Download as pdf or txt
Download as pdf or txt
You are on page 1of 22

International Journal of Production Research

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tprs20

Assembly line balancing with parallel workstations

Eduardo Álvarez-Miranda , Sebastián Chace & Jordi Pereira

To cite this article: Eduardo Álvarez-Miranda , Sebastián Chace & Jordi Pereira (2020): Assembly
line balancing with parallel workstations, International Journal of Production Research, DOI:
10.1080/00207543.2020.1818000

To link to this article: https://doi.org/10.1080/00207543.2020.1818000

Published online: 14 Sep 2020.

Submit your article to this journal

Article views: 15

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tprs20
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
https://doi.org/10.1080/00207543.2020.1818000

Assembly line balancing with parallel workstations


Eduardo Álvarez-Mirandaa,b , Sebastián Chaceb,c and Jordi Pereirac,d
a Faculty of Engineering, Department of Industrial Engineering, Universidad de Talca, Curicó, Chile; b Instituto Sistemas Complejos de Ingeniería,

Santiago, Chile; c Universidad Adolfo Ibáñez, Santiago, Chile; d Faculty of Engineering and Science, Universidad Adolfo Ibáñez, Villa del Mar, Chile

ABSTRACT ARTICLE HISTORY


The simple assembly line balancing problem (SALBP) considers work division among different work- Received 19 May 2020
stations of a serially arranged assembly process to maximise its efficiency under workload (cumu- Accepted 21 August 2020
lative) and technological (precedence) constraints. In this work, we consider a variant of the SALBP KEYWORDS
which allows parallel workstations. To study the effect of parallel stations, we propose a new problem Line balancing; parallel
(the parallel station assembly line balancing problem or PSALBP) in which the objective is to minimise stations; dynamic
the number of parallel stations required to obtain the maximum theoretical efficiency of the assem- programming; hybrid
bly process. We study the complexity of the problem and identify a polynomially solvable case. This metaheuristic; matheuristic
result is then used as a building block for the development of a heuristic solution procedure. Finally,
we carry out a computational experiment to identify the characteristics of assembly lines that may
benefit from station paralleling and to evaluate the performance of the proposed heuristic.

1. Introduction and motivation an earlier workstation or to the same one as the second
An assembly line is a production system frequently used task). Among the feasible assignments, the SALBP aims
for the mass production of standardised commodity to optimise line efficiency (i.e.minimise total idle time).
goods. In an assembly line, unfinished products flow The most common variants of the SALBP are the SALBP-
through consecutive workstations. Each station succes- 1 (minimise m for a given c), the SALBP-2 (minimise c for
sively performs a subset of the tasks required to assemble a given m) and the SALBP-E which jointly optimises cycle
the product until the finished product departs the line. time and the number of workstations for a specified range
The assembly line balancing problem, ALBP, comprises on the number of workstations (Scholl and Becker 2006).
a broad family of problems that try to study the optimal While the SALBP is a simplification of the real
assignment of tasks to the workstations of the assembly problem faced in industrial settings, it offers a basic
line. framework to build more realistic models denoted as
Among the different ALBP formulations, the most general assembly line balancing problems, GALBPs (Bay-
thoroughly studied formulation is the simple assembly bars 1986). Consequently, GALBP formulations usually
line balancing problem, referred to as the SALBP in build upon the SALBP formulation by altering some of
the literature. The SALBP depicts a basic common for- its assumptions.
mulation among line balancing problems. The SALBP In this work, we investigate one of the SALBP
(see Scholl and Becker 2006) considers the assembly assumptions. Specifically, we analyse the effect of serially
of a single product whose production process has been arranged workstations on overall line efficiency. While
divided into a set of elementary operations with deter- each serially arranged workstation sequentially performs
ministic and known task times. These tasks are to be per- a different set of tasks on each unit, multiple parallel
formed in m serially arranged, equally equipped work- workstations perform the same set of tasks on different
stations, each with an identical amount of allotted time units simultaneously. As a result, the time allotted to each
known as cycle time c. A solution corresponds to an parallel workstation is larger than the cycle time while
assignment of tasks to workstations such that each work- still achieving the desired assembly line throughput.
station can perform its allotted work, i.e. the sum of the As an example of the implications of parallel work-
assigned task times is no larger than c, and precedence stations, consider the case in which each workstation
constraints among tasks are satisfied (i.e. if a task must performs all the tasks required to assemble the product.
precede another task, then the first task is assigned to The configuration corresponds to m parallel workstations

CONTACT Jordi Pereira jorge.pereira@uai.cl Faculty of Engineering and Science, Universidad Adolfo Ibáñez, Villa del Mar, Chile

© 2020 Informa UK Limited, trading as Taylor & Francis Group


2 E. ÁLVAREZ-MIRANDA ET AL.

(the term parallel refers to, in this context, the lack of a reward from introducing parallel stations within an
serial arrangement since station ordering does not affect assembly line. An evaluation of the improvements intro-
behaviour in the assembly process). In such a configu- duced by station paralleling provides some insights on
ration, the time required by each workstation equals the the involved trade-offs: reduced idle times at the expense
sum of task times; the workstations have no idle time and of additional equipment and labour costs required to
the line throughput is the maximum for m workstations perform more tasks in a single workstation.
(i.e.the time to manufacture a unit, the cycle time, equals This work considers the PSALBP from both perspec-
the time allotted to each workstation divided by the num- tives. As an optimisation problem, we propose a for-
ber of workstations rounded up). In other words, parallel mulation and a solution method. As a tool to test the
workstations partially circumvent cumulative and prece- opportunities raised by introduced limited degrees of sta-
dence constraints of the SALBP and help generate better tion paralleling, we compare optimal solutions from the
solutions. SALBP and the PSALBP to identify conditions where
Even if the configuration may be optimal in terms of station paralleling may be beneficial to line efficiency.
line efficiency, it may not be viable from a practical stand- The remainder of this work is organised as follows.
point, because no amount of parallel workstations would Section 2 provides a literature review on assembly line
achieve the equivalent throughput of a serial assembly balancing with parallel stations and a succinct analysis
line for hard-to-assemble products. This shows some lim- of the state-of-the-art on solution methods for assem-
itations of the SALBP formulation. Specifically, assembly bly line balancing problems. Section 3 formally intro-
lines rely on work specialisation to maximise efficiency, duces the PSALBP and identifies a class of problems that
but the SALBP does not explicitly account for work spe- are polynomially solvable through a dynamic program-
cialisation within its assumptions. It fails to properly con- ming formulation. Section 4 describes a variable neigh-
sider the role of learning curves in assembly lines (Otto bourhood search (VNS) metaheuristic for the PSALBP
and Otto 2014), and to properly evaluate the investment that makes use of previous results (Hoffmann 1963;
costs associated to adequately equipping each worksta- Kramer, Dell–Amico, and Iori 2017; Pereira, Ritt, and
tion with its required tools (Pereira 2018). Vásquez 2018) and the results from Section 3 to define
Some GALBP formulations have tried to remedy this its neighbour structures. Section 5 reports the results of
shortcoming by either acknowledging the opportunity of the computational experiments conducted to compare
using parallel workstations and changing other assump- the SALBP and the PSALBP and to evaluate the qual-
tions to encourage work division (Daganzo and Blu- ity of the proposed solution procedure. Finally, Section 6
menfeld 1994; Becker and Scholl 2009); allowing parallel gives some conclusions and identifies future research
workstations only when serial workstations could not opportunities.
deal with the workload derived from some tasks (i.e.
when operation times of some tasks exceed the cycle time
2. Literature review
(Vilarinho and Simaria 2006); or by introducing alter-
native objectives that encourage division of work, like Work parallelisation is a common practice in assembly
cost-oriented objectives (Bukchin and Rubinovitz 2003; lines and the literature covers multiple works consid-
Tuncel and Topaloglu 2013; Tiacci 2015a)). This work fol- ering different types of parallelisation within assembly
lows an alternative approach. More precisely, if we were lines. A simple classification divides work parallelisation
to allow station paralleling to maximise line efficiency strategies into two main groups. A first group in which
and obtain the additional degrees of efficiency, flexibil- multiple workers perform different operations for a single
ity, and reliability provided by parallel workstations, one product concurrently, and a second group in which mul-
can assume that we would still like to take advantage of tiple workers perform the same operations for multiple
work specialisation by minimising the number of parallel products concurrently.
stations. This leads to an optimisation problem in which The first group corresponds to line balancing config-
line efficiency is a parameter (both the number of work- urations in which multiple workers share a workplace.
stations and the cycle time are known) and we aim to Two-sided assembly lines (Bartholdi 1993; Li, Kucukkoc,
minimise the number of parallel workstations required to and Zhang 2020), multi-manned assembly lines (Dimi-
achieve a line configuration under the remaining SALBP triadis 2006; Cevikcan, Aslan, and Yeni 2020) and vari-
assumptions. We refer to this problem as the parallel sta- able worker assembly lines (Becker and Scholl 2009) are
tion assembly line balancing problem, or PSALBP for among the most frequently studied configurations within
short. this group. The rationale behind the use of these con-
The PSALBP formalises not only the described situ- figurations varies among problems (e.g. products too
ation but also provides a tool to evaluate the expected large to operate on every section, or space limitations
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3

forcing multiple workers to operate simultaneously in The literature also covers station paralleling with other
each workstation) but, in every case, it involves additional common GALBP features like parallel assembly lines
complexities to the basic line balancing problem since (Süer 1998), mixed-model assembly lines (Vilarinho and
the problem must consider interactions among different Simaria 2002, 2006), tasks with stochastic task times
workers in a single workstation. (McMullen and Frazier 1997, 1998) and line balancing
The second group corresponds to assembly line con- problems with embedded scheduling decisions (Öztürk
figurations in which two or more independent worksta- et al. 2013; Lopes et al. 2019).
tions perform the same set of operations for different The literature covers a broad range of methods to
units of product. Parallel workstations provide several solve line balancing problems with parallel worksta-
benefits (Buxey 1974; Bard 1989) including efficiency tions. Among these methods, we highlight constructive
improvements (through the reduction of idle times). heuristics (Sarker and Shanthikumari 1983; Askin and
These benefits come at the cost of increasing equipment Zhou 1997; McMullen and Frazier 1997), enumerative
requirements (more workstations may require specific methods (Buxey 1974; Pinto, Dannenbring, and Khu-
tools), and a loss in terms of worker specialisation. Since mawala 1981; Bard 1989; Bukchin and Rubinovitz 2003),
the problem under study falls within this group, we focus local search methods (McMullen and Frazier 1998),
our review on this group. off-the-shelf integer programming software (Tuncel and
Duplicating a workstation increases the total time Topaloglu 2013) and metaheuristics like ant colony opti-
allotted to these workstations to perform tasks. Usually misation (Vilarinho and Simaria 2006), genetic algo-
the time allotted to each parallel workstation equals the rithms (Akpınar and Bayhan 2011; Tiacci 2015a, 2015b)
number of identical stations times the cycle time of the or simulated annealing (Simaria and Vilarinho 2001).
assembly line (Bukchin and Rubinovitz 2003). Alterna- Table 1 summarises previous work on parallel work-
tively, one can assume that parallel workstations is a pro- stations (other station duplication options are not con-
cess selection assembly line balancing problem (Bukchin sidered). As the table shows, the PSALBP departs from
and Tzur 2000) in which we equip each workstation with the previous models by imposing the maximum line effi-
a certain set of tools that define task times. Consequently, ciency as a hard constraint (hence limiting the total num-
two parallel workstations lead to task times equal to half ber of workstations in the assembly line) and by consid-
the original task times. ering the minimisation of the number of parallel work-
The literature sometimes limits station duplication stations as its main objective. The limit on the number of
when strictly required (Vilarinho and Simaria 2006). For workstations enables us to maintain an efficiency-based
instance, the task time of some task is larger than the cycle objective function (facilitating the comparison between
time, but in most cases station paralleling focuses on cost the solutions of the PSALBP and the SALBP) and to avoid
reduction (Pinto, Dannenbring, and Khumawala 1981) previous constraints on the solution topology impos-
or work imbalance reduction among stations (Sarker and ing which tasks could be parallelised. Finally, this work
Shanthikumari 1983). also offers an alternative metaheuristic solution approach
Cost-oriented objectives (Askin and Zhou 1997; Ege, where the neighbourhood structures make use of spe-
Azizoglu, and Özdemirel 2009; Tuncel and Topaloglu cific problem characteristics like the optimal resolution
2013) focus on the trade-off between station efficiency of a special polynomially solvable case. The resulting
and installation costs. Parallel workstations can only method, which can be seen as a hybrid metaheuristic
improve line efficiency at the expense of additional costs. (Blum et al. 2011) or as a matheurisitic (Maniezzo, Stüt-
Daganzo and Blumenfeld (1994) provides an analytical zle, and Voß 2010), is also a novel contribution within
model to study these trade-offs and concludes that station the literature on assembly line balancing with parallel
paralleling decisions depend on the trade-offs between workstations.
installation and manpower costs. Given the similarity of the PSALBP with other line
Smoothing objectives (McMullen and Frazier 1998) balancing formulations, we briefly address the state-of-
consider improvements on work division introduced by the-art solution methods for the SALBP and GALBP
the additional flexibility gained from parallel lines. In formulations we later use in this work.
Akpınar and Bayhan (2011), McMullen and Tarasewich State-of-the-art methods for the SALBP are usu-
(2006), Simaria and Vilarinho (2001), the authors ally constructive. Both complete (Morrison, Sewell, and
simultaneously optimise cost and smoothing objectives Jacobson 2014) and truncated (Fleszar and Hindi 2003;
while Tiacci (2015a) evaluates these objectives through Hoffmann 1963; Sternatz 2014) versions of the branch-
simulation. and-bound algorithm are common. In Pape (2015), the
4 E. ÁLVAREZ-MIRANDA ET AL.

Table 1. Summary of previous works on parallel workstations. For each paper, we consider its objective function, assembly line char-
acteristics, when they accept parallel workstations and if they impose any limit on them, and the solution procedure described in the
paper.
Reference Objective Special Line Features Parallelism Solution Procedure
Akpınar and Bayhan (2011) line length and work zoning limited to large tasks hybrid genetic algorithm
smoothness
Askin and Zhou (1997) cost mixed model always possible greedy heuristic
Bard (1989) cost none maximum two parallel dynamic programming
stations per stage
Bukchin and Rubinovitz (2003) cost none always possible branch and bound
Buxey (1974) idle time zoning limited per production ad hoc heuristics
stage
Daganzo and Blumenfeld (1994) cost none always possible analytical model
Ege, Azizoglu, and cost equipment always possible branch and bound
Özdemirel (2009)
Lopes et al. (2019) cycle time space limitations and limited by space mixed integer programming
scheduling
McMullen and Frazier (1997) multiple metrics stochastic task times always possible greedy heuristic
McMullen and Frazier (1998) cost, smoothness, and stochastic task times always possible simulated annealing
variability
McMullen and Tarasewich (2006) multiple metrics stochastic task times always possible ant colony optimisation
Öztürk et al. (2013) completion time mixed model and disjunctive always possible mixed integer programming
constraints
Pinto, Dannenbring, and cost none limited to two workstations branch and bound
Khumawala (1981) per stage
Sarker and Shanthikumari (1983) idle time none according to idle time goal greedy
Simaria and Vilarinho (2001) idle time and workstations zoning limited in number and to simulated annealing
large tasks
Süer (1998) cycle time none global limit two phase heuristic
Tiacci (2015a) cost and performance stochastic task times always possible genetic algorithm and
simulation
Tiacci (2015b) cost and performance stochastic task times and always possible genetic algorithm and
buffers simulation
Tuncel and Topaloglu (2013) parallel workstations task assignment constraints limited to large tasks mixed integer programming
Vilarinho and Simaria (2002) line efficiency and incompatibility constraints limited to large tasks simulated annealing
workload balance
Vilarinho and Simaria (2006) line efficiency and mixed model limited to large tasks ant colony optimisation
workload balance
This work parallel workstations maximum efficiency constraint always possible variable neighbourhood search

author provides a comparison of different heuristics 3. Problem formulation


showing the applicability of constructive approaches, but
Since the PSALBP builds upon the SALBP, we first
some local search-based methods are also competitive.
introduce the SALBP and then describe the differences
More recently, Sternatz (2014) improved the Hoffmann
between both problems.
heuristic (Hoffmann 1963), increasing the overall effi-
The SALBP tries to divide set V of the n = |V| ele-
ciency of the method. Recently, Kramer, Dell–Amico, and
mentary tasks required to assemble the final product into
Iori (2017) discuss a novel local search method based on
M disjoint subsets (S1 , . . . , S|M| ) of tasks, and to define an
guiding the search with an auxiliary function.
ordering among the subsets to identify each subset with
For GALBPs, most of the literature focuses on heuris-
a workstation. Each task i ∈ V has an operation time ti , a
tics, especially if there are no tight bounds and dom-
set Pi ⊂ V of immediate preceding tasks and a set Fi ⊂ V
inance rules for the problem on hand (Scholl and
of immediate following tasks. Precedence relations may
Becker 2006; Pereira 2015). Given the relationship
be represented through a directed acyclic graph in which
between most GALBPs and the SALBP, it is com-
the set of vertices correspond to the set of tasks and there
mon to use some know-how learned from the SALBP
is an arc from i to j if task i is a predecessor of task j (i.e. if
to design effective heuristics for GALBPs (Becker and
j ∈ Fi , then there is an arc from i to j). The set of transi-
Scholl 2006; Sternatz 2014; Pereira 2018; Pereira, Ritt, and
tive predecessors (and successors) of task i derived from
Vásquez 2018).
immediate relations are denoted as Pi∗ (and Fi∗ ).
To conclude, we refer the interested reader to (Scholl
A feasible solution to the SALBP satisfies three condi-
and Becker 2006; Dolgui and Battaïa 2013) for detailed |M|
tions: (1) each task is assigned to a workstation (∪j=1 Sj =
reviews of the literature on other line balancing prob-
V), (2) the sum oftask operation times must not exceed
lems and a review of different metaheuristic approaches
the cycle time ( i∈Sj ti ≤ c ∀j ∈ M), and (3) we can
for their resolution.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 5

order workstations in such a way that if task i is a


predecessor of task i then the workstation that per-
forms i has smaller or equal lexicographic order than the
workstation that performs i (if i ∈ Pi , i ∈ Sj and i ∈ Sj
then j ≤ j ). The objective function of the SALBP-1, the
most studied SALBP model (Scholl and Becker 2006;
Dolgui and Battaïa 2013), aims to minimise the cardi-
nality of M, that is, minimise the number of required
workstations, for a production rate (i.e. for fixed cycle Figure 1. Bowman instance.
time c).
A common lower bound on the optimal objective
value of the SALBP-1 corresponds to the trivial bound while a solution in which each workstation performs all
or LB1 (Scholl and Becker 2006), Equation (1). the tasks has one production stage.
Then, the objective of the PSALBP corresponds to the
 
i∈V ti minimisation of the total number of additional worksta-
LB = (1)
c tions among all the production stages of the assembly
line subject to the following constraints: (1) each task is
One could think of the trivial bound as the number of assigned to a production stage, (2) the sum of task oper-
workstations required to achieve maximum efficiency or ation times in each in each production stage does not
as the optimal solution to a problem in which each work- exceed the number of workstations in this stage multi-
station performs all the tasks in V; we aim to find the plied by the cycle time c, (3) the production stages can be
minimum number of such workstations that output no ordered in such a way that precedence constraints among
less than a unit of product each c time units. tasks in different production stages are fulfiled, and (4)
The last description fits the PSALBP description the the minimum number of workstations is used (i.e. the
best, as a SALBP-1 solution corresponds to a completely total number of workstations equals (1)).
serially arranged solution and a solution with LB parallel We give an example of the PSALBP, Section 3.1,
workstations corresponds to a parallel arranged solution. and provide a mathematical formulation, Section 3.2,
The difference between the number of workstations in before addressing the resolution of special PSALBP
both solutions defines the maximum possible gain from cases.
introducing parallelism within the assembly line.
As discussed in the introduction, the PSALBP removes
condition (6) from the SALBP. If one substitutes a group 3.1. An example
of consecutive workstations from the SALBP-1 solu- Consider the classical Bowman instance (Bowman 1960),
tion by multiple parallel workstations, the number of see Figure 1, with cycle time 20. The optimal number
workstations of the former will be no smaller than the of stations for the SALBP-1 is 5 as both task 1 and task
number of workstations of the latter; however, there is 2 cannot be performed with any other task due to the
an opportunity for improvement as the solution may precedence constraints (the remaining six tasks require
require fewer workstations to perform the same set of a minimum of three stations).
tasks. According to the trivial bound, we need a minimum
Specifically, if workstations Su , Su+1 , . . . , Sv are sub- of 4 workstations (the sum of task times is 75, so LB =
stituted by some parallel workstations, such a change will 75
20 = 4). Thus, duplicating some stations may lead to
lead to efficiency improvements if (2) holds. an increase in line efficiency through the reduction of
 v   the number of workstations in one unit. For instance, if
k=u a∈Sk ta the first production stage comprises two parallel stations
< (v − u + 1) (2)
c each performing tasks 1, 2 and 3 (workload equal to 37)
and two additional serial workstations (one performing
Let us refer to the group of workstations that perform the tasks 4, 6 and 8 and another performing tasks 5 and 7),
same subset of tasks as a production stage, also referred the solution has only four workstations and it is optimal
to as a station group according to Daganzo and Blumen- for the PSALBP.
feld (1994), and let us refer to the number of workstations Clearly, the example shows that some degree of station
minus one in a production stage as the number of addi- paralleling can lead to better productivity rates without
tional (or parallel) workstations. Note that a SALBP-1 renouncing to the advantages of task division and worker
solution has as many production stages as workstations, specialisation provided by traditional serial assembly
6 E. ÁLVAREZ-MIRANDA ET AL.

lines as the solution departs only slightly from the SALBP with the ti ≤ c condition. The modified instance has the
solution. same set of precedence relations and task times ti =
ti − ( tci − 1)c. The solution to the modified instance
can be automatically be mapped to the original instance
3.2. An integer programming formulation
with
 theti objective between both workstations varying in
The PSALBP can be formulated using an integer pro- i∈V c − 1 units.
gram, IP, in which the binary variables xik indicate if While not strictly required, constraint sets (11)
production stage k performs task i, binary variables yk and (12) help remove symmetries from the formulation.
represent if the solution requires production stage k, and The experiments reported in Section 5 make use of these
non-negative integer variables zk identify the number of constraints. Also, the cardinality of M is set to be equal to
additional stations associated with production stage k. LB, Equation (1), as constraint (7) ensures that LB is an
Using the previous notation, model (3)–(10) pro- upper bound on the number of production stages.
vides a valid integer programming formulation for the
y1 = 1 (11)
PSALBP.
 yk ≥ yk+1 ∀k ∈ 1, . . . , |M| − 1 (12)
z∗ = min zk (3)
k∈M If a feasible solution is available, i.e. using the methods
 described in Section 4, constraint (11) can be strength-
xik = 1 ∀i ∈ V (4) ened to take into account that a better solution is sought.
k∈M Note that constraint (7) imposes the total number of
 
kxi k ≤ kxik ∀i ∈ V, ∀i ∈ Pi (5) workstations. To reduce the number of parallel worksta-
k∈M k∈M tions, the number of production stages must increase.
  As a result, if a solution with z additional workstations
ti xik ≤ c yk + zk ∀k ∈ M (6) is known, (13) substitutes (11) removing non-improving
i∈V
   solutions from the space of feasible solution.
i∈V ti
yk + z k = (7) yi = 1 ∀i ∈ 1, . . . , LB − z + 1 (13)
c
k∈M

xik ∈ {0, 1} ∀i ∈ V, ∀k ∈ M (8) Note that, as in many other line balancing problems
(Pereira 2015), the linear relaxation of model (3)–(10)
zk ∈ Z≥0 ∀k ∈ M (9) provides a weak lower bound, and thus its applicabil-
yk ∈ {0, 1} ∀k ∈ M (10) ity to solve PSALBP instances is limited to small-sized
instances. Specifically, the linear relaxation of (3)–(10)
Objective (3) minimises the number of additional work- has an optimal solution with objective value 0.
stations per production stage. The objective is equal to To prove the statement, notice that if zk = 0 for each
maximising k ∈ M and we drop constraint set (7) from the PSALBP
 the number of production stages, i.e. max-
imising k∈M yk . Constraint set (4) ensures that all
formulation, the resulting constraints correspond to the
tasks are assigned to a production stage, while constraint constraint set of the SALBP-1. The objective value of the
set (5) ensures technological (i.e.precedence) constraints. linear
 relaxation of the SALBP-1 is known to be equal to
Constraint set (6) ensures the cycle time condition and i∈V ti
c . Consequently, the optimal solution to the linear
enforces enforces use of the production stage if it per- relaxation of the SALBP-1 also fulfils the set of con-
forms any task. Notice that any addition station in pro- straints of the PSALBP, including (7), and has an objective
duction stage k increases the production capacity of the function value equal to 0.
said production stage. Constraint (7) ensures the effi-
ciency condition (that is, the solution contains the min-
3.3. Some polynomially solvable special cases
imum number of stations). Finally, conditions (8), (9)
and (10) define the domain of the variables. The PSALBP is NP-hard as it subsumes the feasibility
The PSALBP model implicitly accounts for the case versions of the bin packing problem (Garey and John-
in which one or more tasks obligatorily require one or son 1979, problem SR1, pg. 226) and the SALBP as special
more parallel workstations (i.e. any task i such that ti > c cases. Moreover, an alternative proof is also given in
holds) as the cycle time constraint (6) will accommodate Appendix. When the order of operations is known (i.e. if
any required parallel workstation. It is also interesting precedence constraints impose a unique feasible order of
to point out that such an instance can be transformed operations among the tasks), the problem is polynomially
into an equivalent instance where every task complies solvable in O(n3 ) time through dynamic programming.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7

To describe the dynamic programming recurrence Table 2. Values for cij .


used to solve the problem in polynomial time, we intro- cij 1 2 3 4 5 6 7 8
duce the following notation. Let (t1 , t2 , . . . , t|V| ) be the 0 0 1 1 2 2 3 3 3
order in which the tasks are performed and let cij (0 ≤ 1 – 0 1 1 1 2 3 3
i < j ≤ |V|) denote the number of additional worksta- 2 – – 0 0 1 1 2 2
3 – – – 0 0 1 1 1
tions required by tasks {ti+1 , ti+2 , . . . tj } if one production 4 – – – – 0 0 1 1
stage performs these tasks, see expression (14). 5 – – – – – 0 1 1
6 – – – – – – 0 0
 j  7 – – – – – – – 0
a=i+1 ta
cij = −1 (14)
c
Table 3. Optimal f (i, k) values.
Expressions (15) and (16) provide the recurrence used fi,k 1 2 3 4
to ascertain the minimum number of additional paral- 1 0 – – –
lel workstations required to perform tasks {1, . . . , i} in k 2 1 0 – –
consecutive production stages, fik . 3 1 1 0 –
4 2 1 0 0
5 2 1 1 0
fi1 = c0i (15) 6 3 2 1 0
7 3 2 2 1
fik = min fj,k−1 + cji (16) 8 3 2 2 1
k−1≤j<i

To check the correctness of (15), note that if the number


of stages equals 1, fi1 corresponds to c0i , Equation (15), To illustrate these calculations, consider the Bowman
and its correctness follows from the definition of c0i . instance introduced in Section 3.1. The task order-
To check the correctness of (16), note that if tasks ing under evaluation corresponds to task numbering
{1, . . . , i} are performed in k consecutive stages, a subset (1, 2, . . . , 8). Table 2 reports the values for cij and fi,k .
{1, . . . j} of tasks are performed in the first k−1 stages and The optimal solution corresponds to 2 as shown in
the remaining tasks, subset {j + 1, . . . i} is performed in entry f8,2 from Table 3 as f8,3 + 3 > 4. To compute entry
stage k. As the optimal number of additional workstations f8,2 , the recurrence considers assigning subsets {1}, {1, 2},
for each subset of tasks in k−1 stages are computed in {1, 2, 3}, {1, . . . , 4}, {1, . . . , 5}, {1, . . . , 6} or {1, . . . , 7} to
advance, and the recurrence compares each alternative, the first stage and the remaining tasks to the second stage
correctness holds. by calculating (16), see (18).
Once the complete recurrence is computed, the opti- ⎧ ⎫ ⎧ ⎫
mal objective value of the special PSALBP is obtained by ⎪
⎪ f1,1 + c1,8 ⎪⎪ ⎪
⎪ 0 + 3⎪⎪

⎪ + ⎪
⎪ ⎪
⎪ + ⎪

identifying the largest k such that f|V|,k + k equals LB and ⎪
⎪f 2,1 c 2,8 ⎪
⎪ ⎪
⎪ 1 2 ⎪

⎪ ⎪ ⎪ ⎪
the optimal objective value corresponds to f|V|,k . ⎨f3,1 + c3,8 ⎬ ⎨1 + 1⎪
⎪ ⎪ ⎪ ⎬
Finally, we verify the complexity of O(n3 ) where n f8,2 = f4,1 + c4,8 = 2 + 1 = 2 (18)
⎪ ⎪ ⎪ ⎪
equals |V|. First, notice that the maximum number of ⎪

⎪f5,1 + c5,8 ⎪ ⎪



⎪ 2 + 1⎪⎪


⎪ ⎪
⎪ ⎪
⎪ ⎪

production stages is bounded by the number of tasks, ⎪
⎪f 6,1 + c ⎪
6,8 ⎪ ⎪
⎪ 3 + 0 ⎪

⎩ ⎭ ⎩ ⎭
hence, there are at most O(n2 ) values for f (i, k), each tak- f7,1 + c7,8 3+0
ing at most O(n) operations to evaluate. Therefore, the
time complexity of the calculations is bounded by O(n3 ). The optimal value comes from f3,1 , and corresponds to
Other operations, like calculating cij and retrieving the performing tasks 1, 2 and 3 in two workstations in the first
optimal objective value, are dominated by the described stage and the other tasks in two workstations in stage 2.
operation. Hence, the above recurrence provides a solu- Note that the optimal solution differs from the one found
tion for the special case under consideration in O(n3 ) in Section 3.1 as only task orderings (1, 2, 3, 4, 6, 8, 5, 7)
time and thus, the problem is polynomially solvable. or (1, 2, 3, 5, 7, 4, 6, 8) may lead to the optimal solution for
Notice that if we impose an additional condition to the PSALBP instance under consideration.
limit the number of stages to two, the problem can be
solved in O(n) as we have to evaluate (17) which can be
easily performed by updating c0,j and cj,|V| while evaluat- 4. A matheuristic-based variable
ing the minimum. neighbourhood search for the PSALBP
Given the complexity of the PSALBP and the lack of
f|V|,2 = min fj,1 + cj,|V| = min c0,j + cj,|V|
1≤j<|V| 1≤j<|V| good lower bounds to rely on an enumeration proce-
(17) dure (the bounds for the SALBP do not apply to the
8 E. ÁLVAREZ-MIRANDA ET AL.

PSALBP), we consider a heuristic approach that com- by verifying precedence constraints and the number of
bines the results provided in Section 3.3, local search workstations required by the solution.
procedures for previous line balancing problems (Scholl Local search procedures based on the stage represen-
and Voss 1997; Kramer, Dell–Amico, and Iori 2017) and tation use the move and swap operators.
a state-of-the-art heuristic for the SALBP (Sternatz 2014) A move, denoted as (i, su , sv ), corresponds to chang-
within the variable neighbourhood search (VNS) meta- ing the production stage of task i from production
heuristic framework (Mladenović and Hansen 1997). stage su to production stage sv . A swap, denoted as
The VNS provides a method to explore different neigh- (i, j, su , sv ) is a combination of two consecutive moves,
bourhoods in an orderly fashion and has been used move (i, su , sv ) and move (j, sv , su ). We also define a spe-
to solve other line balancing problems (Kellegöz 2016; cial move, denoted as (i, su , −) or (i, su , +), in which we
Polat et al. 2016; Sadeghi, Rebelo, and Ferreira 2018). We remove task i from production stage su , create a new
can see the resulting method as a hybrid metaheuristic production stage before (after) production stage su , and
(Blum et al. 2011) or a matheuristic (Maniezzo, Stüt- assign the task to the new production stage.
zle, and Voß 2010) as it integrates different exact and To decide if we accept a move or a swap, one
heuristic components within a single solution approach. checks its feasibility and its suitability (objective value
The implementation uses three different neighbourhood improvements).
structures. Checking for feasibility requires verifying the prece-
The first neighbourhood structure, Section 4.1, tries dence constraints and the number of workstations
to improve a solution by changing the assignment of required by the solution. Both operations are per-
tasks between different production stages. The neigh- formed by updating pre-calculated values (see Scholl and
bourhood stems from previous classical neighbourhood Voss 1997; Bautista and Pereira 2007).
structures for assembly line balancing problems (Scholl Regarding the suitability of a move or a swap, note that
and Voss 1997; Scholl and Becker 2006) and uses an aux- only backward and forward special moves may improve
iliary function to guide the search (Kramer, Dell–Amico, the objective function value of the incumbent solution,
and Iori 2017). i.e. the PSALBP ensures maximum efficiency by impos-
The second neighbourhood structure, Section 4.2, ing that the assembly line is made up of exactly the min-
explores alternative task orderings obtained through imum possible number of workstations. Consequently,
move and swap operators and evaluated using the and to reduce the number of parallel workstations, one
dynamic programming formulation described in has to increase the number of production stages while
Section 3.3. maintaining the total number of workstations. Hence, the
The third neighbourhood structure, Section 4.3, basic move and swap operators do not change the number
adopts ideas from the Hoffmann heuristic, the state- of stages if a solution is to remain feasible.
of-the-art constructive heuristic for the SALBP and The SALBP-1 suffers from a similar issue as the solu-
some GALBPs (Hoffmann 1963; Sternatz 2014), as a tion landscape has few improving moves (Scholl and
local search procedure. The method optimises reduced Voss 1997). To remedy this issue, one may solve a
instances of the PSALBP using the enumerative compo- sequence of problems with an alternative solution land-
nent of the Hoffmann until local optimality is reached. scape (Scholl and Voss 1997; Bautista and Pereira 2007)
Section 4.4 discusses implementation details includ- or one may use an auxiliary function to guide the search
ing the conversion between different solution represen- (Kramer, Dell–Amico, and Iori 2017).
tations used by each of the neighbourhoods. In this work, we adopt the second approach and focus
on an auxiliary function similar to the one proposed in
Kramer, Dell–Amico, and Iori (2017). The auxiliary func-
4.1. Stage-based neighbourhoods
tion z(s|V| ), see Equation (20), tries to create production
Let s|V| be a vector of size |V| such that i-th position of the stages favourable for the application of special moves (i.e.
vector encodes the production stage that performs task those that can improve the solution). Let Tk denote the
i. We refer to this representation as the stage-based rep- sum of task times of production stage k according to the
resentation of a solution and it has been widely used in task assignment s|V| , then (19) provides the contribution
the literature for line balancing and other related prob- to the auxiliary function for production stage k.
lems (Scholl and Voss 1997; Bautista and Pereira 2007;
Pape 2015; Kramer, Dell–Amico, and Iori 2017). Note
that not every stage-based encoding provides a feasible   2
Tk
solution, but all feasible solutions can be encoded using |V| c − Tk if T k > c
wk (s )= c (19)
the stage representation and feasibility is easily checked 0 otherwise
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 9

wk (s|V| ) provides smaller (better) values for production As in the stage-based neighbourhoods, one has to
stages with two or more workstations and small idle check the feasibility and suitability of a solution gener-
times. Moreover, the square operator amplifies the pref- ated through a move or a swap. We check the feasibility
erence for such extreme solutions. As a side effect of of the solution by incrementally updating pre-calculated
prioritising the generation of these production stages, values as in Section 4.1. We check suitability by evaluating
other production stages will tend to have larger idle times, the list using one of the dynamic formulations described
increasing the chances of finding feasible special moves. in Section 3.3, leading to different neighbourhood
The auxiliary function also encourages moves and swaps structures.
that lead to increased idle time of production stages with We consider two different neighbourhood structures,
one workstation. In order to avoid such an undesired namely (1) a neighbourhood structure that simultane-
effect, we reject moves or swaps that increase the idle time ously explores a restricted subset of swaps and moves
of a production stage with only one workstation and we involving tasks assigned to one production stage, and
accept any move or swap that decreases the idle time of (2) a neighbourhood structure that considers moves and
such a production stage. swaps involving tasks in any production stage. The first
structure uses the linear time recurrence to evaluate

z(s|V| ) = wk (s|V| ) (20) the solution and the second structure uses the general
k∈M recurrence.
Exploration of the search space becomes impractical
The procedure simultaneously explores the move and for very large size instances (instances with n = 1000).
swap neighbourhoods while trying to optimise (20). Evaluating the objective function of any neighbourhood
After performing any move or swap, special moves are solution is O(n3 ) and each solution has a number of
tested and if one is found to be feasible, it is implemented. neighbours quadratic to the number of tasks, hence the
time order of the heuristic and the size of the instances
make the method impracticable. Consequently, for very
4.2. List-based neighbourhoods
large size instances, we do not explore this neighbour-
The PSALBP can be seen as the combination of a hood.
sequencing problem (defining the assembly sequence)
and an assignment problem (the problem of assigning
4.3. The Hoffmann-based neighbourhood
tasks to production stages). The results from Section 3.3
show that the second problem, assignment, is polynomi- The Hoffmann heuristic is a truncated enumeration
ally solvable if the solution to the first problem is known. scheme originally proposed as a heuristic method for the
Hence, it makes sense to consider the search for different SALBP (Hoffmann 1963). It is a state-of-the-art heuris-
assembly sequences and to evaluate their performance by tic for the SALBP (Morrison, Sewell, and Jacobson 2014;
optimally assigning tasks to production stages. Sternatz 2014) and for some GALBP formulations (Ster-
Consequently, a solution is represented as a list, a per- natz 2014; Pereira 2018; Pereira, Ritt, and Vásquez
mutation, of tasks. The list is said to be feasible if the order 2018).
in the list complies with the precedence constraints, i.e. if The method is a truncated branch-and-bound that
task i precedes task j, task i must appear before task j in generates a large number of feasible assignments for
the list. the first (last) unassigned workstation, chooses the best
The list-based (or permutation-based) representation among the generated assignments according to some
enables the use of classical neighbourhood operators for heuristic selection rule, and repeats the previous steps
these problems. Specifically, we use the move and swap until it finds a complete solution. While the theoreti-
operators. A swap corresponds to exchanging the posi- cal complexity of the method does not improve over
tion of two elements within the list, and a move corre- the complexity of a branch-and-bound algorithm, several
sponds to sliding an element from its current position in improvements over the original proposal avoid falling
the list to a different position in the list. into its worst-case behaviour.
Let σi be the task in the ith position (i = 1, . . . , |V|) of Recently, the enumeration component of the heuris-
the list and let σi−1 be the position of task i in the list. tic has been used to define a local search procedure for
A swap (i, j) between tasks i and j involves exchanging a GALBP formulation (Pereira, Ritt, and Vásquez 2018).
σi and σj (with the resulting changes to σi−1 and σj−1 . The method in Pereira, Ritt, and Vásquez (2018) re-
A move of task i to position k is a sequence of swaps optimises the assignment of tasks and resources of two
(i, i + 1) ∪ (i + 1, i + 2) ∪ . . . ∪ (k − 1, k) or (i − 1, i) ∪ consecutive workstations. The method acts as a local
(i − 2, i − 1) ∪ . . . ∪ (k, k + 1). search method whose neighbourhood corresponds to all
10 E. ÁLVAREZ-MIRANDA ET AL.

the solutions involved in a subproblem found within the 4.4. Implementation details
original problem. The method proposed for the PSALBP
The VNS metaheuristic implementation is defined by
tries to divide one production stage from the incum-
its neighbourhood structures, its order of application,
bent solution into two production stages, a partition
a method to generate an initial solution and a restart
which leads to improving the objective function value
mechanism when the method reaches local optimality.
by one unit (i.e. adding one production stage to the
Our implementation also needs a mechanism to convert
solution reduces the number of parallel workstations
solutions between different representations as different
by one).
neighbourhood structures use different representations.
The partition of a production stage into two consecu-
We proceed to illustrate each component separately.
tive stages can be formalised as follows. We are given the
Initial solution. i We use the relationship between the
set of tasks Sk from production stage k, each with its cor-
SALBP and the PSALBP to obtain the initial solution. The
responding task time and precedence relations, and we
procedure starts from a feasible solution for the SALBP
aim to find two subsets Sk1 ∪ Sk2 = Sk such that prece-
and iteratively groups on consecutive production stages
dence constraints are fulfilled (i.e. if i ∈ Sk1 , j ∈ Sk and j ∈
until we obtain a feasible PSALBP solution. The method
Pi then j ∈ Sk1 ). For the partition to improve the incum-
can be seen as repair heuristic. The repair heuristic is
bent solution, the total number of workstations should
initialised with a feasible solution for the SALBP. Each
not change; i.e. Equation (21) must hold.
workstation of the initial solution is mapped to a produc-
      tion stage of the PSALBP and if the number of production
i∈Sk ti i∈Sk1 ti i∈Sk2 ti stages coincides with the trivial bound, i.e.constraint (7)
= + (21)
c c c is satisfied, the solution is already feasible for the PSALBP
and we stop. Otherwise, the method groups a subset
Finding if such a partition exists is NP-hard in strong of consecutive stages to reduce the number of worksta-
sense, see Appendix 1. Nevertheless, the method enumer- tions, i.e. we seek k consecutive production stages such
ates alternative assignments for the first stage (the second that the number of workstations required to perform
stage corresponds to the remaining tasks) until a speci- their assigned tasks is strictly smaller than the number
fied number of alternative assignments has been enumer- of workstations required by the production stages sepa-
ated. Consequently, it heuristically solves the problem by rately. Such a grouping always exists (grouping all pro-
truncating the exact search. duction stages into a single production stage always ren-
The implementation enumerates the set of assign- ders a solution with the desired number of workstations)
ments for the first production stage using the recur- but multiple groupings may exist. In such a case, the
sive method described in Fleszar and Hindi (2003). The method selects a random grouping among those with the
recursion enumerates feasible assignments by adding smallest cardinality. Our implementation uses the solu-
tasks to a tentative assignment of tasks for the first stage. tion for the SALBP provided by the Hoffmann heuristic
Tasks are added according to their order in a list that as implemented in Morrison, Sewell, and Jacobson (2014)
constitutes an input of the method and the results may and available at http://www.assembly-line-balancing.de.
vary depending on that given order. The ordering used in Neighbourhood implementation details. First, we
this work orders the tasks of the considered production explore the stage-based neighbourhoods described in
stage according to precedence constraints (i.e. a preced- Section 4.1. The implementation first considers all pos-
ing task is located earlier in the list than a succeeding sible task swaps involving tasks according to their lexi-
task) with ties broken lexicographically (i.e. according to cographic order (provided by the instance file); that is,
task number in the instance file). Whenever the method swaps (i, j, su , sv ) are evaluated according to the non-
identifies a subset of tasks, it is checked for feasibil- decreasing lexicographic order of i with ties broken
ity, Equation (21). This method stops after finding any according to the lexicographic order of j. Moreover, and
improving assignment or reaching a limit on the num- to avoid symmetries, we impose i < j for a swap to be
ber of assignments enumerated (much like in Morrison, considered. Whenever an improving move is found, it is
Sewell, and Jacobson 2014). Consequently, if the limit implemented and special moves are tested. Only special
truncates the search, the ordered list explicitly defines moves from the production stages involved in the change
which subsets of the assignments are considered. The are evaluated, and the order of evaluation corresponds
number of candidate assignments is a parameter of the to their lexicographic order (earlier production stages
algorithm and fine-tuned in a preliminary experiment, first). Special moves are then tested according to the lex-
see Section 5. icographic order of the task with (i, su , −) special moves
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 11

being evaluated first and (i, su , +) moves being evaluated differences between the lists. Hence, it is beneficial to
afterwards. consider changes in later positions first because they
After reaching local optimality for the swap require less computations.
neighbourhood, the method considers moves. The order The proposed order among the three neighbour-
in which moves are considered is according to the lexi- hoods is based on preliminary experiments to reduce the
cographic order (tasks with smaller lexicographic order time required to reach local optimality for each neigh-
are considered first). For any given task, moves are evalu- bourhood. Moreover, experiments showed that the swap
ated in increasing order of the receiving production stage. neighbourhood is smaller and thus faster to compute,
Whenever a move improves the auxiliary objective func- hence the proposed order for task- and list-based neigh-
tion, it is automatically implemented and both the receiv- bourhoods.
ing and the departing production stages are explored Perturbation mechanism. The perturbation mecha-
for special moves. As in the previous case, the earliest nism tries to force a move that ensures that the solution
production stage is considered first, with tasks within escapes from its current local optimum. The perturbation
each production stage being considered in lexicographic performs multiple random swaps in the list-based repre-
order, first for a (., ., −) move, and then for a (., ., +) sentation of the solution. Specifically, the local solution is
move. Whenever a special move improves the incum- restarted by performing k random swaps.
bent solution, it is automatically implemented. Once a Solution representation. The method uses two alter-
swap is implemented, the procedure continues evaluat- natives to represent a solution. The stage-based and the
ing other tasks according to the proposed rules until local Hoffman-based methods use a stage-based representa-
optimality is verified. tion where the solution indicates which production stage
Moreover, if any move improves the solution, feasible performs each task, while the list-based method encodes
swaps are explored again (see Algorithm). the solution as a sequence, a permutation, of tasks. Con-
After reaching local optimality for the stage-based sequently, we have to devise methods to switch between
neighbourhood, we explore the Hoffmann-based these representations as needed.
neighbourhood. The Hoffmann neighbourhood consid- There is no one-to-one correspondence between these
ers each production stage in order starting from the first representations. While it is possible to map each list
production stage to the final production stage. As in the representation to a unique stage representation through
stage-based neighbourhoods, whenever an improving the resolution of the dynamic programming recurrence
neighbour solution is found, it becomes the incumbent described in Section 3.3, there is more than one list for
solution, the search is resumed with the next production any stage encoding. In fact, any ordered list of tasks such
stage until no more production stages can be improved that (2) preceding tasks appear in the list before succes-
with the said neighbourhood. sors, and (2) tasks in earlier production stages appear in
Finally, we consider the list-based neighbourhood. the list before tasks in later production stages, is a valid
Like the stage-based neighbourhoods, we first consider list representation of the original stage-based solution.
swaps and then moves, but in this case, the order in which Besides this multiplicity, there is no guarantee that decod-
tasks are considered follows their order in the solution. ing the list back to a stage representation provides the
Specifically, tasks are considered for a move in reverse original stage representation, and it is even possible that
order to their position in the list. Pairs of tasks are con- the decoding step improves the objective function value
sidered for a swap in reverse order to the last position of the solution.
of the two tasks in the list with ties broken according to Consequently, when the method converts a solution
the position of the other task. Moreover, we first limit from stage to list representation, we generate multiple
the case study to swaps affecting tasks in two consecutive alternative orders complying with the imposed condi-
production stages before considering general swaps and tions and select the one with the best objective value to
moves. become the list-based incumbent solution. The number
The proposed ordering tries to minimise the effort of lists generated is an algorithmic parameter controlled
spent evaluating the list representation using the dynamic in a preliminary experiment reported in Section 5.1.
programming recurrences described in Section 3.3. First,
if only two stages are involved, there is a linear time
4.5. Overall structure of the algorithm
evaluation method. Second, given the particularities of
the dynamic programming recurrence, the calculations Algorithm 1 outlines the final algorithm. Line 1 obtains
required to evaluate the incumbent solution can be an initial solution and becomes the best-known solution.
reused to evaluate any neighbour solution. Specifically, The functions executed in lines 3, 4, 5, 7, and 8 apply
the recurrence table does not change until there are the neighbourhood structures according to increasing
12 E. ÁLVAREZ-MIRANDA ET AL.

computational loads. Once local optimality is verified, memory requirements are low, the 32 processors were
the best-known, is updated. Finally, the solution is per- used to solve 32 independent instances simultaneously.
turbed, line 10, and the loop is repeated until the termi- The benchmark set for the SALBP described in Otto,
nation condition is met. Otto, and Scholl (2013) is used in the experiment.
The instances are available at http://www.assembly-line-
Algorithm 1 Variable Neighbourhood Search balancing.de and follow a factorial design that enables
implementation evaluation of the impact of different instance character-
Input: A PSALBP instance. istics in algorithm performance and the structure of the
Output: A solution to the PSALBP. optimal solutions
1: Generate an initial solution through the Hoffmann-
based heuristic. Initialise best-known. • Instance size. The instances are divided into four
2: while Termination condition is not satisfied do
groups according to the number of tasks (n): small
3: Explore the stage-based swap neighbourhood (n = 20), medium (n = 50), large (n = 100), and very
(Section 4.1) according to the description given large (n = 1000).
in Section 4.4. Repeat this step until no further • Task time distribution. The instances are divided into
changes are possible. three groups: bottom task times in which tasks have an
4: Explore the stage-based move neighbourhood average operation time equal to 10% of the cycle time;
(Section 4.1) according to the description given middle task times in which the average task time is 50%
in Section 4.4. Repeat this step until no further of the cycle time; and mixed task times in which both
changes are possible and go to line 3 if an improv- bottom and middle tasks are present.
ing move was found. • Order strength of the precedence relations. The order
5: Explore the Hoffmann-based neighbourhood strength (OS) corresponds to the ratio between the
(Section 4.3) according to the details outlined in number of precedence relations and the maximum
Section 4.4. Go to line 3 if the step improves the number of precedence relations among tasks, (22).
incumbent solution. Consequently, the larger the order strength, the
6: Convert the incumbent solution from the stage larger the number of precedence relations among
representation to the list representation tasks. Three values are used: low (OS = 0.2), medium
7: Explore the list-based swap neighbourhood (OS = 0.6), and high (OS = 0.9) instances are
(Section 4.2) according to the details provided in considered.
|V| ∗
i=1 |Fi |
Section 4.4.
Explore the list-based move neighbourhood (22)
8: |V|(|V| − 1)/2
(Section 4.2) according to the details provided
in Section 4.4. If line 7 or line 8 improved the • Structure of the precedence relations. The set contains
incumbent solution, go to line 3. three different precedence structures: chain struc-
9: Convert the incumbent solution from the list rep- tures in which precedence relations were generated to
resentation to the stage representation. ensure that there are ‘chains’ of multiple tasks with
10: Perturb solution using the restart mechanism (see one predecessor and one successor; bottleneck struc-
Section 4.4). tures in which precedence relations were generated
11: end while
to ensure that there are ‘bottleneck’ tasks (i.e. there
are some tasks that have multiple predecessors and
multiple successors); and mixed structures in which
precedence relations were randomly generated. Chain
and bottleneck structures try to emulate common pro-
5. Computational experiment
duction patters (a ‘chain’ corresponds to subsets of
To test the proposed methods and determine possi- operations that need to be performed in a fixed order,
ble improvements derived from workstation paralleling, while ‘bottleneck’ tasks correspond to subsets of tasks
the models and algorithms described in this work were that can be performed in any order followed by a
implemented in C++, compiled using the GCC com- common successor, like a quality control operation.
piler version 8.3.1 and run on a 32 processor Intel Xeon
E5-2650 2 GHz CPU with 128GB RAM running the For each combination of characteristics, the instance
Linux operating system. The IBM CPLEX library ver- set contains 25 instances (with the exception of instances
sion 12.7.1 is used to solve integer linear programming with high order strength in which only mixed prece-
models. Since the code does not use any parallelism and dence graphs are feasible). Consequently, the set contains
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 13

a total of 2,100 instances. For details on the instances, Table 4. Algorithmic parameters for 600s time limit. For each
the construction method and the rationale behind the parameter of the VNS (rows), the initial range of values provided
different characteristics is provided in Otto, Otto, and to irace, column Initial range, and the best combination of param-
eters, column Best, found by the method are reported.
Scholl (2013). It is also worth mentioning that the opti-
mal SALBP solution for all small and medium instances, Best

as well as the majority of large and very large instances, Parameter Initial range n ≤ 100 n = 1000
is known. Moreover, for a large subset of these instances, Perturbation [1,25] 1 1
Stage to list [1,100] 4 69
the optimal SALBP objective value is equal to the triv- Hoffmann [1,10000] 4040 9622
ial bound (i.e. there is an assignment of tasks to stages
with the minimum number of workstations even with-
out requiring parallel workstations), and thus the optimal parameter settings. Irace can be seen as a heuristic
PSALBP solution is 0. Note that the existence of such approach to obtain parameter configurations for a heuris-
solutions does not imply that these instances are trivial tic that repeatedly performs the following steps: (1) it
(Otto, Otto, and Scholl 2013, for a discussion), but rather samples the parameters used in the metaheuristic from
that the optimal solution is known. a given distribution (in our case, a discrete distribution
We report the results of three different experiments: for each parameter); (2) it selects the best parameter
configurations by evaluating their performance in some
• A first experiment, see Section 5.1, to calibrate the instances (i.e. solving the instance for a fixed amount of
parameters of the proposed VNS using an iterated time with the said set of parameters) and then comparing
Friedman as implemented in the R irace package their results by performing a Friedman non-parametric
(López-Ibáñez et al. 2016). test; and (3) it updates the distribution to favour the best
• A second experiment, see Section 5.2, to evaluate the performing (statistically better) configurations found in
performance of the VNS when compared to three the second step. These steps are repeated until a compu-
alternative procedures that act as baseline methods. tational budget (in our case, a total of 10, 000 evaluations
• A third experiment, Section 5.3, to compare the opti- each with a 600s time limit) is consumed. We conducted
mal solutions of the PSALBP and the SALBP in small- two separate experiments, a first one to identify param-
sized instances. This experiment tries to draw some eters to solve small, medium and large instances and a
insights on the possible benefits of station paralleling second to identify parameters for very large instances.
on assembly line efficiency. The difference tries to account for changes derived from
the inability to use list neighbourhoods in very large
We do not report results with instances where task instances.
times are longer than the cycle time because these The experiments use random instances with middle
instances can be solved by solving a related instance task time distributions. We chose these instances as they
where the task times are no longer than the cycle time, are hard to solve for the SALBP and we expect similar
see Section 3.2. behaviour for the PSALBP.
Table 4 reports the initial parameter ranges provided
to irace and summarises the results reported by each
5.1. Parameter settings
experiment as well as the best parameter settings found in
The proposed VNS uses three parameters; namely: the experiment. These settings highlight the differences
between instance size and deserve some analysis.
• The number of swaps to perturb an incumbent solu- First, the larger the instance size, the larger the num-
tion after reaching local optimality (row ‘Perturbation’ ber of partial assignments to enumerate with the Hoff-
in Table 4). mann method. Very large instances benefit from test-
• The number of alternative lists to evaluate when con- ing a larger number of trial permutations to move from
verting a solution from the stage to the list representa- stage to list representation. We believe that both results
tion (row ‘Stage to list’ in Table 4). relate to the lack of list-based neighbourhoods when
• The number of candidate assignments considered by solving very large instances. As an alternative, the VNS
the Hoffmann heuristic before stopping the search relies on alternative methods. Specifically, the method
(row ‘Hoffmann’ in Table 4). to transform a solution between representation can be
seen as a search method that tests similar (i.e. neighbour)
To fine-tune the parameter values, we use the irace lists. This interpretation is confirmed by the degrada-
package (López-Ibáñez et al. 2016). Irace implements tion in solution quality if a smaller number of lists is
an iterated Friedman race to compare among candidate used.
14 E. ÁLVAREZ-MIRANDA ET AL.

Second, the number of perturbations required to Table 5. Results from the proposed algorithms, columns IP, ran-
restart the search is significantly small. In both cases, dom, repair and VNS. The table reports for each group of tasks (the
we perform a single random swap between two tasks in rows), the number of instances, column #, and the number of best-
known solutions found, column # best, and average absolute gap,
the list-based representation. We believe that the small column gap, provided by each algorithm.
number of perturbations should be attributed to the use
IP random repair VNS
of a random method to move between solution rep-
resentations. As randomness plays an important role Parameter # #best gap #best gap #best gap #best gap

in the transformation, minor changes are sufficiently n


20 525 523 0.00 522 0.01 382 0.64 525 0.00
strong to escape from the local optimality reached by 50 525 342 2.79 310 1.86 324 3.10 525 0.00
the structured neighbourhoods. Consequently, the con- 100 525 240 9.67 116 7.49 303 7.44 525 0.00
1000 525 0 201.97 0 156.36 322 62.41 525 0.00
tribution of the traditional restart mechanism is marginal
Task Times
and moving between alternative representations provides Small 700 501 34.13 462 22.77 685 0.26 700 0.00
enough diversity to the search method. Middle 700 174 71.82 185 57.51 49 50.96 700 0.00
Bimodal 700 430 54.88 301 44.00 597 3.97 700 0.00
Third, the above parameters have been fine-tuned
OS
considering a time limit of 600s. Consequently, they 0.2 900 490 57.35 391 45.35 620 15.73 900 0.00
account for the trade-offs between methods for this time 0.6 900 475 53.15 406 41.05 583 19.36 900 0.00
0.9 300 140 43.78 151 30.80 128 23.51 300 0.00
limit. For instance, limiting the number of candidate
Graph
assignments in the Hoffmann heuristic reduces the time Block 600 326 56.73 269 44.41 405 16.69 600 0.00
spent on that heuristic. As a result, the method may per- Chain 600 319 53.96 264 42.15 395 17.94 600 0.00
form a larger number of VNS iterations at the expense Mixed 900 460 51.29 415 38.96 531 19.85 900 0.00

of not finding an improving solution in the Hoffmann


neighbourhood even when one exists.
we report the number of best-found solutions and the
average absolute gap (average number of additional par-
5.2. Results for the VNS
allel workstations required when compared to the best-
To test the quality of the results obtained by the VNS known) among the group of instances.
implementation, we compare its performance with three The results suggest that the improvements of the VNS
alternative methods. metaheuristic are significant for all instance character-
The first method, which we will refer to as IP, istics and sizes. The VNS outperforms each method in
corresponds to resolving the formulation depicted in every group of instances, but it is worth mentioning
Section 3.2 using the IBM CPLEX commercial solver that the IP is a viable option to solve smaller (n = 20)
with no initial solution provided. instances.
The second method, which we will refer to as ran- Table 5 hints at the limitations of the different meth-
dom, generates random lists of tasks that comply with ods. The applicability of the IP is limited to small
the precedence constraints and evaluates them using instances, and the random method fails to outperform
the dynamic programming recurrence described in the IP in terms of best-known solutions (it does consis-
Section 3.3. tently improve average absolute gaps). When we compare
The third method, which we will refer to as repair, the random approach and the repair heuristic, we observe
corresponds to the initial solution provided by the repair that the results of the random heuristic are more uniform
heuristic used to initialise the VNS. across instance characteristics, while the repair heuristic
These methods are referred to as baseline methods provides very good solutions except for the middle task
since they compare the performance of the VNS to (1) time distribution in which its performance deteriorates
resolution through a general purpose solver, (2) applica- significantly. We observe the same behaviour for every
tion of the theoretical properties described in this paper method with the exception of the VNS, but it is more
without the metaheuristic framework and (3) use of exist- remarkable for the repair heuristic.
ing SALBP procedures to tackle the PSALBP. When we evaluate the performance of the VNS meta-
Each method, except for the repair heuristic, is run heuristic, we observe that it obtains the best-know solu-
with a time limit of 600 s. The repair heuristic has no time tions for every instance, a logical result that it contains the
limit and takes much less than the allotted time for the repair heuristic as the initial solution and a guided ran-
rest of heuristics, always less than a few seconds in our dom heuristic as a component of the method. The differ-
computing environment. ences between the VNS and the other methods increase
Table 5 provides a comparison among these methods as the instances become more difficult. Specifically, the
according to instance characteristics. For each algorithm, major gains provided by the VNS correspond to instances
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 15

Table 6. Results for the proposed algorithms, columns IP, ran- distribution of task times of the related instance. The
dom, repair and VNS, for large (n = 100) instances. The table VNS results are no worse than the alternative method for
reports for each group of instances grouped according to order every task time distribution, the recommended solution
strength, column OS, graph type, column graph, and task time
distribution, column Task times, the average minimum number method would not vary.t
of production stages, column Bound, and the average number of Finally, a repeated measures ANOVA test was con-
production stages provided by each method. ducted to conclude whether there were significant differ-
OS Graph Task times Bound IP random repair VNS ences among the four proposed methods. Each instance
0.2 Block bottom 14.12 13.52 13.12 14.12 14.12
was considered as a block and the four algorithms as
middle 50.40 14.40 19.76 18.60 38.80 the treatments. The results of the ANOVA test indi-
bimodal 22.64 20.48 17.32 22.64 22.64 cate significant differences (p-value < 0.0005), but the
Chain bottom 14.28 14.08 13.56 14.28 14.28
middle 49.76 10.48 17.56 15.28 36.04 residuals were not normally distributed. Consequently,
bimodal 22.88 17.92 16.44 22.64 22.80 we performed a Friedman non-parametric test to ensure
Mixed bottom 14.12 13.40 13.08 14.12 14.12
middle 50.48 13.84 19.92 20.00 38.84 that the differences were significant even under non-
bimodal 22.76 20.24 16.64 22.76 22.76 normality assumptions. The test reports significant dif-
0.6 Block bottom 14.24 13.60 13.24 14.24 14.24 ferences among solution methods with a p-value <
middle 50.32 13.84 22.20 19.24 39.32
bimodal 22.52 17.92 16.32 21.64 22.32 0.0005.
Chain bottom 13.92 13.36 13.08 13.92 13.92
middle 50.52 7.68 16.40 10.16 30.92
bimodal 22.76 16.60 16.36 19.20 22.48 5.3. Comparison between the PSALBP and the
Mixed bottom 14.32 14.04 13.56 14.32 14.32
middle 50.16 10.04 17.52 12.32 33.40 SALBP solutions
bimodal 22.92 18.00 16.84 19.60 22.68
0.9 Mixed bottom 14.12 11.28 12.40 10.28 13.08
This section further investigates the differences between
middle 50.80 7.60 17.60 10.12 25.28 the SALBP and the PSALBP to identify assembly process
bimodal 23.16 11.68 16.88 11.36 20.72 characteristics that may benefit from station paralleling.
To avoid interference from specific solution methods
in the analysis, we limit the study to instances solved to
with very large (n = 1000) and middle task time distri- optimality. The results from Table 5 suggest that CPLEX
butions. These results agree with previous results for the can solve small instances to optimality. Consequently, we
SALBP as these are the characteristics found in the most solved all small instances to optimality and compare their
difficult instances for the SALBP-1. number of production stages with the optimal number
In order to further illustrate the results and the VNS of workstations for the SALBP, (Morrison, Sewell, and
improvements, Table 6 compares the random, repair and Jacobson 2014).
VNS for large instances (n = 100) grouped according to Table 7 reports the results of the comparison. The
each of its characteristics. The table reports the average table groups instances according to their characteristics,
number of production stages obtained by each method, and reports the trivial lower bound, column Bound, the
and the average trivial bound among instances from each optimal number of workstations for the SALBP, column
group. As the difference between a solution and the triv- SALBP, and the optimal number of production stages for
ial lower bound corresponds to the number of parallel the PSALBP, column PSALBP. While the SALBP tries
workstations, the higher the reported number, the better. to minimise the number of workstations, we report the
Table 6 stresses the relevance of task time distribution PSALBP as a maximisation problem whose objective is to
to assess instance difficulty and the potential gains from maximise the number of production stages. This objec-
station paralleling. The gaps between the trivial bound, tive is equivalent to minimising the number of parallel
column Bound, and the solutions found by the differ- workstations, the objective proposed in Section 3.2, and
ent methods are null or very small for most instance offers a fair comparison between the SALBP and the
groups except for middle task time distributions and high PSALBP (if both values coincide, the optimal SALBP is
order strength instances with bimodal task time distribu- also optimal for the PSALBP and the difference between
tions. For these instances, the VNS improves upon both the bound and each optimal value measures the ineffi-
the random and the repair methods and provides solu- ciencies of both solutions).
tions in which the number of required parallel stations, The results from Table 7 follow a similar trend as those
i.e. the difference between the bound and the number of found in Table 6. Namely, the SALBP and the PSALBP
production stages, is significantly reduced. solutions coincide with the trivial bound for most
Note that if task times were longer than the cycle instance characteristics. The results reported in Morri-
time, the PSALBP solution would vary accordingly, and son, Sewell, and Jacobson (2014), Pereira (2015) were
the expected results would follow the results from the similar and allow us to identify task time distribution
16 E. ÁLVAREZ-MIRANDA ET AL.

Table 7. Results for the optimal methods. For each grouping of ANOVA permutation test where the same factors and
instances, the rows, defined by its order strength, graph struc- interactions were found to be statistically significant.
ture and distribution of task times, the average lower bound on Given that the previous results imply that the distribu-
the number of workstations, column Bound, the average number
of workstations required by the optimal solutions for the SALBP, tion of task times is the most important feature to assess
column SALBP, and the average number of production stages the gains derived from station paralleling, we derive a
required by the optimal solutions for the PSALBP, column PSALBP secondary instance set to further study this effect.
are reported. To understand the logic behind the generation of these
OS Graph Task times Bound SALBP PSALBP new instances, of these new instances, an alternative
0.2 block bottom 3.00 3.00 3.00 description of the bottom and middle task time distri-
middle 10.80 12.16 9.00 butions considers them different ideal ratios of tasks per
bimodal 4.84 4.84 4.84
chain bottom 3.08 3.08 3.08 workstation. Under such a description, bottom instances
middle 10.48 11.72 8.84 have an average of 10 tasks per workstation while middle
bimodal 5.08 5.08 5.08 instances have an average of 2 tasks per workstation. By
mixed bottom 3.08 3.08 3.08
middle 10.52 12.00 8.68 increasing the cycle time, middle instances would become
bimodal 5.20 5.20 5.20 bottom instances (under the assumption that the stan-
0.6 block bottom 2.96 2.96 2.96 dard deviation used to draw tasks times was identical).
middle 10.44 12.04 8.28
bimodal 5.24 5.28 5.20 Consequently, we generate new instances by changing the
chain bottom 3.00 3.00 3.00 cycle time to desired ratios of tasks per workstation.
middle 10.48 11.96 8.00
bimodal 5.08 5.28 4.84 The original set contains 175 instances with n = 20
mixed bottom 3.08 3.08 3.08 and middle task time distribution. These instances are
middle 10.32 11.84 7.72
bimodal 5.24 5.28 5.12
solved with 9 different cycle times (from 1000 to 5000 in
0.9 mixed bottom 3.04 3.04 3.04 steps of 500 time units). Table 8 compares the optimal
middle 10.40 12.56 6.84 solutions found for the SALBP, using the branch, bound
bimodal 5.28 5.40 5.12
and remember method described in Morrison, Sewell,
and Jacobson (2014), and for the PSALBP, using the for-
mulation discussed in Section 3.2. Given the large num-
as the main source of inefficiencies (idle times) in the ber of parameter combinations (order strength, graph
SALBP. Specifically, as the difference between the triv- structure and cycle times), we only report those groups
ial bound and the number of production stages in the in which the optimal solution for the SALBP and the
PSALBP correspond to the number of parallel stations PSALBP does not coincide (i.e. instances in which the
required to achieve maximum efficiency, we conclude trivial bound is not equal to the SALBP solution).
that small degrees of parallelism suffice to reach maxi- The results in Table 8 further stress the relevance of
mum efficiency. task time distributions. Slight changes in the number of
The order strength of the instances also plays an tasks per workstation significantly decrease the ineffi-
important role in this trade-off. For instances with mid- ciencies of the SALBP. Moreover, the number of parallel
dle task time distribution and OS = 0.2, an average 1.9 workstations required to obtain the maximum theoret-
(out of 10.6) parallel workstations are required; while ical efficiency is also minimal. This results show that a
for OS = 0.6, the number increases to 2.4 and to 3.56 limited use of station paralleling can lead to optimal,
for OS = 0.9. If we consider the number of workstations in terms of idle time, assembly lines. Table 8 also high-
saved by station paralleling, these values are respectively lights the role of order strength and graph structure in
1.36, 1.5 and 2.16 for OS = 0.2, OS = 0.6 and OS = 0.9. the results, as instances with low order strengths and an
The previous analysis was confirmed by a three-way average number of tasks per workstation equal to 5 fea-
ANOVA analysis in which the response variable was ture no inefficiencies, while instances with a higher order
set to the difference between the number of production strength still lead to inefficient solutions with larger ratios
stages of the SALBP and the PSALBP. The p-value for of tasks per workstation.
the OS and task time distribution factors were statistically While this study is limited to small instances in which
significant (p-value < 0.0005), while the structure factor we can solve the SALBP and the PSALBP to optimal-
was not significant. Moreover, the interaction between ity, we believe that results can be extrapolated for larger
OS and task time distribution was statistically signifi- instances since the observed pattern for medium, large
cant (p-value < 0.0005), while the remaining interactions and very large instances showed that only instances with
were not significant. a low average number of tasks per workstation have
The residuals of the ANOVA did not satisfy normal- an optimal solution that differs from the trivial lower
ity assumptions. Thus, we performed a non-parametric bound.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 17

Table 8. Comparison of optimal solutions for different cycle Finally, and in order to identify the cases in which
times (i.e.ideal average number of tasks per workstation) and parallel stations may improve line efficiency, the optimal
graph characteristics. For each combination of cycle time, column solutions for the SALBP and the PSALBP are compared.
cycle, order strength, column OS, and graph structure, column
graph, the trivial lower bound, column bound, the optimal num- The results show that assembly lines with a small average
ber of workstations for the SALBP, column SALBP, and the optimal number of tasks per workstation may benefit from station
number of production stages for the PSALBP, column PSALBP, are paralleling regardless of other assembly process charac-
reported. We report only the groups in which there are differences teristics. Moreover, instances with large order strengths
between the optimal solutions and the trivial lower bound. (i.e. many precedence relations) also benefit from station
cycle OS Graph bound SALBP PSALBP paralleling. For other production processes, the results
1000 0.2 block 10.80 12.16 9.00 hint at possible efficiency improvements without resort-
chain 10.48 11.72 8.84 ing to parallel stations.
mixed 10.52 12.00 8.68
0.6 block 10.44 12.04 8.28 These results imply that assembly processes with
chain 10.48 11.96 8.00 fewer assembly alternatives and tighter assembly con-
mixed 10.32 11.84 7.72
0.9 mixed 10.40 12.56 6.84 straints (either precedence or cycle time constraints)
1500 0.2 block 7.20 7.56 6.80 benefit the most from improving the process. Such
chain 7.08 7.24 6.88
0.6 block 7.12 7.32 6.72
improvements may come from reengineering the prod-
chain 7.08 7.36 6.64 uct itself (i.e.modifying assembly conditions by alter-
mixed 7.00 7.28 6.56 ing task times and other technological constraints) or
0.9 mixed 7.12 7.64 6.16
2000 0.2 block 5.68 5.76 5.60 improving the production process (increasing process
chain 5.52 5.56 5.48 flexibility through parallel workstations or U-shaped
0.6 block 5.44 5.52 5.32
chain 5.52 5.64 5.24 assembly lines, among other alternatives), and the final
mixed 5.36 5.56 4.96 decision highly depends on the intrinsic conditions of the
0.9 mixed 5.44 5.76 4.92
2500 0.6 block 4.44 4.52 4.36
product and the production process.
chain 4.52 4.64 4.32 This result presents further evidence of the opportuni-
mixed 4.36 4.52 4.12 ties provided by station paralleling. For instance, station
0.9 mixed 4.44 4.64 4.12
3000 0.6 mixed 3.92 3.96 3.88 paralleling is essential when some task times are longer
3500 0.6 block 3.12 3.16 3.08 than the cycle time, but it also helps when the task times
0.9 mixed 3.16 3.20 3.08
5000 0.6 chain 2.52 2.56 2.48 of the tasks are close to the cycle time. Such assem-
mixed 2.36 2.44 2.28 bly processes correspond to difficult-to-divide, elaborate,
0.9 mixed 2.44 2.52 2.36
assembly operations. Here, the number of tasks per work-
station is small and the possible losses due to station
paralleling, specifically the loss of specialisation when
workers perform multiple tasks, is not significant when
6. Conclusions
compared to the efficiency improvements provided by the
In this paper, we consider the effect of station parallel- new assembly considerations. Obviously, there is still a
ing in assembly line efficiency. To study this effect, we trade-off involved in additional equipment costs when
put forward a variant of the classical simple assembly compared to efficiency increases. The potential benefits
line balancing problem, the parallel station assembly line should be especially considered in the mass-production
balancing problem or PSALBP, in which parallel stations of commodity goods where larger production runs help
are allowed, and the aim is to find the minimum num- firms to recover from larger investment costs.
ber of parallel stations required to reach the maximum To conclude, we would like to stress the limitations of
theoretical efficiency of the line. this work as they suggest possible directions for future
In order to solve the problem, a polynomially solvable work:
case is identified and a VNS metaheuristic that integrates
the previous result with different ideas stemming from • The PSALBP minimises the total number of paral-
the SALBP is developed. The metaheuristic is then tested lel workstations; however, in a practical setting, an
in a benchmark set derived from the SALBP and the assembly line with one production stage with three
different characteristics that influence the performance workstations and an assembly line with one worksta-
of the method are ascertained. The VNS turns out to tion is not equivalent to a line configuration with two
be superior to alternative approaches, especially for the production stages with two workstations each. Given
most difficult instances from the instance set. For these the small number of parallel workstations required in
instances, the VNS provides significant improvements by our results, we consider that the above limitation is
reducing the number of required parallel workstations. not essential to the main conclusions of this work but
18 E. ÁLVAREZ-MIRANDA ET AL.

may be relevant in some circumstances. This point is focused on the development and application of mathematical
especially true when tasks are naturally longer than the optimisation techniques for problems from different applica-
cycle times or when the parallel workstations are eval- tion contexts.
uated from an investment standpoint. In both these Sebastián Chace is an Operations Research
cases, a more fine-grained analysis should help to split M.Sc. candidate and Industrial Engineer-
parallel workstations along the line. ing B.Sc. at Universidad Adolfo Ibáñez. He
currently works at Ernst&Young as Ana-
• The analysis of different cycle times reveals that lytics consultant.
line balancing problems only benefit from additional
degrees of freedom to deal with its constraints when
the task times are very large or the precedence rela-
tions are high. Given that most SALBP instances are Jordi Pereira is Associate Professor at
easily solved, more attention needs to be drawn to the Faculty of Engineering and Science,
sources of comparison that apply to hard line balanc- Universidad Adolfo Ibáñez in Chile. He
ing problems. received his PhD in Engineering in 2004
• This work only considers the application of paral- from the Universitat Politècnica de
Catalunya, Spain. His research considers
lel workstations to assembly lines that comply with
the development of solution procedures
the SALBP assumptions. Given the variety of alter- for combinatorial optimisation problems
native general assembly line formulations, it would in different application areas with special interest in assembly
be interesting to ascertain which of these character- line balancing problems.
istics modify the problem in order to make worksta-
tion paralleling an option to improve assembly line References
efficiency.
Akpınar, S., and G. M. Bayhan. 2011. “A Hybrid Genetic
• Finally, and from a managerial point of view, there Algorithm for Mixed Model Assembly Line Balancing Prob-
is an important research gap regarding the expected lem with Parallel Workstations and Zoning Constraints.”
benefits from different production process alterna- Engineering Applications of Artificial Intelligence 24 (3):
tives. U-shaped assembly lines and parallel work- 449–457. doi:10.1016/j.engappai.2010.08.006.
stations are among the most common methods to Askin, R. G., and M. Zhou. 1997. “A Parallel Station Heuris-
tic for the Mixed-model Production Line Balancing Prob-
increase the flexibility of a production process, but
lem.” International Journal of Production Research 35 (11):
their benefits have not been extensively compared 3095–3106.
with the scientific literature in terms of efficiency gains Bard, J. 1989. “Assembly Line Balancing with Parallel Worksta-
or cost savings. tions and Dead Time.” International Journal of Production
Research 27: 1005–1018.
Bartholdi, J. J. 1993. “Balancing Two-sided Assembly Lines: A
Disclosure statement Case Study.” International Journal of Production Research 31
No potential conflict of interest was reported by the authors. (10): 2447–2461.
Bautista, J., and J. Pereira. 2007. “Ant Algorithms for a Time
Funding and Space Constrained Assembly Line Balancing Prob-
lem.” European Journal of Operational Research 177 (3):
This research has been partially funded by the National Com- 2016–2032.
mission for Scientific and Technological Research CONICYT, Baybars, I. 1986. “A Survey of Exact Algorithms for the Simple
Chile, through research grant N. 1191624 ‘Assembly Line Bal- Assembly Line Balancing Problem.” Management Science 32
ancing for Industry 4.0’. E. Álvarez-Miranda also acknowl- (8): 909–932.
edges the support of the National Commission for Scien- Becker, C., and A. Scholl. 2006. “A Survey on Problems and
tific and Technological Research CONICYT, Chile, through Methods in Generalized Assembly Line Balancing.” Euro-
funding from FONDECYT N. 1180670, FONDECYT N. pean Journal of Operational Research 168 (3): 694–715.
1191624 and through the Complex Engineering Systems Insti- Becker, C., and A. Scholl. 2009. “Balancing Assembly Lines with
tute PIA/BASAL AFB180003. Variable Parallel Workplaces: Problem Definition and Effec-
Notes on contributors tive Solution Procedure.” European Journal of Operational
Research 199 (2): 359–374.
Eduardo Álvarez-Miranda is Associate Blum, C., J. Puchinger, G. L. Raidl, and A. Roli. 2011. “Hybrid
Professor of Industrial Engineering, at the Metaheuristics in Combinatorial Optimization: A Survey.”
Faculty of Engineering, Universidad de Applied Soft Computing 11 (6): 4135–4151.
Talca in Chile. He is also a researcher of Bowman, H. E. 1960. “Assembly-Line Balancing by Linear
the Complex Systems Engineering Insti- Programming.” Operations Research 8 (3): 385–389.
tute, Chile. He received his PhD in Opera- Bukchin, J., and J. Rubinovitz. 2003. “A Weighted Approach for
tions Research in 2014 from the University Assembly Line Design with Station Paralleling and Equip-
of Bologna, Italy. His research interests are ment Selection.” IIE Transactions 35: 73–85.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 19

Bukchin, J., and M. Tzur. 2000. “Design of Flexible Assem- Task Durations and Parallel Stations.” International Journal
bly Line to Minimize Equipment Cost.” IIE Transactions 32: of Production Economics51 (3): 177–190.
585–598. McMullen, P. R., and G. V. Frazier. 1998. “Using Simulated
Buxey, G. M. 1974. “Assembly Line Balancing with Multiple Annealing to Solve a Multiobjective Assembly Line Bal-
Stations.” Management Science 20 (6): 1010–1021. ancing Problem with Parallel Workstations.” International
Cevikcan, E., D. Aslan, and F. B. Yeni. 2020. “Disassembly Line Journal of Production Research 36 (10): 2717–2741.
Design with Multi-manned Workstations: A Novel Heuristic McMullen, P. R., and P. Tarasewich. 2006. “Multi-objective
Optimisation Approach.” International Journal of Production Assembly Line Balancing Via a Modified Ant Colony Opti-
Research 58 (3): 649–670. mization Technique.” International Journal of Production
Daganzo, C. F., and D. E. Blumenfeld. 1994. “Assembly Sys- Research 44 (1): 27–42.
tem Design Principles and Tradeoffs.” International Journal Mladenović, N., and P. Hansen. 1997. “Variable Neighbor-
of Production Research 32 (3): 669–681. hood Search.” Computers and Operations Research 24 (11):
Dimitriadis, S. G. 2006. “Assembly Line Balancing and Group 1097–1100.
Working: A Heuristic Procedure for Workers’ Groups Oper- Morrison, D. R., E. C. Sewell, and S. H. Jacobson. 2014.
ating on the Same Product and Workstation.” Computers and “An Application of the Branch, Bound, and Remember
Operations Research 33 (9): 2757–2774. Algorithm to a New Simple Assembly Line Balancing
Dolgui, A., and O. Battaïa. 2013. “A Taxonomy of Line Balanc- Dataset.” European Journal of Operational Research 236 (2):
ing Problems and Their Solution Approaches.” International 403–409.
Journal of Production Economics 142: 259–277. Otto, C., and A. Otto. 2014. “Extending Assembly Line Bal-
Ege, Y., M. Azizoglu, and N. Özdemirel. 2009. “Assembly Line ancing Problem by Incorporating Learning Effects.” Interna-
Balancing with Station Paralleling.” Computers & Industrial tional Journal of Production Research 52 (24): 7193–7208.
Engineering 57: 1218–1225. Otto, A., C. Otto, and A. Scholl. 2013. “Systematic Data Gener-
Fleszar, K., and K. S. Hindi. 2003. “An Enumerative Heuris- ation and Test Design for Solution Algorithms on the Exam-
tic and Reduction Methods for the Assembly Line Balancing ple of SALBPGen for Assembly Line Balancing.” European
Problem.” European Journal of Operational Research 145 (3): Journal of Operational Research 228 (1): 33–45.
606–620. Öztürk, C., S. Tunali, B. Hnich, and A. Örnek. 2013. “Bal-
Garey, M. R., and D. S. Johnson. 1979. Computers and ancing and Scheduling of Flexible Mixed Model Assem-
Intractability: A Guide to the Theory of NP-Completeness. bly Lines with Parallel Stations.” International Journal
New York, NY: W. H. Freeman. of Advanced Manufacturing Technology 67 (9–12): 2577–
Hoffmann, T. R. 1963. “Assembly Line Balancing with a Prece- 2591.
dence Matrix.” Management Science 9 (4): 551–562. Pape, T. 2015. “Heuristics and Lower Bounds for the Simple
Johnson, D. S., and K. A. Niemi. 1983. “On Knapsacks, Assembly Line Balancing Problem Type 1: Overview, Com-
Partitions, and a New Dynamic Programming Technique putational Tests and Improvements.” European Journal of
for Trees.” Mathematics of Operations Research 8 (1): Operational Research 240 (1): 32–42.
1–14. Pereira, J. 2015. “Empirical Evaluation of Lower Bounding
Kellegöz, T. 2016. “Balancing Lexicographic Multi-Objective Methods for the Simple Assembly Line Balancing Prob-
Assembly Lines with Multi-Manned Stations.” Mathematical lem.” International Journal of Production Research 53 (11):
Problems in Engineering 2016: 9315024. doi:10.1155/2016/ 3327–3340.
9315024. Pereira, J. 2018. “Modelling and Solving a Cost-oriented
Kramer, R., M. Dell–Amico, and M. Iori. 2017. “A Batching- Resource-constrained Multi-model Assembly Line Balanc-
move Iterated Local Search Algorithm for the Bin Pack- ing Problem.” International Journal of Production Research
ing Problem with Generalized Precedence Constraints.” 56 (11): 3994–4016.
International Journal of Production Research 55 (21): 6288– Pereira, J., M. Ritt, and Ó. C. Vásquez. 2018. “A Memetic
6304. Algorithm for the Cost-oriented Robotic Assembly Line
Li, Z., I. Kucukkoc, and Z. Zhang. 2020. “Branch, Bound and Balancing Problem.” Computers & Operations Research 99:
Remember Algorithm for Two-sided Assembly Line Balanc- 249–261.
ing Problem.” European Journal of Operational Research 284 Pinto, P. A., D. G. Dannenbring, and B. M. Khumawala. 1981.
(3): 896–905. doi:10.1016/j.ejor.2020.01.032. “Branch and Bound and Heuristic Procedures for Assembly
Lopes, T. C., A. S. Michels, C. G. S. Sikora, and L. Magatão. Line Balancing with Paralleling of Stations.” International
2019. “Balancing and Cyclical Scheduling of Asynchronous Journal of Production Research 19 (5): 565–576.
Mixed-model Assembly Lines with Parallel Stations.” Jour- Polat, O., C. B. Kalayci, O. Mutlu, and S. M. Gupta. 2016. “A
nal of Manufacturing Systems 50: 193–200. Two-phase Variable Neighbourhood Search Algorithm for
López-Ibáñez, M., J. Dubois-Lacoste, L. Pérez Cáceres, T. Stüt- Assembly Line Worker Assignment and Balancing Problem
zle, and M. Birattari. 2016. “The Irace Package: Iterated Type-II: An Industrial Case Study.” International Journal of
Racing for Automatic Algorithm Configuration.” Operations Production Research 54 (3): 722–741.
Research Perspectives 3: 43–58. Sadeghi, P., R. D. Rebelo, and J. S. Ferreira. 2018. “Balancing
Maniezzo, V., T. Stützle, and S. Voß, eds. 2010. Matheuristics Mixed-model Assembly Systems in the Footwear Industry
– Hybridizing Metaheuristics and Mathematical Program- with a Variable Neighbourhood Descent Method.” Comput-
ming. Vol. 10 of Annals of Information Systems. Boston, MA: ers and Industrial Engineering 121: 161–176.
Springer. Sarker, B. R., and J. G. Shanthikumari. 1983. “A Generalized
McMullen, P. R., and G. V. Frazier. 1997. “A Heuristic for Solv- Approach for Serial Or Parallel Line Balancing.” Interna-
ing Mixed-model Line Balancing Problems with Stochastic tional Journal of Production Research 21 (1): 109–133.
20 E. ÁLVAREZ-MIRANDA ET AL.


Scholl, A., and C. Becker. 2006. “State-of-the-art Exact and which W = v∈V  w(v) ≥ M. Otherwise, the problem is triv-
Heuristic Solution Procedures for Simple Assembly Line Bal- ial.
ancing.” European Journal of Operational Research 168 (3): We construct an instance I of the PSALBP as follows: Prece-
666–693. dence relations are represented using a directed acyclic graph
Scholl, A., and S. Voss. 1997. “Simple Assembly Line Balanc- G (T, A), where the nodes correspond to the set of tasks T =
ing – Heuristic Approaches.” Journal of Heuristics 2 (3): V ∪ {α, ω} and the arcs correspond to the relations. Set A con-
217–244. tains set A from G(V, A ) as well as an arc from α to each of the
Simaria, A. S., and P. M. Vilarinho. 2001. “The Simple Assem- other vertices and an arc from each vertex other than ω to ω.
bly Line Balancing Problem with Parallel Workstations – Each task i ∈ V has a processing time pi equal to w(v). tα and
A Simulated Annealing Approach.” International Journal of tω are obtained as follows:
Industrial Engineering: Theory Applications and Practice 8 Case i. If W−M = m, then task times are set according
(3): 230–240. to (A1)
Sternatz, J. 2014. “Enhanced Multi-Hoffmann Heuristic for
Efficiently Solving Real-world Assembly Line Balancing tα = tω = 0. (A1)
Problems in Automotive Industry.” European Journal of Case ii. If W−M > m, then define task times according to (A2)
Operational Research 235 (3): 740–754. and (A3).
Süer, G. A. 1998. “Designing Parallel Assembly Lines.” Com-
puters & Industrial Engineering 35 (3): 467–470. tα =W − M − m (A2)
Tiacci, L. 2015a. “Coupling a Genetic Algorithm Approach tω =0. (A3)
and a Discrete Event Simulator to Design Mixed-model
Un-paced Assembly Lines with Parallel Workstations and Case iii. If P−W < m, then task times are set according to (A4)
Stochastic Task Times.” International Journal of Production and (A5).
Economics 159: 319–333.
tα = 0 (A4)
Tiacci, L. 2015b. “Simultaneous Balancing and Buffer Alloca-
tion Decisions for the Design of Mixed-model Assembly tω = m − (W − M). (A5)
Lines with Parallel Workstations and Stochastic Task Times.”
International Journal of Production Economics 162: 201–215. The cycle time c is set to tα + M.
Tuncel, G., and S. Topaloglu. 2013. “Assembly Line Balancing We claim that instance I has an optimal solution with 2 pro-
with Positional Constraints, Task Assignment Restrictions duction stages if and only if there exists a solution to the original
and Station Paralleling: A Case in An Electronics Company.” POK instance.
Computers & Industrial Engineering 64 (2): 602–609. A non-technical overview of the proof follows. Consider
Vilarinho, P. M., and A. S. Simaria. 2002. “A Two-stage Heuris- Case i. The solution to the POK instance corresponds to the
tic Method for Balancing Mixed-model Assembly Lines with tasks assigned to the first production stage of the PSALBP. We
Parallel Workstations.” International Journal of Production have to ensure that (i) these tasks are closed under precedence,
Research 40: 1405. (ii) the sum of their weights (values) is no larger than M, and
Vilarinho, P. M., and A. S. Simaria. 2006. “ANTBAL: An (iii) no smaller than m.
Ant Colony Optimization Algorithm for Balancing Mixed- Condition (i) is ensured through the precedence relations
model Assembly Lines with Parallel Workstations.” Interna- of the PSALBP, while condition (ii) is ensured by the definition
tional Journal of Production Research 44: 291–303. of the cycle time. Condition (iii) follows from the cycle time
constraint imposed on the tasks assigned to stage 2 (the subset
of tasks not assigned to stage 1). Figure 1 visualises that the sum
of the weights of the tasks assigned to stage 2 may be no larger
Appendix. Complexity of partitioning a than c, which forces the sum of the weights of stage 1 to be no
production stage smaller than m. For the other two cases, we increase the task
time of either dummy task α or ω and manipulate c in order to
This appendix provides proof to the NP-hard status of the ensure that the same relationship holds.
PSALBP even when the number of workstations equals two Now we proceed with the proof. Converting a feasible POK
(hence, no more than two production stages are required). The into a feasible PSALBP solution with two production stages
proof reduces the Partially Ordered Knapsack, POK, which is and one workstation per production stage is performed as
known to be NP-complete in the strong sense (Johnson and follows. Given a solution V  ⊆ V to the POK instance, we
Niemi 1983) for the problem under study and follows a similar assign tasks V  ∪ {α} to the first production stage and V \ V  ∪
proof outlined in Pereira, Ritt, and Vásquez (2018). {ω} to the second stage. Clearly, the solution complies with
A formal definition of the POK (Johnson and Niemi 1983) the PSALBP precedence constraints and tα + i∈V  pi ≤ tα +
follows: M = c, hence one workstation is required for the first stage.
Instance: Given a directed acyclic graph G(V, A ), a weight In order to show that a single
 workstation is required for
w(v) ∈ Z+ +
0 and a value p(v) ∈ Z0 for each vertex v ∈ V, a the second stage, note that i∈V  ti + i∈V\V  ti = W and,
knapsack capacity M ∈ Z+ , and a bound m ∈ Z+ . 
according to the feasibility of the POK solution i∈V  ti ≥ m.
Question: Isthere a subset V  ⊆ V,closed under predeces- Consequently, (A6) holds.
sors, such that v∈V  w(v) ≤ M and v∈V  p(v) ≥ m?

The POK is NP-complete in the strong sense even if ti ≤ W − m (A6)
w(v) = p(v), ∀v ∈ V. In such a case, the problem aims to find i∈V\V 
asubset V  ⊆ V closed under predecessors such that m ≤
v∈V  w(v) ≤ M. Also, we limit our attention to the case in Now we consider each case separately.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 21

Case i. The relation that defines the case is W−M = m, by


substitution from (A6), we get (A7).

ti ≤ W − m = M = c (A7)
i∈V\V 

Case ii. The condition that defines the case can be rewritten
as (A8).
c = tα + M = (W − M − m) + M = W − m (A8) Figure A1. Outline of relationships among m, M, W and c. The
central line locates the values within the real line. The range of
Consequently, replacing (A8) in (A6) leads to (A7). feasible workloads for the first and second production stage are
 Case iii. The workload of the second stage is (A9) also represented.
i∈V\V  ti + tω .
   
ti + tω = ti + m − (W − M) (A9) Case i. i∈V\V  ti ≤ c = M holds. Substitute in i∈V  ti +

i∈V\V  i∈V\V  i∈V\V  ti = W to obtain (A10) which proves the result.
Replacing (A6) into (A9), we reach the conclusion that the 
ti ≥ P − W = m (A10)
total workload of the second stage is no larger than (W −
i∈V 
m) + (m − W + M) = M, showing that only one workstation 
is required. Case ii. i∈V\V  ti ≤ c = tα + M holds. By definition of tα ,

To show that a feasible PSALBP solution with two produc- see
 (A2), we get (A11). Now replace  in W = i∈V  ti +
tion stages and one workstation per production stage is feasible i∈V\V  ti to get the desired result, i.e. i∈V  ti ≥ m.
for the POK, we proceed as follows. In such a solution to the 
PSALBP, the first stage contains task α and a subset V  of tasks ti ≤ (W − M − m) + M (A11)
from V closed
 under precedence relations with total processing i∈V\V 
time tα + i∈V  ti ≤ c which implies  i∈V  ti ≤ M. Conse- 
quently, it only remains to show that i∈V  ti ≥ m holds. Each Case iii. i∈V\V  ti + tω ≤ c = M holds. Replacing tω = m −
  
case is considered separately. W + M and P = i∈V  ti + i∈V\V  ti , we get i∈V ti ≥ m.
This concludes the proof.

You might also like