Journal of Manufacturing Systems 50 (2019) 180–192

Contents lists available at ScienceDirect

Journal of Manufacturing Systems

journal homepage: www.elsevier.com/locate/jmansys

An effective integer program for a general assembly line balancing problem T

with parallel workers and additional assignment restrictions
⁎
Bryan W. Pearce, Kavit Antani, Laine Mears, Kilian Funk, Maria E. Mayorga, Mary E. Kurz
Clemson University, United States

A R T I C LE I N FO A B S T R A C T

Keywords: The scope of the assembly line balancing problem in research is clear, with well-defined sets of assumptions,
General assembly line balancing parameters, and objective functions. In application, these borders are frequently transgressed. Many of these
Heuristics deviations are internal to the assembly line balancing problem itself, arising from any of the physical or tech-
integer programming models nological features in modern assembly lines. Other issues are founded in the tight coupling of assembly line
balancing with external production planning and management problems, as assembly lines are at the intersec-
tion of multiple related problems in job sequencing, part flow logistics, worker safety, and quality. General
assembly line balancing is devoted to studying the solution techniques necessary to model these applied line
balancing problems.
This article presents a complex line balancing problem based on the real production environment of our
industrial partner, featuring several extensions for task-to-task relationships, station characteristics limiting
assignment, and parallel worker zoning interactions. A heuristic, combining rank-position-weighting, last-fit-
improvement and iterative blocking scheme, and an integer program that can manage multiple constraint types
simultaneously, are developed. An experiment is conducted testing each of these new solution methods upon a
battery of testbed problems, measuring solution quality, runtime, and achievement of feasibility. Results indicate
that the integer programming model provides a viable solution method for those industries with access to
commercial solvers.

1. Introduction stations given a fixed cycle time; this optimization problem is NP-hard
[3].
The traditional form of an assembly line, as described by Scholl [1], ALB appeared in the literature in the 1950s. By the early 1970s,
is a production system consisting of a configuration of consecutive algorithmic ALB methods had proliferated, yet a survey at that time
workstations, typically using some material handling equipment to found that only approximately 5% of companies were using published
transport workpieces down the line at a constant rate. The total work to methods to solve their internal ALB problems [4]. Many articles attest
be performed along the assembly line is subdivided into the smallest to the continuing prominence of intuitive methods over algorithmic
indivisible elements of work, called tasks, and each task i possesses an ones developed by the research community, covering all decades of the
associated task time (ti). Tasks are related to one another by precedence intervening time period [5–8]. A field book published as recently as
attributes, i.e. some tasks must be finished before others can begin, 2012 [9] makes no mention of algorithmic methods at all, instead re-
usually due to the physical architecture of the workpiece. Stations are commending a manual approach, in consultation with a process expert
spaced along the line such that there is one workpiece present at each to ensure the balance is feasible. The simplest explanation for the lack
station, and all stations are allotted a fixed cycle time (C) to execute of algorithms implemented in industry may be the fact that finding a
tasks before the conveyor moves the workpiece to the next station. feasible solution to an ALB can usually be accomplished by hand. The
The assembly line balancing problem (ALB) is a production plan- manual solution will perhaps not be optimal, but might at least be good
ning problem concerned with allocating tasks to the stations on the enough to seem acceptable to management. Companies with expertise
assembly line, first proposed and formulated as a mathematical pro- in product design, as opposed to assembly system design, may not have
gramming problem in [2]. A solution to the ALB is an assignment of the information infrastructure to support computational methods of line
tasks to stations. One common objective is to minimize the number of balancing (perhaps precedence data is unavailable, for example).

⁎
Corresponding author.
E-mail address: [email protected] (M.E. Kurz).

https://doi.org/10.1016/j.jmsy.2018.12.011
Received 9 November 2018; Accepted 24 December 2018
0278-6125/ © 2019 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Additionally, there are certain normal translational difficulties for any used to minimize a composite fatigue score to address the effects of
new theoretical work. Industry adoption requires potential adopters to physical fatigue on workers from tasks that vary in difficulty in a system
learn that the theoretical methods exist, overcome organizational in- with a given number of stations and cycle time [15].
ertia resisting change, and, of course, financial investment to imple- Assembly lines with parallel workers allow multiple workers per
ment the change. station and are appropriate for workpieces large enough to permit
The real-world ALB problem may possess features that either by several workers with simultaneous access, e.g. larger home appliances,
themselves, or in conjunction with one another, are not modeled by any cars or airplanes. Allowing multiple workers per station may reduce the
published solution procedure. No published contribution offers ALB number of stations required, resulting in corresponding improvements
modeling methods with the constraint detail necessary to capture op- to factory floor space utilization and facility capital construction costs.
erational dynamics at a facility manufacturing complex products in a Consolidation of workers into the same station as one another may also
mixed-model environment. The use of ALB solution methods with in- allow sharing of fixed tooling resources between them, reducing capital
sufficient constraint modeling renders any generated solution vulner- costs. Material handling costs may also be reduced, as there are fewer
able to infeasibility. A practitioner might develop new methods as destinations to support with part delivery, spread across a smaller
needed, with appropriate background and skill. Given the deliverable- footprint. Lastly, multiple worker lines may realize superior line bal-
oriented nature of many process engineering job duties, however, it is ancing solutions due to reduced worker movement around the work-
perhaps uncommon that such research would be undertaken. piece, as each worker can be assigned tasks that only appear in a spe-
In this paper, we describe related literature and the specific problem cific zone [16,17].
environment of interest in Sections 2 and 3. We propose the MRPW-LFI- The first parallel ALB considered lines in which there are two dis-
WZBlock Heuristic in Section 4, followed by an integer programming tinct sides to the assembly line, and proposed a priority heuristic to
formulation in Section 5. In Section 6, we present the test bed of pro- generate a solution [16]. Tasks are classified into the sets {left, right,
blems, based on industry problem sets, upon which the heuristic and either} corresponding to the side of the line to which they may be as-
integer program are evaluated. Section 6 also contains the discussion of signed. A genetic algorithm methodology was later introduced for this
the results. Section 7 concludes the paper. gALB problem [18]. An important consideration for simultaneous work
is that precedence related tasks might be assigned to separate sides of
2. Related literature the line, resulting in one side waiting for the other to finish a task in
order to begin work on the next one. Two supplementary objective
The classification and terminology suggestions from the surveys of function were offered to address this issue: 1) work relatedness, which
Refs. [1,10,11,12] have been herein adopted, with any conflict favoring promotes tasks that have an immediate precedence relationship to be
the more recent publication. We focus this review on the features most assigned to the same station, and 2) work slackness, which promotes
relevant to our industrial setting. Salveson’s initial formulation is those tasks to have a large time gap between them if they cannot be
known as the simple assembly line balancing problem (sALB) [10,11], assigned to the same station [17]. A stochastic single-pass prioritization
as it features a number of simplifying assumptions: heuristic was created for a two-sided assembly line that manufactures
appliances [19]. An ant-colony metaheuristic was proposed for a two-
1 Mass-production of one homogenous product. sided assembly line that produces domestic products [20], which was
2 All tasks are processed in a predetermined mode (no processing extended to include secondary vertical and horizontal balancing ob-
alternatives exist). jectives [21]. Previous two-sided ALB models were extended and ad-
3 Paced line with a fixed common cycle time according to a desired dressed by a multi-objective particle swarm optimization approach that
output quantity. incorporates work relatedness, utilization, and vertical smoothing [22].
4 The line is considered to be serial with no feeder lines or parallel Four work zones and additional constraints to require certain sets of
elements. tasks be assigned to the same station were considered and addressed by
5 The processing sequence of tasks is subject to precedence restric- a hybrid DP and tabu search algorithm to vertically smooth workloads
tions. over a given number of stations [23].
6 Deterministic (and integral) task times. Solution methods for ALB problems are often based on the Ranked
7 No assignment restrictions of tasks besides precedence constraints. Positional Weight (RPW) heuristic [24]. RPW is designed for fixed cycle
8 A task cannot be split among two or more stations. time version of SALB, and attempts to pack stations with tasks by as-
9 All stations are equally equipped with respect to machines and signing them one at a time starting at the beginning of the assembly
workers. line. Each task is sorted by a weight that is equal to its task time plus the
task times of all of its successor tasks, which ensures that no task will be
Deviations from the above assumptions are classified as general ALB before one of its predecessors in the sorted list. Each task, considered in
(gALB) problems. Regardless of the deviation, almost all gALB problems the resulting order, is placed at the earliest station that meets both
still require the following minimal set of constraints to be satisfied: constraint criteria: 1) sufficient time capacity exists at the station to
place the task, and 2) no predecessor of the task is at a later station. The
• All tasks are assigned to some station, and the workpiece is finished Inverse Positional Weight heuristic is identical to RPW except that tasks
upon exiting the final station. are weighted as the sum of their task time and all predecessor task
• All precedence relationships must be satisfied. times. The assignment sequence begins at the final station instead of the
• The sum of task times at each station cannot exceed the cycle time. first [24]. The Related Activity heuristic scores tasks in the same way as
the Inverse Positional Weight heuristic but restricts task selection to the
In addition to precedence constraints, many real-world problems task with the largest score that is smaller than the cycle time [25]. The
exhibit tooling, zoning, worker skill, and other characteristics that re- selected task is assigned to a new worker, along with all predecessor
strict the assignment of tasks to stations and/or the relative assignment tasks. The algorithm then recalculates scores for all unassigned tasks,
of tasks to one another. Tooling availability and worker skill have been and repeats the assignment process. Upon completion the set of workers
addressed in a branch and bound model [13]. High complexity tasks are are then sequenced according to precedence validity.
grouped together such that high skill operators can be assigned to those Tonge [26,27], developed a heuristic that groups sets of tasks that
stations. A hybrid adaptive search and genetic algorithm has examined are directly connected in the precedence graph and treats these groups
task sets that are ineligible for assignment within the same station [14]. as single tasks. The resulting groups are approximately equal in ag-
Additionally, a hybrid heuristic/genetic algorithm approach has been gregate time. A valid assignment is sought using the much-reduced task

