sequen e-dependent setups and utility workers
on an assembly line
May 22, 2009
Abstra t
This paperpresentsan integer programmingformulation for the
sequen -ingproblem in mixed-model assembly lines where the number of temporarily
hired utility workersand the numberof sequen e-dependent setups areto be
optimizedsimultaneouslythrougha ost fun tion. Theresultantmodeloers
an operational way to implement the utility work needed to avoid line
stop-pages, unlikeprevious papersaddressingthe goalof smoothingthe workload.
Thepresentresear hhasanimmediateappli ationtotheautomotiveindustry,
namelyto the ar-sequen ingproblem. Simulation resultsshow that the
pro-posed formulationleadsto theoptimumin areasonabletime forinstan esup
to15itemsandtosatisfa toryfeasiblesolutionsforsomeofthelargerproblems
we onsideredwithinamoderatetimelimit.
Keywords: Mixed-model assembly lines; Sequen ing; Spa ing onstraints;
Sequen e-dependentsetups;Utilityworkers;E onomi evaluation.
1 Introdu tion
Past resear h on the sequen ing problem in a mixed-model assembly line has
pri-marilyfo usedon two goals. Prior to the 1980's, themain goalwas (1): to smooth
the workload at ea h workstation through the line. This goal seeks to redu e line
stoppagesor ine ien ies like work ongestion or utility work (see Mitsumori,1969
or Xiaobo andOhno, 1997 for optimal models and for heuristi s, seeThomopoulos,
1967;Ma askill,1973;Sumi hrastetal.,1992;Smithetal.,1996andGottliebetal.,
2003).
TheemergingJIT on eptbythemid1980'shasraisedase ondgoalthat onsists
inkeepinga onstantrate ofpart usageto avoidlarge inventories. Goal (2)maybe
a hieved by: (2a)syn hronizing theprodu tion of ea hmodel withits demand and
(2b)levellingtheusageofea hpartatea hlevelofthemanufa turingpro ess. Ea h
goal(2a)and(2b)isusuallyexpressedasanobje tiveofminimizationofthevariation
(squareddeviations)ofthea tualprodu tion/usagefromthedesiredamount. Goals
(2a)and (2b)arestri tlyequivalent whenallprodu tsrequirethesamenumber and
mix of parts (see Miltenburg, 1989). Under this assumption, Goal (2)is optimally
Goal(2a)isoptimallysolvedbyKubiakandSethi,1991andbyBautistaetal.,2000;
heuristi methods tohandleGoal (2b)maybe foundinBautistaetal.,1996;Leuet
al.,1996 andMonden, 1998.
Goals(1)and(2)aresimultaneouslyaddressedinoptimalformulationsby
Kork-mazelandMeral,2001andbyDrexlandKimms, 2001 whoalsoprovide aheuristi
solutionapproa htotheproblembyusing olumn-generationte hniquestosolvethe
orrespondingLP-relaxation. Goals(1)and(2)areheuristi ally onsideredthrough
bi- riteriaapproa hesbyChoiandShin,1997;Tamuraetal.,1999 andbyKorkmazel
andMeral,2001.
Theadditionalobje tiveofminimizingtheset-up ost,namelyGoal(3),israrely
examined, possibly be ause it oni ts with both previous goals: a sequen e that
smooths the workloadneither ensuresa onstant rate ofpart usagenor a minimum
number of model swit hes (see Burns and Daganzo, 1997 for a dis ussion). A
heuristi approa hhandling both Goals(2)and(3)isdevelopedbyMansouri,2005.
Multi-obje tivemodelsin ludingthethreegoalsarequites ar e(Hyunetal.,1998).
The present paper onsiders Goals (1)and (3) inan optimal formulation ofthe
problemwheretheobje tivefun tionisa ostfun tionasso iatedwithhiring
tempo-rary/utilityworkerstoavoidlinestoppagesandwithmodelswit hesinthesequen e.
Theresear hpresentedherediersfromearlierstudiesinthesequen ingareain
sev-eral ways. First, as an e onomi ost fun tion is asso iated with the goals, the
hallenging task of assigning weights to the dierent obje tiveswhose measuring
unitsarenotidenti alisavoided(seeTamura etal.,1999orMurataetal.,1996for
adis ussiononmethodstosetweightstomulti-obje tive fun tions). Se ond,papers
thataddress the goal of smoothing the workload oftenexpress it asan obje tive of
minimizationofthetotalutilitywork(seeforinstan eHyunetal.,1998). Thislatter
obje tive onsistsinminimizingthenumberofhoursthatadditional workersshould
handletoavoidlinestoppagessonoindi ationonwhenandhowmanyworkersmust
temporarily be hired is provided. By ontrast, our formulation leads to assigning
utility workers to stations so as to pro ess some items in the sequen e. Spa ing
onstraintsareutilized hereto a ount for workstations apa itylimitations, likein
fewpapers (Smithet al.,1996; Choi and Shin, 1997; Drexl andKimms, 2001;
Got-tliebetal., 2003)although they aremore intuitive and operational than pro essing
time-basedrules.
Our work has a dire t appli ation to the ar sequen ing problem that onsists
in nding a sequen e of vehi les that minimizes the set-up osts while satisfying
apa ity ontraints along theassembly line. The set-up ostsof swit hingfrom one
model to another involve the ost of tuning up the ma hines and possibly the
ostof leaningspraygunsifthe olouris hanged. Assolvent pur hase onstitutesa
majorexpenseforautomotive bodyshops,thereisastrongin entive togroup
same- oloured vehi les to minimize the number of purges. In our experimental study,
we onsider this type of appli ation to generate instan es on whi h we apply the
proposedformulationthatissolvedusingthe ommer ialsoftwareMIPsolver
ILOG-Cplex9.0. Consideringthe optionof hiring utilityworkers insu h a ontext reates
The next se tion provides a general des ription of the sequen ing problem as
well as the usual onstraints relative to the demand satisfa tion and the apa ity
limitations. Se tion 3 presents the integer programming formulation designed to
a hieve Goals(1) and (3). Se tion 4 summarizes the optimization problemand the
assumptionswemade. These tionalsoin ludesausefultableofnotationsthereader
mayreferto. InSe tion 5wedes ribetheexperimentalframeworkandwe omment
thesimulation results. We on lude inSe tion 6.
2 Problem statement
Thisse tion isdedi ated to a general des ription of the sequen ing problem.
Con-straints relative to demand satisfa tion and apa ity limitations arethen provided.
Thenotationswewillusethroughout aresummarizedinTables8and9(seeSe tion
4).
2.1 General des ription
Inthe mixed-modelassembly line,some models of items have to be assembled over
ahorizon
T
(usuallya day) for whi ha total demand ofN
items must besatised. The y le timeθ
isdened asθ = ⌊T /N ⌋
whi h means that items arelaun hed in xedθ
time intervals to the assembly line.The line is partitioned into
K
stations of three types. SubsetK
C
groups the stationswith axed apa ity andsubsetK
F
is omposedof exible stationswhose apa ity an be extendedwithutility workers temporarily hired. Stationsthatnei-ther belongs to
K
C
norK
F
are those for whi h pro essing times never ex eed the y letime. Operationson su h stations involve doingexa tly thesame thing to allitems(installing safetybelts on allvehi lesis anexample).
A model or variant results from a ombination of options. Ea h option may
be fullypro essedonasinglestationor mayundergoseveraloperationsondierent
workstationstobe ompleted. Inthefollowing,wewill onsiderea hitemseparately
byletting
y
i,h
beabinaryvariablethattakesavalueof1
ifitemi
isinpositionh
in the sequen e,and0
otherwise. Ourin entive isto allowany ombinationofoptions forthe ustomersothespa eof hoi es isnot limitedto apredeterminednumber ofvariants.
Solving the sequen ing problem involvesnding an arrangement of the itemsin
thesequen e,saya
0 − 1
matrixY = {y
i,h
}
ofdimensionN × N
soastomeetseveral obje tiveslikeminimizingthenumberofutilityworkersandthesetups,levellingtheutilization rate, subje t to a set of onstraints relative to the demand satisfa tion
and apa ity limitations.
2.2 Usual onstraints: demand satisfa tion and spa ing
require-ments
Werstprovidethe onstraintsrelativetothedemandsatisfa tion. Wethengivethe
2.2.1 Demand satisfa tion
The following onstraints ensure that the demand is satised over horizon
T
(ea h itemis sequen ed) and thatexa tly one produ t is assigned to ea h position inthesequen e(see for instan eDrexl andKimms, 2001)
N
X
h=1
y
i,h
= 1, i = 1, . . . , N,
(1)N
X
i=1
y
i,h
= 1, h = 1, . . . , N.
2.2.2 General form of the spa ing onstraints
Spa ing onstraints are spe ied to avoidline stoppages. They di tate the spa ing
(i.e. thenumberofitemswithoutoption)betweentwo onse utiveitemswithoption.
Thespa ing onstraint
ν
o
: η
o
stipulates that `theremust be at mostν
o
items with optiono
inanysequen e ofη
o
onse utive items'.A spa ing onstraint for option
o
is required as soon as its pro essing time on one stationk
, denoted byθ
o,k
is su h thatθ
o,k
> λ
k
· θ
, whereλ
k
is the number of permanent workers dedi ated to stationk.
When the previous inequality holds for several stationsk,
the spa ing onstraint for optiono
is di tated by the most apa itated stationk
o
pro essing this optiono,
withk
o
∈ arg max
k∈K
C
∪K
F
{θ
o,k
},
and we usually have
ν
o
= λ
k
o
andη
o
= ν
o
/d
o
,
whered
o
is the average rate of demandfor optiono.
Consider for instan e one option
o = 1
and 3 stationsk = 1, 2, 3
. The y le time isθ = 60
se . The number of permanent workers on ea h station is given byλ
1
= 2; λ
2
= 4; λ
3
= 3
. We assume the demand rate for option 1 equal to 20%. Table1 providesfor ea h stationk
thepro essingtimeθ
1,k
andthevalueofλ
k
· θ
.Station
k
Pro ess. timeθ
1
,k
λ
k
· θ
Spa ing onstraintν
1
: η
1
1
127
120
2
200
240
3
190
180
3 : 15
Table1: Spa ing onstraintsforanoption
o
-illustrationTwo stations
k = 1
andk = 3
aresu h thatθ
1,k
> λ
k
· θ
, indeed fork = 1
we have127 > 120
andfork = 2,
we have190 > 180
. Amongthesetwostations,stationk
1
= 3
isthestationforwhi hθ
1,k
ismaximum. Withλ
3= 3
wehave
ν
1
= 3
. Sin ed
1= 1/5
then
η
1= 15
. For option 1, themost apa itated stationis station 3 and
thespa ing onstraintis
ν
1
: η
1
= 3 : 15
meaningthattheremustbeatmost3items withoption1 inany sequen eof 15items.When a station
k
provides the tightest spa ing onstraint for more than one option,theseoptions aregrouped inasubsetandthetightestofthetightest spa ingommon assumption that ea h option is fully pro essed on a single station only
dedi atedto the installation of this option. Thus, inthefollowing,wewill make no
dieren ebetween option
o
and thestationk
onwhi h theoptionistreated andwe willkeepthe notationk
to designatetheoption pro essedon stationk.
Forinstan e,followingthepreviousexample,ifase ondoption
o = 2
isalsosu h thatstation3isthemost apa itatedstationwithaspa ing onstraintν
2
: η
2
= 3 : 4
, wethuswill onsideroption1or2asasingle`arti ial'optiono
pro essedonstationk = 3
with a spa ing onstraint ofν
o
: η
o
= 3 : 15
(sin e this is tighter thanν
2
: η
2
= 3 : 4
).Letting
δ
i,k
beaparameter thattakesavalueof1
ifoptionk
isrequiredbyitemi
and0
otherwise, the spa ing onstraint is written ash
X
l=h−η
k
+1
N
X
i=1
δ
i,k
· y
i,l
≤ ν
k
, h = 1, . . . , N ; k ∈ K
C
∪ K
F
.
(2)Thespa ing onstraint(2)appliesforanyexiblestationaslongasnoutilityworker
operateson this station. This onstraint an alsoiteratively be dened as
m
k,h
= m
k,h−1
+
P
N
i=1
δ
i,k
· y
i,h
−
P
N
i=1
δ
i,k
· y
i,h−η
k
≤ ν
k
,
h = 1, . . . , N ; k ∈ K
C
∪ K
F
,
(3)
where
m
k,h
isthetotal numberofitemswithoption re orded intheset ofpositions (or window){h − η
k
+ 1, . . . , h} .
For thesake of larity, we will use expression (3) whenwe adaptthespa ing onstraint to the ase of exiblestations hosting utilityworkers (see paragraph3.1.1, p. 8).
As an illustration, onsider
N = 6
items to be produ ed during horizonT
. Four models are produ ed:A, B, C, D
and modelsB
andD
require optionk = 1
. The spa ing onstraint isν
1
: η
1
= 1 : 2
. Table 2 provides a sequen e of models, whereh = −1, 0
are negative indi es hosen for theprevioushorizonT − 1
. Indi esh = 1, ..., 6
represents items'position during horizonT.
For ea h positionh,
the value ofP
N
i=1
δ
i,k
· y
i,h
is simply equal to 1 if item in positionh
requires optionk
(withk = 1
inour example)and zerootherwise. Constraint(2) onsistsin ounting the number of optionsk
in any window ofη
k
items and he king if this number
is inferior to
ν
k
.
In onstraint (3), the number of items withoption inthe window{h − η
k
+ 1, . . . , h}
equalsthe numberof items withoption intheprevious window to whi h we add the last element of the new window and from whi h we subtra ttherstelementoftheformer window, aswindows roll. Considering
h = 1
,we have itemD
inthe window{0, 1}
thusm
1,1
= 1
. For
h = 2,
we onsider window{1, 2}
, we addP
N
i=1
δ
i,k
· y
i,2
= 1
(last element of the new window) tom
1,1and sustra t
P
N
i=1
δ
i,k
· y
i,0
= 0
(rstelement oftheformer window)to obtainm
1,2
= 2.
Notethe
Horizon
T − 1
T
h
...−1
0
1
2
3
4
5
6
Item
D
C
D
B
A
B
D
C
P
N
i=1
δ
i,k
· y
i,h
1
0
1
1
0
1
1
0
Table2: Che kingthespa ing onstraints
2.2.3 Derivingthespa ing onstraintforstationswithasinglededi ated
worker
For the stations designed to host a single permanent worker
(λ
k
= 1),
spa ing on-straints anbederived from apa itylimitationsdened intermsofoperation time.We note
θ
i,k
the pro essing timeof itemi
on stationk.
The operation time of theh
th
item on stationk
isP
N
i=1
θ
i,k
· y
i,h
.
We an reasonably assume that there are only two possible values for thepro essing time on stationk
:θ
max
k
andθ
min
k
whi h respe tively designates the operation time when optionk
is installed on any itemi
and when it is not, withθ
min
k
< θ < θ
max
k
. Th e pro essing time an therefore be written asP
N
i=1
θ
i,k
· y
i,h
= θ
max
k
P
N
i=1
δ
i,k
· y
i,h
+ θ
k
min
P
N
i=1
(1 − δ
i,k
) · y
i,h
,re allingδ
i,k
= 1
ifitemi
hasoptionk
and zero otherwise. Utilizing onstraint (1), we haveP
N
i=1
θ
i,k
· y
i,h
= (θ
max
k
− θ
k
min
)
P
N
i=1
δ
i,k
· y
i,h
+ θ
k
min
. Thus, ifitemi
has optionk,
itspro essing time onstationk
equalsθ
max
k
;itequalsθ
min
k
otherwise. LetR
max
k
be thedieren e between the total timededi ated to operations per-formed on stationk
andθ.
This represents the extra time that an be spent on stationk, ompared to the averageθ.
To pro ess oneitemwithoption onstationk,
this extratimeR
max
k
mustbe su hthatR
max
k
≥ θ
k
max
− θ.
The ex ess workingtime omparedto the y letime,R
k,h
,
after ompletion onstationk
ofoperationsontheh
th
itemis given byR
k,h
= max
(
R
k,h−1
+ (θ
k
max
− θ
min
k
)
N
X
i=1
δ
i,k
· y
i,h
+ θ
min
k
− θ, 0
)
.
(4)The apa ityof station
k
isviolated bytheiteminpositionh
assoonasR
k,h
>
R
max
k
andR
k,h−1
≤ R
max
k
, ∀ k ∈ K
C
∪ K
F
| λ
k
= 1
andh = 1, . . . , N.
To illustrate, onsidera station
k
hosting asinglededi ated workerwithθ
max
k
=
80
se .,θ
min
k
= 50
se .,θ = 60
se . andR
max
k
, theextra timethat an be spent on stationk
isR
max
k
= 40
se . Asoptionk
onsumesanadditionaltimeofθ
max
k
−θ = 20
se . then 2 items withoption
k
may be produ ed onse utively. To at h up these40
extra se onds, it is then ne esary to produ e 4 items without option, sin e the underuseof su h itemson stationk
isθ − θ
min
k
= 10
se .More formally, let us onsider the situation in whi h option
k
is required by the rst itemin the sequen e(
P
N
i=1
δ
i,k
· y
i,1
= 1)
andR
k,0
= 0
. From (4) we haveR
k,1
= θ
k
max
− θ.
Ase onditemwithoptionmayfollowifR
k,2
= 2(θ
max
k
− θ) ≤ R
max
k
. Finally, the maximum numberν
k
of onse utive items with optionk
is su h thatν
k
=
$
R
max
k
θ
max
k
− θ
%
.
(5)After
ν
k
onse utive items withoption, we haveR
k,ν
k
= R
max
k
and a numberµ
k
of itemswithoutoption mustfollowbeforesequen ingon eagainν
k
onse utive items requiringtheoption. Thisnumberµ
k
issu hthat(θ − θ
min
k
)µ
k
≥ (θ
k
max
− θ)ν
k
whi h nallyleads toµ
k
=
&
ν
k
·
θ
max
k
− θ
θ − θ
min
k
'
.
(6)The spa ing onstraint (2) or (3) still applies with
ν
k
as determined by (5) andη
k
= ν
k
+ µ
k
withµ
k
dened by(6)forall stationsk
su hthatλ
k
= 1
. Itshouldbe notedthatthisspa ing onstraintisane essaryandsu ient onditionnottoex eedthe apa ityonlywhen
ν
k
= 1
orequivalently whenR
max
k
< 2·(θ
max
k
− θ).
Otherwise, thespa ing onstraintisonlyasu ient ondition: awindow{h − η
k
+ 1, . . . h}
may ontainmorethanν
k
itemswiththeoptionwithoutviolatingthe apa ity onstraint ofthe station,R
k,l
≤ R
max
k
for alll ∈ {h − η
k
+ 1, . . . h}
,withR
k,l
dened by (4).Letus onsideragaintheexampleofastation
k
hostingasinglededi atedworker withθ
max
k
= 80
se .,θ
min
k
= 50
se .,θ = 60
se . andR
max
k
= 40
se . From (5) and (6), we getν
k
= 2
andµ
k
= 4
leading toη
k
= 6.
The spa ing onstraint states ina stri t sense that after 2 onse utive items withoption, 4 items without optionmustbesequen edsoasto geta zeroex eeding pro essingtimewhi h allows again
for the sequen ing of 2 onse utive items with option and so on. The su ient
onditionnottoex eedthe apa itylimitationistosequen eatmost2vehi leswith
optionout of6 vehi les. Table3 displays a sequen eof items for whi h the spa ing
onstraint an be he ked for
h = 6
andh = 7
. Forh = 6,
onstraint (2) givesP
6
l=1
P
N
i=1
δ
i,k
· y
i,l
= 2 ≤ 2.
Forh = 7,
we getP
7
l=2
P
N
i=1
δ
i,k
· y
i,l
= 2 ≤ 2.
Note the values ofR
k,l
never ex eedR
max
k
= 40,
meaning that the apa ity limitation is not violated. However, pla ing an itemwithoption inposition7
wouldhave led toR
k,7
= 40 ≤ R
max
k
,
witha numberof3
items withoption in thewindow{2, . . . , 7}.
Thisillustratesthe fa t thatthe spa ing onstraint isa su ient ondition to meetthe apa ity onstraint but it isnot ane essary ondition when
ν
k
> 1.
Position
l
1
2
3
4
5
6
7
P
N
i=1
δ
i,k
· y
i,l
0
1
0
0
0
1
0
R
k,l
10
30
20
10
0
20
10
Table3: Spa ing onstraint with no utility worker
3 Optimal formulation of the sequen ing problem
Theobje tivefun tion onsideredhereisa ostfun tioninvolvingtwoelements: the
without entailing line stoppages, somore grouping possibilities areavailable, hen e
redu ing thenumberof setups.
Werstprovidethe analyti aldes riptionoftheproblemofoptimizingthe
num-berofutilityworkerstobetemporarilyhired. Wethenturntotheformaldes ription
oftheproblemrelativeto theminimization ofthenumberofsetupsinthesequen e.
3.1 Obje tive and onstraints relative to utility workers
Paststudieshave onsideredthe obje tiveofminimizingtotalutilityworkexpressed
asaglobalworkingtimeandminimizedassu h,withoutevaluatingitse onomi
im-pa tasnowage ostsareintrodu ed (seeforinstan eHyunetal. 1998). Wehandle
theobje tive dierently: our in entive is to minimize thenumberof utilityworkers
required to implement sequen ing de isions taken for horizon
T
. Theresultant for-mulationgivesthe optimal assignment of utilityworkers on stationsand items. Werst provide an adaptation of onstraints (3) to the ase of exible stations (from
the set
K
F
)
. We thenderive theoptimalnumberof utilityworkers.3.1.1 Spa ing onstraints forstations with utility workers
Weassumethatea hexiblestationhosts asinglededi atedworker
(λ
k
= 1)
sothe spa ing onstraintν
k
: η
k
is dened by using Eq. (5) to determineν
k
and Eq. (6) togetthe valueofη
k
= ν
k
+ µ
k
. Ea hexible stationk ∈ K
F
is on eivedto hosta singleutilityworker by y le time.Let
w
k,h
beabinaryvariablethattakesavalueof1
ifautilityworkerisrequired onstationk
forthe iteminpositionh
and0
otherwise. Weassumethat workers do not ollaborate on the same task: the item in positionh
is fully pro essed by the utilityworkerwhiletheregularworkernishestheoperationsonthe(h − 1)
th
item.
A utility worker is required on station
k
to pro ess the item in positionh
in orderto avoida violationof apa itywhi ho urswhenR
k,h−1
+ θ
max
k
− θ > R
max
k
. Theitemin positionh
isne essarilyan itemwithoption (see Eq. (4)). We assume the regular worker at hes up on thetotal ba klogR
k,h−1
(while theutility worker fullyhandlestheiteminpositionh
). Thisassumptionmakespossibleanadaptation of the spa ing onstraints otherwise a reasoning in terms of ex ess working timemustbeadopted andspa ing onstraintsmust be abandoned. Formally,we assume
R
k,h−1
≤ θ
(the ex ess working time an not be superior to the `average' timefor produ inganitem). AsastraightforwardupperboundforR
k,h−1
equals(θ
max
k
−θ)ν
k
,
wenallyassume
(θ
max
k
− θ)ν
k
≤ θ.
Thisimpliesthatthepresen eofautilityworker on stationk
for item inpositionh
leads to a redu tion of the ex ess working time by an amount of(θ
max
k
− θ)(ν
k
+ 1)
where(θ
max
k
− θ)ν
k
is handled by the regular worker andθ
max
k
− θ
isabsorbed bytheutilityworker himself. Thisisequivalentto ` an elling'ν
k
+ 1
itemswithoptioninthe window{h − η
k
+ 1, . . . , h}
sin ethenext iteminpositionh + 1
maybe hosenasiftherewasnoitemwithoptioninpositionsh − η
k
+ 1
toh.
To illustrate thereasoning, onsider a station
k
withθ
max
k
= 80
se .,θ
min
k
= 40
se .,
θ = 60
se . andR
max
aught up. The regularworker handlesthis rstitemand needsto work 20se onds
morethanthe y letimeto omplete theitem. Ifase onditemwithoptionistobe
pro essed, the umulated ex ess working time would be
40
se onds, whi h is twi e the extratime thatis allowed. A utility worker istherefore ne esary to handle thisse onditem, sowe have
w
k,2
= 1
. Theutilityworkerfullypro esses this item, thus absorbinganextratimeof20se . whiletheregularworker annishhisworkontherst item, therefore absorbing also an extratime of 20 se onds. A third itemwith
option anthereforebepro essed, asthepresen eoftheutilityworkernallyallows
forabsorbingthe equivalent ofthe extratime neededtopro ess 2itemswithoption
(oneextratimeisdonebytheutilityworkerhimselfandtheotheronebytheregular
worker). Thepresen eoftheutilityworkerfor
h = 2
(se onditem)isthusequivalent to an ellingν
k
+ 1 = 2
itemswithoptioninthewindow{h − η
k
+ 1, . . . , h} = {1, 2}.
Thus,toadapt onstraint(3),wemustde reasem
k,h
by(ν
k
+ 1) w
k,h
andrepla eP
N
i=1
δ
i,k
· y
i,h−η
k
with a proper binary variable, namelyq
k,h−η
k
,
that takes into a ount the fa t that a possible utility worker has led to the ` an ellation' of theoption,ifany,on the
(h − η
k
)
th
item. Letting
q
k,h−η
k
be su ha binaryvariable,the spa ing onstraints forexible stations an nowbe written asm
k,h
= m
k,h−1
+
P
N
i=1
δ
i,k
· y
i,h
− q
k,h−η
k
− (ν
k
+ 1) w
k,h
≤ ν
k
,
h = 1, . . . , N ; k ∈ K
F
,
(7)
where
q
k,h−η
k
takesa value of1
ifitem inpositionh − η
k
has theoption and none ofthe items inpositions{h − η
k
, . . . , h − 1}
was handledbya utility worker, and0
otherwise. Formally,thevariableq
k,h−η
k
orrespondsto thefollowing denitionq
k,h−η
k
=
N
X
i=1
δ
i,k
· y
i,h−η
k
h−1
Y
l=h−η
k
(1 − w
k,l
).
(8)Thisexpression islinearizedusing thethreeinequalities
q
k,h−η
k
≤
P
N
i=1
δ
i,k
· y
i,h−η
k
,
q
k,h−η
k
≥
P
N
i=1
δ
i,k
· y
i,h−η
k
−
P
h−1
l=h−η
k
w
k,l
,
2q
k,h−η
k
≤ 1 +
P
N
i=1
δ
i,k
· y
i,h−η
k
−
P
h−1
l=h−η
k
w
k,l
.
(9)
To illustrate, onsider the example in Table 4 for a station
k ∈ K
F
andν
k
=
µ
k
= 1
sothespa ing onstraint isν
k
: η
k
= 1 : 2.
Table4 providesthe value ofthe variables in luded inEq. (7)for asequen e of3
itemsea h requiringtheoption.For
h = 1
onstraint (3)givesm
k,1
=
P
N
i=1
δ
i,k
· y
i,1
− 2w
k,1
,
sin em
k,0
andq
k,−3
have negative indi es, they are ignored. Assuming there is no initial ondition, wesetto zeroallvariables withnegative indi esinEq. (7). However, initial onditions
(i.e. past de isions) will be taken into a ount in theexperiment sothis point will
be dis ussed further in Se tion 5. We have
m
k,1
= 1 ≤ ν
k
= 1
withw
k,1
= 0.
Notethatw
k,1
ould have been equal to 1without violatingthespa ing onstraint. However,the obje tiveistominimizethe ostasso iatedwithtemporary workerssoPosition
h
1
2
3
P
N
i=1
δ
i,k
· y
i,h
1
1
1
m
k,h
1
0
1
w
k,h
0
1
0
q
k,h−η
k
/
/
0
Table 4: Spa ing onstraintforaexiblestation
k ∈ K
F
hostingautilityworkervariables
{w
k,h
}
will always take their minimum value while satisfying the spa ing onstraints. Forh = 2,
we havem
k,2
= m
k,1
+
P
N
i=1
δ
i,k
· y
i,2
− 2w
k,2
= 0
withm
k,1
= 1
andw
k,1
= 1.
A utility worker is required to pro ess the se ond item as the spa ing onstraint is violated sin e items in position 1 and 2 both requirethe option. For
h = 3,
we getm
k,3
= m
k,2
+
P
N
i=1
δ
i,k
· y
i,3
− q
k,1
− 2w
k,3
withq
k,1
=
P
N
i=1
δ
i,k
· y
i,1
Q
2
l=1
(1 − w
k,l
)
as dened by Eq. (8). We haveq
k,1
= 0
sin e therstitemhastheoptionbutautilityworkerwaspresentfor these onditemandthis leads to ` an elling' the option on item 1 and 2. We have
m
k,3
= 1 ≤ 1
withw
k,3
= 0.
3.1.2 Optimal number of utility workers
In this paragraph we rst provide a formulation of the obje tive fun tion relative
to the ost of hiring temporary workers for horizon
T
whi h is usually a day of produ tion. Thisobje tivefun tionin ludesvariablesrepresentingthetotalnumberofutilityworkershiredperhalf-days. Wethenderivesuitable onstraintstolinkthese
variables to the binary variables
{w
k,h
}
ree ting the presen e (or the absen e) of utility workers onsome station anditem.Let
W
1
(T )
bethenumberofa tivetemporaryworkersinthersthalfofhorizonT.
Analogously,W
2
(T )
designates thenumber ofutilityworkersrequired to pro ess items in the se ond half of horizonT.
We nally deneW
3
(T )
as the number of utility workers thatareneededto implement sequen ing de isions made for horizonT
butphysi allypresentonlyinthenexthorizonT + 1
. Theobje tiveistominimizeω · [W
1
(T ) − W
3
(T − 1) + W
2
(T ) + W
3
(T )] ,
(10)where
ω
is the ost of hiring a utility worker for one half-day, and the dieren eW
1
(T ) − W
3
(T − 1)
represents the number of utility workers resulting from hiring de isions made forT
and a tually present during the rst half of horizonT.
The e onomi fun tion (10) therefore refers to hiring de isions aimed at implementingthe sequen ing of items for horizon
T
only. We shall now spe ifythe relationships between variables{w
k,h
}
and variablesW
1
(T ), W
2
(T ), W
3
(T ).
Anytemporary workerrequiredonstation
k
hasa workingtimeofθ
max
k
on ea h itemrequiringtheoption. Thisimpliesthatheisassignedtostationk
foranumber of y letimesequalstoβ
k
=
θ
max
k
/θ
.
Thusautilityworkerthathandlesonstationk
theh
th
iteminthe sequen ebe omesanewavailable forperformingoperationson
the
(h + β
k
)
th
item, ifneed be, eitheron the same station or on any other station,
β
k
≥ 2
sin eθ
max
k
> θ.
From the assumption(θ
max
k
− θ)ν
k
≤ θ
we made in the previous paragraph, we getθ
max
k
/θ ≤ 1 + 1/ν
k
with1 + 1/ν
k
taking a maximum value of2
whenν
k
= 1
. We therefore haveβ
k
= 2, ∀k.
Whenν
k
> 1,
it should be notedthat the utility worker spends lessthan2
y le times on stationk
sothe remaining time an be onsidered as a free time for moving from one station toanother.
Re allthehorizon isdividedinto
N
y letimes. During theh
th
y letime,item
inposition
h − k + 1
ispro essedonstationk.
Forinstan e,withnoinitial ondition,N = 6
andK
F
= {1, 2},
in the rst y le time, the rst item in the sequen e is pro essed on station1
and no item is pro essed on station2.
In the se ond y le time, the se ond item is pro essed on station 1 while item 1 rosses station 2 et .Asautility workerstays
β
k
= 2
y letimesonanystationstationk
ea h timehe is required, the sumw
k,h−k
+ w
k,h−k+1
givesthe number of utility workers operating on stationk
in theh
th
y le time. The sum
P
k∈K
F
(w
k,h−k
+ w
k,h−k+1
)
therefore representsthe total numberof a tiveutilityworkerson all stationsintheh
th
y le.
Itfollows that thenumberof utilityworkers neededinthe rsthalf ofhorizon
T
isW
1
(T ) = max
h=1,...,[N/2]
n
P
k∈K
F
(w
k,h−k
+ w
k,h−k+1
)
o
. A similar reasoning holds
for
W
2
(T )
andW
3
(T )
. Thelinearversionof the onstraintsasso iated withW
1
(T ),
W
2
(T ), W
3
(T )
isP
k∈K
F
(w
k,h−k
+ w
k,h−k+1
) ≤ W
1
(T ), h = 1, . . . ,
N
2
,
P
k∈K
F
(w
k,h−k
+ w
k,h−k+1
) ≤ W
2
(T ), h =
N
2
+ 1, . . . , N,
P
k∈K
F
P
h−k+1
l=h−k
l≤N
w
k,l
≤ W
3
(T ), h = N + 1, . . . , N + K − 1.
(11)Let us note that at the end of horizon
T,
part of the sequen e that has been de idedinT
isonly pro essed and ompleted inT + 1.
Thethird inequalityin(11) ree ts the hiring de isions referred to the sequen ing of these items although theorrespondingutilityworkerswill onlybe present atthebeginningofhorizon
T + 1.
3.1.3 Illustration
Let us onsider a simple example with
K = 3
stations, where stationsk = 1
andk = 3
areexibleandk = 2
isneitheraxed- apa itystationnoraexibleone. We assume thatN = 6
items are sequen ed during ea h horizonT − 1, T
andT + 1.
Three horizons are examined to larify the interdependen e between past, presentand future de isions. With two options (
k = 1
,k = 3)
, 4 models of items are available: itemA
has no option, itemB
requires optionk = 1
only; optionk = 3
is ex lusively installed on itemC
and itemD
in ludes both options. We assumeν
1
= ν
3
= 1
andµ
1
= 1, µ
3
= 2.
Table 5 illustrates the satisfa tion of onstraints (7) for stations
k = 1
andk = 3,
for ea hpositionh
whereh = −5, . . . , 0
are thepositions of items for whi h sequen ing de isions are made during horizonT − 1,
de isions forh = 1, . . . , 6
are made for horizonT,
et . We indeed adopt the onvention that negative positionsh ≤ 0
refer to past sequen ing de isions whereas positionsh > N
are relative to futuresequen ing. Table 5also displaysthevalues of{w
k,h
} .
For instan e, we havew
1,2
= 1
meaning that a temporary worker must pro ess on station1
the item in these ondpositioninthe sequen ewhi h isde ided duringhorizonT.
Thisitema modelB
in ludesoptionk = 1
justlikethepreviousone (amodelD
)whi hwould haveledtoaviolationofthespa ing onstraintν
1
: η
1
(i. e.1 : 2)
ifnoworker ould have temporarily been hired.P
N
i=1
δ
i,k
y
i,h
m
k,h
≤ν
k
= 1
w
k,h
q
k,h−η
k
Hor.
h
Modelk = 1
k = 3
k = 1
k = 3
k = 1
k = 3
k = 1
k = 3
−5
D
1
1
/
/
0
0
/
/
−4
A
0
0
1
/
0
0
/
/
T −1
−3
B
1
0
1
1
0
0
1
/
−2
D
1
1
0
1
1
0
0
1
−1
C
0
1
0
0
0
1
0
0
0
C
0
1
0
1
0
0
0
0
1
D
1
1
1
0
0
1
0
0
2
B
1
0
0
0
1
0
0
0
T
3
A
0
0
0
0
0
0
0
0
4
B
1
0
1
0
0
0
0
0
5
D
1
1
0
1
1
0
0
0
6
C
0
1
0
0
0
1
0
0
7
D
1
1
1
1
0
0
0
0
8
B
1
0
0
1
1
0
0
0
T +1
9
A
0
0
0
1
0
0
0
0
10
D
1
1
1
1
0
0
0
1
11
B
1
0
0
1
1
0
0
0
12
A
0
0
0
1
0
0
0
0
Table 5: Spa ing onstraintsfortwof lexiblestations
Table 6showsfor ea h y letimethe items positions thatarepro essedon ea h
station. For instan e, itemin position
h = −5
ispro essed on stationk = 1
inthe rst y le time of horizonT − 1.
This item leads over stationk = 2
in the se ond y letime. It undergoesoperations on stationk = 3
in the third y letime and is nally ompletedinthefourth y letime,asindi atedinthepenultimate olumnofTable6. Re allthata utilityworkerrequired onstation
k
toperform operations on theh
th
itemisbusyfor
2
y letimes,∀k = 1, 3.
Thea tivityofevery utilityworker is represented by a re tangle in Table 6. As an example,w
1,−2
= 1
is symbolized byare tanglearoundpositions−2
and−1
indi atingthatthe orrespondingutility worker is busy on station1
inthe fourthand fth y letime of horizonT − 1.
We alsoreportinea hre tanglethehorizoninwhi hthehiringde isionhasbeen taken.Withsu ha table,we anvisualize thea tivity ofutilityworkers inthetimesothe
number of ne essary workers an easily be derived. The last olumn provides the
numberof a tive workers per y le time. Considering horizon
T,
one utility worker is a tive on workstationk = 3
in the rst y le time and he is engaged for2
y le times. Thismeans that this worker an not beallo ated to station1
inthe se ond y le time to perform operations on the item in position2,
thus2
utility workers mustbe a tive inthe se ond y letime.Horizon Cy le
k = 1
k = 2
k = 3
item(pos.) u. workers1
−5
0
2
−4
−5
0
T − 1
3
−3
−4
−5
0
4
−2 (T − 1)
−3
−4
−5
1
5
−1
−2
−3
−4
1
6
0
−1
−2
−3
0
1
1
0
−1 (T − 1)
−2
1
2
2 (T )
1
0
−1
2
T
3
3
2
1 (T )
0
2
4
4
3
2
1
1
5
5 (T )
4
3
2
1
6
6
5
4
3
1
1
7
6
5
4
0
2
8 (T + 1)
7
6 (T )
5
2
T + 1
3
9
8
7
6
2
4
10
9
8
7
0
5
11 (T + 1)
10
9
8
1
6
12
11
10
9
1
12
11
10
12
11
12
Table6: Itemspositionsonea hstationanda tivityofutilityworkers
Applying onstraints (11) to our example yields
w
1,0
+ w
1,1
+ w
3,−2
+ w
3,−1
≤
W
1
(T )
forh = 1,
wherethevalueofthesumw
1,0
+w
1,1
+w
3,−2
+w
3,−1
anbereadin thelast olumn ofTable6intherowrelative totherst y letimeofhorizonT.
We getW
1
(T ) ≥ 1, W
1
(T ) ≥ 2, W
1
(T ) ≥ 2
forh = 1, 2, 3
respe tively. Thisleads to an optimal valueW
∗
1
(T ) = 2,
assuming thesequen e ofitems given forT
inTable5 is anoptimalone given the ostparameters. AsW
1
(T )
isto be minimized,ittakesits minimumpossiblevalue. SimilarlyweobtainW
∗
2
(T ) = 1.
Thethirdinequalityin(11) appliesforh = 7, 8.
Wehavew
1,6
+ w
3,4
+ w
3,5
≤ W
3
(T )
forh = 7
andw
3,5
+ w
3,6
≤
W
3
(T )
forh = 8.
NotethatW
3
(T )
isonlylinked tovariables{w
k,h
}
asso iatedwith itemspositionsthatdonotex eedN = 6.
WehaveW
∗
3
(T ) = 1.
Theoptimalnumber of utility workers whose e onomi onsequen e is attributable to the sequen ingde isions made for horizon
T
equalsW
∗
1
(T ) − W
3
∗
(T − 1) + W
2
∗
(T ) + W
3
∗
(T ) = 3.
In the rst half of
T
,2
operators are required but one of them has been de ided inT − 1
so its wage is in luded in thee onomi fun tion related to de isions made for horizonT − 1.
A single operator is required inthese ond half-dayand a utility worker thatwill be present only inthenext dayis a tually paidinhorizonT.
Many operationsne essitate sequen e-dependent setups. For instan e, inthe
auto-motive industry, vehi les of a same olour are grouped in an attempt to minimize
the ost of leaningsprayguns. We will fo us here on su h setup osts and we will
onsider a single paint station. The formulation below may easily be generalized
to anynumber ofstations and anytype of setup osts(see Giard, 2003). A typi al
setup ostinvolvesthe ostasso iatedwithma hinetuneupstoswit hfromamodel
toanother, e.g. atwo-door vehi letoa four-door one. Vehi lesof asame olourare
groupedinbat hesbutthe bat hsizehasanupperboundfortworeasons. First,
af-tersome vehi les,spraygunsgetpaint-en rusted andneed apurge. Se ond, a olour
hange is required to avoid eye tiredness responsible for a less ee tive paint aw
dete tion. Welet
L
bethe maximum number of ars inanyidenti al- olour group. The binary variableu
h
takes a value of1
if the olour of the ar in positionh
is dierent from the olour of the previous one inthesequen e (positionh − 1)
and0
otherwise. Thesumofvariablesu
h
overthe wholesequen eexpressesthenumberof purges. Lettingγ
betheunit ostof leaningpaint fromsprayguns (solvent,labour andstation downtime), the obje tive isto minimizeγ ·
N
X
h=1
u
h
.
(12)The onstraint on the maximum group size is equivalent to enfor e at least one
olour swit h between positions
h − L + 1
andh, ∀h
. Thisiswritten ash
X
l=h−L+1
u
h
≥ 1, h = 1, . . . , N,
(13)whi hindi atesthenumberof olour hangesanysubsequen eof
L
vehi lesmust be at leastequal to one.We must impose additional onstraints on variables
u
h
so they take a value of1
ea h time there is a olour swit h in the sequen e, and zero otherwise. The olour index of vehi lei
is denoted byρ
i
so the olour index whi h is applied to the vehi le in positionh
isP
N
i=1
ρ
i
· y
i,h
. Variableu
h
will take a value of 1 if a olour hange o urs between positionh − 1
and positionh,
whi h is translated by anonzerodieren e of olour indi esbetweenpositionh
andpositionh − 1
,thatisP
N
i=1
ρ
i
· y
i,h
−
P
N
i=1
ρ
i
· y
i,h−1
6= 0.
This onditionisfullledwhenthetwo following inequalitieshold:P
N
i=1
ρ
i
· y
i,h
−
P
N
i=1
ρ
i
· y
i,h−1
≤ P · u
h
, h = 1, . . . , N,
P
N
i=1
ρ
i
· y
i,h
−
P
N
i=1
ρ
i
· y
i,h−1
≥ −P · u
h
, h = 1, . . . , N.
(14)
To illustrate, onsider
N = 6
vehi les andP = 2
olours (an index value of1
is hosen for the blue olour and a value of 2 orresponds to the red hue). Themaximum number of ars with same olours in a sequen e is
L = 3
. In table 7, we give the values of relevant variables for a given sequen e of vehi les for whi holoursareindi ated. Forea hposition
h,
thevalueofP
N
i=1
ρ
i
· y
i,h
simplygivesthe olourindexofthevehi leinpositionh
(thisisgiveninthe4
th
rowofTable7). The
dieren e
P
N
i=1
ρ
i
· y
i,h
−
P
N
i=1
ρ
i
· y
i,h−1
betweenthe olourindi esinpositionh
andh − 1
isnon zerowhenthe olourhas hanged between the urrent positionandthe previousone. Thisdieren e isgiven for ea hpositionh
inthe5
th
rowof thetable
( olour dieren e). During horizon
T,
this happens in positionsh = 1
andh = 3
(and alsoforh = 0
whi h is the last position inhorizonT − 1)
. Appropriate values foru
h
arethenobtainedutilizing onstraint(14). Forinstan e, inpositionh = 3,
we haveP
6
i=1
ρ
i
· y
i,3
−
P
6
i=1
ρ
i
· y
i,2
= −1
andwe wantu
3
su hthat−2u
3
≤ −1 ≤ 2u
3
whi himpliesu
3
= 1.
Thus,a olourdieren ene esarilyimpliestheimplementation ofa olour hange. ThelastrowofTable7 he ks onstraint(13). This onstraintisviolated for
h = 6
asno olour hange hasbeendone for thelast 3vehi les sothere is more than 3 onse utive vehi les with same olour. Lettingγ = 10
euros be the unit ostof apurge, thetotal ostof leaninggunsduringhorizonT
amountsto20
eurosasP
6
h=1
u
h
= 2.
Horizon
T − 1
T
Position
h
...−2
−1
0
1
2
3
4
5
6
Colourname red red blue red red blue blue blue blue
Colourindex
2
2
1
2
2
1
1
1
1
Colourdieren e
0
−1
1
0
−1
0
0
0
Colour hange
(u
h
)
0
1
1
0
1
0
0
0
P
h
l=h−2
u
h
2
2
2
1
1
0
Table7: Constraintson olour hanges-illustration
4 Summary: notations, assumptions and the
optimiza-tion problem
Tables8and9providethenotationsweusedsofar. Theoptimizationproblemnally
onsistsinminimizingthe sumoftheobje tivefun tionsin(10)and (12)subje tto
onstraints (1), (3) for all xed- apa ity stations in the set
K
C
, (7) for all exible stationsinK
F
,(8), (9),(11), (13), (14). We madethefollowing assumptions.Assumption 1. There are only two possible values for the pro essing time on
station
k, θ
max
k
andθ
min
k
whereθ
max
k
= max
i=1...,N
{θ
i,k
}
(with at least one itemi
requiringoptionk
)andθ
min
k
= max
i=1...,N
θ
i,k
| θ
i,k
< θ .
Thus, installing option
k
takes always the same pro essing time whatever the itemtype. Allitemswithoutoptionk
areassumedtohavethesame pro essingtimeθ
k
min
onstationk
althoughsomeofthemmayrequireazerotimewhi ha tuallyleads toagreaterde reaseintheex essworkingtimethantheonewe onsidered. Withoutthis assumption howeverthe logi ofspa ing requirements onexible stations must
be abandonedinfavor ofa reasoninginterms ofpro essingtimes.
Assumption 2. For any option
o
pro essed on several stationsK
o
withK
o
=
k| θ
o,k
> θ
T
Sequen inghorizon,usuallyadayN
Numberofitemstobesequen edoverT
θ = ⌊T /N ⌋
Cy letimeK
SetofallstationsinthelineK
C
Subsetof xed- apa itystationsK
F
Subsetof exiblestationsO ≡ K
C
∪ K
F
Setofoptionsλ
k
Numberofdedi atedoperatorsonstationk
ν
k
: η
k
Spa ing onstraint: `atmostν
k
itemswithoptionk
outofη
k
onse utiveitems'
δ
i,k
Booleantakingavalueof1
ifitemi
requiresoptionk
and0
otherwiseθ
i,k
Pro essingtimeofitemi
onstationk
θ
max
k
Pro essingtimeonstationk
ofanyitemrequiringtheoptionk
θ
min
k
Pro essingtimeonstationk
ofanyitemwithoutoptionk
R
max
k
Maximumex essworkingtimeallowedonstationk
µ
k
Numberofitemswithoutoptiontobesequen edafterν
k
onse utiveitemswithoption
k
toavoidex eeding apa ity,whenλ
k
= 1
β
k
Numberof y letimesautilityworkerspends onstationk
to pro essoneitemwithoption
ω
Costofhiringautilityworkerperhalf-dayγ
Set-up ostL
Maximumnumberof onse utiveitemswithsame olourP
Numberof oloursρ
i
Colourindex hosenforitemi, 1 ≤ ρ
i
≤ P
Table 8: Notationsforparameters
Itshouldbenotedthatwhenautilityworkerisallowedtoworkonsu hstation
k
, thisrelaxesthespa ing onstraintsoallstationsinK
o
mustbe onsideredseparately. A simplifying assumption onsists in onsidering that any option is fullypro essedbya singlespe ialized workstation.
Assumption3. Ea h exible stationhasasinglededi ated workerand mayhost
asingleutilityworkerper y letime. Assumption4. The maximumex ess working
time,
(θ
max
k
− θ)ν
k
,
without violating the apa ity of any exible stationk
is su h that(θ
max
k
− θ)ν
k
≤ θ
so the regular worker fully absorbs the ex ess working time whilethe utilityworkersolely pro essesan itemwithoption.y
i,h
∈ {0, 1}
Binarythat takesavalueof1
ifitemi
ispla edinpositionh
m
k,h
∈ N
Numberofitemswithoptionk
in thewindow{h − η
k
+ 1, . . . , h}
w
k,h
∈ {0, 1}
Binarythat takesavalueof1
ifautilityworkerisrequiredonstationk
topro esstheh
th
item
q
k,h
∈ {0, 1}
Binarythat takesavalueof1
ifiteminpositionh
hastheoptionandnoneoftheitemsin positions
{h − η
k
, . . . , h − 1}
ishandledbya utilityworkerW
1
(T ) ∈ N
Numberofa tivetemporaryworkersinthersthalf ofhorizonT
W
2
(T ) ∈ N
Numberofutilityworkersrequiredtopro essitemsinthese ondhalfofhorizon
T
W
3
(T ) ∈ N
Numberofutilityworkersneededtoimplementsequen ingde isionsmadefor
T
but physi allypresentonlyinthenexthorizonT + 1
u
h
∈ {0, 1}
Binarythat takesavalueof1
ifaset-upisin urredto pro esstheh
th
item(e.g. a olourswith)
Table9: Notations
5 Computational study
We rst dis uss the initial onditions. We then des ribe the experimental design.
We nallyturnto the omment of theresults.
5.1 Initial onditions
In pra ti e, sequen ing de isions are made on a rolling horizon basis: the sequen e
implementedwithinhorizon
T − 1
providesinitial onditionstothesequen ing prob-lem for horizonT.
In the experiment, we have generated a random sequen ing of itemsfor horizonT − 1
andwehave optimized thesequen ing forhorizonT.
Inthe random sequen ing, we have in orporated some rules in order to obtain a feasiblesequen e that satises the spa ing onstraints for xed- apa ity stations only. We
then have omputed the values of all variables orresponding to this feasible
ran-domsequen e,someofthembeingin ludedasinitialparametersintheoptimization
program relative to horizon
T.
5.2 Experimental design
Thenumber
N
ofvehi lestobesequen edduringthehorizon tookthevaluesinthe set{10, 15, 20, 25, 30, 35}.
ForN = 10
and15,
thenumberP
of olourswas setto3
withamaximumofL = 3
onse utive arswithsame olour. ForN = 20
and25,
we hoseP = 4
andL = 5.
Finally,forN = 30
and35,
we setP = 5
andL = 6.
Note thatfor ea hvalueofN,
parametersP
andL
aresu h thatareasonable numberof subset of onse utive identi al- olour ofvehi les maybe obtained.K
F
of exible stations took the values of2
and4,
i.e.(K
C
, K
F
) = {(1, 2), (2, 4)}
leading to a total number of either3
or6
options. We setν
k
= 1
for allk,
sothe spa ing onstraint for exiblestations is a ne essaryand su ient onditionnot toex eed the apa ity. We let
η
k
U {2, 3, 4, 5}
for allk
so we get reasonable values forthe spa ing onstraints withrespe tto thenumberof itemsto besequen ed.Letting
N
k
bethenumberofvehi leswithoptionk,
wesetN
k
= 0.75N ν
k
/η
k
for allk ∈ K
C
andN
k
= 2N ν
k
/η
k
for allk ∈ K
F
.
This makes easier the sear h for a feasiblesequen e satisfying thespa ing onstraints for xed- apa ity stations whilemakingne essarythe appointment ofsome utility workers.
Ea h item is uniformly pi ked at random to be endowed with option
k
until the number of itemswith optionk
rea hesthe valueofN
k
.
The olours of ars are identi allyindependentlyuniformlydistributedoverthesetofpossible olours. Thusthereis on average a proportion
1/P
of ars ofea h olour. In a ordan e withthe fren h wage levels we setω = 70
euros. In thefren h automotive industry, the ost of leaning the sprayguns approximates10
eurossowesetγ = 10
throughout.For ea h problem dened in terms of values for
N
and for the pair(K
C
, K
F
),
we made 3 repli ations and we obtained30
instan es sin e we only onsidered(K
C
, K
F
) = (1, 2)
forN ≥ 30.
Intheoptimizationprograms,thespa ing onstraints forexible stations werenally written ash
X
l=1
N
X
i=1
δ
i,k
· y
i,l
− (1 + ν
k
) · w
k,l
!
−
h−η
k
X
l=1
q
k,l
, h = 1, . . . , N ; k ∈ K
F
,
(15)instead of using (7) whi h would lead to a higher number of variables with the
introdu tionofthe `intermediary'variables
{m
k,h
}.
Analogously,weusedthespa ing onstraints for apa itated stations asgiven by(2).5.3 Results
Table 10 reports the results for the small instan es (up to 15 vehi les) for whi h
we get an optimal solution ina reasonable timethrough theuse of the ommer ial
software MIPsolverILOG-Cplex 9.0. The rstthree olumnsgive thevalues of
N
,K
C
andK
F
whi h dene a type of instan e. The next two olumns provide the number of variables and of onstraints. For ea h type of problem and ea h of thethreerepli ations,wereportthequalityoftherstfeasiblesolutionmeasuredasthe
relative deviation (gap)to the optimalvalue. The penultimate olumn displaysthe
exe ution time to rea h the rst feasible solution whereas the last olumn exhibits
the time to obtain the optimum. Computing times are expressed in CPU se onds
andrefer to aBi-Xeon 3.4GHz with4Goof mainmemory.
Ex ept for one of the largest problems, it takes less than 2 se . to get a rst
integer (feasible)solution. Thequalityof these rstfeasible solutions is rather
het-erogeneous, with a deviation to the optimal value ranging from about 15% to 75%
withno obvious relationship between the deviation and theproblem size. The
exe- utiontimeto rea htheoptimumseemsto stronglydependonthestru tureofea h
N
K
C
K
F
Var. Const. Gap(%) Time Time, Opt.10
1
2
151
166
34.38 0.01 0.67 32.26 0.010.27
72.73 0.030.65
10 2 4 189 273 19.15 0.98 14.33 17.02 1.99 79.99 35.90 1.26 240.65 15 1 2 301 251 54.55 0.04 3.83 75.00 1.41 648.98 57.58 0.03 0.23 15 2 4 359 413 18.03 0.24 66.17 46.00 15.87 1026.38 14.52 0.12 47.19Table 10: Feasible andoptimal solution tosmall instan es
the largest time is asso iated with the biggest problem and the smallest instan es
arealsothe fastest problemsto be solvedto optimality.
Table 11 displays the results for larger instan es (up to 35 vehi les) for whi h
Cplex wasnot able to nd theoptimal solution within the timelimit of 7200 CPU
se onds. Like in the previous table, the rst ve olumns des ribe the instan es.
The next two olumns report the quality of the rst feasible solution and thetime
required byCplex to rea h it. Thequalityis measuredbythe relative gap between
the rstfeasiblesolution andthebestbound. The penultimate olumn displaysthe
gapbetweenthebestintegersolutiondeliveredbyCplexafter7200se . andthebest
bound. Thelast olumn providestheimprovement oftheobje tive valueafter 7200
se . ompared to the rst feasible solution(for instan e, thebestobje tive valueis
23.40% lowerthan therst feasiblesolution to therstproblem).
For instan es up to 20 vehi les Cplex delivers solutions of good quality within
the 7200-se ond timelimit. For 13 ases out of 18, Cplex is able to provide a rst
feasible solution in less than 3 min. Even if the solution quality after 7200 se .
rapidly degrades aslarger instan es are onsidered, the solution to 5 instan es over
18exhibitsagapbelow5%. Inall ases,therstfeasiblesolutionissigni antly
im-provedduringsubsequent iterationswithade reaseintheobje tivevalueof47.75%
onaverage.
6 Con lusion
Inthispaperwe have developedan optimalformulationfor thesequen ing problem
where simultaneous minimization of the number of utility workers and setups is
desired. The resultant optimization model is more realisti than those of earlier
papersdealingwithsu hgoalssin eitprovidesanoperationalwaytoimplementthe
utility work needed to avoid line stoppages. Although theproposed approa h here
an hardly deal with real-size problems, it may lay down the basis for developing
N
K
C
K
F
Var. Const. Gap(%) Time Gap(%) Imp. (%) 20 1 2 501 336 26.09 0.06 2.17 23.40 40.63 0.20 3.13 36.36 65.63 0.15 3.13 60.61 20 2 4 579 553 60.38 468.59 5.00 34.92 49.06 4.06 14.46 29.51 50.94 14.60 16.67 14.29 25 1 2 751 420 275.00 11.09 150.00 50.00 291.67 93.34 95.61 95.83 66.67 0.63 3.03 61.76 25 2 4 849 693 139.39 1703.85 77.67 33.90 80.00 424.21 70.00 5.88 78.79 23.45 66.67 7.27 30 1 2 1051 505 308.33 12.64 158.33 58.06 308.33 16.37 108.33 96.00 308.33 163.80 108.33 96.00 35 1 2 1401 590 225.00 735.13 158.33 25.81 140.74 16.55 39.06 58.54 361.54 422.38 169.23 71.43Table 11: Feasibleand bestinteger solutionto larger instan es
Referen es
[1℄ J.Bautista,R.Companys,andA.Corominas.Heuristi sandexa talgorithmsto
solve the monden problem. European Journal of Operational Resear h, 88:101
113,1996.
[2℄ J. Bautista, R. Companys, and A. Corominas. Note on y li sequen es in
the produ tratevariationproblem. European Journalof Operational Resear h,
pages 468477, 2000.
[3℄ L. D. Burns and C. F. Daganzo. Assembly line job sequen ing prin iples.
In-ternational Journal of Produ tion Resear h, 25:7199,1987.
[4℄ W. Choiand HShin. Areal-time sequen e ontrolsystemfor thelevel
produ -tion of the automobile assembly line. Computers ind. Engng, pages 769772,
1997.
[5℄ F.-Y. Ding and L. Cheng. An ee tive mixed model assembly line sequen ing
heuristi for just-in-time produ tion systems. Journal of Operations
Manage-ment,pages 4550,1993.
[6℄ A. Drexl and A. Kimms. Sequen ing jit mixed-model assembly lines under
station-loadandpart-usage onstraints.ManagementS ien e,47:480491,2001.
Colony Optimisation Approa hes for Car Sequen ing Problems. Le ture Notes
inComputer S ien e,Springer, 2003.
[9℄ C.J.Hyun,Y.Kim,andY.K.Kim. A geneti algorithm formultipleobje tive
sequen ingproblemsinmixedmodelassemblylines. ComputersandOperations
Resear h, pages675690, 1998.
[10℄ T.KorkmazelandS.Meral. bi riteriasequen ing methods forthemixed-model
assemblylineinjust-in-timeprodu tionsystems.europeanjournalofoperational
resear h, pages188207, 2001.
[11℄ W. Kubiak and S. Sethi. A note on level s hedules for mixed-model
assem-bly linesinjust-in-timeprodu tion systems. Management S ien e, 37:121122,
1991.
[12℄ Y.Y.Leu, A.L.Matheson,andL.P.Rees. Sequen ing mixed-model assembly
lines withgeneti algorithms. Computers andIndustrial Engineering, 30:1027
1036, 1996.
[13℄ J. L.Ma askill. Computer simulation for mixed-modelprodu tion lines.
Man-agement S ien e, 20:341348,1973.
[14℄ S.A.Mansouri.Amulti-obje tivegeneti algorithmformixed-modelsequen ing
on jit assembly lines. European Journal of Operational Resear h, 167:696716,
2005.
[15℄ J. Miltenburg. Level s hedules for mixed-model assembly lines in just-in-time
produ tion systems. Management S ien e, 35:192207, 1989.
[16℄ S.Mitsumori. Optimals hedule ontrolof onveyor line.IEEE Transa tions on
Automati Control,10:633639, 1969.
[17℄ Y.Monden.ToyotaProdu tionSystem,anIntegratedApproa htoJust-in-Time.
Engineering and Management Press, Nor ross,Georgia,1998.
[18℄ T. Murata, H. Ishibu hi, and H. Tanaka. Multi-obje tive algorithm and its
appli ations to owshop s heduling. Computers and Industrial Engineering,
30:957968, 1996.
[19℄ K.Smith,M.Palaniswami,andM.Krishnamoorthy.Traditionalheuristi versus
hopeld neural network approa hes to a ar sequen ing problem. European
Journal of Operational Resear h, pages300316, 1996.
[20℄ R. T. Sumi hrast, R. S.Russell, and B. W. Taylor. A omparative analysis of
sequen ingpro eduresformixedmodelassemblylinesinjust-in-timeprodu tion
system. International Journal of Produ tion Resear h, pages 199214, 1992.
[21℄ T. Tamura, H. Long, and K. Ohno. A sequen ing problem to level part usage
rates and work loads for a mixed model assembly line with a bypass subline.