describing a process like in our case study.
Astateisaglobaldefinitioninoursystem,butaglobaldefinition
is not necessary a state. Therefore, the description of a casecan be different depending on the userand his interpretation of the reality.Indeed,a state described an objectorconcept inthe real worldwhichislinkedtootherstatesbyrelationswhereasaglobal definitioncan be the descriptionof a state orthe description of anabstractobjet belongingto thereasoningworldofthesystem. Theseabstractobjectsaretypesofrepresentationofstatebutthey donotrepresenttheconcreteobject.
3.2.Thelink,anelementcomposingthecase
Thispartdeals withthe connectionsbetweentwo states,that istosaythelink.Itrepresentsdifferentkindsofrelations.Indeed, arelationcandescribeafactoranactionbetweentwostates.For example, ifthere is a state describing a father, andanother one describing hisson, a relation is_son_of canlinked the two states. Therefore,it ispossibletoconsiderthiskindofrelationasan ex-tensionofthedescriptionofastate.However,themaindifferences comefromthe role ofthisrelation inthe CBRsystem. Indeed,if therearetwostates,eachstatecanplaytheroleofproblemor so-lutionduring theresolution process. Therefore,the choice of the description, i.e. if facts are described by different states or only one depends on the kind ofproblem submitted to the CBR sys-tem. Consequently, the model proposed can be used for several problems. Its specification dependson the representation of the information during the learning process. A relation can also de-scribethe resultof a process ora transformation.The difference withthefirstrelationdescribedisthetime.Withthefirstrelation, fatherandson, thetwostatescanexistinthesametime i.e. isa staticfact.However,inthesecondkindofrelation,astatewill ex-istafteranotherone. Thisreasoningisinaccordancewithhuman mind. But,in the CBR system, the relation is defined differently. It distinguishesrelations representingfactsand others represent-ing results of processes. For example,if there is tyre, “enhanced value” tyre_powder, itis possibleto describe thisrelationby two states,onefortyre,onefortyre_powderandalinkenhanced_value
asinan ontology orconcept map (Fig.2). Withthesame idea a productAcanbelinked toaproductBby transformation_1andit meansthatAistransformedintoBbytransformation_1.The advan-tageofthisstructureofknowledgerepresentationisthepossibility tolinkastatetosomeotherstates.Asaconsequence,itallows ex-pressingallthepossiblerepresentationsforastateunderdifferent
Fig. 3. Example of enhance number of possible cases using inference mechanisms.
Fig. 4. Example of dynamic cases.
meanings.InourCBRsystemthisconceptallowsdescribing differ-entpossibilitiesofwastetreatmentsorprocessesfromanelement orasetofelements.Italsopermitsbuildingdifferentlevelsof re-lations.Forexample,aprocesscanbedefinedbetweentwostates, andother processesincludingother statescan describe thesame knowledgeunderalower levelofrepresentationasinFig.3.This capacityofrepresentationisthebasisofourmethodtorevise so-lutions.
3.3. Dynamiccase
The last point is the description of a case. By definition, a case in CBR is the couple problem/solution. However, in some CBR systems there are two spaces inthe knowledge base as ex-plain in the Section 2. In our system, information needs to be sometimes used as problems, andsometimes as solution. There-fore, authors propose to define a case, i.e. a couple problem/so-lution as a knowledge structure composed by two states and a link.
Definition 3. A case isdefined asa set composed by two states andalinkwhereonestaterepresentsaninitialstate(theobjectof thequestion),thelinkiswhatiswanted(theverbofourquestion) andthelast staterepresentsthesolution.Onlytwoofthesethree elementsarenecessarytoconstitutetheproblem.
For example in (Fig. 4) there is the data: state1 (tyre)→ en-hanced_value→state3(tyrepowder).From thisdata,itispossible toinfer3problems:
• Whatisthefinalstatetoenhance_valueofatyre?
• Whatistheinitialstateofthetyrebeforetoenhanceitsvalue
?
• Howtoreachthestate3(tyre_powder)fromthestate 1(tyre)
?
Torealisethispart,someinferencemechanismsareusedwhich encapsulate statesaspartof thesolutionandothers asapartof theproblem.AuthorscallthisDynamiccases.
Definition4. InaCBRsystem,acaseisdynamicwhenthe identi-fication oftheproblempartandthesolutionpartisrealised dur-ing each Retrieve step. Thatis to say, the systemdoes not store casesbutonlyknowledge onstatesandtheir relations.Moreover, thisknowledgewithsomemechanismsproducecases correspond-ingtothecurrentproblemastheneedarises.
This mechanism has several advantages. One of them is the possibility to exploit moreinformation ofthe knowledge thanin a classicaldivisionofthespaces.Anotherone isthepossibilityto useinformation(statesorlinks)aspartofaproblemandasapart ofthesolutionduringthesameresolutionprocessi.e.forthesame problem.Thiscapacityofthemechanismisthebasisoftherevise stepinoursystem(notdetailedinthispaper).Indeed,thisstepis basedonthedecompositionofproblemintosub-problems,which allows adapting the solution using different cases and not only one.
4. Caseretrievalandsimilarity
In the literature, there are manyexamples describing the re-trieval step in CBR. In addition to a similarity measure, there is often a mechanism to try to identify the most similar case by limiting the exploration with for example, filters or indexation techniques. This part explains how the process of selection of similar cases occurs as well as the different mechanisms which allow reducing the time of research. The presented methodol-ogy is divided into two phases. The first one explains how the knowledge isstoredandprocessedinordertoapply theresearch step. The second one describes the research algorithm. The ma-jor difference betweenthetwo partsof themethodologyis their runtime. Indeed, the first part is realised as a learning step i.e. during the introduction of new knowledge. On the contrary, the research step occursduring theresolutionprocess. Therefore,the realisation of these two parts can be separated in time and in processes.
4.1. Pre-phase:learningphase
This partof the process isrealised during the introduction of newknowledgeinthesystem.Itcantakeplaceduring the initial-isation of the CBRsystem or during its utilisation thanks to the retain step. It is possible to divide thisprocess into three parts. Thefirstonerequirestheinterventionoftheexpertinknowledge managementandthetwoothersareautomaticallyoperatedbythe system.
