The International Journal of Advanced Manufacturing Technology

https://doi.org/10.1007/s00170-019-04756-1

ORIGINAL ARTICLE

Optimization of Kanban systems using robust parameter

design: a case of study
W. L. M. Braga 1,2 & F. L. Naves 3 & J. H. F. Gomes 2

Received: 29 August 2019 / Accepted: 26 November 2019

# Springer-Verlag London Ltd., part of Springer Nature 2019

Abstract
The current competitive market demands that manufacturing companies have a survival strategy which should totally focus on
providing high-quality products, being consistent in the level of service provided, having lower inventory levels and higher
flexibility in operations. In this context, Kanban can be classified as a tool that assists in the proper sizing of inventory levels and
production control of the system. However, numerous studies have been developed with the aim of reliably reducing stock levels
in Kanban supermarkets, yielding different optimization techniques, but for fixed demands. This work proposes the minimization
solution of the supermarket from randomly different demands, using environments with simulated experiments from the
ARENA® software, from a response surface methodology (RSM) along with a weighted minimization of the mean standard
error and compare with the results provided by the use of optimizer Arena OptQuest®, thus, it is possible to evaluate the
robustness question.

Keywords Kanban . Design of experiments . Robust project design . Simulation experiments . OptQuest®

1 Introduction systems through simulated experiments brings an alternative

to the demand variability gap, making it possible to solve the
The concept of Kanban since its inception in the Toyota trade-off between the service level and the stock level; there-
Production System (TPS) has been a very helpful tool in pro- fore, it is necessary to increase the stock level to ensure a
duction scheduling and control. Its emphasis on pulling the higher level of service. In essence, the RPD concept makes
production system from a level demand provides a minimiza- the system more robust to demand variations and can be pre-
tion of inventory levels. However, it is not always possible to sented as a set of engineering techniques created by Taguchi,
work in a constant demand scenario, so the concept of Robust with the purpose of determining the levels of a set of factors,
Parameters Design for modeling and optimization of Kanban such as the sensitivity of the system. Process to noise varia-
tions is reduced, thus increasing the robustness or its ability to
remain stable in situations of continuous change in the envi-
ronment [6] (Gijo and Scaria, 2012). In order to prove the
* W. L. M. Braga effectiveness of the proposed method, this paper presents an
[email protected] example of the application of the concept in the case of the
literature proposed by Hurrion [8] and compares the results
F. L. Naves
obtained with those provided by the OptQuest® optimizer.
[email protected]
OptQuest® is an optimization tool incorporated into Arena®
J. H. F. Gomes software to automate and control the execution of the simula-
[email protected]
tion system to the corresponding values of the factor levels,
1
Pontifícia Universidade Católica de Minas Gerais (PUC), Poços de
looking for a combination of the best solutions, which are
Caldas, MG, Brazil used as points of analysis starting for a new application of
2
Institute of Industrial Engineering and Management, Federal
heuristic processes, once these steps are repeated until the
University of Itajubá, Itajubá, MG, Brazil criteria defined in the objective function and constraints are
3
Department of Chemical Engineering and Statistics, Federal
found [14]. Comparing the results of the proposed method
University of São João del-Rei, Ouro Branco, MG, Brazil with the OptQuest® optimizer provided greater reliability in
 Int J Adv Manuf Technol

