Mixed-integer linear programming models for

the type-II resource-constrained assembly

line balancing problem
Adalberto Sato Michels and Alysson M. Costa
School of Mathematics and Statistics, University of Melbourne, Melbourne, Australia

Abstract
Purpose – Resource-constrained assembly lines are widely found in industries that manufacture complex products. In such lines, tasks may require
specific resources to be processed. Therefore, decisions on which tasks and resources will be assigned to each station must be made. When the
number of available stations is fixed, the problem’s main goal becomes the minimisation of cycle time (type-II version). This paper aims to explore
this variant of the problem that lacks investigation in the literature.
Design/methodology/approach – In this paper, the authors propose mixed-integer linear programming (MILP) models to minimise cycle time in
resource-constrained assembly lines, given a limited number of stations and resources. Dedicated and alternative resource types for tasks are
considered in different scenarios.
Findings – Besides, past modelling decisions and assumptions are questioned. The authors discuss how they were leading to suboptimal solutions
and offer a rectification.
Practical implications – The proposed models and data set fulfil more practical concerns by taking into account characteristics found in real-world
assembly lines.
Originality/value – The proposed MILP models are applied to an existing data set, results are compared against a constraint programming model,
and new optimal solutions are obtained. Moreover, a data set extension is proposed due to the simplicity of the current one and instances up to 70
tasks are optimally solved.
Keywords Assembly line balancing, Resource constraints, Type-II problem, Mixed-integer linear programming
Paper type Research paper

1. Introduction by Scholl and Becker (2006), which also list various exact
algorithmic solution approaches and practical applications.
The design and operation of production lines have originated Another simplification hypothesis of SALBPs is assuming
many complex combinatorial problems in academic research. that any given task can be performed in any station. The
The most popular one is the assembly line balancing problem underlying assumption is that all stations are homogeneously
(ALBP), which is broadly discussed in the literature (Boysen manned or equipped with the necessary resources, i.e. with the
et al., 2021) under several classifications. The most basic same workers, tools and/or machines (Baybars, 1986).
version of the problem, the simple assembly line balancing Nonetheless, real-world assembly lines usually rely on limited
problem (SALBP), was formally defined by Baybars (1986)
heterogeneous resources, e.g. usage of special equipment and/
with several simplification hypotheses. The SALBP comprises
or professional workers. To complete the task set, specific
the problem of deciding the assignment of a set of tasks to
resources must also be assigned to different stations (Agpak and
stations. These tasks cannot be arbitrarily assigned to any
Gökçen, 2005). Several examples of this application can be
station; they must respect precedence constraints commonly
found in factories: cutting tools that make metal sheets in a
represented by a precedence graph. Moreover, stations are
specific width and shape, welding tools that can access certain
organised in a straight, serial line. Such layout enforces the
locations in a car body, drilling tools for particular materials
line’s pace to take the time of the most loaded station, which is
and workers/robots that are skilled to operate such machinery
known as the cycle time. Finally, the assembly line produces a
(Boysen et al., 2007).
unique model of a single product. The most common goals in a
By extending the SALBP to contemplate this feature, Agpak
SALBP are the minimisation of the number of stations
and Gökçen (2005) proposed the resource-constrained
(SALBP-1) and the minimisation of the cycle time (SALBP-2).
Important contributions to the SALBP literature were surveyed
This research was partially funded by the Australian Government through
the Australian Research Council Industrial Transformation Training
The current issue and full text archive of this journal is available on Emerald Centre in Optimisation Technologies, Integrated Methodologies and
Insight at: https://www.emerald.com/insight/0144-5154.htm Applications (OPTIMA), Project ID IC200100009.

Received 27 October 2021

Assembly Automation
Revised 22 April 2022
42/5 (2022) 585–594 10 June 2022
© Emerald Publishing Limited [ISSN 0144-5154] 14 June 2022
[DOI 10.1108/AA-10-2021-0140] Accepted 13 July 2022