Fig. 5. Creation and storage of common definitions in the learning step.
4.1.1. Pre-phasestep1:introductionofknowledge
The learning phasestarts when auserintroduces new knowl-edge containingasetofstateslinkedby relationsasexplainedin
Section 3.Toreduce knowledgeengineeringefforts,theuser pre-viously describeseach state, the relations betweenstatesand he completes thetaxonomiesifnewconceptsareintroduced.Finally, this knowledge is integratedin the systemsharing the same se-manticnetwork.
4.1.2. Pre-phasestep2:enhancerelationswithinferences
Once newknowledge is introduced in the system, the learn-ing stepstarts. The systemstarts to enhancethe numberof rela-tions betweenthestates.Thisprocessisbasedon theuseof tax-onomies for the relations andallows generalizing theserelations i.e. theprocessrealisesa conceptualisationofthenewknowledge focused on its relations. Inference mechanisms are used for this task withdifferentrules introduced during theinitial conception of the CBRsystem. There are two kinds of rules.The aim ofthe first category istoconceptualise relations.These rulespermit re-placing arelationby anotheronedefinedinahigherlevelofthe taxonomy.Theaimofthesecondoneistomakeinferencesonthe relations intheinput knowledge (Fig.3). Inother words,whena relationbetweentwostatesisdefinedwithotherrelations,authors propose to includetheserelations asapart ofthefirst one. This mechanismisrelevantbecauseitincreasesthenumberofpossible cases.
Forexample,ifthereisAis_transformintoB,Bis_transforminto C,Cis_transform intoD,andAis_recovered_inD,insome kindsof problems, it is possible to affirm that B is_recovered_in D and C is_recovered_inD.Inthesameway,ifarelationisdefinedina tax-onomy,itispossibletoenhancetherelationofthestatewithmore conceptual links.ForexampleifAisfixed_with_gluetoB,andina taxonomythereisfixed_with_glueisfixedthereforeauthorspropose toinferthatAisfixedtoB.
4.1.3. Pre-phasestep3:completionofcommondefinitionstructures
Oncenewknowledgeisintroducedanditsrelationshavebeen inferred,theCBRsystempreparestheresearch stepwithaphase of“learning”.Thisphase isa kindofindexationand conceptuali-sationoftheinformationasit ispossibletofindintheliterature (Section 2). However, there are manydifferences with the tradi-tional methods. The conceptualisation is not focused on case as forBichindaritz (2008) because casesdo not exist in the system (theyareinferred), butinstates.Moreprecisely, foreach relation (thebothdirectionsarepossible)betweenstates(inferredornot), statesaregroupedtogetherinacommondefinitionstructurewhich representseach kindofresolvableproblem(Section3).Acommon definitionstructureiscomposed bydifferentlevel ofstates.In the lowerone,therearethestatesintroducedinthesystem.Therefore, thelower levelrepresentsthereality.When anewstate is intro-ducedinthestructure,thesystemwillcreateacommondefinition
asresultofthecombinationofthisstateandthestatesexistingin thelowerlevel.
Definition 5. A common definition is a state arising from two states“origins” anditisaglobaldefinitionapplicabletothesetwo origins.Ifanobjectsatisfiesthepropertiesofthiscommon defini-tionthenthisobjectcansatisfythepropertiesofoneoriginbutis notsure.Conversely,ifitdoesnotsatisfytheproperties,itwillnot satisfythepropertiesofetherorigins.
Themechanismcreatesa secondlevelcomposedonlyby com-mon definitions. Then, the mechanism works with the new ele-mentscreatedinthislevelanditgeneratesnewcommondefinitions
ina higherlevel andso forth(Fig.5). Some system’sparameters permitdefiningtherateofmixingforeachlevel.
A common definition is generated using abstraction mecha-nismandconceptualisationmechanism.Thefirstoneallows delet-ing properties, i.e. relations as defined in part 3.1, which are not present in the two original states. The second one, the
Fig. 6. Mechanism of research.
conceptualisation, isbasedon theuseoftaxonomies ofconcepts. If the two original states have the same property but defined at different levels of conceptualisation (for concept, relation or both),thismechanismcreatesorselectsthepropertyinthehigher level.Therefore,thisnewproperty satisfiesthetwooriginal ones. For example if there are the two original properties: is in cop-per, is in led, the property is in metal satisfies the two origi-nal ones if, in a taxonomy, copper andled are defined as metal (copper is_a metal and led is_a metal). For numbers, the fuzzy logic is used with range value which contains the two original values.
Thestructureobtainedisnotaclusterbecausethereisnoroot. Indeed,fora combinationoftwo states,the mechanismcan pro-duce different common states ifthey contain differentobjects. In thiscase, a kindof matchingprocess is realisedandit generates allthemappingpossibilities.Allthecombinationswithobjectsare possible ifthetwo matched objectsshared atleastone common property.However,inthestructure,implausiblecommonstatesare not saved inhigherlevels because theydo not sharedproperties with the others. Therefore, the more a level is higher, the more the common states contained properties shared by all the states. Then,two processesaresupportedbythestructure.First,thereis an indexing which organises states. Second, there is a filterand weighting system. Fortherelation,higherlevels contain onlythe mainpropertiesandthemoreapropertyispresentinhigherlevel, moreisimportant.
Inconclusion,inthelearningphasesomeknowledgestructures areenrichedbythenewstateandbythecreationofcommon defi-nitionsandthereisonestructurebyrelations(links)originatingin thisstate.Thelevelzeroofthisstructureisthereality.Thehigher thelevelis,themoretheconceptualisationandabstractiondegree isimportantandthelevelsarecomposedbystateswithverylarge definition.
4.2. Retrievephase:researchofsimilarstates
This partof the process isrealised during the resolutionof a problem.Itisbasedontheuseofcommondefinitionstructures cre-ated orcompleted inthelearning phase. It iscomposed ofthree steps.
4.2.1. Retrievephasestep1:selectionofthecommondefinition structure
Asthe problemisdefinedasthecombinationofa stateanda relation,thefirststepoftheretrievepartistoselectthecommon definitionstructure correspondingto therelationandthetype of thesubmittedproblem,(Fig.4).