181
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

set size. If an assignment is found, the algorithm then seeks to improve and a WZ in order to ascribe them to a unique worker.
the solution by smoothing tasks from one station into an adjacent one, Each task i is encoded with one of nine Product Zones (PZ) L=
while preserving precedence and cycle time constraint satisfaction. If {RV,MV,LV,RM,MM, LM,RH,MH,LH} corresponding to location on the
no solution can be found with a given task grouping, then the groups vehicle with which the task interacts, divided into a 9-zone grid. Fig. 1
are broken into smaller subsets and the process repeats. Pinto et al. also displays the WZ and PZ zoning divisions.
utilized the network inspired procedure, in which tasks are grouped Workers may only be assigned a task if the worker’s WZ overlaps
together via adjacencies on the precedence graph. These super-nodes with the PZ of the task. Each WZ is initially eligible to cover three or
are then sequenced and fitted into stations to evaluate the resulting more PZ, as shown in Fig. 2. For example, a worker in the V WZ has the
balance. The procedure is iterated with differing heuristic rules for RV, MV, and LV product zones within reach.
constructing the super-nodes [28]. While Fig. 2 shows all potential matches between PZ and compa-
Baybars developed a set of preprocessing steps oriented to reducing tible WZ, some pairings cannot be activated simultaneously. To avoid
problem complexity [10,11]. The input data is analyzed for complexity interference problems between workers, each PZ may only be assigned
saving opportunities, such as decomposing the original problem into to one WZ within each station. For example, while the LV PZ may be
smaller sub-problems, or detection of implicit assignment constraints. assigned to either the L or the V WZ, it may not be assigned to both
After preprocessing a heuristic procedure assigns tasks one at a time within any given station. All tasks that are located in the LV PZ must be
beginning at the end of the assembly line. The task prioritization metric assigned to the same worker, either the L or the V worker.
considers the subset of tasks with no currently unassigned successors, Fig. 3 illustrates this zoning conflict. The lightning bolt flags 1 and 3
and then chooses the one with the highest number of predecessors. The show workers attempting to perform tasks upon the same area of the
chosen task is then assigned to the latest possible station. vehicle.
Tonge utilized a suite of prioritization metrics coupled with a one-
at-a-time task assignment approach [29]. After each task assignment, a 3.2. Task to task constraints
randomly selected heuristic is used to select the next task. Tonge re-
ported a competitive result from executing multiple runs of this heur- Adjacency requires the involved tasks to be performed consecutively
istic, though it is likely this performance was due to the wider solution by the same worker. Consider two tasks: Task A requires collecting a
space search afforded by the element of randomness in each run. part, and Task B installs that part on the vehicle. In addition, Task A
The COMSOAL algorithm [30] expanded upon the work of [29]. On must immediately precede Task B, because the worker’s hands are full
the first pass the selection probabilities are uniform between all tasks with the part and no other tasks should intervene. A same-takt (worker)
that have no unassigned predecessors, though these probabilities may constraint requires the involved tasks to be performed by the same
change as the algorithm iterates. COMSOAL can be characterized as a worker, but not necessarily consecutively. A same-station constraint
form of learning algorithm, as objective function performance is used as requires the involved tasks to be performed on the same station, but not
an input between iterations to bias the probabilistic selection steps in necessarily by the same worker. Consider the example of headliner
future iterations. Nine methods are given for inducing bias into the installation. Due to the size of the part, installation requires several
probabilistic search. workers to hold the part during the install. Each of these workers is
assigned a different task, but all tasks must be done in tandem.
3. Problem environment and additional constraints Collectively these tasks must be assigned to the same station. In con-
trast, the not-same-takt constraints forbid the same worker from per-
The motivating context for this research is automotive engineering. forming the related tasks, e.g. an inspection task that must be performed
We consider a vehicle to be the representative workpiece. Table 1 by a separate worker from the install task.
provides an overview summary of the features of the problem under
consideration in this paper. The classification system of Boysen et al. 3.3. Station constraints
[31] is noted for each feature in the right-most column.
This research utilizes the Type 1 objective function, to minimize the Resource constraints involve stations that possess fixed resources
number of workers, as cycle time is fixed by consumer demand and not that are necessary to complete certain tasks, e.g. robotic lift support,
subject to the line balancing process. While holding cycle time constant, pneumatic tooling. There may be one or more of each of these resources
the number of stations in the solution is a function of the optimization distributed along the line, and any task that interacts with a certain
performance. resource is forced to be assigned to a station that possesses the resource.
Fixed resources generally are located on one side of the line or the
3.1. Parallel workers and zoning constraints other, and can only be used upon the side of the vehicle that is facing it.
Fixed resource coverage zones (TZ) are defined as shown in Fig. 4. A
Zoning constraints are introduced to prevent interference between fixed resource covers the six PZ that are on the same side of the vehicle.
parallel workers in the same station. The physical space that the vehicle The fixed resource may satisfy the resource needs of any task in those
occupies on the conveyor is partitioned into a set W of five work zones PZ that is assigned at the station.
(WZ): {V (front), R (right), L (left), H (rear), and I (center)}. Every Vehicles may assume one of eight orientations relative to the con-
station will have between zero and five assigned workers, each of which veyor, and the orientation may change between stations. For example,
is responsible for a single WZ. Tasks must be assigned to both a station the vehicle may be oriented nose-first in one station, then rotate 90

Table 1
gALB Problem Features Considered.
Feature Description Class

Parallel workers Up to five workers may be assigned at each station, each with a non-overlapping work area dynamically determined by the set of tasks β3 = pwork 5
assigned
Mixed model Intermixed sequences of different models are produced on the assembly line α1 = mix
Grouped tasks Task groups define tasks that must be performed by the same worker, or in the same station α5 = link
Stationary resources Tasks that require fixed resources can only be assigned to stations that possess the resource α5 = fix
Task exclusion Some tasks cannot be assigned to certain stations α5 = excl

182
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Fig. 1. Work Zones (WZ) and Product Zones (PZ).

Fig. 2. Product Zones Eligible in each Work Zone.

Fig. 3. Zone Conflicts.

degrees such that the vehicle is sideways for the next station. Vehicle 3.4. Notation convention
orientation strongly affects zoning relationships, dictating which WZ,
PZ, and TZ are associated with one another at each station. Appendix B Table 2 presents the sets over which the parameters and variables in
in Supplementary material: IP Preprocessing shows zoning interactions the gALB problem are indexed, and the indexing variable typically used
in detail. when quantifying over each set. Table 3 presents input parameters for
Each PZ and WZ at a station may be inaccessible, due to the each problem instance.
structural layout of the station. An inaccessible WZ may not have a
worker assigned to the zone. Similarly, an inaccessible PZ at a station 4. Constructive and improvement heuristics
indicates that tasks of that PZ may not be assigned at that station.
Physical obstructions such as pillars, robotic machinery, or the work- Concepts from RPW have been extended to manage the features of
piece carrier itself are common causes for inaccessible zone constraints. the production environment in question, including zoning constraints,
The number of workers at any station may be capped at any value task groupings, and resource constraints. Several scoring metrics are
0–5 (there are 5 WZs). introduced that prioritize tasks that have difficult satisfaction require-
ments. The metrics prioritize tasks that require or support successor
tasks that require stationary resources on the assembly line.

183
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Fig. 4. Fixed Resource Coverage Zones (TZ).

Table 2
Sets.
Symbol Description Index

I set of all tasks {1, …, n} i, j

K set of all stations{1, …, |K |} k
W set of all work zones {L, R, V , H , I , P } b
L set of all product zones {LV , MV , RV , LM , MM , RM , LH , MH , RH } l
F set of all fixed resources (such as tools) {1, …, |F|} f
A set of assignment restriction types {adj, st , ss, nt } (adjacency, same takt, same station, not same takt) u

Table 3 task i. The ranked positional weight score (ri ) for each task is its own
Problem Input Parameters. time plus the sum of all successor task times, as shown in (1).
Symbol Description
ri = ti + ∑ tj
C Cycle time (sec)
j ∈ Qi (1)
ti Time of task i (sec)
A first-fit decreasing (FFD) assignment algorithm assigns tasks one
li Product zone of task i; li ∈ L
Fi Set of fixed resources required by task i; each member is an element of F at a time, in descending order of ri score. The first station that satisfies
pij Precedence relation between tasks i and j both precedence and cycle time constraints is assigned the task. The
pij = 1 if i is an immediate predecessor of j; 0 otherwise RPW prioritization scheme ensures all predecessors are assigned prior
aiju Assignment restriction of type u between tasks i and j to their successors, as ri > r j if i is a predecessor of j, as j ∈ Qi and Qi ⊃
aiju = 1 if i and j have relationship u ∈ A ; 0 otherwise Qj.
Fkp Set of fixed resources available at station k in product zone p; each
member is an element of F
Qifu Fixed resource requirement for task i 4.1.1. Extension: grouping constraints
Qifu = 1 if task i requires fixed resource f; 0 otherwise Adjacency, same-takt, and same-station relationships require the
W
skb W = 1 if WZ b is accessible in station k; 0
Work zone accessibility: skb related tasks to be assigned to the same station and/or worker. All tasks
otherwise in a same-takt group must also be assigned to the same station, and so
sklP Product zone accessibility: slpP = 1 if PZ l is accessible in station k; 0 may be considered to be within a same-station group. Adjacency groups
otherwise generalize to same-takt groups using the same logic. Adjacency related
skmax Maximum number of workers that may be assigned to station k tasks, then, may also be considered to be in a same-takt group together
as well as being in a same-station group together. The term Gi detailed
in (2), represents the group of tasks that are related to task i, after fully
Two improvement heuristics are developed in conjunction with the extending the domain of each task relationship.
constructive heuristic. The first, LFI, leverages the task prioritization
metrics from MRPW to consolidate tasks and remove lightly loaded j ∈ Gi iff aijadj = 1, aijst = 1, or aijss = 1 (2)
workers. The second improvement heuristic, termed the WZBlock ap-
proach, considers the problem of first selecting work zones then as- See the example in Fig. 5. Tasks 4–6 are in a same-takt group (da-
signing tasks. Two new work zone scoring metrics are developed, or- shed line oval), and tasks 6 and 8 are in an adjacency group together
iented towards superior selection of the work zones available for (solid line oval). Task 8 is implicitly included in the same-takt group.
activation. The adjacency group, however, does not expand to include tasks 4 or 5.
All four tasks are mutually involved in the same G group.
Eqs. (1) and (3) define a responsibility set of tasks that require task i,
4.1. Modified ranked positional weight heuristic either directly or indirectly. First define Q̄i as the set that contains all
tasks that succeed task i or any task grouped with task i, but not any of
The Modified Ranked Positional Weight (MRPW) heuristic, an the tasks within the group Gi itself.
adaptation of RPW [24], seeks to identify a solution to the gALB pro-
blem described in Section 3. Let Pi be the set of all tasks that are pre- Q¯ i = ( ⋃ Qg ) \Gi
g ∈ Gi (3)
decessors to task i, and Qi be the set of all tasks that are successors to

184
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

{
rir = max Zres ∀ res: res ∈ ⋃ =
j ∈ Qi
Fj }
(7)

4.1.3. MRPW algorithm

The MRPW algorithm applies the metrics from (1), (5), and (7) to
prioritize tasks according to task grouping, resource, and precedence
constraints. These metrics are combined in a hierarchy such that rir
dominates, followed by rig , then ri (Fig. 7).

4.1.4. MRPW remarks

The algorithm calculates ri , rig , and rir for each task i, by application
Fig. 5. Overlapping Task Groups.
of (1-(7. In lines 10 and 11 of the MRPW algorithm, the set of unas-
signed tasks is sorted, first filtering tasks by maximum rir . Ties are
The responsibility set Qi̿ recursively defines the set of all tasks that broken first by the largest rig , then the largest ri , and then arbitrarily
are dependent on task i, either directly by precedence relationships or from the candidates. Next, all tasks that are linked to task i via ad-
indirectly by grouping with tasks that have precedence relationships. jacency, same-takt, or same-station constraints are assigned to a group
which will all be assigned to the same WZ or station as appropriate.
Qi̿ = ( ⋃ Qh̿ ) ∪ Gi Lines 14 and 15 describe finding the station at which to begin the
h ∈ Q¯ i (4)
search, by considering precedence. On line 17, three conditions are
The set Gi is removed from set Q̄i in (3) to prevent self-referencing considered for assigning task i. First, if there is a task j at this station
recursion in the definition of (4), as would occur in the case of pre- with the same PZ as task i, then the only WZ at this station to which task
cedence relationships between tasks of the same group. Qi̿ contains task i can be assigned is the same WZ to which j is assigned. The second
i and all tasks in the group Gi , enabling calculation of a modified ranked condition checks whether the WZ that contains j has sufficient capacity
positional weight score rig that includes all tasks within the responsi- to add task i. The third condition checks whether the resource needs of
bility set of i, shown in (5) task i are met at this location. If all of these conditions hold then task i is
assigned to this station and WZ.
rig = ∑ tj Lines 20–23 consider all WZ at this station that are not empty
j ∈ Qi̿ (5) (possess at least one task). The motivation here is to attempt to add task
i to an existing WZ if possible, rather than open a new WZ. If there exists
This metric is analogous to combining grouped tasks into single an already open WZ that can hold task i’s PZ, and that WZ has sufficient
super-tasks, and as a result all tasks g ∈ Gi will be scored equivalently time capacity, and the resource needs of task i are met at this location,
by rig . Relative scoring to break these ties between in-group tasks can be then assign task i to that WZ.
measured with ri . Fig. 6 shows the growth of responsibility sets from The logic on line 25 considers relaxing the restriction that the WZ be
task groupings. non-empty. If the count of WZ with tasks assigned has not yet hit the
skmax limit at this station, then perhaps a new WZ can be opened to hold
task i. The time capacity and resource satisfaction assignment condi-
4.1.2. Extension: resource constraints
tions must be met here.
Recall that Fi is the set of fixed resources required by task i and |K | is
Line 30 increments to the next station, as a feasible assignment at
the number of stations available. Let Fkl be the set of resources available
the current station was not found. Line 31 checks whether the new
at station k in PZ l. Eq. (6) is an urgency score that measures the relative
station index exceeds the number of stations given as input. If so, then
importance of fixed resource res by the last station to possess res. For
the new station under consideration will not have any information re-
example, if there are 17 stations and a resource last appears on station
garding resource availability. This is considered a failure mode, as the
15, then that resource has Zres = 2 .
task will be assigned to a station beyond the bounds of the given input
Zres = |K | − max {k| ∀ l, res ∈ Fkl } data. All known resources are given to the new (dummy) station to
(6)
ensure that task i can be assigned.
Given two fixed resources res1 and res2, if the station number of the
final appearance of res1 is less than the final appearance of res2 then res1 4.2. Last-Fit-Increasing improvement heuristic
will have a higher urgency score, reflecting the fact that tasks that re-
quire res1 have fewer opportunities to assign their predecessors during a The Last-Fit-Increasing (LFI) improvement heuristic is a one-at-a-
first fit decreasing heuristic. To impose prioritization of these pre- time task reassignment method that seeks to consolidate tasks from
decessors each task i inherits the maximum Zres from their responsibility lightly loaded workers. It intends to improve the FFD solution by
set Qi̿ , as shown in (7). moving tasks to as late in the assembly line as possible without

Fig. 6. Group Definition of Responsibility Sets.

185
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Fig. 7. MRPW Algorithm.

Fig. 8. Last Fit Increasing Improvement Heuristic.

disrupting the current allocation of tasks, hopefully resulting in at least assembly line, relative to other tasks, as constraints have not inflated its
one empty WZ. Fig. 8 presents an algorithmic view of this logic. LFI MRPW metric. This task is a good candidate for pushing as far to the
begins with a feasible ALB solution, and utilizes the MRPW task metrics end of the line as possible, so that it might be far out of the way of other
ri , rig , and rir . In contrast to FFD, however, LFI in lines 8 and 9 considers tasks with more demanding constraints. The task is moved to the last
all tasks in increasing order of priority, first selecting the lowest-priority, already-active worker that can feasibly accept it (new workers may not
last-assigned task during MRPW. The small priority score for this task be activated during LFI). Iteration then continues to the task with the
indicates that it is maximally free for assignment anywhere on the next-lowest MRPW priority.

186
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Fig. 9. Work Zone Blocking Heuristic.

Consider all the stations in which successors of tasks in Gi are as- is a measure of the maximum degree to which WZ b at station k is
signed. In line 12 the initial station to begin the search, k, is chosen to needed by any task, normalized on the [0,1] scale. A measure of 0
be the earliest of those stations. This is the farthest that group Gi might uniqueness indicates that WZ b at station k is not particularly important
go toward the end of the line, due to precedence. Whichever successor to any task, as any task that is compatible with m b is also compatible
task is found during this search has already moved previously in the with every other WZ. On the other hand, a measure of 1 uniqueness
course of the LFI, as it has a lower MRPW score. Station k is examined to indicates that there exists some task for which WZ b at station k is the
determine if active workers can feasibly absorb group Gi . If so, the tasks only possible assignment.
are assigned and the loop proceeds to the next task group. If not, then
∑b'∈ W , b'≠ b Compi, b, k
the previous station will be considered. If no later station is found to Uniquenessb, k = Max i ∈ I {1 − |Compi, b, k = 1}
which group Gi can move, then the group will simply be reassigned at |W | × |K | − |W | (8)
their originally assigned station. The second WZ metric is the “flexibility” score, given in (9).
Flexibilityb, k is the proportion of tasks that may be assigned at WZ b at
4.3. Work zone blocking improvement heuristic station k. A measure of 1 flexibility indicates that the WZ is compatible
with every task. Zero flexibility indicates that the WZ is compatible
Solutions to the gALB problem specified in Section 3 typically ac- with no tasks.
tivate only a fraction of the WZs available. Solving the problem may be
considered in two sequential stages: first, to choose which WZ are ac- ∑i ∈ I Compi, b, k
Flexibilityb, k =
tivated, then to assign tasks to them. In the spirit of leveraging ag- n (9)
gregate task PZ to WZ compatibility patterns, two metrics are in-
Both the flexibility and uniqueness metrics provide insight into the
troduced that provide insight into the relative quality of activating each
relative usefulness and importance of each WZ. Both metrics are nor-
WZ.
malized to the scale [0,1]. To support the WZBlock heuristic, the two
metrics are added together to create a single composite score for each
4.3.1. Work zone metrics
WZ:
Let Compi, b, k be an indicator variable on whether task i is compatible
with WZ b at station k, based on the following constraints: Scoreb, k = Uniquenessb, k + Flexibilityb, k (10)

1 Fixed resources. If task i requires any fixed resources, then the TZ at

WZ’s station must provide them. 4.3.2. WZBlock heuristic algorithm
2 Accessibility. This checks both WZ and PZ accessibility. The PZ of The WZBlock algorithm proceeds by iteratively blocking work zones
task i must not be blocked at this station, nor may WZ b itself be from usage, by simulating accessibility constraints additional to any
blocked. that may be in the original problem data. The approach aspires to
3 Zone overlap. The PZ of task i must be associated with WZ b for identify and forbid the WZs that, if chosen for activation, are most likely
possible assignment, as per the mapping provided in Fig. 1. to cause infeasibilities or sub-optimality in the objective function. The
sum of the flexibility and uniqueness metrics presented in Section 4.3.1
Moreover, these conditions must be met for all tasks that are is used to discriminate between WZs. The algorithm shown in Fig. 9
grouped with task i via adjacency, same-takt, or same-station linkages, details the procedure.
not just for task i itself. The variable LBbest retains the best solution found through the
The first WZ metric is a “uniqueness” score, given in (8). The in- course of the algorithm. During each iteration, WZ b at station k with
ternal term ∑b′∈ W , b′≠ b Compi, b′, k counts how many other WZ are com- the smallest positive Flexibility + Uniqueness score is chosen, and WZ k
W
patible with a given task i. This quantity is divided by the total number is blocked (skb = 0 ). For redundancy purposes, composite scores of zero
of other WZ on all stations combined, |W | × |K | − |W |, and subtracted are not targeted, as a zero flexibility implies that no task can be as-
from 1, yielding the fraction of non-compatible WZs for task i. This signed to the WZ regardless. The MRPW heuristic is then applied with
value is then maximized over the set of all i, subject to i being com- the targeted WZ blocked from activation, followed by the LFI_Improve
patible with WZ b at station k. The final value delivered, Uniquenessb, m , heuristic. The forthcoming solution retained if it is an improvement

187
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Table 4 5.1.2. Objective

Criteria for Retention of Incumbent Solution. The objective function, z, minimizes the number of workers, as
LBtest LBbest ACTION measured by the count of active WZ, modifying the classic objective to
conflate WZ (instead of stations) with workers.
FEASIBLE FEASIBLE Retain higher utilization
FEASIBLE INFEASIBLE Retain LBtest minimize z = ∑ ∑ ykb
INFEASIBLE FEASIBLE Retain LBbest k∈K b∈W
INFEASIBLE INFEASIBLE Retain fewest dummy stations. If tied, retain higher
utilization
5.1.3. Constraints
Table 6 presents formulas for all constraints in the IP. The ID
upon the best-yet-found solution, whereupon the WZBlock algorithm column is referenced in subsequent text to provide description for the
considers the WZ to block for the next iteration. Looping continues until mechanics of each constraint.Constraint set (C1) enforces each task to
the iteration count exceeds a user-defined maxIter hyper-parameter, or be assigned to exactly one worker. Constraint set (C2) ensures the
no WZs are identified for potential blockage. average workload assigned to each active worker cannot exceed the
Heuristic solutions might not be feasible. Infeasible solutions con- cycle time. Constraint set (C3) ensures no more than skmax workers may
tain “dummy” stations with tasks that could not otherwise be assigned be active at station k. Constraint set (C4) enforces precedence con-
to any station. The “update LBbest” command checks whether candi- straints. Constraint set (C5) enforces adjacency and same-takt con-
date LBtest is superior to the incumbent LBbest. Feasible solutions are straints. Constraint set (C7) enforces same-station constraints. Con-
preferred over infeasible ones, using utilization (efficiency) as a metric straint set (C8) enforces not-same-takt constraints; this is a clique
of goodness. Table 4 presents the four possible scenarios, and corre- inequality, enforced only for cliques defined by each not-same-takt
sponding action. group. Constraint set (C9) enforces fixed resource constraints. Con-
straint set (C10) enforces WZ accessibility constraints. Constraint set
(C11) enforces PZ accessibility constraints. Constraint set (C12) en-
4.4. MRPW-LFI-WZBlock Heuristic
forces zoning compatibility. Tasks may only be assigned a WZ at some
station if the PZ of the task is compatible with that WZ. See Appendix B
The heuristic proposed includes the three elements previously de-
in Supplementary material: IP Preprocessing for derivation of the B
scribed and is refered to as the MRPW-LFI-WZBlock Heuristic. WZBlock
parameter. Constraint set (C13) enforces zone overlap constraints; this
is used as the outer wrapper heuristic, with maximum iteration hyper-
constraint considers all task pairs i and j that share the same PZ, in-
parameter set to 10. Within each iteration, the MRPW constructive
dicated by the preprocessing parameter wij = 1. If i and j are assigned to
heuristic is executed, followed by the LFI improvement heuristic.
the same station, and task i is assigned to WZ b, then task j must also be
The computational complexity of each iteration of the WZBlock
assigned to WZ b. This is accomplished by restricting j from assignment
heuristic is O(n |K |), where n is the number of tasks and |K | is the
to any other WZ that is not b. This constraint prevents workers from
number of stations. The computational complexity is O(n2 |K |) for the
interfering with one another. Without this constraint it would be pos-
MRPW and LFI heuristics.
sible for two workers to simultaneously attempt tasks within the same
PZ. Constraint sets (C14) and (C15) ensure all decision variables are
5. Integer programming model binary.

In this section an integer programming (IP) formulation is presented 6. Application of solution methodologies
that models the gALB problem introduced in Section 3. Notation for the
sets and input parameters are shown in Section 3.4. As discussed in This section describes a computational experiment for the MRPW-
Appendix B in Supplementary material: IP Preprocessing, many of these LFI-WZBlock Heuristic and the IP formulation. For the IP, performance
input parameters are preprocessed before execution of the IP. Outputs criteria of interest relate to the computer time requirements necessary
from the preprocessing routine are shown in Table 5. Section 5.1 pre- to solve the problem, and the scaling of this time as a function of
sents the IP formulation in three parts: decision variables, objective problem size. The primary criteria of interest for the heuristic method is
function, and constraints. the quality of the solution, as measured by optimality gap.

6.1. Test data sets

5.1. IP formulation

The ALB literature provides no testbed data sets that exhibit all
5.1.1. Decision variables
constraints desired for this paper. Three sets of test data were collected
1 if task i is assigned to station k and WZ b during the development of these methods, in conjunction with our in-
x ikb = ⎧ dustrial partner. The three data sets corresponding to assembly line
⎨
⎩ 0 else
sections are labeled “Band 1”, “Band 26”, and” Band 30.” Table 7
summarizes some properties of each of these initial data sets.
1 if WZ b at station k is active
ykb = ⎧ These three initial data sets form the testbed basis of the experi-
⎨
⎩ 0 else ment. In addition, an array of ALB sub-problems are created from the

Table 5
Input Parameters Derived During Preprocessing.
Symbol Description

c
Qklf Fixed Resource Coverage:
c = 1 if station k has fixed resource f that can be used in PZ l
Qklf
Bkbl WZ−PZ compatability: Bkbl = 1 if WZ b can contain PZ l at station k ; 0 else
wij Indicates that tasks i and j have the same product zone. wij = 1 if li = l j ; 0 else

188
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Table 6
IP Constraints.
Constraint Formula Quantification and Condition Set

∑k ∈ K ∑b ∈ W xikb = 1 ∀i∈I C1
∑i ∈I ti xikb ≤ Cykb ∀ k ∈ K, b ∈ W C2
∑b ∈ W ykb ≤ skmax ∀k∈K C3
|K| v
∑k = v + 1 ∑b ∈ W xjkb ≤ 1 − ∑k = 1 ∑b ∈ W xikb ∀ v = 1…|K | − 1, if pij = 1 C4
xikb = xjkb ∀ i, j ∈ I , k ∈ K , b ∈ W , if aijadj = 1 or aijst =1 C5

∑b ∈ W xikb = ∑b ∈ W xjkb ∀ i, j ∈ I , k ∈ K , if aijss = 1 C7

xikb + xjkb ≤ 1 ∀ k ∈ K , b ∈ W , for i and j where aijnt = 1 C8

Qifu (∑b ∈ W xikb) ≤ Qkc, l , f ∀ i ∈ I, k ∈ K, f ∈ F C9

i
xikb = 0 W = 0
∀ i ∈ I , k ∈ K , b ∈ W , if skb C10
∑b ∈ W xikb = 0 P
∀ i ∈ I , k ∈ K , if skb =0 C11
xikb = 0 ∀ i ∈ I , k ∈ K , b ∈ W , if Bkbli = 0 C12
xikb + ∑b′∈ W \ b xjkb′ ≤ 1 ∀ i, j ∈ I , k ∈ K , b ∈ W , if wij = 1 C13
xikb ε {0,1} ∀ i ∈ I, k ∈ K, b ∈ W C14
ykb ε {0,1} k ∈ K, b ∈ W C15

Table 7 Table 8
Test Data Set Properties. Summary Results (Feasible and Infeasible for MRPW-LFI-WZBlock).
Data Set # Stations # Tasks # Unique FIXED RESOURCES Band Number of Number of Average AMPL MRPW-LFI-WZBlock
Stations Instances Elapsed Time
Band 1 13 396 12 (s) Average Average gap
BAND 26 9 317 12 Time (s)
Band 30 10 300 3
1 3 11 0.16 24.00 0.30
4 9 0.32 30.90 0.28
4 1 0.09 24.43 ∞
initial data sets using the tasks assigned in a provided feasible solution,
5 8 0.65 13.74 0.22
from every possible partition of three or more adjacent stations. This 5 1 0.21 28.37 ∞
resulted in 125 datasets with between 58 and 395 total tasks over three 6 8 1.07 7.70 0.26
to 13 stations. Data sets are available from the authors upon request. 7 7 3.68 57.33 0.28
8 6 5.37 59.55 0.28
9 5 7.54 60.81 0.23
6.2. Experiment execution 10 4 9.67 60.75 0.38
11 3 11.75 60.19 0.44
The IP was applied to each data instance and executed on the Linux- 12 2 15.85 48.49 0.56
based Palmetto Cluster at Clemson University. The IP formulation is 13 1 20.44 60.16 0.53
26 3 3 0.21 26.27 0.33
modeled in AMPL, and run using the Gurobi 5.0 Linux 64 solver. For
3 2 0.23 29.47 ∞
each problem instance 8 processors and 120gb of RAM are allocated. 4 2 0.54 30.59 0.48
MRPW-LFI-WZBlock is applied to each data instance with the hyper- 4 2 0.38 30.44 ∞
parameter set to 10. This setting is chosen because offline testing has 5 1 1.53 10.44 0.57
shown limited improvement in objective function for values beyond 10. 5 3 0.77 1.66 ∞
6 4 1.23 9.21 ∞
MRPW-LFI-WZBlock is implemented in VBA, and executed on a 64-bit
7 3 1.99 60.37 ∞
Windows PC with 2.40 GHz processor and 2GB RAM. The experimental 8 2 3.18 10.36 ∞
data is summarized in Table 7. 9 1 6.30 52.63 ∞
30 3 8 0.32 1.60 0.00
4 7 0.87 11.82 0.00
6.2.1. Heuristic feasibility
5 6 1.88 19.04 0.00
Some instances of the data were not amenable to solution with 6 5 3.35 31.47 0.00
MRPW-LFI-WZBlock, and entered into the aforementioned “failure 7 4 5.02 42.15 0.04
mode”. The instances that did not result in solution from MRPW-LFI- 8 3 9.24 48.40 0.14
WZBlock are included in Table 8, with a separate row, italicized font, 9 2 12.47 54.72 0.13
10 1 20.62 60.39 0.13
and an average gap of ∞. MRPW-LFI-WZBlock found feasible solutions
for 106 out of 125 problem instances. Of the 19 heuristic-infeasible
instances, 17 of them are sourced in Band 26 data, and 2 are from Band bands. It appears that there are some characteristics particular to each
1. band which carry strong implications towards the IP runtime.

6.2.2. IP runtime
The average time to execute the IP model was 3.21 s, aggregated 6.2.3. MRPW-LFI-WZBlock optimality gap and time
z i − z opt , i
across all datasets; the IP model successfully solved all 125 instances. The optimality gap for instance i is measured by Gapi = z opt , i
,
The maximum time is under 25 s. It is remarkable to have solved a 400- where z opt , i is the optimal value of the objective function, as determined
task problem, easily a middle-sized problem by ALB standards, in only by the IP solution, and z i is the value of the objective function found by
20 s. the heuristic. If the heuristic finds an optimal solution, then the gap is
Fig. 10 presents AMPL Elapsed Time versus task and station counts, zero. The heuristic found 31 optimal solutions of 125, or approximately
with each band’s datasets collected separately. Fig. 10 shows particu- 25% of problem instances. Of the 31 optimal instances, 29 were from
larly strong differentiation between bands, and consistency within Band 30, two were from band 1 and none were from Band 26. The

189
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

Fig. 10. AMPL Elapsed Time by Band and Task, Station Count.

average gap in the cases in which the optimal solution is not found is measuring constraint satisfaction scarcity. Responsibility sets are in-
0.204, and in the worst instance is 0.62. troduced to encapsulate task-to-task precedence and assignment
The average time to execute MRPW-LFI-WZBlock is 31.9 s. While linkage constraints. Urgency score is introduced to measure assignment
this is difficult to compare with the IP running time, due to the plat- limitations due to resource constraints. Tasks are weighted by a com-
forms on which each is executed, the 10-fold increase is notable. posite prioritization score based on these new metrics, and assigned
according to a first-fit-decreasing single-pass heuristic. The MRPW
6.3. Discussion heuristic is oriented toward creation of feasible solutions, with effi-
ciency being a secondary consideration. The second component, the
The runtime plots of the IP suggest accelerating growth in runtime Last Fit Increasing improvement heuristic leverages the task prior-
with respect to the size of the problem. Indeed, ALBs are NP-hard, and itization rankings of MRPW, and consolidates flexible tasks into
extremely large problems will certainly be intractable. The instances otherwise lightly-packed workers. The goal of the LFI improvement
solved here span up to 400 tasks and 13 stations, what might be con- heuristic is to improve the efficiency of a feasible solution. The Work
sidered mid-sized problems in ALB, with runtimes under one minute. Zone Blocking heuristic focuses on the first work zone selection sub-
Each band represents an independent production process, with problem of the bifurcated ALB problem. The purpose of this approach is
several key differences that might help illuminate band-specific differ- to address zoning difficulties encountered in the MRPW heuristic. Two
entiation in performance. All tasks in Band 1 belong to one of the four new metrics are introduced to support the heuristic, measuring work
corner PZs: {LV, RV, LH, RH}, and in most stations workers may only be zone flexibility and uniqueness. The metrics are developed in con-
assigned to the L or R work zones. Only two stations support three sideration of each WZ’s offerings in terms of satisfying task needs with
parallel workers (the rest cannot have more than two), limiting the respect to zoning, fixed resource, and accessibility constraints.
potential for overlapping work zones. There are many fixed resources in An IP is developed to manage the zoning and worker parallelization
Band 1, but for most tasks that require fixed resources there is only one aspects of the problem. Preprocessing transformations render several
WZ which can satisfy both fixed resource and zoning needs, simplifying complex facets of the problem into representations amenable for a tight
the decision problem by forcing task assignment. Band 26 is relatively IP formulation.
complex, with the full complement of up to 5 parallel workers per- Each solution methodology is applied to a testbed of 125 instances
mitted at many stations. Tasks are located in every PZ. Fixed resources derived from real ALB data collected in conjunction with our industrial
are common, though many are duplicated across two or more stations, partner. The IP is benchmarked primarily according to the runtime
permitting tasks requiring fixed resources to be assigned in one of required relative to the size of the instance, to which it performs sur-
several WZs. Band 30 is a single-sided assembly line, with only one prisingly well, needing only 22 s at most to solve an instance. The IP
worker permitted per station. Every task is located in the same PZ, and solution is also used to benchmark heuristic performance. The heur-
fixed resources are sparse on the line. Task grouping is relatively istics were able to find feasible solutions for 83.2% of problem in-
common in Band 30, but otherwise this band is easily the simplest of stances. The heuristics averaged an optimality gap of approximately
the three with respect to constraint complexity. 20%, and found the optimal solution for 25% of the instances. Due to
Fig. 10 shows Band 30 instances requiring the most IP time to solve, superior performance and adaptability, the IP is recommended for in-
Band 26 the least, and Band 1 in the middle. For any given problem size dustrial application.
(in terms of either task or station count), Band 30 instances required
approximately 4x runtime relative to Band 26 instances, and approxi- 7.1. Implications for future work and industry implementation
mately 2x runtime relative to Band 1 instances. It appears that addi-
tional constraints lower IP runtime in general, as a highly constrained A manufacturer with operations encapsulated by the gALB char-
instance has a smaller feasible region within the solution space. acteristics for this problem can implement these methods. The algo-
rithmic approach is suitable for embedding within a commercial line
7. Summary and conclusions balancing visualization software tool. Direct use cases include support
for initial (product launch) line balancing, and periodic rebalancing to
The MRPW-LFI-WZBlock Heuristic heuristic incorporates three respond to forecasted demand changes. Secondary use cases include
methods. The first, MRPW, introduces a prioritization scheme driven by exploration of input parameter variation, e.g. cycle time changes, or

190
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

relocation of fixed equipment. logistics, or facility design. The IP does not consider task sequencing, so
Industry application of ALB methods commonly encounter diffi- an infeasibility of this kind may be present in the solution produced by
culties in extending existing methods to account for gALB features the IP. Through post-processing review, the data sets used did not result
specific to the problem environment. For the gALB environment con- in any violations. It is possible to extend the IP to manage task se-
sidered by the methods here, no existing ALB methods were suitable for quencing, and thereby prevent such infeasibilities, but at the cost of
immediate application, largely due to the unique zoning features. introducing a new set of decision variables related to sequencing.
During the course of this research, the heuristic methods were created
first, and the IP formulation later. More than a year was spent working 7.1.2. Heuristics
with our industry partner, both to collect data and to understand the Relative to the IP, the heuristic methods are ill-suited for extensions
various constraints that appear in the problem. Some constraint types, that add constraints or other gALB features. The first ancestral version
e.g. not-same-takt constraints, are especially rare on the assembly lines of the MRPW heuristic was developed early in the research project,
under study, and were not discovered until late in the process. The ALB before discovering many of the constraints now represented. Since that
methods detailed in this work are certainly extensible for application to time, each constraint added has induced excessive difficulty when
problem domains outside our industrial partner’s, for which the adapting the MRPW method
methods were specifically designed. Issues related to industrial appli- Further, the experiment has shown significant performance pro-
cation of the IP and heuristic methods are discussed separately in blems for the heuristics, in terms of finding feasible solutions and in the
Sections 7.1.1 and 7.1.2, respectively. quality of those solutions. MRPW-LFI-WZBlock is possibly useful in a
few cases. The first is the case of extraordinarily large problem size. The
runtime of heuristics scales in a polynomial fashion with respect to
7.1.1. IP formulation result
problem size and will experience a slower growth rate than the IP. The
The IP performed exceptionally well for all problem instances in the
second scenario for application of heuristic methods is if resources for
experiment. Assuming availability of a solver such as Gurobi, it is the
solving IPs are unavailable.
recommended solution to any industry customer with an applicable and
comparably-sized ALB problem. It is difficult to speculate on IP runtime
Acknowledgment
performance for problems larger than in the experiment, as runtime will
certainly experience combinatorial growth rates at some size. Perhaps
Funding for this research was provided by BMW Manufacturing Co.
problems up to an order of magnitude of the largest instances in this
experiment (similar in size to the largest ALB problems considered in
Appendix A. Supplementary data
any literature) would still find acceptable runtimes.
The IP is particularly well-suited for constraint extensions that in-
Supplementary material related to this article can be found, in the
volve task-to-task or task-to-station assignment compulsion or for-
online version, at doi:https://doi.org/10.1016/j.jmsy.2018.12.011.
biddance. Several constraints of this variety already exist within the
current gALB problem, implemented with relatively clear, direct, and
References
tight IP constraints. Presumably, extending the IP for another gALB
environment by adding more constraints of this type would be rela-
[1] Scholl A. Balancing and sequencing assembly lines. 2nd ed. Heidelberg: Physica-
tively simple. Indeed, during development of this IP the not-same-takt Verlag.; 1999.
constraints were added late in the timeline, but were easily modeled in [2] Salveson M. The assembly line balancing problem. J Ind Eng 1955;6(3):18–25.
the IP structure. The not-same-takt task sets in this project were limited [3] Wee T, Magazine M. Assembly line balancing as generalized bin packing. Oper Res
Lett 1982;1(2):56–9.
to sets of two tasks, and as such, the provided constraint C8 is sufficient.
[4] Chase R. Survey of paced assembly lines. Ind Eng 1974;6(2):14–8.
However, if a not-same-takt task set contained more than two tasks, this [5] Schöniger J, Spingler J. Planung der Montageanlage. Technica 1989;14:27–32.
constraint set can be expanded to take advantage of the clique structure [6] Milas G. Assembly line balancing… let’s remove the mystery. Ind Eng
as follows. Let Ω̄ be the set of all not-same-takt sets, each of which is 1990;22:31–6.
[7] Erel E, Sarin S. A survey of the assembly line balancing procedures. Prod Plan
labeled Ωo , o ∈ {1, …, |Ω̄|} . None of the tasks i ∈ Ωo can be assigned to Control 1998;9(5):414–34.
the same takt, so the following constraint set C8’ can be utilized: [8] Boysen N, Fliedner M, Scholl A. Production planning of mixed-model assembly
lines: overview and extensions. Prod Plann Control: Manage Oper
∑i ∈ νΩo x ikb ≤ 1 ∀ k ∈ Κ , b ∈ W , for o ∈ {1, …, |Ω̄|} C8’ 2009;20(5):455–71.
[9] Townsend B. The basics of line balancing and JIT kitting. Boca Raton, FL: Taylor &
Francis Group; 2012.
The IP features three distinct zone types: work zones, product zones, [10] Baybars I. A survey of exact algorithms for the simple assembly line balancing
and fixed resource coverage zones. The implementation details of these problem. Manage Sci 1986;32:909–32.
zones, such as their mapping relationships, are easily customizable. It is [11] Baybars I. An efficient heuristic method for the simple assembly line balancing
problem. Int J Prod Res 1986;24:149–66.
possible to add, remove, or re-map any of the zoning features with [12] Becker C, Scholl A. A survey on problems and methods in generalized assembly line
reasonable effort. Such changes would require no alteration of the IP balancing. Eur J Oper Res 2006;183:694–715.
formulation itself, only redefinition of the preprocessing parameters, in [13] Johnson R. A branch and bound algorithm for assembly line balancing problems
with formulation irregularities. Manage Sci 1983;29:1309–24.
which zone relationships are encapsulated. [14] Bautista J, Suarex R, Mateo M, Companys R. Local search heuristics for the as-
Implementing task sequencing constraints would require adding sembly line balancing problem with incompatibilities between tasks. Proceedings of
new decision variables to the IP to ensure that task start/stop times are the 2000 IEEE international conference on robotics and automation 2000:2404–9.
[15] Carnahan B, Norman B, Redfern M. Incorporating physical demand criteria into
properly managed. Adding these variables and associated sequencing assembly line balancing. Iie Trans 2001;33:875–87.
constraints to the IP is relatively direct in terms of formulation, but [16] Bartholdi J. Balancing two-sided assembly lines: a case study. Int J Prod Res
might present significant consequences in terms of runtime. Adding 1993;31:2447–61.
[17] Lee T, Kim Y, Kim Y. Two-sided assembly line balancing to maximize work relat-
decision variables might always be expected to add runtime, but in
edness and slackness. Comput Ind Eng 2001;40:273–92.
particular start/stop time variables are quantified over the real num- [18] Kim Y, Kim J, Kim Y. Two-sided assembly line balancing: a genetic algorithm ap-
bers. All other variables currently in the IP are binary, significantly proach. Prod Plan Control 2000;11:44–53.
[19] Lapierre SD, Ruiz AD. Balancing assembly lines: an industrial case study. J Oper Res
restricting the size of the solution space. Timing variables changes the
Soc 2004;55:589–97.
IP from a BIP to a MILP, and runtime penalties should be expected. [20] Baykasoglu A, Dereli T. Two-sided assembly line balancing using ant-colony-based
Problem extensions that permit the IP to touch on related produc- heuristic. Int J Adv Manuf Technol 2008;36(5-6):582–8.
tion planning problems would necessitate large-scale adaptations to the [21] Simaria A, Vilarinho P. 2-ANTBAL: an ant colony optimization algorithm for bal-
ancing two-sided assembly lines. Comput Ind Eng 2009;56:489–506.
IP. Examples include extensions to accommodate job sequencing, part

191
B.W. Pearce et al. Journal of Manufacturing Systems 50 (2019) 180–192

[22] Chutima P, Chimklai P. Multi-objective two-sided mixed-model assembly line bal- Laine Mears, Ph.D., P.E. is the SmartState Endowed Chair of Automotive Manufacturing
ancing using particle swarm optimization with negative knowledge. Comput Ind at Clemson University. He earned a BS in mechanical engineering from Virginia Tech and
Eng 2012;62:39–55. MS and Ph.D. degrees in mechanical engineering from Georgia Tech. He is a Fellow of
[23] Pastor R, Corominas A. Assembly line balancing with incompatibilities and bounded both the American Society of Mechanical Engineers and SME, and associate editor of the
workstations. Ricerca Operativa 2000;30:23–45. ASME Journal of Manufacturing Science and Engineering and the SME Journal of
[24] Helgeson W, Birnie D. Assembly line balancing using the ranked positional weight Manufacturing Systems. His research in state estimation and control of manufacturing
technique. J Ind Eng 1961;12:394–8. equipment is sponsored by the National Science Foundation, Office of Naval Research and
[25] Agrawal P. The related activity concept in assembly line balancing. Int J Prod Res numerous industrial partners.
1985;23:403–21.
[26] Tonge F. Summary of a heuristic line balancing procedure. Manage Sci
1960;7:21–42. Kilian Funk, Ph.D., is a senior software developer with the BMW Group. He earned his
[27] Tonge F. A heuristic program for assembly line balancing. New Jersey: Prentice- diploma in mechanical engineering and his Ph.D. in technical mechanics from the
Hall; 1961. Technische Universität München (TUM). He spent many years with BMW research
[28] Pinto P, Dannenbring D, Khumawala B. A heuristic network procedure for the as- working on driver assistance systems, mechatronic system design, assembly and pro-
sembly line balancing problem. Naval Res Logist Rev 1978;25:299–307. duction planning methods. His is currently working in the field of autonomous driving.
[29] Tonge F. Assembly line balancing using probabilistic combinations of heuristics.
Manage Sci 1965;11:727–35. Maris E. Mayorga, PhD is a Professor in the Fitts Department of Industrial and Systems
[30] Arcus A. COMSOAL: a computer method of sequencing operations for assembly Engineering at North Carolina State University. She earned her BS in Mechanical
lines. Int J Prod Res 1966;4:259–77. Engineering from The George Washington University and her MS and PhD degrees in
[31] Boysen N, Fliedner M, Scholl A. A classification of assembly line balancing pro- Industrial Engineering and Operations Research from the University of California,
blems. Eur J Oper Res 2007;183:674–93. Berkeley. She is a member of IISE and INFORMS. She is an associate editor for IISE
Transactions, OMEGA (the International Journal of Management Science), IISE Transactions
Bryan W. Pearce, Ph.D., is a technical process engineering at First Quality. He earned his on Healthcare Systems Engineering, Systems Science and Service Science, and the INFORMS
BS, MS and PhD in Industrial Engineering from Clemson University. His interests are in Journal on Computing. Her research interests are modeling and optimization of complex
optimization and risk management for capital project planning and production planning systems under uncertainty, including healthcare and logistics.
applications.”
Mary E. Kurz, PhD is an associate professor in Industrial Engineering at Clemson
Kavit Antani, Ph.D. is an assembly quality manager at the BMW Manufacturing plant in University. She earned her BS and MS in Systems Engineering and her PhD in Systems and
Spartanburg, SC responsible for X3, X4, X5, X6 and X7 sports activity vehicles. He earned Industrial Engineering from the University of Arizona. She is a Senior Member of IISE and
his BS in Production Engineering from University of Mumbai, India, his MS in Industrial a member of INFORMS and SME. She is a member of the editorial board for the Journal of
and Systems Engineering from Auburn University and his Ph.D. in Automotive Manufacturing Systems. Her research interests are in optimization for assembly-related
Engineering from Clemson University. His research interests are in automotive assembly, tasks, using exact techniques, heuristics and metaheuristics.
manufacturing complexity and its effects on product quality, and powertrain engineering.
He is an active member of SME, SAE and ASQ.

192