You are on page 1of 13

Int J Adv Manuf Technol (2008) 38:965–977

DOI 10.1007/s00170-007-1138-1

ORIGINAL ARTICLE

Determining the parameters of dual-card kanban system:


an integrated multicriteria and artificial neural network
methodology
Ozlem Uzun Araz & Ozgur Eski & Ceyhun Araz

Received: 1 December 2006 / Accepted: 20 June 2007 / Published online: 14 August 2007
# Springer-Verlag London Limited 2007

Abstract In this study, we proposed a methodology for inventory reduction, higher throughput times, better deliv-
determining the design parameters of kanban systems. In ery performance, higher machine utilization, lower costs,
this methodology, a backpropagation neural network is and greater profits.
used in order to generate simulation meta-models, and a In traditional manufacturing systems, the materials are
multi-criteria decision making technique (TOPSIS) is pushed through the system and the finished products are
employed to evaluate kanban combinations. In order to stocked so as to meet expected demand. On the other hand,
reflect the decision maker’s point of view, different weight JIT systems are ‘pull’ production systems which withdraws
structures are used to find the optimum design parameters. the appropriate quantity of material needed satisfying the
The proposed methodology is applied to a case problem final schedule requirements. Therefore the ‘pull’ production
and the results are presented. We also performed several system produces only the quantity of parts required to
experiments on different types of problems to show the replace those that have been withdrawn.
effectiveness of the methodology. A kanban system is one of the important elements of JIT
philosophy. Kanban means card in Japanese. In a kanban
Keywords Kanban . Multicriteria decision making . system, cards are used systematically to control the produc-
Simulation metamodeling . Artificial neural networks tion within stations and the movement of parts between
stations. They also help to visualize and control work-
in-process (WIP) inventories [1]. Therefore, kanban system
1 Introduction can be defined as a decentralized system that harmoniously
controls the production of the necessary products in the
Just-in-time (JIT) is a management philosophy that encour- necessary quantities at the necessary time [2].
ages improvement through inventory reduction and elimi- According to their functions, different types of kanbans
nates manufacturing wastes by producing only the right can be formed such as supply kanban, procurement kanban,
quantity at the right time. The use of JIT philosophy results subcontract kanban, auxiliary kanban [3]. Two types of
in some important improvements to the company, such as kanbans are generally used: withdrawal kanban and
production kanban. A withdrawal kanban defines the
quantity that the succeeding workstation should withdraw
from the preceding one while a production kanban defines
the quantity which the preceding workstation must produce.
O. U. Araz (*) : O. Eski : C. Araz
Industrial Engineering Department, Dokuz Eylul University,
In kanban systems, the efficiency of the manufacturing
35100 Izmir, Turkey system depends on several factors such as the total number
e-mail: ozlem.uzun@deu.edu.tr of kanbans, transport sizes (container size), safety storage
O. Eski sizes, the sequencing rules, etc. [4]. For instance, the
e-mail: ozgur.eski@deu.edu.tr inventory levels can be controlled by the total number of
C. Araz kanbans. If too many kanbans are used, the result is large
e-mail: ceyhun.araz@deu.eu.tr amount of work-in process inventory at each workstation.
966 Int J Adv Manuf Technol (2008) 38:965–977

Conversely, reducing the number of kanbans causes a experiments are also performed in the same section. Final
decrease in work-in-process inventory but also leads to a section includes concluding remarks.
reduction in the system throughput rate [5]. Hence, deter-
mining the best combination of kanban system parameters is
a critical decision. 2 Literature review
In general, simulation is used as a tool to analyze such
complex problems. However, the use of simulation is time Several studies have been done to find the number of
consuming, due to the combinatorial nature of the problem. kanbans required in JIT system. A detailed review on JIT
In order to overcome these limitations of simulation, meta- and kanban systems can be found in Akturk and Erhun [1],
modeling approaches have been widely used. A metamodel and Huang and Kusiak [3]. In most of the existing studies,
is a mathematical approximation of a simulation model that kanban systems are modeled by using mathematical
represents the relationship between inputs and outputs. The programming, Markov chains or simulation approaches.
use of metamodeling approach is helpful when the Kimura and Terada [11] developed the first mathematical
simulation becomes very large and costly. In general, model for the kanban system. In an extended version of
regression analysis has been combined with simulation for Kimura and Terada’s study, Bitran and Chang [12] used a
building simulation metamodels [6]. However, regression nonlinear integer formulation for kanban determination
metamodels have some drawbacks: it is limited to approx- problem. Bard and Golany [13] developed a mixed integer
imate a subset of the simulation domain, sensitive to linear program to determine the number of kanbans.
deviations from statistical model assumptions, and it is Philipoom et al. [14] developed a nonlinear integer
possible to select incorrect functional form during the mathematical model for the multi-item, multi-stage, multi-
analysis [7]. On the other hand, some researchers have period, capacitated kanban system. Husseini et al. [15]
stated that meta-modeling through artificial neural network proposed a methodology using an integer linear program-
(ANN) models may overcome these disadvantages of ming technique to flexibly determine the number of
traditional regression approaches [7, 8]. kanbans while minimizing total inventory cost. Wang and
Performances of kanban systems are generally measured Sarker [16] developed a mixed integer nonlinear mathe-
by more than one performance measures, usually in conflict. matical model for determining the optimal number of
If a decision maker is interested in optimizing all objectives kanbans, batch size, the number of batches and total
simultaneously, an aggregated weighted function has to be quantity over one period for a multistage supply chain
defined. Some researchers have used cost function as system.
aggregated weighted function in order to evaluate all Markov chains have been widely used to describe the
possible kanban combinations (see e.g., [5, 9]). However, behavior of kanban system and to find the optimum number
in real life, construction of the cost function is a difficult of kanbans. Askin et al. [17] determined the optimum
task because of incommensurable nature of some perfor- number of production kanbans by using a continuous time,
mance measures such as average customer delay time, steady state Markov model for multi-item, multi-stage serial
maximum customer delay time, the number of demand production system. Nori and Sarker [18] used Markov
backordered, etc. Multi-criteria decision making (MCDM) Chains to model a single card kanban system and found the
approaches may be helpful in evaluating the alternative optimum number of kanbans between two adjacent stations.
kanban combinations with regards to multiple performance Tardif and Maaseidvaag [19] modeled a new adaptive
measures. kanban-type pull control mechanism which determines
This paper presents an integrated MCDM and ANN when to release or when to reorder raw parts based on
methodology in order to determine the optimum kanban customer demands, inventory and backorders using contin-
parameters. A backpropagation neural network is used for uous Markov chains.
generating simulation metamodels. Instead of using an Simulation is another tool used to analyze kanban
aggregated cost function, TOPSIS [10], which is a well- systems. Gupta and Gupta [20] simulated a single-item,
known MCDM technique, is employed to evaluate kanban multi-stage dual card kanban system in order to investigate
combinations. We also performed some experiments in the effect of the number of kanbans and kanban sizes on the
order to investigate the effectiveness of the methodology. system performance. Sarker [21] and Savsar [22] presented
This paper is further organized as follows. Section 2 simulation models to analyze JIT systems under various
presents a summary of the related literature. In Sect. 3, a operating conditions. Andijani [23] used a stochastic
brief description about ANN is given. Section 4 contains a system simulation to generate the set of efficient kanban
description of TOPSIS method. Section 5 is devoted to combinations. Analytic hierarchy process (AHP) is utilized
explain the proposed methodology. The proposed method- to identify the most-preferred allocation. Sengupta et al.
ology is applied on a case problem in Sect. 6. Some [24] presented a simulation-based direct search approach in
Int J Adv Manuf Technol (2008) 38:965–977 967

order to determine the optimal number of kanbans. They signals weighted by the respective synapses of the neuron.
also studied on simultaneous allocation of kanbans to (3) An activation function that limits the amplitude of the
different part types. Paris and Pierreval [4] proposed a output of a neuron.
distributed simulation optimization method based on ANNs have a number of features that offer some useful
evolutionary principles, which is capable of simultaneously properties and capabilities. First, the ANN computations
considering various parameters in the configuration of may be carried out in parallel which allows for fast
multi-product kanban systems. Kochel and Nielander [25] computations. Second, ANNs can represent both linear
proposed an evolutionary optimization approach which and non-linear relationships. Finally, ANNs have the ability
combines simulation and genetic algorithms in order to to adapt their synaptic weights to changes in the behavior of
find optimum number of kanbans and the volume of the surrounding environment [30].
containers. Shahabudeen et al. [26] used a simulated In general, two types of learning are used in ANN
annealing technique to select the number of production models namely, supervised learning and unsupervised
and withdrawal kanbans at each workstation and the lot size learning. The ability to learn from example is one of the
for each part type. They developed an object-oriented most important characteristics of supervised learning. The
simulation model for a multi-product, dual card dynamic backpropagation (BP), which is the most widely used
kanban system. supervised learning algorithm, was presented by Rumelhart
Some researchers have used simulation metamodeling et al. [31]. In BP learning, error signals which are
approaches. Aytug et al. [27] determined the number of calculated by subtracting the actual response of the network
kanbans in a JIT system using a simulation metamodeling from a desired response, propagated backward through the
approach. They developed a set of metamodels for a two- network against the direction of synaptic connections. In
stage kanban using regression metamodels. As mentioned order to improve the performance of the network, the
before, the use of artificial neural networks is another weights are adjusted during the learning phase [29]. When
effective approach for metamodeling. Hurrion [28] devel- the network reaches a satisfactory level of performance, it
oped a neural network metamodel to search for the optimal means that network learns the relationship between input
kanban combination for a two-station pull system. Savsar factors and desired outputs. Backpropagation neural net-
and Choueiki [5] extended the study of Hurrion [28] and work is one of the most frequently used tool in simulation
proposed a generalized systematic procedure that integrates metamodeling. A neural network-based simulation meta-
experimental design concepts with simulation and neural model is a neural network whose training set (i.e., input-
networks for solving kanban allocation problem. output pairs) is provided by a simulation model [32]. There
are several studies in the literature that have demonstrated
that the neural network based simulation metamodels can
3 Artificial neural network modeling be effectively used for estimation of the system perfor-
mance [7, 8, 32–36].
Artificial neural network is a computing architecture that is
motivated by the biological nervous systems. ANNs are
composed of a large number of simple processing elements, 4 TOPSIS
all connected and in parallel. There are three basic elements
of an ANN model [29], which are illustrated in Fig. 1: (1) TOPSIS, which was developed by Hwang and Yoon [10], is
Set of synapses or connected links which is represented by one of the well-known MCDM techniques. The basic
a weight of its own. (2) An adder which sums the input concept of this method relies on the selection of the best
alternative that has the shortest distance from the ideal
solution and the farthest distance from the negative-ideal
x1 wk1 solution. The TOPSIS method consists of the following
steps [37, 38]:
Summing
x2 wk2
junction Step 1 Normalization of performance measures. This step
Input
uk
Output converts the various attribute dimensions into non-
ϕ (.)
signals . Synaptic ∑ yk dimensional attributes. Normalized value of ith
. weights
Activation alternative in terms of the jth criterion (aij) is
. function
calculated as follows:
xm wkm
Bias xij
bk aij ¼ i ¼ 1;2;:::;m: j ¼ 1;2;:::; n: ð1Þ
maxaij
Fig. 1 Non-linear model of a neuron [29] i
968 Int J Adv Manuf Technol (2008) 38:965–977

where, xij is the performance measure of the ith


alternative in terms of the jth criterion.
Step 2 Determine the ideal and the negative-ideal solutions.
The ideal aM and the negative-ideal solutions am
are defined as follows:
 
aM ¼ aM 1 ; :::; a M
j ; ::; a n :
M
ð2Þ

where aM j is the best value for the jth attribute


among all available alternatives. The composite of
all best attribute ratings is the ideal solution,
whereas the negative-ideal solution is composed
of all worst attribute ratings. The negative-ideal
solution am is given as
 
am ¼ am 1 ; ::; aj ; ::; an :
m M
ð3Þ

where am j is the worst value for the jth attribute


among all alternatives.
Step 3 Calculate the separation measure. The weighted
distances of each alternative from (aM) and (am)
are then calculated respectively, according to the
chosen metric p:
hX  p i1=p
p M 
d M pð ai Þ ¼ w
j j
a j  a ij ð4Þ

hX  p i1=p
p m 
d m p ð ai Þ ¼ w
j j
a j  a ij ð5Þ

where wj is the weight of jth criteria.


Step 4 Calculate the relative closeness (similarity ratio) to
the ideal solution. The relative closeness of ith
alternative with respect to the ideal solution aM is
defined as follows:

d m p ð ai Þ
D p ð ai Þ ¼ ð6Þ
dM p ð ai Þ þ d p ð ai Þ
m Fig. 2 The proposed methodology

This ratio varies from Dp(am)=0 for the anti-ideal


Step 1 Selection of possible kanban combinations. First,
point to Dp(aM)=1 for the ideal point.
subject to some system constraints such as
Step 5 Rank the preference order. Rank the alternatives
maximum number of kanbans in a station, the set
according to the descending order of similarity
of all possible kanban combinations is identified.
ratio Dp(ai).
Selection of a limited set of combinations can be
made by using randomized, one-factor-at-a-time or
full factorial experimental design.
5 Proposed methodology for determining Step 2 Simulation modeling of manufacturing system. The
the parameters of kanban systems second step of the methodology includes building
a simulation model of JIT manufacturing system,
In this study, we proposed a design methodology incorpo- setting the parameters of system such as arrival
rating BP neural network-based meta-modeling and TOP- time and processing time distributions and defining
SIS multi-criteria decision making technique for the performance measures to be evaluated. Finally,
determining number of kanbans and kanban size. The flow the simulation model is run for each kanban
of this methodology is illustrated in Fig. 2. combinations selected in step 1.
Int J Adv Manuf Technol (2008) 38:965–977 969

Step 3 Building, training and testing of neural network station and controls production while the WK circulates
model. An ANN is built, trained and tested by between workstations and controls the material flow.
using the selected set of kanban combinations as The system under consideration is in stochastic nature
inputs and performance measures obtained by and has following assumptions:
simulation as outputs.
– An infinite supply of raw materials.
Step 4 Full enumeration: The performance measures of
– Transfer time for raw materials is 1 minute between
all possible kanban combinations are obtained by
warehouse and the input buffer of the first station.
using ANN model built in step 3.
– Transfer time for PKs is negligible.
Step 5 Evaluation of kanban combinations by using a
– Transfer time for WKs is 1 minute between workstations.
multi-criteria decision making approach. The
– The kanbans are released as soon as their containers are
performance measures conflicting in nature are
empty.
evaluated by TOPSIS approach, instead of using
– The container size is allowed to change within the
an aggregated cost function.
range [1-3].
– Including PKs and WKs, a minimum of two kanbans
In the following section, the proposed methodology will and a maximum of five kanbans can be assigned to
be applied on a case problem in order to find the optimum each station.
kanban allocations. – Demand is satisfied on FCFS rule.
The time between two successive demands is randomly
chosen according to an exponential distribution with a
6 Numerical example
mean of 11.25. Processing times are normally distributed
with a mean of 8, 10, 6 and a standard deviation of 1.15,
The manufacturing system under consideration is an
2.75, and 1.25, respectively.
assembly line composed of three workstations and produces
There are seven decision variables to be considered in
a single item. Figure 3 represents an assembly line which is
this case: WK1, PK1, WK2, PK2, WK3, PK3, CS. WKi
controlled by dual card kanban system. Each workstation
and PKi represent the number of withdrawal and production
(SN) consists of one machine, one input (IBN) and output
kanbans assigned to the workstation i, respectively. Con-
buffer (OBN). Containers which are located in input and
tainer size is denoted by CS and assumed to be same for all
output buffers are used to transport the products and one
containers in the system.
kanban is attached to each container. When the demand
In this study, four important performance measures,
arrives to output buffer of the last station, the demand is
namely average work-in process level (WIP), average
fulfilled from the output buffer. If there is no available
customer delay time (CDT), number of containers (# of
finished product in the output buffer, the demand waits until
con.), percentage of the number of tardy orders (%ntardy),
a container becomes available. The production kanban (PK)
are considered. It should be noted that the number of
is detached from the container as soon as their containers
containers is equal to the number of kanbans.
are empty. The detached PK is sent to the input buffer of
A JIT manufacturing system is expected to respond
the station to start the production providing that there are
quickly to demand with minimum average WIP. Customer
available materials in the input buffer. At this time,
delay time and the WIP levels are in conflicting nature. In
withdrawal kanban (WK) and emptied container are sent to
other words the lower the WIP levels, the higher the
the output buffer of the preceding workstation and replaced
average customers delay time and vice versa. However
with a full one. This process continues throughout the whole
minimization of the average WIP is also necessary.
preceding stations. The PK always circulates within the
Additionally, if demand has not been met in time it causes

Fig. 3 N-station dual-card kan- WK1 PK1 WK2 PK2 WK PKN


ban system [40]

S1 SN
WH S2

IB1 OB1 IB2 OB2 IBN


Raw Finished
Material Product
Warehouse Buffer
Flow of parts and Kanban
Flow of Kanban
970 Int J Adv Manuf Technol (2008) 38:965–977

dissatisfaction for the customer. In the same manner, the A feed-forward architecture including one layer of
number of tardy orders is another critical performance hidden units has been employed for each performance
measure for JIT systems even if the average customer delay measure. WK1, PK1, WK2, PK2, WK3, PK3 and CS were
time is low. Moreover, the number of containers should be introduced as input nodes.
decreased for reducing material handling cost. It is clear that The performance of neural network depends on several
other performance measures can be added with respect to the design parameters such as number of hidden layers, the
characteristic of the manufacturing system considered. number of nodes in the hidden layer, the type of transfer
According to the manufacturing system constraints, the function, learning rate, momentum rate, etc. In general,
number of kanban combinations for each station is 10 as identification of these design parameters is done by trial
seen in Table 1. Additionally, container size can take values and error [5]. Using trial and error method, we set the
between 1 and 3. Thus the total number of kanban design parameters in Table 2.
combinations to be simulated is 10×10×10×3=3,000. For neural network model, 300 and 50 input-output pairs
Simulating 3,000 combinations (15 replications) requires were randomly taken and used for training and cross
approximately 150 hours. A neural network-based meta- validation, respectively. When the number of learning
model requires much less time than the related simulation epochs is larger than 10,000 or the mean square error is
model to generate the simulation outputs. less than 0.0001 the learning process stops. The remaining
Considering the above system, the optimum number of 50 combinations were used for testing.
kanbans for each work station and the optimum container size The mean square error was used as the measure of
are determined using the proposed methodology as in follows: accuracy of the neural network. Figure 5 displays the
simulation and neural network results of test data.
Step 1 In this step, the set of all possible kanban
As seen in Fig. 5, the neural network model provides
combinations was identified considering some
accurate estimates for customer delay time, WIP levels and
system constraints such as maximum number of
percentage of number of tardy orders. Table 3 shows some
kanbans in a station. Among 3,000 possible
measure of accuracy such as mean square error; mean
combinations, 400 combinations were selected
absolute error and root mean square error. Although the
using randomized experimental design.
accuracy measures show that artificial neural network
Step 2 The simulation model of three station JIT
models have good performance for all performance mea-
production line was constructed using ARENA
sures, we compared artificial neural network metamodels
10.0 Simulation Software and run for each
and regression metamodels with respect to their deviations
combinations selected in Step 1.The replication
from the simulation results.
length was chosen as 35,000 minutes with a
The quadratic regression model was fitted using the
warm-up period of 5,000 minutes. The variance
training set data (300 examples) of neural network model.
reduction technique of common random numbers
In the comparison, the test data set (50 examples) of the
[39] is used for synchronization of random
neural network models was used. According to the results
numbers so that the combinations are compared
in Table 3, one can conclude that, for the problem
under similar conditions. Fifteen replications
considered, the neural network models give better results
were made for each combination and the average
in terms of all performance measures.
WIP levels, percentage of number of tardy orders
Step 4 The average WIP levels, percentage of number of
and customer delay times were computed.
tardy orders and average customer delay times for
Step 3 After obtaining the corresponding performance
each kanban combination were obtained by using
measures for 400 kanban combinations, we built a
the neural network model built in step 3. The
multilayer feed-forward neural network model for
fourth performance measure, number of containers,
each performance measure as illustrated in Fig. 4.
is equal to total number of kanban since one
kanban is attached to each container.
Step 5 After obtaining performance measures for all
Table 1 Total kanban combinations kanban combinations, TOPSIS technique was
WKi PKi used in order to choose optimum kanban combi-
nation considering the relative importance of the
1 1, 2, 3, 4 performance measures. All of these performance
2 1, 2, 3 measures were wanted to be minimized.
3 1, 2
4 1
Firstly, all data are normalized between 0 and 1.
Total number of combinations for each workstation 10
Secondly, ideal and anti-ideal points for all criteria were
Int J Adv Manuf Technol (2008) 38:965–977 971

Fig. 4 Neural network model of


the JIT manufacturing system

WK1

PK1

WK2 Performance
Measures
PK2

WK3 Output Layer

PK3

CS

Input Layer Hidden Layer

identified as shown in Table 4. Thirdly, the weighted with the optimum solution. All of the solutions seen in
euclidean distances (p=2) d2M ðaiÞ and d2m ðaiÞ to the ideal Table 6 were obtained for the weight structure (0.8, 0.1,
and anti-ideal points were computed from Eqs. (4) and (5). 0.05, 0.05) and can be chosen to implement, since all of
At the fourth step, similarity ratio is calculated from Eq. (6). them provides very close performances on all criteria. It is
Finally, the ranking of the alternatives was obtained obvious that a decision maker can find different solutions
according to the descending order of similarity ratio D2(ai). by using different weight structure that reflects her/his
We assumed that the customer delay time has the highest preferences.
priority among all measures. In order to reflect the relative
importance of performance measures for different decision 6.1 Illustrative experiments and results
makers (DM), different weight structures were generated.
For given set of weights, the alternative with the highest Dp As mentioned before, allocation of kanbans to the stations
was chosen. Table 5 shows the alternative solutions is a complex manufacturing design problem due to the
according to DM’s point of view. variations in demand, processing times and product types.
As seen in Table 5, if customer delay time is used as a In order to show the effectiveness of the proposed
sole criterion for choice, the alternative (2-3-1-4-2-3-3) for methodology, we performed several experiments on differ-
(WK1, PK1, WK2, PK2, WK3, PK3 and CS) is selected. ent types of problems in the next subsections.
This alternative has maximum number of kanbans (15) and
maximum container size (3) and provides minimum 6.1.1 Experiment 1–effect of changing in problem size
customer delay time. It should also be noted in Table 5
that when the weight of customer delay time is decreased, In the first experiment, the proposed methodology is
the alternatives which have lower container size and fewer applied on two problem instances with different sizes; four
kanbans are chosen. It causes an improvement on the WIP and five workstations, respectively. The same assumptions
performance. given in previous section have been used for each problem
Furthermore, for a given weight structure, there are case. Tables 7 and 8 show the processing times and the
several solutions that have close TOPSIS similarity ratio networks design parameters, respectively.
Some measure of accuracy such as mean square error;
Table 2 Design parameters of neural network mean absolute error and root mean square error were also
Design parameters %ntardy CDT WIP
calculated for each instance in order to show the accuracy
of the neural network models trained. Table 9 gives the
Number of hidden layer 1 1 1 values of these measures.
Number of nodes in hidden layer 8 9 8 Using the proposed methodology and considering the
Learning rate 0.1 0.1 0.1 same weight structure, we determined the optimum kanban
Momentum rate 0.7 0.7 0.7
allocation for these two problem instances. The results are
Transfer function Hyperbolic tangent
given in Table 10. It can be seen from the results that the
972 Int J Adv Manuf Technol (2008) 38:965–977

Fig. 5 Simulation and neural % ntardy WIP


network results of test data 35
1
30
0.8
25
0.6 20
15
0.4
10
0.2 5
0 0
1 5 9 13 17 21 25 29 33 37 41 45 49 1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
Simulation Results Simulation Results
ANN Results ANN Results

CDT
40
35
30
25
20
15
10
5
0
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
Simulation Results
ANN Results

procedure gives reasonable solutions for each case. For 6.1.2 Experiment 2–effect of increasing the number
example, for 5-workstation instance with weight structure of products
of (0.8, 0.1, 0.05, 0.05), the optimum solution is the kanban
combination (1,1,1,1,1,4,1,4,1,4,2) for (WK1, PK1, WK2, In this experiment, the three-workstation case is again used.
PK2, WK3, PK3, WK4, PK4, WK5, PK5 and CS). The However, it is assumed that production line can handle
associated customer delay time is 1.379 while the average three different types of product, A, B, and C. In this case,
WIP level is 24.1715. When the weight of WIP is increased each part type has its own production and withdrawal
by 0.1 (from 0.1 to 0.2) and the weight of customer delay kanbans. In addition to the base case assumptions,
time is decreased by 0.2 (from 0.8 to 0.6), the performance following assumptions are also considered:
of customer delay time is deteriorated while WIP level is
– The size of container is the same everywhere and fixed
decreased. In this case, the total number of kanban is
at two for each product type.
decreased from 19 to 17 in order to increase WIP
– Including PKs and WKs, a minimum of six kanbans
performance.
and a maximum of eight kanbans can be assigned to
Furthermore, it should also be noted in Table 9 that
each station for all product types.
although the number of decision variables and the number
– Each type of job has its own demand interarrival time
of possible kanban combinations are increased in some
distribution.
extend, artificial neural network models gives accurate
– All the jobs are processed in all three stations.
results for each performance measure in a reasonable time.

Table 3 Some measure of accuracy for%ntardy, CDT and WIP Levels (MSE: mean square error; RMSE=root mean square error) for ANN and
regression model

ANN Regression

Error measure %ntardy CDT WIP %ntardy CDT WIP

Mean absolute error 0.0047 0.179 0.125 0.123 1.412 0.855


% Absolute error 3.582 2.151 0.886 17.34 25.99 7.20
MSE 3.862e–05 0.056 0.037 0.025 3.716 1.432
RMSE 0.006 0.236 0.194 0.158 1.928 1.197
Int J Adv Manuf Technol (2008) 38:965–977 973

Table 4 Ideal and anti-ideal points for all criteria number of containers and container size. This is probably
Criteria Ideal points Anti-ideal points because the procedure increases the number of kanbans
(and number of containers) in order to decrease the
%ntardy 0.0106 1 customer delay time. In the same manner, when demand
CDT 0.0384 1 inter-arrival time is low, the number of tardy orders is
WIP 0.1116 1
increased because of an imbalance between production and
# of con. 0.4000 1
demand rates.

The processing times and demands inter arrival distri-


butions of each product type are given in Table 11. 7 Concluding remarks and future research
Here, there are 18 decision variables (WKim, PKim
where i represents number of stations and m represents Determining the number of kanbans and the container size is
product type ) to be considered. Therefore the total number a critical decision for JIT manufacturing systems. Because of
of kanban combinations was increased to 21,952. Ten the complex nature of such systems, generally simulation
percent of the total combinations was selected for training models are used to evaluate the possible design alternatives.
and testing. The network parameters given in Table 2 were However simulation may become a cumbersome process
also used for the multi-product case. Some accuracy when the number of alternatives to be evaluated is increased.
measures of neural network models are given in Table 12. It is also difficult to evaluate alternative solutions when there
The results show that neural network models are capable of is more than one performance measure to be considered.
estimating the performance measures accurately for multi- Generally, a cost function is used as aggregated weighted
product case either. The optimum kanban allocation for the function, in order to evaluate all possible kanban combina-
multi-product case is given in Table 13. tions with respect to the objective function. However, in real
It should be noted in Table 13 that, in all the solutions, life, construction of the cost function is a difficult task
the number of production kanbans of product A at the because of incommensurable nature of some performance
output buffer of the last station is higher than the others. It measures such as average customer delay time, maximum
is probably due to the output buffer of the last station serves customer delay time.
as a buffer in meeting the demand and the product A has In this study, we proposed a methodology for determin-
the highest demand arrival rate. ing the optimal number of kanbans for each station and the
optimal container size for whole manufacturing system. In
6.1.3 Experiment 3–effect of demand variation this methodology, a backpropagation neural network is
used in order to generate simulation meta-models, and a
In this experiment we considered the effect of the variation multi-criteria decision making technique (TOPSIS) is
of the demand on the performance measures. To study the employed to evaluate kanban combinations considering
effect of this factor on the proposed procedure, the base the priorities of the decision maker. The proposed method-
case (with three station, single product) was considered ology is applied to a case problem and the results are
with two different demand patterns of EXPO(10.25) and presented. We also performed several experiments on
EXPO(12.25). For each instance, the proposed methodol- different types of problems to show the effects of problem
ogy was applied to find the optimum number of kanbans. size, variations in demand and types of products on the
The results are given in Tables 14 and 15. effectiveness of the methodology. The results show that the
As can be seen from the Tables 14 and 15, any decrease proposed methodology can solve the kanban allocation
in the demand inter-arrival time causes an increase in the problem effectively and efficiently.

Table 5 Optimum solutions for different set of weights

Weights Alternatives Criteria

CDT WIP # of con. %ntardy WK1 PK1 WK2 PK2 WK3 PK3 CS %ntardy CDT WIP # of con.

0.25 0.25 0.25 0.25 1 1 1 1 1 3 3 0.142 8.353 11.302 8


0.4 0.2 0.2 0.2 1 1 1 1 1 4 3 0.079 5.087 13.639 9
0.6 0.2 0.1 0.1 1 1 1 4 2 3 2 0.068 3.450 14.180 12
0.8 0.1 0.05 0.05 1 1 1 4 1 3 3 0.024 2.115 19.231 11
1 0 0 0 2 3 1 4 2 3 3 0.013 1.329 30.699 15
974 Int J Adv Manuf Technol (2008) 38:965–977

Table 6 First 6 alternative solutions according to weight structure (0.8, 0.1, 0.05, 0.05) (SR: similarity ratio)

Weights Alternatives Criteria SR

CDT WIP # of % WK1 PK1 WK2 PK2 WK3 PK3 CS % CDT WIP # of
con. ntardy ntardy con.

0.8 0.1 0.05 0.05 1 1 1 4 1 3 3 0.024 2.115 19.231 11 0.931


0.8 0.1 0.05 0.05 1 1 1 3 2 3 3 0.024 2.113 19.246 11 0.931
0.8 0.1 0.05 0.05 1 1 1 3 1 4 3 0.025 2.378 19.086 11 0.928
0.8 0.1 0.05 0.05 1 1 1 4 1 4 3 0.009 1.589 22.029 12 0.924
0.8 0.1 0.05 0.05 1 1 1 4 2 3 3 0.014 1.379 22.254 12 0.924
0.8 0.1 0.05 0.05 1 1 1 3 1 3 3 0.046 3.277 16.414 10 0.921

Table 7 The distributions of processing times for 4 and 5 machine cases

Station1 Station2 Station3 Station4 Station5

4 machine cases Normal (8,1.15) Normal (10,2.75) Normal (6,1.25) Normal (7,1.25) –
5 machine cases Normal (8,1.15) Normal (10,2.75) Normal (6,1.25) Normal (7,1.25) Normal (6,1.25)

Table 8 Design parameters of neural network model for 4 and 5 machine cases

Design parameters 4 machine case 5 machine case

%ntardy CDT WIP %ntardy CDT WIP

Number of hidden Layer 1 1 1 1 1 1


Number of nodes in hidden layer 28 28 18 40 40 20
Learning rate 0.1 0.1 0.1 0.1 0.1 0.1
Momentum rate 0.7 0.7 0.7 0.7 0.7 0.7
Total number of alternative 30,000 30,000 30,000 30,0000 30,0000 30,0000
The number of training set data 2,400 2,400 2,400 4,000 4,000 4,000
The number of test set data 300 300 300 500 500 500
The number of cross validation set data 300 300 300 500 500 500

Table 9 Some measure of accuracy for performance measures

Error measure 4 machine case 5 machine case

%ntardy CDT WIP %ntardy CDT WIP

Mean absolute error 0.039 0.441 0.222 0.003 0.324 0.278


% Absolute error 4.016 3.018 1.77 5.38 10.38 1.75
MSE 0.002 0.002 0.003 0.0001 0.0002 0.0002
RMSE 0.013 0.052 0.055 0.011 0.0133 0.0130
Int J Adv Manuf Technol (2008) 38:965–977 975

# of con
Table 11 The distributions of interarrival time and processing times
for multi-product case

12
14
13
17

13
12
17
19
23
11
Type A Type B Type C
Arrival rate Expo(14.5) Expo(22) Expo(50)

13.855
19.046
20.492
24.715
50.456
12.53
15.85
22.35
34.58
11.07
WIP

Processing times
Station1 Normal Normal Normal
(8,1.15) (5,1.25) (6,1.25)

8.727
5.136
2.872
1.379
0.662
CDT

9.62
7.78
4.93
2.81
0.17
Station2 Normal Normal Normal
(10, 2.75) (4,1.2) (4, 2.25)
Criteria

Station3 Normal Normal Normal

0.081
0.055
0.035
0.021
0.037
0.08
0.05
0.05
0.03
0.11

(6,1.25) (3,1.15) (5, 2.25)


%nt
CS

2
2
2
3
3

2
3
2
2
3 Since the proposed methodology is based on ANNs, it
PK5

inherits all the disadvantages of ANNs. First, the construc-






3
3
4
4
3

tion of training set becomes difficult when the problem size


is increased. Second, determining the optimum levels of
WK5

ANNs parameters such as number of processing elements,






1
1
1
1
2

number of hidden layers, learning rate, momentum rate, etc.


is also difficult. Third, ANN models should be retrained
PK4

4
4
4
3
3

2
1
2
4
3

when the system state parameters change.


Besides these disadvantages, the use of ANNs provide
WK4

quick and accurate estimates for kanban allocation and the


1
1
1
1
2

1
1
2
1
2

use of TOPSIS allows one to evaluate alternative solutions


Table 10 Optimum solutions for different set of weights for 4 and 5 machine cases (%nt:%ntardy)

considering multiple objectives with respect to the decision


PK3

maker’s preferences. So far, the results of the proposed


1
1
1
1
3

1
1
3
4
3

methodology are promising; however, further research is


required in the following directions. First, the production
WK3

system state parameters (such as number of machines,


1
1
4
4
2

1
1
1
1
2

demand rate, etc.) can be considered as input parameters of


PK2

the ANN models and the methodology can be used for re-
1
2
1
1
4

1
1
1
1
4

allocation of kanbans in a dynamic manner. Second,


different multi-criteria decision-making techniques such as
WK2

Promethee, Electre, etc. can be employed instead of


1
1
1
1
1

1
1
1
1
1

TOPSIS. Third, the effectiveness of the proposed method-


ology can be investigated for different kanban based
PK1

production control systems such as CONWIP, hybrid


Alternatives

1
1
1
1
1

1
1
1
1
2

systems, etc. Fourth, the optimum parameters of ANNs


can be determined by using some metaheuristic approaches
WK1

such as genetic algorithms, tabu search, etc.


1
1
1
1
1

1
1
1
1
1
0.25

0.05

0.25

0.05
%nt

0.2
0.1

0.2
0.1
0

0
#of con.

Table 12 Some measure of accuracy for performance measures–


multiproduct case
0.25

0.05

0.25

0.05
0.2
0.1

0.2
0.1
0

Error measure %ntardy CDT WIP


4 machine case

5 machine case
WIP

0.25

0.25
0.2
0.2
0.1

0.2
0.2
0.1

Mean absolute error 0.007 1.153 0.219


0

% absolute error 1.346 2.409 0.652


Weights

MSE 0.0003 4.656 0.066


0.25

0.25
CDT

0.4
0.6
0.8

0.4
0.6
0.8

RMSE 0.016 2.157 0.257


1

1
976 Int J Adv Manuf Technol (2008) 38:965–977

Table 13 Optimum solutions for different set of weights for multi-product case

Weights Alternatives Criteria

CDT WIP # of % WK1A ,PK1A, WK1B, PK1B, WK1C, PK1C, WK2A, PK2A, WK2B, % CDT WIP # of
con. ntardy PK2B, WK2C, PK2C, WK3A, PK3A, WK3B, PK3B, WK3C, PK3C ntardy con.

0.25 0.25 0.25 0.25 1, 1, 1, 1, 1, 1, 1, 3, 1, 1, 1, 1, 1, 3, 1, 1, 1, 1 0.3049 31.221 27.227 22


0.4 0.2 0.2 0.2 1, 1, 1, 2, 1, 1, 1, 3, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1 0.3504 29.465 27.723 22
0.6 0.2 0.1 0.1 1, 1, 1, 3, 1, 1, 1, 3, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1 0.3483 25.375 29.301 23
0.8 0.1 0.05 0.05 1, 1, 1, 3, 1, 1, 1, 3, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1 0.3483 25.371 29.301 23
1 0 0 0 1, 1, 1, 3, 1, 1, 1, 3, 1, 1, 1, 1, 1, 3, 1, 1, 1, 1 0.3047 24.761 30.147 24

Table 14 Optimum solutions for different set of weights for arrival rate 10.25

Weights Alternatives Criteria

CDT WIP # of con. %ntardy WK1 PK1 WK2 PK2 WK3 PK3 CS %ntardy CDT WIP # of con.

0.25 0.25 0.25 0.25 1 1 1 1 1 4 3 0.641 165.852 7.622 9


0.4 0.2 0.2 0.2 1 1 1 4 1 3 3 0.523 130.805 9.813 11
0.6 0.2 0.1 0.1 1 1 1 4 1 4 3 0.498 118.859 11.104 12
0.8 0.1 0.05 0.05 1 1 1 4 2 3 3 0.493 118.817 11.182 12
1 0 0 0 1 3 1 4 2 3 3 0.491 116.825 17.386 14

Table 15 Optimum solutions for different set of weights for arrival rate 12.25

Weights Alternatives Criteria

CDT WIP # of con. %ntardy WK1 PK1 WK2 PK2 WK3 PK3 CS %ntardy CDT WIP # of con.

0.25 0.25 0.25 0.25 1 1 1 2 1 3 2 0.065 2.905 10.119 9


0.4 0.2 0.2 0.2 1 1 1 2 1 3 3 0.027 1.408 15.645 9
0.6 0.2 0.1 0.1 1 1 1 2 1 3 3 0.027 1.408 15.645 9
0.8 0.1 0.05 0.05 1 1 1 3 2 3 3 0.004 0.166 21.637 11
1 0 0 0 1 4 1 4 2 3 3 0.000 0.022 33.161 15
Int J Adv Manuf Technol (2008) 38:965–977 977

References 21. Sarker BR (1990) Service time distributions and the performance
of a pull system. Prod Plan Control 3:155–165
22. Savsar M (1996), Effects of kanban withdrawal policies and other
1. Akturk MS, Erhun F (1999) An overview of design and operational
factors on the performance of JIT systems: a simulation study. Int
issues of Kanban Systems. Int J Prod Res 37:3859–3881
J Prod Res 34:205–214
2. Ramanan GV, Rajendran C (2003) Scheduling in kanban- 23. Andijani A (1997) Trade-off between maximizing throughput rate
controlled flowshops to minimise the makespan of containers. and minimizing system time in kanban systems. Int J Oper Prod
Int J Adv Manuf Technol 21:348–354 Manage 17:429–445
3. Huang CC, Kusiak A (1996) Overview of kanban systems. Int J 24. Sengupta S, Sharief S, Dutta P (1999) Determination of the
Comput Integr Manuf 9:169–189 optimal number of kanbans and kanban allocation in a FMS: a
4. Paris JL, Pierreval H (2001) A distributed evolutionary simulation simulation based study. Prod Plan Control 10:439–447
optimisation approach for the configuration of multiproduct 25. Kochel P, Nielander U(2002) Kanban optimization by simulation
kanban systems. Int J Comput Integr Manuf 14:421–430 and evolution. Prod Plan Control 13:725–734
5. Savsar M, Choueiki MH (2000) A neural network procedure for 26. Shahabudeen P, Krishnaiah, Narayanan MT (2003) Design of a
kanban allocation in JIT production control systems. Int J Prod two-card dynamic kanban system using a simulated annealing
Res 38:3247–3265 algorithm. Int J Adv Manuf Technol 21:754–759
6. Law AM, Kelton WD (1991) Simulation modeling and analysis 27. Aytug H, Dogan CA, Bezmez G (1996) Determining the number
(2nd Edition). Singapore: McGraw-Hill of kanbans: A simulation metamodeling approach. Simulation
7. Fonseca DJ, Navaresse DO, Moynihan GP (2003) Simulation 67:23–32
metamodeling through artificial neural networks. Eng Appl Artif 28. Hurrion RD (1997) An example of simulation optimization using
Intell 16:177–183 a neural network metamodel: finding the optimum number of
8. Kilmer RA, Smith A, Shuman LJ (1997) An emergency kanbans in manufacturing system. J Oper Res Soc 48:1105–1112
department simulation and neural network metamodel. Journal 29. Haykin S (1999) Neural Networks, Prentice-Hall, N. Jersey NY
of the Society for Health Systems 5:63–79 30. Ham FM, Kostanic I (2000) Principles of Neurocomputing for
9. Mitwasi MG, Askin RG (1994) Production planning for a multi- Science & Engineering. McGraw-Hill Inc., USA
item, multi-stage kanban system. Int J Prod Res 32:1173–1195 31. Rumelhart DE, McClelland J; PDP Research Group (1986)
10. Hwang CL, Yoon K (1981) Multiple attribute decision making: Learning internal representations by error propagation. Parallel
methods and applications, a State-of-the-art survey. Springer– Distributed Processing 1:318–362
Verlag, Berlin 32. Sabuncuoglu I, Touhami S (2002) Simulation metamodeling with
11. Kimura O, Terada H (1981), Design and analysis of Pull System, a neural networks: an experimental investigation. Int J Prod Res
method of multi-stage production control. Int J Prod Res 19:241–253 40:2483–2505
12. Bitran GR, Chang L (1987) A mathematical programming approach 33. Pierreval H (1992) Training a neural network by simulation for
to a deterministic Kanban System. Manag Sci 33:427–441 dispatching problems. In Proceedings of the Third Rensselaer
13. Bard JF, Golony B (1991) Determining the number of kanbans in International Conference on Computer Integrated Engineering,
a multiproduct, multistage production system. Int J Prod Res New York, 332–336
29:881–895 34. Pierreval H, Huntsinger R (1992) An Investigation on neural
14. Philipoom PR, Rees LP, Taylor BW (1996) Simultaneously network capabilities as simulation metamodels. In Proceedings of
determining the number of kanbans, container sizes, and the the 1992 Summer Computer Simulation Conference, California,
final-assembly sequence of products in a just-in-time shop. Int J CA, 413–417
Prod Res 34:51–69 35. Badiru A, Siege D (1998) Neural network as a simulation
15. Husseini MS, O’brien C, Hosseini ST (2006) A method to metamodel in economic analysis of risky projects. Eur J Oper
enhance volume flexibility in JIT production Control. Int J Prod Res 105:130–142
Econ 104(2):653–665 36. Araz UO (2005) A simulation based multi-criteria scheduling
16. Wang S, Sarker B (2006) Optimal models for a multi-stage supply approach of dual resource constrained manufacturing systems with
chain system controlled by kanban under just-in-time philosophy. neural Networks, AI2005. Lec Notes Artif Intell 3809:1047–1052
Eur J Oper Res 172:179–200 37. Pomerol JC, Romero SB (2000) Multicriterion decision in
17. Askin RG, Mitwasi MG, Goldberg JB (1993) Determining the management: principles and practice. Kluwer Academic Publish-
number of kanbans in multi-item just-in-time systems. IIE Trans- ers, USA
actions 25:89–97 38. Triantaphyllou E, Lin CT (1996) Development and evaluation of
18. Nori VS, Sarker BR (1998) Optimumum number of kanbans five fuzzy multi-attribute decision making methods. Int J Approx
between two adjacent stations. Prod Plan Control 9:60–65 Reason 14:281–310
19. Tardif V, Maaseidvaag L(2001) An adaptive approach to 39. Pegden CD, Shannon RE, Sadowski RP (1990) Introduction to
controlling kanban systems. Eur J Oper Res 132:411–424 simulation using siman. McGraw-Hill Inc., USA
20. Gupta YP, Gupta MC (1989), A system dynamics model for a 40. Gupta SM, Al-Turki AY (1997), An algorithm to dynamically
multi-stage, multi-line dual-card JIT-Kanban system. Int J of adjust the number of kanbans in stochastic processing times and
Production Research 27:309–352 variable demand environment. Prod Plan Control 8:133–141

You might also like