4.2.2. Retrievephasestep2:evaluationofstates
The next step is to check if the problem’s state satisfied the commondefinitionspresentintheselectedstructure.Theresearch mechanismconvertsthedefinitionofthesestatesintoruleswhere objects areconvertedintovariables.Then,itstartstocheck ifthe rules are applicable to the problem’s state, i.e. if the problem’s state contains all the properties contained in the rules withthe same level of conceptualisation or a lower one. The mechanism begins with the common definitions of the higher level of the structure. If a common definition is satisfied, it continues with the original ones. Ifitisnot satisfied,the mechanismchecks an-othercommondefinitioninthehigherleveluntilthereisnomore (Fig. 6). The mechanism stops the exploration when a common definition fromthe lower level is verified (andit continues with anotherone fromthehigherlevel)orwhenallthecommon defi-nitionsfromthehigherlevelweretested.
Duringthisphase,eachverifiedstateisstoredinalistwithits associatedlevel.
4.2.3. Retrievephasestep3:selectionofthemostsimilarstate
Oncetheexplorationisfinished,anascendingsortonthelevel is realised withthe stored elements. Logically andfollowing the descriptionofcommondefinitions,ifthereareverifiedstates com-ing fromthelevelzero(lowerlevel),that meansthat thecurrent problem can be defined assolved problems existing in the data base. Therefore,they can be considered assimilar to the current problemandusedtosolveit.Ifthereisnostatecomingfromthe levelzero,thesystemwillselectthestatewiththelower level.It isnotarealstatebutastategeneratedduringthelearningphase. However, authors proposeto defineit asthemostsimilar oneto thecurrentproblemandtouseittosolveit.Here,itispossibleto measurethesimilarity.
Definition 6. In a commondefinitionstructure, the morea state isverifiedwithalower level,themoresimilaritistothecurrent one.
Finally,ifthereareseveralverifiedstatesinthelowerlevel,the systemcan apply differentpolicies.Asthissystemcan not deter-minewhichoneisthemostsimilar(allareinthesamelevel),the systemcanrandomlyselectoneorproposeseachonesasa possi-blesolution.
5. Casestudy:recoverytreatment
Authors implement themethodpreviously described in aCBR system dedicated to generate new recovery processes for wastes
treatment. Theideaistouseknownrecovery processforawaste oratypeofwastetoproposenewonesforotherdifferentwastes. The aim of the CBR is not to define each unit process with all the parameters and to give solutions ready for use, but to give the main steps of a new process and to try to estimate the fi-nal product and what will be its functions or potential applica-tions. Therefore, authors developed a CBR based on logical pro-gramming paradigm withthe SWI-Prolog implementation,which is a free software1 andit comes forLinux2 .SWI-Prologextends Prologlanguage(Wielemaker,2014).
5.1. Data
The datausedcomesfromdifferentknownrecoveryprocesses for6typesofwaste.Theselectedwastes andtheir solutionshare somecommonstepsandgenerallysomerelationsallowingtheuse ofCBRsystem.The6kindsofwasteare:
• Wastescomposedofpolypropylene
• CRT television composed of elements containing glass, metal,
plasticelements
• Neon tube composed of elements containing glass, metal,
chemicalcompounds,gas
• Glassbottle
• Wastescomposedofaluminium • Carbattery
Forexample,the definitionsofsome wastesintroduced inthe systemarethefollowing:
( d e f ( compose , bottle˙cap ,˙,˙,˙, p o l y p r o p y l e n e ) , d e f ( has , bottle˙cap ,˙,˙,˙, metal ) ,
d e f ( s i z e , bottle˙cap , 3 ,˙, cm ,˙) , d e f ( form , bottle˙cap ,˙,˙,˙, tube ) ) ,
or
( d e f ( has , neon˙tube ,˙,˙,˙, g l a s s ˙ t u b e ) , d e f ( composed , glass˙tube ,˙,˙,˙, g l a s s ) , d e f ( has , neon˙tube ,˙,˙,˙, powder˙PhM ) , d e f ( composed , powder˙PhM ,˙,˙,˙, phosphorus ) , d e f ( composed , powder˙PhM ,˙,˙,˙, mercury ) , d e f ( has , neon˙tube ,˙,˙,˙, piece˙metal ) , d e f ( composed , piece˙metal ,˙,˙,˙, metal ) , d e f ( form , glass˙tube ,˙,˙,˙, tube ) ) ,
Each waste can have several recovery processes, furthermore each process can be divided into other sub-processes in func-tion of separation steps. To complete the knowledge base ofthe system, authors include taxonomies on operations and concepts. In this system, a taxonomy can be completed progressively de-pending ontheconcepts usedin theknowledgebase.The taxon-omy on operationsisa tree structureordering processes in fam-ilies and sub-families. With the same idea, taxonomies on con-ceptsaredividedintotwostructures.Oneconcernsthecomponent where it is possible to find concepts as glass, metal, aluminium, etc. Theotheroneisaboutgeometryandallowsdescribingforms and architecture ofobjects.Each transformationstep ismodelled
1http://www.swi-prolog.org/license.html.
2http://www.fraber.de/university/prolog/comparison.html.
in the system asa link and each product orintermediate prod-uct is defined as a state. Moreover, authors define relation be-tweenstatescreatingacrudeapproximationoftheprocess.These linksareincluded inthe taxonomy.Forexample,a waste is con-nectedtotheendproductwithalinkdefiningthewholerecovery process.
5.2.Experimentsandresults
All the data are introduced into the CBR system and the learning phase is launched. Authors only present a fragment of the database obtained. It contains a huge number of infor-mation, therefore, only three kinds of fragment are presented below:
• Fragment of Def, which describe the definition of states (the
informationiscondensed).
• Fragmentoftaxonomies.
• Fragment of Relations, which describe the relations between
twostates.
