Escolar Documentos
Profissional Documentos
Cultura Documentos
a r t i c l e
i n f o
Article history:
Received 24 March 2013
Received in revised form 1 December 2013
Accepted 17 December 2013
Available online 25 December 2013
Keywords:
Multiple criteria analysis
Sorting problem
Inventory classication
a b s t r a c t
The ABC method is a well-known approach to classify inventory items into ordered categories, such as A,
B and C. As emphasized in the literature, it is reasonable to evaluate the inventory classication problem
in the multi-criteria context. From this point of view, it corresponds to a sorting problem where
categories are ordered. Here, one important issue is that the weights of the criteria and categorization
preferences can change from industry to industry. This requires the analysis of the problem in a specic
framework where the decision maker (expert)s preferences are considered. In this study, the preferences
of the decision maker are incorporated into the decision making process in terms of reference items into
each class. We apply two utility functions based sorting methods to the problem. We perform an experiment and compare results with other algorithms from the literature.
2014 Elsevier Ltd. All rights reserved.
1. Introduction
The ABC method has been used by business rms to manage a
huge number of distinct items, referred to as stock keeping units
(SKUs). As the size of the inventory increases, controlling the items
needs time and additional expenditure. A reasonable strategy in
this case is to classify the SKUs in terms of inspection priority.
Therefore, the SKUs with high priority may be inspected more frequently to prevent stock-out cases and the resulting losses, while
the SKUs with low priority may be inspected less frequently in order to reduce inspection costs. This tradeoff reveals that each SKU
should have an inspection priority to minimize total inventory
costs. Inventory classication is not only related to inventory management but also related to production planning. Selection of the
appropriate production planning technique should also be considered together with the inventory classication.
Traditionally, the ABC analysis classies items into three categories, namely, A (very important), B (moderately important) and
C (least important) according to the annual usage value. Here,
the priority criterion is the annual usage value which determines
the class of the SKU. Since closely monitoring all items is too costly,
a suitable inventory management policy is to concentrate more on
class A items than on classes B and C items. Here, class C items also
receive a more relaxed control policy than class B items (Silver,
Pyke, & Preterson, 1998). However, authors agree that other
characteristics of the inventory like criticality, lead time, ordering
cost, commonality, repairability, and durability may affect and
q
change the class of items (Flores & Whybark, 1987; Gvenir & Erel,
1998; Partovi & Anandarajan, 2002; Ramanathan, 2006). To take
into account these characteristics we require multiple criteria decision analysis. Evaluating the problem in the multi-criteria context
may also improve inventory investments (see Bhattacharya, Sarkar, & Mukherjee, 2007, for a comparison study of traditional and
multi-criteria inventory classication via simulation). This problem has already been addressed in the literature. Recently, Ramanathan (2006) proposed a weighted linear optimization model
where an LP model maximizing the weighted sum of criteria for
the considered inventory item is solved and the resulting weights
are assigned as favorable weights of this item. After that, Zhou and
Fan (2007) presented an extended version of Ramanathans model.
Another weighted linear optimization model was given by Ng
(2007) and Hadi-Vencheh (2010).
One different study about this problem is the case-based distance model by Chen, Kevi, Marc Kilgour, and Hipel (2008). This
was originally a sorting procedure, which classies actions into ordered classes. The decision maker (DM) assigns in advance some
reference items into each class, and a mathematical model determines the thresholds and weights. (Fuzzy) AHP based approaches
have also been presented for this problem in the literature (Partovi
& Hopton, 1993; akr & Canbolat, 2008). Genetic algorithms
(Gvenir & Erel, 1998), particle swarm optimization (Tsai & Yeh,
2008), fuzzy classication (Chu, Liang, & Liao 2008; Keskin & zkan, 2013), TOPSIS (Bhattacharya et al., 2007) and neural networks
(Partovi & Anandarajan, 2002) have also been applied to this
problem.
All the studies mentioned above present different perspectives
for the multi-criteria inventory classication. Despite the advantages of these methods, it should be noted that the importance of criteria and the categorization may change from industry to industry
13
and even from company to company. Several authors also emphasize this issue (Taylor, Sewart, & Bolander, 1981; Kampen, Akkerman, & Donk, 2012). For instance, the lead time criterion is more
important in the drug industry than in the furniture industry.
The adaptation of a methodology to these differences might be
possible by adjusting the problem parameters in accordance with
the decision maker (DM)s judgment. The DMs judgment can be
obtained directly by assigning values to the problem parameters
like weights, thresholds, etc. Alternatively, this information can
be indirectly inferred from some classication examples (see for
instance, Mousseau & Slowinski, 1998; Soylu, 2011). The latter
has some advantages since it is more understandable by the DM
and more agreeable in the case of multiple DMs.
In this study, we apply UTADIS (Doumpos & Zopounidis, 2004)
based sorting methods to the multi-criteria ABC inventory classication problem. We assume that the DM assigns in advance some
reference items for each class. It is assumed that these assignments
actually reect the characteristics of the industry involved. That
means the weights given to criteria in this industry can be inferred
from these category examples. A mathematical model constructs a
utility function and determines thresholds between classes with
respect to reference items. The aim of the model is to minimize
the total classication error over reference items. That means the
corresponding utility value of an item should place it in its correct
class otherwise the classication error occurs. Based on this information, unclassied items are placed manually. We evaluate two
function types, which are linear and piece-wise linear utility
functions.
The paper is organized as follows. We explain the utility function based approaches in Sections 2 and 3. We present computational results in Section 4. We conclude with a discussion and
further research directions in Section 5.
zji
If j 2 C Ref
then uA > Uzj P uB
B
If j 2 C Ref
then uB > Uzj
C
The following LP determines the utility function parameters, w, and
thresholds, uA and uB.
P
LP 1 :
Min f
ej
8j2C ref
A
jC ref
A j
ej
8j2C ref
B
jC ref
B j
ej
ej
8j2C ref
C
jC ref
C j
m
X
wi zji uA ej P 0 8j 2 C ref
A
2
3
i1
m
X
wi zji uB ej P 0 8j 2 C ref
B
i1
m
X
i1
m
X
5
6
i1
m
X
wi 1
i1
uA uB P s
wi P 0 8i 1; 2; . . . ; m
uA ; uB P 0
Z ji
If j 2 C Ref
then Uzj P uA
A
ej ; ej P 0 8 j 1; 2; . . . ; n
^zji zmin
i
zmax
zmin
i
i
In other words, if the utility of an SKU is in the rst range, this SKU
can be placed either in CA or CB. We will call this set as CA/CB.
Similarly, if the utility of an SKU is in the second range, this SKU
14
ui zji
wip
p1
r
zji g i ji
r ji 1
r
gi
g i ji
wirji
h r r 1 i
to which the zji belongs, and
where rji is the subinterval g i ji ; g i ji
p1
p
wip ug i g i .
The global utility function of alternative zj is given in the following equation:
Uzj
m
X
ui zji
10
i1
can be placed either in CB or CC. We will call this set as CB/CC. One of
the following two strategies can be considered in this case.
Strategy 1: Select the middle of the range as a threshold. This
strategy divides the mid region into two equal parts.
uA
uB
2
min8j2C Ref fuzj g max8t2CRef fuzt g d
B
The classication of the non-reference alternatives is then performed by comparing U(zj) with thresholds ud , d 2; 3; . . . ; h 1 as
follows:
If
Uzj P u1 then j 2 C 1
If
If
m
X
wi zji
UTADIS :
Min f
i1
jC ref
A j
ej
8j2C ref
B
jC ref
B j
ej
ej
8j2C ref
C
jC ref
C j
11
s:t:
Uzj uA ej P 0 8j 2 C ref
A
12
C ref
B
13
ej
ej
ej
Uz uB
Uz uA
Uz uB
8j 2 S
ej
8j2C ref
A
P 0 8j 2
6 d 8j 2
C ref
B
14
6 d 8j 2
C ref
C
15
bi
m X
X
wip 1
16
i1 p1
uA uB P s
17
If Uz < uB then j 2 C C
ui
u i (g ibi +1)
ui ( gi2 )
ui ( gi1 )
0.0
g i1
gi2
b +1
gi i
Zi
15
u1
u2
ER
P
j
# of misclassified SKUs
8j2S x
150
150
8
1;
if
the
class
of
an
SKU j and the class
>
<
where xj
u h1
wip P 0 8i 1; 2; . . . ; m; 8p 1; 2; . . . ; bi
ej ; ej P 0 8j 1; 2; . . . ; n
uA ; uB P 0
d 0:0001; s 0:0002 user defined small positiv e constants
The explanation of constraints is the same with LP1 except for the
utility function denition.
4. Computational results
In the rst part of this section, we design an experiment to analyze the classications by the R-model (Ramanathan, 2006), ZFmodel (Zhou & Fan, 2007), case-based model (Chen et al., 2008),
linear utility model and UTADIS (Doumpos & Zopounidis, 2004).
The brief review of rst three methods is given in Appendix A.
When running the UTADIS method, we use the Heuristic 2 to dene the number of subintervals (the reader may refer to Doumpos
& Zopounidis, 2004 for the denition of Heuristic 2). We used
Cplex 12.1 solver (Cplex, 2010) for the solution of all models. In
the second part of this section, we also apply the utility function
based models to the ABC classication problem given in the literature and discuss results.
4.1. Experimental analysis
For our experiments, we consider a system with 150 SKUs, and
we select three evaluation criteria, which are annual usage, average unit cost and lead time, from the literature. The criteria values
of each SKU in each class are generated articially from lognormal
distribution with the following mean and standard deviation.
Class CA ! lA 4;
rA 4
Class CB ! lB 2; rB 2
Class CC ! lC 1; rC 1
>
:
17
0;
Otherwise
Since the linear utility model, UTADIS and the case-based model
require reference SKUs, all 150 SKUs are also assumed as references for these algorithms.
Table 1 presents the mean error rate and standard deviation results of each algorithm over 30 runs. We also test different k levels
of the ZF-model. According to these results UTADIS gives the best
performance in both sets. In the ZF-model, k = 0.25 is better than
other k levels. For further statistical comparisons, we apply the
paired-t test to analyze whether the mean ER of UTADIS is signicantly different from the mean ER of another method. Actually, we
test the following hypothesis.
H0 : lUTADIs lMethod
H1 : lUTADIs lMethod
In both sets, p-values are less than 0.05, which means the H0
hypothesis is rejected, concluding that UTADIS is signicantly better than the others except for the ZF-model k 0:50 in Set 2.
When we compare Set 1 and Set 2, we observe that the error rate
decreases signicantly in Set 2. This is because in Set 2 the cardinalities of classes A and B decrease but the cardinality of class C increases. As the sizes of two sets decrease, the probability of
misclassications due to classes A and B also decreases, i.e. misclassications due to A ? B, B ? A, A ? C, B ? C. However, since
the size of the class C increases, the probability of misclassications (due to C ? B, C ? A) increases. The decrement in misclassications is higher than the increment.
Additionally, we perform an analysis on inventory costs and ll
rate as presented in Teunter, Babai, and Syntetos (2010) and Lajili,
Babai, and Ladhari (2012). Here it is assumed that the system is
controlled with an (s, Q) policy. We compare the total holding cost
and ll rate for the classication of each method. The following
notation (Lajili et al., 2012) and corresponding settings are used
in inventory cost calculations.
N
n
Dj
rj
Lj
Qj
CSLj
kj
e 2 kj 1 Ukj
Gkj p1
2p
(continued on next page)
16
Table 1
Error rate results of methods.
Method
Set 1
Set 2
Mean
Std. dev.
p-Value
Mean
Std. dev.
p-Value
R-model
0.36263
0.02744
0.000
0.27683
0.03208
0.000
ZF-model
k = 0.0
k = 0.25
k = 0.50
k = 0.75
Case-based
Linear U.
