mit2 852s10 control

Upload: mine-rh

Post on 08-Aug-2018

222 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/22/2019 MIT2 852S10 Control

    1/108

    MIT2.852

    ManufacturingSystemsAnalysis

    Lectures1921

    Scheduling:Real-TimeControlofManufacturingSystems

    StanleyB.GershwinSpring,2007

    Copyrightc2007

    Stanley

    B.

    Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    2/108

    DefinitionsEventsmaybecontrollable ornot,andpredictable

    ornot.controllable uncontrollable

    predictable

    loadinga

    part

    lunch

    unpredictable ??? machinefailure

    Copyright2007StanleyB.Gershwin.c 2

  • 8/22/2019 MIT2 852S10 Control

    3/108

    DefinitionsSchedulingistheselectionoftimesforfuturecontrollableevents.

    Ideally,schedulingsystemsshoulddealwithallcontrollableevents,andnotjustproduction.Thatis,theyshouldselecttimesforoperations,

    set-upchanges,preventivemaintenance,etc.Theyshouldatleastbeaware ofset-upchanges,

    preventivemaintenance,etc.whentheyselecttimesforoperations.

    Copyright2007StanleyB.Gershwin.c 3

  • 8/22/2019 MIT2 852S10 Control

    4/108

    DefinitionsBecauseofrecurringrandomevents,schedulingis

    anon-goingprocess,andnotaone-timecalculation.

    Scheduling,orshopfloorcontrol,isthebottomofthescheduling/planninghierarchy.Ittranslatesplansintoevents.

    Copyright c 42007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    5/108

    IssuesinFactoryControlProblemsaredynamic;currentdecisionsinfluence

    futurebehaviorandrequirements.Therearelargenumbersofparameters,time-varying

    quantities,andpossibledecisions.Sometime-varyingquantitiesarestochastic.Somerelevantinformation(MTTR,MTTF,amountof

    inventoryavailable,etc.) isnotknown.Somepossiblecontrolpoliciesareunstable.Copyright c 52007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    6/108

    Dynamic ExampleProgramming Problem

    DiscreteTime,DiscreteState,DeterministicF6

    8

    A1

    B L5

    6

    22

    G5

    1064M

    9

    2D 7

    72

    4 N

    594E

    5 I 86

    136 2C

    H

    J

    1

    4Z

    K 3 O

    Problem: findtheleastexpensivepathfromAtoZ.Copyright c 62007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    7/108

    Dynamic ExampleProgramming Problem

    Letg(i,j)bethecostoftraversingthelinkfromitoj.Leti(t)bethetthnodeonapathfromAtoZ.Thenthepathcostis

    Tg(i(t1), i(t))

    t=1whereT isthenumberofnodesonthepath,i(0)=A,andi(T) =Z.T isnotspecified;itispartofthesolution.

    Copyright c 72007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    8/108

    Dynamic ExampleProgramming SolutionApossibleapproachwouldbetoenumerateallpossiblepaths

    (possiblesolutions).However,therecanbealotofpossiblesolutions.

    Dynamicprogrammingreducesthenumberofpossiblesolutionsthatmustbeconsidered.Goodnews: itoftengreatly reducesthenumberofpossiblesolutions.Badnews: itoftendoesnotreduceitenoughtogiveanexactoptimalsolutionpractically(ie,withlimitedtimeandmemory).Thisisthecurseofdimensionality.

    Goodnews: wecanlearnsomethingbycharacterizingtheoptimalsolution,andthatsometimeshelpsingettingananalyticaloptimalsolutionoranapproximation.

    Goodnews: ittellsussomethingaboutstochasticproblems.Copyright2007StanleyB.Gershwin.c 8

  • 8/22/2019 MIT2 852S10 Control

    9/108

    Dynamic ExampleProgramming Solution

    InsteadofsolvingtheproblemonlyforAastheinitialpoint,wesolveitforall possibleinitialpoints.Foreverynodei,defineJ(i)tobetheoptimalcosttogo fromNodeitoNodeZ(thecostoftheoptimalpathfromitoZ).Wecanwrite

    TJ(i) = g(i(t1), i(t))

    t=1wherei(0)=i;i(T) =Z;(i(t1), i(t))isalinkforeveryt.Copyright c 92007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    10/108

    Dynamic ExampleProgramming Solution

    ThenJ(i)satisfiesJ(Z) = 0

    and,iftheoptimalpathfromitoZ traverseslink(i,j),J(i) =g(i,j) +J(j).

    i j

    Z

    Copyright2007StanleyB.Gershwin.c 10

  • 8/22/2019 MIT2 852S10 Control

    11/108

    Dynamic ExampleProgramming Solution

    SupposethatseverallinksgooutofNodei.

    4j

    3j

    2j

    5j

    1j

    6j

    i

    Z

    Supposethatforeachnodej forwhichalinkexistsfromitoj,theoptimalpathandoptimalcostJ(j)fromj toZ isknown.Copyright2007StanleyB.Gershwin.c 11

  • 8/22/2019 MIT2 852S10 Control

    12/108

    Dynamic ExampleProgramming Solution

    ThentheoptimalpathfromitoZ istheonethatminimizesthesumofthecostsfromitojandfromj toZ.Thatis,

    J(i)=min [g(i,j) +J(j)]j

    wheretheminimizationisperformedoveralljsuchthatalinkfromitojexists.ThisistheBellmanequation.This

    is

    a

    recursion

    or

    recursive

    equation

    because

    J()

    appears

    onbothsides,althoughwithdifferentarguments.J(i)canbecalculatedfromthisifJ(j)isknownforeverynodejsuchthat(i,j)isalink.Copyright c 122007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    13/108

    Dynamic ExampleProgramming SolutionBellmansPrincipleofOptimality: ifiandjarenodesonanoptimalpathfromAtoZ,thentheportionofthatpathfromAtoZbetweeniandj isanoptimalpathfromitoj.

    A

    j

    i

    Copyright2007StanleyB.Gershwin.c 13

  • 8/22/2019 MIT2 852S10 Control

    14/108

    Dynamic ExampleProgramming Solution

    Example: AssumethatwehavedeterminedthatJ(O) = 6andJ(J)=11.TocalculateJ(K),

    g(K,O) +J(O)J(K)= ming(K,J) +J(J)

    3 + 6= min = 9.9+11

    Copyright2007StanleyB.Gershwin.c 14

  • 8/22/2019 MIT2 852S10 Control

    15/108

    Dynamic ExampleProgramming Solution

    Algorithm1.SetJ(Z) = 0.2.Findsomenodeisuchthat

    J(i)hasnotyetbeenfound,andforeachnodej inwhichlink(i,j)exists,J(j)isalready

    calculated.AssignJ(i)accordingto

    J(i)=min [g(i,j) +J(j)]j3.RepeatStep2untilallnodes,includingA,havecosts

    calculated.Copyright c 152007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    16/108

    K

    Dynamic ExampleProgramming Solution

    F

    A

    B11 L

    5

    6HC

    14

    D

    EI

    O

    J

    9

    G

    M

    N48

    14

    13

    11

    17

    12

    136

    11

    Z

    Copyright2007StanleyB.Gershwin.c 16

  • 8/22/2019 MIT2 852S10 Control

    17/108

    Dynamic ExampleProgramming SolutionTheimportantfeaturesofadynamicprogrammingproblemarethestate(i);thedecision (togotojafteri);

    theobjectivefunction T

    t=1g(i(t1), i(t))thecost-to-gofunction(J(i));theone-steprecursionequationthatdeterminesJ(i)

    (J(i)=minj[g(i,j) +J(j)]); thatthesolutionisdeterminedforeveryi, notjustAandnotjustnodesontheoptimalpath;

    thatJ(i)dependsonthenodestobevisitedafteri, notthosebetweenAandi.Theonlythingthatmattersisthepresentstateandthefuture;

    thatJ(i)isobtainedbyworkingbackwards.Copyright2007StanleyB.Gershwin.c 17

  • 8/22/2019 MIT2 852S10 Control

    18/108

    Dynamic ExampleProgramming Solution

    Thisproblemwasdiscretetime,discretestate,deterministic.

    Otherversions:discretetime,discretestate,stochasticcontinuoustime,discretestate,deterministiccontinuoustime,discretestate,stochasticcontinuoustime,mixedstate,deterministiccontinuoustime,mixedstate,stochastic

    instochasticsystems,weoptimizetheexpected cost.Copyright c 182007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    19/108

    ProgrammingDiscretetime,discretestate

    StochasticDynamic

    Supposeg(i,j)isarandomvariable;orifyouareatiandyouchoosej,youactuallygotokwith

    probabilityp(i,j,k).Thenthecostofasequenceofchoicesisrandom.Theobjectivefunctionis

    TE g(i(t1), i(t))

    t=1

    andwecandefineJ(i) =Emin[g(i,j) +J(j)]

    jCopyright2007StanleyB.Gershwin.c 19

  • 8/22/2019 MIT2 852S10 Control

    20/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Context: Theplanning/schedulinghierarchyLongterm: factorydesign,capitalexpansion,etc.Mediumterm:demandplanning,staffing,etc.Shortterm:

    responsetoshorttermeventspartreleaseanddispatchIn

    this

    problem,

    we

    deal

    with

    the

    response

    to

    short

    term

    events.

    Thefactoryandthedemandaregiventous;wemustcalculateshorttermproductionrates;theseratesarethetargetsthatreleaseanddispatchmustachieve.

    Copyright c 202007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    21/108

    Dynamic

    ProgrammingContinuousTime,MixedState

    StochasticExamplex1

    x2

    d1

    d2

    u (t)1

    u (t)2 Type 2

    Type 1

    r, p

    Perfectlyflexiblemachine,twoparttypes.i timeunitsrequiredtomakeTypeiparts,i= 1,2.

    Exponentialfailuresandrepairswithratespandr.Constantdemandratesd1,d2.Instantaneousproductionratesui(t), i= 1,2control

    variables.Downstreamsurplusesxi(t).Copyright c 212007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    22/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Objective: Minimizethedifferencebetweencumulativeproductionandcumulativedemand.Thesurplussatisfiesxi(t) =Pi(t)Di(t)

    t

    CumulativeProduction

    and Demand production P (t)

    surplus x (t)

    i

    i

    idemand D (t) = d ti

    Copyright c2007StanleyB.Gershwin. 22

  • 8/22/2019 MIT2 852S10 Control

    23/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Feasibility:Fortheproblemtobefeasible,itmustbepossibletomake

    approximatelydiT TypeipartsinalongtimeperiodoflengthT, i= 1,2.(Whyapproximately?)

    ThetimerequiredtomakediT partsisidiT.Duringthisperiod,thetotaluptimeofthemachineie,the

    timeavailableforproductionisapproximatelyr/(r+p)T.Therefore,wemusthave1d1T +2d2Tr/(r+p)T,or

    2

    ridi

    r+pi=1

    Copyright2007StanleyB.Gershwin.c 23

  • 8/22/2019 MIT2 852S10 Control

    24/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Ifthisconditionisnotsatisfied,thedemandcannotbemet.Whatwillhappentothesurplus?Thefeasibilityconditionisalsowritten

    2di r

    i=1 i r+pwherei =1/i.Iftherewereonlyoneparttype,thiswouldber

    dr+p

    Lookfamiliar?Copyright2007StanleyB.Gershwin.c 24

  • 8/22/2019 MIT2 852S10 Control

    25/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Thesurplussatisfiesxi(t) =Pi(t)Di(t)

    where

    tPi(t) = u

    i(s)ds; D

    i(t) =d

    it

    0Therefore

    dxi(t)=ui(t)di

    dt

    Copyright2007StanleyB.Gershwin.c 25

  • 8/22/2019 MIT2 852S10 Control

    26/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Todefinetheobjectivemoreprecisely,lettherebeafunctiong(x1, x2)suchthatgisconvexg(0,0)=0 lim g(x1, x2) =; lim g(x1, x2) =.

    x1 x1 lim g(x1, x2) =; lim g(x1, x2) =.x2 x2

    Copyright2007StanleyB.Gershwin.c 26

  • 8/22/2019 MIT2 852S10 Control

    27/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Examples:g(x1, x2) =A1x12 +A2x22g(x1, x2) =A1|x1|+A2|x2|g(x1, x2) =g1(x1) +g2(x2)where

    + gi(xi) =g(i+)xi +g(i)xi , xi+ =max(xi,0),xi =min(xi,0),

    g(i+) >0, g(i) >0.

    Copyright c 272007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    28/108

    DynamicProgramming

    Objective:

    ContinuousTime,MixedStateStochasticExample

    Tmin

    E g(x1(t), x2(t))dt

    0

    x1

    g(x ,x )1 2

    x2

    Copyright2007StanleyB.Gershwin.c 28

  • 8/22/2019 MIT2 852S10 Control

    29/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Constraints:u1(t)0; u2(t)0

    Short-termcapacity:Ifthemachineisdownattimet,

    u1(t) =u2(t) = 0

    Copyright2007StanleyB.Gershwin.c 29

  • 8/22/2019 MIT2 852S10 Control

    30/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleAssumethemachineisupforashortperiod[t,t+t].Lett

    Dynamic

    besmallenoughsothatui isconstant;thatisui(s) =ui(t), s[t,t+t]

    Themachinemakesui(t)tpartsoftypei.ThetimerequiredtomakethatnumberofTypeipartsisiui(t)t.Therefore

    iui(t)tt 1/i

    oriui(t)1

    1/1

    2

    u2

    0

    iCopyright2007StanleyB.Gershwin.c 30

  • 8/22/2019 MIT2 852S10 Control

    31/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamic

    Machinestatedynamics: Define(t)tobetherepairstateofthemachineattimet.(t) = 1meansthemachineisup;(t) = 0meansthemachineisdown.

    prob((t+t) = 0|(t)=1)=pt+o(t)prob((t+t) =1|(t)=0)=rt+o(t)

    Theconstraintsmaybewritteniui(t)(t); ui(t)0

    iCopyright2007StanleyB.Gershwin.c 31

  • 8/22/2019 MIT2 852S10 Control

    32/108

    ProgrammingContinuousTime,MixedState

    StochasticExampleDynamicDynamicprogrammingproblemformulation: T

    minE g(x1(t), x2(t))dtsubjectto: 0

    dxi(t)=ui(t)di

    dtprob((t+t) = 0|(t)=1)=pt+o(t)prob((t+t) =1|(t)=0)=rt+o(t)

    iui(t)(t); ui(t)0i

    x(0), (0)specified

    Copyright2007StanleyB.Gershwin.c 32

  • 8/22/2019 MIT2 852S10 Control

    33/108

    Dynamic ElementsofaDPProblemProgrammingstate: xalltheinformationthatisavailabletodeterminethe

    futureevolutionofthesystem.control: utheactionstakenbythedecision-maker.objectivefunction: J thequantitythatmustbeminimized;dynamics: theevolutionofthestateasafunctionofthecontrol

    variablesandrandomevents.constraints: thelimitationsonthesetofallowablecontrolsinitialconditions: thevaluesofthestatevariablesatthestart

    ofthetimeintervaloverwhichtheproblemisdescribed.Therearealsosometimesterminalconditions suchasinthenetworkexample.

    Copyright c 332007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    34/108

    Dynamic ElementsofaDPSolutionProgrammingcontrolpolicy: u(x(t), t). Astationary ortime-invariant policyisoftheformu(x(t)).

    valuefunction: (alsocalledthecost-to-go function)thevalueJ(x,t)oftheobjectivefunctionwhentheoptimalcontrolpolicyisappliedstartingattimet,whentheinitialstateisx(t) =x.

    Copyright c 342007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    35/108

    Bellmans Continuousx,tEquation Deterministic

    T

    Problem: min g(x(t), u(t))dt+F(x(T))u(t),0tT 0

    suchthatdx(t)

    =f(x(t), u(t), t)dt

    x(0)specified

    h(x(t), u(t))0

    xRn, uRm, fRn, hRk,andgandF arescalars.Data: T,x(0),andthefunctionsf,g,h,andF.Copyright c 352007StanleyB.Gershwin.

  • 8/22/2019 MIT2 852S10 Control

    36/108

    Bellmans Continuousx,tEquation DeterministicThecost-to-gofunctionis

    TJ(x,t)=min g(x(s), u(s))ds+F(x(T))

    t

    T

    J(x(0),0)=min g(x(s), u(s))ds+F(x(T))0 t1 T

    = min g(x(t), u(t))dt+ g(x(t), u(t))dt+F(x(T)) .u(t), 0 t1

    0tT

    Copyright2007StanleyB.Gershwin.c 36

    C i

  • 8/22/2019 MIT2 852S10 Control

    37/108

    Bellmans Continuousx,tEquation Deterministic

    t1 T= min g(x(t), u(t))dt+ min g(x(t), u(t))dt+F(x(T))

    u(t),

    0 u(t), t1

    0tt1 t1tT t1 = min g(x(t), u(t))dt+J(x(t1), t1) .

    u(t), 00tt1

    Copyright2007StanleyB.Gershwin.c 37

    C ti

  • 8/22/2019 MIT2 852S10 Control

    38/108

    Bellmans Continuousx,tEquation Deterministic

    where T

    J(x(t1), t1)= min g(x(t), u(t))dt+F(x(T))u(t),t1tT t1

    suchthatdx(t)

    =f(x(t), u(t), t)dt

    x(t1)specifiedh(x(t), u(t))0

    Copyright2007StanleyB.Gershwin.c 38

    C ti t

  • 8/22/2019 MIT2 852S10 Control

    39/108

    Bellmans Continuousx,tEquation Deterministic

    Breakup[t1, T]into[t1, t1 +t][t1 +t,T] :t1+tJ(x(t1), t1)=min g(x(t), u(t))dt

    u(t1) t1+J(x(t

    1 +t), t

    1 +t)}

    wheretissmallenoughsothatwecanapproximatex(t)andu(t)withconstantx(t1)andu(t1),duringtheinterval.Then,approximately,J(x(t1), t1)=min g(x(t1), u(t1))t+J(x(t1 +t), t1 +t)

    u(t1)

    Copyright2007StanleyB.Gershwin.c 39

    C ti t

  • 8/22/2019 MIT2 852S10 Control

    40/108

    Bellmans Continuousx,tEquation Deterministic

    Or,J(x(t1), t1)=min g(x(t1), u(t1))t+J(x(t1), t1)+

    u(t1)J J

    (x(t1), t1)(x(t1 +t)x(t1)) + (x(t1), t1)tx tNotethat

    dxx(t1 +t) =x(t1) + t=x(t1) +f(x(t1), u(t1), t1)tdt

    Copyright2007StanleyB.Gershwin.c 40

    Continuous t

  • 8/22/2019 MIT2 852S10 Control

    41/108

    Bellmans Continuousx,tEquation Deterministic

    ThereforeJ(x,t1) =J(x,t1)

    J J+min g(x,u)t+ (x,t1)f(x,u,t1)t+ (x,t1)t

    u x twherex=x(t1);u=u(t1) =u(x(t1), t1).Then(droppingthetsubscript)

    J J (x,t)=min g(x,u) + (x,t)f(x,u,t)t u x

    Copyright2007StanleyB.Gershwin.c 41

    Continuous x t

  • 8/22/2019 MIT2 852S10 Control

    42/108

    Bellmans Continuousx,tEquation DeterministicThisistheBellmanequation. Itisthecounterpartoftherecursionequationforthenetworkexample.IfwehadaguessofJ(x,t)(forallxandt)wecouldconfirmitbyperformingtheminimization.

    IfweknewJ(x,t)forallxandt,wecoulddetermineubyperformingtheminimization.Ucouldthenbewritten

    Ju=U x, , t .

    xThiswouldbeafeedbacklaw.

    TheBellmanequationisusuallyimpossibletosolveanalyticallyornumerically.Therearesomeimportantspecialcasesthatcanbesolvedanalytically.Copyright2007StanleyB.Gershwin.c 42

    Continuous x t

  • 8/22/2019 MIT2 852S10 Control

    43/108

    Bellmans Continuousx,tEquation ExampleBang-BangControl

    min |x|dt0

    subjecttodx

    =udtx(0)specified1u1

    Copyright2007StanleyB.Gershwin.c 43

    B Continuous x t

  • 8/22/2019 MIT2 852S10 Control

    44/108

    Bellmans Continuousx,tEquation ExampleTheBellmanequationis

    J J (x,t)= min |x|+ (x,t)u .t xu,

    1u1J(x,t) =J(x)isasolutionbecausethetimehorizonisinfiniteandtdoesnotappearexplicitlyintheproblemdata(ie,g(x) =|x|isnotafunctionoft.Therefore

    dJ0= min |x|+ (x)u .

    dxu,1u1

    J(0)=0becauseifx(0)=0wecanchooseu(t) =0forallt.Thenx(t) = 0foralltandtheintegralis0.Thereisnopossiblechoiceofu(t)thatwillmaketheintegrallessthan0,sothisistheminimum.Copyright2007StanleyB.Gershwin.c 44

    B ll Continuous x t

  • 8/22/2019 MIT2 852S10 Control

    45/108

    Bellmans Continuousx,tEquation Example

    Theminimumisachievedwhen

    u=

    Why?

    1 if dJ(x)>0dx

    1 if dJ(x)

  • 8/22/2019 MIT2 852S10 Control

    46/108

    Bellmans Continuousx,tEquation Example

    ConsiderthesetofxwheredJ/dx(x)0andu=1,dJ

    (x) =|x|dx

    Copyright2007StanleyB.Gershwin.c 46

    B ll Continuous x t

  • 8/22/2019 MIT2 852S10 Control

    47/108

    Bellmans Continuousx,tEquation Example

    Tocompletethesolution,wemustdeterminewheredJ/dx>0,0forallx 0because|x|>0sotheintegralof|x(t)|mustbepositive.=SinceJ(x)> J(0)forallx 0,wemusthave=

    dJ(x)0dx

    Copyright2007StanleyB.Gershwin.c 47

    B ll Continuous x, t

  • 8/22/2019 MIT2 852S10 Control

    48/108

    Bellmans Continuousx,tEquation Example

    ThereforedJ

    (x)>=xdxso

    1J = x22

    and

    1 if x 0

    Copyright2007StanleyB.Gershwin.c 48

    Continuous x, t,Discrete Bellmans

  • 8/22/2019 MIT2 852S10 Control

    49/108

    Continuousx,t,DiscreteStochastic

    T

    BellmansEquationJ(x(0), (0),0)=minE g(x(t), u(t))dt+F(x(T))

    u 0

    suchthatdx(t)

    =f(x,,u,t)dt

    prob [(t+t) = = =ijtforalli,j,ii|(t) j] =jx(0), (0)specified

    h(x(t), (t), u(t))0Copyright2007StanleyB.Gershwin.c 49

    Bellmans Continuous x, t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    50/108

    BellmansEquation

    Continuousx,t,DiscreteStochastic

    GettingtheBellmanequationinthiscaseismorecomplicatedbecausechangesbylargeamountswhenitchanges.LetH()besomefunctionof.Weneedtocalculate

    EH((t

    +

    t))

    =

    E

    {H((t

    +

    t))

    |(t)}

    =

    H(j)prob {(t+t) =j |(t)}

    j

    Copyright2007StanleyB.Gershwin.c 50

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    51/108

    Bellman sEquation

    Co t uous , , sc ete

    Stochastic

    = H(j)j(t)t+H((t))1 j(t)t+o(t)j j=(t)=(t)

    = H(j)j(t)t+H((t)) 1 +(t)(t)t +o(t)j=(t)

    E{H((t+t))|(t)}=H((t))+

    H(j)j(t)t+o(t)

    jWeusethisinthederivationoftheBellmanequation.

    Copyright2007StanleyB.Gershwin.c 51

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    52/108

    Bellman sEquation

    , ,

    Stochastic

    TJ(x(t), (t), t)= min E g(x(s), u(s))ds+F(x(T))

    u(s), tts

  • 8/22/2019 MIT2 852S10 Control

    53/108

    Bellman sEquation

    t+t= min E g(x(s), u(s))ds

    , ,

    Stochastic

    u(s),

    t0st+t

    T

    +

    min

    E g(x(s), u(s))ds+

    F

    (x(T

    ))

    u(s), t+t

    t+tsT

    Copyright2007StanleyB.Gershwin.c 53

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    54/108

    Bellman sEquation Stochastic

    = minu(s), E

    t+tt g(x(s),u(s))ds

    tst+t

    +J(x(t+t), (t+t), t+t)

    Next,

    we

    expand

    the

    second

    term

    in

    a

    Taylor

    series

    about

    x(t).

    Weleave(t+t)alone,fornow.

    Copyright2007StanleyB.Gershwin.c 54

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    55/108

    Bellman sEquation Stochastic

    J(x(t), (t), t) =minE g(x(t), u(t))t+J(x(t), (t+t), t) +u(t)

    J J(x(t), (t+t), t)x(t) + (x(t), (t+t), t)t +o(t)

    x twhere

    x(t) =x(t+t)x(t) =f(x(t), (t), u(t), t)t+o(t)

    Copyright c 552007StanleyB.Gershwin.

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    56/108

    Bellman sEquation Stochastic

    UsingtheexpansionofEH((t+t)),J(x(t), (t), t)=min g(x(t), u(t))t

    u(t)

    +J(x(t), (t), t) + J(x(t),j,t)j(t)tj

    J J

    + (x(t), (t), t)x(t) + (x(t), (t), t)t +o(t)x t

    Wecancleanupnotationbyreplacingx(t)withx,(t)with,andu(t)withu.Copyright2007StanleyB.Gershwin.c 56

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    57/108

    Bellman sEquation Stochastic

    J(x,,t)=minu

    g(x,u)t+J(x,,t)+

    jJ(x,j,t)jt

    J

    J

    + (x,,t)x+ (x,,t)t +o(t)x t WecansubtractJ(x,,t)frombothsidesandusetheexpressionforxtoget...

    Copyright2007StanleyB.Gershwin.c 57

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    58/108

    Bellman sEquation Stochastic

    0=min g(x,u)t+ J(x,j,t)jtu j

    J J +

    x(x,

    ,

    t)f(x,

    ,

    u,

    t)t

    +

    t(x,,t)t+o(t)

    or,

    Copyright2007StanleyB.Gershwin.c 58

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    59/108

    Bellman sEquation Stochastic

    J

    (x,

    ,

    t) =

    J(x,

    j,

    t)j+

    t j

    Jmin

    g(x,

    u) +

    (x,

    ,

    t)f(x,

    ,

    u,

    t)

    u x

    Badnews: usuallyimpossibletosolve;Goodnews: insight.

    Copyright c 592007StanleyB.Gershwin.

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    60/108

    Bellman sEquation Stochastic

    Anapproximation:whenT islargeandf isnotafunctionoft,typicaltrajectorieslooklikethis:

    x

    t

    Copyright2007StanleyB.Gershwin.c 60

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    61/108

    e a s

    Equation StochasticThatis,inthelongrun,xapproachesasteady-stateprobabilitydistribution.

    Let

    J

    be

    the

    expected

    value

    of

    g(x,

    u),

    where

    u

    is

    theoptimalcontrol.Supposewestartedtheproblemwithx(0)arandomvariablewhoseprobabilitydistributionisthesteady-statedistribution.Then,forlargeT,

    T

    EJ =minuE 0 g(x(t), u(t))dt+F(x(T))JT

    Copyright2007StanleyB.Gershwin.c 61

    Bellmans Continuousx,t,Discrete

  • 8/22/2019 MIT2 852S10 Control

    62/108

    Equation StochasticForx(0)and(0)specified

    J(x(0), (0),0)JT +W(x(0), (0))or,moregenerally,forx(t) =xand(t) =specified,

    J(x,,t)J(Tt) +W(x,)

    Copyright2007StanleyB.Gershwin.c 62

    FlexibleManufacturing

  • 8/22/2019 MIT2 852S10 Control

    63/108

    ManufacturingSystemControlSinglemachine,multipleparttypes.x,u,dareN-dimensionalvectors.

    TminE g(x(t))dt

    subjectto: 0dxi(t)

    =ui(t)di, i= 1,...,Ndt

    prob((t+t) = 0|(t)=1)=pt+o(t)prob((t+t) =1|(t)=0)=rt+o(t)

    iui(t)(t); ui(t)0i

    x(0), (0)specified

    Copyright2007StanleyB.Gershwin.c 63

    FlexibleManufacturing

  • 8/22/2019 MIT2 852S10 Control

    64/108

    ManufacturingSystemControl

    Define() ={u| iiui}.Then,for= 0,1,J (x,,t) = J(x,j,t)j+t

    j

    Jmin g(x) + (x,,t)(ud)u() x

    Copyright2007StanleyB.Gershwin.c 64

    FlexibleManufacturing

  • 8/22/2019 MIT2 852S10 Control

    65/108

    ManufacturingSystemControl

    ApproximatingJ withJ(Tt) +W(x,)gives:J = (J(Tt) +W(x,j))j+

    jW

    min g(x) + (x,,t)(ud)u() xRecallthat

    j = 0...j

    Copyright2007StanleyB.Gershwin.c 65

    FlexibleManufacturing

  • 8/22/2019 MIT2 852S10 Control

    66/108

    ManufacturingSystemControl

    soJ = W(x,j)j+

    jW

    min g(x) + (x,,t)(ud)u() xfor= 0,1

    Copyright2007StanleyB.Gershwin.c 66

    FlexibleManufacturing

  • 8/22/2019 MIT2 852S10 Control

    67/108

    ManufacturingSystemControl

    Thisisactuallytwoequations,onefor= 0,onefor= 1.WJ =g(x) +W(x,1)rW(x,0)r (x,0)d,x

    for= 0,WJ =g(x) +W(x,0)pW(x,1)p+ min (x,1)(ud)

    u(1) xfor= 1.

    Copyright2007StanleyB.Gershwin.c 67

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    68/108

    ManufacturingSystemControl Technically,notflexible!

    Now,xanduarescalars,and(1)=[0,1/]=[0, ]

    dWJ =g(x) +W(x,1)rW(x,0)r (x,0)d,

    dxfor= 0,dW

    J =g(x) +W(x,0)pW(x,1)p+ min (x,1)(ud)0u dxfor= 1.

    Copyright2007StanleyB.Gershwin.c 68

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    69/108

    ManufacturingSystemControl

    Seebook,Sections2.6.2and9.3;seeProbabilityslides#91120.

    When= 0, u= 0.When= 1,ifdW 0, u= 0.dx

    Copyright c 692007StanleyB.Gershwin.

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    70/108

    ManufacturingSystemControl

    W(x,)hasbeenshowntobeconvexinx. IftheminimumofW

    (x,

    1)

    occurs

    at

    x

    =

    Z

    and

    W

    (x,

    1)

    is

    differentiable

    for

    all

    x,

    thendW 0x > Zdx

    Therefore,ifx < Z, u=,ifx=Z, uunspecified,ifx > Z, u= 0.Copyright c 702007StanleyB.Gershwin.

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    71/108

    Manufacturing

    SystemControlSurplus,orinventory/backlog:Productionpolicy: ChooseZ(thehedgingpoint)Then,if= 1,

    ifx < Z, u=,ifx=Z, u=d,ifx > Z, u=0;

    if= 0, u

    = 0.

    HowdowechooseZ?

    dx(t)=u(t)d

    dtCumulative

    Production and Demand production

    d t + Z

    hedging point Z

    surplus x(t)

    demand dt

    t

    Copyright2007StanleyB.Gershwin.c 71

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    72/108

    g

    SystemControl DeterminationofZ

    ZJ =Eg(x) =g(Z)P(Z,1)+ g(x) [f(x,0)+f(x,1)]dx

    inwhichP andf formthesteady-stateprobabilitydistributionofx.WechooseZ tominimizeJ. P andf aregivenby

    f(x,0)=Aebx

    f(x,1)

    =

    A

    debx

    d

    P(Z,1)=ApdebZ

    Copyright2007StanleyB.Gershwin.c 72

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    73/108

    g

    SystemControl DeterminationofZwhere

    r p

    b=

    d d

    andAischosensothat

    Z[f(x,0)+f(x,1)]dx+P(Z,1)=1

    Aftersomemanipulation,A= bp(d) ebZ

    db(

    d) +

    pand

    db(d)P(Z,1)=

    db(d) +pCopyright2007StanleyB.Gershwin.c 73

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    74/108

    g

    SystemControl DeterminationofZSinceg(x) =g+x+ +gx ,

    ifZ0,then ZJ =gZP(Z,1) gx[f(x,0)+f(x,1)]dx;

    ifZ >0,

    0

    J =g+ZP(Z,1) gx[f(x,0)+f(x,1)]dx

    Z+ g+x[f(x,0)+f(x,1)]dx.

    0

    Copyright2007StanleyB.Gershwin.c 74

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    75/108

    g

    SystemControl DeterminationofZTominimizeJ:

    ln Kb(1+g)

    ifg+Kb(g+ +g)

  • 8/22/2019 MIT2 852S10 Control

    76/108

    g

    SystemControl DeterminationofZThatis,wechooseZsuchthat

    ebZ =min 1, Kb g+ +gg+or

    bZ 1 g+e =max 1,Kb g+ +g

    Copyright2007StanleyB.Gershwin.c 76

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    77/108

    SystemControl DeterminationofZ

    0prob(x0) = (f(x,0)+f(x,1))dx

    0d=A 1 + ebxdx

    d d 1 =A 1 + =A

    d b b(d)bp(d) bZ = e

    db(d) +p b(d)p bZ= e

    db(d) +pCopyright2007StanleyB.Gershwin.c 77

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    78/108

    SystemControl DeterminationofZOr,

    prob(x0)=

    pmax 1,

    1 g+db(d) +p Kb g+ +g

    Itcanbeshownthatp

    Kb= p+bd(d)Therefore

    prob(x0)=Kbmax 1, 1 g+Kb g+ +gp g+

    =max ,p+bd(d) g+ +g

    Copyright2007StanleyB.Gershwin.c 78

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    79/108

    SystemControl DeterminationofZThatis,if

    p>

    g+,thenZ = 0andp+bd(d) g+ +g

    prob(x0)= p ;p+bd(d)

    if p < g+ ,thenZ >0andp+bd(d) g+ +g

    prob(x0)= g+ .g+ +gThislooksalotlikethesolutionofthenewsboyproblem.Copyright2007StanleyB.Gershwin.c 79

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    80/108

    Zvs.dSystemControlBasevalues:g+ = 1,g =10d=.7,= 1.,r=.09,p=.01.

    10090807060

    Z 50403020100

    d0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

    Copyright2007StanleyB.Gershwin.c 80

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    81/108

    SystemControl Zvs.g+Basevalues:g+ = 1,g =10d=.7,= 1.,r=.09,p=.01.

    70605040

    0 0.5 1 1.5 2 2.5 3g+

    Z3020

    100

    3.5

    Copyright2007StanleyB.Gershwin.c 81

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    82/108

    SystemControl Zvs.gBasevalues:g+ = 1,g =10d=.7,= 1.,r=.09,p=.01.

    14

    12

    10

    8

    Z

    6

    4

    2

    00 1 2 3 4 5 6 7 8 9 10 11

    g

    Copyright2007StanleyB.Gershwin.c 82

    Flexible Single-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    83/108

    SystemControl Zvs.pBasevalues:g+ = 1,g =10d=.7,= 1.,r=.09,p=.01.

    140012001000800

    Z600400

    2000

    0.005 0.01 0.015 0.02 0.025 0.03 0.035p0 0.04

    Copyright2007StanleyB.Gershwin.c 83

    Flexible Two-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    84/108

    SystemControlx1

    x2

    d1

    d2

    u (t)1

    u (t)2 Type 2

    Type 1

    r, p

    u11/1

    1/

    2

    u2

    0

    Capacityset(1)whenmachineisup.Copyright2007StanleyB.Gershwin.c 84

    Flexible Two-part-typecaseManufacturing

  • 8/22/2019 MIT2 852S10 Control

    85/108

    SystemControl

    Wemustfindu(x,)tosatisfyW

    min (x,,t) u

    u() xPartialsolutionofLP:IfW/x1 >0andW/x2 >0,u1 =u2 = 0.IfW/x1

  • 8/22/2019 MIT2 852S10 Control

    86/108

    1

    0

    2

    SystemControlCase:ExactsolutionifZ = (Z1, Z2) = 0x

    2

    x1

    1 2u = u = 0

    dx

    dt

    2

    2u = 0

    u = 01

    1u =

    1

    2

    u =

    u1

    1/

    1/2

    u2

    0

    u1

    1/1

    1/2

    u2

    1/1

    1/

    u2

    0 u1

    Copyright2007StanleyB.Gershwin.c 86

    Flexible Two-part-typecaseManufacturing

    S C l

  • 8/22/2019 MIT2 852S10 Control

    87/108

    1

    2

    0

    SystemControlCase:ApproximatesolutionifZ >0x

    2

    1 2u = u = 0

    dx

    dt

    2

    2u = 0

    u = 01

    1

    u =1

    2

    u =

    u1

    1/

    1/2

    u2

    0

    1/1

    1/

    u2

    0

    u1

    1/1

    1/2

    u2

    x1

    u1

    Copyright2007StanleyB.Gershwin.c 87

    Flexible Two-part-typecaseS t C t l

    Manufacturing

  • 8/22/2019 MIT2 852S10 Control

    88/108

    45

    SystemControlTwoparts,multiplemachineswithoutbuffers:

    e12

    4

    61e

    e

    34e

    23e

    3

    12

    x2

    Z

    6x

    1

    556e

    6

    1 2

    3

    4

    5

    u2

    e34

    12e

    u1

    Ie23

    e45

    e56

    d

    61e

    Copyright2007StanleyB.Gershwin.c 88

    Flexible Two-part-typecaseManufacturingS t C t l

  • 8/22/2019 MIT2 852S10 Control

    89/108

    SystemControlProposedapproximatesolutionformultiple-part,

    singlemachinesystem:Rankordertheparttypes,andbringthemtotheirhedgingpointsinthatorder.

    Copyright2007StanleyB.Gershwin.c 89

    Flexible Single-part-typecaseS t C t l

    ManufacturingS rpl s and tokens

  • 8/22/2019 MIT2 852S10 Control

    90/108

    SystemControl SurplusandtokensOperatingMachineMaccordingtothehedgingpointpolicyisequivalenttooperatingthisassemblysystem

    according

    to

    a

    finite

    B

    bufferpolicy.D

    M

    S

    FG

    Copyright2007StanleyB.Gershwin.c 90

    Flexible Single-part-typecaseS t C t l

    ManufacturingSurplus and tokens

  • 8/22/2019 MIT2 852S10 Control

    91/108

    SystemControl SurplusandtokensDisademandgenerator.

    Whenever a demand arrives, DsendsatokentoB.

    S isasynchronizationmachine. S is perfectly reliable and in

    finitelyfast.

    M

    D

    S

    FG

    B

    F Gisafinitefinishedgoodsbuffer.

    Bis

    an

    infinite

    backlog

    buffer.

    Copyright c 912007StanleyB.Gershwin.

    Flexible Single-part-typecaseManufacturingSystem Control Material/token policies

  • 8/22/2019 MIT2 852S10 Control

    92/108

    SystemControl Material/tokenpoliciesOperator

    AnoperationcannottakeMachineplace

    unless

    there

    is

    a

    Part Part tokenavailable.Operation

    Consumable Waste

    TokensauthorizeToken Token

    production.Thesepoliciescanoftenbeimplementedeither withfinite

    bufferspace,orafinitenumberoftokens.Mixturesarealsopossible.

    Bufferspacecouldbeshelfspace,orfloorspaceindicatedwithpaintortape.

    Copyright c 922007StanleyB.Gershwin.

    Multi-stage Proposedpolicysystems

  • 8/22/2019 MIT2 852S10 Control

    93/108

    systems

    TocontrolM B M B M

    1 1 2 2 3

    addaninformationflowsystem:B

    1M B M

    2 2 3M

    1

    S S2 3

    D

    S1

    BB1

    SB2

    BB2

    SB3

    BB3

    SB1

    Copyright2007StanleyB.Gershwin.c 93

    Multi-stage Proposedpolicysystems

  • 8/22/2019 MIT2 852S10 Control

    94/108

    systemsB

    1M B M

    2 2 3M

    1

    S S2 3

    D

    S1

    BB1

    SB2

    BB2

    SB3

    BB3

    SB1

    Bi arematerial buffersandarefinite.SBi aresurplus buffersandarefinite.BBi arebacklog buffersandareinfinite.ThesizesofBi andSBi arecontrolparameters.Problem: predictingtheperformanceofthissystem.Copyright c 942007StanleyB.Gershwin.

    Multi-stagesystems

    ThreeViewsofScheduling

  • 8/22/2019 MIT2 852S10 Control

    95/108

    systemsThreekindsofschedulingpolicies,whicharesometimes

    exactly

    the

    same.

    Surplus-based: makedecisionsbasedonhow

    muchproductionexceeddemand.Time-based: makedecisionsbasedonhowearlyor

    lateaproductis.Token-based: makedecisionsbasedonpresence

    orabsenceoftokens.Copyright c 952007StanleyB.Gershwin.

    Multi-stage ObjectiveofSchedulingsystems

    Surplusand

    time

  • 8/22/2019 MIT2 852S10 Control

    96/108

    Surplus and time

    and Demand

    earliness

    production P(t)

    demand D(t)

    surplus/backlog x(t)

    Objectiveistokeepcumulativeproductionclosetocumulativedemand.

    CumulativeProduction

    Surplus-basedpolicieslookatverticaldifferencesbetweenthegraphs.

    Time-basedpolicieslookatthehorizontal

    t

    differences.Copyright2007StanleyB.Gershwin.c 96

    Multi-stage Otherpoliciessystems

    CONWIPkanban

    and

    hybrid

  • 8/22/2019 MIT2 852S10 Control

    97/108

    CONWIP, kanban, and hybrid

    CONWIP: finitepopulation,infinitebufferskanban: infinitepopulation,finitebuffershybrid: finitepopulation,finitebuffers

    Copyright c 972007StanleyB.Gershwin.

    Multi-stage Otherpoliciessystems

    CONWIP,kanban,

    and

    hybrid

  • 8/22/2019 MIT2 852S10 Control

    98/108

    CONWIP, kanban, and hybrid

    CONWIPSupply Demand

    Token flow

    Demandislessthancapacity.How

    does

    the

    number

    of

    tokens

    affect

    performance

    (production

    rate,inventory)?

    Copyright c 982007StanleyB.Gershwin.

    Multi-stage Otherpoliciessystems

    CONWIP,kanban,

    and

    hybrid

  • 8/22/2019 MIT2 852S10 Control

    99/108

    0.85

    0.855

    0.86

    P

    CONWIP, kanban, and hybrid

    0.835

    0.84

    0.845

    0.865

    0.87

    0.875

    0 20 40 60 80 100 120

    0

    5

    10

    20

    25

    30

    AverageBufferLevel

    n1n2n3

    15

    0 20 40 60

    Population Population

    80 100 120

    cCopyright2007StanleyB.Gershwin. 99

    Multi-stage Otherpoliciessystems

    Basestock

  • 8/22/2019 MIT2 852S10 Control

    100/108

    Basestock

    Demand

    Copyright2007StanleyB.Gershwin.c 100

    Multi-stage Otherpoliciessystems

    FIFO

  • 8/22/2019 MIT2 852S10 Control

    101/108

    First-In,FirstOut.Simpleconceptually,butyouhavetokeeptrackofarrivaltimes.Leavesoutmuchimportantinformation:duedate,valueofpart,currentsurplus/backlog

    state,etc.

    Copyright2007StanleyB.Gershwin.c 101

    Multi-stage Otherpoliciessystems

    EDD

  • 8/22/2019 MIT2 852S10 Control

    102/108

    Earliestduedate.Easytoimplement.Doesnotconsiderworkremainingontheitem,value

    oftheitem,etc..

    Copyright2007StanleyB.Gershwin.c 102

    Multi-stage Otherpoliciessystems

    SRPT

  • 8/22/2019 MIT2 852S10 Control

    103/108

    ShortestRemainingProcessingTimeWheneverthereisachoiceofparts,loadtheonewithleastremainingworkbeforeitisfinished.Variations: includewaitingtimewiththeworktime.

    Useexpectedtimeifitisrandom.

    Copyright2007StanleyB.Gershwin.c 103

    Multi-stage Otherpoliciessystems

    Criticalratio

  • 8/22/2019 MIT2 852S10 Control

    104/108

    Widelyused,butmanyvariations.Oneversion:Processingtimeremaininguntilcompletion

    DefineCR= Duedate- CurrenttimeChoosethejobwiththehighestratio(provideditispositive).Ifajobislate,theratiowillbenegative,orthedenominator

    willbezero,andthatjobshouldbegivenhighestpriority.Ifthereismorethanonelatejob,schedulethelatejobsin

    SRPTorder.

    Copyright2007StanleyB.Gershwin.c 104

    Multi-stage Otherpoliciessystems

    LeastSlack

  • 8/22/2019 MIT2 852S10 Control

    105/108

    Thispolicyconsidersapartsduedate.Defineslack =duedate- remainingworktimeWhenthereisachoice,selectthepartwiththeleast

    slack.Variationsinvolvedifferentwaysofestimating

    remainingtime.

    Copyright c 1052007StanleyB.Gershwin.

    Multi-stage Otherpoliciessystems

    Drum-Buffer-Rope

  • 8/22/2019 MIT2 852S10 Control

    106/108

    DuetoEliGoldratt.

    Basedon

    the

    idea

    that

    every

    system

    has

    a

    bottleneck.

    Drum: thecommonproductionratethatthesystemoperates

    at,whichistherateofflowofthebottleneck.

    Buffer:

    DBRestablishes

    a

    CONWIP

    policy

    between

    the

    entranceofthesystemandthebottleneck.ThebufferistheCONWIPpopulation.

    Rope: thelimitonthedifferenceinproductionbetweendifferentstagesinthesystem.

    But:Whatifbottleneckisnotwell-defined?Copyright c 1062007StanleyB.Gershwin.

    Conclusions

  • 8/22/2019 MIT2 852S10 Control

    107/108

    Manypoliciesandapproaches.Nosimplestatementtellingwhichisbetter.Policiesarenotallwell-definedintheliteratureorinpractice.Myopinion:

    Thisisbecausepoliciesarenotderived fromfirstprinciples.Instead,theyaretestedandcompared.Currently,wehavelittleintuitiontoguidepolicydevelopment

    andchoice.

    Copyright2007StanleyB.Gershwin.c 107

    MIT OpenCourseWarehttp://ocw.mit.edu

    2.852 Manufacturing Systems Analysis

    Spring 2010

    http://ocw.mit.edu/http://ocw.mit.edu/
  • 8/22/2019 MIT2 852S10 Control

    108/108

    Spring 2010

    For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

    http://ocw.mit.edu/termshttp://ocw.mit.edu/terms