FragmentofDefdatabase:
:− dynamic d e f / 7 . d e f ( [ 2 ] , t a i l l e , bouchon , 3 , ˙, cm , ˙) . d e f ( [ 3 ] , t a i l l e , pare˙choc , 2 , ˙, m, ˙) . d e f ( [ 4 ] , c o n t i e n t ˙ t r a c e , b r o y a t ˙ p o l y p r o p y l e n e ˙ s a l e , ˙, ֒→ ˙, ˙, metal ) . d e f ( [ 7 ] , compose , poudrette˙polypropylene˙humide , ˙, ˙, ֒→ ˙, p o l y p r o p y l e n e ) . d e f ( [ 7 ] , t a i l l e , poudrette˙polypropylene˙humide , 1 , 3 , ֒→mm, ˙) . d e f ( [ 7 ] , forme , poudrette˙polypropylene˙humide , ˙, ˙, ˙ ֒→ , p o u d r e t t e ) . d e f ( [ 7 ] , c o n t i e n t ˙ t r a c e , poudrette˙polypropylene˙humide ֒→ , ˙, ˙, ˙, eau ) . d e f ( [ 8 ] , compose , p o u d r e t t e ˙ p o l y p r o p y l e n e , ˙, ˙, ˙, ֒→ p o l y p r o p y l e n e ) . d e f ( [ 8 ] , t a i l l e , p o u d r e t t e ˙ p o l y p r o p y l e n e , 1 , 3 , mm, ˙) . d e f ( [ 8 ] , forme , p o u d r e t t e ˙ p o l y p r o p y l e n e , ˙, ˙, ˙, ֒→ p o u d r e t t e ) . d e f ( [ 9 ] , c o n t i e n t , tv , ˙, ˙, ˙, tube˙cathodique ) . d e f ( [ 1 0 , 9 ] , compose , tube˙verre , ˙, ˙, ˙, verre˙plomb )
֒→ . d e f ( [ 1 3 ] , compose , b r o y a t ˙ p l a s t i q u e , ˙, ˙, ˙, p l a s t i q u e ֒→ ) . d e f ( [ 1 3 ] , forme , b r o y a t ˙ p l a s t i q u e , ˙, ˙, ˙, b r o y a t ) . d e f ( [ 1 4 ] , compose , r e s i d u , ˙, ˙, ˙, s a l e t e ) . d e f ( [ 1 4 ] , forme , r e s i d u , ˙, ˙, ˙, b r o y a t ) . d e f ( [ 1 6 ] , compose , verre˙broye˙sale , ˙, ˙, ˙, v e r r e ) . d e f ( [ 1 7 ] , compose , verre˙broye , ˙, ˙, ˙, v e r r e ) . d e f ( [ 1 9 ] , compose , verre˙plomb˙broye˙sale , ˙, ˙, ˙, ֒→ verre˙plomb ) . [ . . . ]
Fragmentoftaxonomydatabase: :− dynamic o n t o l o g i e / 2 . o n t o l o g i e ( t r a n s f o r m a t i o n ( broyage˙bouchon ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( broyage˙parechoc ) , t r a i t e m e n t ) ֒→ . o n t o l o g i e ( t r a n s f o r m a t i o n ( b r o y a g e ˙ f i n ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( broyage˙2 ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( broyage˙3 ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( broyage˙4 ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( broyage˙5 ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( b r o y a g e ˙ b o u t e i l l e ) , t r a i t e m e n t ֒→ ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( broyage˙alu ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( b r o y a g e ˙ b a t t e r i e ) , t r a i t e m e n t ) ֒→ . o n t o l o g i e ( t r a n s f o r m a t i o n ( f l o t a t i o n ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( t r i ˙ a ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( t r i˙ b ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( separation˙poudre˙a ) , ֒→ t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( separation˙poudre˙b ) , ֒→ t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( separation˙chimique˙a ) , ֒→ t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( separation˙chimique˙b ) , ֒→ t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( t r i 2 ) , t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( separation˙metal˙a ) , ֒→ t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( separation˙metal˙b ) , ֒→ t r a i t e m e n t ) . o n t o l o g i e ( t r a n s f o r m a t i o n ( s e p a r a t i o n ˙ v e r r e ) , t r a i t e m e n t ) ֒→ . o n t o l o g i e ( t r a n s f o r m a t i o n ( separation˙oxyde˙plomb ) , ֒→ t r a i t e m e n t ) . [ . . . ]
Fragmentofrelationdatabase: :− dynamic r e l a t i o n / 3 . r e l a t i o n ( t r a i t e m e n t , 2 , 6 ) . r e l a t i o n ( t r a i t e m e n t , 3 , 6 ) . r e l a t i o n ( p r e t r a i t e m e n t , 6 , 8 ) . r e l a t i o n ( r e v a l o r i s a t i o n ( m a t i e r e ) , 2 , 8 ) . r e l a t i o n ( r e v a l o r i s a t i o n ( m a t i e r e ) , 3 , 8 ) . r e l a t i o n ( p r e t r a i t e m e n t , 9 , 1 1) . r e l a t i o n ( t r a i t e m e n t , 1 1 , 1 3 ) . r e l a t i o n ( t r a i t e m e n t , 1 1 , 1 4 ) . r e l a t i o n ( p r e t r a i t e m e n t , 9 , 1 5) . r e l a t i o n ( p r e t r a i t e m e n t , 9 , 1 8) . r e l a t i o n ( t r a i t e m e n t , 1 5 , 1 6 ) . r e l a t i o n ( t r a i t e m e n t , 1 8 , 1 9 ) . r e l a t i o n ( p r e t r a i t e m e n t , 1 6 , 1 7) . r e l a t i o n ( p r e t r a i t e m e n t , 1 9 , 2 0) . r e l a t i o n ( r e v a l o r i s a t i o n ( m a t i e r e ) , 9 , 13 ) . r e l a t i o n ( r e v a l o r i s a t i o n ( m a t i e r e ) , 9 , 14 ) . r e l a t i o n ( r e v a l o r i s a t i o n ( m a t i e r e ) , 9 , 17 ) . r e l a t i o n ( r e v a l o r i s a t i o n ( m a t i e r e ) , 9 , 20 ) . r e l a t i o n ( t r a i t e m e n t , 2 1 , 2 5 ) . r e l a t i o n ( t r a i t e m e n t , 2 1 , 2 6 ) . r e l a t i o n ( t r a i t e m e n t , 2 1 , 2 8 ) . [ . . . ]
Then, different questions are submitted to the CBR system to try to assess the efficiency and the capabilities of the proposed methods.Thefirstelementtestedisthelearningphase.Duringthis phase, the systemproduced some structures containing common definitionsasexpected.Itshowsthatthemechanismisrealisable. First,all the statescomposingthedifferentprocesses linkedwith
Fig. 7. Example of result, where each number represents a state. different connexions had been inferred. Authors have filled the knowledgebasewith45statesandtheinferencemechanisms gen-erated 137 differentreal possible relations.Then, it creates some structures witharound5000commondefinitions.ThentheCBRis used to solve some problemsdescribing newstates and the de-sired link.Inthefirstrequestsomeonedemands solvedproblems introduced inthesystemasdata.Allquerieswork welland, dur-ingthetests,thereusestepgivesrelevantanswerswhenthereare no differencesbetweenthe desiredproblemandaknown solved problems. Ina second phase,the systemistestedwithproblems where inference mechanisms are required, for example, by sub-mitting problemscomposed by intermediate productand general definitionofthe process(the linkoriginatingfromthe wasteand finishing tothenewproduct). Oneexampleisillustrated inFig.7
whereaconcreteproblemissolved:
1. Thewaste whichis a smallmaterial composed by polypropy-lene(state2)isfoundtoberecoveredintopolypropylene pow-der readyforuse (state8) via two mainprocesses definedas treatment and pretreatment. Treatment groups correspond to materialmodificationoperationsandpretreatmentis condition-ing.
