ISTANBUL TECHNICAL UNIVERSITY
INSTITUTE OF SCIENCE AND TECHNOLOGY

MULTI CRITERIA ABC CLASSIFICATION OF LIGHT RAILWAY SPARE

PARTS USING ARTIFICIAL NEURAL NETWORK APPROACH

M.Sc. Thesis by
Jarkyn JUMABAEVA

Department: Management Engineering

Programme: Management Engineering

JUNE 2011
ĐSTANBUL TECHNICAL UNIVERSITY
INSTITUTE OF SCIENCE AND TECHNOLOGY

MULTI CRITERIA ABC CLASSIFICATION OF LIGHT RAILWAY SPARE

PARTS USING ARTIFICIAL NEURAL NETWORK APPROACH

M.Sc. Thesis by
Jarkyn JUMABAEVA
(507071030)

Date of submission : 06 May 2011

Date of defence examination: 06 June 2011

Supervisor (Chairman) : Prof. Dr. Demet BAYRAKTAR (ITU)

Members of the Examining Committee : Assoc. Prof. Dr.Đlker TOPCU (ITU)
Asst. Prof. Dr. Dilay ÇELEBĐ (ITU)

JUNE 2011
ĐSTANBUL TEKNĐK ÜNĐVERSĐTESĐ
FEN BĐLĐMLERĐ ENSTĐTÜSÜ

HAFĐF RAYLI SĐSTEMLERDE YEDEK PARÇA STOKLARININ

SINIFLANDIRILMASI ĐÇĐN ÇOK ÖLÇÜTLÜ ABC ANALĐZĐ ĐLE YAPAY
SĐNĐR AĞI YAKLAŞIMI

YÜKSEK LĐSANS TEZĐ

Jarkyn JUMABAEVA
(507071030)

Tezin Enstitüye Verildiği Tarih : 06 Mayıs 2011

Tezin Savunulduğu Tarih : 06 Haziran 2011

Tez Danışmanı : Prof. Dr. Demet BAYRAKTAR (ĐTÜ)

Diğer Jüri Üyeleri : Doç.Dr. Đlker TOPCU (ĐTÜ)
Yrd. Doç. Dr. Dilay Çelebi (ĐTÜ)

HAZĐRAN 2011
FOREWORD

I am very grateful to Prof. Dr. Demet Bayraktar for her supervision and helpful
directions throughout the study. I would also like to express my deep appreciation
and thanks forAsst.Prof.Dr. Dilay Çelebi for helping with gathering data for this
thesis.
I also thankful to my husband and family for their continuous support and
encouragement during the preperation of this thesis.

June 2011 Jarkyn Jumabaeva

Management Engineering

v
vi
TABLE OF CONTENTS

Page

TABLE OF CONTENTS......................................................................................... vii

ABBREVIATIONS ................................................................................................... ix
LIST OF TABLES .................................................................................................... xi
LIST OF FIGURES ................................................................................................ xiii
SUMMARY .............................................................................................................. xv
ÖZET....................................................................................................................... xvii
1. INTRODUCTION .................................................................................................. 1
1.1 Purpose of the Thesis ......................................................................................... 1
1.2 Research Scope .................................................................................................. 3
1.3 Research Methodology....................................................................................... 4
2. INVENTORY MANAGEMENT .......................................................................... 5
2.1 Inventory Classification ..................................................................................... 5
2.2 ABC Analysis .................................................................................................... 8
2.3 Multi Criteria ABC Analysis ........................................................................... 10
3. AN ARTIFICIAL NEURAL NETWORK ......................................................... 13
3.1 Application Areas Of Artificial Neural Networks ........................................... 14
3.2 Benefits of Neural Networks ............................................................................ 15
3.3 Artificial Neural Network Structure: An Explanation Based
onBiologic Nervous System ............................................................................. 16
3.4 Learning Process in Artificial Neural Networks .............................................. 20
3.4.1 Supervised learning ................................................................................... 21
3.4.2 Unsupervised learning............................................................................... 22
3.4.3 Reinforcement learning ............................................................................. 22
3.5 Multi-Layer Perception (MLP) and Back Propagation
Algorithms ........................................................................................................ 23
3.5.1 Multi layer perception ............................................................................... 23
3.5.2 Backpropagation algorithm ....................................................................... 24
3.6 Types of Activation Functions ......................................................................... 27
4. LITERATURE REVIEW.................................................................................... 31
4.1 Inventory Classification Techniques ................................................................ 31
4.2 Multi Criteria Decision Making (MCDM) ....................................................... 34
4.3 Analytical Hierarchy Process (AHP) ............................................................... 34
4.4 Data Envelopment Analysis ............................................................................. 36
4.5 Artificial Neutral Network ............................................................................... 37
4.6 Genetic Algorithms .......................................................................................... 41
4.7 Other Multi Criteria Classification Techniques Widely Used in
Abc Analysis ..................................................................................................... 42
4.7.1 Ng model ................................................................................................... 42
4.7.2 Case based model for multi criteria abc classification .............................. 42
4.7.3 Abc fuzzy classification ............................................................................ 43

vii
5. AN ARTIFICIAL NEURAL NETWORK MODEL FOR ABC
CLASSIFICATION OF SPARE PARTS IN RAILWAY SYSTMES ............ 45
5.1 Identification of Evaluation Criteria for ANN Model ...................................... 45
5.2 Network Building ............................................................................................. 50
5.2.1 Network structure ...................................................................................... 51
5.2.2 Multi layer feedforward networks ............................................................. 51
5.2.3 Learning rule in neural network ................................................................ 52
5.2.4 Choice of transfer function ........................................................................ 53
5.2.5 Determination of hidden layer number and nodes in hidden layer ........... 54
5.2.6 The stopping criteria, learning rate and momentum ................................. 55
5.3 Neurosolutions 5.0 Software Used for Classification Model ........................... 57
5.4 Training Process by Neurosolutions ................................................................ 58
5.4.1 Five-fold cross validation method ............................................................. 58
5.4.2 Training proposed artificial neural network model ................................... 60
5.5 Testing of the Neural Network ......................................................................... 61
6. CONCLUSION AND RECOMMENDATIONS ............................................... 65
REFERENCES ......................................................................................................... 67
APPENDICES .......................................................................................................... 73
CURRICULUM VITAE .......................................................................................... 93

viii
ABBREVIATIONS

ABC-FC : ABC–Fuzzy Classification

AHP : Analytical Hierarchy Process
ANN : Artificial Neural Network
App : Appendix
BPA : Back Propagation Algorithm
DEA : Data Envelopment Analysis
EFNN : Enhanced Fuzzy Neural Network
FAHP : Fuzzy Analytical Hierarchy Process
GA : Genetic Algorithm
k-NN : k-nearest neighbour
MBPN : Moving Back-propagation neural network
MC : Multi criteria
MCDA : Multi criteria decision aid
MCDM : Multi Criteria Decision Making
MDA : Multiple Discriminant Analysis
METRIC : Multi-echelon technique for recoverable item control
MFNN : Moving Fuzzy Neuron Network
MLP :Multi Layer Perception
MSE : Mean Squared Error
PE : Processing Element
RCM : Reliability Centered Maintenance
SVM : Support Vector Machine
VLSI :Very-large-scale-integrated

ix
x
LIST OF TABLES

Page

Table 2.1: ABC analysis grouping .............................................................................. 8

Table 4.1: Academic researches about multi criteria ınventory classification .......... 32
Table 5.1: Subject criteria definitions (Çakır and Canbolat, 2008) .......................... 46
Table 5.2: Dimension of evaluation criteria for classification of spare parts ........... 50
Table 5.3: Network parameteres for the ann model .................................................. 57
Table 5.4: Average minimum mse for training and cross validation
data with hyperbolic tangent function ..................................................... 59
Table 5.5: Average minimum mse for training and cross validation data with
sigmoid function ...................................................................................... 60
Table 5.6: Results of training for hyperbolic tangent and sigmoid transfer functions
................................................................................................................. 61
Table 5.7: Data for 11 spare part used for testing of the network ............................. 62
Table 5.8: Results of testing process for proposed neural network .......................... 62
Table 5.9: Output matrix of the neural network for 11 testing data .......................... 63
Table 5.10: Performance of classification of network for 11 testing
data according to 3 desired output sets, a, b and c .................................. 63
Table 5.11:Desired data and output of neural network ............................................. 64
Table A.1: Evaluation data for spare parts in excel file ............................................ 75
Table A.2: Randomized training data divided into 5 disjoint subsets....................... 77
Table B.1: Minimum mse for training data for five subsets with
different hidden node numbers ................................................................ 79
Table B.2: Minimum mse for cross validation data for five subsets
with different hidden node number ......................................................... 82
Table B.3: Calculation of average minimum mean squared error for
two node of cross validation data ............................................................ 85

xi
xii
LIST OF FIGURES

Page

Figure 3.1: Block diagram representation of nervous system(Haykin, 1999) .......... 17