the study, thus stating that the proposal is more robust and can mutually used with simulations. This combination can lead
be implemented. This work can be subdivided into the follow- to solutions that are limited, sensible to the defined statistical
ing steps: model, and prone to error.
Hurrion [8] applied DOE using the factorial technique,
Step 1: Initially, the experimental design was defined using considering 6 factors which yielded a scenario with 64 com-
the safety stock equations, which generated the re- binations of experiments. Lavoie et al. [12] study Kanban
sponse surface and the experimental arrangements. production control in homogeneous transfer lines, analyzing
Step 2: Data were collected from discrete event simulation the impact of machines that are prone to failures on the inven-
using Arena software, considering production times tory level, backlog costs, and customer satisfaction. The au-
and random monthly demands. The mean and vari- thors propose a heuristic method to reduce the number of
ance were modeled using the MSE (mean square parameters in the optimization problem and apply the concept
error), thus making it possible to size the optimal of DOE with CCD (central composite design) to plan the
quantity of Kanban in supermarkets. experiments. Hou and Hu [7] propose using a system based
Step 3: Optimization of optimal weights in mixing arrange- on the integrated Multiple Objective Genetic Algorithm
ments for better sizing of Kanban quantities in su- (MOGA), to specify the Kanban optimal number and size in
permarkets and minimization of associated variance. a company that produces sunroofs for the automotive sector.
Step 4: Simulation and determination of optimal Kanban
quantities using Optquest®.
2.2 Robust parameter design
Step 5: Comparison of the results of the proposed method
with those provided by OptQuest® and verification
Robust Parameter Design (RPD) can be defined as a set of
of robustness.
engineering techniques created by Taguchi. RPD is an effi-
cient method used to improve the quality of products and
processes and its main objective to guarantee the system’s
robustness in terms of quality, productivity, and reliability at
2 Literature review
a low cost. RPD focuses on the appropriate definition of con-
trollable factor levels, ensuring that the average response
2.1 Kanban optimization
achieved is in line with the desired level and that variability
is reduced to the lowest possible levels [7].
Most studies related to Kanban optimization mention its enor-
This method assumes that there are controllable and uncon-
mous applicability in controlling supplies, where the correct
trollable factors (noise variables) that interfere in the quality
dimensioning of the numbers of cards allows the company to
characteristics of a process. RPD defines the level of these
reach JIT manufacturing, ensuring better production planning
control variables in order to minimize the variation that the
and scheduling [4]. From the application viewpoint, studies
noise imposes on the process.
related to Kanban optimization are technologically based and
It is understood that the RPD is an RSM technique, which
can be divided into: production control, stock management,
seeks to guarantee the process robustness. When using RPD, it
and supply chain management [5]. The studies published by
is necessary to define average and variance as the systems
Price et al. [13] and Ahmed and Sultana [1] can be highlighted
variations in order to find adequate conditions of the output
among the few theoretical Kanban optimization reviews found
parameters of the examined process. This situation leads to an
in literature.
optimization of the Dual Response Surface (DRS), where the
The relevant number of publications demonstrates that
dual refers to both the average and the process standard devi-
Kanban optimization has been widely explored academically.
ation [8].
It is important to understand which methods have been used
for this purpose. This section presents a thorough analysis on
the publications regarding Kanban optimization. 2.3 Optimization by MSE
According to the articles published over the last 30 years, it
was observed that 26.8% of them proposed DOE and 56.4% Problems involving DOE can usually be optimized in terms of
as an optimization method. minimizing the variance associated with a particular expected
The total percentage considers how DOE and simulation response. The MSE (mean square error) becomes an appro-
was mentioned, regardless of whether it was applied indepen- priate method to achieve this objective.
dently or as a complement to other methods. The percentage The MSE indicated in Eq. 1 can be defined as a meth-
analysis according to the type of method studied by Araz et al. od used to combine average and variance in the optimi-
[2] reports that a great part of the articles which studied zation of a given response variable without the need to
Kanban optimization mention that regression models are consider noise [8].
Int J Adv Manuf Technol

 h
i2
 is detailed, seeking to maximize or minimize any performance

MSE y j ¼ μ y j −T y j þ σ2 y j ð1Þ metric evaluated during the simulation.

where:
3 Methodology
MSE (yj) mean squared error for the j-th response
μ(yj) average model of the j-th response
The proposed method for optimizing Kanban systems has the
Tyi target of the j-th response, obtained from the in-
purpose of contributing with the following aspects: use an
dividual optimization of μ(yj.)
experimental approach based on the service level and safety
σ2(yj) variance model established for the j-th response
stock; treat demand as a noise variable; reduce variability in
Minimization of the MSE function improves productive machine processing times; and robust optimization through
processes by reducing variability in quality, productivity, and MSE. The method uses the robust optimization concept to
costs. It also allows simultaneous optimization of average and allow analysis of experimental strategies by considering de-
variance of a single response [9]. In situations with more than mand as a noise variable. Figure 1 describes the steps taken to
one quality characteristic, Köksoy and Yalcinoz [10] proposes carry out the robust optimization, intended to adequately di-
weighted or unweighted functions to minimize the MSE, as mension the supermarkets.
described in Eq. 2. The procedure described in Fig. 1 discusses questions re-
m m n o lated to the conceptual model, computational model, its vali-
MSE G ¼ ∑ wi :MSE i ¼ ∑ wi : ½μi ðyÞ−T i 2 þ σ2i ðyÞ ð2Þ dation, and its verification, which are in accordance with the
i¼1 i¼1
simulation methods in the literature. Experimental planning of
where: the supermarket considers the security stocks to define the
different levels of service offered, so that it can define a plan
MSEG global mean square error that guarantees a more adequate level of service. The criterion
m number of considered responses adopted for this study guarantees a level of service of 99%.
wi assigned weights The number of experimental replicas is defined, and the sim-
It is important to mention that MSE will be used in this ulated experiments are executed. The control variables are the
article, and no other methods will be presented for solving amount of Kanbans in the supermarket, and they are measured
dual problems. by the average, variance, and MSE results. The quantities of
undelivered Kanbans after the simulation are record, accord-
ing to the defined experimental plan, right after are calculates
2.4 OptQuest® the average, variance, and MSE values. Then model mathe-
matically the system for average, variance, and MSE, consid-
The advantage of this method is that it saves time in the sim- ering the significant coding coefficients encountered after the
ulation process, as each solution is analyzed and simulated variance analysis. The robust optimization is performed using
individually. The finishing criteria used in the optimization the MSE and considering the variance equation (σ2), so that
process performed by OptQuest® vary according to a speci- the values are not negative and the experimental space con-
fied number of simulations; close the simulation until there is straint (xTx) is respected for the MSE. Checked whether the
no better result in the objective function for the number of system was or not robust. If robustness was achieved, the
simulations defined; as well as the combination of both pre- system was checked for consistency with the simulated data.
vious rules [9]. Aiming to find the best viable solution among Otherwise, the weighting would be applied in the next step, in
all possibilities, software heuristic procedures are employed order to find the ideal average and variance weight values. The
using concepts for providing viable solutions in a given situ- criterion adopted in this method establishes that the total sum
ation. The algorithms used by OptQuest® to search for solu- of the weights must be equal to one. In this step, each combi-
tions are scatter search, tabu search, and neural networks [9]. nation of weights yields a new optimization problem. The
To use OptQuest®, the optimization model control variables optimization results were confirmed through power and sam-
must be specified, which may be the number of resources or ple size tests using a 95% confidence interval for the power of
the information defined by the person responsible for the cre- the test. Ten experiments were generated in order to confirm
ation of the simulation model. Other variables are the response whether the simulation results were consistent with those of
variables, which affect the optimization model and can be the statistical module.
used both in constraints and in the objective function, acting To apply the proposed method, a case of Hurrion [8] was
as results of model output. There are also system constraints selected. The data was adapted and described in Table 1.
obtained from controls and/or responses. And at the end, the The simulation model was developed in the Arena® soft-
objective function, determined by a mathematical expression, ware. The conceptual model validation steps and the
 Int J Adv Manuf Technol

Fig. 1 Proposed method for

optimizing Kanban systems START

P1: Model the P6: Robust

simulated system opmizaon by MSE

P2: Experimental
planning NO
P7: Was
robustness
achieved?