2. The treatment phase is identified by three operations: crush-ing_cap which corresponds to an operation of crushing little pieces in polypropylene, separationof non-polypropylene ma-terial and another phase of crushing to obtain a powder of polypropylenewithsomeimpurities.
3. Thepretreatmentphasewhichconsistsinwashingthepowder obtainedandtodryit.
Heretoo,foralltheteststhesystemgivesrelevantanswersthat means each step ofthe process andan estimationof each states representingintermediateproductsworkcorrectly.Afterthese ex-periments tovalidateasimplereusemechanism,thesystemhave tosolvetwocategoriesofstateswhichwerenotinthedatabase.
The firstone isdescribing states,copiesofexisting states,but wherepropertiesweremodifiedwithnewconceptsderiving from theoriginalone.Forexample,ifthereisapropertyintheoriginal state asAis_fixed_withconceptA,authorsdefinea newstate with thefollowingpropertiesAis_fixed_with conceptBwhereconceptB
is_a conceptAinthetaxonomy.The mechanismof conceptualisa-tion istestedand, moreprecisely, thesystem’scapacityto affirm that two elements which are not in thesame level oftaxonomy descriptioncanbesimilar.Forexample:
1. Ais_fixed_withconceptA
2. Ais_fixed_withconceptB
3. conceptBis_aconceptA
itispossibletoevaluatethefollowingassertions: Role A Role B
2 is true ⇒ 1 is true
2 is false 6⇒ 1 is false
1 is true 6⇒ 2 is true
Fig. 8. Example of fail return in the resolution process. Therefore, the system tries to determine if a property of the
current state satisfied a property of a known state, the current state plays the Role A and the known state the Role B. The sys-tem triesto establishthat thecurrentstate hasthepropertiesto satisfy thedefinitionoftheknown state.Thisiswhyauthorscall this kindof description Globaldefinition.Here again, as expected the mechanismwas validatedandgoodresultswerefoundin ac-cordancewiththetheory.
The secondone istoincrease thenumberofpropertiesinthe state comparedtotheoriginal one.Theideatestedhereisto en-surethattheabstractionmechanismworkscorrectlyandtherefore, that the interpretation of the state by the user does not impact theglobalresolutionifitsatisfiestheminimaldescriptiondefined by the Global definition. For thistest, authors create some states ascopiesofstatescomingfromtheknowledgebase.Thenauthors add some properties tothem toensure that thegenerated states arenotexactcopiesofexistingones.Forexample,state_1hasthe followingdefinition:
1. Arelation_1Concept1
2. Arelation_2Concept2
3. Arelation_3Concept3
Anewstate_2canbecreatedwiththefollowingdefinition: 1. Arelation_1Concept1
2. Arelation_2Concept2
3. Arelation_3Concept3
4. Arelation_4Concept4
5. Arelation_5Concept5
The first definition is included in the second one. Let define
Def1 the set ofproperties describing state_1 andDef2 the set of
propertiesdescribing state_2.Thefollowing assertions canbe es-tablished: De f1 ⊂ De f2 De f1 is true 6⇒ De f2 is true De f1 is false ⇒ De f2 is false De f2 is false 6⇒ De f1 is false De f2 is true ⇒ De f1 is true
Therefore, themechanismofcomparisonverifies thatthe def-inition of thecurrentstate isequal orincludingthedefinition of the state comingfromthedata baseandafterthat all properties of thislast state are verified by thecurrent one. Duringthe test where severalproblems have been submitted, all solutions have beenfoundshowingthatthemechanismgivesrelevantresultsand
thereforethat a morecomplete descriptionthan theoriginal one doesnotstoptheresolutionstep.
The third test is the opposite of the second one. The idea is to propose to the systemto solve states forwhich, compared to thestatesinknowledgebase,somepropertiesaremissing.Forthe example of the second test, roles are reversed. This test tries to assesstheanswersofthesystemwherethedefinitionofthe cur-rentstate isincompletetoverifyaknownsolved state.Therefore, authorssubmitdifferentstatesdeletingdifferentpartsofthe def-inition.Forall thetests, noknownstate comingfromthelevel 0 hasbeenfound.Therefore,thesystemgivesthemostsimilarstates found(Fig.8).
