Large production line optimisation using simulated annealing (cid:1)y DiomidisSpinellis z ChrissoleonPapadopoulos x J.MacGregorSmith (cid:1) November20,2000 Abstract 1 Introduction and literature review Alargeamountofresearchhasbeendevotedtotheanaly- sis and design of production lines. A lot of this research Wepresentarobustgeneralisedqueuingnetworkalgorithm concerns the design of manufacturing systems with con- as an evaluative procedure for optimising production line siderableinherentvariabilityintheprocessingtimesatthe configurationsusingsimulatedannealing. Wecomparethe various stations, a common situation with human opera- resultsobtainedwithouralgorithmtothoseofotherstudies tors/assemblers. The literature on the modellingand opti- and findsome interestingsimilarities but also strikingdif- misationof productionlines is vast, allowingus toreview ferencesbetweenthemintheallocationofbuffers,numbers only the most directly relevant studies. For a systematic ofservers,andtheirservicerates.Whilecontextdependent, classification ofthe relevantworkson thestochastic mod- thesepatternsofallocationareoneofthemostimportantin- elling of these and other types of manufacturing systems sightswhichemergeinsolvingverylongproductionlines. (e.g.,transferlines,flexiblemanufacturingsystems(FMS), The patterns, however, are often counter-intuitive, which andflexibleassemblysystems(FAS)),theinterestedreader underscoresthedifficultyoftheproblemwe address. The isreferredtoareviewpaperbyPapadopoulosandHeavey mostinterestingfeatureofouroptimisationprocedureisits (1996) and some recently published books, such as those boundedexecutiontime,whichmakesitviableforoptimis- byAskinandStandridge(1993),BuzacottandShanthiku- ingverylongproductionlineconfigurations. Basedonthe mar(1993),Gershwin(1994),Papadopoulosetal. (1993), boundedexecutiontime property,we have optimisedcon- ViswanadhamandNarahari(1992),andAltiok(1997). figurationsofupto60stationswith120buffersandservers Therearetwobasicproblemclasses: inlessthanfivehoursofCPUtime. 1. the evaluation of production line performance mea- Keywords:BufferAllocation,Non-linear,Stochastic,In- suressuchasthroughput,meanflowtime,workstation teger,NetworkDesign,SimulatedAnnealing meanqueuelength,andsystemutilisation,and 2. the optimisation of the decision variables of these lines. InternationalJournalofProductionResearch,38(3):509–541,Febru- Examples of decision variables that have been considered (cid:1) ary2000. are: Thisisamachine-readablerenderingofaworkingpaperdraftthatled y toapublication. Thepublicationshouldalwaysbecitedinpreferenceto 1. the sizes of the buffers placed between successive this draft using the reference inthe previous footnote. Thismaterial is workstationsofthelines, presentedtoensuretimelydisseminationofscholarlyandtechnicalwork. Copyrightandallrightsthereinareretainedbyauthorsorbyothercopy- 2. the number of servers allocated to each workstation, rightholders.Allpersonscopyingthisinformationareexpectedtoadhere and tothetermsandconstraintsinvokedbyeachauthor’scopyright. Inmost cases,theseworksmaynotberepostedwithouttheexplicitpermissionof 3. theamountofworkloadallocatedtoeachworkstation. thecopyrightholder. DepartmentofInformationandCommunicationSystems,University z The corresponding optimisation problems are named, re- oftheAegean,83200KarlovasiSamos,Greece. (Correspondingauthor); spectively,(1)thebufferallocationproblem,(2)theserver contact address: Myrsinis 1, GR-145 62Kifisia, GREECE.Tel. +301 2719710,Fax+3012719712,email:[email protected] allocationproblem, and(3) the workloadallocationprob- Department of Business Administration, University of the Aegean, lem,inaproductionline. x 82100Chios,Greece.email:[email protected] InPapadopoulosetal. (1993)bothevaluativeandgener- Department of Mechanical and Industrial Engineering, Univer- (cid:2) ative(optimisation)modelsaregivenformodellingthevar- sity of Massachusetts, Amherst Massachusetts 01003, USA. email: [email protected] ious typesof manufacturingsystems. This workfalls into 1 the second category. Evaluative and optimisation models Yamazakietal.,1992),amongothers,andtheserveralloca- canbecombinedbyclosingtheloopbetweenthem;thatis, tionproblem(MagazineandStecke, 1996, Hillier andSo, onecan use feedbackfroman evaluativemodelto modify 1989). HillierandSo(1995)studiedvariouscombinations thedecisionstakenbytheoptimisationmodel. of these three design problems. Other references may be Oneofthekeyquestionsthatthedesignersfaceinase- foundtherein(BuzacottandShanthikumar,1992,Buzacott rialproductionlineisthebufferallocationproblem(BAP), andShanthikumar,1993),amongothers. i.e., how muchbufferstorage to allow andwhere to place The present work deals with the same design problems it within the line. This is an important question because (buffer allocation, server allocation, and workload alloca- bufferscanhaveagreatimpactontheefficiencyofthepro- tion)butforlongproductionlineswithmulti-machinesta- duction line. They compensate for the blocking and the tions. starving of the line’s stations. For this reason, the buffer As the problem being investigated is combinatorial in allocation problem has received a lot more attention than nature, traditional OperationsResearch techniquesare not theothertwodesignproblems. Bufferstorageisexpensive as practical for obtaining optimal solutions for long pro- duebothtoitsdirectcost,andtotheincreaseofthework- ductionlines. We proposea simulatedannealing(SA)ap- in-process(WIP)inventories. Inaddition,therequirement proachasthesearchmethodinconjunctionwiththeexpan- tolimitthebufferstoragecanalsobearesultofspacelim- sionmethoddevelopedbyKerbacheandMacGregorSmith itationsintheshopfloor. TheliteratureontheBAP is ex- (1987) as the evaluative tool. Simulated annealing is an tensive. Asystematicclassificationoftheresearchworkin adaptationofthesimulationofphysicalthermodynamican- thisareaisgiveninSinghandMacGregorSmith(1997)and nealingprinciplesdescribedbyMetropolisetal. (1953)to Papadopoulosetal. (1993). Theworksaresplitaccording thecombinatorialoptimisationproblems(Kirkpatricketal., to: 1983,Cerny,1985).Similartogeneticalgorithms(Holland, 1975,Goldberg,1989)andtabusearchtechniques(Glover, Themethodusedtosolvethebufferallocationproblem 1990)itfollowsthe‘localimprovement’paradigmforhar- searchmethods,(AltiokandStidham,1983,Smithand nessingtheexponentialcomplexityofthesolutionspace. Daskalaki, 1988, Seong et al., 1995, Hillier and So, 1991a,HillierandSo,1991b),dynamicprogramming Thealgorithmisbasedonrandomisationtechniques.An methods, (Kubat and Sumita, 1985, Jafari and Shan- overview of algorithms based on such techniques can be found in the survey by Gupta et al. (1994). A complete thikumar,1989,YamashitaandAltiok,1997),among others,andsimulationmethods,(Conwayetal.,1988, presentationofthemethodanditsapplicationsisdescribed Hoetal.,1979). by Van Laarhoven and Aarts (1987) and accessible algo- rithmsforitsimplementationarepresentedbyCoranaetal. Thetypeofproductionline balanced/unbalanced,(Pow- (1987)andPressetal. (1988). Acriticalevaluationofdif- ell(1994)presentsaliteraturereviewaccordingtothis ferentapproachestoannealingschedulesandothermethod scheme), or reliable/unreliable; the majority of pa- optimisationsaregivenbyIngber(1993). Asatoolforop- pers deal with reliable lines. Only a few algorithms erationalresearchSAispresentedbyEglese(1990),while have been developed to calculate the performance Koulamas et al. (1994) provide a complete survey of SA measures of unreliable productionlines (Glassey and applicationstooperationsresearchproblems. Hong,1983,ChoongandGershwin,1987,Gershwin, The use of the Simulated Annealing algorithm appears 1989, Heavey et al., 1993). Hillier and So (1991a) tobeapromisingapproach. Webelievethatthisalgorithm and Seong, Chang, and Hong (1995)have dealt with couldbeappliedin conjunctionwith a fast decomposition thebufferallocationprobleminunreliableproduction algorithmtosolveefficientlyandaccuratelytheaforemen- lines. tioned optimisation problems in much longer production lines. Thenumberofmachinesavailableateachworkstation The remainderof the paperis organisedas follows: we Similarly, few researchers have dealt with the buffer firstdescribetheproductionlinemodelandtheproblemof allocation problem in production lines with multiple ourinterest followedby the methodologyof ourapproach (parallel) machines at each workstation (Hillier and namely: the performance model, the expansion method So, 1995, Magazine and Stecke, 1996, Singh and used for evaluating the line performance, an overview of Smith,1997). thecombinatorialoptimisationmethods,thesimulatedan- Apart fromthe bufferallocationproblem,the othertwo nealingoptimisationmethod,andthecompleteenumeration interesting design problems have also been considered by method; we then describe our experimental methodology someresearchers,e.g.,theworkallocationproblem(Hillier andpresentanoverviewofthenumericalandperformance andBoling,1979,HillierandBoling,1966,HillierandBol- resultsforshortandlongproductionlines.IntheAppendix ing, 1977, Ding and Greenberg, 1991, Huang and Weiss, following the concluding section, we provide a full tabu- 1990,Shanthikumaretal.,1991,w.WanandWolff,1993, latedsetoftheexperimentalresults. 2 andinteger , M1 B2 M2 B3 … BN M(cid:1) qi (cid:2)(cid:6) (cid:4)i(cid:2)(cid:3)(cid:1)(cid:7)(cid:1)(cid:2)(cid:2)(cid:2)(cid:1)N(cid:5) andinteger , si si (cid:2)(cid:1) (cid:4)i(cid:2)(cid:1)(cid:1)(cid:3)(cid:1)(cid:2)(cid:2)(cid:2)(cid:1)N(cid:5) Mi : Station i consisting of si servers , wi (cid:3)(cid:6)(cid:4)i(cid:2)(cid:1)(cid:1)(cid:3)(cid:1)(cid:2)(cid:2)(cid:2)(cid:1)N(cid:5) where and are fixed constants and , and are Bi : Buffer i thedecQisi(cid:1)oSnvectoNrs. Thethirdconstraintindqicsatesthwatthe sumoftheexpectedservicetimesisafixedconstant,which bynormalisationcanbeequalto . Figure1:An -workstationmulti-machineproductionline The objectivefunction of throNughput, , is not with inNtermediatebuffers theonlyperformancemeasureofinterest.RTh(cid:4)eq(cid:1)asv(cid:1)ewra(cid:5)geWIP, N (cid:1)(cid:1) the flow time, the cycle time, the system utilisation, the 2 The production line model and the averagequeuelengths andothermeasures are equallyim- portantperformancemeasures. However,throughputisthe optimal design problem mostcommonlyusedperformancemeasureintheinterna- tionalliterature. Wedefineanasynchronouslineasoneinwhicheverywork- station can pass parts on when its processing is complete as longas bufferspace is available. Such a line is subject 3 Optimal allocation methodology to manufacturing starving and blocking. We assume that thefirststationisneverstarvedandthelaststationisnever 3.1 Performance models blocked. Therefore we can say that the line operates in a The queuing model that we use refers to a pushmode: partsarealwaysavailablewhenneededatthe queuingsystemwhereM: (cid:4)M(cid:4)C(cid:4)K first workstation and space is always available at the last workstationtodisposeofcompletedparts. itisassumedthatthearrivalsaredistributedaccording An -station line consists of workstations in series, (cid:3) tothePoissondistribution(orequivalentlythatthein- labelleNd and N locationsforbuffers, termediate times between two successive arrivals are labelledM(cid:1)(cid:1)M(cid:2)(cid:1)(cid:2)(cid:2)(cid:2)(cid:1)MN, is illNust(cid:1)ra(cid:1)ted in figure 1. Each exponentiallydistributed), station Bha(cid:2)s(cid:1)B(cid:3)(cid:1)s(cid:2)e(cid:2)r(cid:2)v(cid:1)eBrsNoperating in parallel. The buffer capacitiies of sthie intermediate buffers , , the service (processing)times follow the exponential are denoted by , whereas the mean sBerivicie(cid:2)tim(cid:3)e(cid:1)s(cid:2)(cid:2)o(cid:2)f(cid:1)tNhe (cid:3) distribution, stations( qi )aredenotedby . i The maiin(cid:2)(cid:1)p(cid:1)e(cid:2)rf(cid:2)o(cid:2)r(cid:1)mNance measure owf ithe production thereare parallelservers(machineswhichareiden- line is the mean throughput, denoted by , (cid:3) ticalateaCchworkstation),and where , R(cid:4)aqn(cid:1)sd(cid:1)w(cid:5) the total capacity of the system is finite and equal to q(cid:2) (cid:4)q(cid:2)(cid:1)q(cid:3)(cid:1).(cid:2)(cid:2)(cid:2)(cid:1)qN(cid:5) s(cid:2) (cid:4)s(cid:1)(cid:1)s(cid:2)(cid:1)(cid:2)(cid:2)(cid:2)(cid:1)sN(cid:5) w(cid:2) (cid:3) . (cid:4)wI(cid:1)f(cid:1)w(cid:2)d(cid:1)e(cid:2)n(cid:2)o(cid:2)t(cid:1)ewsNth(cid:5)etotalnumberofavailablebufferslotsto K beallQocatedtothe buffersand thetotalnumberof While our focus in this paper is on ap- availableservers(mNac(cid:1)hi(cid:1)nes)to allocaStedto the stations proximationsfor openqueuingnetworksMof(cid:4)sMeri(cid:4)eCs-(cid:4)pKarallel thenthegeneralversionoftheoptimisationmodNel(firstre- topologies, we also briefly discuss some of the available portedbyHillierandSo,1995)is: approachesused for modelling systems since mostoftheliteraturehasfocusedMo(cid:4)nM(cid:4)(cid:1)(cid:4)K systems. (1) BothopenandclosedsystemshavMeb(cid:4)eMen(cid:4)(cid:1)st(cid:4)uKdiedbyex- maxR(cid:4)q(cid:1)s(cid:1)w(cid:5) subjectto: actanalysisalthoughresultshavebeenlimited.Exactanal- ysesofopentwo,three,andfournode-servermodelswith exponentialservicearelimited bytheexplosivegrowthof N (2) theMarkovchainmodelsforanalysingthesesystems. The analysisofverylargeMarkovchainmodelshasledtoeffec- qi (cid:2)Q(cid:1) i(cid:4)(cid:2) tive aggregation techniques for these models (Schweitzer XN (3) and Altiok, 1989, Takahashi, 1989) but the computation si (cid:2)S(cid:1) timeandpowerrequiredfortheseexactresultsleavesopen i(cid:4)(cid:1) theneedforapproximationtechniques. X N (4) Van Dijk and his co-authors (1988, 1989) have devel- wi (cid:2)N(cid:1) oped some bounding methodologies for both i(cid:4)(cid:1) M(cid:4)M(cid:4)(cid:1)(cid:4)K X 3 and systemsandhavedemonstratedtheiruse- fulneMss(cid:4)iMnt(cid:4)hCe(cid:4)dKesignofsmallqueuingnetworks.Ofcourse, M/M/C/K M/M/C/K whendoingoptimisationofmediumandlongqueuingnet- Λ i j R works, bounds can be far off the optimum, so robust ap- proximation techniques close to the optimal performance measuresaremostdesirable. Most approximation techniques appearing in the litera- P'K turerelyondecomposition/aggregationmethodstoapprox- imateperformancemeasures. Oneandtwonodedecompo- M/M/C/K M/M/∞ M/M/C/K sitionsofthenetworkhavebeencarriedout,allwithvary- ingdegreeofsuccess. Λ i h j R PK Thefewapproximationapproachesavailableintheliter- ature can be classified as follows: Isolation methods, Re- peated Trials, Node-by-node decomposition, and Expan- sionmethods. IntheIsolationmethod,thenetworkissub- 1 - PK divided into smaller subnetworks and then studied in iso- lation(LabetoulleandPujolle,1980,BoxmaandKonheim, Figure2:TypeIBlockinginFiniteQueues 1981).ThismethodwasusedbyKuehn(1979)andGelenbe andPujolle(1976)buttheyfailedtoconsidernetworkswith finitecapacity. Smith,1997). Closely related to the Isolation method is the Repeated Trials Method, a class of techniques based upon repeat- 3.2 Expansionmethod edlyattemptingtosendblockedcustomerstoaqueuecaus- ingtheblocking(CaseauandPujolle,1979,Fredericksand TheExpansionMethodisarobustandeffectiveapproxima- Reisner,1979,Fredericks,1980). tion technique developed by Kerbache and Smith (1987). In Node-by-nodedecomposition, the network is broken As described in previous papers, this method is charac- down into single, pairs, and triplets of nodes with aug- terised as a combinationof Repeated Trials and Node-by- mentedserviceandarrivalparameterswhicharethenstud- node Decomposition solution procedures. Methodologies ied separately (Hillier and Boling, 1967, Takahashi et al., for computing performance measures for a finite queuing 1980, Altiok, 1982, Altiok and Perros, 1986, Perros and networkuseprimarilythefollowingtwokindsofblocking: Altiok, 1986, Brandwajn and Jow, 1988). More general service time approximations appear in the work by Gun TypeI: The upstream node gets blocked if the service and Makowski (1989). The Expansion method is the ap- onacustomeriscompletiedbutitcannotmovedown- proach argued for in this paper for computing the perfor- streamduetothequeueatthedownstreamnode be- mance measures of finite queuing networks ingfull.ThisissometimesreferredtoasBlockingjAf- (Kerbache,1984,KerMba(cid:4)cMhe(cid:4)aCnd(cid:4)KSmith,1987,Kerbacheand terService(BAS)(Onvural,1990). Smith, 1988). It can be characterised conceptually as a combinationofRepeatedTrialsandNode-by-NodeDecom- TypeII: The upstream node is blocked when the down- positionwherethekeydifferenceisthata‘holding’nodeis stream node becomes saturated and service must be added to the network to register blocked customers. The suspended on the upstream customer regardless of addition of the holdingnode ‘expands’the network. This whetherserviceiscompletedornot.Thisissometimes approach transforms the queuing network into an equiva- referredtoas BlockingBeforeService(BBS) (Onvu- lent Jackson network which is then decomposed allowing ral,1990). for each node to be solved independently. We have suc- cessfullyusedtheExpansionMethodtomodel The Expansion Method uses Type I blocking, which (KerbacheandSmith,1988), (JainMa(cid:4)nMdS(cid:4)m(cid:1)(cid:4)itKh, is prevalent in most production and manufacturing, trans- 1994, Han and Smith, 1992)M, (cid:4)M(cid:4)C(cid:4)K (Kerbache and portationandothersimilarsystems. Smith, 1987), and most recentMly(cid:4)G(cid:4)(cid:1)(cid:4)K (Cheah and Considerasinglenodewithfinitecapacity (including Smith,1994,Smith,1991,Smith,M19(cid:4)9G4)(cid:4)qCu(cid:4)eCuesandqueu- service).ThisnodeessentiallyoscillatesbetweKentwostates ingnetworks.Inaddition,wehavealsousedourExpansion — the saturated phase and the unsaturated phase. In the Methodologytomodelrouting(DaskalakiandSmith,1989, unsaturatedphase,node hasatmost customers(in Daskalaki and Smith, 1986, Gosavi and Smith, 1990)and serviceorinthequeue).jOntheotherhKan(cid:1)d,(cid:1)whenthenode optimalresourceallocationproblems(SmithandDaskalaki, issaturatednomorecustomerscanjointhequeue.Referto 1988, Smith, 1991, Smith and Chikhale, 1995, Singh and figure2foragraphicalrepresentationofthetwoscenarios. 4 The Expansion Method consists of the following three 3.2.2 StageII:parameterestimation stages: Thisstageessentiallyestimatestheparameters , and (cid:1) StageI:NetworkReconfiguration. utilisingknownresultsforthe pmKodpelK. (cid:3) (cid:6)h M(cid:4)M(cid:4)C(cid:4)K StageII:ParameterEstimation. : Analytical results from the model (cid:3) (cid:3) pprKovidethefollowingexpressionfMor(cid:4)M:(cid:4)C(cid:4)K StageIII:FeedbackElimination. pK (cid:3) The following notation defined by Kerbache and K (5) Smith (1987, 1988)shall be used in furtherdiscussion re- (cid:1) (cid:5) gardingthismethodology: pK (cid:2) cK(cid:2)cc(cid:11) (cid:6) p(cid:5) wherefor (cid:1) (cid:2) (cid:4)(cid:5)(cid:4)c(cid:6)(cid:5)(cid:2)(cid:1)(cid:5) The holding node established in the Expansion h (cid:8)m(cid:2)ethod. c(cid:2)(cid:1) n c K(cid:2)c(cid:6)(cid:1) (cid:2)(cid:1) :=ExternalPoissonarrivalratetothenetwork. (cid:1) (cid:5) (cid:4)(cid:5)(cid:4)(cid:6)(cid:5) (cid:1)(cid:1)(cid:4)(cid:5)(cid:4)c(cid:6)(cid:5) (cid:9) p(cid:5) (cid:2) (cid:12) :=Poissonarrivalratetonode . (cid:3)n(cid:4)(cid:5)n(cid:11)(cid:1)(cid:6)(cid:2) c(cid:11) (cid:1)(cid:1)(cid:5)(cid:4)c(cid:6) (cid:4) (6) X (cid:5)j j andfor , :=Effectivearrivalratetonode . (cid:4)(cid:5)(cid:4)c(cid:6)(cid:2)(cid:1)(cid:5) (cid:5)(cid:10)j j :=Exponentialmeanservicerateatnode . (cid:6)j j c(cid:2)(cid:1) n c (cid:2)(cid:1) (7) (cid:1) (cid:5) (cid:4)(cid:5)(cid:4)(cid:6)(cid:5) :=Effectiveservicerateatnode duetoblocking. p(cid:5) (cid:2) (cid:12) (cid:4)K(cid:1)c(cid:12)(cid:1)(cid:5) (cid:6)(cid:10)j j (cid:3)n(cid:4)(cid:5)n(cid:11)(cid:1)(cid:6)(cid:2) c(cid:11) (cid:4) :=Blockingprobabilityoffinitequeueofsize . : SXince there is no closed form solution for this (cid:1) pK K (cid:3) qpuKantityanapproximationisusedgivenbyLabetoulle :=Feedback blocking probability in the Expansion (cid:1) and Pujolle obtained using diffusion techniques (La- pK method. betoulleandPujolle,1980). :=Unconditional probability that there is no customer j p(cid:5) intheservicechannelatnode (eitherbeingservedor beingheldafterservice). j K K K(cid:2)(cid:1) K(cid:2)(cid:1) (cid:2)(cid:1) :=Meanthroughputrate. (cid:1) (cid:6)j (cid:12)(cid:6)h (cid:5)(cid:13)(cid:4)r(cid:2) (cid:1)r(cid:1) (cid:5)(cid:1)(cid:4)r(cid:2) (cid:1)r(cid:1) (cid:5)(cid:14) R pK (cid:2)(cid:3) (cid:6)h (cid:1) (cid:6)h(cid:13)(cid:4)r(cid:2)K(cid:6)(cid:1)(cid:1)r(cid:1)K(cid:6)(cid:1)(cid:5)(cid:1)(cid:4)r(cid:2)K (cid:1)r(cid:1)K(cid:5)(cid:14)(cid:4)(8) :=Servicecapacity(buffer)atnode ,i.e. fora where and aretherootstothepolynomial: Sj singlequeue. j S (cid:2)K r(cid:1) r(cid:2) (9) (cid:2) 3.2.1 StageI:networkreconfiguration (cid:5)(cid:1)(cid:4)(cid:5)(cid:12)(cid:6)h(cid:12)(cid:6)j(cid:5)x(cid:12)(cid:6)hx (cid:2)(cid:6) while, and and are the (cid:1) Usingtheconceptoftwophasesatnode ,anartificialnode actualarri(cid:5)val(cid:2)rat(cid:5)ejst(cid:1)ot(cid:5)hhe(cid:4)fi(cid:1)n(cid:1)itepaKn(cid:5)dartifi(cid:5)cijalhold(cid:5)ihngnodes is added for each finite node in the njetwork to register respectively. bhlocked customers. Figure 2 shows the additional delay, Infact, thearrivalratetothefinitenodeisgivenby: causedtocustomerstryingtojointhequeueatnode when (cid:5)j it is full,withprobability . Thecustomerssuccejssfully (10) join queue with a probabpiKlity . Introductionof (cid:5)j (cid:2)(cid:5)(cid:10)i(cid:4)(cid:1)(cid:1)pK(cid:5)(cid:2)(cid:5)(cid:10)i(cid:1)(cid:5)h anartificialjnodealsodictatesthe(cid:4)(cid:1)ad(cid:1)diptiKon(cid:5)ofnewarcswith Letusexaminethefollowingargumenttodeterminethe and astheroutingprobabilities. service time at the artificial node. If an arrivingcustomer pKThe b(cid:4)l(cid:1)oc(cid:1)kepdKc(cid:5)ustomerproceedsto the finite queuewith is blocked, the queue is full and thus a customer is being probability onceagainafterincurringadelayatthe serviced, so thearrivingcustomertothe holdingnodehas (cid:1) artificialno(cid:4)d(cid:1)e(cid:1). IpfKth(cid:5)equeueisstillfull,itisre-routedwith to remain in service at the artificial holding node for the probability totheartificialnodewhereitincursanother remainingservicetimeintervalofthecustomerinservice. (cid:1) delay.ThispprKocesscontinuestillitfindsaspaceinthefinite Thedelaydistributionofablockedcustomerattheholding queue. A feedback arc is used to model the repeated de- nodehasthesamedistributionastheremainingservicetime lays.Theartificialnodeismodelledasan queue. ofthecustomerbeingservicedatthenodedoingtheblock- The infinite numberof servers is used simMpl(cid:4)yMto(cid:4)(cid:4)serve the ing.Usingrenewaltheory,onecanshowthattheremaining blockedcustomeradelaytimewithoutqueuing. servicetimedistributionhasthefollowingrate : (cid:6)h 5 forsolvingequation17with usedasadummyparameter (11) forsimplicityofthesolution.zLastly,equation21givesthe (cid:3)(cid:6)j (cid:6)h (cid:2) (cid:2) (cid:2) approximationtotheblockingprobabilityderivedfromthe (cid:1)(cid:12)(cid:7) (cid:6)j exactmodelforthe queue. Hence,weessen- where, is the service time variance given by Klein- rock(1975(cid:7))(cid:2). Notice that if theservice time distributionat tiallyhavefiveequatMio(cid:4)nMsto(cid:4)Cso(cid:4)lKve,viz.14to17and21. To recapitulate, we first expandthe networkwith an artificial thefinitequeuedoingtheblockingisexponentialwithrate holdingnode;thisstageisthenfollowedbytheapproxima- ,then: (cid:6)j tion of the routing probabilities, due to blocking, and the (cid:6)h (cid:2)(cid:6)j service delay in the holding node; and, finally, the feed- the service time at the artificial nodeis also exponentially backarcattheholdingnodeiseliminated.Oncethesethree distributedwithrate . stages are complete, we have an expandednetworkwhich (cid:6)j canthenbeusedtocomputetheperformancemeasuresfor 3.2.3 StageIII:feedbackelimination the original network. As a decomposition technique this approachallows successive additionof a holdingnodefor Duetothefeedbacklooparoundtheholdingnode,thereare everyfinite node, estimation of the parameters and subse- strong dependencies in the arrival processes. Elimination quenteliminationoftheholdingnode. ofthesedependenciesrequiresreconfigurationofthehold- ingnodewhichisaccomplishedbyrecomputingtheservice time atthenodeandremovingthefeedbackarc. Thenew 3.3 Combinatorialoptimisationmodels servicerateisgivenby: TheBufferAllocationProblem(BAP) is perhapsbestfor- (cid:1) (cid:1) (12) mulated as a non-linear multiple-objective programming (cid:6)h (cid:2)(cid:4)(cid:1)(cid:1)pK(cid:5)(cid:6)h problemwherethedecisionvariablesaretheintegers. Not Theprobabilitiesofbeinginanyofthetwophases(satu- onlyistheBAPadifficult hardcombinatorialoptimi- ratedorunsaturated)are and .Themeanservice timeatanodeiprecedingpKthefini(cid:4)t(cid:1)e(cid:1)npodKe(cid:5)is wheninthe sationproblem,itismadeNalPlt(cid:1)hemoredifficultbythefact Tunhsuast,uroanteadnpahvaesreaagned, t(cid:4)h(cid:6)e(cid:2)im(cid:1)e(cid:12)an(cid:6)(cid:1)hs(cid:2)e(cid:1)r(cid:5)viicnethtiem(cid:6)sea(cid:2)itau(cid:1)trattheednpohdaesei. tthoaitnttehrereolbajteectthiveeinftuengcetriodneciissinoontvoabrtiaaibnlaebsleianndcltohseepdefroformr- mancemeasuressuchasthroughput ,wox(cid:15)rk-in-process , precedingafinitenodeisgivenby: totalbuffersallocated ,andothRersystemperformanLc(cid:15)e measuressuchassystemiuxtiilisation foranybutthemost (13) (cid:6)(cid:10)(cid:2)i (cid:1) (cid:2)(cid:6)(cid:2)i (cid:1)(cid:12)pK(cid:6)(cid:1)h(cid:2)(cid:1) trivial situations. ComPbinatorial Op(cid:8)(cid:15)timisation approaches Similarequationscanbeestablishedwithrespecttoeach forsolvingproblemslike the BAP are generallyclassified ofthefinitenodes. Ultimately,wehavesimultaneousnon- aseitherexactoptimalapproachesorheuristicones. linearequationsinvariables , , alongwithaux- (cid:1) (cid:2)(cid:1) Exactapproachesareappropriateforsolvingsmallprob- iliaryvariablessuchas anpdK .pKSol(cid:6)vhingtheseequations leminstancesorforproblemswithspecialstructuree.g.the simultaneously we can(cid:6)cojmput(cid:5)e(cid:10)iall the performancemea- Travelling Salesman Problem, which admit optimal solu- suresofthenetwork. tions. Classical approachesforachievinganoptimalsolu- tionincludeBranch-and-Bound,Branch-and-Cut,Dynamic Programming,ExhaustiveSearch,andrelatedimplicitand (14) (cid:3) (cid:1) (cid:1) (cid:1)j(cid:1)(cid:1)h(cid:2)(cid:3)(cid:1)pK(cid:4) explicitenumerationmethods. Thedifficultywithutilising (15) (cid:1)j (cid:1) (cid:1)(cid:5)i(cid:2)(cid:3)(cid:1)pK(cid:4) these exact approaches for the BAP such as Branch-and- (16) (cid:1)j (cid:1) (cid:1)(cid:5)i(cid:1)(cid:1)h Boundisthatthesubproblemsforwhichoneseekstocom- p(cid:3)K (cid:1) (cid:2)j(cid:2)(cid:6)h(cid:2)h (cid:1) (cid:2)(cid:1)h(cid:7)(cid:2)(cid:7)(cid:2)rr(cid:1)K(cid:1)K(cid:3)(cid:1)(cid:2)r(cid:1)(cid:2)K(cid:4)r(cid:2)K(cid:1)(cid:3)(cid:2)(cid:2)r(cid:4)(cid:1)K(cid:1)(cid:4)(cid:2)(cid:2)r(cid:1)(cid:1)Kr(cid:1)(cid:2)K(cid:4)r(cid:2)K(cid:2)(cid:4)(cid:4)(cid:8)(cid:8)(1(cid:4)7(cid:2)) psdtuioftficehcuuapslttpiaces,rntahonend-ollroiniwgeieanrrabplorpourgonrbdalsmeommnisntohgeliptortbloejbelicestmigvaseiwnfuehdniccbhtyioantrheaearsees (cid:5) (cid:1) (cid:6)(18) z (cid:1) (cid:2)(cid:1)(cid:6)(cid:9)(cid:2)h(cid:4) (cid:1)(cid:10)(cid:1)(cid:2)h exactproblemdecompositionmethods. (cid:1) (cid:7)(cid:2)(cid:1)(cid:6)(cid:9)(cid:2)h(cid:4)(cid:1)z(cid:2)(cid:8) (19) This dilemma implies that heuristic approaches are the r(cid:2) (cid:1) (cid:9)(cid:2)h only reasonable methodology for large scale problem in- (cid:1)(cid:2) (20) stances of the BAP problem. Heuristic approaches can (cid:7)(cid:2)(cid:1)(cid:6)(cid:9)(cid:2)h(cid:4)(cid:6)z (cid:8) r(cid:1) (cid:1) be classified as either classical Non-linear Programming (cid:9)(cid:2)h searchmethodsorMetaheuristics. K (21) pK (cid:1) K(cid:4)(cid:3)c (cid:1) p(cid:4) Non-linear Programming (derivative-free)search (Him- c c(cid:11) (cid:2) melblau, 1972) methods such as, to name a few, Hooke- (cid:1) (cid:2) Equations 14 to 17 are related to the arrivals and feed- Jeeves,Nelder-Meadsimplexmethods,PARTAN,Powell’s backintheholdingnode. Theequations18to20areused ConjugateDirectionmetods,FlexibleTolerance,theCom- 6 plexMethodofBox,andotherrelatedtechniqueshavemet attemperaturesnearthefreezingpoint. Ifthisprocedureis with variedlevels ofsuccess intheBAP literatureandare notfollowedtheresultingsubstancemayformaglasswith viablemeansofdealingwiththeBAPbecauseofthenon- notcrystallineorderandonlymetastable,consistingoflo- closed form nature of the non-linear objective function. cally optimalstructures(Kirkpatricket al., 1983). During While many researchers feel that the objective functionis thecoolingprocessthesystemcanescapelocalminimaby concave or pseudo-concave in the decision variables, the movingtoathermalequilibriumofahigherenergypoten- discretenatureofthedecisionvariablesmakestheproblem tialbasedontheprobabilisticdistribution ofentropy : discontinuousandsonoderivativeinformationisavailable. w S (22) Metaheuristic methods such as Simulated Annealing, S (cid:2)klnw Tabu Search, Genetic Algorithms, and related techniques where isBoltzmann’sconstantand theprobabilitythat havenothistoricallybeenutilisedtosolvetheBAP;inthis thesysktemwillexistinthestateitiswinrelativetoallpos- paperweshallexploretheuseofSimulatedAnnealing. siblestatesitcouldbein. Thusgivenentropy’srelationto energy andtemperature : 3.4 Simulatedannealing E T (23) Simulatedannealingisanoptimisationmethodsuitablefor d(cid:15)E dS (cid:2) combinatorial optimisation problems. Such problems ex- T wearriveattheprobabilisticexpression ofenergydistri- hibit a discrete, factorially large configuration space. In butionforatemperature : w commonwithallparadigmsbasedon‘localimprovements’ T the simulated annealing method starts with a non-optimal (24) initialconfiguration(whichmaybechosenatrandom)and (cid:1)E w (cid:6)exp(cid:4) (cid:5) worksonimprovingitbyselectinganewconfigurationus- kT This so-called Boltzmann probability distribution is illus- ing a suitable mechanism (at randomin the simulated an- tratedin figure3. The probabilistic‘uphill’energymove- nealingcase)andcalculatingthecorrespondingcostdiffer- mentthatismadepossibleavoidstheentrapmentinalocal ential( ). Ifthecostisreduced,thenthenewconfigu- minimumandcanprovideagloballyoptimalsolution. rationis(cid:16)aRccKeptedandtheprocessrepeatsuntilatermination Theapplicationoftheannealingoptimisationmethodto criterionis satisfied. Unfortunately,such methodscan be- otherprocessesworksbyrepeatedlychangingtheproblem come‘trapped’inalocaloptimumthatisfarfromtheglobal configurationandgraduallyloweringthetemperatureuntil optimum. Simulated annealingavoids this problemby al- aminimumisreached. lowing ‘uphill’ moves based on a model of the annealing The correspondence between annealing in the physical processinthephysicalworld. worldandsimulatedannealingas usedforproductionline optimisationisoutlinedintable1. 3.5 Completeenumeration Asabase-ratebenchmarkforourapproach,weusedcom- Probability pleteenumeration(CE)ofallpossiblebufferandserveral- 1 locationpossibilities.Thenumberoffeasibleallocationsof bufferslotsamongthe intermediatebufferlocations 0.5 Qand serversamongtheN(cid:1)s(cid:1)tationsincreasesdramatically withS , ,and andisgNivenbytheformula: 0 Q S N 0.5 0 Energy0 .s5tate 0.1 Te0m.3perature QN(cid:12)N(cid:1)(cid:3)(cid:1)(cid:3) S(cid:12)NN(cid:1)(cid:1)(cid:1) (cid:1) (cid:2) (25) 1 (cid:1)(cid:7)Q(cid:6)(cid:1)(cid:8)(cid:7)Q(cid:6)(cid:2)(cid:8)(cid:3)(cid:3)(cid:3)(cid:7)Q(cid:6)(cid:2)N(cid:1)(cid:2)(cid:2)(cid:8)(cid:7)S(cid:6)(cid:1)(cid:8)(cid:7)S(cid:6)(cid:2)(cid:8)(cid:3)(cid:2)(cid:3)(cid:3)(cid:7)S(cid:6)N(cid:2)(cid:1)(cid:8) (cid:7)N(cid:2)(cid:2)(cid:8)(cid:9) (cid:7)N(cid:2)(cid:1)(cid:8)(cid:9) All bufferandservercombinationscanbe methodically enumeratedbyconsideringavector denotingtheposition within the productionline of each bpufferor server. Given Figure3: Probabilitydistribution ofenergy (cid:2)E the vector we can then easily map to or using the statesaccordingtotemperature. w (cid:6) exp(cid:4)kT (cid:5) followingepquation(forthecaseof ):p q s q Physical matter can be brought into a low-temperature ground state by careful annealing. First the substance is Q if (26) (cid:1) othpejrw(cid:2)isie melted,thenitis graduallycooledwithalotoftimespent qi (cid:2) (cid:6) j(cid:4)(cid:1)(cid:7) X 7 Thegeneralisedqueuingnetworkthroughputevaluation PhysicalWorld Production Line algorithm was initially ported from a VAX VMS operating Optimisation system,toaPC-basedIntelarchitecture. Mostchangesin- Atomplacement Lineconfiguration volvedtheadjustmentofnumericalconstantsaccordingto Random atom Buffer space, server, theIEEE488floatingpointrepresentationusedontheIntel movements serviceratemovement platforms. Subsequently, the algorithmwas rewritten in a Energy Throughput pure subroutine form so that it could be repeatedly called EnergyEdifferential ConfiguratioRn through- from within a program and semi-automatically converted (cid:16)E putdifferential fromFORTRANtoANSIC. Energystateprobability Changes accor(cid:16)diRng to TheSA algorithmused fortheproductionline resource distribution the Metropolis crite- optimisationisgiveninfigure4. rion, (cid:2)(cid:10)E TheSAprocedurewasrunwiththefollowingcharacter- exp(cid:4) ,T i(cid:5)mpl(cid:3)e- isticsbasedonthenumberofstations : mraenndti(cid:4)n(cid:6)g(cid:2)(cid:2)th(cid:2)e(cid:1)(cid:5)Boltzmann N Maximumtrialsatgiventemperature probabilitydistribution (cid:1)(cid:6)(cid:6)(cid:7)N Temperature Variable for establish- Maximumsuccessesatgiventemperature ing configuration ac- (cid:1)(cid:6)(cid:7)N ceptancetermination Initialtemperature (cid:6)(cid:2)(cid:17) Coolingschedule Exponential: Table 1: Correspondencebetween annealing in the physi- TK(cid:6)(cid:1) (cid:2)(cid:6)(cid:2)(cid:18)TK Initiallineconfiguration Equal division of buffers and calworldandsimulatedannealingusedforproductionline servers among stations with any remaining resources optimisation. placedonthestationinthemiddle. Reportedtime Elapsedwallclocktimeinseconds. Giventhenabufferorserverconfiguration weadvanceto thenextconfiguration usingthefunctionpg: (cid:1) Thefollowingfactsclarifytheuseoftheevaluationalgo- p rithm: (27) (cid:1) p (cid:2)g(cid:4)p(cid:1)(cid:1)(cid:5) thethroughputusedasthelinemetricandpresentedin wheregisrecursivelydefined(for buffers)as: (cid:3) the tabulated results is the throughputat the last sta- Q if tion, (cid:2)p(cid:2)(cid:3)(cid:4)(cid:4)(cid:4)(cid:3)pl(cid:6)(cid:3)(cid:3)(cid:4)(cid:4)(cid:4)(cid:3)pQ(cid:4) pl(cid:6)(cid:3)(cid:2)N g(cid:2)p(cid:3)l(cid:4)(cid:1) (cid:2)wph(cid:3)(cid:2)e(cid:3)r(cid:4)e(cid:4)(cid:4)(cid:3)p(cid:3)l(cid:4)(cid:2)(cid:3)p(cid:3)l(cid:3)(cid:2)(cid:3)(cid:4)(cid:4)(cid:4)(cid:3)p(cid:3)l(cid:3)(cid:2)(cid:4) otherwise the line topology graph is a linear series of stations (cid:8)(cid:9) p(cid:3) (cid:1)g(cid:2)p(cid:3)l(cid:6)(cid:3)(cid:4) (28) (cid:3) allowingforparallelservers,and Essentially,gmapsthevectorofpositionstoanewonerep- theinitialandtheeffectivearrivalrateatthefirstserver resenting(cid:10)anotherlineconfiguration. Whentheposition (cid:3) aresetto . ofbufferorserverresource ,asitisincremented,reachpesi (cid:1)(cid:2)(cid:17) thelastplaceintheline (i )gisrecursivelyap- Threebatchesoftestswereplannedandexecuted. One, pliedtothebufferorserNverpiin(cid:12)p(cid:1)os(cid:2)itioNn setting – presented in tables 2–13, was planned in order to com- tothenewvalueof .Thecompletepeni(cid:6)um(cid:1) erationpteirmpiQ- pare the results of our approach with those of Hillier and nateswhenallbuffepris(cid:6)o(cid:1)rserversreachthelineposition . So (1995). A second set, presented in tables 14–17, was ToenumerateallbufferandservercombinationsonecoNm- plannedinordertocomparetheapproachtotheresultsob- plete serverenumerationis performedforeach line buffer tainedusingadifferentevaluativeprocedure(Spinellisand configuration. Papadopoulos,1997). Finally,thethirdbatchoftests, pre- sentedintables18–27,aimedatestablishingourmethod’s efficacyindeterminingoptimalconfigurationsoflargepro- 4 Experimental design ductionlines. Tables14–27containanadditionalcolumn, ,presentingtheprogramexecutiontime( ). ThemaingenerativeprocedureweusedwasSimulatedAn- T s nealing (Kirkpatrick et al., 1983, Laarhoven and Aarts, 1987, Spinellis and Papadopoulos, 1997) with complete 5 Experimental results enumeration used where practical in order to validate the SAresults.Inordertotesttheapproach,339casesofsmall Thefollowingparagraphsoutlinetheresultsweobtainedby andlargeproductionlinesfromawidecombinationofallo- utilising the described methodology on a variety of prob- cationoptionsweresearchedusingSAorCE. lems. Where applicable we compare the obtained results 8 withotherpublisheddata. Inourdescriptionoftheresults 1. Set initial line configuration. Set , set wesometimesrefertothephenomenaofthe bowlandthe (cid:13) (cid:14) q,ise(cid:8)t bQ(cid:4)Nc , -shape. N qseNt(cid:1)(cid:2) (cid:8) qN(cid:1)(cid:2)(cid:12)Q(cid:1) i(cid:4)(cid:2)bQ(cid:4)Nc ,sesti (cid:8) bS(cid:4)Nc. L Thebowlphenomenonoccurswhenfor , N sN(cid:1)(cid:2) (cid:8)sN(cid:1)(cid:2)(cid:12)SP(cid:1) i(cid:4)(cid:1)bS(cid:4)Nc wi (cid:8)(cid:1)(cid:4)N j (cid:2)(cid:1)(cid:1)(cid:3)(cid:1)(cid:2)(cid:2)(cid:2)(cid:1)N 2. Setinitialtemperature . Set . (29) (cid:4) (cid:4) (cid:13) TP(cid:5)(cid:14) T (cid:8)(cid:6)(cid:2)(cid:17) wj (cid:2) wN(cid:6)(cid:1)(cid:2)j 3. Initializestepandsuccesscount. Set ,set (30) (cid:13). (cid:14) S (cid:8)(cid:6) U (cid:8) wj(cid:4) (cid:3) wj(cid:4)(cid:6)(cid:1) (cid:1)(cid:9)j (cid:9) N (cid:1)(cid:1) (cid:6) (cid:3) (cid:1) (cid:2) 4. Createnewlinewitharandomredistributionofbuffer (31) (cid:4) (cid:4) N (cid:1)(cid:1) (cid:13)space,servers,orservicerate. wj (cid:9) wj(cid:6)(cid:1) (cid:9)j (cid:9)N (cid:12)(cid:1) (cid:3) (cid:14) (cid:1) (cid:2) Thenamearisesfromthefactthatthevalueof is (a) Createa copyoftheconfigurationvectors. Set (cid:4) (cid:4) (cid:13) ,set ,set . (cid:14) considerablylargerthantheother (cid:4),whereaswth(cid:1)eo(cid:2)thwerN (cid:4) q(cid:1) (cid:8)q s(cid:1) (cid:8)s w(cid:1) (cid:8)w are relatively close, so that plottinwgjthe (cid:4) versus givwejs (b) Determine which vectorto modify. Set theshapeofthecross-sectionoftheinterwiojrofabowjl. (cid:13) . (cid:14) Rv (cid:8) Correspondingly,the -phenomenonistheallocationof brand(cid:13)(cid:6)(cid:2)(cid:2)(cid:2)(cid:3)(cid:14)c (c) if Create new line with a random extraserverstothefirstLstationoftheproductionline. This redRisvtrib(cid:2)utio(cid:6)n(cid:13)of buffer space. Move space leadstoalmostthemaximumthroughputoftheline,which froma source buffer to a(cid:14)destinatiRonnbuffer is attained when the extra servers are allocated to the last : set qRs , set stationoftheline,havingtheadvantagethatitleadstoless qRd Rs (cid:8), sebtrand(cid:13)(cid:3)(cid:2)(cid:2)(cid:2)N(cid:5)c Rd (cid:8) work-in-processinventory. bra,nsde(cid:13)t(cid:3)(cid:2)(cid:2)(cid:2)N(cid:5)c Rn ,(cid:8)set brand(cid:13)(cid:1)(cid:2)(cid:2)(cid:2)qRs (cid:12). (cid:1)(cid:5)c qRs (cid:8)qRs (cid:1)Rn qRd (cid:8)qRd (cid:12)Rn (d) if Createnewlinewitharandomredis- 5.1 Short lines triRbuvtio(cid:2)n(cid:1)of(cid:13)serverallocation. Move servers from source station to a(cid:14)destinatRionn station Tables 2–13 present results for the optimal allocation us- : set sRs , set ing SA for lines similar to the ones examined by Hillier sRd Rs (cid:8), sebtrand(cid:13)(cid:1)(cid:2)(cid:2)(cid:2)N(cid:5)c Rd (cid:8) andSo(1995).InallcasesCEresultshavebeencalculated bra,nsde(cid:13)t(cid:1)(cid:2)(cid:2)(cid:2)N(cid:5)c Rn ,(cid:8)setbrand(cid:13)(cid:1)(cid:2)(cid:2)(cid:2)sRs (cid:12). forcomparisonpurposes.Incomparingtheresultsobtained (cid:1)(cid:5)c sRs (cid:8)sRs(cid:1)Rn sRd (cid:8)sRd(cid:12)Rn withthosepresentedbyHillierandSothefollowingimpor- (e) if Create new line with a random re- tantobservationscanbemade: disRtrvibu(cid:2)tio(cid:3)n(cid:13)of service rate. Move service ratefromsourcestation (cid:14)toadestRinnationsta- 1. Theservicerateallocation(table2)doesnotfollowthe tion : set sRs , set bowlphenomenon,butdiminishestowardstheendof sRd Rs,s(cid:8)et brand(cid:13)(cid:1)(cid:2)(cid:2)(cid:2)N(cid:5)c Rd,s(cid:8)et theline. brand(cid:13)(cid:1)(cid:2)(cid:2)(cid:2)N(cid:5)c ,sRetn (cid:8) rand(cid:4)(cid:6)(cid:2)(cid:2)(cid:2)wR.s(cid:14) wRs (cid:8)wRs (cid:1)Rn wRd (cid:8)wRd (cid:12)Rn 2. Thebufferallocation(table3)doesnotfollowthebowl 5. Calculate energy differential. Set phenomenon, but increases monotonically across the (cid:13) . (cid:14) (cid:1) (cid:1) (cid:1) line. (cid:16)E (cid:8)R(cid:4)q(cid:1)s(cid:1)w(cid:5)(cid:1)R(cid:4)q (cid:1)s(cid:1)w (cid:5) 6. Decide acceptance of new configuration. Accept all 3. The server allocation (table 4) for a small numberof (cid:13)new configurations that are more efficien(cid:14)t and, fol- servers follows a pattern strikinglysimilar to the one lowing the Boltzmann probability distribution, some presented in Hillier and So (1995). However, as the that are less efficient: if or (cid:2)(cid:10)E numberof servers increases, servers tend to accumu- ,set (cid:1),(cid:16)seEt (cid:9) (cid:6)(cid:1),seetxp(cid:4) T (cid:1)(cid:5),s(cid:3)et latetowardsthebeginningoftheline. rand(cid:4)(cid:6)(cid:2)(cid:2)(cid:2)(cid:1).(cid:5) q (cid:8) q s(cid:8) s w (cid:8)w U (cid:8)U (cid:12)(cid:1) 4. Theserverandservicerateallocation(table5)donot 7. Repeat forcurrenttemperature. Set . If (cid:13) maximumnumberofsteps,(cid:14)gotosStep(cid:8)4.S (cid:12)(cid:1) exhibitthe -phenomenon. L S (cid:9) 8. Lowertheannealingtemperature. Set ( 5. Thebufferandservicerateallocation(table6)results (cid:13) ). (cid:14) T (cid:8) cT (cid:6) (cid:9) areverysimilartotheresultspresentedinHillierand c(cid:9)(cid:1) So(1995).Asfarasthebufferallocationisconcerned, 9. Check if progress has been made. If , go to bufferstendtoaccumulatetowardstheendoftheline. (cid:13)step3;otherwisethealgorithmterm(cid:14)inatUes.(cid:3) (cid:6) The allocationof workhowever, does not exhibit the (cid:1) bowlphenomenonpresentedintheotherstudyandfol- lowstheusualdescendingrateacrosstheline. Figure4: Simulatedannealingalgorithmfordistributing bufferspace, servers,and servicerateinan -statioQn 9 line. S N N 6. The buffer and server allocation (tables 7, 8) results are roughly similar to those presented in Hillier and 18000 So (1995)inboththeserverandthebufferallocation 16000 q vectors.Serverstendtoaccumulatetowardsthebegin- q, w 14000 q, s ning and middle of the line, whereas buffers tend to q, s, w 12000 accumulatetowardsthelineends. w ) (s 10000 s 7. Finally, the buffer, server, and service rate allocation e s, w m 8000 (table9)roughlyfollowstheshapeofservicerateal- Ti locationpresentedinHillierandSo(1995),buttheal- 6000 locationofbuffersandserversisquitedissimilar. 4000 2000 ConcerningthebehaviourofSAcomparedtoCEthere- sultsofthetwomethodsareasfollows: 0 0 10 20 30 40 50 60 70 1. Buffer allocation (tables 3, 10) is the same using SA Station number andCE. 2. Theserverallocationofalargenumberofservers(ta- Figure5:Executiontimeforoptimalconfigurationcalcula- bles4,11)usingSAdiffersfromtheallocationusing tionsusingSA. CE but only in the last two experiments, and this is probablyduetoincorrectresultsreturnedbytheeval- uationalgorithm. 120servers,andtheservicerateconfigurationin hours ona120MHzPentium-processorsystem. Thissol(cid:19)u(cid:2)t(cid:17)ionus- 3. For buffer and server allocation for up to 4 stations ing SA required the testing of 238,248 different configu- (tables7,12)SAandCEgivethesameresults. rations. The dominant factor of the algorithm is the line configurationevaluationtakingon oursystem about 70ms 4. Forbufferandserverallocationof5stations(tables8, for every configuration test. Multiplying this time factor 13)SAandCEdifferintheallocatedvectorsofsome by the number of all possible buffer and server allocation configurations,butwithonlyaslightimpactonthere- configurationsfortheabovecase—accordingtoequation sultingthroughput. 25—givesusanapproximateevaluationtimeusingCEof 5.2 Longlines (cid:1)(cid:3)(cid:6)(cid:12)(cid:21)(cid:6)(cid:1)(cid:3) (cid:1)(cid:3)(cid:6)(cid:12)(cid:21)(cid:6)(cid:1)(cid:1) (cid:11)(cid:12) or (cid:6)(cid:2)(cid:6)(cid:20)s (cid:2) (cid:1)(cid:6) s y(cid:21)ea(cid:6)rs(cid:1). (cid:3) (cid:21)(cid:6)(cid:1)(cid:1) Thetables(16–27)presentresultsforoptimalallocationus- (cid:1)(cid:11)(cid:2) (cid:2)(cid:1) (cid:2) (cid:1)(cid:6) ing SA for lines of stations with buffers and N s(cid:2)erv(cid:19)er(cid:1)s. (cid:20)W(cid:6) here practicalQ(i.(cid:2)e. (cid:3)fNor 6 Summary and conclusions small lines)SCE(cid:2)res(cid:3)uNlts have been calculated for compari- sonpurposes. Theoptimalallocationofavariablenumber of buffers to a fixed number of servers (9 servers in table Our approach of using the expansion method for evaluat- 14; 15 servers table 16) follows a regular pattern: buffers ing the production line configurations generated by simu- areaddedtothelinefromtherighttotheleft. Theresults lated annealing optimisation allowed us to explore small oftheCE(table15;table17)confirmtheSAresults.How- and large line configurations in bounded execution time. ever, this placement strategy is different from the optimal Theresultsobtainedwithourapproach,comparedtothose strategyobtainedusingthedecompositionmethod(Dallery of other studies exhibit some interesting similarities but andFrein,1993)astheevaluativeprocedure(Spinellisand alsostrikingdifferencesbetweentheminthe allocationof Papadopoulos,1997). buffers,numbersofservers,andtheirservicerates. While context dependent, these patterns of allocation are one of the most important insights which emerge in solving very 5.3 Performanceanalysis long production lines. The patterns, however, are often The execution time of the SA optimisation algorithm in- counter-intuitive, which underscores the difficulty of the creasedexponentiallydependingonthenumberofstations problemweaddressed. (figure 5). This increase was a lot better than the combi- Havingdemonstratedtheviabilityofusinganalgorithm natorial explosionof CE, and allowedus to producenear- based on randomisation techniques for optimising large optimalconfigurationsforrelativelylargeproductionlines productionlineswenowplantoexploreotheroptimisation inreasonabletime. Attheextremeendfora60stationline methods based on similar principles such as genetic algo- SA produced a solution for the allocation of 120 buffers, rithmstofindhowtheycomparetoourapproach.Thesim- 10
Description: