Proceedigs

Information of the 15th

Control IFAC Symposium
Problems on
in Manufacturing
Proceedigs
Information
Proceedigs
May of
of the
Control15th IFAC
Problems
theOttawa,
11-13, 2015. 15th IFAC Symposium on
in Manufacturing
Symposium
Canada on
Information
Proceedigs
May Control
of
11-13, 2015.
Information Control Problems
theOttawa,
15th IFAC
Problems in
in Manufacturing
Symposium
Canada on
Available
Manufacturingonline at www.sciencedirect.com
May
May 11-13,
11-13, 2015.
Information Ottawa,
Control
2015. Canada
Problems
Ottawa, in Manufacturing
Canada
May 11-13, 2015. Ottawa, Canada
ScienceDirect
Calculation
Calculation of
of the
the Throughput-Time in
IFAC-PapersOnLine 48-3
Throughput-Time Simple
in (2015)
Simple Assembly
Assembly Lines
314–319
Lines with
with Learning
Learning Effect
Effect
Calculation of the Throughput-Time in Simple Assembly Lines with Learning Effect
Calculation of the Throughput-Time in Simple
Tamás Koltai*
Tamás Koltai*
Noémi Assembly
Kalló**
Noémi Kalló** Lines with Learning Effect
TamásRita
Tamás
Györkös***
Koltai*
Koltai*
Rita Noémi
Noémi Kalló**
Györkös*** Kalló**
TamásRita Koltai* Noémi Kalló**
Rita Györkös***
Györkös***
* Budapest University of Technology and Economics, Rita Györkös***
Magyar tudósok körútja 2. 1117 Budapest, Hungary
* Budapest University of Technology and Economics, Magyar tudósok körútja 2. 1117 Budapest, Hungary
** Budapest (Tel: +36-14632432; e-mail: koltai@mvt.bme.hu).
Budapest University
University ofof Technology
Technology and
and Economics,
Economics,
(Tel: +36-14632432; Magyar
Magyar
e-mail: tudósok
tudósok körútja
körútja 2.
koltai@mvt.bme.hu). 2. 1117
1117 Budapest,
Budapest, Hungary
Hungary
**
* BudapestUniversity
Budapest UniversityofofTechnology
Technology and
and Economics,
Economics, Magyar
Magyar tudósokkörútja
tudósok körútja 2.1117
1117 Budapest,Hungary
Hungary
(Tel:
(Tel: +36-14632432;
** Budapest University of Technology +36-14632432;
and Economics, e-mail:
e-mail: tudósok körútja2.
koltai@mvt.bme.hu).
koltai@mvt.bme.hu).
Magyar 2. 1117Budapest,
Budapest, Hungary
** (Tel:
(Tel: +36-14631057;
+36-14632432; e-mail:
e-mail: kallo@mvt.bme.hu).
koltai@mvt.bme.hu).
** Budapest
Budapest University
University of
of Technology
Technology and
and Economics,
Economics,
(Tel: +36-14631057; Magyar
Magyar
e-mail: tudósok
tudósok körútja
kallo@mvt.bme.hu).körútja 2.2. 1117
1117 Budapest,
Budapest, Hungary
Hungary
*** BudapestUniversity
** Budapest University ofTechnology
Technology andEconomics,
Economics, Magyar tudósokkörútja
körútja 2. 1117Budapest,
Budapest, Hungary
*** Budapest Universityof (Tel:
of Technology and
(Tel: +36-14631057;
+36-14631057;
and Economics, Magyar
e-mail:
e-mail: tudósok
tudósok körútja2.2.1117
kallo@mvt.bme.hu).
kallo@mvt.bme.hu).
Magyar 1117 Budapest,Hungary
Hungary
*** (Tel:
(Tel:+36-14631092;
+36-14631057; e-mail:
e-mail: gyorkos@mvt.bme.hu).
kallo@mvt.bme.hu).
*** Budapest
Budapest University
University of
of Technology
Technology
(Tel: and
and Economics,
+36-14631092; Economics,
e-mail: Magyar
Magyar tudósok
tudósok körútja
gyorkos@mvt.bme.hu).körútja 2.2. 1117
1117 Budapest,
Budapest, Hungary
Hungary
*** Budapest University of Technology
(Tel: and Economics,
(Tel: +36-14631092;
+36-14631092; e-mail:Magyar
e-mail: tudósok körútja 2. 1117 Budapest, Hungary
gyorkos@mvt.bme.hu).
gyorkos@mvt.bme.hu).
(Tel: +36-14631092; e-mail: gyorkos@mvt.bme.hu).
Abstract: Calculating the throughput-time of a production run in simple assembly lines is important for several
Abstract: Calculating the throughput
reasons. For example, the length of time -timeto of a production
complete run in simple
a production order isassembly
required lines is important
for planning and for several
scheduling
Abstract:
Abstract:
reasons. For Calculating
example,
Calculating the
the
the throughput
length of
throughput -t
-t
timeime
imeto of aa production
complete
of a run in
production
production run in simple
order
simple isassembly
required
assembly lines
for
lines is important
planning
is and
important for several
scheduling
for several
production or the evaluation of throughput-time is required by any effort to improve assembly operations. Classical
Abstract:
reasons. For
production
reasons. Calculating
Fororexample, thelength
the
the evaluation
example, the throughput
of -time
of time to of
to
of throughput-time
length time aisproduction
complete
completerequired byrun
anyineffort
aa production
production simple
order isassembly
ordertois required
improve lines
for is important
for planning
requiredassembly and for
and several
scheduling
operations.
planning Classical
scheduling
assembly line balancing techniques assume a constant cycle time. In this case, the calculation of the throughput-time
reasons.
production
assembly Fororexample,
or
line the the
the evaluation
balancing length ofassume
time to complete
of throughput-time
techniques is a production
is required by any order
anyIneffort toisimprove
to required for planning
assembly and scheduling
operations. Classical
production
is straightforward. evaluation
In case ofofthe presence aofconstant
throughput-time learning cycle
required time.
by
effect, this
effort
however, case,
cycle the
improve
timecalculation
assembly
changes of the
twothroughput-time
foroperations.
main Classical
reasons.
production
assembly
is or
line the
straightforward.
assembly line evaluation
balancing
In
balancing case ofof throughput-time
techniques
the
techniques assume
presence aof
assumebecause is required
constant
learning
a constant cycle by
effect,
cycle any
time. Ineffort
this
however,
time. Inofthis to
case,
cycleimprove
the
time
case, thetime assembly
calculation
changes
calculation of
foroperations.
the
two Classical
throughput-time
main reasons.
of the throughput-time
First, cycle time continuously decreases of the decrease operation of the bottleneck station as a
assembly
is
First, line
cycle balancing
straightforward.
time In techniques
In case of
of the
continuously assumebecause
presence
decreases aof
ofconstant
learning
of cycle
the time.
effect, Inofthis
however,
decrease case,
cycle
operation the
timecalculation
time changes of
for the
twothroughput-time
two
of the bottleneck main reasons.
station
is straightforward.
consequence case
of learning. the
Second, presence
bottleneck learning
may shift effect,
from however,
one station cycle time
to another, changes
causing for
further main
changes ofasthea
reasons.
is straightforward.
First, cycle time
consequence ofthis In caseSecond,
of the
continuously
learning. presence
decreases ofmay
learning
because
bottleneck of the
shift effect, however,
decrease of cycle
operation time
timechanges
time of the for two changes
theabottleneck main reasons.
station as aa
First,
cycle cycle time
time. In continuously decreases
paper, an algorithm isbecause
presented of to from
the one
decrease
determine station
of to another,
theoperation
throughput-time causing
of of further
bottleneck
production run of
atasthe
station the
First,
cycle cycle
consequence
time.
consequence time
Inofthiscontinuously
learning.
paper, Second,
an decreases
algorithm isbecause
bottleneck may
presented of the
shift
to decrease
from one
determine of operation
station
the to time causing
another,
throughput-time of of
theabottleneck
further station
changes
production run of
atasthea
presence of learning based on the residence time of the workstations in the bottleneck. The algorithm and the
of learning. Second, bottleneck may shift from one station to another, causing further changes of its
consequence
cycle time.
presence
cycle In
ofare
time. ofthis
learning.
learning
In this paper,
based
paper, Second,
an thebottleneck
anonalgorithm is
residence
algorithm may
timeshift
is presented
presented ofto from
tothe one station
determine
determine the to
in another,
the throughput-time
workstations causing
the bottleneck.
throughput-time of further
of aaThe changes
production
algorithm
production run of
at the
at
runand its
the
application illustrated with a simple example.
cycle time.
presence
application
presence of In
ofare this paper,
learning based
illustrated
learning anon
with
based onalgorithm
the
a simple is presented
the residence
example.
residence time
time ofoftothe
determine
the the throughput-time
workstations
workstations in
in the
the bottleneck.of aThe
bottleneck. production
The algorithm
algorithmrunand
at the
and its
its
presence
applicationofare
© 2015, IFAC
application
Keywords: learning
are based
illustrated withon
with
(International
illustrated
Capacity and the residence
aaFederation
simple
simple
Performance example. time of the
of Evaluation;
example. Automatic workstations
Control)
Line Hostinginand
Design bytheElsevier
bottleneck. The
Ltd. All
Balancing; algorithm
rights and its
reserved.
Human-Automation
Keywords:areCapacity
application illustratedand
withPerformance
a simple example. Evaluation; Line Design and Balancing; Human-Automation
Integration; Learning Curve; Bottleneck Analysis.
Keywords: Capacity
Keywords:
Integration; Capacity
Learning Curve; and
and Performance
Performance Evaluation;
Evaluation; Line
Bottleneck Analysis. Line Design
Design and and Balancing;
Balancing; Human-Automation
Human-Automation
Keywords:
Integration; Capacity
Learning and
Curve; Performance
Bottleneck
Integration; Learning Curve; Bottleneck Analysis. Evaluation;
Analysis. Line Design and Balancing; Human-Automation
Integration; Learning Curve; Bottleneck Analysis. learning and forgetting on bottleneck shifts was studied by
1. INTRODUCTION learning and forgetting on bottleneck shifts was studied by
1. INTRODUCTION Glock
learning andand Jaber (2013)onin bottleneck
forgetting a two-stage production
shifts was system.
learningand
Glock andJaberforgetting
(2013)onin bottleneck
a two-stage shifts was studied
production studied
system.by
by
A simple assembly1. INTRODUCTION
1. line consists of several consecutive The
INTRODUCTION
Glock
The
widespread
learning
Glock and
andand
widespreadJaber
Jaber
effect
forgetting
(2013)
(2013)
effect
of
on
of in
in
learning
bottleneck
a
a
on
two-stage
two-stage
learning on
scheduling
shifts was
production
production
scheduling
problems
studied
system.
system.
problems by
A simple assembly1. line INTRODUCTION
consists of several consecutive has been reviewed by Biskup (2008).
workstations. At each station, an operator performs the same Glock
The
Thebeen and
widespread
widespreadJaber
reviewedeffect (2013)
effect of in a two-stage
learning
of learning on production
scheduling
(2008).on scheduling problems system.
problems
A
A simple
simple assembly
workstations. At each line
assembly line
station, consists
consists
an operatorof
of several
several
performs the same has
consecutive
consecutive by Biskup
tasks
A simplerepetitively.
workstations. assembly
At Tasks
line assigned
consists toof the workstations
several consecutive are has The widespread
has been
been reviewed
reviewed effect
by of learning
by Biskup
Biskup (2008).on
(2008). scheduling problems
workstations.
tasks At each
repetitively. each station,
station,
Tasks an
an operator
assigned operator
to the performs
performs
workstationsthe
the sameare The
same effect of learning on assembly line balancing has also
determined
workstations. by
tasks repetitively. At
repetitively. precedence
each station,
Tasks assignedconstraints
an
assigned operator
to andand by several
performs
the workstations
workstationsthe other
same has been reviewed by Biskup
are been studied previously. Cohen andbalancing
The effect of learning on (2008).
assembly line has also
tasks
determined Tasks
by precedence constraints to the by several other are The effect of learning on assembly line Dar-El has
balancing (1998)
also
conditions
tasks like
repetitively.
determined by capacity,
Tasks cycle
assigned time, to workforce
the and
workstations zoning
are The
been effect
studiedof learning on
previously. assembly
Cohen line
and balancing
Dar-El has also
(1998)
determined like
conditions by precedence
precedence
capacity, cycle constraints
constraints and
and by
time, workforce by several
several other
other formulated
and zoning The
been effect
studiedofseveral
learning nonlinear
on
previously. mathematical
assembly
Cohen line
and balancingprogramming
Dar-El has also
(1998)
conditions.
determined
conditions Finished
by parts/products
precedence constraintsleave andthe assembly
by several line
other at been
formulatedstudied severalpreviously.
nonlinear Cohen and
mathematical Dar-El (1998)
programming
conditions like
conditions. like
Finishedcapacity,
capacity, cycle
cycle time,
parts/products time,
leave workforce
workforce
the assemblyand
and zoning
line at models
zoning been
formulated for optimizing
studied several the number
previously.
nonlinear Cohen of stations
and
mathematical inprogramming
an assembly
Dar-El (1998)
the final
conditions station.
like The
capacity, classical
cycle time, assembly
workforce line balancing formulated several nonlinear mathematical
models for optimizing the number of stations in an assembly programming
conditions.
conditions.
the Finished
Finished
final station. parts/products
parts/products
The classical leave
leave
assembly the lineand
the assembly
assembly zoning
line
line at
balancing at line withforlearning.
formulated severalIn these models, however, they assumed
models
conditions.
the finalassume
Finished
station. constant station times,
parts/products
The classical leave
assembly that
the is, regardless
assembly
line line
balancing of
at models
models
line optimizing
withforlearning.
optimizing Innonlinear
the
the number
these number mathematical
models, of stations
stations in
of however, programming
an
an assembly
inthey assembly
assumed
the finalassume
models station. The classical
constant station times,assembly that is,line balancing
regardless of that models
line the
withline
for is balanced,
optimizing
learning. In theand
these the
number bottleneck
models, of stations
however, station
in an
theyis the final
assembly
assumed
how
the frequently
final a
station. task
The is repeated,
classical the task
assembly time is unchanged line
that with
the linelearning.
is In
balanced, these
and models,
the however,
bottleneck stationtheyis assumed
the final
models
models
how assume
assume
frequently aconstant
constant
task station
station times,
is repeated, times,
the that
task is,line
is,
thattime is balancing
regardless
regardless
unchanged of
of station.
line
that with
the Cohen,
linelearning.
is Vitner
In
balanced, and
these
and Sarin
models,
the (2006) showed,
however,
bottleneck stationthey that
is when
assumed
the final
(see
models for example Scholl and Becker (2006); Battaïa and that the
station. line
Cohen, is balanced,
Vitner and
and the
Sarin bottleneck
(2006) station
showed, is thewhen
that final
how
(see forassume
how frequently
frequently
example aaconstant
task
task is
Scholl station
is repeated,
repeated, times,
and Becker the
the task that
task is, is
time
time
(2006); regardless
is unchanged
unchanged
Battaïa of the throughput-time is minimized in an assembly line
and that the
station. line
Cohen, is balanced,
Vitner and
and the
Sarin bottleneck
(2006) station
showed, is the
that with
final
when
Dolgui
how
(see (2013)
frequently
for example ora Koltai,
task is
Scholl Tatay and
repeated,
and Kalló
the
Becker task (2014)).
time
(2006); is unchanged
Battaïa and station.
the Cohen,
throughput-time Vitneris and Sarin
minimized (2006)
in an showed,
assembly that
line when
with
(see for example Scholl and
Dolgui (2013) or Koltai, Tatay and Kalló (2014)).Becker (2006); Battaïa and learning,
station.
the then
Cohen,
throughput-time the optimal
Vitneris and solution belongs
Sarin (2006)
minimized in an to
showed,
assembly an unbalanced
that
line when
with
(see
Dolgui for example Scholl and Becker (2006); Battaïa and the throughput-time
learning, then the is
optimal minimized
solution in an
belongs assembly
to an line
unbalancedwith
In case(2013)
Dolgui of theor
(2013) Koltai,
Koltai, Tatay
orpresence Tatay and
and Kalló
of learning, Kalló (2014)).
(2014)).station time learning,
however, line,throughput-time
the and then
bottleneck
the shifts
is
optimal from thein
minimized
solution final
an
belongs station
assembly
to an toward
towardwith
line
unbalanced the
Dolgui
In case(2013)
of theorpresence
Koltai, Tatay and Kalló
of learning, (2014)).station time learning,
however, line, and then
first station.
the optimal
bottleneck
They
shifts solution
assumed,
from thebelongs
however, that
to an unbalanced
final station
total production
the
decreases
In case
case of of when tasks assigned
the presence
presence of learning, to ahowever,
learning, station are repeated
station time time learning,
line, and then
and
line, station.
first the optimal
bottleneck
bottleneck
They shifts solution
shifts
assumed, from
from thebelongs
the
however, final
final tototal
thatstation
stationan unbalanced
toward
toward
production the
the
In
decreases the
when tasks assignedof to ahowever,
station are station time
repeated can bottleneck
be Theydistributed among thefinal
stations without any
several
In case times.
of the The first
presence description
of learning, of the
however,decrease
stationof task
time line,
first and
station.
first station.
can be They shifts
assumed,
assumed, from the
however,
however, thatstation
total toward
production
that totalwithout
production the
decreases
decreases
several whenThetasks
when
times. tasks assigned
first assigned
description to of
to aa station
station
the decrease of task time
are repeated
are repeated distributed among the stations any
time
several is times.
decreases attributed
when The to assigned
tasks
first Wright
description (1936)
to a
of who
station
the analyzed
are
decrease repeated
of the limitation.
task first station.
time
time can
can
limitation.
They
be
be They
They
also
distributed
also
noted
assumed,
distributednoted
that
among
that
when
however,
among the
the
when
precedence
that
stations
stations
precedence
relations
totalwithout
production
without any
any
relations
severalis times.
time The first
attributed to Wrightdescription (1936) of the whodecrease
analyzed of taskthe and other combinatorial problems limit workload allocation,
operation timesThe in an airplane assembly process. Since of then, time can combinatorial
be distributed among the
limitstations without any
several
time
time is times.
is
operation attributed
attributed
times in an first
to description
Wright
toairplane
Wright (1936)
(1936)
assembly of the who
whodecrease
process. analyzed
analyzed
Since the limitation.
task
the
then, limitation.
and
the
other
problem
They
They
gets
also
also noted
very
noted that
that when
problems
complicated.
when precedence
precedence
workload
Assembly line
relations
relations
allocation,
balancing
the
time existence of
is attributed the learning function (or progress function)the and limitation.
other They
other combinatorial also noted that
problems when
problems limit precedence
workload relations
lineallocation,
operation
operation
the existencetimes
times of in
in anto
thean Wright
airplane
airplane
learning (1936)(orprocess.
assembly
assembly
function who
process. analyzed
progress Since
Since then,
then,
function) and problem
the
problems
combinatorial
gets very complicated. limit workload
Assembly allocation,
balancing
and its practical relevance have been (or illustrated infunction)
several and other with learning
combinatorial has also been
problems limitconsidered
workload by Toksary
allocation,
operation
the existence
the existence
and times
its practical of in
the
of the an airplane
learning
learninghave
relevance assembly
function
function
been (or process.
progress Since several et al. First, they proposed some heuristics balancing
progressinfunction)
illustrated then, the
the problem
problem
problems with gets
gets very
very
learning complicated.
complicated.
has also been Assembly
Assembly
considered line
lineby balancing
Toksary
for the
industrial areas. See, for example, Conway andinfunction)
Schultz the problem gets very proposed
complicated. Assembly lineby
bybalancing
the existence
and
and its
its practical
industrial practicalof the
areas. learning
relevance
relevance
See, function
have
have
for example, been (or progress
illustrated
beenConway
illustrated several problems
several
andin Schultz problems
et
minimization
with
with
al. First,
of
learning
learning
they
total flow
has
has also
also been
time in
been
some considered
considered
simple
heuristics
and U-shaped
Toksary
Toksary
for the
lines
(1959),
and its Hirschmann
practical (1964) and
haveYellebeen(1979). several et problems
et al. withoflearning
al. First, they hastime
proposed alsoinbeen
some considered
heuristics by for
Toksarythe
industrial
industrial
(1959), areas. relevance
areas.
Hirschmann See,
See, for
for and
(1964) example,
example,
Yelle illustrated
Conway
Conway
(1979). andin Schultz
and Schultz minimization
(Toksary
First,
et al,
they
2008).
proposed
total flow
Next,
some
they
simple heuristics
and U-shaped
presented
for
a nonlinear
the
lines
industrial
(1959), areas.
Hirschmann See, for
(1964) example,
and Yelle Conway
(1979). and Schultz et al.
minimization
minimization
(Toksary First,
et of
of
al, they
total
total
2008). proposed
flow
flow time
time
Next, in
in some
theysimple
simple heuristics
and
and
presented U-shaped
U-shaped
a for the
lines
lines
nonlinear
(1959), Hirschmann (1964) and Yelle (1979).
When learning is present, several classical models of integer programing model when learning and task time
(1959), learning
When Hirschmann is (1964)
present, andseveral
Yelle (1979).classical models of (Toksary minimization
(Toksary et
et of
al,
al, total
2008).
2008).flow time
Next,
Next, in
they
they
integer programing model when learning and task time simple and
presented
presented U-shaped
a
a lines
nonlinear
nonlinear
operations management mustseveral
be revised. Themodels
effect of of deterioration
(Toksary et is present (Toksary et al, 2010). They illustrated
When learning
When
operations
learning
learning
on
is present,
is
management present,
cost-volume-profit
mustseveral
be revised.
analysis
classical
classical
has
Themodels
been
effect of
studied by
integer
integer programing isal,present
programing
deterioration 2008).model
model Next,
(Toksary when
whenthey presented
learning
learning
et al, 2010). They anda task
and nonlinear
task time
time
illustrated
When
operationslearning is
management present, must several
be classical
revised. The models
effect of the performance
integer programing of their
model model
when with the
learning solution
and task oftime
the
operations
learning on management
cost-volume-profit must analysis
be revised. has beenThe studied
effect by of deterioration
deterioration
the performance is
is present
present
of (Toksary
(Toksary
their model et
et al,
al,
with 2010).
2010).
the They
They
solutionillustrated
illustrated
of the
McIntyre
operations
learning on (1977)
management and a must
cost-volume-profit nonlinearbe
analysis model
revised.
has for
The
been break-even
effect
studied of
by classical
deteriorationJacksonis 11 problem
present (Toksary with a
et heuristics.
al, 2010). They illustrated
learning on(1977)
McIntyre cost-volume-profit
and a nonlinear analysis modelhas been studied by the
for break-even the performance
performance
classical Jackson of
of
11 their
their
problem model
model
with a with
with the
the
heuristics. solution
solution of
of the
the
analysis
learning has been andsuggested. The economic order quantity
McIntyreon
McIntyre
analysis has cost-volume-profit
(1977)
(1977)
been and aa nonlinear
suggested. analysis
nonlinear
The model
model
economic has been
for
for studied
break-even
break-even
order quantity by classical
the performance
classical
It can be Jackson
Jackson
concluded
of problem
11
11 their model
problem
that when with
with
totalaawith the solution of the
heuristics.
heuristics.
workload can be allocated
(EOQ)
McIntyre model(1977)has been
and arevised by Jaber
nonlinear model and for
Bonney (1999) It
break-even
analysis
analysismodel
(EOQ) has
has been
been suggested.
suggested.
has been revised The
The economic
economic
by Jaber order
order
and Bonney (1999) classical
quantity
quantity can be Jackson
concluded 11 that
problem
whenwith totalaworkload
heuristics.can be allocated
and
(EOQ) a
analysis new has
model approach
been
has been has
suggested. been
revised proposed
The
by economic
Jaber when
and learning
order
Bonney and to
quantity
(1999) It
It
to
thebestations
can
can
the be concluded
concluded
stations
without
that
that
without
any precedence
when
when
any total
total workload
workload
precedence
andcancombinatorial
be
be allocated
andcancombinatorial
allocated
(EOQ) model has been revised by Jaber and Bonney (1999)
and a new approach has been proposed when learning and restrictions, then line optimization is tractable but
forgetting
(EOQ) is
model present
has (Jaber
been and
revised Bonney,
by Jaber 2007).
and The
Bonney effect
(1999) of It
to can
the be concluded
stations
to the stations that
without
thenwithout when
any total workload
precedence
any precedence and can
and but bethe results
allocated
combinatorial
combinatorial
and
and aa new
newisapproach
forgetting approach
present (Jaberhas
has beenbeen proposed
proposed2007).
and Bonney, when
when The learning
effectand
learning of restrictions,
and line optimization is tractable the results
to the stations thenwithout any precedence and but combinatorial
and a newis
forgetting
forgetting isapproach
present
present (Jaberhas been
(Jaber and proposed2007).
and Bonney,
Bonney, when The
2007). learning
The effectand
effect of restrictions,
of restrictions, then line
line optimization
optimization is
is tractable
tractable but the
the results
results
Copyright
forgetting©is2015 IFAC (Jaber and Bonney, 2007). The effect of335 restrictions, then line optimization is tractable but the results
present
Copyright © 2015 IFAC 335
2405-8963 © 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
Copyright
Peer review©
Copyright 2015
©under IFAC
2015 responsibility
IFAC 335
335Control.
of International Federation of Automatic
Copyright © 2015 IFAC
10.1016/j.ifacol.2015.06.100 335
INCOM 2015
Tamás Koltai et al. / IFAC-PapersOnLine 48-3 (2015) 314–319 315
May 11-13, 2015. Ottawa, Canada
have little direct practical relevance. When precedence
relations and other combinatorial problems are considered,
then the problem is practically relevant but only heuristic
solutions can be found.
The objective of this paper is to determine the throughput-
time of the production of a given quantity in simple assembly
lines with learning. Since bottlenecks determine the output
rate of the line, the description of the shifts of bottleneck is
important information for the calculation. Consequently, the
residence time of the stations in the bottleneck is also
determined, which may provide useful information when the
start-up period of a line is examined.
The paper is structured as follows. First, the basic definitions,
notations and the limiting conditions of the analysis are
presented. Next, some underlying properties of the
exponential learning curve and characteristics of the
bottleneck shifts in simple assembly lines are described. An
algorithm is provided that determines those periods, in which
different stations are in the bottleneck during the completion
of a production order. Based on the residence time of the
stations in the bottleneck the throughput-time is calculated.
The performance of the algorithm is illustrated with a
numerical example. Finally, the practical relevance of the
algorithm is discussed and some future research possibilities
are outlined.
2. NOTATIONS AND DEFINITIONS
Let us consider a simple assembly line depicted in Figure 1
(notations used in Figure 1 and in the following part of this
paper are listed in Table 1). Parts proceed from the first
station to the last station visiting all the stations. Stations are
numbered with consecutive integers, denoted by j (j=1,…,J).
The stations following station j are the upstream stations, and
the stations preceding station j are the downstream stations.
For the sake of simplicity, it is assumed that work-in-process
inventory cannot accumulate between the stations.
s1 s2 sk sl sv
1 2 k l v J
Fig. 1. Illustration of station indices and station times of a
simple assembly line.
The station time of station j is denoted by sj. The values of sj
can be determined with assembly line balancing methods. In
this paper, it is assumed that sj is given. If the presence of
learning effect is assumed, then sj is the station time at the
first performance of the operation at station j. The sj(Q)
function shows the station time at station j at the Qth
performance of the operations. Consequently, if Q indicates
production quantity, then sj(Q) is the operation time of the
last part at station j when Q number of parts are produced.
Applying the classical exponential learning function, the
value of sj(Q) is the following,
s j (Q ) = s j Q b ,
where b<0 determines the decrease of station time in case of
learning effect, and it is assumed identical for each station. In
the following part of the paper, sj denotes the duration of the
first performance of the tasks at station j, which will be
briefly called initial station time of station j.
Table 1. List of notations
Indices:
j – index of workstations,
k – index of workstations,
l – index of workstations,
v – index of workstations,
i – index of the workstation entering first in the
bottleneck,
f – index of the workstation entering last in the
bottleneck.
Parameters and variables:
J – total number of workstations,
Q – total production quantity,
Qj – quantity produced at station j,
Q(k,l) – production quantity at which station k enters
and station l leaves the bottleneck,
Qj(k,l) – production quantity of station j, at which
station k enters and station l leaves the
bottleneck,
sj – initial station time at station j,
sj(Q) – station time of station j as a function of
production quantity,
Ll – production quantity at which station l leaves
the bottleneck,
Ek – production quantity at which station k enters
the bottleneck,
Tc – cycle time of the assembly line,
d – difference of the station indices of two
workstations,
b – power of the learning curve function,
L – learning rate (L=2b),
TH(Q) – throughput time function.
Sets:
S – index set of those workstations which might be
in the bottleneck,
sJ I – set of those index pairs which belong to
potential bottleneck changes,
ISol – set of those index pairs which belong to real
bottleneck changes.
Note, that each station of an assembly line processes different
parts at the same time. When station j works on part Q, then
the preceding station (station j−1) works on the next part
(part Q+1) and the succeeding station (station j+1) works on
the previous part (part Q−1). To identify the parts
manufactured at a station, the station index is used together
with the production quantity. Consequently, Qj indicates the
part manufactured at station j. When the identification of the
station is not significant, then the station index is ignored to
simplify notation.
(1) Finished parts leave the assembly line one after the other. The
time between two consecutive finished parts is the cycle time
(Tc). The cycle time of an assembly line is determined by the
bottleneck station, that is, by the station with the longest
336
INCOM 2015
316 Tamás Koltai et al. / IFAC-PapersOnLine 48-3 (2015) 314–319
May 11-13, 2015. Ottawa, Canada
station time. In case of learning effect, cycle time constantly
changes for two main reasons:
1) If station j is the bottleneck of an assembly line, then, as a
consequence of the learning effect, cycle time decreases
exponentially according to the sj(Q) function.
2) In case of the presence of learning effect, bottleneck may
shift from one station to another at certain point of time
causing further changes in the cycle time
The cycle time change is continuous (exponentially
decreasing) in the case of point 1). In the case of point 2,
however, the change of the cycle time is not continuous, as it
is illustrated in Figure 2.
sj(Ql)
station l
sl(Ql)
station k
sk(Ql)
Ql
Ql(k,l)
Fig. 2. Illustration of bottleneck shift.
Figure 2 shows the station time functions of station l and k.
Let d denote the difference between station l and k, that is,
d=l−k, and assume that station k is a downstream station with
respect to station l. When station l works on part Ql, then
station k works on part Qk= Ql+d. To depict the station time
functions of stations k and l in the same diagram, a common
independent variable must be selected. For practical reasons,
we select the production quantity of the latest workstation as
independent variable, which is Ql in this case. Consequently,
sk(Ql) denotes the station time of station k as a function of the
part manufactured at station l.
According to Figure 2, at first, station l is in the bottleneck,
since sl(Ql)>sk(Ql), and cycle time exponentially decreases.
Station time of stations k and l are equal at quantity Ql(k,l).
After this quantity, station k is in the bottleneck of the line
because sl(Ql)<sk(Ql). Consequently, Ql(k,l) indicates
bottleneck shift. At this quantity, station l leaves the
bottleneck and station k enters the bottleneck. The value of
Ql(k,l) can be determined making equal the station times of
stations k and l at the same corresponding production
quantities, that is,
s l Q l b = s k (Q l + d )b .
Eq. (2) can be solved using numerical methods.
In the following part of the paper, the station which enters the
bottleneck is called entering station and the station which
leaves the bottleneck is called leaving station. The index pair
belonging to the bottleneck shift in Figure 2 is denoted by
(k,l). Several station time functions may have intersections in
a long assembly line. Not all of these intersections indicate,
however, a bottleneck shift. The set of index-pairs belonging
to the intersections of station time functions is denoted by I.
The set of index-pairs belonging to real bottleneck shifts is
denoted by ISol. Our objective is to determine those
production quantities, at which bottleneck shift occurs, that
is, to determine all Q(k,l) values for which (k,l)ϵISol.
Knowing the Q(k,l) values, the calculation of the throughput-
time of a given production batch consisting of Q number of
parts is as follows,
j = i −1  q = Q (k ,l )  j=J
 b
TH (Q ) = ∑s j + ∑ ∑
 s l q ∑
 + s jQb . (3)
j =1 (l , v ),(k ,l )∈I Sol  q = Q (l ,v )  j = f +1
Eq. (3) consists of the sum of three elements. The first
element calculates the time required by the first part to reach
the station which enters first in the bottleneck. The second
element calculates the length of time that the different
bottleneck stations spend in the bottleneck. For this
calculation, the Q(k,l) entering and leaving quantities are
required. Finally, the third element calculates the time the last
part spends in the line after leaving the bottleneck station.
It can be seen that the calculation of the Q(k,l) values must be
known to get the throughput-time. In the next section, some
basic relationship necessary to get the Q(k,l) values are
derived.
3. BASIC RELATIONSHIPS
The operational characteristics of simple assembly lines and
the strictly monotone decreasing property of the exponential
function determine some general characteristics of bottleneck
shifts. In the following, some of these general properties of
bottleneck shifts are discussed.
Theorem 1. If station time changes according to an
exponential function in a simple assembly line, then only
those stations can be in the bottleneck for which sk>sl for all
l>k.
Proof. An earlier (k) station always manufactures a later
(Q+d) part compared to a later (l) station. Since station time
exponentially decreases, the station time of a later part is
always smaller than the station time of an earlier part at the
same station. If the initial station time of the earlier station
(sk) is smaller than the initial station time of the later station
(sk<sl), then station k will never enter the bottleneck. In this
case, sk is smaller than sl at the beginning, and it stays smaller
later as well, because station k always processes later parts.
If, however, sk>sl, then it is possible, that this relationship of
the initial station times of stations k and l changes after a
certain production quantity.
(2)
Based on Theorem 1, the initial station time of those stations
which can potentially be bottlenecks create a strictly
monotone decreasing series if station times are ordered
according to the increase of station indices. Those stations,
which cannot be fitted in this strictly monotone decreasing
series of initial station times, never enter the bottleneck. Let
set S contain the indices of those stations which can
337
INCOM 2015
Tamás Koltai et al. / IFAC-PapersOnLine 48-3 (2015) 314–319 317
May 11-13, 2015. Ottawa, Canada
potentially be bottlenecks. In the followings, stations with
indices not belonging to set S can be ignored.
Theorem 2. If station time changes according to an
exponential function in a simple assembly line, then, in case
of bottleneck shift, the station index of the entering station is
always smaller than the station index of the leaving station.
Proof. Bottleneck shift occurs, when the decreasing station
time of the entering station is equal to the decreasing station
time of the leaving station. This occurs only if the decrease of
station time of the entering station is larger than the decrease
of the station time of the leaving station. The gradient of the
station time of station j is the following:
ds j (Q)
= s j bQ b−1 .
dQ
In the case of an earlier station, initial station time is larger
than the initial station time of a later station according to
Theorem 1, and Q is higher, because earlier station processes
later parts. Consequently, in a certain point of time, the
gradient of station time of an earlier station can be higher
than the gradient of a later station. The opposite, however, is
not true. The initial station time of a later station is always
smaller than the initial station time of an earlier station
according to Theorem 1, and Q is also smaller, because later
station processes earlier parts. Consequently, bottleneck can
shift only downstream.
According to Theorem 2, the indices of the bottleneck
stations are strictly monotone decreasing if the indices are
ordered according to the time of occurrence of the bottleneck
shifts.
According to Theorem 2, bottleneck moves downstream if
there is bottleneck change during production. Consequently,
production quantities belonging to bottleneck shifts form a
strictly monotone decreasing series if these quantities are
ordered according to the index of entering (or leaving)
stations, that is, QJ(k,l)>QJ(l,v) if k<l<v.
Theorem 3. If station time changes according to an
exponential function in a simple assembly line as a
consequence of learning, then there is only one intersection
of any two station time functions.
Proof. The intersection of the station time functions of
stations l and k can be determined with (2) if d=l−k.
Rearranging (2) we get the following formula,
sk Ql b
= .
sl (Ql + d )b
The right-hand-side of (5) is the ratio of the two values of the
same exponential function with negative power. This
exponential function is strictly monotone decreasing;
therefore, the ratio of the right-hand-side of (5) is also strictly
monotone decreasing, and the limit value with respect to Ql is
equal to 1. Consequently, each value of the right hand side of
(5) is taken only once.
At the Q(k,l) value, determined by (2) or (5), the station times
at station k and l are equal. Below this value, the station time
of station l is higher than the station time of station k. Above
this value, the station time of station k is higher than the
station time of station l. Consequently, Q(k,l) may belong to
bottleneck shift, but not necessarily belongs to any shift. The
intersection of the station time functions of stations k and l
with other station time functions must also be considered. If a
station time function intersects more than one other station
time functions, then the following two corollaries of Theorem
3 can help determining the real bottleneck shifts.
If the station time function of station l intersects several other
station time functions from above, then only one of these
intersections may belong to bottleneck shift. Let Ll denote the
production quantity at which station l leaves the bottleneck.
(4) Corollary 1. If the station time function of station l intersects
several other station time functions from above, then the
quantity which belongs to bottleneck shift (Ll) can be
determined in the following way,
Ll = Min [Q( j, l )] . (6)
( j,l )∈I
At quantity Ll, station l leaves the bottleneck.
Proof. The station time function of station l is strictly
monotone decreasing and, according to Theorem 3, it has
only one intersection with another station time function. If
the station time function of station l intersects a station time
function from above, then it stays below this function forever.
Consequently, only the first of such intersections belongs to
bottleneck shift. This first intersection occurs at the smallest
Q(j,l) quantity.
If the station time function of station k intersects several other
station time functions from below, then only one of these
intersections may belong to bottleneck shift. Let Ek denote
the production quantity at which station k enters the
bottleneck.
Corollary 2. If the station time function of station k intersects
several other station time functions from below, then the
quantity which belongs to bottleneck shift (Ek) can be
determined in the following way,
Ek = Max [Q(k , j )] . (7)
(k , j )∈I
At quantity Ek, station k enters the bottleneck.
Proof. The station time function of station k is strictly
monotone decreasing and, according to Theorem 3, it has
only one intersection with another station time function. If
(5) the station time function of station k intersects a station time
function from below, then it stays above this function forever.
Consequently, only the last of such intersections belongs to
bottleneck shift. This last intersection occurs at the highest
Q(k,j) quantity.
Corollary 1 and 2 determines those quantities at which a
station enters and leaves the bottleneck. Bottleneck change
involving stations k and l occurs only if the quantity at which
the leaving station leaves the bottleneck (Ll) is equal to the
quantity at which the entering station enters the bottleneck
(Ek). In this case, station k enters the bottleneck and station l
338
INCOM 2015
318 Tamás Koltai et al. / IFAC-PapersOnLine 48-3 (2015) 314–319
May 11-13, 2015. Ottawa, Canada

leaves the bottleneck. Consequently, the index pairs Table 2: Elements of the Q5(k,l) matrix
belonging to the bottleneck shift is determined as follows,
Q5(k,l) l=1 l=2 l=3 l=4 l=5 Row max.
(k, l ) ∈ I Sol Ek = Ll . (8) (Ek)
k=1 - - 8.29 - 5.25 8.29
When the length of time spent in the bottleneck by a station is k=2 - - - - - -
calculated the Ek values must be rounded up ( Ek  ) and the Ll k=3 - - - - 4.21 4.21
values must be rounded down ( Ll  ). Knowing the Ek and Ll k=4 - - - - - -
values (3) can be written as follows: k=5 - - - - - -
Col. min.
j = i −1  q = Ll   j=J - - 8.29 - 4.21 -
( L l)
 b
TH (Q ) = ∑ ∑ ∑
sj +  s l q  + ∑
s jQb (9)
j =1

l ∈S  q = E l 

 j = f +1
Eq. (9) shows, that the total time that workstation l spends in Step 5: The maximums of the rows of the Q5(k,l) matrix in
Table 2 show the quantities at which the corresponding
the bottleneck is calculated as the sum of the station time of
stations enter the bottleneck. For example, the station time
parts El  ,…, Ll  at station l.
function of station 1 (k=1) intersects several station time
functions from below as can be seen in Table 2. Since
4. ILLUSTRATION OF THE CALCULATION stations 2 and 4 cannot be in the bottleneck, their station time
To illustrate the calculation, a simple assembly line functions are ignored. The station time function of station 3
consisting of five workstations (J=5) is used. Initial station and 5, however, must be considered. According to Corollary
time of the five stations are given in minutes and are the 2, the latest intersection must be determined. Since
following, s1=18, s2=17, s3=17, s4=15, s5=15. An identical, Q5(1,3)=8.29 is greater than Q5(1,5)=5.25, station 1 enters the
L=0.8 learning rate is assumed for each station. In this case bottleneck when part 8.29 is processed at station 5,
b=−0.3219. The values of the Q5(k,l) matrix and the consequently, E1=8.29. All the Ek values are presented in the
corresponding station indices can easily be determined with last column of Table 2.
the presented algorithm. Step 6: Based on the equivalence of the column minimums
Step 1: The series of the five station times is not strictly and row maximums of the Q5(k,l) matrix the ISol set can be
monotone decreasing. Discarding, however, s2 and s4, a easily determined. The result of the bottleneck shift analysis,
strictly monotone deceasing series of station times is that is, the production quantities belonging to bottleneck
obtained. Consequently, station 2 and 4 cannot be in the shifts and the corresponding station indices are summarized
bottleneck of the line and indices 1, 3 and 5 are the elements in Table 3.
of set I. Table 3. The Q5(k,l) quantities of the sample problem
Step 2: The bottleneck always shifts downstream, Station 1 2 3 4 5
consequently, the elements of set I are the following index Entering (Ek) 8.29 0 4.21 0 1
pairs: (1,3), (1,5) and (3,5). Leaving (Ll) ∞ 0 8.29 0 4.21
Step 3: Solving (2) for the 3 elements of set I the following
values are obtained: Q5(1,3)=8.29, Q5(1,5)=5.25 and
Two technical comments must be added to the results of
Q5(3,5)=4.21. The elements of the Q5(k,l) matrix are
Table 3. First, note that the quantities in Table 3 indicate the
presented in Table 2. There are no values in columns and
parts processed at station 5, that is, the Q5(k,l) values are
rows 2 and 4, because station 2 and 4 cannot be in the
given. For example, bottleneck shifts from station 3 to station
bottleneck of the line. According to Theorem 2, there are no
1 when station 5 processes part 8.29. The Q5(k,l) quantities,
data below the main diagonal. Data in the main diagonal are
however, can be easily transformed into the quantities
also infeasible, because bottleneck shift requires two different
processed at any station using the station index differences.
stations.
The first bottleneck shift occurs when station 3 processes part
Step 4: The minimums of the columns of the Q5(k,l) matrix in 6.21, which is calculated as 4.21+(5−3). The second
Table 2 show the quantities at which the corresponding bottleneck shift occurs when station 1 processes part 12.29,
stations leave the bottleneck. For example, the station time which is calculated as 8.29+(5−1). Second, the numbers
function of station 5 (l=5) intersects several station time given in Table 3 indicate which part is processed; therefore,
functions from above as can be seen in Figure 2. Since these numbers should be integers. The non-integer values
stations 2 and 4 cannot be in the bottleneck, their station time indicate that bottleneck shift occurs during the processing of
functions are ignored. The station time function of station 3 a part, and not at the end of the corresponding operation. For
and 5, however, must be considered. According to Corollary example, bottleneck shifts from station 5 to station 3, when
1, the earliest intersection must be selected. Since station 5 processes part 5.
Q5(3,5)=4.21 is smaller than Q5(1,5)=5.25, station 5 leaves
Step 7: Using (9), the three element of the throughput-time is
the bottleneck when part 4.21 is processed, consequently,
calculated. In the sample problem, station 5 is in the
L5=4.21. All the Ll values are presented in the last row of
bottleneck first. The first part requires 67 seconds to reach
Table 2.

339
INCOM 2015
Tamás Koltai et al. / IFAC-PapersOnLine 48-3 (2015) 314–319 319
May 11-13, 2015. Ottawa, Canada
station 5 (18+17+17+15). Stations 5, 3, and 1 stay in the
bottleneck for 47.132, 34.271, and 86.086 minutes,
respectively. The last station in the bottleneck is station 1.
The last part (part 20), after leaving station 1, passes through
all the remaining stations in 23.007 minutes. Consequently,
according to (3) the throughput-time of the production of 20
parts in the sample assembly line is 256.595 minutes.
4. CONCLUSIONS
The presence of learning effect in assembly lines requires the
revision of the traditional assembly line balancing methods.
When task times do not decrease with the frequent repetition
of tasks, a constant cycle time belongs to a line configuration.
The bottleneck of the line is formed by the stations which
station time is equal to the cycle time. The cycle time
determines the capacity of the line and the throughput-time of
a given production order. Consequently, a frequent objective
of the optimization of assembly lines is the minimization of
cycle time.
If, however, task time decreases with the frequent repetition
of tasks, the cycle time of the line decreases as well. The
decrease of task time can be different at different stations;
consequently, the bottleneck of the line may shift from
station to station. In case of learning, line capacity and the
throughput-time of a given production order can be
determined only if bottleneck shifts are known and
quantitatively characterised. Therefore, the direct
minimization of cycle time, used in traditional simple
assembly line optimizations, cannot be applied in the case of
lines with learning.
In this paper, we summarised the major characteristics of
bottleneck shifts in simple assembly lines with homogeneous
learning in case of exponential learning curve. Based on the
strictly monotone decreasing property of the exponential
learning function and on the operational characteristics of
simple assembly lines, three theorems were formulated. With
the help on these theorems, the quantities at which bottleneck
shifts occur and the residence time of the stations in the
bottleneck are determined. Knowing the residence time of the
stations in the bottleneck, the throughput-time of a
production order can easily be calculated.
The calculation of the throughput-time of a given production
order based on the three theorems of Section 3 is illustrated
with a simple example in Section 4. The calculation was
performed in an Excel environment. The implementation of
the algorithm requires the use of simple cell functions and the
repeated application of the goal seeking tool of Excel for
solving (2). The repeated goal seeking was organized by
Visual Basic for Applications (VBA) macros.
Based on the presented results, further important research
questions can be raised:
− In practice, it is difficult to determine the exact value of the
learning rate (L), but in some special cases, the throughput
time can be very robust to the change of L, especially in case
of inhomogeneous learning. The analysis of the robustness of
the throughput time is a topic of further research.
340
− The presented algorithm is based on the mathematical
properties of the exponential learning curve. In case of other
learning functions, a similar approach can be applied to
derive the appropriate algorithm.
− The direct minimization of the throughput time in case of
ALB models with learning results in very complex
combinatorial problems. The systematic evaluation of the
throughput time of different task assignments, however, can
be a basis of ALB heuristics in case of learning.
REFERENCES
Battaïa, O., and Dolgui, A. (2013). A taxonomy of line
balancing problems and their solution approaches.
International Journal of Production Economics, 142 (2),
259–277.
Biskup, D. (2008). A state-of-the-art review on scheduling
with learning effects. European Journal of Operational
Research, 188 (2), 315-329.
Cohen, Y., and Dar-El, M.E. (1998). Optimizing the number
of stations in assembly lines under learning for limited
production. Production Planning & Control, 9 (3), 230-
240.
Cohen, Y., Vitner, G., and Sarin, S. (2006). Optimal
allocation of work in assembly lines for lots with
homogeneous learning. European Journal of
Operational Research, 168 (3), 922-931.
Conway, R.W., and Schultz, A.Jr. (1959). The manufacturing
progress function. The Journal of Industrial Engineering,
10 (1), 39-54.
Glock, H.G., and Jaber, M.Y. (2013). Learning effects and
the phenomenon of moving bottlenecks in a two-stage
production system. Applied Mathematical Modelling, 37,
8617-8628.
Hirschmann, W.B. (1964). Profit from the learning curve.
Harvard Business Review, 42 (1), 125-139.
Jaber, M.Y., and Bonney, M. (1999). The economic
manufacture/order quantity (EMQ/EOQ) and the
learning curve: Past, present, and future. International
Journal of Production Economics, 59 (1-3), 93-102.
Jaber, M.Y., and Bonney, M. (2007). Economic manufacture
quantity (EMQ) model with lot-size dependent learning
and forgetting rates. International Journal of Production
Economics, 108 (1-2), 359-367.
Koltai, T., Tatay, V., and Kalló, N. (2014). Application of the
results of simple assembly line balancing models in
practice: the case of a bicycle manufacturer.
International Journal of Computer Integrated
Manufacturing, 27 (9), 887-898.
McIntyre, E.V. (1977). Cost-Volume-Profit analysis adjusted
for learning. Management Science, 24 (2), 149-160.
Scholl, A., and Becker, C. (2006). State-of-art exact and
heuristic solution procedures for simple assembly line
balancing. European Journal of Operational Research,
168, 666-693.
Yelle, L.E. (1979). The learning curve: Historical review and
comprehensive survey. Decision Sciences, 10, 302-328.
Wright, T.P. (1936). Factors affecting the cost of airplanes.
Journal of Aeronautical Sciences, 3 (4), 122-128.