Theseresultscanbeexplainedbecauseknownstatesfoundare
Commondefinitionsfromthelevel1orupper.However,thesystem describesinthe previous part,returnsthe origins(real states) of thecommondefinitionwiththelowerlevel.Someofthemare ran-domlyselectedandthesolvingprocess (notdescribed inthis pa-per)usedthemtoestimate thegeneratedproduct.The generated resultsaredifferent.Someofthemcorrespond totheoriginal so-lution.However,othersareunexpectedandproposeprocessesnot very compatiblewiththe original state. But, all the returned so-lutionsarelogiccompared tothedescriptionofthecurrentstate. Finally, authors conclude that the proposed system works under thelogicofGlobaldefinition(thefactthataGlobaldefinitionisthe minimaldescriptionofanobject).
6. Discussion
Themethoddetailedinthispaperallows realizingsome steps ofaCBRsystemdesignedto generatenewrecovery processesfor waste. In a first time, it permits describing the knowledge un-der twolevels. Thestate representsa situationora thingwitha modelbasedonconnectionsusingrelations,conceptsand numer-icalproperties.Thisrepresentationallowsaflexible descriptionof asituationanditallowsrepresentingaverywidevarietyof situa-tion.Thesecondlevelisthenetworkcomposedbystatesandlinks orrelations.Itrepresentstherelationsbetweenstatesanditisthe levelofproblemresolution.
Theproposedmethoddoes notstoreknowledgeinaspacefor problemsandanother forsolutions.There is onlyone containing thenetworkofstatesandlinks. TheCBR’scasesaregeneratedby asetofinferencemechanismswhichdefineapartasproblemand anotherassolution.Thismethodallows modifyingthestatusofa setofknowledge (problemorsolution) duringthe sameproblem resolution.
Theapplicationofinferencemechanismspermitsenhancingthe knowledge storedwithuseoftaxonomies.Therelations and con-ceptsdescribing statesorrelations betweenstatesare more con-ceptualised. Therefore, it increases the possibilities to retrieve a similarcase.
The proposed retrieve method is not common for traditional CBRsystem. Indeed,itdoes not measurea distancebetweentwo points of the knowledge space and does not use any weight or similarity values.The method isbased on the assumption that a state encompasses the minimal set of properties needed to de-scribeasituation.Therefore,asituationverifyingalltheproperties ofastateisconsideredasdescribingthesamestate.Theclassical CBRsystemsprovideaminimumrangesetofproperties.Thesesets describe problemsandposition them toeach other.Almost, they are limited tothe minimal set which allows doing this position-ing usedby theRetrieve step.However, theproposed description methodlinksasetofpropertiestoamajorconceptrepresentedby a state, that is to say, a setdescribes a state whichis a concept asanobjectorawaste.Therefore,themethodtriestoidentify ob-jects orconcepts,whereas classicalCBRstrytopositionthe prob-lems retrievetoeach other.Theretrievemethod isbasedonthis assumption. Structures of knowledge containing states and more conceptual andabstract statesare buildingforeach relation dur-ing the learning phase. These structures allow reducing the time ofresearch,filteringtheimportantpropertiesforthe relationand weighting them. Therefore, the method permits a retrieve CBR’s stepbasedonlogicaldeduction.
The resultof thismethod is theacquisition ofa list ofstates orderedindecreasingsimilarityvalues.Theinferencemechanisms allowenhancingtheflexibilityoftheresearchandpermit consider-ingasequalelementsindifferentlevelsofdescriptionand enhanc-ing thecreativityofthe CBRsystembytherealisation oforiginal combinations ofobjects orconcepts. These original combinations arealogicalconsequenceofthemechanismsdescribed inSection
5.2,i.e., a state or a partof a state can be considered assimilar to amore conceptualor moreabstractobjet. Therefore,solutions which are not known toresolve the currentproblem can be ap-pliedtothisstate.Asanexample,aglassbottlecanbeconsidered asglassmaterialandtheglassmaterial’s solutionscanbeapplied totheglassbottle.However,aglassbottlecanbeconsideredasa container likea woodboxoraflower potforexample.Therefore some solutions applied to a boxwoodor to aflower pot can be usedtoresolvetheglassbottleproblem. Authorsthinkthatthese kindsofreasoningcanleadtocreativeprocesses.
Inaddition,acomparisonbetweensomemethodsfoundinthe literatureandtheproposedmethodcanbesummarised.Themain distinctionisthefactthatthereisnotaproblemspaceanda solu-tionspaceintheexposedmethod.ThemajorityofCBRsystemsare basedonthisdistinctionwhichreduce thepossibilityofthiskind of system. Moreover, the presented method contains some com-mon elements with traditionalCBR. As explain in the Section 2,
Amailef and Lu (2013) use an ontology to improve the compre-hensionof a submittedproblem andits representation. However, intheproposedmethod,theontologycontainsalltheinformation anditisthesupportofthegenerationofcase.Therefore,the ontol-ogyevolvesduringthetime.Anothercomparisonisthesimilarity measure. Theproposed measure isnot basedon adistance mea-surewiththeuseofmathematicalformulaasinmostoftheCBR systemsbutifthesubmitted problemisa partofarealgroupof caseormoreconceptual groups.As forXiong(2011),themethod adaptsbythecreationofnewgroups,commondefinitions,andby the introduction of new concepts in the taxonomies. Finally, the flexibilitytodescribeacaseisamajordistinctionwithamainpart ofCBRsystempresentedintheliteraturereview.
However, the proposed method has some limitations. As the knowledge is described by statesand links, the methodimposes
that aproblemcanbe describedunderthisform. Inthesystema state hastodescribeastaticsituationintheintellectualapproach oftheproblemresolution.Inother words,astatehastorepresent astepinthisresolution.
Anotherlimitation isthe mainassumption ofthismethod,i.e. the capacity to describe a situation with a minimal set of prop-ertiestakingintoaccountthatallsituations withtheseproperties will be considered assimilar. Whereas, it increasesthecreativity of the problemresolution, thispoint is alsoa limitation because itcanconsiderequaltwodifferentsituationsbecausethe descrip-tion ofthe stateis notstrict enough andthereforeitcan leadto inconsistentassociation.