P3: Execute
Experiments
YES
P8: Robust
opmizaon using
weighted MSE
P4: Record Data

P5: Model average,

variance and MSE
equaons

P9: Verify consistency

with the simulated
data in the model

END

verification of the computational model were also performed The control variables studied in the object of study were the
using the same software. quantities of Kanbans in the supermarkets, and the measured
response was the total of undelivered Kanbans. A standard
Table 1 Current scenario proposed

Product Daily Monthly Processing Total standard Lead

demand demand time (min/ deviation - σT time Table 2 Number of Kanbans with security stock
(Kanban) (Kanban) Kanban) (min/Kanban) required
(min) Type of Number of Kanban without SS Number of Kanban with
product SS (−α) SS (+α)
A 96 2880 28,8 3,9 82,944
B 24 720 117 11,03 84,240 A 96 35 131
C 38 1140 75,6 7,43 86,184 B 24 9 33
D 82 2460 34,5 2,49 84,870 C 38 14 53
D 82 30 112
Source: Adapted of Hurrion, 1997
Int J Adv Manuf Technol

Table 3 Levels used for the experiment planning Table 5 Statistical data
R-sq (adj) p value
Kanban −α −1 0 +1 +α
Average 70.93% 0.050
A 96 105 114 122 131 Variance 43.33% 0.030
B 24 26 29 31 33
C 38 42 46 49 53
D 82 90 97 105 112 In order to determine the variable levels for the CCD ar-
rangement, the proposed method used the calculation of the
security stocks for the quantities of Kanbans as a reference.
Equation 3 describes the Security Stock, as suggested by [13].
deviation of 10% was attributed during a period of 30 days. pffiffiffiffiffiffi
Therefore, the experimental arrangement was organized ac- SS ¼ F S :σD : LT ð3Þ
cording to a central composite design (CCD) to compose the
response surface experiments, generating 20 experiments. The where:
experiments were simulated with five replicates, where in SS Security stock
each replicate, new demands were generated during a period FS Service level factor
of 30 days, according to the standard deviation of 10%. σD Standard deviation of demand

Table 4 Experiment matrix

Experiment Product A Product B Product C Product D Average Variance

1 105 26 42 90 348,20 3731,70

2 122 26 42 90 299,00 9900,50
3 105 31 42 90 336,40 7781,30
4 122 31 42 90 386,80 3721,20
5 105 26 49 90 353,00 8753,50
6 122 26 49 90 277,80 7265,70
7 105 31 49 90 357,20 2011,70
8 122 31 49 90 285,40 4591,30
9 105 26 42 105 306,00 9272,50
10 122 26 42 105 283,00 3028,50
11 105 31 42 105 335,00 10,575,50
12 122 31 42 105 291,40 5914,80
13 105 26 49 105 320,20 3134,20
14 122 26 49 105 249,80 4893,20
15 105 31 49 105 301,20 3873,70
16 122 31 49 105 321,00 5306,00
17 96 29 46 98 329,20 5070,70
18 131 29 46 98 280,60 5517,80
19 114 24 46 98 242,00 3476,00
20 114 33 46 98 293,00 3768,50
21 114 29 38 98 305,00 4053,50
22 114 29 53 98 285,40 12,061,30
23 114 29 46 83 415,20 7103,70
24 114 29 46 112 360,60 7322,30
25 114 29 46 98 319,00 3494,00
26 114 29 46 98 340,40 4642,30
27 114 29 46 98 288,00 5294,00
28 114 29 46 98 377,60 7301,80
29 114 29 46 98 326,20 1123,70
30 114 29 46 98 273,80 4533,70
31 114 29 46 98 312,80 1374,20
 Int J Adv Manuf Technol

