Computers and Operations Research 115 (2020) 104863

Contents lists available at ScienceDirect

Computers and Operations Research

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

A simheuristic approach for throughput maximization of

asynchronous buffered stochastic mixed-model assembly lines
Thiago Cantos Lopes a, Adalberto Sato Michels a, Ricardo Lüders a, Leandro Magatão a,∗
Graduate Program in Electrical and Computer Engineering (CPGEI), Federal University of Technology - Paraná (UTFPR), Curitiba 80230-901, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Mixed-model assembly lines are large scale production layouts that often operate under uncertainties
Received 16 May 2019 such as stochastic product sequences. Balancing such lines can be particularly challenging as throughput
Revised 28 October 2019
estimation can be difficult to determine, especially when asynchronous pace and buffers are considered.
Accepted 1 December 2019
Recent works have addressed problem variants with a given target throughput, but few authors consider
Available online 2 December 2019
a variant of the throughput maximization of mixed-model assembly line balancing problem. This paper
Keywords: addresses the balancing optimization problem for an assembly line with a given number of workstations
Mixed-model assembly line balancing and buffers between them. A make-to-order environment is considered, modeled as stochastic sequence
Unpaced asynchronous lines of products with known demand rates. A novel specialized cycle time simulator (CTS) is introduced, as
Stochastic product sequences well as a simheuristic approach (PSH) that exploits CTS to assess the cycle time of an assembly line and
Simulation provide good balancing solutions. The proposed simheuristic PSH is applied to a dataset with several
Simheuristic
buffer layouts, and its solutions are then compared to those of literature benchmarks. Performance com-
Bowl phenomenon
parisons show that PSH’s solutions outperform the benchmarks’ ones, with statistically significant dif-
ferences. Furthermore, the solution quality difference was greater for instances with more buffers, high-
lighting PSH capacity to conveniently exploit buffers in assembly lines. Lastly, analyses on the average
processing times of stations, obtained for each buffer layout, partially verifies and question established
results of the “bowl phenomenon” on unpaced assembly lines.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction In mixed-model assembly lines, it is assumed that products

can be mixed without setup times. However, differences in the
Assembly Line Balancing (ALB) problems consist of distributing processing times of tasks for different models can make the effec-
operations (tasks) to workers or robots (stations) in a manner that tive cycle time (or throughput) difficult to compute. The effective
maximizes a given performance measure under a set of practical cycle time is not only a function of line balancing, but also of
constraints. Assembly line balancing is key to efficient design and assembly line’s physical characteristics and of how products are
operation of production systems (Scholl, 1999). The simplest ALB sequenced. This paper assumes that the assembly line is reliable
problem which is widely studied in the literature is the Simple and has an asynchronous line control with a given number of
Assembly Line Balancing Problem (SALBP) whose assembly line is stations and buffers between them. An infinite input buffer before
assumed to be composed of simple serial stations producing a sin- the first station and an infinite output buffer after the last one are
gle product model. Furthermore, it is assumed that operations have also assumed. In asynchronous lines, a workpiece is allowed to
given deterministic processing times and precedence relations, and independently and discretely move to the following station when
that the sum of processing times in each station is bounded by processing is completed at the current station and the next station
the cycle time (either variable or parameter) which determines the is empty. Mixed-model sequencing is often treated as a separate
line pace (Baybars, 1986). While several variants of these problems optimization problem (Boysen et al., 2009). However, some au-
have been studied, this paper is focused on the Mixed-Model As- thors also focus on simultaneous balancing-sequencing problems
sembly Lines Balancing (MMALB), i.e. lines that are shared by a set (Lopes et al., 2018a; Sawik, 2012) or consider product sequences
of similar product models. as parameters (Biele and Mönch, 2018; Lopes et al., 2018b). In this
paper, a make-to-order context is considered as usually defined in
the assembly line balancing literature (Bukchin et al., 2002; Tiacci,
∗
Corresponding author. 2015b; Venkatesh and Dabade, 2008). The model of the next work-
E-mail address: [email protected] (L. Magatão). piece to enter the line is a random variable and the probability of

https://doi.org/10.1016/j.cor.2019.104863
0305-0548/© 2019 Elsevier Ltd. All rights reserved.
2 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863

being an instance of a specific model matches the model’s demand affect such throughput calculation, there are four main approaches
rate. This means that even if processing times are considered commonly used in the literature: worst case, best case, indirect
deterministic for each task and product model, there is a stochas- evaluation, and direct evaluation.
tic sequence of product models to be produced that makes the The most widely used analysis is the worst case one, in which
throughput of the assembly line hard to predict. processing times are controlled such that a given cycle time is
A series of recent publications (Tiacci, 2012; 2015a; 2015b; valid for all product models regardless of product sequence and
2017; Tiacci and Mimmi, 2018) have addressed variants of this line control: Reference (Karabati and Sayin, 2003) employs that
problem with simulation-based evaluation functions. However, concept for synchronous lines, references (Akpinar and Baykasoglu,
the throughput maximization variant of this problem (given the 2014a; 2014b) apply it for a variant with set-ups between tasks,
number of workstations and buffers) has not been addressed by reference (Roshani and Nezami, 2017) employs it for multi-manned
these works. To the best of the authors’ knowledge, few works mixed-model lines, reference (Dong et al., 2018) presents a chance-
have addressed this issue. In references (Bukchin et al., 2002; constrained version of this rationale for problems with stochastic
Venkatesh and Dabade, 2008), goal functions are described and task times, references (Kucukkoc and Zhang, 2015; 2016) use it for
surrogates discussed such for horizontal balancing (Becker and problems with parallel assembly lines.
Scholl, 2006). However, reference (Tiacci, 2015b) points out that Some authors consider the best case scenario (usually associ-
these are not objectives in themselves, but rather supposed means ated with large buffer availability), in which effective cycle time
to achieve a high and stable throughput. A different body of liter- is approximated by the highest station-wise weighted average of
ature (McNamara et al., 2016) discusses the “bowl phenomenon” processing times: Reference (Sawik, 2002) uses it as the first stage
in unpaced lines. It states that adequately unbalanced lines might of a decomposition procedure in a balancing-sequencing problem,
outperform balanced ones as a result which has been shown to be reference (Alghazi and Kurz, 2018) employs it in a problem variant
robust (Hillier and So, 1996). Even though the bowl phenomenon incorporating parallel workstations and ergonomic constraints,
literature presupposes stochastic components other than ones of reference (Merengo et al., 1999) presents a vertical balancing for-
the make-to-order context studied here, some of the obtained mulation that incorporates that core idea. Another very common
results present similar findings of the literature. approach is the use of indirect goal functions, in particular work-
This paper addresses the throughput maximization (or cycle load smoothing (Decker, 1993; Hamzadayi and Yildiz, 2012; 2013;
time minimization) of possibly buffered mixed-model assembly Özcan et al., 2010) and deviations from an ideal average (Battini
lines with stochastic product sequences. The main contribution is et al., 2009; Kim et al., 20 06; 20 0 0). The most accurate, and most
to propose a new simheuristic (Juan et al., 2015) procedure (PSH) problem-specific, is to explicitly take sequencing and line features
as optimization method that utilizes a new cycle time simulator into account to provide a direct measure of throughput: References
(CTS) to evaluate solution quality. CTS is a specialized simulator (Öztürk et al., 2015; 2013; Sawik, 2012) combine balancing and
that runs faster than other benchmark simulators. It allows PSH to sequencing to minimize the makespan of a product set, reference
find good solutions within practical runtime. Results are compared (Biele and Mönch, 2018) present a cost-oriented optimization prob-
to benchmarks of the literature, and to what is expected from the lem in which specific product sequences are given, recent works
bowl phenomenon. have also aimed at stable cycle time minimizations on the context
This paper is structured as follows. Section 2 presents an of given cyclical product sequences (Lopes et al., 2018b) and
overview of related works. Section 3 presents a formal definition of integrated balancing with cyclical sequencing (Lopes et al., 2019;
the problem and notation. Section 4 presents the new cycle time 2018a).
simulator CTS designed to simulate effective cycle time of mixed- In the specific case of asynchronous lines with a make-to-
model assembly lines. Section 5 presents the simheuristic method order context (Bukchin et al., 2002), in which product sequences
PSH developed to employ CTS and maximize the throughput of are stochastic, typical indirect performance measures have been
mixed-model assembly lines. Section 6 presents the undertaken mostly replaced by simulation-based direct evaluation procedures:
computational experiments, and Section 7 summarizes the papers’ A recent series of works have applied these methods to lines with
findings and contributions. parallel stations (Tiacci, 2015a), buffer allocation (Tiacci, 2015b), U-
shaped lines (Tiacci, 2017), and ergonomic constraints (Tiacci and
2. Related works Mimmi, 2018). These works, however, employ a cost-oriented
framework with a target throughput, and do not aim directly at
Readers can refer to review papers (Battaïa and Dolgui, 2013; an effective cycle time minimization (throughput maximization).
Becker and Scholl, 2006; Boysen et al., 20 07; 20 08) for more This paper, therefore, presents an optimization method to bridge
details on basic elements of assembly line balancing. Some spe- this specific gap: minimizing the effective cycle time of (possibly
cific topics of particular interest for this paper are: reductions buffered) asynchronous mixed-model lines on a make-to-order
from mixed-model to single model (Becker and Scholl, 2006) and context and analyze the impact of buffer allocation on efficient
common approaches tied to mixed-model lines; classifications and balancing solutions.
discussions of line features (Boysen et al., 20 07; 20 08) such as In this paper, simheuristics are chosen as the method to achieve
line control and buffers; an overview of problems, common goal such simulation-based optimization (Juan et al., 2015). They are
functions and solution approaches (Battaïa and Dolgui, 2013). a rather recent research line with a diverse range of applications
Mixed-model assembly lines often employ a conveyor with (Gonzalez-Neira et al., 2017; Guimarans et al., 2018; Hatami et al.,
given constant speed and an evenly spaced product launch dis- 2018) that go beyond using a simulation as a black-box evaluation
cipline. In that case, throughput is stable and the optimization function of a meta-heuristic. Simheuristics use problem-specific in-
problem consists of providing balancing and sequencing solutions formation to closely integrate optimization and simulation, using
that minimize the required number of stations and work overload information on feasibility and solution quality prior to simulation
(Boysen et al., 2009). However, for some types of line, the conveyor runs (Juan et al., 2015).
does not move with constant speed; such is the case in unpaced
synchronous and asynchronous lines. For these lines, or when the 3. Problem statement
optimization problem targets throughput maximization, determin-
ing throughput (or its inverse, the effective cycle time) is chal- The assembly line consists on series of stations and buffers,
lenging. While product sequencing and line characteristic factors represented by the set of integers S = 1, 2, . . . , |S|. Workstations
 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863 3

are equally manned and equipped so that any task can be assigned Algorithm 1 CycleTimeSimulator(sol, nK, nP).
to any station. Buffers are represented by the set SB , a sub-set of S,
1: T xM,S ← ProcessingTimesFromSolution(sol)
and no task can be assigned to them. The line produces a set M of
2: sumCT = 0
different product models, each model m of which has a given de-
3: for k = 1, k ≤ nK; k + + do
mand rate Om . Product models share a set of tasks T, and each task
4: T f ree1...nS ← 0
t has a deterministic duration Dt,m for each product model m. For
5: for p = 0; p ≤ nP ; p + + do
specialization purposes, tasks must be assigned to the same sta-
6: m ← NextModel()
tion for all product models and have a set of precedence relations
7: T f ree1 ← max(T f ree1 + T xm,1 , T f ree2 )
R. Each task must be assigned to a single station, and the process-
8: for s = 2; s < |S|; s + + do
ing time Txm,s at each station s for each product model m is the
9: T f rees ← max(T f rees−1 + T xm,s , T f rees+1 )
sum of processing times of the tasks assigned to the station.
10: end for
Product flow is asynchronous: each workpiece can move for-
11: T f ree|S| ← T f ree|S|−1 + T xm,|S|
ward discretely when processing at the current station is com-
12: if p = 0 then
pleted and the next station (or buffer) is empty. Starvations occur
13: F irst ← T f ree|S|
when a station is empty and the previous one has not completed
14: end if
processing of the next piece while blockages occur when a piece
15: end for
has completed processing and the following station is occupied by
16: sumC T ← sumC T + (T f ree|S| − F irst )/nP
a product. By hypothesis, an infinite input buffer exists before the
17: end for
first station and an infinite output buffer exists after the last one,
18: return sumCT /nK
meaning the first station is never starved and the last one is never
blocked. Product sequence is stochastic, i.e. the model of the next
product to enter the line is a random variable dictated with prob- Algorithm 2 NextModel().
abilities equal to the demand rates Om of each model m. 1: ran ← Random()
The optimization problem consists of defining a task-station 2: sum ← 0
assignment that respects precedence relations (feasible) and that 3: for m = 1; m < nM; m + + do // nM: number of product mod-
maximizes the expected throughput (i.e. minimizes the effective els
cycle time). While precedence relations can be quickly checked, 4: sum ← sum + Om
throughput must be simulated in order to be measured. 5: if sum ≥ ran then
6: return m
7: end if
4. Cycle time simulator (CTS) 8: end for
9: return nM
Consider the first workpiece p to enter a straight asynchronous
assembly line without parallel stations: its schedule will not de-
pend on subsequent workpieces. Furthermore, departure times CTS is designed to simulate assembly lines with stochastic
from each station are the only information needed to describe product sequence and deterministic processing times. However, it
blockages or starvations tied to the second piece. For instance, if can be easily adapted to incorporate deterministic sequences and
the first piece leaves a station after the second piece is ready to stochastic processing times. For instance, the processing time Txm,s
depart the previous one, the second piece will be blocked; on the at lines 6, 8, and 10 of Algorithm 1 is set as a parameter, but it
other hand, if the first piece leaves a station before the second one can be easily replaced by a random variable. Furthermore, a warm-
can enter it, that station will be starved. This behavior can be gen- up period can be implemented at line 11 by considering the first
eralized by knowing the departure times of workpiece p from each (valid) workpiece as being different from zero, i.e., a warm-up pe-
station. These departure times are enough to describe the depar- riod is measured in number of pieces rather than time units.
ture times (scheduling) of workpiece ( p + 1 ). An illustrative example of the proposed cycle time simulator
Algorithm 1 synthesizes the proposed cycle time simulator (CTS) for an assembly line with 2 product models and 5 stations
CTS. Given a balancing solution sol, nK replications of simulations is presented by Table 1. Demand rates are set to 50% of each prod-
with nP workpieces are conducted. TxM,S stores the processing uct model with production of 10 workpieces (nP = 10), and the
times of each model m ∈ M at each station s ∈ S. The subroutine number of simulation replications is set to one (nK = 1). Processing
NextModel() sets the next product model to enter the assembly times for each model m at each station s are presented on the left
line either randomly (based on product demand rates, as pre- side of Table 1. Columns of the right side of Table 1 presents values
sented by Algorithm 2) or according to a given fixed sequence. of Tfrees for each station when each product leaves the station. In
The time each station s is available to receive the next workpiece this short simulation, the average cycle time (Algorithm 1, lines 15
is stored in Tfrees . Initially, these values are set to zero. Each and 17) is (86 − 20 )/10 = 6.6 time units per product. Notice that
simulated workpiece i updates these values one station at a time: product workpiece 0 functions as a warm-up for the simulation, as
at each station s, these values are updated by considering possible described in Algorithm 1 (Lines 12–14).
starvations by station (s − 1 ) and blockages by station (s + 1 ).
Naturally, the update rules are slightly different for the first and 4.1. Comparison to literature benchmark
last stations (lines 6 and 10 in Algorithm 1, respectively). The first
station is never starved and the last station is never blocked. The Cycle Time Simulator (CTS) presented by Algorithm 1 is
The NextModel() function is presented by Algorithm 2. In hereby compared to the most recent benchmark Assembly Line
line 1, it generates random number generator, with uniform dis- Simulator (ALS) presented by reference (Tiacci, 2012). CTS is more
tribution between 0 and 1, which is used to determine the next specialized than ALS, which can incorporate other features such as
product to enter the line given the demand rates Om . It is assumed parallel workstations and U-shaped lines. However, this section re-

that m Om = 1, meaning the demand rates add up to 100%. This stricts the comparison to the specific capabilities of CTS: namely
function is employed by the proposed Cycle Time Simulator (CTS) straight lines without parallel workstations. There are two vari-
in the case of stochastic product sequences. able dimensions that both CTS and ALS can process: stochastic
4 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863

Table 1
Illustrative example of CTS execution.

Processing (TxM,S ) Tfrees for each station s after each product p

Model m Product p 0 1 2 3 4 5 6 7 8 9 10

Station s 1 2 Model m 1 2 1 1 1 1 2 2 1 1 2

1 7 3 s=1 7 10 17 24 31 38 41 45 50 57 64
2 2 4 s=2 9 14 19 26 33 40 45 50 57 64 69
3 2 5 s=3 11 19 26 31 36 42 50 57 64 69 74
4 4 7 s=4 15 26 31 36 41 46 57 64 69 74 81
5 5 5 s=5 20 31 36 41 46 51 62 69 74 79 86

Table 2
Algorithm 3 Simheuristic.
Average cycle time (CT) in time units and simulation runtime in seconds for ALS
and CTS. 1: while TimeLimitNotReached do
2: best ← GenerateInitialSolution(|T |, |S|)
ALS CTS
3: ini ← null
Processing Sequencing CT Runtime CT Runtime
4: Zinc ← ∞
Deterministic Deterministic 1605.7 2.66 1605.7 0.26 5: Zbest ← ∞
Deterministic Stochastic 1592.6 2.99 1592.4 0.29 6: while ini = best do
Stochastic Deterministic 1617.5 8.08 1617.3 0.80
7: ini ← best
Stochastic Stochastic 1617.3 8.01 1618.5 0.85
8: for all s1 , s2 ∈ S \ SB such that s1 is adjacent to s2 do
9: sol ← ini
10: UnassignTasks(sol, s1 , s2 , Ltasks) // Ltasks: list of
vs. deterministic processing times, and stochastic vs. deterministic
unassigned tasks
product sequences. ALS is made available for download by its au-
11: IterativeSearch(sol, best, Zbest, s1 , s2 , Ltasks, 1)
thor at http://doi.org/10.13140/RG.2.2.16680.72961. The data used
12: end for
in these simulation comparisons originates from a practical case
13: end while
study (Lopes et al., 2018b), and is made available in this paper’s
14: LocalMinima.Add (best ) // LocalMinima: list of local minimal
supplementary material. A very long simulation length (2.5 × 106
solutions
workpieces/loads) was used to reduce the impact of the transient
15: Zsol ← CycleTimeSimulator(best, nKlarge , nPlarge ) // nKlarge ,
behavior and provide the total simulation time in seconds. Table 2
nPlarge : external parameters
presents the results.
16: if Z sol < Z inc then
According to Table 2, the computed cycle times for both sim-
17: Z inc ← Z sol
ulators are very similar. Such small differences (less than 0.1%)
18: Incumbent ← best
are easily attributable to the sampling process using random vari-
19: end if
able generators. However, runtimes are quite different. CTS is ap-
20: end while
proximately ten times faster than ALS. This is expected as CTS
21: return (Incumbent, Zinc)
is more specialized than ALS. However, the development of such
fast specialized simulator is key to the simheuristic presented in
Section 5 as simulation is extensively used.
tasks with high priority and whose predecessors were already as-
5. The proposed simheuristic (PSH) signed are selected first. The task i to be selected is then assigned
to the station 1 + (i − 1 ) · |S|/|T | th in SࢨSB . For instance, if there
In order to exploit CTS presented in Section 4, the following are 20 tasks and five stations, the first four task selected will be
simheuristic procedure is proposed. It is based on Iterated Local assigned to the first non-buffer station, the next four tasks to the
Search (ILS) (Lourenço et al., 2003), which starts with randomized second non-buffer station, and so on. At Lines 8–12, Algorithm 3 se-
feasible solutions and performs local searches until a local min- lects pairs of adjacent non-buffer stations to remove all tasks and
imum is found. Local minima are then stored in a solution list test all possible re-assignments of such tasks. This makes feasibil-
that prevents duplicate solutions, effectively acting as a Tabu list ity verification easier, as only precedence between tasks assigned
(Glover, 1986). to the station pair (s1 , s2 ) must be checked. At Line 15, CTS
The proposed simheuristic assumes a given number of stations is used to measure the performance of the most recent local
and buffers between them. Buffers are modeled as dummy sta- minimum, which is saved in the LocalMinima list (Line 14) and
tions with no task assigned to them. The neighborhood search is stored as incumbent if appropriate (Lines 16 to 19). Notice that
based on repeated two-station optimization: tasks are fixed on all the CTS takes into account buffers between stations, meaning that
but two stations, and all relevant ways to reassign tasks within solution performance will be dependent on the buffer layout. The
that station pair are tested for all adjacent station pairs. Block- IterativeSearch() subroutine (Line 11) is further explained
ages and starvations happen between adjacent pairs. Therefore, by Algorithm 4.
it is expected that neighborhoods defined in terms of adjacent The IterativeSearch() procedure is the local search as-
pairs of stations will contain solutions with different amounts of pect of the proposed simheuristic. Tasks unassigned from stations
blockages and starvations. This means these neighborhoods will s1 and s2 are stored in list LTasks. Algorithm 4 begins with i = 1,
lead to incremental improvements as required by ILS algorithms taking the first task of the list and attempting to re-assign it to
(Lourenço et al., 2003). each station. Lines 2–14 present such attempt for s1 and Lines 15–
Algorithm 3 presents the high level operations of the proposed 27 present the analogous attempt for s2 . The Assignable() test
simheuristic. The initial solution at Line 2 is provided by a simple at Line 2 serves a dual feasibility-optimality role: on one hand it
randomized priority rule: tasks are given random priorities and re- tests whether the assignment of task t to s1 violates a precedence
ordered accordingly. They are then assigned one at time such that relation in regard to a task assigned to s2 (feasibility); on the
 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863 5

Algorithm 4 IterativeSearch(sol, best, Zbest, s1 , s2 , Ltasks, i). randomized re-starts. The Local Minima restriction prevents the
method from converging to previous solutions and, therefore, en-
1: t ← Ltasks(i )
hances solution diversity.
2: if Assignable(sol, t, s1 , Zbest) then
Notice that, while both Algorithms 3 and 4 use the proposed
3: Assign(t, s1 , sol)
CTS (Algorithm 1), the parameters change. During the local search,
4: if i = Ltasks.Count then
simulations are shorter and, therefore, less accurate (for experi-
5: if not LocalMinima.Contains(sol ) then // LocalMinima:
ments, nKsmall , nPsmall were set to 1 and 10 0 0, respectively). While
external list of local minimal solutions
in the outer loop of the algorithm, simulations are longer (nKlarge ,
6: Zsol ← CycleTimeSimulator(sol, nKsmall , nPsmall ) //
nPlarge were set to 50 and 50,0 0 0, respectively). This is justified as
nKsmall , nPsmall : external parameters
it is not desirable to overload the simheuristic with long simula-
7: if Z sol < Z best then
tions. In order to allow the search procedure to converge, quick
8: best ← sol
approximate evaluations are used in the lower-level loops, while
9: end if
longer, more precise, and exhaustive simulations are reserved for
10: end if
the subset of promising solutions (Chica et al., 2017).
11: else IterativeSearch(sol, best, Zbest, s1 , s2 , Ltasks, i + 1)
Lastly, this procedure can be easily parallelized. Multiple lo-
12: end if
cal searches can be executed on different CPU cores with several
13: Unassign(t, s1 , sol)
threads. The only objects that must be shared between threads are
14: end if
the solution Incumbent, its performance value (Zinc), and the list
15: if Assignable(sol, t, s2 , Zbest) then
LocalMinima. The later can be implemented as a hash table to im-
16: Assign(t, s2 , sol)
prove the computational performance.
17: if i = Ltasks.Count then
18: if not LocalMinima.Contains(sol ) then
19: Zsol ← CycleTimeSimulator(sol, nKsmall , nPsmall ) 6. Computational experiments
20: if Z sol < Z best then
21: best ← sol The proposed simheuristic PSH was applied to an expanded
22: end if dataset (Lopes et al., 2018a; 2018b) with 2625 instances, in which
23: end if five buffer layouts were considered for each of the 525 task-
24: else IterativeSearch(sol, best, Zbest, s1 , s2 , Ltasks, i + 1) property data vectors, with five product models each. These in-
25: end if stances are based on the dataset provided in reference (Otto et al.,
26: Unassign(t, s2 , sol) 2013). Instance data is made available in the paper’s supple-
27: end if mentary material, as well as the solutions obtained by PSH and
the benchmarks. All tests were conducted with a Core i7-8700
(3.2 GHz, 12 CPUs) with 32GB RAM. The time limit was set ac-
cording to the number of tasks (3 s per task) in each instance for
other hand, it tests whether the sum of processing times of tasks each PSH execution using 12 available threads.
already assigned to s1 is too large if t is assigned to it (optimality The one-way ANOVA analyses have been computed with a con-
condition). The optimality condition of this test is based on the fidence level of 99% for comparing results obtained by simulation,
observation that each workstation is a potential line bottleneck indicating that there are statistically significant differences when-
(Becker and Scholl, 2006). Therefore, the expected assembly line ever p-value < 1%. Bonferroni corrections and Tukey’s post-hoc
performance (and, hence, cycle time) is bounded by the expected test were used for multiple comparisons.
performance of each station, which is trivially determined by Tests were separated into two categories: (i) small and (ii)
the weighted averages of processing times. Such comparison is medium and large instances. For the 875 small instances, PSH
dynamically strengthened during the iterative search as it is based is compared with four benchmarks. Its convergence behavior
on the performance of the current best neighbor (Zbest). This is analyzed and the influences of buffers on efficient solu-
allows substantial reductions on the number of simulations on tions are discussed in light of the bowl phenomenon literature
unpromising neighbors. By doing this, simulation and optimization (McNamara et al., 2016). For 1750 large instances, the performance
are more closely integrated (Juan et al., 2015): answers reported of PSH is compared with the best benchmark of item (i) in terms
by simulations are used to accelerate neighborhood searches of average cycle time (inverse of throughput).
by informing performance requirements, and problem-specific
information is used to reduce the number of simulations runs on
low-quality solutions. 6.1. Small instances
A logical test at Line 4 of Algorithm 4 verifies whether all tasks
were reassigned. If so, the current solution sol is tested as a mem- In total 875 instances were generated combining 175 task-
ber of LocalMinima list (Line 5). If it is a new solution, sol is tested property data vectors to 5 buffer layouts. These all have 20
for performance, and saved as the best solution of the current tasks, and the number of workstations was fixed to seven.
iteration (best), if appropriate (Lines 6–9). Back to Line 4, if not Section 6.1.1 presents the tested buffer layouts and the respective
all tasks were assigned, the search iterates by incrementing the results. Section 6.1.2 compares the solutions found by PSH with
depth i as stated at Line 11. Tasks are unassigned at Line 13 from four benchmarks of the literature. Section 6.2 presents the conver-
s1 so that they can be tested as possible assignments for s2 (Lines gence behavior of PSH.
15–27).
In that sense, the neighborhood of each solution tested in 6.1.1. Influence of buffer layout
Algorithm 3 is the set of all relevant (feasible, new, and promis- This section describes the observed impact of buffer layout
ing) two-station reassignments: they must be feasible regarding in light of the “bowl phenomenon” introduced by Hillier and
precedence relations, new in regard to the Local Minima List, and Boiling (1979). Several authors have shown that unpaced lines
promising in the sense that their performance lower bound must with stochastic processing times can have better performance
be better than the current solution (best). Until the time limit when adequately unbalanced. Readers can refer to a recent review
or a stopping criterion is reached, this procedure is iterated by (McNamara et al., 2016) for more information.
6 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863

Fig. 1. Tested buffer layouts. Darker squared represent workstations, lighter ones represent buffers.

Fig. 2. Weighted average processing time profiles for homogeneous buffer layouts.

In order to verify how buffers affect the processing time dis- times at the first and last stations. In other words, the second and
tribution among stations, five buffer layouts are tested. They are second-last station act as “partial buffers” for their heavy loaded
presented by Fig. 1. Buffers are naturally expected to enhance the neighbors. This difference in regard to standard bowl phenomenon
performance of asynchronous assembly lines even when balancing literature can be explained by the aforementioned differences in
solutions are kept unchanged (Boysen et al., 2008). The simulator the nature of the problem’s stochastic dimensions.
CTS presented in Section 4 considers buffers as dummy stations The fully buffered profile (Fig. 2b) present a much milder bowl
with zero processing times. Therefore, PSH takes the given buffer phenomenon, with processing time averages differing at most 2.5%
layout explicitly into consideration. compared to the 6.5% observed in the unbuffered layout. Indeed,
ANOVA and post-hoc tests confirm that there are differences in post-hoc tests Appendix A still show statistically significant differ-
processing time distributions for each buffer layout Appendix A. ences in processing time averages of the extreme stations (1 and 7)
In order to visualize the impact of buffer layout on balancing so- and their neighbors (2 and 6). This verifies a result from the bowl
lutions found by PSH, the results are averaged by the following phenomenon literature, as the ideal degree of imbalance decreases
procedure: (i) the weighted (by demand rates) average processing when increasing of internal storage (Hillier and Boiling, 1979). Fur-
time in each station is divided by sum of weighted average pro- thermore, this resembles the desired results of vertical balancing
cessing times in the whole line; (ii) such normalized values are (Merengo et al., 1999) by homogenizing the average workloads
averaged over 175 instances and used to determine the average across workstations.
normalized processing time profile for a given buffer layout, i.e., a Fig. 3 helps explaining the differences between the None and
perfectly balanced assembly line has average normalized process- Full layouts. For the three remaining layouts, processing times are
ing time of 100% in every station. Fig. 2 presents the profiles for concentrated around buffers: whether they are around the en-
fully buffered and unbuffered layouts. trance of the line (Fig. 3a), its middle (Fig. 3b), or its exit (Fig. 3c).
The unbuffered profile in Fig. 2a presents an interesting bowl Incumbent solutions concentrated, for each layout, processing
phenomenon variant: processing times are increased near the en- times around buffers. Indeed, post-hoc tests Appendix A verify
trance and exit stations of the assembly line. As expected, post-hoc statistically significant differences between key station pairs such
tests Appendix A show statistically significant differences between as stations 2 and 6 for both begin and end layouts, and stations
the processing time averages of the extreme stations (1 and 7) 4 and 6 for the middle layout. A possible explanation for such
and their neighbors (2 and 6). Interestingly, this happens despite phenomenon corroborates the standard understanding of why the
the differences in stochastic components between the studied bowl phenomenon exists (McNamara et al., 2016): buffers can
problems and the bowl phenomenon literature. However, the ob- compensate for differences in processing times between models,
served profile violates the “monotonicity” assumption (Hillier and therefore, buffered stations can afford to have longer processing
So, 1991) of the bowl phenomenon literature. The second and times without the same risk of starving and blocking adjacent
second-last stations have lower average processing times (around stations. The assumed infinite input and output buffers, before
97%) than the center ones (around 99%). Such “ripple effect” might the first station and after the last one, further explain the profile
be justified as a way to better exploit the high average processing observed in Fig. 2a. More importantly, the proposed simheuristic
 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863 7

Fig. 3. Weighted average processing time profiles for partially buffered layouts.

is shown to be able to properly take advantage of buffer layouts, station-wise weighted average of processing times, which can be
hinted by the differences in average solution profiles, which are seen as a best-case scenario optimization. Expression (3) mini-
further discussed in the following sections. mizes the highest processing time value, which can be seen as a
worst case scenario optimization. Expression (4) attempts to bal-
6.1.2. Comparison to alternative goal functions ance the minimization of variability in regard to the highest value
Due to the stochastic component of the problem, a direct of processing time and the minimization of processing time vari-
measure of effective cycle time is challenging. Consequently, sev- ability for each model. In it, Cm states the theoretical minimum
eral indirect approaches were developed to optimize throughput cycle time of model m, defined by the ceiling of the division of to-
indirectly. This is the case of four benchmarks considered here for tal task processing time by the number of workstations (Scholl and
comparison with PSH results: Horizontal Balancing (HBal), Ver- Becker, 2006).
tical Balancing (VBal), Maximum Sub-cycle Time (MST), and the The first three benchmarks (Hbal, VBal, and MST) are either
squared Base Model deviation plus smoothness Index (BMI). These linear or easily linearizable goal functions. Given the size of in-
formulations were presented by references (Decker, 1993; Karabati stances, it is possible to compare PSH solutions to the optimal
and Sayin, 2003; Merengo et al., 1999; Venkatesh and Dabade, solutions of these benchmarks that have linearizable goal func-
2008), respectively. The formal definitions of these measures are tions (Hillier and Lieberman, 2015). Therefore, they were con-
presented by the Expressions (1)–(4). verted into mixed-integer programming instances and solved with
   Gurobi 8.0 under the same hardware conditions PSH was executed.
 Dt,m BMI, on the other hand, cannot be straightforwardly linearized and
HBal: Minimize max max T xm,s − (1)
m∈M s∈S |S | t∈T
was instead tested using the simheuristic procedure described in
    Section 5 by replacing CTS by Expression (4) as the performance
   evaluation routine on Algorithm 4. The first three benchmarks pro-
VBal: Minimize max Om · T xm,s − Om · T xm,s1 duced MILP instances that were solved to optimality. The forth one
s∈S
s1 ∈S m∈M m∈M (BMI) was executed with a time limit of 3 s per task per instance,
(2) the same used by PSH.
MST: Minimize max T xm,s (3) The average cycle time obtained for each benchmark solution
m∈M, s∈S was compared with the cycle time obtained for PSH solution us-
 
  2 ing long simulations of CTS (50 replications with 50,0 0 0 products
BMI: Minimize Om · T xm,s − max T xm ,s each). The average cycle time improvements (avg) and standard de-
s ∈S, m ∈M viations (sdv) of PSH solutions compared to each benchmark for
m∈M s∈S
 different buffer layouts are presented in Table 3. It also presents
the results of a one-way ANOVA, which show that the performance
+ (T xm,s − Cm )2 (4)
differences between benchmarks are statistically significant for all
buffer layouts. Post-hoc Tukey’s HSD tests Appendix A confirmed
Expression (1) seeks to homogenize station-wise workload dis- (p-value < 0.1%) that the performance difference between PSH
tributions for each model. Expression (2) minimizes the highest and all benchmarks is statistically significant. Furthermore, these
8 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863

Table 3 Table 4
Cycle time improvement (%) of PSH solutions compared to each benchmark for dif- Cycle time improvement (%) of PSH and benchmark solutions for each buffer layout
ferent buffer layouts. compared with the respective unbuffered solutions.

Benchmark PSH MST Vbal Hbal BMI

Buffer Layout Rate MST Vbal Hbal BMI ANOVA Begin 4.6% 1.6% 3.6% 2.7% 2.7%
Middle 4.1% 2.9% 5.5% 3.4% 3.1%
None avg 3% 10.9% 4.6% 2% F-value = 306
End 4.6% 1.1% 2.2% 1.5% 1.0%
sdv 1.8% 4.9% 2.5% 1.8% p-value < 0.01%
Full 12.0% 3.5% 7.5% 4.9% 4.7%
Begin avg 6.3% 12.1% 6.7% 4.1% F-value = 166
sdv 2.8% 4.9% 3.3% 2.5% p-value < 0.01%
Middle avg 4.3% 9.3% 5.3% 4.1% F-value = 102 obtained better results than all benchmarks on 853 (97.5% of
sdv 2.5% 4.2% 3.1% 2.5% p-value < 0.01%
them), performing 3.7% better in average than the instance-wise
End avg 6.7% 13.7% 7.9% 6.5% F-value = 149
sdv 2.7% 5.4% 3.2% 2.8% p-value < 0.01% best out of the four benchmarks.
Full avg 13% 16.5% 13.1% 8.2% F-value = 131 Lastly, the impact of each buffer layout in each approach was
sdv 3.7% 4% 4% 4% p-value < 0.01% measured by dividing the performance of buffered solutions with
unbuffered ones. The cycle time improvement (%) of PSH and
benchmark solutions for each buffer layout compared with the re-
tests confirmed (p-value < 0.1%) that BMI outperforms the other spective unbuffered solutions are presented in Table 4. Notice that
benchmarks, hence justifying its use for comparisons in larger in- the proposed method was able to better exploit buffers than all
stances. benchmarks as the improvement rates for PSH were greater than
Notice that all benchmarks’ relative performances worsen when those of benchmarks. The single exception was VBal for middle
more buffers are added. This further highlights the proposed algo- layout buffer (improvement of 5.5% > 4.1%), but that can be at-
rithm’s better capacity to exploit buffers as resources. For the fully tributed to that benchmark’s poor performance on the unbuffered
buffered case, PSH had solutions with homogenized station-wise layout (Table 3). Furthermore, notice that for all benchmarks, the
averages (Fig. 2b). However, the poor performance of VBal shows “Middle” layout led to greater improvements than the “Begin”
that balancing average station-wise workloads is not sufficient to and “End” ones. This echoes established results on the bowl phe-
ensure good stable cycle time, even for the fully buffered case. The nomenon literature as (Hillier and So, 1991) states that buffers are
other cases (MST, VBal and BMI) show that homogenizing model- better assigned to the center of the assembly line. However, this
wise workload distribution can often be better than homogenizing was not observed for PSH: buffer allocations in the Begin and End
station-wise average workload distribution, but are also insuffi- of the line led to better performance than in the Middle (4.6% ver-
cient to guarantee a good performance, especially when buffers sus 4.1%). A one-way ANOVA confirms that there are statistically
are added to the line. significant differences between the performances of PSH for Begin,
Fig. 4 presents the distribution of cycle time improvement (%) Middle, and End layouts (F-value = 5.27, p-value = 0.54%). This
of PSH solutions compared to each benchmark for all buffer lay- can be explained by: (i) the differences in balancing solutions that
outs. Notice that, while PSH is outperformed by the benchmarks accumulate processing times around buffers (Fig. 3); (ii) the afore-
on a few instances, for the vast majority of them, it outperforms mentioned difference nature of the stochastic component of the
the benchmarks substantially. Out of 875 small instances, PSH problem.

Fig. 4. Distribution of cycle time improvement (%) of PSH solutions compared to each benchmark.
 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863 9

Fig. 5. PSH convergence behavior for each dataset.

6.2. Convergence behavior Fig. 5a compares the best benchmark’s (BMI) average perfor-
mance to the convergence behavior for small instances. This high-
During PSH computational experiments, the time limit was set lights that PSH quickly finds solutions that outperform the bench-
to 3 s per task, meaning 60, 150 and 300 s for small, medium mark’s average relative performance. Fig. 5b and c confirm that
and large instances, respectively. During the optimization proce- the 3 s per task time limit is reasonable compromise: Medium
dure, for each instance, all incumbents were saved along with the and large datasets display slower drops for all percentiles, meaning
elapsed time necessary to find them. For most instances, few solu- more instances took longer to converge, which suggests that larger
tions were found until the best incumbent. However, time required time limits could lead to further improvements. However, for half
to find them was different for each instance. In order to evaluate of the instances (50% curves), no improvement was found after half
the algorithm’s convergence behavior, the following steps were un- the time limit and, for most of those that did, the improvement
dertaken: The objective value of each incumbent was divided by its was lower than 2%. In the following section, PSH’s performance is
instance’s last incumbent’s one, defining a rate to the best solution. compared to BMI, i.e. the benchmark with best performance ob-
These rates define instance-wise convergence steps, which allow a served in Section 6.1.2.
convergence curve to be drawn for each percentile. Such curves are
presented by Fig. 5 with a logarithmic (base 2) time axis. They can 6.3. Medium and large instances
be interpreted by generalizing the following: the 10% curve indi-
cates that only 10% of the instance convergence curves are above Medium instances had 50 tasks and 13 workstations, while
it and 90% are below it. large ones had 100 tasks and 25 workstations. Buffer layouts were
10 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863

Table 5 is much faster than the simulation performed by CTS and, there-
Cycle time improvement (%) of PSH solutions compared with BMI and number of
fore, the number of neighborhoods explored using it is consistently
iterations for different buffer layouts.
larger than that of PSH. However, this does not necessarily reflect
Improvement Iterations in better solution quality, especially when buffers are added be-
Instances avg sdv PSH BMI tween stations. Fig. 6 presents the distribution of cycle time im-
provement (%) of PSH solutions compared to BMI.
Small Empty 2.0% 1.8% 2000 73,800
Begin 4.1% 2.5% 2850 Notice that Fig. 6a, similarly to Fig. 4d, reports few instances in
Middle 4.1% 2.5% 2450 which the difference in solution quality obtained by PSH and BMI
End 6.5% 2.8% 2350 favors BMI (difference smaller or equal than zero). However, as re-
Full 8.2% 4.0% 5750 ported by Fig. 6b, this was more common for larger instances. PSH
Medium Empty 2.4% 1.8% 1500 5100
Begin 4.7% 2.2% 1700
outperformed BMI in 870 instances, and 661 instances (out of 875)
Middle 4.4% 2.3% 1750 from the medium, and large dataset, respectively. Table 5 indicates
End 4.5% 2.1% 1700 that this difference is primarily due to the instances with ‘Empty’
Full 6.3% 3.2% 3100 and ‘Middle’ layouts, indicating that it is hard to outperform BMI
Large Empty 0.8% 1.1% 750 2700
for these assembly line layouts. These results suggest that, for ex-
Begin 1.6% 1.7% 700
Middle 1.2% 1.5% 800 tremely large instances (more than 100 tasks), neighborhood-based
End 1.8% 1.7% 680 searches are overwhelmed by the amount of neighboring solutions,
Full 8.3% 2.2% 1100 especially if time-consuming performance measures are required.
This would explain a reduced improvement of PSH over BMI by
increasing the instance sizes.
defined similarly to the small instances with partially buffered lay- Despite difficulties observed for larger instances, both in aver-
outs representing buffer allocations at the first, central and last age performance (Table 5) and number of instances (Fig. 6), PSH
third part of stations for Begin, Middle and End layouts, respec- outperformed BMI for all datasets. Furthermore, a one-way ANOVA
tively (4 buffers for medium instances and 8 buffers for large confirms that the difference in performance is statistically signifi-
ones). Given the results reported in Section 6.1.2, BMI was chosen cant for every instance size and buffer layout (F-value > 14, p-
as benchmark due to its better performance compared to other al- value < 0.02%). This difference was more substantial for fully
ternatives (HBal, VBal, MST). buffered cases, highlighting the capacity of CTS to explicitly take
A total of 1750 instances (875 medium and 875 large) were buffer information into account.
tested with the same hardware and time limit used for small in-
stances. For each instance, the cycle time improvement of PSH 6.4. A case-study on the impact of demand rates
compared with BMI was measured as a percentage. Furthermore,
the number of local minima (number of iterations) found by both In order to evaluate the proposed simheuristic approach re-
PSH and BMI were also saved. Table 5 presents a comparison be- garding different demand rates, an instance (small dataset, num-
tween PSH and BMI for medium and large instances with different ber 1, available in this article’s supplementary material) was solved
buffer layouts. For convenience, results for small instances (from by PSH with two demand scenarios: (i) Equal Demands (ED), in
Table 3) are also included in Table 5. which all five models have 20% demand rate, and (ii) Dominant
One of the noteworthy results reported by Table 5 is that more Model (DM) scenario, in which product model 1 has a 90% de-
neighborhoods are explored for the ‘Full’ layout of each dataset. mand rate, and all others have 2.5%. This leads to two different
This reflects the optimality nature of the valid reassignment con- solutions, whose processing times for each product model (M1 to
cept described in Section 5. Only reassignments that are not triv- M5) at each station are presented by Fig. 7.
ially dominated by the incumbent are simulated. Given that more Fig. 7 highlights product model 1 for both scenarios using a
buffers tend to lead to better cycle times, it is only natural that darker blue color and solid fill. Notice that its processing times are
better incumbents allow a larger portion of neighborhoods to be much better balanced in the DM scenario (Fig. 7b), at the cost of
dominated. Another expected result reported by Table 5 is that higher variance of other product models. This happens in the DM
increasing the instance size leads to fewer iterations. This is ex- scenario because model 1 is much more common than the other
pected, as neighborhoods of each solution tend to increase signif- models. Thus, PSH balances the processing time of model 1 across
icantly in size. The computation of the BMI performance measure stations regardless of the impact on other product models.

Fig. 6. Distribution of cycle time improvement (%) of PSH solutions compared to BMI.
 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863 11

Fig. 7. Processing time of solutions obtained for each demand scenario.

Table 6 to the proposed simheuristic PSH. The later produced better so-
Realized cycle time for each solution at each scenario.
lutions for 90.8% of instances, outperforming the best benchmark
Solution from Scenario: Performance in average by 5.0%, 4.2%, and 2.3% for small, medium, and large
Tested Scenario: Equal Demands Dominant Model difference
instances, respectively. Lastly, a case study has illustrated the rele-
vance of demand rates in solution quality, as well as PSH capacity
Equal Demands 811.5 852.3 4.8%
to adequately take them into account.
Dominant Model 914.5 794.7 13%
Further works should also focus on producing exact perfor-
mance evaluations and analytical lower bounds for the consid-
ered problem. This is particularly challenging due to the stochas-
This is corroborated by applying the solution obtained in one
tic product sequences. However, overcoming such challenge would
scenario to the other and comparing the realized cycle times, as
allow an absolute measurement of solution quality rather than a
depicted by Table 6. Notice that applying the balancing solution
relative one: lower bounds could partially replace the use of al-
obtained for one demand scenario to the other leads to worse per-
ternative formulations as benchmarks. Furthermore, an exact per-
formance. This case study confirms that solution quality is sensible
formance evaluation could allow exact solution procedures for this
to the demand rates, and illustrates PSH’s capacity to take them
stochastic problem.
into account.
Acknowledgments
7. Conclusions
This study was funded in part by the Coordenação de Aper-
This paper deals with a variant of the throughput maximization feiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance
of mixed-model assembly line balancing problem. It presupposes Code 001. The authors also thank the financial support from Fun-
a given number of stations, asynchronous line pace, number of dação Araucária (Agreement 041/2017 FA – UTFPR – RENAULT), and
buffers between stations, and deterministic processing times. How- CNPq (Grants 406507/2016-3 and 307211/2017-7). The authors also
ever, the product sequence is considered stochastic in a make-to- acknowledge Celso Gustavo Stall Sikora for his constructive criti-
order environment as defined in the assembly line balancing litera- cism and suggestions.
ture. A specialized simulator CTS is proposed to measure the solu-
tion quality as well as a simheuristic PSH that finds good solutions Appendix A. ANOVA analysis
using CTS as a performance measure.
CTS outperforms a literature object and event-oriented bench- In Section 6.1.1, the influence of buffer layout in processing time
mark simulator by an order of magnitude as, expected due to its distributions is discussed. The detailed results of the ANOVA post-
higher degree of specialization. However, due to fast simulation, it hoc test is presented by Table A.7. Entries of Table A.7 above the
allows a more extensive use of simulations by PSH. Even not repre- diagonal are the Least Significant Difference (LSD) T-test statistics,
senting features as U-shaped assembly lines and parallel stations, while the ones below the diagonal are the Bonferroni-corrected p-
CTS can easily be adapted to include features such as stochas- values. All p-values lower than 1% (confidence level of 99%) were
tic processing times and deterministic product sequence. Further considered statistically significant differences and are highlighted
works can also extend the proposed CTS to include other assembly in boldface.
line features. In Section 6.1.2, PSH and the benchmarks are compared
Computational experiments with PSH have shown that buffers (Table 3). By combining all layouts and all methods (including
affect task distributions. They allow processing times to be more PSH), a one way ANOVA indicates that there are statistically signif-
substantially concentrated around buffers. Interestingly, this lead icant differences between them (Table A.8). The post-hoc multiple
to a bowl phenomenon variant which partially verifies established comparisons are presented by Table A.8. The Tukey Honest Signif-
results and concepts of the bowl phenomenon literature. Further- icant Difference statistic (HSD) and corresponding p-values above
more, surrogate benchmark goal formulations have been compared and below the diagonal, respectively, for each comparison.
12 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863

Table A.7
Post-hoc test: LSD T-test statistic (above diagonal) and corresponding p-values (below diagonal) for multiple comparisons.

None Layout - F-test: F-value = 11.9, p-value < 0.01%

Station 1 2 3 4 5 6 7

1 - 7.05 4.33 4.33 4.24 5.89 1.38

2 < 0.1% - 2.72 2.71 2.8 1.15 5.66
3 < 0.1% 13.1% - 0 0.09 1.56 2.95
4 < 0.1% 13.2% 100% - 0.09 1.56 2.95
5 < 0.1% 10.2% 100% 100% - 1.65 2.86
6 < 0.1% 99.8% 92.9% 93% 88.8% - 4.51
7 97.8% < 0.1% 6.7% 6.6% 8.7% < 0.01% -
Begin Layout - F-test: F-value = 111.7, p-value < 0.01%
Station 1 2 3 4 5 6 7
1 - 1.32 7.84 16.4 14.01 17.02 10.51
2 98.7% - 9.16 17.72 15.33 18.34 11.83
3 < 0.1% < 0.1% - 8.56 6.17 9.18 2.67
4 < 0.1% < 0.1% < 0.1% - 2.39 0.62 5.89
5 < 0.1% < 0.1% < 0.1% 30.4% - 3.01 3.5
6 < 0.1% < 0.1% < 0.1% 100% 5.5% - 6.51
7 < 0.1% < 0.1% 14.9% < 0.1% 1% < 0.1% -
Middle Layout - F-test: F-value = 100.2, p-value < 0.01%
Station 1 2 3 4 5 6 7
1 - 10.29 0.57 10.06 1.37 10.16 0.96
2 < 0.1% - 10.86 20.35 11.66 0.14 9.33
3 100% < 0.1% - 9.49 0.81 10.72 1.53
4 < 0.1% < 0.1% < 0.1% - 8.69 20.21 11.02
5 98% < 0.1% 100% < 0.1% - 11.53 2.34
6 < 0.1% 100% < 0.1% < 0.1% < 0.1% - 9.19
7 100% < 0.1% 94.1% < 0.1% 33.9% < 0.1% -
End Layout - F-test: F-value = 138.4, p-value < 0.01%
Station 1 2 3 4 5 6 7
1 - 10.6 7.35 5.37 1.72 10.61 10.23
2 < 0.1% - 3.25 5.23 12.31 21.21 20.83
3 < 0.1% 2.5% - 1.98 9.07 17.96 17.58
4 < 0.1% < 0.1% 64.3% - 7.09 15.98 15.6
5 85.1% < 0.1% < 0.1% < 0.1% - 8.89 8.51
6 < 0.1% < 0.1% < 0.1% < 0.1% < 0.1% - 0.38
7 < 0.1% < 0.1% < 0.1% < 0.1% < 0.1% 100% -
Full Layout - F-test: F-value = 7.6, p-value < 0.01%
Station 1 2 3 4 5 6 7
1 - 3.54 1.52 3.4 0.07 3.48 1.1
2 0.87% - 2.02 0.14 3.47 0.06 4.64
3 94.5% 60.7% - 1.89 1.45 1.96 2.61
4 1.4% 100% 72.4% - 3.34 0.07 4.5
5 100% 1.1% 96.4% 1.8% - 3.41 1.16
6 1.1% 100% 66.4% 100% 1.4% - 4.57
7 99.9% < 0.1% 17.4% < 0.1% 99.7% < 0.1% -

Table A.8 Alghazi, A., Kurz, M.E., 2018. Mixed model line balancing with parallel stations,
Post-hoc test: Tukey HSD-test statistic (above diagonal) and corresponding p-values zoning constraints, and ergonomics. Constraints 23, 123–153. doi:10.1007/
(below diagonal) for multiple comparisons. s10601-017-9279-9.
Battaïa, O., Dolgui, A., 2013. A taxonomy of line balancing problems and their solu-
F-test: F-value = 1121, p-value < 0.01% tion approaches. Int. J. Prod. Econ. 142 (2), 259–277. doi:10.1016/j.ijpe.2012.10.
020.
Method PSH MST VBal HBal BMI
Battini, D., Faccio, M., Persona, A., Sgarbossa, F., 2009. Balancing - sequencing pro-
PSH - 49.5 92.8 55.9 37.1 cedure for a mixed model assembly system in case of finite buffer capacity. Int.
MST < 0.1% - 43.3 6.4 12.4 J. Adv. Manuf. Technol. 44 (3–4), 345–359. doi:10.10 07/s0 0170-0 08-1823-8.
VBal < 0.1% < 0.1% - 36.9 55.7 Baybars, I., 1986. Survey of exact algorithms for the simple assembly line balancing
HBal < 0.1% < 0.1% < 0.1% - 18.8 problem. Manag. Sci. 32 (8), 909–932. doi:10.1287/mnsc.32.8.909.
Becker, C., Scholl, A., 2006. A survey on problems and methods in generalized as-
BMI < 0.1% < 0.1% < 0.1% < 0.1% -
sembly line balancing. Eur. J. Oper. Res. 168 (3), 694–715. doi:10.1016/j.ejor.
2004.07.023.
Biele, A., Mönch, L., 2018. Hybrid approaches to optimize mixed-model assembly
lines in low-volume manufacturing. J. Heuristics 24 (1), 49–81. doi:10.1007/
Supplementary material s10732- 017- 9357- 6.
Boysen, N., Fliedner, M., Scholl, A., 2007. A classification of assembly line balancing
Supplementary material associated with this article can be problems. Eur. J. Oper. Res. 183 (2), 674–693. doi:10.1016/j.ejor.2006.10.010.
Boysen, N., Fliedner, M., Scholl, A., 2008. Assembly line balancing: which model to
found, in the online version, at 10.1016/j.cor.2019.104863
use when? Int. J. Prod. Econ. 111 (2), 509–528. doi:10.1016/j.ijpe.2007.02.026.
Boysen, N., Fliedner, M., Scholl, A., 2009. Sequencing mixed-model assembly lines:
References Survey, classification and model critique. Eur. J. Oper. Res. 192 (2), 349–373.
doi:10.1016/j.ejor.2007.09.013.
Akpinar, S., Baykasoglu, A., 2014. Modeling and solving mixed-model assembly line Bukchin, J., Dar-el, E.M., Rubinovitz, J., 2002. Mixed model assembly line design
balancing problem with setups. Part I: a mixed integer linear programming in a make-to-order environment. Comput. Ind. Eng. 41, 405–421. doi:10.1016/
model. J. Manuf. Syst. 33, 177–187. doi:10.1016/j.jmsy.2013.11.004. S0360-8352(01)0 0 065-1.
Akpinar, S., Baykasoglu, A., 2014. Modeling and solving mixed-model assembly line Chica, M., Juan Pérez, A.A., Cordon, O., Kelton, D., 2017. Why simheuristics? Benefits,
balancing problem with setups. Part II: a multiple colony hybrid bees algorithm. limitations, and best practices when combining metaheuristics with simulation.
J. Manuf. Syst. 33 (4), 445–461. doi:10.1016/j.jmsy.2014.04.001. SSRN doi:10.2139/ssrn.2919208.
 T.C. Lopes, A.S. Michels and R. Lüders et al. / Computers and Operations Research 115 (2020) 104863 13

Decker, M., 1993. Capacity smoothing and sequencing for mixed-model lines. Int. J. Lopes, T.C., Sikora, C.G.S., Michels, A.S., Magatão, L., 2018. Mixed-model assem-
Prod. Econ. 30–31, 31–42. doi:10.1016/0925- 5273(93)90079- Z. bly line balancing with given buffers and product sequence: model, formu-
Dong, J., Zhang, L., Xiao, T., 2018. A hybrid PSO / SA algorithm for bi-criteria stochas- lation comparisons and case study. Ann. Oper. Res. 1 (1), 1–26. doi:10.1007/
tic line balancing with flexible task times and zoning constraints. J. Intell. s10479- 017- 2711- 0.
Manuf. 29 (4), 737–751. doi:10.1007/s10845-015-1126-5. Lourenço, H.R., Martin, O.C., Stützle, T., 2003. Iterated local search. In: Glover, F.,
Glover, F., 1986. Future paths for integer programming and links to artificial intelli- Kochenberger, W., Gary, A. (Eds.), Handbook of Metaheuristics. Springer US,
gence. Comput. Oper. Res. 13 (5), 533–549. doi:10.1016/0305-0548(86)90048-1. pp. 320–353. doi:10.1007/b101874.
Gonzalez-Neira, E.M., Ferone, D., Hatami, S., Juan, A.A., 2017. A biased-randomized McNamara, T., Shaaban, S., Hudson, S., 2016. Fifty years of the bowl phenomenon. J.
simheuristic for the distributed assembly permutation flowshop problem with Manuf. Syst. 41, 1–7. doi:10.1016/j.jmsy.2016.07.003.
stochastic processing times. Simul. Model. Pract. Theory 79, 23–36. doi:10.1016/ Merengo, C., Nava, F., Pozzetti, A., 1999. Balancing and sequencing manual mixed-
j.simpat.2017.09.001. model assembly lines. Int. J. Prod. Res. 37 (12), 2835–2860. doi:10.1080/
Guimarans, D., Dominguez, O., Panadero, J., Juan, A.A., 2018. A simheuristic approach 002075499190545.
for the two-dimensional vehicle routing problem with stochastic travel times. Otto, A., Otto, C., Scholl, A., 2013. Systematic data generation and test design for
Simul. Model. Pract. Theory 89, 1–14. doi:10.1016/j.simpat.2018.09.004. solution algorithms on the example of SALBPGen for assembly line balancing.
Hamzadayi, A., Yildiz, G., 2012. A genetic algorithm based approach for simultane- Eur. J. Oper. Res. 228 (1), 33–45. doi:10.1016/j.ejor.2012.12.029.
ously balancing and sequencing of mixed-model U-lines with parallel worksta- Özcan, U., Çerçioglu, H., Gökçen, H., Toklu, B., 2010. Balancing and sequencing
tions and zoning constraints. Comput. Ind. Eng. 62 (1), 206–215. doi:10.1016/j. of parallel mixed-model assembly lines. Int. J. Prod. Res. 48 (17), 5089–5113.
cie.2011.09.008. doi:10.1080/00207540903055735.
Hamzadayi, A., Yildiz, G., 2013. A simulated annealing algorithm based approach Öztürk, C., Tunali, S., Hnich, B., Örnek, A., 2015. Cyclic scheduling of flexible mixed
for balancing and sequencing of mixed-model U-lines. Comput. Ind. Eng. 66 (4), model assembly lines with parallel stations. J. Manuf. Syst. 36 (3), 147–158.
1070–1084. doi:10.1016/j.cie.2013.08.008. doi:10.10 07/s0 0170-012-4675-1.
Hatami, S., Calvet, L., Fernández-Viagas, V., Framiñán, J.M., Juan, A.A., 2018. A Öztürk, C., Tunali, S., Hnich, B., Örnek, M.A., 2013. Balancing and scheduling of
simheuristic algorithm to set up starting times in the stochastic parallel flow- flexible mixed model assembly lines. Constraints 18 (3), 434–469. doi:10.1007/
shop problem. Simul. Model. Pract. Theory 86, 55–71. doi:10.1016/j.simpat.2018. s10601-013-9142-6.
04.005. Roshani, A., Nezami, F.G., 2017. Mixed-model multi-manned assembly line balancing
Hillier, F.S., Boiling, R.W., 1979. On the optimal allocation of work in symmetrically problem: a mathematical model and a simulated annealing approach. Assem.
unbalanced production line systems with variable operation times. Manag. Sci. Autom. 37 (1), 34–50. doi:10.1108/AA- 02- 2016- 016.
25 (8), 721–728. doi:10.1287/mnsc.25.8.721. Sawik, T., 2002. Monolithic vs. hierarchical balancing and scheduling of a flexible
Hillier, F.S., Lieberman, G.J., 2015. Introduction to Operational Research, tenth ed. assembly line. Eur. J. Oper. Res. 143 (1), 115–124. doi:10.1016/S0377-2217(01)
McGrawHill, New York, NY doi:10.2307/2077150. 00328-9.
Hillier, F.S., So, K.C., 1991. The effect of the coefficient of variation of operation times Sawik, T., 2012. Batch versus cyclic scheduling of flexible flow shops by mixed-
on the allocation of storage space in production line systems. IIE Trans. 23 (2), integer programming. Int. J. Prod. Res. 50 (18), 5017–5034. doi:10.1080/
198–206. doi:10.1080/07408179108963854. 00207543.2011.627388.
Hillier, F.S., So, K.C., 1996. On the robustness of the bowl phenomenon. Eur. J. Oper. Scholl, A., 1999. Balancing and Sequencing Assembly Lines, second ed. Physica, Hei-
Res. 89, 496–515. doi:10.1016/0377- 2217(94)00287- 8. delberg.
Juan, A.A., Faulin, J., Grasman, S.E., Rabe, M., Figueira, G., 2015. A review of Scholl, A., Becker, C., 2006. State-of-the-art exact and heuristic solution procedures
simheuristics: extending metaheuristics to deal with stochastic combinatorial for simple assembly line balancing. Eur. J. Oper. Res. 168 (3), 666–693. doi:10.
optimization problems. Oper. Res. Perspect. 2, 62–72. doi:10.1016/j.orp.2015.03. 1016/j.ejor.2004.07.022.
001. Tiacci, L., 2012. Event and object oriented simulation to fast evaluate operational
Karabati, S., Sayin, S., 2003. Assembly line balancing in a mixed-model sequenc- objectives of mixed model assembly lines problems. Simul. Model. Pract. Theory
ing environment with synchronous transfers. Eur. J. Oper. Res. 149 (2), 417–429. 24, 35–48. doi:10.1016/j.simpat.2012.01.004.
doi:10.1016/S0377-2217(02)00764-6. Tiacci, L., 2015. Coupling a genetic algorithm approach and a discrete event sim-
Kim, Y.K., Kim, J.Y., Kim, Y.K., 2006. An endosymbiotic evolutionary algorithm for the ulator to design mixed-model un-paced assembly lines with parallel worksta-
integration of balancing and sequencing in mixed-model U-lines. Eur. J. Oper. tions and stochastic task times. Int. J. Prod. Econ. 159, 319–333. doi:10.1016/j.
Res. 168 (3), 838–852. doi:10.1016/j.ejor.2004.07.032. ijpe.2014.05.005.
Kim, Y.K., Kim, S.J., Kim, J.Y., 20 0 0. Balancing and sequencing mixed-model U-lines Tiacci, L., 2015. Simultaneous balancing and buffer allocation decisions for the de-
with a co-evolutionary algorithm. Prod. Plan. Control 11 (8), 754–764. doi:10. sign of mixed-model assembly lines with parallel workstations and stochastic
1080/09537280 0750 038355. task times. Int. J. Prod. Econ. 162, 201–215. doi:10.1016/j.ijpe.2015.01.022.
Kucukkoc, I., Zhang, D.Z., 2015. Balancing of parallel U-shaped assembly lines. Com- Tiacci, L., 2017. Mixed-model U-shaped assembly lines: balancing and comparing
put. Oper. Res. 64, 233–244. doi:10.1016/j.cor.2015.05.014. with straight lines with buffers and parallel workstations. J. Manuf. Syst. 45,
Kucukkoc, I., Zhang, D.Z., 2016. Integrating ant colony and genetic algorithms in the 286–305. doi:10.1016/j.jmsy.2017.07.005.
balancing and scheduling of complex assembly lines. Int. J. Adv. Manuf. Technol. Tiacci, L., Mimmi, M., 2018. Integrating ergonomic risks evaluation through OCRA
82, 265–285. doi:10.10 07/s0 0170-015-7320-y. index and balancing/sequencing decisions for mixed model stochastic asyn-
Lopes, T.C., Michels, A.S., Sikora, C.G.S., Magatão, L., 2019. Balancing and cyclical chronous assembly lines. Omega 78, 112–138. doi:10.1016/j.omega.2017.08.011.
scheduling of asynchronous mixed-model assembly lines with parallel stations. Venkatesh, J.V.L., Dabade, B.M., 2008. Evaluation of performance measures for repre-
J. Manuf. Syst. 50, 193–200. doi:10.1016/j.jmsy.2019.01.001. senting operational objectives of a mixed model assembly line balancing prob-
Lopes, T.C., Michels, A.S., Sikora, C.G.S., Molina, R.G., Magatão, L., 2018. Balancing and lem. Int. J. Prod. Res. 46 (22), 6367–6388. doi:10.1080/00207540701383164.
cyclically sequencing synchronous, asynchronous, and hybrid unpaced assembly
lines. Int. J. Prod. Econ. 203, 216–224. doi:10.1016/j.ijpe.2018.06.012.