UTADIS
0.34290
0.33460
0.34087
0.34647
0.39020
0.32890
0.25267a
0.03463
0.03273
0.03342
0.02867
0.04069
0.03890
0.03508
0.000
0.000
0.000
0.000
0.000
0.000
0.25107
0.24330
0.24737
0.25847
0.2779
0.30390
0.23197a
0.02927
0.03395
0.03800
0.03308
0.06270
0.04808
0.05134
0.000
0.022
0.103
0.039
0.001
0.000
FRj
FRT
rj
p
Lj
Qj
Gkj
PN
FRj Dj
j1
Overall ll rate of the inventory system. FRT P
N
j1
hj
Wj
CT
Dj
zj1 5:12
Uz 0:07872
210 5:12
!
!
zj2 25:38
zj 1
0:40999 3
5840:64 25:38
71
!
j
z 0:01
0:41969 4
1:0 0:01
0:09161
18
Table 2
Inventory cost and ll rate results of methods.
Method
Inventory cost
Fill rate
Set 1
Set 2
Set1
Set 2
Mean
Std. dev.
Mean
Std. dev.
Mean
Std. dev.
Mean
Std. dev.
R-model
12046
1505
4703
600
0.98627
0.00091
0.98157
0.00097
ZF-model
k = 0.00
k = 0.25
k = 0.50
k = 0.75
12019
12072
12085
12069
1552
1536
1522
1511
4682
4728
4730
4719
646
638
614
612
0.98637
0.98657
0.98663a
0.98637
0.00113
0.00097
0.00085
0.00093
0.98163
0.98190
0.98170
0.98163
0.00103
0.00076
0.00095
0.00096
Case-based
Linear U.
UTADIS
11548a
12098
12000
1595
1538
1520
4662a
5069
4920
649
679
656
0.98390
0.98363
0.98613
0.00271
0.00299
0.00148
0.98120
0.98187
0.98417a
0.00262
0.00232
0.00109
17
1
2a
3
4c
5
6
7
8
9
10b
11
12
13a
14
15
16
17
18
19
20
21
22
23
24
25c
26
27c
28
29b
30
31
32
33
34c
35
36b
37
38
39
40
41
42
43
44
45b
46
47
Lead time
Critical factor
AHP
R-model
ZF (0.5)
Case based
Linear U.
UTADIS
49.92
210
23.76
27.73
57.98
31.24
28.2
55
73.44
160.5
5.12
20.87
86.5
110.4
71.2
45
14.66
49.5
47.5
58.45
24.4
65
86.5
33.2
37.05
33.84
84.03
78.4
134.34
56
72
53.02
49.48
7.07
60.6
40.82
30
67.4
59.6
51.68
19.8
37.7
29.89
48.3
34.4
28.8
8.46
5840.64
5670
5037.12
4769.56
3478.8
2936.67
2820
2640
2423.52
2407.5
1075.2
1043.5
1038
883.2
854.4
810
703.68
594
570
467.6
463.6
455
432.5
398.4
370.5
338.4
336.12
313.6
268.68
224
216
212.08
197.92
190.89
181.8
163.28
150
134.8
119.2
103.36
79.2
75.4
59.78
48.3
34.4
28.8
25.38
2
5
4
1
3
3
3
4
6
4
2
5
7
5
3
3
4
6
5
4
4
4
4
3
1
3
1
6
7
1
5
2
5
7
3
3
5
3
5
6
2
2
5
3
7
3
5
1
1
1
0.01
0.5
0.5
0.5
0.01
1
0.5
1
0.5
1
0.5
1
0.5
0.5
0.5
0.5
0.5
1
0.5
1
1
0.01
0.01
0.01
0.01
0.01
0.01
0.5
1
0.01
0.01
0.01
1
0.01
0.5
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
A
A
A
C
B
C
C
C
A
B
B
B
A
B
A
C
B
A
B
B
A
B
A
A
C
C
C
C
B
C
B
B
C
C
C
B
C
C
C
C
C
C
C
C
B
C
C
B
C
C
A
C
C
C
C
C
C
C
C
C
A
A
A
B
A
A
B
B
B
B
C
C
C
B
A
A
C
B
C
B
A
C
C
B
C
B
B
C
C
B
C
A
C
B
A
A
A
C
B
C
C
B
A
A
C
B
A
A
C
C
C
A
B
B
C
B
B
C
C
C
C
A
A
C
B
C
B
B
C
C
B
C
B
B
C
C
C
C
B
C
C
A
A
A
C
B
C
C
C
A
B
B
C
A
C
B
C
C
B
C
C
B
C
B
B
C
C
C
C
B
C
C
B
C
C
C
B
C
C
C
C
C
C
C
C
B
C
C
A
A
A
C
B
B
B
C
A
B
B
B
A
B
B
B
B
B
B
B
B
B
B
B
C
C
C
C
B
C
B
B
C
C
C
B
C
B
C
C
C
C
C
C
B
C
C
A
A
B
C
C
C
C
C
B
B
B
C
A
B
B
C
C
B
C
C
B
B
B
B
C
C
C
B
B
C
B
B
C
C
C
B
C
C
C
C
C
C
C
C
B
C
C
C Ref
for case based model, linear utility model and UTADIS.
A
C Ref
for case based model, linear utility model and UTADIS.
B
C Ref
for case based model, linear utility model and UTADIS.
C
18
Table 4
Classication results (average of 5-fold cross-validation, %).
AHP
LP1
A
Avg.
Min.
Max.
Avg.
Min.
Max.
Avg.
Min.
Max.
A
B
C
100
5
0
100
0
0
100
25
0
0
95
4
0
80
0
0
100
2
0
0
96
0
0
80
0
0
100
Avg. accuracy
97.0
50
0
0
100
30
0
0
70
13
0
50
0
0
100
50
10
17
87
0
0
50
50
50
100
UTADIS
A
B
C
90
13
0
Avg. accuracy
82.3
solution to the LP1 or UTADIS model and all SKUs are classied
according to the optimal solution. If the following modications
are done in the reference set, the nal classication will not change.
If a new reference SKUj e Cn is added to the C ref
(for instance,
n
SKU11 e CB is added to the C ref
),
then
the
nal
classication
will
B
not change.
Remark 1. The optimal solution will always satisfy the new constraint added to the LP1 or UTADIS.
utility functions depending on the industry. Different utility functions should be considered and evaluated.
Acknowledgement
This study was supported by the Erciyes University, Scientic
Research Unit under Grant No. BAP-FBY-11-3457.
Appendix A
R-model (Ramanathan, 2006): assigns favorable weights for
each SKU by using the following model. The objective function of
the model is the inventory score (IS(j)) of the corresponding SKUj
(IS(j)). SKUs are classied by selecting user-dened thresholds.
Max ISj
m
X
wi zji
19
8j 2 S
i1
s:t:
m
X
wi zji 6 1 8j 2 S
i1
wi P 0 8i 1; 2; . . . ; m
ZF model (Zhou & Fan, 2007): Least favorable weights are also computed by another model in addition to R-model. Finally, a combined
index CIk j is computed.
Min BIj
m
X
wbi zji
8j 2 S
i1
s:t:
m
X
wbi zji P 1 8j 2 S
i1
wbi P 0 8i 1; 2; . . . ; m
CIk j k
ISj IS
BIj BI
1 k
IS IS
BI BI
Fig. A3. The joint matrix for classications according to A+ and A.
20
Mousseau, V., & Slowinski, R. (1998). Inferring an ELECTRE TRI model from
assignment examples. Journal of Global Optimization, 12, 157174.
Ng, W. L. (2007). A simple classier for multiple criteria ABC analysis. European
Journal of Operational Research, 177, 344353.
Partovi, F. Y., & Anandarajan, M. (2002). Classifying inventory using an articial
neural network approach. Computers & Industrial Engineering, 41, 389404.
Partovi, F. Y., & Hopton, W. E. (1993). The analytic hierarchy process as applied to
two types of inventory problems. Production and Inventory Management Journal,
35(1), 1319.
Ramanathan, R. (2006). ABC inventory classication with multiple-criteria using
weighted linear optimization. Computers & Operations Research, 33, 695700.
Silver, E. A, Pyke, D. F, & Preterson, R. (1998). Inventory management and production
planning and scheduling. John Wiley & Sons.
Soylu, B. (2011). A multi-criteria sorting procedure with Tchebycheff utility
function. Computers & Operations Research, 38(8), 10911102.