Table 6 Matrix with MSE values Table 8 Values after

MSE optimization Average Variance
Experiment Product A Product B Product C Product D EQM
Results 275 4013
1 105 26 42 90 20,683,74 Target 218 1808
2 122 26 42 90 16,461,50
3 105 31 42 90 21,799,86
4 122 31 42 90 32,214,64 calculated for the safety stock, where −α corresponds to the
5 105 26 49 90 26,978,50 lower level of the array and +α represents the upper level with
6 122 26 49 90 10,841,74 security stock, as shown in Table 2.
7 105 31 49 90 21,388,34 Table 3 describes the Kanban levels generated from the
8 122 31 49 90 9134,06 amount of Kanbans and their respective safety levels. The
9 105 26 42 105 17,016,50 others levels (−1, 0 e + 1) were obtained using the DOE
10 122 26 42 105 7253,50 (RSM).
11 105 31 42 105 24,264,50 Since the work proposal is focused on the use of the robust
12 122 31 42 105 11,302,36 parameter design, the replicate arrangement was used as an
13 105 26 49 105 13,579,04 experimental technique.
14 122 26 49 105 5904,44
15 105 31 49 105 10,795,94
16 122 31 49 105 15,915,00
4 Results and discussion
17 96 29 46 98 17,436,14
The Kanban quantities were considered as control variables to
18 131 29 46 98 9436,56
simulate the experiments. This step measured the average and
19 114 24 46 98 4052,00
variance results, the MSE values, and the defined Kanban
20 114 33 46 98 9393,50
levels, which experimental arrangement was selected and
21 114 29 38 98 11,622,50
how many experiments will be necessary. The average and
22 114 29 53 98 16,604,06
variance values of the undelivered Kanbans were calculated
23 114 29 46 83 45,991,54
from the results generated by the simulation (Table 4), for each
24 114 29 46 112 27,657,06
knitwear combination defined in the experimental planning.
25 114 29 46 98 13,695,00
26 114 29 46 98 19,624,06
4.1 Modeling average, variance, and MSE equations
27 114 29 46 98 10,194,00
28 114 29 46 98 32,773,96
The resulting data was analyzed using the Minitab® software,
29 114 29 46 98 12,830,94
and the following statistical results were found for the average
30 114 29 46 98 7647,34 and variance models (Table 5).
31 114 29 46 98 10,361,24 Equations 4 and 5 were found for the average and variance
values of the reduced model, respectively:

LT Lead time μ ¼ 305; 65−15x1 þ 11; 64x2 −6; 64x3 −14; 39x4 −9; 78x22

þ 20; 32x24 þ 10; 79x1 x2 −8; 26x1 x3 ð4Þ

According to Krajewski et al. [11], for a service level of
99.99%, there is a value for FS of 3.62. σ2 ¼ 4111; 65−71; 64x1 −793; 99x3 þ 486; 96x4
From the quantities calculated for the safety stock, the ex- þ 828; 57x24 þ 1630; 47x1 x3 −800; 83x1 x4 ð5Þ
periment planning was defined, as shown in Table 2. The
experimental planning was defined based on the quantities
Table 9 Experiment matrix for combinations of weights

Experiment (W1) (W2) MSE calculated

Table 7 Number
optimum of Kanban Product Number of Kanbans
1 1.0 0.0 4380,34
A 124 2 0.5 0.5 4457,61
B 26 3 0.0 1.0 16,489,98
C 46 4 0.75 0.25 4045,39
D 103 5 0.25 0.75 6018,08
Int J Adv Manuf Technol

Table 10 Optimized S:a : σ2 ≥ 0; 01

number of Kanbans Product Number of Kanbans
MSE ≥ 0; 01
A 123
xt x ≤ 4
B 26
C 46 The optimization problem was programmed using the
D 102 Generalized Reduced Gradient (GRG) algorithm. The identi-
fied optimum point, which resulted in the best combination of
Kanbans, is described in Table 7.
The modeling of the average was not considered satisfac- The data was simulated using a total of 30 replicates. The
tory, since it presented a R-sq (adj) smaller than 70% results of the total undelivered Kanbans were recorded, and
(Table 5). The next step was to calculate the MSE (Table 6) the average and variances values for each replicate were cal-
and model it for each experimental condition. culated and can be seen in (Table 8).
The MSE model was developed using the same procedure After optimization, the encountered results were tested in
used for the average and variance models, resulting in Eq. 6 of terms of robustness. The average values were considered sat-
the reduced model: isfactory when compared to the target values. However, the
solution presented high variance values, which does not meet
MSE ¼ 13019; 91−1986; 48x1
the standards of the robust parameter design.
þ 2274; 24x2 −4414; 29x4 −1527; 59x22
4.3 Robust optimization using weighted MSE
þ 5997; 80x24 þ 2744; 81x1 x2 ð6Þ
It is important to point out that each combination of weights
Statistical analysis considering nonsignificant interactions
yields a new optimization problem, because new coefficients
presented adequate values: a p value of 0.014 and an R-sq
are generated, and new MSE models are tested in an optimi-
(adj) of 80.26%.
zation routine. Five experimental procedures for weights were
used, and the results are described in Table 9.
4.2 Robust optimization by MSE The data in Table 9 was statistically analyzed, and the
model with the best results, after disregarding the nonsignifi-
In this step, the robust optimization is developed by MSE, as cant interactions, was the full cubic model, described in Eq. 8,
described by Eq. 7: which presented a p value of 0.046 and an R-sq (adj) of
Min MSE ¼ 13019; 91−1986; 48x1 99.89%.
þ 2274; 24x2 −4414; 29x4 −1527; 59x22 EQMP ¼ 4275; 17w1 þ 16384; 81w2 −26013; 66w1 w2
þ 5997; 80x24 þ 2744; 81x1 x2 ð7Þ þ 21771; 36w1 w2 ðw1 −w2 Þ ð8Þ

Fig. 2 Optimal theoretical values

(unweighted versus weighted) Distribution Plot
Poisson
Mean
0,025 235
237

0,020
Probability

0,015

0,010

0,005

0,000
200 220 240 260 280 300
X
 Int J Adv Manuf Technol

Fig. 3 Samples for validation of

optimized values Power Curve for 1-Sample t Test
1,0
Sample
Size
28
0,8 32
37
45

Assumptions
0,6 α 0,05

Power
StDev 65,1057
Alternative ≠

0,4

0,2

0,0
-50 -25 0 25 50
Difference

Table 11 Total
undelivered Kanbans Sample Results The WMSE minimization was used to obtain the optimal
weights, described in Eq. 9, using the spreadsheet with the
1 323 GRG algorithm.
2 349
3 219
4 290 Min WMSE ¼ 4275; 17w1
5 281
6 299 þ 16384; 81w2 −26013; 66w1 w2
7 249
8 206
9 384 þ 21771; 36w1 w2 ðw1 −w2 Þ ð9Þ
10 320
11 301 s:a : 0≤ w1 ≤ 1
12 381
13 229 0 ≤ w2 ≤ 1
14 320
15 388 w1 þ w2 ¼ 1
16 231
17 241 The resulting optimal values for the weights w1 and w2,
18 328
19 263 respectively, were 0.1179 and 0.8821. After the weights were
20 308 calculated, the values were inserted into the formulated prob-
21 261
22 242 lem, and the results of the optimal weights were found with
23 280 their respective model described in Eq. 10:
24 376
25 355
26 203 Min WMSE ¼ 6467; 01−651; 63x1
27 294
28 316 þ 966; 50x2 −2699; 13x4 −838; 29x22
29 283
30 408
31 248
þ 3068; 28x24
32 280
33 284 þ 2032; 59x1 x2 −853; 68x2 x4 ð10Þ
34 201
35 236
36 298
37 293 Table 12 Confirmation
38 265 results Target Result
39 362
40 377 Average 218 295
μ 295,41
Standard deviation 43 56
σ 55,64
σ2 3095,95 Variance 1808 3096
Int J Adv Manuf Technol

Table 13 Information used in programming the OptQuest® Table 15 Comparison of results proposed method versus OptQuest®

Informations Description A B C D TOTAL

Control variables Kanban product A; Kanban product B; Proposed method 123 26 46 102 297
Kanban product C; Kanban product D OptQuest® 131 33 53 112 329
Response variable Total Delivery Product A (TEA); total delivery Difference −6% −21% −13% −9% −10,77%
product B (TEB); total delivery product C
(TEC); total delivery product D (TED);
Purpose Maximize total delivered Kanban products (Z)

Source: Prepared by the author (2019) allowing to search for optimal solutions in the simulation
model. The information used for optimizer configuration were
s:a : σ2 ≥ 0; 01 the control variables, the response variables, constraints, and
objective function, described in Table 13 with the mathemat-
WMSE ≥ 0; 01
ical modeling shown in Eq. 11.
xT x≤ 4

From the coefficients of Eq. 10, the optimization of the Max Z ¼ 1*ðTEAÞ þ 1*ðTEBÞ þ 1*ðTEC Þ þ 1*ðTEDÞ ð11Þ
number of Kanbans was performed using the GRG algorithm.
The results can be seen in (Table 10). S.a:
Figure 2 illustrates the comparison of the first optimization Kanban A ≤ 131
without weights (represented by the blue curve) with the op- Kanban A ≥ 96
timization after weighting (shown in the green curve). Kanban B ≤ 33
The results were considered consistency in adequately Kanban B ≥ 24
sizing the supermarkets obtained in the simulation model kanban C ≤ 53
and then were verified through the number of confirmation kanban C ≥ 38
experiments (Table 10); this was performed using the Power kanban D ≤ 112
and Sample size test, and the values for test power definition kanban D ≥ 82
were 0.80 0.85 0.90 0.95. According to the results, 40 samples
are suggested to reach a 95% confidence interval (Fig. 3). About 100 simulations were performed with a number of
Confirmatory experiments were performed in the simulated three replicates, default values of the system, in order to better
model itself to validate the results and verify if the simulation estimate the results delivered by the optimizer. In the end, we
values would match the results of the statistical model. found the optimal results described in Table 14.
Table 11 shows the results yielded after simulating the The values found from the simulator bring a worse config-
computational model with 10 confirmation replicates. uration in terms of quantities of Kanbans in stock (+ 10.77%),
Table 12 shows the confirmation results compared to the when compared to the proposed method (Table 15).
calculated maximum limits for a 95% confidence interval. Moreover, it does not present a robust solution when eval-
uating the mean and variance behavior (Table 16), where the
proposed method allows to achieve a more robust value with a
smaller variance.
5 Comparison of results with Arena software
According to Table 16, the results show that the proposed
optimizer
method enables the production system described by Hurrion
[8] to obtain greater stability in the service level and inventory
The adapted study object of Hurrion [8] contributed in the
comparison between the proposed optimization method and
the results obtained from the Arena® optimizer OptQuest®,

Table 14 Optimal values OptQuest® Table 16 Comparison results of average and variance
Product description Quantity of Kanban Maximum limit Value achieved Value
(95% CI) proposed method OptQuest®
Kanban A 131
Kanban B 33 Average 218 295 645
Kanban C 53 Standard deviation 43 56 79
Kanban D 112 Variance 1808 3096 6156
 Int J Adv Manuf Technol

level indicators, thus being able to behave better in a scenario A suggestion for future studies is to study the implementa-
of demand variation. tion of the proposed method in scenarios with a greater mix of
products and a great number of workstations and to verify the
extension of the method’s effectiveness.
6 Final considerations

The present study was developed with the purpose of recom-

mend an alternative to sizing of Kanban supermarkets based
on robust optimization and analyzed the application of the References
method in the context of a literature case with simulated ex-
periments, comparing the results with the Arena® OptQuest® 1. Ahmed I, Sultana I (2014) A state of art review on optimization
techniques in just in time. Uncertain Supply Chain Manag 2, 15:–26
optimizer. This objective was fulfilled since the study problem
2. Araz OU, Eski O, Araz C (2008) Determining the parameters of
was mathematically modeled and the proposal validity was dual-card Kanban system: an integrated multicriteria and artificial
verified using a computational model. neural network methodology. Int J Adv Manuf Technol 38:965–
Opportunities to develop methods that present more suit- 977. https://doi.org/10.1007/s00170-007-1138-1
able configurations in the definition of experimental planning 3. Ardakani MK, Noorossana R (2008) A new optimization criterion
for robust parameter design – the case of target is best. Int J Adv
as well as the use of the robust parameter design to reduce the Manuf Technol 38:851–859
inventory levels in Kanban systems were identified after a 4. Bezzerra MA, Santelli RE, Oliveira EP, Villar LS, Escaleira LA
careful review of the literature on Kanbans. (2008) Response surface methodology (RSM) as a tool for optimi-
Robust design approaches often present difficulty when zation in analytical chemistry. Talanta 76:965–977
modeling the average and variance equations due to the ran- 5. Brenneman, W.; Myers, W. Robust parameter design with categor-
ical noise variables. J Qual Technol. Vol. 35, n 4, 2003
domness disturbances caused by the noise variables that are
6. Gijo, EV Scaria, J (2012) Product design by application of
purposefully experienced and randomized for experimenta- Taguchi's robust engineering using computer simulation.
tion processes. MSE modeling was an effective strategy in International Journal of Computer Integrated Manufacturing.
overcoming these difficulties which are crucial for the appli- 25(9): 761–773.
cation of the robust design, and weighting before using MSE 7. Hou TH, Hu WC (2011) An integrated MOGA approach to deter-
mine the Pareto-optimal KANBAN number and size for a JIT sys-
was important in prioritizing the desired response tem. Expert Syst Appl 38:5912–5918
The experimental approach was fundamental in modeling, 8. Hurrion RD (1997) An example of simulation optimization using a
analyzing, and optimizing the Kanban supermarkets. This ap- neural network metamodel: finding the optimum number of
proach was also useful in evaluating the applicability of the Kanbans in a manufacturing system. J Oper Res Soc 48:1105–1112
proposed method, and the GRG algorithm was able to solve 9. Jerbi A, Ammar A, Rid M, Salah B. Performance optimization of a
flexible manufacturing system using simulation: the Taguchi meth-
the modeled equations with a satisfactory accuracy. od versus OptQuest®
Considering demand as a noise variable made it possible to 10. Köksoy O, Yalcinoz T (2006) Mean square error criteria to
evaluate the behavior of the production system and introduces multiresponse process optimization by a new genetic algorithm.
a new approach to Kanban supermarket sizing, since in real Appl Math Comput 175:1657–1674
situations, most companies consider the level demand for such 11. Krajewski LJ, Ritzman L, Malhotra M (2009) Operations manage-
ment, 9th edn. Edição, Pearson Prentice Hall
a task. 12. Lavoie P, Gharbi A, Keneé JP (2010) A comparative study of pull
Designing the experiments based on the security stock control mechanisms for unreliable homogenous transfer lines. Int J
equations made it possible to balance common a trade-off in Prod Econ 124:241–251
organizations: the difficulty in ensuring the level of service 13. Price W, Gravel M, Nsakanda AL (1994) A review of optimization
models of Kanban-based production systems. Eur J Oper Res 75:1–
without increasing the stock levels.
12
The approach proposed in this paper is applicable to the 14. Rossetti MD (2016) Simulation modeling and Arena®, 2nd edn.
physical dimensioning of Kanban controlled supermarket Edição. John Wiley & Sons
stocks using both the traditional system and the electronic
Kanban without any interference in the product variety or Publisher’s note Springer Nature remains neutral with regard to jurisdic-
the production scale. tional claims in published maps and institutional affiliations.