The use of taxonomies can also be a limitation in the CBR system. Ataxonomy is a data structure wheredifferent concepts are organised ina hierarchicalstructure. However, this hierarchi-calstructuredeterminesaninterpretationofthereality.This inter-pretation impactsandlimitsall themechanisms usingthese tax-onomies during theproblemresolution phase andtherefore, this phaseisorientedtofollowthisinterpretation.Inother words,the useoftaxonomiesreducesthequantityofsolutiongeneratedand requiresasharpknowledgeontheapplicationdomain.Inaddition, thismethodneedstobeabletocreatethesetaxonomies.
Finally, the realisation of this method raises different prob-lems. To realise the retrieve part,the CBR systembuilds knowl-edge structure containing states and common definitions which are combinationofthepropertiesofthesestates.Inaddition,the combination oftwo states can produce differentcommon defini-tion allowing a kindof creativity and the inference mechanisms increasethenumberofpossiblecombinations.Therefore,the num-ber ofcommondefinitiongrowsexponentially withthe introduc-tion of newstates. In ourapplication, for a numberof 45 states described and137possiblecasesgeneratedbyinferences,the sys-temduringthelearningphaseproducedaround5000Common def-initions.Thestructure’sparameters,asthenumberofslicesorthe rateofmixing,arenotoptimallydefined.Theconsequencesofthis is the tremendous computational time ofthe learning step. Also, this original number of possible usable cases is small (147) but the trajectories described share part of solutions or some com-monstates.Thissharingbringstolightthepossibilitiesofthis pre-sentedmethodby apossible recombinationofsolutions and pro-ducingacreativeprocess.Nevertheless,thismethodwillbetested witha significantnumberofcaseswhena secondversion ofthis methodwillbedeveloped.
Anotherdifficult pointis thepossible randomselection ofthe mostsimilarknownstate.Infact,ifunderthelogicofthesystem there is no doubt, it appears to be important to develop a good policyofselectiondependingoftheresolvingmethod.Thus,itcan be interestingtoselecteveryknownstatesfromthelowerlevelif thenumberofpossiblecombinationsisnotimportantinthe reso-lutionprocess.Onthecontrary,therandomselectioncanproduce non-deterministicsolvingprocessandsomegoodsolutionsmaybe lost.
Finally, another difficulty can appear if the state is described with alotof slices.Forexampleif astate is described wherean objectiscomposedbyother objectsdefinedwithother objectsor concepts,therearenoproblemsduringtheresearchstep.However, once a source caseis selected fromanother level than the level 0,themechanismofadaptationhastoresolvearandommapping process because the satisfied common definition was done with theabstractionofsomeproperties.
7. Conclusionandoutlook
Thispaperdeals withthesimilaritymeasures inCBRandalso withtherepresentationandmemorisationofknowledge.Casesare not described withthe classical feature-value representation, but
itproposestodescribethemwithanetworkstructure.Inthe pro-posedmethod,knowledgeisstoredsothatitenablesthe genera-tionofdynamiccaseandtheapplicationofinferencemechanisms. Thesemechanismspermitincreasingtheflexibilityofthesystem’s logic andthereforeto give manyoriginal solutions.It alsoallows weighting the importanceof some propertiestakingintoaccount thecontextoftheproblembutalsothekindofsolution.Toreach thisgoaltwodefinitionsofconceptsareintroduceswhicharethe base of thismethod. Moreover, itpresents a structure composed by commondefinitions whichplays the role ofindexation mech-anism andfilter.All thesepoints enable to designa flexible CBR which can be used with very different kinds of problems. How-ever, two limitationsare identified.Firstly, to be ableto describe theknowledgeunderthestructureofstate-relation-statewhere thepropertiesofthesituationortheobjectarecontainedinstate. Thesecond istohaveseveralstateslinkedwiththesamerelation to providethenecessaryelementsto generatethestructure com-posedbycommondefinitions.
One way to improvethis methodis to reduce the number of
commondefinitionsgeneratedby thesystemtonotincrease expo-nentiallythetimeofthelearningstep.Anotherpointshouldbethe introduction of full ontology andnot one limitedto a taxonomy structure.Therefore,moredevelopedinferencemechanismsshould be introduced toincrease thepossibilities ofthe systemwithout reducingitsflexibility.
References
Aamodt, A. , & Plaza, E. (1994). Case-based reasoning: foundational issues, method- ological variations, and system approaches. AI communications, 7 (1), 39–59 .
Amailef, K., & Lu, J. (2013). Ontology-supported case-based reasoning approach for intelligent m-government emergency response services. Decision Support Sys-
tems, 55 (1), 79–97. doi: 10.1016/j.dss.2012.12.034 .
Armaghan, N., & Renaud, J. (2012). An application of multi-criteria decision aids models for case-based reasoning. Information Sciences, 210 , 55–66. doi: 10.1016/j. ins.2012.04.033 .
Avramenko, Y., & Kraslawski, A. (2006). Similarity concept for case-based design in process engineering. Computers & Chemical Engineering, 30 (3), 548–557. doi: 10. 1016/j.compchemeng.2005.10.011 .
Bareiss, R. (2014). Exemplar-based knowledge acquisition: a unified approach to con-
cept representation, classification, and learning . Academic Press .
Bichindaritz, I. (2008). Memory structures and organization in case-based reasoning. In D. P. Perner (Ed.), Case-based reasoning on images and signals number 73 . In
Studies in Computational Intelligence (pp. 175–194). Springer Berlin Heidelberg . Birkhoff, G. (1940). Lattice theory . American Mathematical Soc .
Bisson, G. (20 0 0). La similarité : une notion symbolique/numérique.
Branting, L. K. , & Aha, D. W. (1995). Stratified case-based reasoning: reusing hierar- chical problem solving episodes. In IJCAI (pp. 384–390). Citeseer .
Bunke, H., & Riesen, K. (2011). Recent advances in graph-based pattern recognition with applications in document analysis , 44 (5), 1057–1067. doi: 10.1016/j.patcog. 2010.11.015 .
Cordier, A. , Mascret, B. , & Mille, A. (2009). Extending case-based reasoning with traces. In Grand challenges for reasoning from experiences, workshop at ijcai: vol. 9 (p. 31) .
Cornuéjols, A. (1996). Analogie, principe d’économie et complexité algorithmique.
Actes des 11èmes Journées Françaises de l’Apprentissage .
Díaz-Agudo, B. , & González-Calero, P. A. (2001a). Classification based retrieval using formal concept analysis. In D. W. Aha, & I. Watson (Eds.), Case-based reason-
ing research and development number 2080 . In Lecture Notes in Computer Science (pp. 173–188). Springer Berlin Heidelberg .
Díaz-Agudo, B., & González-Calero, P. A. (2001b). Formal concept analysis as a sup- port technique for CBR. Knowledge-Based Systems, 14 (3), 163–171. doi: 10.1016/ S0950-7051(01)0 0 093-4 .
Falkenhainer, B., Forbus, K. D., & Gentner, D. (1989). The structure-mapping en- gine: algorithm and examples. Artificial Intelligence, 41 (1), 1–63. doi: 10.1016/ 0 0 04-3702(89)90 077-5 .
Garey, M. R., & Johnson, D. S. (2002). Computers and intractability.
Gennari, J. H., Langley, P., & Fisher, D. (1989). Models of incremental concept forma- tion. Artificial Intelligence, 40 (1), 11–61. doi: 10.1016/0 0 04-3702(89)90 046-5 .
Goldstone, R. L., & Barsalou, L. W. (1998). Reuniting perception and conception. Cog-
nition, 65 (2), 231–262. doi: 10.1016/S0010- 0277(97)00047- 4 .
Kokinov, B. (1994). A hybrid model of reasoning by analogy. Advances in connection-
ist and neural computation theory, 2 , 247–318 .
Kolodner, J. L. (2014). Retrieval and organizational strategies in conceptual memory
(PLE: Memory): a computer model . Psychology Press .
Kuo, T. C. (2010). Combination of case-based reasoning and analytical hierarchy pro- cess for providing intelligent decision support for product recycling strategies.
Expert Systems with Applications, 37 (8), 5558–5563. doi: 10.1016/j.eswa.2010.02. 057 .
Leake, D. , Kinley, A. , & Wilson, D. (1996). Linking adaptation and similarity learning. In Proceedings of the eighteenth annual conference of the cognitive science society . Lieber, J. (2002). Strong, fuzzy and smooth hierarchical classification for case-based
problem solving. In ECAI (pp. 81–85) .
Liu, K. F.-R., & Yu, C.-W. (2009). Integrating case-based and fuzzy reasoning to qual- itatively predict risk in an environmental impact assessment review. Environ-
mental Modelling & Software, 24 (10), 1241–1251. doi: 10.1016/j.envsoft.2009.04. 005 .
López-Arévalo, I., Bañares Alcántara, R., Aldea, A ., Rodríguez-Martínez, A ., & Jiménez, L. (2007). Generation of process alternatives using abstract models and case-based reasoning. Computers & Chemical Engineering, 31 (8), 902–918. doi: 10.1016/j.compchemeng.2006.08.011 .
Markman, A. B. , & Gentner, D. (1993). Structural alignment during similarity com- parisons. Cognitive Psychology, 25 (4), 431–467 .
McFee, B. , & Lanckriet, G. (2011). Learning multi-modal similarity. The Journal of Ma-
chine Learning Research, 12 , 491–523 .
Montani, S. (2011). How to use contextual knowledge in medical case-based reason- ing systems: a survey on very recent trends. Artificial Intelligence in Medicine,
51 (2), 125–131. doi: 10.1016/j.artmed.2010.09.004 .
Mougouie, B. , & Bergmann, R. (2002). Similarity assessment for generalizied cases by optimization methods. In S. Craw, & A. Preece (Eds.), Advances in case-based
reasoning number 2416 . In Lecture Notes in Computer Science (pp. 249–263). Springer Berlin Heidelberg .
Napoli, A. , Lieber, J. , & Curien, R. (1996). Classification-based problem-solving in case-based reasoning. In I. Smith, & B. Faltings (Eds.), Advances in case-based rea-
soning number 1168 . In Lecture Notes in Computer Science (pp. 295–308). Springer Berlin Heidelberg .
Nesme, M., & Hidalgo, P. (2013). PhiloSophie. definitions : ressemblance et analogie.
Perner, P. (2003). Incremental learning of retrieval knowledge in a case-based rea- soning system. In Case-based reasoning research and development (pp. 422–436). Springer .
Peschl, M. F. , & Riegler, A. (1999). Does representation need reality , 9–17 .
Reyes, E. R. , Negny, S. , Robles, G. C. , & Le Lann, J. (2015). Improvement of online adaptation knowledge acquisition and reuse in case-based reasoning: applica- tion to process engineering design. Engineering Applications of Artificial Intelli-
gence, 41 , 1–16 .
Richard, A. (1979). La resolution de problemes . Universite Laval. Ph.D. thesis. . Richter, M. M. (1993). Classification and learning of similarity measures . Springer . Rifqi, M. (2010). Mesures de similarité, raisonnement et modelisation de l’utilisateur . Sorlin, S. , & Solnon, C. (2005). Reactive tabu search for measuring graph similarity.
In Graph-based representations in pattern recognition (pp. 172–182). Springer . Wielemaker, J. (2014). Swi-prolog version 7 extensions. In Workshop on implemen-
tation of constraint and logic programming systems and logic-based methods in
programming environments 2014 (p. 109) .
Wu, Z., & Palmer, M. (1994). Verbs Semantics and Lexical Selection. In Proceedings of
the 32nd Annual Meeting on Association for Computational Linguistics . In ACL ’94 (pp. 133–138). Stroudsburg, PA, USA: Association for Computational Linguistics. doi: 10.3115/981732.981751 .
Xiong, N. (2011). Learning fuzzy rules for similarity assessment in case-based rea- soning. Expert Systems with Applications, 38 (9), 10780–10786. doi: 10.1016/j.eswa. 2011.01.151 .
Yang, C. J., & Chen, J. L. (2011). Accelerating preliminary eco-innovation design for products that integrates case-based reasoning and TRIZ method. Journal of
Cleaner Production, 19 (9), 998–1006. doi: 10.1016/j.jclepro.2011.01.014 . Zeid, I. , M. Gupta, S. , & Bardasz, T. (1997). A case-based reasoning approach to