Figure 3.2: Block diagram representation of ann architecture .................................. 17
Figure 3.3: Structure of a typical neuron (Partovi and Anandarajan, 2002) ............. 18
Figure 3.4: Model of a typical processing element (Haykin, 1999) ......................... 19
Figure 3.5: Multi-layer perception model ................................................................. 24
Figure 3.6: Threshold function.................................................................................. 27
Figure 3.7: Piecewise-linear function ....................................................................... 28
Figure 3.8: Sigmoid function .................................................................................... 28
Figure 5.1: Artificial neural network structure (Partovi and
Anandarajan, 2002) ................................................................................. 52
Figure 5.2: Variation of average minimum mse on training dataset
with step size of hidden layer. ................................................................. 56
Figure 5.3: Variationof average minimum mse on training dataset
with momentum output layer. ................................................................. 56
Figure 5.4: Proposed artificial neural network model............................................... 61
Figure B.1: Selection of input and desired parameters in excel................................ 87
Figure B.2: Neurobuilder tool ................................................................................... 88
Figure B.3: Selection of hidden layer number in neurosolutions ............................. 88
Figure B.4: Selection of training parameters for hidden layer in neurosolutions ..... 89
Figure B.5: Selection of training parameters for output layer in neurosolutions ...... 90
Figure B.6: Selection of performace criteria in neurosolutions ................................ 90
Figure B.7: Setup for probe configuration in neurosolutions ................................... 91
Figure B.8: Neurosolutions network structure .......................................................... 92
Figure B.9: Training process in neurosolutions ........................................................ 92

xiii
xiv
MULTI CRITERIA ABC CLASSIFICATION OF LIGHT RAILWAY SPARE
PARTS USING ARTIFICIAL NEURAL NETWORK APPROACH

SUMMARY

Management of spare parts inventories is crutial for successfull execution of the

maintenance processes. Main attention is paid to maintenance operations in railway
systems management because it directly affects the performance of railway machines
and play important role in railway service, in order to provide uninterrupted and high
quality service to passengers. The most important characteristics of railway services
are the availability and reliability of spare items, which are necessary for
maintenance operations. The role of spare parts is to support the maintenance
operations in keeping vehicles in a functioning condition. Supplying an adequate and
yet efficient amount of spare parts to support the maintenance activities in railway
service is a tough management problem.
In this thesis, artificial neural network model was developed to classify spare parts
inventory based on multi criteria ABC analysis in light railway systems. After
thorough literature review on inventory classification, the relevant classification
criteria and control chracteristics of maintenance spare parts are identified and
discussed in terms of their effects on maintenance operations, purchasing
characteristics, positioning of materials, responsibility of control and control
principles. The classification model of railway systems spare parts have been
performed based on following criteria: annual unit cost, lead time, value usage,
substitutability, commonality and stock-out penalty.
Most widely used neural network structure is chosen, which contains one input, one
hidden and one output layer. For the training of the network which will be used for
classification of items into A, B and C class items, a multi layer perception network
with back propagation algorithm was used. The optimum architechture is determined
by estimating number of nodes in hidden layer by five-fold cross validation method.
The comparison of classification accuracy was done between neural networks with
hyperbolic tangent function and sigmoid transfer function. As a result of training,
neural network with hyperbolic tangent function gave more accurate results, by
giving the lowest the minimum mean squared error, MSE. The proposed artificial
neural network architecture has 6-8-3 structure, which constitutes six nodes in the
input layer, eight nodes in the hidden layer and three nodes in the output layer.
The evaluation data of 71 maintenance spare part items were utilized for training and
testing processes, specifically, 60 item data for training and 11 items data for testing
the developed artificial neural network.

xv
The main contribution of this thesis is that, it applies a different approach, which is
ANN, for classification of spare parts inventory in railway systems, which was not
widely investigated area in the literature. The previous researches in literature on
multi criteria inventory management have been done using analytical hierarchy
process and data envelopment analysis). The developed network has shown good
classification accuracy and this network can be used for classification of other real-
world inventory data in railway systems maintenance.

xvi
HAFĐF RAYLI SĐSTEMLERDE YEDEK PARÇA STOKLARININ
SINIFLANDIRILMASI ĐÇĐN ÇOK ÖLÇÜTLÜ ABC ANALĐZĐ ĐLE YAPAY
SĐNĐR AĞI YAKLAŞIMI

ÖZET

Bakım işlerinin başarılı bir şekilde gerçekleşmesi için yedek parça stoklarının
yönetimi çok önemlidir. Raylı sistemlerin yönetiminde bakım işlemlerine büyük
önem verilmesinin sebebi, bakım işlemlerinin raylı sistemlerde kullanılan araçların
performasını direk etkilemesi ve yolculara kesintisiz ve yüksek kalitede hizmet
sağlanmasında çok önemli rol oynamaktadır. Raylı sistemlerin hizmetinde en önemli
iki husus, bakım işleri için gerekli olan yedek parçaların yeterli miktarda elde
bulundurulması ve güvenilir olmasıdır. Yedek parçaların amacı araçların çalışır
vaziyette tutulması için gerekli bakım işlerinde kullanılması. Raylı sistemlerin
hizmetinde yapılan bakım işleri için yeterli ve aynı zamanda etkin bir şekilde yedek
parça sağlamak, yönetim için zor bir iştir.
Bu çalışmanın amacı hafif raylı sistemlerde kullanılan yedek parça stoklarının çok
kriterili ABC analizi ile sınıflandırılması için yapay sinir ağı modeli geliştirmektir.
Envanter sınıflandırma üzerine kapsamlı literatür incelemesi yapıldıktan sonra,
bakım için kullanılan yedek parçaların sınıflandırılma ölçütleri ve kontrol
parametreleri belirlendi. Belirlenen sınıflandırma ölçütlerin bakım işleri, satın alma
özellikleri, malzemenin konulmandırması, kontrol sorulluluğu ve denetim prensipleri
açılarından tartışıldı. Hafif raylı sistemlerin yedek parça sınıflandırma modelinde
değerlendirme süreci için elde bulundurmama maliyeti, ikame edilebilirlik, kullanım
adedi, maliyet, tedarik süresi ve ortak nokta kriterileri belirlenmiştir.
En yaygın kullanılan, tek girdi, tek gizli ve tek çıktı katmanlı, yapay sinir ağı modeli
seçilmiştir. Yedek parçaların A,B ve C gruplarına sınıflandırmasın için kullanılan
ağın eğitililmesi için çok katmanlı algılama ağ modeli ile geri yayılım
algoritmasından yararlanılmıştır. Uygun ağ yapısını blirlenebilmesi için beşli çapraz
doğrulama yöntemi ile gizli katmandaki sinir hücre sayısı tespit edilmiştir.
Sınıflandırma doğrulaması için, hiperbolik tanjantlı ve sigmoid transfer fonsiyonlu
iki farklı yapay sinir ağı modeli karşılaştırılmıştır. Ağın eğitimi sonucunda,
hiperbolik tanjant fonksiyonlu yapay sinir ağının daha düşük hata oranı gösterdiği,
ortalama karesel hatadan belirlenmiştir. Önerilen YSA modelinde; yedek parça
envanteri sınıflandırma kriterilerini temsil eden 6 adet girdi sinir hücresi, tek gizli
katmanda yer alan 8 adet sinir hücresi ve sınıflandırma kümelerini gösteren 3 çıktılı
ileri beslemeli bir ağ yapısı oluşturulmuştur.

xvii
Değerlendirmede eğitim ve test işlemleri için toplam 71 adet yedek parça verisi
kullanılmıştır, bunların, 60 adet yedek parça verisi eğitim işlemi ve 11 adet yedek
parça verisi test işlemi için değerlendirilmiştir.
Yapılan çalışmanın en önemli katkısı, hafif raylı sistemlerin bakım işlerinde
kullanılan yedek parça envanter sınıflandırması çalışmasına farklı yöntemle
yaklaşılmış olması ve bu alanın yapılan çalışmalar arasında daha önce incelenmemiş
olması. Daha önce yapılan çalışmalar arasında çok kriterili ABC analizi ile hafif
raylı sistemlerin yedek parça stokların sınıflandırmasında analitik hiyerarşi prosesi ve
veri zarflama analizi yöntemleri kullanılarak incelenmiştir. Bu çalışmada, hafif raylı
sistemlerinde kullanılan envanter sınıflandırması için yapay sinir ağı yaklaşımı
önerilmiştir. Yedek parça sınıflandırma modeli olarak geliştirilen YSA modeli
başarılı sonuçlar göstermiştir. Geliştirilen YSA modeli diğer raylı sistemlerin de
bakım işleri için kullanabileceği bir yapıya dönüştürülebilir.

xviii
INTRODUCTION

Inventory management is one the main objectives of production and operations

management. The role of inventory managers is to maintain sufficient inventoires to
meet demand and achieve productivity, while at the same time to incur the lowest
possible cost. In the literature there are several models and approaches that have been
developed for planning and controlling inventories (Leenders et.al., 1989).
ABC analysis is used to classify items into three groups for management and control
of inventories with large number of different items. Items are classified according to
their annual dollar usage in traditional ABC analysis. Howewer, there are instances
where criteria other than unit price and anuual usage are important. Thus, inventory
managers have to decide how to take these different criteria into account. It has been
suggested by Flores and Whybark (1986) that multiple criteria ABC classification
can provide a more comprehensive managerial approach.Withmultiple criteria certain
mechanical methods are required to reduce the classification to an ABC grouping.
An artificial neural network approach, pioneered by McCulloch and Pitts (1943), can
be used to classify large amount of items in terms of multi criteria classification in
order to reach inventory management objectives.

1.1 Purpose of the Thesis

For effective functioning of any organization, it is essential to implement efficient

performance measurement system. Performance in a company must be controlled
throughout every operation and functioning units. There are several performance and
operation control methods and almost in every organization main attention is paid to
maintenance policies, which are known as age replacement, periodic repair, failure
limit, etc. Each of these maintenance policies have many advantages and
disadvantages. Several known maintenance techniques are productive maintenance,
condition based maintenance and predictive maintenance, which all somehow
necessitate the importance of effective inventory management, to provide efficient
operating conditions in any organization.

1
Inventory management is one of the major points in production scheduling and
operations management. It is very important to maintain sufficient amount of
inventory to meet operational demand and achieve productivity and at the same time
maintain possible costs at the minimum level. There are several techniques
developed to plan and control inventory levels in the organization to minimize costs
of keeping inventory and diminish levels of stock-out of inventory. Methods
developed are mostly based on mathematical and classification approaches.
Mathematical approaches developed for inventory management are linear
programming, dynamic programming, goal programming, simulation and etc.
During the last decades, various models have been examined and developed for
inventory management based on mathematical optimization approaches. Multi-
echelon technique for recoverable item control (METRIC) developed by Sherbrooke
(1968) is the early study, which concentrates on optimization of inventory costs and
service levels connected with a potential spare parts management policy based on
economic order quantity, reorder point, safety stock and so on. There are several
drawbacks in these developed models like complexity, abstract and oversimplified
structures and they do not include factors like obsolescence, standard characteristics
of the item, type and quality of suppliers, etc.

Main attention is given to inventory classification schemes as a spare parts

management, because of its need to keep ample sizes of inventories to provide for
unexpected breakdowns and at the same time keeping minimum inventory level in
terms of stock costs. Generally, classification problems are often encountered in a
variety of fields including finance, marketing, environmental and energy
management, human resources management, medicine, etc. Many investigations
have been done on inventory classification in various industries like production,
pharmaceuticals, services industries like hospitals, and many other areas. However,
little research has been done on inventory management in railway systems.
Specifically, the objective of this study is to develop multi criteria inventory
classification model based on ABC classification for maintenance spare parts in
railway systems.

The research methodology will include developing an artificial neural network which
will be used to classify spare parts inventories in railway industry based on multi
criteria ABC classification. The reason why neural network approach is chosen for

2
the classification in this study is that there is learning process that enable the model
to catch up the nonlinear relation in the problem pattern.

1.2 Research Scope

In the new millennium, the transportation of people and materials has even greater
significance world-wide than it had in the immediate past. The birth and evolution of
railways during the last two centuries have had an enormous influence on industrial,
social and economic development (Bonnett,2005).

Every railway needs to establish for itself a regime for the regular maintenance of
rolling stock. The various activities, involving both inspection and work to
components, will necessitate some time when stock is not available to get back to
normal operation. It is difficult to get the balance right between over and under
maintenance and this must be watched carefully, allowing suitable adjustments to be
made in both procedures and frequencies when these are shown to be necessary
(Bonnett,2005).

A little investigation has been done on inventory management in the area of railway
systems. The importance to provide reliable and effective performance in railway
systems is based on continuous or scheduled maintenance and unplanned repair. In
other words, maintenance operations directly affect the performance of railway
vehicles and play a crucial role in railway services to provide uninterrupted and high
quality services to passengers.

Inventory management of spare parts in light railway systems has been a researh
topic for studies done by Çelebi et al. (2008) using analytical hierarchy process
(AHP) with data envelopment analysis (DEA) and later, by Sönmeztürk et.al (2008)
using fuzyy AHP and DEA. In the context of our research knowledge through the
literature, spare parts inventory classification with ABC analysis has not been
mentioned to be implemented by the usage of artficial neural network for
classification of spare parts inventory in light railway systems. So, the contribution
of this research will be to propose a multi criteria classification model and develop
artificial neural network model based on ABC analysis for classification of spare
parts inventory of company performing maintenance operations in light railway
systems.

3
1.3 Research Methodology

In this study artificial neural network model will be proposed for the classification of
spare parts inventory in light railway systems. The design of the study will be
conducted in three steps. First, a literature review will be done on inventory
classification based on multi criteria ABC analysis and artificial neural network
topologies will be researched. Within the literature review, findings on inventory
classification will be analyzed and evaluation criteria for the study in light railway
systems will be selected. Then, artificial neural network model that is suitable for
inventory classification will be chosen and its structure and parameters will be
analyzed. Two network structures will be trained and the more successful one will be
chosen as the proposed model for the study. Finally, the developed neural netwrok
will be tested to validate the performance of the network and results will be
discussed.

4
2. INVENTORY MANAGEMENT

The objective of inventory management is to replace a very expensive asset called

“inventory” with a less-expensive asset called “information.” In order to accomplish
this objective, the information must be timely, accurate, reliable and consistent
(Viale, 1996). There are five basic types of inventory: raw material, work-in-process,
finished goods, distribution inventory and maintenance, repair, and operating
supplies which are present in almost every type of organizations. All of the following
types of inventory act to buffer fluctuation in supply and demand. They add no value
to the process and result in additional cost to carry, store, etc. However, they are
necessary in order to ensure high levels of customer service. All types of businesses
(retailers, manufacturers, banks) are challenged with the conflicting objectives of
minimizing inventory and ensuring high levels of customer service (Viale, 1996).

Inventory management in most companies necessitates maintenance operations of the

most inventories because efficient and effective management helps to maintain
competitive advantage to keep up in a time of accelerating globalization and
innovations in management techniques. The number of stock keeping units that is
held by large companies may easily approach ten of thousands and it may not seem
economical to design inventory management policy to each of the inventory unit.
There is a need to differentiate the inventory according to the business model of the
firm, because almost every business company differentiates in a way they manage
and operate company. So, several techniques have been developed to classify the
inventory according to their importance in the firm to enable meaningful
categorization and thus provide to manageable way to control all inventory items in
each category.

2.1 Inventory Classification

Inventory classification should be done in management of inventory because there

are so many types of materials and spare parts and it is easy to lose sight of managing
materials effectively (Çelebi et al., 2008; Parvoti and Anandarajan, 2002). To prevent
a company from misallocating its materials management resources, planning and

5
control systems must be implemented. Effective procurement planning and control
system maintenance must be developed so that it would keep the balance between the
keeping inventory costs to the minimum achievable level and protect the company
from critical stock-outs of raw materials, work-in-progress inventories and finished
goods.

Multi criteria decision aid (MCDA) has several distinctive and attractive features,
involving mainly its decision support orientation which has found its crucial
applications in inventory classification. The significant advances in multi criteria
decision aid over the last three decades constitute a powerful non-parametric
alternative methodological approach to study inventory classification problems.
Although the MCDA researches until the late 1970s, have been mainly oriented
towards the fundamental aspects of this field, as well as to the development of choice
and ranking methodologies, during the 1980s and 1990s significant research has been
undertaken on the study of the classification problem within the MCDA framework
(Doumpos, 2002).

Classification itself is referred as to assign a finite set of alternatives into predefined

groups. In this case the alternatives belonging into different groups have different
characteristics, without being possible to establish any kind of preference relation
between them (i.e., the groups provide a description of the alternatives without any
further information). One of the most well-known problems of this form is the iris
classification problem used by Fisher (1936) with a pioneering work on the linear
discriminant analysis. This problem involves the distinction between three species of
flowers, iris setosa, iris versicolor and iris virginica, given their physical
characteristics (length and width of the sepal and petal). Obviously, each group
(specie) provide a description of its member flowers, but this description does not
incorporate any preferential information. Pattern recognition is also an extensively
studied problem of this form with numerous significant applications in letter
recognition, speech recognition, recognition of physical objects and human
characteristics (Doumpos, 2002).

Doumpos (2002) has also made attention to the difference between classification and
clustering of alternatives in decision making problems because they are usually
confused. In classification the groups are defined a priori, whereas in clustering the
objective is to identify clusters (groups) of alternatives sharing similar

6
characteristics. In other words, in a classification problem the analyst knows in
advance what the results of the analysis should look like, while in clustering the
analyst tries to organize the knowledge embodied in a data sample in the most
appropriate way according to some similarity measure.

The significance of the classification problems extends to a wide variety of practical

fields of interest. Some characteristic examples are the following:

Medicine: medical diagnosis to assign the patients into groups (diseases) according
to the observed symptoms (Tsumoto, 1998).

Pattern recognition: recognition of human characteristics or physical objects and

their classification into properly defined groups (Nieddu and Patrizi, 2000).

Human resources management: personnel evaluation on the basis of its skills and
assignment to proper working positions.

Production management: monitoring and control of complex production systems for

fault diagnosis purposes (Shen, 2000).

Marketing: selection of proper marketing policies for penetration to new markets,

analysis of customer characteristics, customer satisfaction measurement, etc. (Siskos
et al., 1998).

Environmental management and energy policy: analysis and in time diagnosis of

environmental impacts, examination of the effectiveness of energy policy measures
(Diakoulaki et al., 1999).

Financial management and economics: bankruptcy prediction, credit risk

assessment, portfolio selection (stock classification), country risk assessment
(Zopounidis, 1998).

As Doumpos (2002) outlined about classification problems application field, it is

widely used in modeling inventory classification problems. As was pointed out in
previous sections, the aim of inventory management is to make right decision
regarding the appropriate level of inventory. In practice, all inventories cannot be
controlled with equal attention. The most widespread technique used in inventory
systems is the ABC classification system.

The utilization of classification techniques for spare part management tool represents
a widely used approach in industrial world. ABC inventory classification is a

7
frequently used procurement planning and control method that was designed to
achieve balance between two conflicting economic forces: inventory costs and
critical stock-outs of spare parts.

2.2 ABC Analysis

ABC analysis was discovered by Pareto, an Italian economist, approximately 100

years ago. He discovered that a small percentage of a population always has the
greatest effect. Because of its easy-to implement nature and remarkable effectiveness
in many inventory systems, this approach is widely used in practice (Chen et.al.
2006). Pareto’s law was further expanded to the ABC classification and is
summarized below. Viale (1996) has explained ABC analysis like this: “When
considering how to apply this tool to establish inventor y levels, consider the
following: From a practical standpoint, ask yourself: “Which products (and which
customers) generate 80 percent of the revenue?” Answer: Approximately 20 percent
of the products and customers generate 80 percent of the revenue. From the Table
2.1, detailed explanation on the A, B and C grouping can be viewed:

Table 2.1: ABC analysis grouping

20% of 80% of the company’s These are called
Customers, products, or revenue and inventory “A” customers
parts= investment “A” products
“A” parts
30% of customers, 15% of the company’s These are called
products, or parts= revenue and inventory “B” customers
investment “B” products
“B” parts
50% of customers, 5% of the company’s These are called
products, or parts= revenue and inventory “C” customers
investments “C” products
“C” parts

The fluctuation in demand for the B and C products causes most of the product mix
problems, the changes on the shop floor (capacity) and the changes in the supplier
due dates Viale (1996).

Consider doing the following:

• Reduce the forecast error by improving the forecast model.

• Use the standard deviation formula A customers and A and B parts.

8
 • For C parts, give a predetermined number of days’ supply and allow the
production floor to build this during the beginning of each quarter.

The standard deviation of forecast error tool can also be used to determine

 The amount of extra (safety) lead time needed to ensure on-time delivery.
 The amount of work to release to the shop floor to make sure machines do not
run out of work. That is, how much “queue” buffer is needed in back of the
bottleneck work center?
 The number of extra (buffer) pieces to start on the initial machining
operation to ensure a particular yield after the final operation (for example,
scrap allowance).
 The amount of machine downtime (safety) to allow for in planning utilization
of available capacity (Viale, 1996).

Classical ABC technique differentiates the inventory into three classes: A – very
important, the inventory small in volume but large in cost, B – moderately important,
inventory intermediate in volume and cost; and last, C – least important, that is
inventory is large in volume and however low in cost usage.
“A” items have the highest value. These are relatively few items (15– 20 percent)
whose value accounts for 75– 80 percent of the total value of the inventory. As a
general rule, 20 percent of the items constitute 80 percent of the annual requirements.
All “A” items are counted monthly.

“B” items have medium value. These are a larger number in the middle of the list,
usually about 30– 40 percent of the items, accounting for about 15 percent of the
value. All “B” items are counted quarterly.

“C” items have low value. These are the bulk of the inventory, usually about 40– 50
percent of the items, whose total inventory value is almost negligible, accounting for
only 5–10 percent of the value. All “C” items are counted annually. Many times
these physical counts are based on estimates (Viale,1996).

The amount of time, effort, money and other resources spent on inventory planning
and control needs to be in accordance with importance of each item. So, the purpose
of this technique is to provide all inventory items with appropriate levels of control.

Traditional ABC classification is based on only one criterion. Generally, the only
criterion is taken as annual cost usage or average unit cost, and sometimes it may be

9
the number of orders and purchasing conditions (Simunovic et al., 2008). However,
there may be other important criteria that represent various important considerations
for management and it is a disadvantage of the classical ABC classification that it
takes into account only one criteria and leaves out many important characteristics
according to which inventory may differentiate. Thus multi-criteria techniques have
been developed to enable efficient decision making which will include crucial
criteria like criticality of an item, stock-out penalty, lead time of supply, part
criticality, availability, average unit price, the scarcity, the rate of obsolescence,
substitutability, etc. (Simunovic et al., 2008). Therefore, it has been usually
recognized that traditional ABC analysis may not be able to provide a good
classification of inventory in practice (Guvenir & Erel, 1998; Huiskonen, 2001;
Partovi & Anandarajan, 2002). There are many instances when other criteria become
important in deciding the importance of an inventory item. This problem becomes a
multi-criteria inventory classification that has been studied by some researchers in
the past. In general, complex computational tools or procedures are needed for multi-
criteria ABC classification (Chu et. al., 2008).

2.3 Multi Criteria ABC Analysis

A typical company holds multitude of items in inventory and only a small fraction
deserves the close attention. The decision maker’s task is determining the inspection
levels on each inventory item; thus, first and foremost, assigning the priorities for
each item and realizing a sensible classification . As Sharaf and Helmy (2001) have
noted, thousands of items may be potentially held in inventory by a typical
organization, however only a small portion of them deserve management’s close
attention and accurate control.

To overcome the limitations of the traditional classification analysis, many

researchers concentrated on incorporating multiple criteria judgments into the
inventory classification procedure. Flores et. al. (1985) have transformed the
classical ABC analysis for the inventory classification, considering another important
criterion to so called bi-criteria inventory classification. They used traditional ABC
analysis by classifying the inventory by first criterion and then by second criterion.
Here the disadvantage of this method was that they took weights of each criterion to

10
be equal. Flores et al. (1985) suggested using multiple criteria like lead time,
criticality, commonality, obsolescence and substitutability.

Since the multi-criteria classification idea was introduced, several authors reported
ondifferent approaches to the problem. Cluster analysis method (Cohen and Ernst,
1988), a classification expert system (Petrovic & Petrovic, 1992), joint criteria matrix
approach (Flores, Olson, & Dorai, 1992), a heuristic procedure employing genetic
algorithms (Guvenir & Erel, 1998), and a weighted linear optimization methodology
(Ramanathan, 2006) have been successfully applied in the past years.

Other developments of multi-criteria of ABC analysis has been done by Ernst and
Cohen (1990). They proposed technique based on statistical clustering that uses a full
range of operationally significant attributes. Their technique can accommodate a
large number of combinations of attributes which is mostly important for strategic
and operational purposes. Yet, their method required utilization of large amount of
data, the use of factor analysis and clustering procedure which may end up to be
impractical for typical stockroom environment. Moreover, the clusters should be
measured again in order to classify new stock items and there is also chance that each
time a new item is added, the clusters previously classified will be re-classified
again. This model may be too sophisticated and might disturb the inventory control
procedure.

11
12
3. AN ARTIFICIAL NEURAL NETWORK

The simplest definition for artificial neural network (ANN) is that, it is a

mathematical or a computational model which is inspired by the structure and/or
functional aspects of biological neural networks. A neural network is an adaptive
system that changes its structure based on external or internal information that flows
through its network and generally is learning to generalize or classify or organize
data in the learning phase. Modern neural networks are non-linear statistical data
modeling tools. They are used to model complex relationships between inputs and
outputs or to find correlations in the data. Unlike traditional computational models,
ANN model has brain like structure and functionality of the brain cells.

Researches on artificial neural networks have been inspired by the recognition of

how human brain computes in an entirely different way from the computer
functioning. The structure of the brain is highly complex, nonlinear and parallel, like
information processing systems in computer. It has the ability to do certain
computations such as pattern recognition, perception and motor control in a very
short term. For instance, it usually accomplishes perceptual recognition tasks such as
memorizing a familiar face in an unknown scene in 0,1-0,2 milliseconds. On the
other hand, computer will spend much more time like several minutes or even hours
to perform less difficult tasks. The reason why human brain has been interesting
topic for researchers is that its structure and functioning gives thorough insight to
how a computer can be developed to perform functions usually completed with
human intelligence such as reasoning, learning and self improvement (Haykin,
1999). Haykin (1999) explains that a brain has great structure and ability to build up
its own rules through time using that is known as experience.

Human brain undergoes a dramatic development within two years after birth and
continues to evolve after that. It has capability to develop its structural elements
known as neurons that allow the building nervous system to adjust to its
surroundings. Neurons play same role, which is information processing unit, in both
human brain and neural networks. In other words, it can be explained that the neural

13
network are designed like a machine to perform particular tasks like human brain
(Haykin, 1999). To specify the above statements made by Haykin (1999), his
definition for artificial neural network is that “it is a massive parallel distributed
processor for storing experiential knowledge and making it available for use. It
resembles the brain in two respects: knowledge is acquired by the network from its
environment through a learning process and secondly interneuron connection
strengths, known as synaptic weights, are used to store the acquired knowledge”.

The model of learning process is known as learning algorithm and its function is to
modify the synaptic weights of the network in an orderly fashion to attain a desired
design objective (Haykin, 1999).

3.1 Application Areas of Artificial Neural Networks

The application area of artificial neural networks can be extended to practically every
kind of problem where regression based models and statistical models can’t be
applied or do not give desired results. Neural networks give the best results in
building nonlinear models.

• Classification:Discrimination of elements based on similarities

• Pattern recognition: Recognition of an output according to given input data.
• Modeling:Generalization of few examples
• Clustering:Grouping the elements according to common attributes or
features.
• Forecasting:Prediction of future outputs by analysis of current and past data.
• Optimization:Satisfying multiple conflicting goals and constraints.
• Finance:ANN’s can be used for funds investment, credit analysis, risk
insurance, options and futures prediction, trend prediction, stock investment
optimization , cash flow forecasting, etc.
• Medicine: ANN’ can be used medical diagnosis, classification of deceases,
genetic mapping, blood mapping, treatment cost estimation, etc.
• Science: ANN’s can be used for modeling complex problems, nonlinear
problems, physical system modeling, chemical compound identification,
botanical classification, odor analysis and identification, etc.
• Sales and Marketing: ANN’s can be used for sales forecasting, targeted
marketing, service usage forecasting, etc.

14
 • Energy: ANN’s can be used for electrical load forecasting, energy demand
forecasting, power control systems, etc.
• Production: ANN’s can be used for quality control, process control,
temperature and force predictions, production routing optimization, etc.

3.2 Benefits of Neural Networks

The ability to learn and generalize is the main advantage of neural networks. The
ability of generalization makes it possible for neural network to give a reasonable
output for inputs that have not been encountered before. Below are the properties and
capabilities of neural networks reported by Haykin (1999):

• Nonlinearity: the nonlinear structure of neural network is a very important

property in cases where a nonlinear relation is required. A regression based
methods are hard to control and implement where a cases of linear relations
takes place.
• Input-Output Mapping: A popular way of training a neural network is to use
supervised training. Supervised training teaches a neural network to produce
the ideal output with a given inputs. Every training iteration calculates how
close the actual output is to the ideal output. The closeness to ideal output is
expressed as an error percent. Each iteration modifies the internal weights of
the neural network to get the error rate as low as possible. The training is
repeated until there no further significant changes in error rate. The neural
network learns from training and constructs an input-output mapping for a
given problem.
• Adaptivity: Neural Networks have a capability to adapt their synaptic weights
to changes in surrounding environment. A neural network trained to operate
in a specific environment can easily be adapted to minor changes
environmental conditions.
• Evidential Response: In pattern classification a neural network can be
designed to provide information not only about the output but also about the
confidence in the decision made. The confidence in decision making can be
used to eliminate ambiguous patterns and improve classification performance
of the network.

15
 • Contextual Information: Knowledge is presented in by the structure and
activation of a neural network. Every neuron in the network is potentially
affected by global activity of all other neurons in the network. Consequently,
contextual information is dealt with naturally by a neural network.
• Fault Tolerance: A neural network which is used in hardware form has the
ability to be intrinsically fault tolerant or capable of robust computation, in
the sense that its efficiency deteriorates gracefully undesirable operating
conditions. That is if the neuron or its connection links are damaged, the
recall of the saved patters in worsened in quality. On the other hand, because
of distributed nature of information saved through the network, the adverse
condition has to be extensive before the overall response of the network is
damaged. Therefore, a neural network shows a graceful degradation rather
than catastrophic failure.
• VLSI Implementability: A neural network has the potential to compute certain
tasks very quickly because of its massively parallel nature. This characteristic
enables neural network well suited for implementation using very-large-
scale-integrated (VLSI) technology. One of the beneficial virtues of VLSI is
that it provides means of recognizing very complex behavior in a highly
hierarchical fashion.
• Uniformity of Analysis and Design: Neural networks have universality as
information processors, because the same notation is used in all domains
involving the application of neural networks. This characteristic reveals itself
in several ways; at first neuron represents an ingredient common to all neural
networks. Then, this commonality makes it possible to share theories and
learning algorithms in different applications of neural networks. And finally,
modular networks can be designed a seamless integration of modules.
• Neurobiological Analogy: the architecture of neural network is motivated by
analogy of brain, which is a living proof that fault tolerant parallel processing
is not only physically possible but also fast and powerful (Haykin,1999).

3.3 Artificial Neural Network Structure: An Explanation Based on Biologic

Nervous System

Haykin (1999) has explained the biologic nervous system in a three staged system as
can be seen from diagram in Figure 3.1. The brain is in the center of neural system

16
which represents the neural net. It continuously receives information, perceives it,
and gives appropriate decision. There are two arrow nets in the figure. The arrows
pointing from left to right indicate the forward transmission of the information-
bearing signals through the system. Moreover, there is the feedback process which is
transmitted from right to the left by arrows. The receptors convert stimuli from the
human body or the external environment into electrical impulses that conveys
information to the neural net (brain). The effector convert electrical impulses
generated by the neural net into discernible responses as system outputs.

Stimulus Receptors Neural Effectors Response

Net

Figure 3.1: Block diagram representation of nervous system (Haykin, 1999).

There are several types of ANN architecture in literature; however a general ANN
architecture can be shown as in Figure 3.2. There is at least one input element in the
input layer. In this layer input elements generate the same values of input without
any processes (Tosun, 2007). There is at least one output element and in spite of
input elements in output elements, there is a process that generates output (Tosun,
2007). Processing layers usually called black box because understanding each
processing elements behaviors is a very difficult task. This layer(s) and functions
used in these layers can change according to the ANN type.

Input layer Processing Output

Inputs Layer(s) layer
Output

Figure 3.2: Block diagram representation of ann architecture.

Neurons are the structural constituents of biological nervous system. The structure of
neuron and its parts are demonstrated in Figure 3.4. With dendrites neuron collects
the stimulus of previous neurons. In cell body this stimulus is evaluated and an
output stimulus is generated. This stimulus is send to following neurons with the
axon. At this point axon can be divided into many parts and send the stimulus to all
connected neuron dendrites. The axon- dendrite connection areas are called synapse.
The most common type of synapse is called chemical synapse, which operates as
follows; a presynaptic process liberates a transmitter substance that diffuses across
the synaptic junction between neurons and then acts on the postsynaptic process

17
(Haykin, 1999). In an adult brain, plasticity, which permits the developing nervous
system to adapt to its surrounding environment, may be accounted for by two
mechanisms: the creation of new synaptic connections between neurons, and the
modification of existing synapses (Haykin, 1999).

Partovi and Anandarajan (2002) defines ANN as follows: “ The ANN in general can
be characterized in the following components: a set of nodes and connection between
nodes. The nodes can be seen as computational units. They receive inputs and
process them to obtain an output. This process might be very simple computation
like summing or it might also be complex as to contain another network. The
interaction of nodes through the connections leads to a global behavior of the
network, which cannot be observed in the elements of the network. This global
behavior is said to be emergent that is the abilities of the network supersede the ones
of its elements, making networks a very powerful tool. In ANNs, the network sees
the nodes as ‘artificial neurons’. An artificial neuron is a computational model
inspired by the natural neurons. Natural neurons receive signals through synapses
located on the dendrites or the membrane of the neuron. When the signals received
are strong enough, the neuron is activated and emits a signal through the axon. This
signal might be sent to another synapse and might activate other neuron”. A typical
neuron structure described by Partovi and Anandarajan (2002) can be seen in Figure
3.3 below:

Axon
Soma Terminal

Nucleou

Figure 3.3: Structure of a typical neuron (Partovi and Anandarajan, 2002).

18
The complexity of the real neuron is highly abstracted when modeling artificial
neurons. These basically consist of inputs (like synapses) which are multiplied by
weights (strength of the respective signals), and then computed by the mathematical
function which determines the activation of the neuron. Another function, which may
be the identity, computes the output of the artificial neuron. ANNs combines
artificial neurons in order to process information.

A processing element can be shown as in Figure 3.4. We can express the following
similarities between processing elements (PEs) or neurons of ANN and neurons of
nervous systems. The weights in processing elements are synapses. The summing
junction is the dendrites that collect the inputs. Activation function is the cell body
that processes the stimulus. And the output element is the axon that transports the
output to the other neurons.
Bias Activation
x1 wk bk function

φ .
x2 Output
Input wk2 ∑ yk
signa
ls
Summing
xm wk junction

Synaptic
weights

Figure 3.4: Model of a typical processing element (Haykin, 1999).

The neuron model shown in Figure 3.4 also includes an externally applied bias,
denoted by bk.. The bias bk has the effect of increasing or decreasing the net input of
the activation function depending on whether it is positive or negative (Haykin,
1999). Mathematically we can describe a neuron k by the following pair of equations
3.1 and 3.2 :

 = ∑

  (3.1)

 = 
 +   (3.2)

Here, 
,  ,  , …  are the input signals; 
,  , …  are synaptic weights of
neuron k;
 is the linear combiner output due to the input signals;  is the bias;
 .  is the activation function; and  is the output signal of the neuron. The use of
bias  has the effect of applying the affine transformation to the output
 of the
linear combiner in the model (Figure 3.4), as shown by equation 3.3 below:

19
 =
 +  (3.3)

The higher the weight of an artificial neuron, the stronger the input will be, which is
multiplied by it. Depending on the weights, the computation of the neuron will be
different. By adjusting the weights of an artificial neuron we can obtain the output
we want for specific inputs. But if we have an ANN of thousand of neurons, their
weights of the ANN may be adjusted only by algorithms in order to obtain desired
output from the network. This adjusting process is called learning or training of the
network. The number of types of ANNs is very high and there are more than hundred
of different models considered as ANNs. Since the functions of ANNs is to process
information, they are used in many fields like study behavior and control in animals
and machines, also as well as engineering purposes like pattern recognition,
forecasting, data compression and classification. For example, Partovi and
Anandarajan (2002) had used two types of learning methods to examine
classification accuracy of ANN as an aid to facilitate the decision making process of
classifying inventory items. More specifically, two learning methods used by Partovi
and Anandarajan (2002) are back propagation (BP) and genetic algorithms. Also,
Wei et al. (1997) had used ANN which takes into account quantitative and
qualitative data and nonlinear problem with multiple inputs in supplier evaluation
process.

3.4 Learning Process in Artificial Neural Networks

Learning in the conext of the neural network has been defined by Haykin (1999) as;
learning is a process by which the free parameters of a neural network are adjusted
through a process of stimulation by the environment in which the network is
embedded and the type of learning is determined by the manner in which the
parameter changes take place. Haykin (1999) also mentioned that this definiton of
learing process implies the following sequence of events:

• The neural network is stimulated by an environment.

• The neural network undergoes changes in its free parameters as a result of the
changes that have occurred in its internal structure.

Çınar (2007) has defined learing process in ANN as a kind of price-penalty system.
If the output of artificial neural network and the desired output are in the same
direction, the weights of ANN are strengthened. If the output of ANN and desired

20
output are in opposite direction, the weights are weakened to teach ANN to respond
differently (Çınar, 2007).

In practice, artificial neural networks that have only one hidden layer may easily
learn the process with few data and continuous functions (Çınar, 2007). Regarding
the hidden layer, Çınar (2007) has mentioned that hidden layer is only required if the
function is not continuous for some points. For many ANN model in literature,
researchers reported that only one hidden layer is enough and that the second hidden
layer slows the learning process (Çınar, 2007).

Feedforward network indicates that data flow from input layer to output layer.
Output of each layer is the input of the next layer and is the functions of its inputs
(Çınar, 2007). Activation fucntion computes the output value of each neuron. In
complex models it is important to utilize nonlinear activation functions. The shape of
the activation function affects the learning performance of the neural network;
however, it does not influence the overall performance of the network (Çınar, 2007).

The learning of network can be in two ways: online or batch learning. First one, in
online learning data used one by one for learning. In second way, batch learning, the
whole data is used at once for learning. Apaydin (2004) mentioned that in batch
learning, learing (changes in free parameters of neural network) accumulated over
the all patterns and the change is made once after a complete pass over the whole
training set is run. A complete pass over all the patterns is called an epoch (Apaydin,
2004).

There are three kinds of learning process: supervised learning, unsupervised learning
and hybrid learing.

3.4.1 Supervised learning

It is know as learning under supervison of teacher, since in conceptual terms a
teacher with the knowledge about environment teaches the neural network with that
knowledge being presented by a set of input-output examples (Haykin, 1999).
Apaydin (2004) gave examples of supervised learning as regression and
classification problems.

The following example gives better explanation of the supervised learning: let the
teacher and the neural network both be exposed to training vector drawn from the
environment. By the virtue of built-in knowledge, the teacher is able to provide the

21
neural network with desired responce for that training vector (indeed, the desired
respense represents the optimum action to be performed by the neural network). The
parameters of the network are modified according to the traininf vector and the signal
error. The error signal is computed by the difference between the desired output and
the actual output of the network. This adjustment is carried out iteratively in a step
by step fashion with objective to eventually make the neural network emulate the
teacher. The emulation process is presumed to be optimum in some statistical sense.

3.4.2 Unsupervised learning

It is also called learning without a teacher, because in spite of supervised learning
only inputs of the problem are known. In unsupervised learning the goal is to
determine the formation along the inputs (Çınar, 2007). Input space has a pattern and
if analyzed it can be deduced which input are more repeated and which are less
repeated. This is called density estimation in statistics (Alpaydın, 2004). When the
patterns are discovered learning is completed; a new input’s cluster can be
determined (Haykin, 1999).

One method for density estimation is clustering where the aim is to find clusters or
groupings of input. The following example of clustering is given by Alpaydın
(2004): in the case of a company with a data of past customers. The data contains the
demographic information as well as the past transactions with the company, and the
company may want to see the distribution of the profile of tis customers, to see what
type of customers frequently occur. The author noticed that in such a case, a
clustering model allocates customer similar in their attributes to the same group,
providing the company with natural groupings of its customer. Alpaydın (2004) also
added that once such groups are found, the company may decide strategies (for
example, specific services and products to different groups).

3.4.3 Reinforcement learning

In some applications, the output of the system is a sequence of actions. In such a
case, a single action is not important; the policy, which is the sequence of correct
actions to reach the goal, is important. In this case, neural network should be able to
assess the goodness of policies and learn from past good action sequences to be able
to generate a policy. Such learning methods are called reinforcement learning
(Alpaydın, 2004).

22
In reinforcement learning, like unsupervised learning, certain outputs are not used to
train the neural network. But the desired outputs are defined as good output or bad
output and then used to train the neural network (Çınar, 2007). But defining good or
bad outputs are somehow similar to supervised learning.

Chess game can be an example of this type of learning because the rules of the game
are limited but in many situations there is large number of possible moves (Aplaydın,
2004). In such a case one move is not important, the series of moves are important to
win the game. For further information about unsupervised and reinforcement
learning please look at Haykin (1999) and Schalkoff (1997).

In this study, supervised learning methodology is going to be described in detail

because it will be utilized in the classification of the spare parts inventory based on
ABC analysis.

3.5 Multi-Layer Perception (MLP) and Back Propagation Algorithms

3.5.1 Multi layer perception

The multi-layer perception is one the the most used neural network topologies.
Principle et al. (1999) have indicated that, the MLP is capable of approximating
arbitrary functions. MLP plays important role in the study of nonlinear dynamics and
other function mapping problems. A common multi layer perception architecture can
be seen in Figure 3.5:

23
 Hidden Layer Output Layer

x1
y1
x2

y2
x3

x4

Inputs Hidden Output Outputs

units units

Figure 3.5: Multi-layer perception model.

The multilayer perception network given above has three layers with one hidden
layer. x1,x2,x3,x4 are input values and y1and y2 are the outputs of the network.

The are two crucial aspects of the multi-layer perception. First one is that nonlinear
processing elements (PEs) that have a nonlinearity which has to be smooth (the
logistic function and the hyperbolic tangent are the most widely used). Second aspect
is that MLP has massive interconnectivity that is any element of a given layer feeds
all elements of the next layer (Principle et al., 1999).

Principle et al. (1999) have indicated that multi-layer perception are generally trained
by backpropagation algorithms.

3.5.2 Backpropagation algorithm

The Backpropagation algorithm propagates the error through the network and allows
adjustions of the hidden nodes, i.e. processing elements. The multi-layer perception
is trained by error correction learning, so it means that desired for the system should
be known.

24
Learning by error correction is done in the following way:

At PE i at the iteration n, the system generates response yi(n) and the difference
between system response, yi(n) and the desired response di(n) for a given pattern
gives us the instantaneous error,  , which is defined by equation 3.4:

  =   −   (3.4)

According to the gradient descent learning theory, each weight in the network can be
adjusted by modifying the present value of the weight with the terms that is
proportional to the present input and error at the weight, that is defined as in equation
3.5:

  + 1 =   +    (3.5)

where   is the local error and can be easily computer from the system error,  , at
the output PE or can be computed as the weighted sum of errors at the internal PEs.
The constant  is the step size and called the learning rate. This procedure is defined
the backpropagation algorithms.

The backpropagation algorithm computes the sensitivity of the cost function with
respect to each weight in the network, and updates each weight proportional to the
sensitivity. This procedure is very easy because it can be implemented with local
information and requires just a few multiplications per weight. Due to the fact that
gradient descent procedure uses only local information, it can be stuckt at local
minima. Principle et al. (1999) have indicated that the procedure is inherently noisy
because usage of poor estimate of gradient may cause slow convergence.

Momentum learning is the improvement to the straight gradient descent in the sense
that a memory term ( the past increment to the weight) is used to speed up and
stabilize convergence. By momentum learning, the weights are updated as following
in equation 3.6:

  + 1 =   +    +   −   − 1 (3.6)

Where is the momentum of learning. The momentum rate is taken to be between

range 0.1 to 0.9 in most literatures (Çelebi and Bayraktar, 2008; Ghate and Dudul,
2010).

25
Weights in the network during the training process can be updated in two ways, on-
line and batch training. In on-line training, the weights are updated continuously each
time as each pattern is input to the network. On the other hand, in batch learning we
present all the patterns in the input file (an epoch), accumulate the weight usdates,
and then update the weights with the average wieght update. Principle et al. (1999)
have noted that online learning and batch learning are theoretically equivalent,
however, on-line learning is superior in tough problems where many similar input-
output pairs exits. Moreover, Haykin (1999) states that it is less likely that
backpropagation algorithm to get stuck in a local mnima with the on-line mode of
training.

Backpropagaion starts by loading an initial value for each weight (normally a small
value) and proceeding until some stopping criterion is met. The three most common
stopping criteria are to cap the number of iterations, to threshold the output mean
square error, or to use cross validations. The cross validation is the most powerful of
the three types of criteria because it stops the training at the point where the best
generalization is obtained that is the performance in the test set. The training data is
divided into two sets: training set and cross validation set. The cross validation set is
used to see how the trained network is doing, for example, every 100 training epochs
cross validation set test the network training performance. When the performance
starts to degrade in the validation set, the training should be stopped (Alpaydın,
2004; Haykin, 1999; Principle et al., 1999).

The fundamental of any iterative training process is the measurement of the learning
process. The learning curve can be good estimeate of the training process. Learning
curve shows the change of mean squared error with the number of epochs. The
difficulty of the task and how to control learning can be judged from the learning
curve. If the learning curve is straight, it means that learning rate should be
increased. Meanwhile, if the learning curve is oscillating up and down, it means that
step size shoulf be decreased. When the learning curve stabilizes at an error level that
is not acceptable, the network topology, like altering number of node in hidden layer,
or number of hidden layers, or the training prosedure like other more sophisticated
search techniques.

26
3.6 Types of Activation Functions

Activation function, shown as  , defines the output of a neuron in terms of the
induced local field v. Here are three basic types of activation functions (Haykin,
1999):

1. Threshold Function: For this type of the activation function in equation 3.7,
described in Figure 3.6,

Figure 3.6: Threshold function

1 #$ ≥ 0(
  = !
0 #$ < 0
(3.7)

is used. In engineering literature, this form of threshold function is commonly

referred to as Heaviside function. Correspondingly, the output of neuron k
employing such a threshold function is expressed as in equation 3.8:

1 #$ ≥ 0(
 = !
0 #$ < 0
(3.8)

where  is the induced local field of the neuron; that is as in equation 3.9:

 = ∑
)
  +  (3.9)

Such a neuron is called in the literature as the MCCulloch-Pitts model in

recognition of the leading work done by McCulloch and Pitts. Within this model,
the output of the neuron takes on the value of 1 if the induced local field of that
neuron is non negative or 0 if negative. This statement describes the all-or-none
property of the McClloch-Pits model.

2. Piecewise-Linear Function: For the piecewise-linear function described in

Figure 3.7, the following equation 3.10 is used:

27
 -1,  ≥ + 

+
  =  +  , +  >  > −  (

,
+0,  ≤

(3.10)

* 

Where the amplification factor inside the linear region of operation is considered to
be unity. This form of an activation function may be considered as an approximation
to a nonlinear amplifier. The following two situations may be viewed as special
forms of the piecewise-linear function (Haykin,1999):

a. A linear combiner arises if the linear region of operation is maintained

without running into saturation.
b. Piecewise-linear functions if the amplifier factor of the linear region is made
infinitely large.

Figure 3.7: Piecewise-linear function

3. Sigmoid Function: This function, which has s-shaped figure, is by far the
most common forms of activation function used in the architecture of
artificial neural networks. Sigmoid function can be seen in Figure 3.8 below:

Figure 3.8: Sigmoid function

28
Sigmoid function is defined as a strictly increasing function that exhibits a graceful
balance between linear and nonlinear behavior (Haykin, 1999). An example of
sigmoid function is the logistic function, defined by equation 3.10 below:
  =

0123 456
(3.11)

Where a is the slope parameter of the sigmoid function. By varying the parameter a,
we obtain sigmoid functions of different slopes, as ilustrated in Figure 3.8. Actually,
5
7
the slope at the origin equals . In the limit, as the slope parameter approached

infinity, the sigmid function becomes simply thresholf function taking the value of 0
or1, a sigmoid function assumes continuous range of value from 0 to 1. Note also
that the sigmoid function is differentiable, whereas the threshold function is not
(Haykin, 1999).

The activation functions defined above the range from 0 to +1. It is sometimes
desirable to have the activation function with range from -1 to +1, in which case the
activation function takes an antisymmetric form with respect to the origin. In other
words,the activation function is an odd function of the induced local field.
Specifically, the threshold function in Equation 3.11 is now as

1 #$ ≥ 0
  = 8 0 #$ = 0 (
−1 #$ < 0
(3.12)

Which is known as the signum function. For the corresponding form of a sigmoid
function we may use the hyperbolic tangent function, defined by equation 3.12:

  = tanh  (3.13)

Letting the activation function of the sigmoid type to assume negative values as
above have some analytic benefits.

29
30
4. LITERATURE REVIEW

Inventory control also known as stock control and inventory management is one of
the important techniques of operational management and it plays important role in
the company management. A well designed approach to inventory management
might have very important influence on the company competitiveness. To achieve
their goals companies should apply optimization and multi-criteria decision making
methods because of the huge amount of inventory items which needs a great
attention to classify these items into different groups. That is different groups might
require application of different management tools and policies. ABC inventory
classification based on the Pareto principle takes into account only one criterion.

There are many contributions to multi-criteria inventory classification in the

literature. Multiple criteria classification necessitates complex decision making tools
(Chen, 2011). For inventory control with many different types of inventory, multi-
criteria decision making methods have been developed for ABC classification. In
literature, various methods include fuzzy analytical hierarchy process (AHP), data
envelopment analysis (DEA), case based distance model, the weighted linear
optimization, the joint criteria matrix, the clustering procedure, the particle swarm
optimization method, principle component analysis, genetic algorithms, artificial
neural network approaches and etc.

4.1 Inventory Classification Techniques

Inventory classification techniques have been developed for the various industries
where large companies accommodate large quantities of inventory items. The
literature on inventory classification includes different methodologies for multi-
criteria inventory classification.

The literatur review hase been done on spare parts inventory classification based on
multi criteria decision makng techniques and short summary of the main research
studies can be seen in Table 4.1.

31
Table 4.1: Academic Researches about Multi Criteria Inventory Classification.
# Citations Technique(s) used Evaluation criteria Application
Area
1 Braglia et al. Reliability Centered Main criteria:spare part Spare parts
Maintenance plant criticality, spare classification
(2004)
(RCM) with AHP. supply characteristics, in paper
inventory problems and operating
usage rate company
Sub criteria: quality
problems, lead time,
internal repair, price, space
required, obsolescence,
deterioration problems,
production loss,
redundancies, etc.

2 Cakir and Fuzzy AHP Price/cost, annual demand, Designing a

blockage effect in case of decision
Canbolat
stock out, availability of support
(2008) substitute material, lead system in a
time small
electrical
appliances
company

3 Simunovic et a. Traditional ABC a1. annual cost usage Classification

b. Multi Criteria b1. annual cost usage, of spare parts
al. (2008)
ABC with AHP criticality, lead time1, in agricultural
lead time 2 machine.

4 Çelebi et al. Multi Criteria ABC Main criteria Criticality, Classification

with AHP and DEA value usage, unit cost, lead of
(2008)
time maintenance
Criticality sub criteria: spare parts in
penalty cost, light railway
substitutability, systems
commonality

32
Table 4.1: Academic Researches about Multi Criteria Inventory classification
(contd.).
# Citations Technique(s) used Evaluation criteria
5 Sönmeztürk et al. Multi Criteria ABC Main criteria
with Fuzzy AHP Criticality, value
(2008)
usage, unit cost, lead
time
Criticality sub criteria:
penalty cost,
substitutability,
commonality

6 Hadi-Venchen and MC ABC with Fuzzy Annual dollar usage,

Mohamanghasemi AHP and DEA limitation of
(2011) warehouse usage,
average lot cost, lead
time
7 Ramanathan (2006) Weighted linear Average unit cost,
optimization based on annual dollar usage,
DEA critical factor, lead
time
8 Ng (2007) Ng model
9 Partovi and Multi Criteria ABC Unit price, ordering
a. Genetic Algorithm cost, demand range,
Anandarajan
b. Artificial Neural lead time
network

10 Li and Kuo (2008) Enhanced Fuzzy Part factor, demand

Neural Network factor, time factor, sale
factor,associated factor

11 Chen et al. (2009) Back-propagation Large variation of

neural network demand, long
(MBPN) and moving purchasing lead time,
fuzzy neuron network necessity to the
(MFNN) operation machine,
cost

33
4.2 Multi Criteria Decision Making (MCDM)

Multiple criteria decision making methods are separated into two basic groups:
multiple attribute decision making methods and multi objective decision making
methods. The first method is used to select the best alternative among the many
feasible alternatives where each alternative is defined by some quantitative and/or
qualitative criteria or attribute. The second method deals with two or more objective
functions subjected to some constraints.

Multi criteria decision aid (MCDA) theory includes multi criteria ABC classification.
According to Roy (1996), multi criteria decision is the aid to assist a single decision
maker to choose, rank or sort alternatives within a finite set according to two or more
criteria. MCDA begins with process of defining objectives, arranging them into
criteria, identifying all possible alternatives and measuring the consequences of each
alternative on each criterion. A consequence is a direct measurement of the success
of an alternative against a criterion and usually is an objective physical measurement
that includes no preferential information (Chen et al., 2008).

The main difficulty with many existing MCDA methods lies in the acquisition of the
decision maker’s preference information in the form of values or weights. Case based
reasoning is an approach to finding preferential information using cases selected by
the decision maker. Chen et. al. (2008) have proposed case-based distance model for
multiple criteria decision aid problem which allows usage of any finite number of
criteria in ABC analysis. In this approach criterion weights and sorting thresholds are
generated mathematically based on the decision maker’s assessment of the case set
and thus, problems related to direct acquisition of preferences information are
avoided. In their evaluation, the criterion includes annual unit cost, annual dollar
usage, critical factor (criticality) and lead time.

4.3 Analytical Hierarchy Process (AHP)

Analytical hierarchy process is the most widely used techniques in almost all sectors
such as economy, traffic, agriculture, information technology and many others. AHP
was developed by Thomas Saaty (1980) and it has a tree like structure that contain
main goal at the top of the hierarchy which is first level, then followed by criteria and
sub-criteria levels. At the bottom of the hierarchy tree various alternatives are present

34
that are compared according to above criteria within the tree levels. Analytical
hierarchy process has been widely used for classification of inventory in several
fields. Partovi and Burton (1993) applied the analytical hierarchy process to
inventory classification in order to include both quantitative and qualitative
evaluation criteria.

Braglia et al. (2004), has proposed a multi attribute decision making classification
model as a tool for spare parts inventory management in a paper operating company
by utilizing two different techniques: the reliability centered maintenance (RCM) and
the AHP. The Reliability centered maintenance analysis defines a decision diagram
which guides decision maker toward the best criticality classification for each type of
spare part and several AHP models are implemented and integrated with each node
of the decision tree. By these technique, Braglia et al. (2004) have avoided using
over-complex and unmanageable decision diagrams and have taken into account
numerous potential attributes influencing the spares inventory management policy in
an easy and rational manner. In their developed model, evaluation criteria taken into
consideration are group under four main classes: spare part plant criticality, spare
supply characteristics, inventory problems and usage rate. Several of sub criteria
analyzed in the model are quality problems, lead time, internal repair, price, space
required, obsolescence, deterioration problems, production loss, redundancies, etc.

Çakir and Canbolat (2008) have proposed an inventory classification system based
on the fuzzy analytic hierarchy process by integrating fuzzy concepts with real
inventory data and designing a decision support system assisting a sensible multi-
criteria inventory classification. Their study was conducted in a small electrical
appliances company and they validated the design of the proposed multi-criteria
inventory classification system and its underlying fuzzy AHP model.

Simunovic et al. (2008) has applied and compared the traditional ABC inventory
classification, based on one criterion, and inventory classification based on multi
criteria decision making technique, which is analytical hierarchy process on spare
parts in agricultural machine. For traditional inventory classification Simunovic et al.
(2008) has analyzed inventory with ABC analysis based on annual cost usage. On the
other hand, evaluation criteria involved in multi criteria inventory classification with
AHP methodology were annual cost usage, criticality and lead time1 and lead time 2.

35
The AHP methodology includes pair wise comparison of criteria, but not the pair
wise comparison of alternatives.

AHP analysis is also widely used in inventory management of railway systems. For
example, Sönmeztürk et al. (2008) have classified maintenance spare parts inventory
using multi criteria ABC analysis and fuzzy AHP. They used fuzzy analytic
hierarchy process model to define the major criteria used in inventory management
in the railway organization and estimated level of significance of the criteria. Then,
they classified the spare parts inventory according to following criteria: lead time,
unit cost, value usage and criticality. They used importance scale from 1 to 9 as in
most evaluation by AHP process models. AHP has been praised for its ease of use
and its inclusions of group opinions; however, the subjectivity resulting from the
pair-wise comparison process of AHP poses problems (Yu, 2011).

4.4 Data Envelopment Analysis

Ramanathan (2006) has developed a weighted linear optimization model that is based
on the concept of data envelopment analysis. In his work, the weighted additive
function is used which covers all the performances in terms of different criteria of an
item and optimization linear model is defined for each item. The model developed by
Ramanathan (2006) generated the optimal inventory score for each item as well as
weighted factor values (weights) for all the criteria. However, in a model with large
inventory items his method is time consuming. Several researchers have modified
and improved Ramanathan’s model. For example, Zhou and Fan (2007) have
extended Ramanathan’s model by incorporating some balancing features for multi-
criteria ABC inventory classification. They developed more reasonable and
encompassing index, since it used two sets of weights that are most favorable and
least favorable for each item and the purpose of the model was to aggregate multiple
performance scores of an item with respect to different criteria into a single score for
the subsequent ABC inventory classification (Zhou and Fan, 2007).

Ng (2007) has improved the method developed by Ramanathan by just simplifying

the model which was easier to apply. In his approach, Ng (2007), instead of using
linear optimizer, has transformed and normalized all the criteria data for each
alternative to the scale from 0 to 1. After that, partial averages of the transformed
criteria values are found and maximal partial average value for each item is chosen to

36
rank the items by the chosen value. At last, traditional ABC analysis could be
applied.

Data envelopment analysis is also widely used with ABC analysis of inventory in
railway systems management. Çelebi et al. (2008) have extended the classical ABC
analysis by developing a multi-criteria inventory classification approach for
supporting the planning and designing of a maintenance system for railway spare
parts inventory. In their research, the criticality model is proposed and ABC analysis
is performed according to the chosen criteria of criticality, lead times, value usage
and units costs. The weights of the evaluation criteria were generated by a DEA like
linear optimization to avoid the subjectivity on weights assignments. The DEA
model used is a simple input-oriented single input, multi output model which is very
similar to the model developed by Ramanathan (2006).

A similar method for estimation of weights of criteria as was done by Çelebi et al.
(2008) have been performed by Hadi-Venchen and Mohamanghasemi (2011) in their
examination of multiple criteria ABC inventory classification model. They have
integrated fuzzy analytical hierarchy process (FAHP) and data envelopment analysis
and applied their methodology with real case study. The proposed FAHP-DEA
methodology used the FAHP to determine the weights of criteria, linguistic terms
such as Very High, High, Medium, Low and Very Low to asses each item under each
criterion and the data envelopment analysis (DEA) method to determine the values of
the linguistic terms and the simple additive weighing method to aggregate item
scores under different criteria in an overall score for each item.

Data envelopment analysis (DEA) maximizes the artificial inventory score that is
used to classify each inventory item. Unlike AHP, the weights given to classified
criteria are solved automatically when the DEA model is optimized. Like statistical
clustering technique, this model must be reprogrammed and solved whenever a new
inventory item is introduced (Yu, 2011).

4.5 Artificial Neutral Network

ABC analysis has been popular and effective method used to classify inventory items
into specific categories that can be managed and controlled separately. Apart from
traditional ABC analysis, multi criteria ABC analysis has gained much attention and
many different models and techniques have been developed to improve classification

37
of inventory in multi item operations and processes. Advancement of computer
technology has enabled usage of complex database methodologies that helps manage
the decision-making process. Moreover, human involvement in the decision making
process has been reduced and thus it enhances the accuracy and consistency of the
decision making process, at the same time decreases the processing time. One of the
methods for multi-criteria inventory management has been artificial intelligence
(Guvenir and Erel, 1998). Many researchers have compared artificial-intelligence
based classification techniques with traditional multiple discriminant analysis. One
of artificial intelligence based technique is an artificial neural network (ANNs) which
is used widely for classification processes. Examples of these artificial neural
network models include support vector machines (SVMs), back propagation
networks (BPNs), and the k-nearest neighbour (k-NN) algorithm, etc. (Yu, 2011).
Neural network models have been used and developed to classify and control
inventory. The ANNs have two major strengths over the more traditional model
fitting techniques. First, ANNs are capable of detecting and extracting nonlinear
relationships and interactions within predictor variables. Second, the inferred
relationships and associated estimates of the precision of the ANN are not related to
the several assumptions about the distribution variables (Partovi and Anandarajan,
2002).

Artificial neural network have been used for classifications purpose, as well as for
forecasting problems in a variety of applications. They are very helpful for finding
nonlinear surfaces and separating the underlying patterns. Paliwal and Kumar (2009)
have performed a comprehensive study on neural network articles, categorizing the
application of networks into categories: accounting and finance, health and medicine,
engineering and manufacturing, and marketing. Accounting and finance is the
category with the greatest number of applications, especially with regards to
bankruptcy prediction, credit evaluation, fraud detection and property.

Partovi and Andarajan (2002) have performed artificial neural network for ABC
classification for stock keeping units in a pharmaceutical company. They compared
the two learning methods in ANN: namely back propagation and genetic algorithm.
Their results showed that there was no significant difference between two learning
methods used to develop ANN. The evaluation criteria used in the models were unit
price ($/unit), ordering cost ($/order), demand range (units/year) and lead time

38
(days). The results showed that neural network based classification models have a
higher predictive accuracy than conventional multiple discriminant analysis (MDA)
technique.

Li and Kuo (2008) have developed an enhanced fuzzy neural network (EFNN) based
decision support system for managing automobile spares inventory in a central
warehouse. In their system, the EFNN is utilized for forecasting the demand for spare
parts.

Chen, et al. (2009) have proposed moving back-propagation neural network (MBPN)
and moving fuzzy neuron network (MFNN) to effectively predict the critical spare
parts requirements in a wafer testing factory.

Other than classification problems, artificial neural network are used in clustering of
inventory. For example, Malakooti and Raman (2000) have utilized unsupervised
learning clustering ANN with variable weights for clustering alternatives and used
feed-forward for the selection of the best alternatives for each cluster of alternatives.
In their model, the learning mechanism of ANN was a generalized Euclidean
distance where by changing its coefficients new formation of clusters of alternatives
was achieved. The main difference between their approach of multi criteria
clustering and the general clustering method lies in the fact that their model deals
with criteria that have to be optimized and the selection of the best alternative.

The use of artificial neural networks in classification problems has become popular
among several fields. ANN are used in classification of three phase induction motors
(Ghate and Dudul, 2010), in planning, scheduling and analyzing a flexible
manufacturing systems (Kurt, 2003). Other researches include the benchmark of the
developed classification method to see their effectiveness, for example, Yu (2011)
has performed comparison of three different artificial intelligence techniques with
the multiple discriminant analysis techniques developed by Reid (1987), Flores et. al.
(1992), Ramanathan (2006) and Ng (2007). In his study, techniques used were back-
propagation network, support vector machine and k-nearest neighbor’s methods.
Back propagation networks (BPN) are the most widely used for classification method
for training an artificial neural network. A BPN utilizes supervised learning methods
and feed forward architecture to perform complex functions such as pattern
recognition, classification and prediction. Support vector machine (SVM) were

39
developed by Vapnik et al. (1995) and it employs structural risk minimization rather
than the empirical risk minimization used by conventional neural networks. SVM use
linear model to implement nonlinear class boundaries by the nonlinear mapping of
input vectors into a high-dimensional feature space (Yu, 2011). K-nearest neighbors
method is a non-parametric technique for classifying observations (Cover and Hart,
1967). Computations of the measure of distance or similarity between observations
are conducted prior to classification and then, newly introduced item is classified to
the group where the majority of k-NNs belong. As a result of his study Yu (2011)
have concluded that forecasting accuracy of classification with artificial intelligence
methods perform better than traditional MDA techniques like traditional
ABC (Reid, 1987), AHP (Flores et al., 1992), optimal score (Ramanathan, 2006), and
scaled score (Ng, 2007).

Ghate and Dudul (2010) have employed neural network classifier in their research
and have proposed optimal multi layer perception (MLP) network model for
classification model. Also, they have compared classification efficiency with
different network structure, the Kohonen self-organizing map (SOM) network. The
main purpose of their study was to reveal the importance of selection of significant
inputs and selection of neural network parameters, which make the structure
compact, and create highly accurate networks. In their work they have applied
different input network parameter and compared the classification accuracy to
determine the optimum network model. Results of network models with different
transfer functions, momentum and learning rates have been evaluated and the
optimum model resulted to be multi layer perception network with hyperbolic
tangent function (Ghate and Dudul, 2010).

Traditional ABC analysis should be replaced with multi-criteria classification

approaches in order to manage inventory more efficiently. Multi-class classification
utilizing multiple criteria requires techniques capable of providing accurate
classification and processing a large number of inventory items. AI-based
classification techniques such as back-propagation network and genetic algorithm
have proven to be efficient methods for classifying inventory items. The use of these
techniques will improve the effectiveness and efficiency of inventory management
(Yu, 2011).

40
4.6 Genetic Algorithms

Guvenir and Erel (1998) have applied genetic algorithm (GA) for multi criteria
inventory classification and have compared their results with classification efficiency
of analytical hierarchy process. Genetic algorithms are general-purpose search
algorithms that use principles inspired by natural population genetics to evolve
solutions to problems. The basic idea in genetic algorithms is to maintain a
population of knowledge structures (also called chromosomes) that represent
candidate solutions to the current problem. A chromosome is a sequence of genes.
The population evolves over time through competition and controlled variation. The
initial population can be initialized using whatever knowledge is available about the
possible solution. Each member of the population is evaluated and assigned a
measure of its fitness as a solution. After evaluating each structure in the populations,
a new population of structures is formed in two steps. First, structures in the current
population are selected for replication based on their relative fitness. High-
performing structures might be chosen several times for reproduction, while poorly
performing structures might not be selected at all. Then, the chosen structures are
modified using idealized genetic operators to form a new set of structures for
evaluation. The primary genetic search operator is cross-over operator, which
combines the features of two parent structures to form two off springs similar to their
parents. The role of the cross over operation is to form new fit chromosomes from fit
parents. There are many possible ways of crossover: the simplest swaps
corresponding segments of a string, list or vector representation of parents. In
generating new structures for testing, the cross over operator usually draws only on
the information present in the structures of the current knowledge base. If specific
information is missing, because of storage limitations or loss incurred during the
selection process of a previous generation, then crossover is not able produce new
structures than contain it. A mutation operator which changes one or more
components of a selected structure provides the way to introduce new information
into the knowledge base. In most cases, mutation serves as a secondary search
operator that ensures the ability to reach all points in the search space. The resulting
off springs are then evaluated and inserted back into the population. This process
continues until either, an absolute fittest structure is detected in the population or a
predetermined stopping criterion (e.g., maximum number of generations or

41
maximum number of fitness evaluation) is reached (Guvenir & Erel, 1998). In their
model, Guvenir and Erel (1998) have proposed a new cross over operation, called
continuous uniform crossover, such that it produces valid chromosomes given that
the parent chromosomes are valid. One other research which compared classification
performance of different artificial intelligence techniques on classificiaton of
inventory with multiple criteria ABC analysis was done by Partovi and Anandarajan
(2002) on stock keeping unit in pharmaceutical industry. Partovi and Anandarajan
(2002) have compared classification performance of back-propagation network and
genetic algorithm with multiple discriminant analysis technique. Their study has
revealed that both artificial neural network models had higher predictive accuracy
than MDA. The results also indicated that there is no significant difference between
the two learning methods used to develop the artificial neural network.

4.7 Other Multi Criteria Classification Techniques Widely Used in ABC

Analysis

4.7.1 Ng model
Ng- model was developed by Ng (2007) and his model is simple classification model
used for multiple criteria ABC classification. According to model all the
measurements are transformed to 0-1 scale according to below equation 4.1:

=>? 4@>AB,C,…? D=>? E

5F>AB,C,…? D=>? E4@>AB,C,…? D=>? E
(4.1)

For more detailed explanation of model please see Ng (2007). More detailed research
on Ng-model for inventory classification has been done by Hadi-Vencheh (2009)
where he has developed a nonlinear programming model which determines a
common set of weights for all items. The developed model incorporated multiple
criteria for ABC classification and at the same time maintained the effects of weight
in the final solution. The evaluation criteria analyzed in the model were annual unit
cost, annual dollar usage and lead time.

4.7.2 Case based model for multi criteria abc classification

Chen et.al. (2008) have introduced a case based multi criteria ABC analysis that
improves traditional ABC analysis by accounting for additional criteria such as lead
time and criticality in order to provide more managerial flexibility with inventory
classification. The main problem with many existing multi criteria decision aid

42
methods lies in the acquisition of the decision maker’s preference information in the
form of value or weights Chen et.al. (2008). Case based reasoning is an approach to
finding preferential information using cases selected by the decision maker (Jacquet-
Lagreze et al., 2004; Slowinski, 1995). The choice of cases may involve: past
decisions taken by the decision maker; decisions taken for a limited set fictitious but
realistic alternatives; decisions rendered for a representative subset of alternatives
under consideration, which are sufficiently familiar to the decision maker that they
are easy to evaluate v Chen et.al. (2008). A main advantage of case based reasoning
is “decision makers may prefer to make exemplary decisions than to obtain them in
terms of specific functional model parameters” (Doumpous, 2002). The model uses
decisions from cases as input, preferences over alternatives are represented
intuitively using weighted Euclidean distances which can be easily understood by a
decision maker.

4.7.3 Abc fuzzy classification

Chu et al. (2008) have purposed a new inventory control approach called ABC–fuzzy
classification (ABC–FC), which can incorporate manager’s experience, knowledge,
and judgment into inventory classification and can be implemented easily. Fuzzy
classification is a technique that uses the available information in a set of
independent attributes to predict the value of a discrete or categorical dependent
attribute. The main difference of their paper from earlier works is as follows: (1) the
approach can handle any combination of item attribute information that is important
for managerial purposes (e.g., the criticality of a stock-out, order size requirement of
the item); (2) manager’s preference for grouping based on operational performance
can be accommodated; (3) fuzzy statistical discrimination criteria are considered; (4)
our ABC–FC approach can be easily implemented on the spreadsheet, which is more
accessible to practitioners.

43
44
5. AN ARTIFICIAL NEURAL NETWORK MODEL FOR ABC
CLASSIFICATION OF SPARE PARTS IN RAILWAY SYSTMES

The usage of artificial neural network in inventory management especially, in spare

part classification is more precise than the traditional classification techniques. We
will extend the model developed by Çelebi et al. (2008) for classification of railway
spare parts by the classical ABC analysis and developing multi-criteria inventory
classification approach. The model will be developed for the inventory classification
using ABC analysis and ANN approach for spare parts in light railway system.
Maintenance operations play crucial role in the performance of railway vehicles and
they are very important for railway services and very significant to provide
uninterrupted and high quality service to passengers. To insure efficiency in railway
services availability and reliability on spare part control method should be provided
by maintenance operations.

5.1 Identification of Evaluation Criteria for ANN Model

Traditional ABC analysis for inventory classification is solely based on annual dollar
usage (Chen et. al., 2008). Although annual dollar usage is an important criterion in
inventory management systems, other criteria may also deserve management’s
attention in classification of inventory. