An integrated model for supplier selection process

Journal of Harbin Institute of Technology( New Series ), Vol. 10 ,No. 1 ,2003An integrated model for supplier selection processOboulhas Conrad Tsahat ONESIME ,XU Xiao-fei , ZHAN De-chen ,SHA Lei徐晓飞，占德臣， 沙磊( Dept. of Computer Science and Engineering , Harbin Institute of Technology , Harbin 150001 , China )Abstract : In today' s highly competitive manufacturing environment , the supplier selection process becomesone of crucial activities in supply chain management. In order to select the best supplier( s )it is not only nec-essary to continuously tracking and benchmarking performance of suppliers but also to make a tradeoff betweentangible and intangible factors some of which may confict. In this paper an integration of case-based reasoning( CBR ) ,analytical network process( ANP ) and linear programming( LP ) is proposed to solve the supplier se-lection problem. .Key words : supplier selection ; case-based reasoning ; ANP ; linear programmingCLC number : TP391Document code :AArticle ID : 1005-9113( 2003 )01 -0043-080 INTRODUCTIONBasically there are two kinds of supplier selectionproblemtI : single sourcing and multiple sourcing. AIn today' s highly competitive manufacturing envi-single source establishes closer contacts with the suppli-ronment , manufacturing companies must constantly asker ,in some cases these contacts extend as far as syn-themselves these tough questions :Do we have the bestchronizing their production delivery schedules to reducesuppliers at the lowest possible prices ? Are we gettinginventory. All these well-intentioned programs , howev-and sending materials as quickly as possible ?Addition-er ,are not without problems. One obvious problem isally , the battlefield is shifting from individual companythe erosion of supply base for the buyer. Relying orperformance to what we call Supply Chain. The supplyone supplier is risky , and often might not produce thechain represents all activities associated with the move-lowest costs for the product. Also , as Newman' 31 sug-ment and transformation of raw materials to finishedgests , single sourcing may lead to loss of technologicalgoods，from primary suppliers ，to assemblers andthrust , excess control , and lack of identity for the sup-through to the end user. Increasingly , managers viewplier. However ,a stream of research in the last decadethe operations and purchasing functions as intimatelyhas focused on how the use of multiple sources can rlinked parts of the supply chain , each with the abilityduce' effective' transit or lead-times that can maketo contribute strategically to the firm. Since the typicaldual or multiple sourcing policy an effective risk strate-manufacturing firm spends approximately 60% of eachgy and cost reducing one.sales dollar on purchased components , materials ancSeveral methods for supporting. the supplier selec-services from external suppliers ，the manufacturingtion process have been proposedt 4，5 .firm' s final products are significantly affected by theLuitzen de Boer , Eva Labro , Pierangela Morlac-performance of external suppliers in terms of cost ,chit61 proposed four steps for supporting the supplierquality and so on' . The purchasing function is largelyselection process :( 1 ) finding out exactly what we wantresponsible for ( 1 ) determining the characteristics ofto achieve by selecting a supplier ,( 2 )defining the cri-purchased materials , components , and services ,( 2 )teria ,( 3 ) pre-qualifying suitable suppliers ,( 4 ) mak-selecting suppliers capable of providing the required i-ing a final choice.tems at the requisite levels of quality and price , andIn this paper , we emphasize much attention in the( 3 ) managing the transaction so that the goods or serv-pre-qualifying of suitable suppliers phase and the finalices are delivered in a timely manner. The purcha-choice phase ，because these phases ( pre-qualifyingsing department ability to identify useful suppliers canand final choice ) often being the most visible phases incontribute an organization' s efficiency and effective-the;中国煤化工ality of these phases isness : cost reduction ，profitability and flexibilty oflargeYHCNMHGY of the steps prior tocompany.theseReceived 2002 -06 -07.Sponsored by the National Natural Science Foundatioin of Chin( Grant No. 60084004 ) and National 863 Research Item/CIMS Program( Grant No.2001AA414610 ,2001 AA4115002 ) and the Cross Century Fund for Person with Ability of Chinese Educational Department.Journal of Harbin Institute of Technology( New Series ), Vol. 10 ,No. 1 ,2003( 1 )Pre-qualification of suitable suppliers phasesionmakingsituations ( such as supplier selec-The pre-qualification phase is the process of reduc-tion ) 45 . Mathematical programming ( MP ) modelsing the set of all' suppliers to a smaller set of accepta-allows the decision- maker to formulate the decisionble suppliers. This process may be caried out in moreproblem in terms of a mathematical objective functionthan one step. However , the first step always consists ofthat subsequently needs to be maximized( e. g. maxi-defining and determining the set of acceptable suppliersmize profit ) or minimized( e. g. minimize costs ) bywhile possible subsequent steps serve to reduce thevarying the values of the variables in the objectivenumber of suppliers to consider. Basically therefore.function( e. g. the amount ordered with supplier X ).pre-qualification is sorting process rather than a rankingWeber and Current' 151 proposed the use of MP. Artifi-process. Several methods have been proposed in the lit-cial intelligence( AI )-based models are based on com-erature for the pre-qualification step but the most impor-puter-aided systems that in one way or another can betant of them are briefly discussed below.trained' by a purchasing expert or historic data. AI-Categorical methods are qualitative models , basedbino and Garavellr' 16 J present a decision support systemon historical data which assigns either good ( + )，based on Neural Networks.neural( 0 ) or unsatisfactory( - ) to each defined cri-teria for all suppliers and then a total rate for each sup-1AN INTEGRATED MODEL FOR SUPPLIERplier is calculated 77. Data envelopment analysisSELECTION( DEA ) allows for the simultaneous analysis of multipleinputs to multiple outputs , a multifactor productivityThe model presented in this article applies firstlyapproachCluster analysis( CA ) is a basic methodthe CBR to make a search for candidate suppliersfrom statistics and was applied by Holt 9. Case -based-based on the past contacts undertaken by them ; sec-reasoning( CBR ) systems : is a method for solvingondly the ANP , which uses pair wise comparison , toproblems by making use of previous , similar situationsmake the trade off between tangible and intangible fac-and reusing information and knowledge about such situ-tors and calculate a rating of suppliers , and then byations' 10]. The basic idea of CBR is to adapt solutionsapplying these ratings as coefficients of an objectivethat were used to solve old problems and use them tofunction in linear programming allocates order quanti-solve new problems and only few systems have been de-ties between the suppliers such that the total value ofveloped for purchasing decision-making. Ng et al].purchasing( TVP ) becomes a maximum. Therefore,Developed a CBR-system for the pre-qualification ofthe main steps of algorithm are discussed below.suppliers.1.1 Case-Based Reasoning( 2 )Decision models for the final choice-phaseCBR is a method for solving problems by makingThe vast majority of the decision models found ap-use of previous , similar situations and reusing informa-ply to the supplier choice phase of the buying process.tion and knowledge about such situations[ 10]CBR isSeveral methods have been proposed in the literatureadopted to sort-listing the candidate suppliers based onfor the supplier qualification step. The most importanttheir previous records and overall performance , capaci-of them are ( briefly reviewed ) discussed below.ty ,i. e. their total weighted score.We propose twoLinear weighting models : In linear weighting mod-case-bases : supplier case-base and bid case-base.els weights are given to the criteria , the biggest weightCBR has a large database in form of a supplier case-indicating the highest importance. The supplier withbase and bidding case-base. If the previous supplierthe highest overall rating can then be selected.performance history is systematic , it will be convertedNarasimhan T2 proposed the use of the analytic hierar-to domain-specific knowledge to form a knowledge-;hy process( AHP ) to deal with imprecision in suppli-based systems( KBS ). Then this knowledge is presen-er choice. J. Sarkis and R. P. Sundarra 13] proposedted in a cased-based form for the manipulation by thethe use of the analytical network process ( ANP ) , aCBR technique. CBR manage the selection of suppliermore sophisticated version of AHP. Total cost of own-basing on the previous performance of the suppliers.ership ( TCO ) models-based models attempt to includeThe supplier case-base contains the supplier case. Itall quantifiable costs in the supplier choice that are in-consists of its type , last updated dated , product pricecurred throughout the purchased item 's life cycle.( quantity and value per part ) , transportation cost , or-Zeger Degraeve and Filip Roodhooft 4 J developed arder process time , manu-mathematical programming model that uses total cost offactul中国煤化工time, target schedule .ownership information to simultaneously select suppli-comrYHCN M H G cost , on time deliveryers and determine order quantities over a multi-periodand order completeness. The type of supplier is classi-time horizon. Discrete Choice Analysis ( DCA-also .fied as the raw materials , supplier , and standard partsknown as choice-based conjoint analysis ) is an effec-of supplier. The whole supplier case is frequent upda-tive methodology for analyzing choices in complex deci-ted from the last bid case related to this supplier , such44.Journal of Harbin Institute of Technology( New Series ), Vol. 10 ,No. 1 ,2003that price and all other related costs can be updated.top level of the hierarchy. The next two levels consist ofThe on time delivery and the other completeness of thethe clusters( criteria ) and various components( sub cri-supplier can be computed by accumulating all historicalteria ) within each cluster for this goal and at the bottombid cases statistically. Some attributes in the technicallevel are the alternatives to be evaluated. The model de-capability , quality assessment and organization profilevelopment will require the delineation of attributes aof the supplier performance are updated by each bideach level and a definition of their relationships. In thisfrom the bid case-base. After retrieve the'potentialexample , the interdependence or feedback occurs be-suppliers' cases which most closely meet the specifica-tween different levels are observed. The summary oltions using the nearest neighbor numerical matchingclusters and components is shown in Table 2.function in case-based reasoning mechanism ( CBRM )( according to Kolodner) , they are assessed to eval-Table 1 Fundamental scale for pair wise comparisonsuate their suitability for a specific task. In this case ，Numericalassessment factors such as the category of technical ca-Verbal scalevaluespacity and quality from the case-base are to be used todifferentiate the capabilities of individual suppliers.Equally important , likely or preferredEach assessment factor carries certain weight ( W ) ,Moderately more important3which signifies the importance of the relevant factor.Finally , each potential supplier will have a Perform-Strongly more important5ance Score( PS ) from each category. The performanceVery strongly more importantscore are then multiplied by the weight assigned to at-tain a total weighted score.Extremely more important9Total Weighted Score= 2 PS,W;Intermediate values to reflect compromise 2 ,4 ,6 ,8To verify the competence level of the potentialTable 2 Summary of clusters and components for suppliersuppliers retrieved from the case-base library ，totalselectionweight score of them is compared with the ideal scoreClusterComponentsassigned by the user group. For example , the score is0.7 and if any of the suppliers have achieved a totalMaterial cosi( MC )score greater than this ideal score , are considered to beTotal cost( TC)Fixed ordering cost( FOC )Holding cost( HC )suitable. If there are no component suppliers , CBRwill advise the users to find news suppliers or level thePerformance( P)qualified threshold point. If there are potential suppli-Quality( Q)Conformance( C)ers , further analysis will be taken in the ANP.Reliability( R )1.2 ANP Analysis and Solution MethodologySpeed( S )Delivery( D)Analytical network process( ANP ) is a more gen-Reliability( Rd )eral form of the AHP ( Dr. Thomas Saaty-18] ). This aAbility to customize( AC )multi-attribute , decision making approach based on theFlexibility( F)New productreasoning , knowledge , experience , and perceptions ofIntroduction( NPI )experts in the field. Whereas , AHP models a decisionmaking framework using a unidirectional hierarchical re-Step 2 : Pair wise comparisons matrices of interde-lationship among decision levels , ANP allows for morependent component levels : The second step is the com-complex interrelationships among the decision levels andparison of various elements of network( i.e. , its clus-attributes. The interdependency among factors and lev-ters，components and alternatives ) using a nine-pointels of factors is defined as a systems with feedback ap-scale suggested by Saaty[19( see Table 1 ). This reproach. AHP does not contain feedback loops among thequires the analyst ( decision maker ) to make pair wisefactors , that can adjust weightings and lessen the possi-comparisons of elements at each level relative to eachbility of the reverse ranking phenomenon. The relativeactivity at the next higher level in the hierarchy. Inimportance or strength of the impacts on a given elementANP , like AHP , pairwise comparisons of the elementsis measured on a ratio scale similar to AHP( Table 1 ).in each levelrnndusted with respect to their rela-The details of ANP steps are given below.tive中国煤化工ntrol citerion. Compa-Step 1 : Model construction and problem structu-ringYCNMHGwhereiisassumedtobering :The first step is to structure a model into a hierar-at least as mportant asJ ) ,give a value ai.chical decision network as depicted in Fig. 1. This net-Of course , we seta; = 1. Furthermore ,if we setwork ,in our case , has four levels. The goal of deci-aj=hi,thenweseta;=-k'sion ,such as" Select the best supplier( sy" ,is at theJournal of Harbin Institute of Technology( New Series ), Vol. 10 ,No. 1 ,2003Select best supplierOverallohjertifCluslers(Criteria)Total CostQualityDeliveryFlexibilityMaterialAbility toFixer| PerformancecustormizeComponentsOrderingNew produet(suberiteria)_CoConformanceSpeedintoduetionHolding CostReliabilityuRQReliabilty(RD)(Innovation)Supplier nAteruativesSupplierSupplier iFig. 1 ANP networkA pair wise comparison matrix will be required fornumber of rows ( components ).each of the four major clusters for calculation of impactsAn example of the component pair wise compari-of each of the components. In addition , ten pair wiseson matrix within a total cost environment is presentedcomparison matrices will need to be determined for cal-in Table 3. The weighted priorities for this matrix isculation of the relative impacts of the criteria on a spe-shown as the last column in Table 3. The weighted pri-cific component. To fully describe these two-way rela-orities for each of cluster matrices ( four in all ) aretionships , 14 pair wise comparison matrices will be re-combined to create a matrix A with four columns andquired. Once the pair wise comparisons are completed ,ten rows( Table4 ).the local priority vector W( defined as the eVector in theAnother example of the cluster pair wise compari-example figures ) is computed as the unique solution to :son matrix within a material cost environment is shownAw=λmxWin Table 5. .where λ max is the largest eigenvalue of A. Saaty-20 pro-The weighted priorities for each of component ma-vides several algorithms for approximating w. In thistrices( eight in all ) are combined to create a matrix Bpaper a two-stage algorithm that involved forming a newwith ten columns and four rows( Table 6 ).n X n matrix by dividing each element in a column byStep 3 : Supermatrix formation : Each row and re-the sum of the column elements and then summing thespective column of the supermatrix represents an ele-elements in each row of the resultant matrix and divid-ment of the decision network. The supermatrix allows aing by the n elements in the row. This is referred to asresolution of the effects of interdependence that existsthe process of averaging over normalized columns. Thisbetween the elements of the ANP network. The suis represented as :permatrix is then composed of the weight vectors deter-min中国煤化工fciliate convergenceSaatof the interdependenceaj)of eaTYHCN MH Gus anI LI x| Ll identi-W，=-Jty matrix is included in the supermatrix. The two com-piled matrices A and B , are combined to form the su-where w; is the weighted priority for component i ;J ispermatrix M shown in Table 7.index number of columns ( components );I is index46.Journal of Harbin Institute of Technology( New Series ), Vol. 10 ,No. 1 ,2003Table 3 Components pair wise comparison matrix for total cost environment and eigenvectorTotal costMCFOC HCCRRd NPIACeVector20.127FOC1/20.108HC1/2 1/1/3 1/3 1/30. 077P0. 2000. 1720.144S0. 056Rd0. 049NPI资资资0.0361/4 1/4 1/31/41/41/41/21/21/20.030Table 4 A matrix formed from eigenvectors ( impactweight ) for clusters implication on componentsA MatrixrCQTable 5 Clusters pair wise comparison matrix for materialcost environment and eigenvectorM0.0930. 102Material cost TCDeV ector0. 0820. 095H0.0770.0670.037TC10.3210.2000. 2270. 1750.19230.450.1720. 1960. 1930. 1660. 1280.1451/30.1420.0560. 0650.0490. 0460. 0840. 057F1/40. 0870.0330.04A(0. 0300.0290.035Table 6 B matrix formed from eigenvector ( relative importance weights ) for components implications on clustersMatrix4C0.3030.3430.2140.220. 2390.2830. 2960.2850.4560. 4460. 5650.5480.5250.4330.4260.4150. 440.1110.1370.140.2390.130.0870.0940.1010.0830.0920. 0930.0890.088.0.1590.119Table 7 Supermatrix M compiled from matrix A and BSuper-s00.3210.303 0.343 0.214 0.22 0.239 0.239 0.283 0. 2960. 2850.45 0.456 0. 4460.565 0. 5480.525 0. 4330.426 0. 4150.44 .0.1420. 146 0.111 0. 1370. 140.143 0.239 0.203 0. 130.1560.087 0.094 0. 101 0. 0830.092 0.093 0.089 0.088 0.159 0. 1190.1270.093 0. 1020. 1080.1080.082 0.077 0. 0950.077 0.067 0.037 0. 0820.200 0. 2270.175 0. 1920.1720. 1960.193 0. 1660.144 0.172 0. 1280. 145中国煤化工0.056 0. 056 0.128 0. 065?d0.049 0.046 0. 0840.05MHCNMHG0.0360.033 0.04 0. 0490.030 0.029 0.035 0. 042Journal of Harbin Institute of Technology( New Series), Vol. 10 , No.1 ,2003Step 4 : Convergence of supermatrix : Here we firstuct and its order quantity( X; ) should be equal or lessnormalize the supermatrix-divide each element by itsthan its capacity , these constraints are :X; ≤V; ,i =column sum一so that each column adds to one ( mak-1 2 p..几. V; is capacity of ith supplier.ing the supermatrix" column stochastic" ). The nor-On the other hand , aggregate suppliers' capacitymalized supermatrix is then raised to a significantlyshould be equal or greater than demand ，therefore ,large power. This execution provides the converged( orstable ) weights of the elements of network on one an-2 V;≥D.other. The number of interest in resulting supermatrix( 4 )Quality constraintare weights of the alternatives on the decision the lastAsQ is the buyer' s maximum acceptable defect| LI x| LI rows of the first column.rate andq; is defect rate of the ith vendor , the qualityStep 5 : Alternative evaluations : Each alternativewill need to be evaluated on each of the components.constraint can be shown as: 2 X;q;≤QD.Since there are 10 components , an additional 10 n X nWe propose a model solution algorithm Fig. 2.pairwise comparison matrices ( where n is the number ofalternatives ) will be needed for evaluation. .Step 6 : Selection of best alternative : The selec-2 CONCLUSION .tion of the best alternative depends on the calculationThe model aims at helping a decision-maker withof the' desirability index' for an alternative i( D; ).useful information and experiences from similar , previ-The equation for D; is defined by : Dous decision situations , for selecting suppliers. By in-2 EPAwSw ,where P, is the relative importancecluding CBRM for captures the knowledge of potentialsuppliers , which performs analysis to find matchedweight of clustersj ,Agis the relative importance weightcandidates and provides information on projected per-for components ki of clustersj ,S海j is the relative impactformance of them through a pair wise comparison usingof alternativei on component k of clusterj ,K; is the in-ANP , more realistic weights for the performance meas-dex set of components for clusterj ,J is the index set ofures can be determined.clusters.References :1.3Linear ProgrammingThe field of mathematical programming arose from[1 ] BURT D N. Managing product quality through strategicthe study of linear programming. In Dantzig' s seminalpurchasing[ J ] Sloan Management Review , 1989 , 30( 1 ):39 -47.textbookn , the concept of linear programming is in-[2] GHODSYPOURS H ,0' BRIEN C. A decision supporttroduced by describing a few different planning prob-system for supplier selection using an integrated analyticallems. If there are some constraints such as suppliers'hierarchy process and linear programming[ J ] Interna-capacity , quality , etc. use the suppliers' ratings astional Journal of Production Economics ,1998( 56-57 ):coefficients of an objective function in linear program-199 -212.ming to assign order quantities to the suppliers such the[3] NEWMAN R G. Single sourcing : short term saving ver-total value of purchasing ( TVP ) becomes maxi-sus long-term problems[ J ] Journal of Purchasing andmum. The objective function and constraints of thisMaterials Management , 1988 ,25( 2 ):20 -25.[4] ANSARI A , MODARRESS B. Just-in-time purchasing :linear programming are as follows.problems and solution[ J] J Purch Mater Mgmt , 1986 ,( 1 )Objective function22 :11 -15.As R; and X; ,respectively , denote the ratings and[5 ] MCFADDEN D. The choice theory apprach to market re-the numbers of purchased units from the ith suppliersearcH[ J] MKTG Sci ,1986 ,5 :275 - 297.and maximizing the total value of purchasing is de-[6] DE LUTTZEN B ,EVA L ,PIERANGELA M. A review ofsired，the objective function is Max( TVP) =methods supporting supplier selection[ J ]. EuropeanJournal of Purchasing & Supply Management 2001 7 752 RX; ,R, is final ratings of ith supplier ,X; is order- 89.[7] TIMMERMAN E. An approach to vendor performance e-quantity for ith supplier ,X;≥0 ,i=1 2...几。valuatior[ J ] Journal of Purchasing and Supply Manage-( 2 )Demand constraintment ,1986 1 :27 - 32.As sum of assigned order quantities to n vendors[8] WFBEB CA : FLL.RAM I. M. Supplier selection usingshould meet the buyer' s demand , it can be stated中国煤化工: a decision support systemMHCNMH(_ournal of Physical Distribu-that: 2 X; = D ,D is demand for the period.,1992 ,23(2):3-14.[9 ] HOLT G D. Which contractor selection methodology ?( 3 )Capacity constraint[J ] International Joural of Project Management , 1998 ,As vendor i can provide up to V; units of the prod-16(3 ):153 - 164.Journal of Harbin Institute of Technology( New Series ), Vol. 10 ,No. 1 ,2003MRP(Material RequirementPlanning)Translate intoIndexing rulesQueryRetrieve similarsupplier CaseSupplier caselibraryAre there anyNoRefine cascspotentialsupplierShort-listingcandidateRetainsuppliersModel constructionand problemstrucruringCulculate relativeimportance of thefactors(" weight")Form1 a supermatrixfrom the weightsForm a converge| supermatrix from thesupermatrixAltemative evadutionsDetermine the finalraling of supplierLChoose the maximumscore supplier and stopconstraints?Yes3Buikd the lincar中国煤化工| programming model |MYHCNMHGSeleet the bestuppliersFig.2 Flowchart of model solutionJournal of Harbin Institute of Technology( New Series ), Vol. 10 ,No. 1 ,2003[ 10] AAMODT A ,PLAZA E. Case-based reasoning :founda-proach to vendor selectior[ J ] European Jourmal of Op-tional issues , methodological variations and system aperational Researeh ,1993 ， 68 :173- 184.proaches[ J ] AICom-Artifcial Intelligence Communica-[ 16] ALBINO V ,CARAVELLI A C. A neural network appli-tions ,1994 ,7( 1 ):39 -59.cation to subcontractor rating in construction firms[ J ][11]NGsT,SMITHNJ,SKITMORERM.Acase-base.Intermational Journal of Project Management , 1998 , 16 .reasoning model for contractor pre qualificatior[ J ] In-( 1 ):9-14.ternational Journal of Construction Information Technolo-[ 17 ] KOLODNER J. Case-based Reasoning[ M ] San Ma-gy ,1998 ,6 :47 -61.teo : Morgan Kaufman 1993.[12] NARASIMHAN R. An analytical approach to supplier[ 18 ] SAATY T L. Axiomatic foundation of the analytic hierar-selectior[ J ] J Purchasing Mater Mgmt , 1983 :27 -chy proces[ J ] Management Science , 198632( 7 ) :1032.[ 13 ] JOSEPH S , SUNDARRAJ R P. Hub location at digital[ 19 ] SAATY T L. The Analytic Hierarchy Process : Plan-equipment corporation :comprehensive analysis oning , Priority Setting , Resouree Allocation[ M ] Newqualitative and quantitative factors[ J ] European Jour-York : MeGraw-Hill , 1980.nal of Operational Research ,2002 ,137 336 - 347.[20] SAATY T L, TAKIZAWZ M. Dependence/ independ-[ 14 ] DEGRAEVE Z , ROODH0OFT F. Improving the ffi-ence : from linear hierarchies to nonlinear network:[ J ]ciency of the purchasing process using total cost of own-European Jourmal of Operational Research , 1986 ,26 :ership information : The case of heating electrodes at229 - 237.Cockerill Sambre S. A[ J ] European Journal of Opera-[21 ] DANTZIG G B. Linear Programming and Extensionstional Research , 1999 ,112 :42 -53.[ M ] Princeton : Princeton University Press , 1963[ 15] WEBER C A , CURRENT J R. A multi-objective ap-中国煤化工MHCNMHG