585
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

assembly line balancing problem (RCALBP). Becker and RCALBP. Correspondingly, an alternative RCALBP would
Scholl (2006) state that the RCALBP is a generalisation of the require at least one of the resource types that can perform the
previous SALBP and surveys it alongside various practical task assigned to that station, meaning that some tasks can be
extensions. Recently, Alakas  et al. (2020) proposed a variant of performed by alternative resources. The proposed MIP
the RCALBP with the goal of minimising both the cycle time approach’s goal was to establish a balance of the assembly line
and the resource usage for a given number of stations. Sample that minimises the number of stations and resources. For this
scenarios were solved with a constraint programming (CP) purpose, two MIP models (dedicated and alternative) are
approach, and optimal solutions were found for all tested developed and applied to numerical examples, which are
instances. Such an approach had been successfully applied in a tackled by using a black-box commercial solver.
previous CP model that dealt with the ALBP with assignment The RCALBP was further generalised when the multiple and
restrictions (Pinarbasi et al., 2019). However, the CP method concurrent versions of the problem were proposed by Corominas
modelled by Alakas  et al. (2020) has one constrictive et al. (2011). Previously, specific assumptions limited the
assumption: the objective of minimising cycle time is achieved problem: dedicated cases assume that tasks need an exclusive
by compulsorily assigning one resource at each station. In resource type (A), whereas the alternative version states that
practice, by allowing resources to be distributed between tasks may require one out of two or more resources (A Ú B).
stations more freely, it may be possible to achieve the same They extended the problem by presenting cases in which each
cycle time with fewer resources or lower cycle times with the task needs dedicated, alternative, multiple and/or concurrent
same number of resources. resources. Some examples are provided (Corominas et al.,
Motivated by such potential improvements, the type-II 2011): (3 A), (A ^ 4B ^ 3 C), (3 A Ú 2B Ú C), (A ^ B) Ú (2 C ^
RCALBP-2 is further investigated in this paper. Past modelling D) or (A Ú B) ^ (2 C Ú D). Out of that, three MIP variations are
decisions and assumptions are questioned: we extend the presented to mathematically model the problem: two in the
previous concepts on the limit and availability of resources and conjunctive normal form (CNF-1 and CNF-2) and one in the
propose new mixed-integer linear programming (MILP) disjunctive normal form (DNF), as defined by Friedman
models to flexibly cope with these more realistic scenarios, such (1985). The equivalent linearised constraints for the
as having a welding and drilling tool in the same station. We generalised RCALBP containing logical expressions in CNF (^
also discuss how these restrictive concepts were leading to of Ús form) or DNF (Ú of ^s form) were developed. These three
suboptimal solutions and offer a rectification. modelling variations were applied to a 560-instance data set
The remainder of this paper is organised as follows. In with up to 80 tasks in a black-box optimiser. Their effectiveness
Section 2, a literature review is conducted for the RCALBP, was compared, and CNF-2 obtained the best results, with a
emphasising particular objective functions, extensions and greater number of instances solved to optimality and fewer
methodologies investigated by each study. Section 3 precisely instances with no feasible solution. Finally, an upper bound
defines and explains the characteristics of an RCALBP-2 limit on the number of available resources was firstly
considered in this paper. Section 4 presents new MILP models, considered in a mathematical model. Due to the complexity
which contemplate different extensions regarding multiple- increase, the CNF-2 model with active resource limits was not
resource types and availability, as well as budget-oriented as efficient as the model without these constraints.
constraints, taking into account characteristics found in real- Quite a few papers have developed ad hoc solution methods
world assembly lines. Computational experiments use the data to solve specific versions of the RCALBP in the literature. Kao
set proposed by Alakas  et al. (2020) as a benchmark. et al. (2010) proposed a shortest path algorithm to tackle the
Furthermore, a data set expansion is proposed to fulfil more dedicated RCALBP. The method was tested on a sample
practical concerns. All instances are solved, and results are problem composed of 51 tasks and obtained the same solution
discussed throughout Section 5. Finally, Section 6 concludes as the one found in Agpak and Gökçen (2005). Following that
this paper by highlighting its main contributions and providing study, Kao et al. (2011) developed a heuristic based on ranked
further research directions. positional weight rules. To show the effectiveness of their
proposed method, numerical results were obtained from
different sample problems for up to 21 tasks. Kamarudin and
2. Literature review
Ab Rashid (2017) applied a genetic algorithm (GA) to the
The RCALBP was introduced by Agpak and Gökçen (2005) dedicated RCALBP with skilled workers. Their GA used two
along with a mathematical MIP model. They generalised the rank-based crossover mechanisms and was tested on
usage of special equipment and/or professional workers as benchmark problems up to 148 tasks. Quyen et al. (2017a)
resources. Each resource is capable of performing different developed a hybrid GA to solve an RCALBP found in the
processes in a factory. Besides, two resource classifications were sewing lines of a footwear company. The experimental results
considered in their study: dedicated and alternative. Although showed that their algorithm outperformed the factory’s manual
the former considers that each task can only be performed by procedure used at that time and Kao et al. (2011)’s heuristic.
one resource type, the latter makes the problem more flexible Immediately after that, Quyen et al. (2017a, 2017b) proposed
by defining tasks that can be assigned to different resources. A dynamic programming (DP) and applied it to the same
task can only be assigned to a station with the required resource footwear company’s sewing lines. Their DP was capable of
(dedicated RCALBP) or one among the possible required obtaining equivalent or improved results for 16 out of 18 tested
resources (alternative RCALBP). For instance, if a task is instances when compared to Quyen et al. (2017a)’s GA.
assigned to a station, the resource type which can perform such Industrial extensions of the RCALBP have also been
task must also be assigned to that station in a dedicated reasonably explored in the literature. For instance, when

586
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

multiple workers are allowed to simultaneously perform tasks cost over all stations and a B&B algorithm was proposed. Both
in the same station, the SALBP is extended to the so-called approaches were tested in a data set with instances of up to 30
multi-manned assembly line balancing problem (MALBP) tasks. Michels et al. (2018) examined robotic assembly lines
(Michels et al., 2019). This concept can be integrated into the with the possibility of parallel stations. Dedicated and
problem under study to create the resource-constrained multi- alternative resources (robots and equipment) were considered
manned assembly line balancing problem (RCMALBP). This in a cost-oriented MILP model. Their objective function
extension raises an interesting practical question: should minimises the purchase cost of the line, which is composed of
resources be assigned to workers or to stations? Yilmaz and platform and transporter robots along with their tools, stations
Yilmaz (2017) considered the former: they proposed a and track-motion systems. Practical extensions to consider
mathematical programming model to minimise the number of welding tasks were modelled, and tests were also conducted in a
workers, stations and resources used in the line. Workers can cost structure ratio-based data set (Askin and Zhou, 1997).
only perform tasks if they hold the necessary resource to do so. Moreover, an automotive robotic assembly line case study was
This model was tested with a numerical example to indicate its solved. Finally, Pereira (2018) studied a cost-oriented, multi-
applicability. On the other hand, Chen et al. (2018) studied a model RCALBP found in the textile industry. Straightforward
practical case found in an automotive factory, in which and alternative MILP formulations were developed. The
resources (machines) are assigned to stations and shared former minimises the weighted sum of costs associated with
among all workers also employed in the same station. A stations and resources whilst the latter minimises the total cost
mathematical programming model was proposed to minimise of the stations. The straightforward MILP uses variable sets to
the number of workers, stations and resources. This model was decide on task-station, item-station and resource-station
then used as the basis for the subsequent development of a assignments. The alternative model makes use of a
suitable GA-based algorithm. Their proposed approach was feasible assignment concept that does not manage item and
compared to the traditional planning method. Results showed resource assignments explicitly. These decisions are transferred
that better configurations could be obtained in terms of to an auxiliary pricing problem that iteratively identifies feasible
solution quality and production efficiency. assignments given a constraint set and calculates
A similar extension is the two-sided assembly line balancing the assignment costs accordingly. A hybrid method that
problem (TALBP), in which both left-hand and right-hand combines a modified Hoffmann heuristic (Hoffmann, 1963)
sides of the line are used for manufacturing products with an estimation of distribution algorithm (Pereira, 2015)
(Bartholdi, 1993). Yang et al. (2020) aimed to study the multi- was also proposed. Computational experiments were
constrained version of the TALBP, which involved positional, conducted using a benchmark-based data set derived from the
zoning and synchronism constraints. They extended the basic literature (Otto et al., 2013), comprising instances with up to
TALBP-1 model and proposed a new mathematical 100 tasks. The case study from the manufacturer that
programming model in which the cycle time is given. The originated the problem was also solved.
weighted sum of the number of mated stations and stations is Thus far, we can observe that researchers had only discussed
minimised. They also designed an improved imperialist the RCALBP by essentially considering the minimisation of
competitive algorithm to solve the problem. Additionally, resources and stations (i.e. RCALBP-1). Nevertheless, Alakas 
Yadav et al. (2020) developed a new mixed-integer et al. (2020) most recently studied the RCALBP-2. Although
mathematical programming model to tackle the parallel the cycle time was considered in the objective, it had a
TALBP, a version in which both tools and tasks could be secondary role. Instead, the minimisation of the total number
shared. The main aim of their model is maximising the of used resources was aimed as the primary goal. However,
workload on each station. Arguably, this goal works in the researchers generally focus on minimising the line’s cycle time
direction of minimising the line length and reducing idle times in type-II ALBPs, a fact that was acknowledged by the authors.
at each station. Their results compared the effects of these tools Furthermore, they claimed that fixing a limit of at most one
and tasks sharing approaches on efficiency, which was used as resource per available station satisfied the condition of resource
the performance measure. minimisation. A new CP model to deal with dedicated and
Other industrial variants considering RCALBPs are related alternative scenarios was proposed, and a new data set was
to line layout, designing costs and multiple product models. generated to test their proposed model. The implicit
These factors are important attributes for the rapid assumption adopted by Alakas  et al. (2020) here was the
advancement of Industry 4.0 Rahman et al. (2020). Ege et al. mandatory use of exactly one resource for each available
(2009) investigated the possibility of parallel stations in station. Optimal solutions were obtained for all instances under
assembly lines containing tasks that require specified dedicated this simplification hypothesis of allowing only one resource per
resources (tooling/equipment). The cost structure ratio station. However, this assumption is further evaluated and
suggested by Askin and Zhou (1997) was used to generate data discussed throughout Sections 3, 4 and 5. We extend the
set instances for up to 70 tasks. They attempted to minimise problem by relaxing the enforcement on having exactly one
costs related to station opening and resources. For that, two resource per station, addressing why this might unnecessarily
branch-and-bound (B&B) algorithms were proposed: one for exclude feasible solutions.
optimal solutions and one for near-optimal solutions. Ogan and
Azizoglu (2015) and Michels et al. (2018) further extended Ege
et al. (2009)’s work in different directions. Ogan and Azizoglu
3. Problem statement
(2015) analysed U-shaped assembly lines with alternative The characteristics of the RCALBP-2 to be considered are
resources. A MILP model that minimises the total equipment introduced using the example depicted in Figure 1. The

587
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

Figure 1 Illustrative example with a feasible solution for the RCALBP-2

Set of tasks
Assign tasks to
B B B B
staon
2 6 8 10 Assign resources
to staon
A A A

1 3 11
Set of resources
B A A

4 7 9 A B

5
Set of staons

2,4,6,
1,3,5 7,9,11
8,10
A B A

Notes: Tasks are represented by Nodes 1 to 11. two resources, named A and B, are available.
the resource requirement of each task is given on the upper right of the node: odd-numbered and
even-numbered tasks require resources A and B, respectively. There are three available stations.
arrows and dotted lines show the assignment of subgroups of tasks and resources to stations,
respectively

notation is detailed in below: indexes, sets, parameters and set of three available stations (S) to which tasks and resources
variables. must be assigned: xik = 1 and yrk = 1 if task i and resource type r
are assigned to station k, respectively. Tasks are indivisible and
Notation
i, j = task indexes; cannot be performed unless the required resource is present in
r = resource index; the station. Finally, the cycle time (c) is determined by the sum
k = station index; of task processing times in the most loaded station.
T = set of tasks; Figure 1 also presents a feasible solution for the illustrative
R = set of resource types; instance. Note that subsets of tasks with identical resource
S = set of stations; requirements are grouped together by dashed lines. Arrows
P = set of precedence relations (i, j) [ Tji p j; assign these subgroups into stations: Tasks 1, 3 and 5 to Station
Tr = set of tasks that can be performed with resource r; 1, Tasks 2, 4, 6, 8 and 10 to Station 2 and Tasks 7, 9 and 11 to
Ri = set of resources required by task i; Station 3. Nonetheless, this is not enough to ensure feasibility;
ti = processing time of task i; we must assign the appropriate resource types to stations as
a = available number of resources; well. Dotted lines assign either resource type A or B to stations.
rr = cost of resource r; Ultimately, we obtained a feasible solution in which three
b = available budget; resource units were used (two of type A and one of type B).
g = number of resources allowed per station; The illustrated version of the problem assumes that resources
xik = binary variable equal to 1 if task i is assigned to station k; are dedicated, i.e. each task only has one possibility of resource
yrk = binary variable equal to 1 if resource r is assigned to type to enable its execution. In Section 4, we model this basic
station k; and case and extend our formulation to contemplate the alternative
c = continuous variable for cycle time. version of the problem, in which multiple-resource possibilities
are considered. Furthermore, we eliminate the simplification
A set of tasks (T) must be performed to assemble a product. hypothesis that imposes exactly one resource to be assigned to
These tasks must respect precedence relations (P) in their each available station, discussing why previous approaches may
execution order. A precedence graph containing task indexes be eliminating conceivable optimal solutions.
from 1 to 11 is presented in Figure 1. This precedence graph
was also used by Agpak and Gökçen (2005) in their illustrative
example. A set of resources (R) is also considered. In this
4. Mixed-integer linear programming models
illustrative example, two resource types (A and B) are We develop the following problem formulation for the
considered and represented in the precedence graph for each dedicated RCALBP-2 and its extensions using the notation
task node. Resource A is necessary to perform odd indexed introduced in Notation, in which definitions of indexes, sets,
tasks and B for even indexed tasks. For this instance, there is a parameters and variables are presented.

588
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

Min z1 ¼ c; (1) readily extended to consider a limit on each type of resource.

Additionally, Constraints (11) are valid inequalities to avoid
XX gaps in station numbering, and thus, reducing the symmetry of
Min z2 ¼ yrk ; (2)
r2R k2S
the model:
XX
subject to: yrk  a; (10)
X r2R k2S
xik ¼ 1; 8 i 2 T; (3)
k2S X X
yrk  jRj  yrk1 ; 8 k 2 S j k > 1: (11)
X X r2R r2R
k  xik  k  xjk ; 8 ði; j Þ 2 P; (4)
k2S k2S

X 4.2 Multiple-resource possibilities

ti  xik  c; 8 k 2 S; (5) As mentioned by Agpak and Gökçen (2005), Constraints (6)
i2T fail to assess scenarios in which alternative resources are
considered. Common tasks that can be performed by different
X resources must be represented by a separate set of alternative
xik  jTr j  yrk ; 8 r 2 R; k 2 S; (6)
i2Tr
resources required by task i (Ri). Constraints (12) may entirely
replace the previous ones (6) when multiple-resource
with: possibilities exist. This feature has not previously been
considered in a MILP model for the RCALBP-2 extension:
xik 2 f0; 1g; 8 i 2 T; k 2 S; (7)
X
xik  yrk ; 8 i 2 T; k 2 S: (12)
yrk 2 f0; 1g; 8 r 2 R; k 2 S; (8) r2Ri

c 2 R0 : (9)
4.3 Simplification hypothesis
Two goals are considered in this model by a lexicographical An implicit simplification hypothesis adopted by Alakas  et al.
objective function. The primary one, given by Expression (1), is (2020) in their CP model was the mandatory use of exactly one
the minimisation of cycle time (c). The minimisation of resource for each available station. They state that by assigning
resources is aimed in Expression (2) as a secondary goal. only one resource at each station, the minimum number of
Because this is a type-2 problem, the importance of c is much resources is guaranteed, and their CP model can focus on
higher than the number of used resources. The literature for achieving the objective of minimising cycle time. In a MILP
generalised type-2 ALBPs supports this approach, e.g. on lines formulation, such an assumption can be integrated by
with two-sided (Kim et al., 2009) or multi-manned stations appending either Constraints (13) or Constraints (14) with g =
(Michels et al., 2020), with task setup times (Zohali et al., 2022) 1. The former is equivalent to the ones used in Alakas  et al.
and resource-constrained lines (Alakas  et al., 2020). (2020), but the latter is more flexible. While Constraints (13)
The model is subjected to Constraints (3)–(6). Task explicitly prohibit the assignment of tasks that require different
assignments are enforced by Constraints (3), ensuring that each resources (i, j [ Ri \ Rj = 1) to the same station, one can easily
task is assigned to exactly one station. Constraints (4) modify the constant on the right-hand side of Constraints (14)
guarantee that the precedence relations are respected. A task i to reflect a more realistic limitation on the number of resources
can only be assigned to a station if all its predecessors jji p j that are allowed in each station. More interestingly, Constraints
have been assigned to that station or earlier ones. Constraints (13) and (14) can be combined to define incompatible
(5) limit the total processing time of tasks for each station to be resources as part of an RCALBP-2 extension. Constraints (13)
at most the cycle time (c). They also measure and identify the may prohibit specific resource pairs from going into the same
highest workload to be minimised. Finally, Constraints (6) station, whereas Constraints (14) restrict the maximum
indicate that a task i can only be assigned to a station k if the number of resources allowed in the same station:
required resource type r to perform task i is also assigned to
xik 1 xjk  1; 8 k 2 S; i; j 2 T j Ri \ Rj ¼ 1; (13)
the referred station k. For these last constraints, a bound on the
number of tasks that can be executed in the station is needed.
X
The cardinality of the subset of tasks that can be performed yrk  g; 8 k 2 S: (14)
with resource r (Tr) is clearly a valid bound. r2R

4.1 Resource availability constraints

An important generalisation proposed by Corominas et al. 4.4 Budget-oriented model
(2011) was to limit the number of resources to an available One last useful modification in the proposed MILP models is
quantity. This condition has not been extended to the taking into account individualised costs for each type of
RCALBP-2. Constraint (10) bounds the total number of resource (rr). Constraints (15) limit expenditure on used
resources to be used in the line to a. Naturally, this can be resources to a given budget amount (b) instead of the total

589
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

quantity [Constraints (10)]. Alternatively, a model that Table 2 Detailed solutions for the comparative instance
prioritises total cost (stations and resources) minimisation
AM RM FM
could also be proposed as a cost-oriented RCALBP-2
Station Task(s) Resource Task(s) Resource Task(s) Resource
extension:
XX 1 1,5 A 1,2 A 1,2 A
rr  yrk  b: (15) 2 4 B 4,5 B 6,8 B
r2R k2S 3 3,7 AÚC 3,7 AÚC 10 B
4 2,9 A 9 A 4,5 B
5 6 B 6,8 B 3,7 AÚC
4.5 Motivational example 6 8 AÚBÚC 10 B 9 A
The first MILP model herein proposed is composed of 7 10 B 11 A 11 A
Expressions (1)–(9) and (10)–(13). As resource assignment to 8 11 A – – – –
each station is limited to one – but not mandated to be one – #R 8 #R 7 #R 7
this model is named the restricted model (RM). The second c 8 c 8 c 8
MILP model is hereafter called the flexible model (FM): it
excludes Constraints (13) from the RM. Therefore, FM does
not impose a limitation on the number of resources that can be CP model (Section 5.1) and the second one on a newly
used in each station. Notice that both RM and FM are able to proposed data set (Section 5.2). All instances were solved on
cope with alternative multiple-resource possibilities. Gurobi 8.1 (Gurobi Optimization, 2019), which was used as
Even though the RM permits only one resource per station, it universal solver. Four threads of a 64-bit Intel CoreTM i5-
does not oblige each available station to take a resource as in the 7200U central processing unit (CPU) (2.5 GHz) with 8 GB of
Alakas et al. (2020) CP model (hereafter called AM) used for random-access memory (RAM) were used to run the models.
comparison. With both proposed models defined, this The models and their interaction with input and output data
advantage can be demonstrated in a sample scenario provided files were coded in Microsoft Visual Basic 2019 programming
by Alakas et al. (2020). In this comparative instance, there are language. The time limit is set to 3,600 s for all cases.
three types of resources (jRj = 3), and some tasks can be
performed by two or more resources (alternative). The number 5.1 Available data set
of available stations is taken as jSj = 8. Consequently, the There is only one data set available in the literature for the
number of available resources is also a = 8 for the comparison. studied RCALBP-2 (Alakas  et al., 2020). This data set samples
The precedence graph with 11 tasks (jTj = 11) from Figure 1 is five classical problem instances (Mitchell, Heskiaoff, Sawyer,
considered. Task processing times and resource requirements Kilbridge and Tonge) from Hoffmann (1990), totalling 40
for each task are given in Table 1. instances consisting of 21–70 tasks (jTj) from a well-known
Detailed solutions for the previous AM and the proposed literature benchmark, which is available for download at
models (RM and FM) are given in Table 2. Columns show <www.assembly-line-balancing.de>. Scenarios with dedicated
which tasks and resources were assigned to each station. The and alternative resource requirements are included, namely, ii
total number of used resources (#R) and the optimised cycle and iii, respectively. For each scenario, instances with 4 and 5
time (c) for each model are also presented. As expected, AM resource types (jRj) are designed. The total number of available
uses eight resources by compulsorily assigning one in each stations (jSj) and resources (a) are fixed to 15, 18, 20, 25, 30 or
available station. With that configuration, AM yields a cycle 35, depending on the instance. The computational experiments
time of eight-time units. The proposed RM and FM, on the were performed by comparing each model’s results in terms of
other hand, are able to obtain the same cycle time (c = 8) with cycle time (c) and the number of used resources along the line
fewer resources (#R = 7). This advantage is made possible by (#R). Finally, we noticed that some precedence arcs were
allowing unused stations and directly considering the use of missing in the instances made available by Alakas  et al. (2020)
resources instead of simply fixing exactly one resource per when compared to the literature benchmark. For fair
station. Indeed, both FM and RM found equivalent solutions comparison purposes, we decided not to include the missing
in terms of objective function value for this case, but FM may arcs to test our models (RM and FM) against the AM. Instead,
yield better solutions for larger instances with more resource we make the instances and detailed results available in the
alternatives, as further shown in the results from Section 5. Supporting Information, highlighting the missing arcs where
appropriate. The summarised results for each instance are
5. Computational results given in Table 3.
This section carries out two computational studies: the first one As AM and RM limit the number of resources to one in each
to compare the proposed models to the previously developed station, the most straightforward comparison is performed
between them. RM reached optimality for the entire data set.
Out of 40 instances, RM obtained three solutions with
Table 1 Task processing times and resource requirements for the
improved cycle times and the same number of resources, 27
comparative instance
solutions with fewer used resources and the same cycle time
Task index 1 2 3 4 5 6 7 8 9 10 11 and 10 equivalent solutions when compared to AM. The
Task time 6 2 5 7 1 2 3 6 5 5 4 instances in which RM obtained better cycle times than AM are
Resources A A, B A, C B A, B, C B A, C A, B, C A B A proven feasible, and the feasibility check is made available in
the Supporting Information. Ultimately, having unnecessary

590
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

 et al. (2020)
Table 3 Result comparison for the data set presented in Alakas resources and the same cycle time and 26 equivalent solutions.
Table 4 exemplifies a solution with improved cycle times and
Problem Scenario jRj jSj = a AM RM FM
the same number of resources. By taking the Mitchell scenario
c #R c #R c #R
ii jRj = 4 instances as an example, Table 4 shows how the cycle
Mitchell ii 4 15 17 15 17 15 13 15 time can be improved with the same number of resources, as
18 13 18 13 16 13 15 the flexibility of admitting more than one resource per station
5 15 13 15 13 15 13 14 allows a range of new assignment possibilities.
18 13 18 13 15 13 14 Finally, adding the flexibility to allow more than one resource
iii 4 15 13 15 13 10 13 10 per station is perfectly reasonable in the context of assembly
18 13 18 13 10 13 10 lines, particularly when large-size products (such as cars, buses
5 15 13 15 13 11 13 11 and trucks) are manufactured. In any case, the maximum
18 13 18 13 11 13 11 number of resources per station can readily be adjusted for each
Heskiaoff ii 4 15 108 15 108 13 108 13 type of resource by modifying Constraints (14). This
18 108 18 108 13 108 13 experiment empirically demonstrated that improved solutions
5 15 108 15 108 13 108 13 can be obtained when the flexibility of assigning more than one
18 108 18 108 13 108 13 resource per station is incorporated into the model.
iii 4 15 108 15 108 11 108 11
18 108 18 108 11 108 11
5.2 Proposed data set
5 15 108 15 108 12 108 12
An extended data set for the RCALBP-2 is herein proposed to
18 108 18 108 12 108 12
test the FM. We consider that the previous data set mentioned
Sawyer ii 4 15 34 15 34 15 34 15
in Section 5.1 did not have the most appropriate and
18 25 18 25 18 25 18
challenging instances for the studied problem. The reason for
5 15 32 15 32 15 32 15
that is mainly related to the excessive number of stations and,
18 26 18 26 18 26 18
consequently, resources. Given the limitation of one resource
iii 4 15 26 15 25 15 25 15
per station, the large number of available instances would
18 25 18 25 15 25 15
always guarantee the existence of feasible solutions. In fact, for
5 15 35 15 26 15 26 15
the simpler SALBP-2 version, the number of available stations
18 25 18 25 16 25 16
was large enough to the point that cycle times were actually
Kilbridge ii 4 20 55 20 55 20 55I 16
25 55 25 55 20 55I 16
bounded only by the largest task processing time.
5 20 55 20 55 20 55I 19
However, with the flexibility implemented on the FM,
25 55 25 55 20 55I 19
infeasible solutions can be easily avoided even if we only have a
iii 4 20 55 20 55 15 55 13 few available stations. This is due to the fact that, in the worst-
25 55 25 55 15 55 13 case scenario, all different resource types could be assigned to
5 20 55 20 55 15 55 14 the same station along with all tasks, which implies a cycle time
25 55 25 55 15 55 14 as high as the sum of all task processing times.
Tonge ii 4 25 231 25 231 25 205I 25 The proposed data set is presented in Table 5, and
30 156 30 156 29 156I 29 instances are made available as Supporting Information. The
5 30 197 30 197 30 180I 30
35 156 35 156 33 156I 33 Table 4 Detailed solutions for the Mitchell scenario ii jRj = 4 instance
iii 4 25 187 25 159 25 159I 25
AM RM FM
30 156 30 156 26 156I 26
Station Task(s) Resource Task(s) Resource Task(s) Resource
5 30 156 30 156 27 156I 27
35 156 35 156 27 156I 27 1 1 A 1 A 1,3 A^C
 I 2 3 C 3 C 4 D
Notes: New optimal solutions. Integer feasible solutions
3 4,12 D 4,12 D 5 A
4 5 A 5 A 7 C
5 2,6 B 2,6 B 2,6,14 B
6 7 C 7 C 8,12 D
stations operating with costly resources along the assembly line
7 8 D 8 D 9,11,13 A^C
is not a desired feature. This experiment demonstrates that
8 9,13,21 A 9,13,21 A 10,15,18,19 B^C
mandatorily fixing one resource in each available station may
9 10 B 10 B 21 A
lead models to obtain suboptimal solutions.
10 11,15 C 11,15 C 16 D
The second half of the experiment is designed to empirically
11 16 D 16 D 17 A
test the theoretical advantages of allowing more than one
12 17 A 17 A 20 D
resource to be assigned to the same station. Under the
13 14,18 B 14,18 B – –
computational processing time limit, the FM was able to yield
14 19 C 19 C – –
optimal solutions for 28 instances and feasible integer solutions 15 20 D 20 D – –
for the remaining 12 cases. Compared to the RM, the FM #R 15 #R 15 #R 15
obtained three solutions with improved cycle times and the c 17 c 17 c 13
same number of resources, 11 solutions with fewer used

591
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

Table 5 Proposed data set extension

Problem Tasks (jTj) Scenarios Resource Available Available
types (jRj) stations (jSj) resources (a)
Mitchell 21 ii,iii 4,5 8 8,10
Heskiaoff 28 ii,iii 4,5 8 8,10
Sawyer 30 ii,iii 4,5 10 10,12
Kilbridge 45 ii,iii 4,5 10 10,12
Tonge 70 ii,iii 4,5 12 12,15

same precedence graphs considered before are used, with jTj Table 6 Experimental results for the proposed data set extension
varying from 21 to 70 tasks. Nevertheless, corrections have
Problem Scenario jRj jSj a c cs Ratio CPU(s)
been made by including the missing arcs from the original
benchmark instances. Dedicated (ii) and alternative (iii) Mitchell ii 4 8 8 53 14 3.79 0.53
scenarios with 4 or 5 resource types (jRj) are once again taken 10 35 2.50 1.03
into account. The number of available stations (jSj) is much 5 8 8 76 5.43 0.47
lower, varying from 8 to 12, depending on the problems’ size. 10 51 3.64 1.00
The last variation is explicitly evaluating the effects of having iii 4 8 8 18 1.29 0.53
different numbers of available resources (a), ranging from 8 10 15 1.07 0.18
to 15 units. These modifications add 40 new instances to the 5 8 8 31 2.21 2.25
RCALBP-2 data set. 10 15 1.07 0.24
Table 6 presents the results obtained by the FM when it is Heskiaoff ii 4 8 8 274 129 2.12 2.64
applied to the proposed data set (Table 5). The comparison 10 153 1.19 1.42
here is made between the minimised cycle time obtained by the 5 8 8 513 3.98 1.11
FM (c) and the optimal SALBP-2 cycle time (cs). As the 10 250 1.94 7.41
number of available resources is scarce, the model always uses iii 4 8 8 139 1.08 4.15
all of them. In Table 6, the ratio stands for the numerical value 10 135 1.05 4.77
obtained by the fraction ccs and CPU(s) for the computational 5 8 8 175 1.36 1.77
processing time, given in seconds. Once again, detailed result 10 139 1.08 2.82
files are provided in the Supporting Information. Sawyer ii 4 10 10 108 34 3.18 14.69
From Table 6, it can be observed that the cycle times 12 59 1.74 27.61
obtained by the resource-constrained version of the problem 5 10 10 133 3.91 8.48
are always higher than when this limitation is not considered, 12 95 2.79 31.21
which is an expected outcome. Nonetheless, there are many iii 4 10 10 37 1.09 7.07
interesting aspects to analyse to extract managerial insights. All 12 35 1.03 8.49
other parameters equal, the optimal cycle time value quickly 5 10 10 41 1.21 15.80
converges to the SALBP-2 lower bound as the available 12 35 1.03 6.95
number of resources increases, which shows the possibility of a Kilbridge ii 4 10 10 163 56 2.91 28.73
trade-off between cycle time and the number of used resources 12 110 1.96 157.20
for practical situations in the decision-making process. As long 5 10 10 266 4.75 12.45
as both the cycle time and used resources are reduced to 12 161 2.88 23.28
monetary units, this trade-off can be promptly evaluated by the iii 4 10 10 78 1.39 132.75
assembly line’s manager by using the budget-constrained 12 59 1.05 118.98
extension proposed in Section 4. Furthermore, whilst 5 10 10 94 1.68 63.72
alternative resources boost assignment possibilities, the 12 64 1.14 145.80
increment of resource type requirements contributes to stricter Tonge ii 4 12 12 1,130 294 3.84 270.78
choices. Finally, the FM was able to achieve and prove optimal 15 835 2.84 1,375.30
solutions for all proposed instances. 5 12 12 1,434 4.88 628.12
15 1,145 3.89 1,474.63
6. Conclusion iii 4 12 12 860 2.93 661.35
15 412 1.40 2,049.35
The RCALBP with the objective of minimising the cycle time 5 12 12 927 3.15 428.08
under a limited number of stations and resources has been 15 481 1.64 2,608.09
addressed in this paper. The existing literature on both types I
and II of the problem has been reviewed, focusing on
mathematical models and including all practical variants. The approach, in which both task and resource assignments are
main contribution herein presented is the extension of a decisions to be taken, avoiding unnecessary use of stations and
recently proposed problem (Alakas  et al., 2020). Instead of resources. This situation is modelled and solved, including the
restricting the assembly line to compulsorily assign exactly one possibility of allocating only one or multiple resources into the
resource to all available stations, we propose a more holistic same station.

592
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

To solve the proposed optimisation problem, two new MILP manned workstations”, Journal of Manufacturing Systems,
models are developed and named RM and FM. These models Vol. 48, pp. 107-119.
ensure that all tasks are performed by the correct resources Corominas, A., Ferrer, L. and Pastor, R. (2011), “Assembly
while minimising cycle time and the total number of resources line balancing: general resource-constrained case”,
used along the line. The RM allows at most one resource to be International Journal of Production Research, Vol. 49 No. 12,
assigned to each station, whereas the FM is free to have one pp. 3527-3542.
resource of each type per station, as long as the total number of Ege, Y., Azizoglu, M. and Ozdemirel, N.E. (2009),
available resources is respected. A cost-oriented extension to “Computers & industrial engineering assembly line
directly evaluate a trade-off between productivity rate and used balancing with station paralleling”, Computers & Industrial
resources is also outlined. The results obtained from a literature Engineering, Vol. 57 No. 4, pp. 1218-1225.
data set indicate that forcing each available station to take one Friedman, A.D. (1985), Fundamentals of Logic Design and
resource may lead to suboptimal solutions. Thus, new optimal Switching Theory, Computer Science Press, New York,
solutions were obtained by explicitly taking into account ISBN:978-0-88175-110-9, p. 262.
resource assignment decisions. A new data set is proposed to Gurobi Optimization (2019), “Gurobi optimizer reference
represent shorter and more realistic lines. The FM was able to manual”.
produce optimal solutions to all new instances. In general, both Hoffmann, T.R. (1963), “Assembly line balancing with a
models were shown to be effective and flexible when dealing precedence matrix”, Management Science, Vol. 9 No. 4,
with ALBPs with resource limitations, providing optimal pp. 551-562.
solutions and applicable results for manufacturing industries, Hoffmann, T.R. (1990), “Assembly line balancing: a set of
as well as valuable managerial insights on the financial trade-off challenging problems”, International Journal of Production
between cycle time and resource usage. Research, Vol. 28 No. 10, pp. 1807-1815.
Following the line of research in the RCALBP-1, the MILP Kamarudin, N.H. and Ab Rashid, M.F.F. (2017), “Assembly
models herein proposed can be expanded to deal with different line balancing with resource constraints using new rank-
line layouts (U-line, parallel stations) and product variety based crossovers”, Journal of Physics: Conference Series,
(multi- and mixed-model lines). Additionally, further research Vol. 908 No. 1, p. 12059.
may also focus on cost-oriented formulations and (meta) Kao, H-H., Yeh, D.-H. and Wang, Y.-H. (2011), “Resource
heuristic algorithms can be developed for very large-size constrained assembly line balancing problem solved with
problems. ranked positional weight rule”, Review of Economics &
Finance, pp. 71-80.
Kao, H-H., Yeh, D-H., Wang, Y-H. and Hung, J-C (2010),
References “An optimal algorithm for type-I assembly line balancing
Agpak, K. and Gökçen, H. (2005), “Assembly line balancing: problem with resource constraint”, African Journal of Business
two resource constrained cases”, International Journal of Management, Vol. 4 No. 10, pp. 2051-2058.
Production Economics, Vol. 96 No. 1, pp. 129-140. Kim, Y.K., Song, W.S. and Kim, J.H. (2009), “A
, H.M., Pınarbas
Alakas ı, M. and Yüzükırmızı, M. (2020), mathematical model and a genetic algorithm for two-sided
“Constraint programming model for resource-constrained assembly line balancing”, Computers & Operations Research,
assembly line balancing problem”, Soft Computing, Vol. 24 Vol. 36 No. 3, pp. 853-865.
No. 7, pp. 5367-5375. Michels, A.S., Lopes, T.C., Sikora, C.G.S. and Magatão, L.
Askin, R.G. and Zhou, M. (1997), “A parallel station heuristic (2018), “The robotic assembly line design (RALD) problem:
for the mixed-model production line balancing problem”, model and case studies with practical extensions”, Computers
International Journal of Production Research, Vol. 35 No. 11, & Industrial Engineering, Vol. 120, pp. 320-333.
pp. 3095-3106. Michels, A.S., Lopes, T.C., Sikora, C.G.S. and Magatão, L.
Bartholdi, J.J. (1993), “Balancing two-sided assembly lines: a (2019), “A benders’ decomposition algorithm with
case study”, International Journal of Production Research, combinatorial cuts for the multi-manned assembly line
Vol. 31 No. 10, pp. 2447-2461. balancing problem”, European Journal of Operational
Baybars, I. (1986), “A survey of exact algorithms for the simple Research, Vol. 278 No. 3, pp. 796-808.
assembly line balancing problem”, Management Science, Michels, A.S., Lopes, T.C. and Magatão, L. (2020), “An exact
Vol. 32 No. 8, pp. 909-932. method with decomposition techniques and combinatorial
Becker, C. and Scholl, A. (2006), “A survey on problems and benders’ cuts for the type-2 multi-manned assembly line
methods in generalized assembly line balancing”, European balancing problem”, Operations Research Perspectives, Vol. 7,
Journal of Operational Research, Vol. 168 No. 3, pp. 694-715. p. 100163.
Boysen, N., Fliedner, M. and Scholl, A. (2007), “A Ogan, D. and Azizoglu, M. (2015), “A branch and bound
classification of assembly line balancing problems”, European method for the line balancing problem in U-shaped assembly
Journal of Operational Research, Vol. 183 No. 2, pp. 674-693. lines with equipment requirements”, Journal of
Boysen, N., Schulze, P. and Scholl, A. (2021), “Assembly line Manufacturing Systems, Vol. 36, pp. 46-54.
balancing: what happened in the last fifteen years?”, Otto, A., Otto, C. and Scholl, A. (2013), “Systematic data
European Journal of Operational Research, Vol. 301 No. 3, generation and test design for solution algorithms on the
pp. 797-814. example of SALBPGen for assembly line balancing”,
Chen, Y.Y., Cheng, C.Y. and Li, J.Y. (2018), “Resource- European Journal of Operational Research, Vol. 228 No. 1,
constrained assembly line balancing problems with multi- pp. 33-45.

593
 Mixed-integer linear programming models Assembly Automation
Adalberto Sato Michels and Alysson M. Costa Volume 42 · Number 5 · 2022 · 585–594

Pereira, J. (2015), “Empirical evaluation of lower bounding balancing”, European Journal of Operational Research,
methods for the simple assembly line balancing problem”, Vol. 168 No. 3, pp. 666-693.
International Journal of Production Research, Vol. 53 No. 11, Yadav, A., Kulhary, R., Nishad, R. and Agrawal, S. (2020),
pp. 3327-3340. “Parallel two-sided assembly line balancing with tools
Pereira, J. (2018), “Modelling and solving a cost-oriented and tasks sharing”, Assembly Automation, Vol. 40 No. 6,
resource-constrained multi-model assembly line balancing pp. 833-846.
problem”, International Journal of Production Research, Yang, M., Ba, L., Xu, E., Li, Y., Liu, Y. and Gao, X. (2020),
Vol. 56 No. 11, pp. 3994-4016. “Optimization of a multi-constraint two-sided assembly line
Pinarbasi, M., Alakas, H.M. and Yuzukirmizi, M. (2019), “A balancing problem using an improved imperialist
constraint programming approach to type-2 assembly line competitive algorithm”, Assembly Automation, Vol. 40 No. 2,
balancing problem with assignment restrictions”, Assembly pp. 273-282.
Automation, Vol. 39 No. 5, pp. 813-826. Yilmaz, H. and Yilmaz, M. (2017), “Multi-manned assembly
Quyen, N.T.P., Chen, J.C. and Yang, C.L. (2017a), “Hybrid line balancing problem: resource constrained case”,
genetic algorithm to solve resource constrained assembly line Proceedings of the International Conference on Industrial
balancing problem in footwear manufacturing”, Soft Engineering and Operations Management, Vol. 1, pp. 752-759.
Computing, Vol. 21 No. 21, pp. 6279-6295. Zohali, H., Naderi, B. and Roshanaei, V. (2022), “Solving the
Quyen, N.T.P., Kuo, R.J., Chen, J.C. and Yang, C.L. (2017b), type-2 assembly line balancing with setups using logic-based
benders decomposition”, INFORMS Journal on Computing,
“Dynamic programming to solve resource constrained
Vol. 34 No. 1, pp. 315-332.
assembly line balancing problem in footwear
manufacturing”, 4th International Conference on Industrial
Engineering and Applications, ICIEA 2017, pp. 66-70. Supplementary material
Rahman, H.F., Janardhanan, M.N. and Nielsen, P. (2020), The supplementary material for this article can be found
“An integrated approach for line balancing and AGV online at: https://data.mendeley.com/datasets/hr95zy2kgy
scheduling towards smart assembly systems”, Assembly
Automation, Vol. 40 No. 2, pp. 219-234. Corresponding author
Scholl, A. and Becker, C. (2006), “State-of-the-art exact and Adalberto Sato Michels can be contacted at: a.satomi
heuristic solution procedures for simple assembly line [email protected]

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: [email protected]

594
Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.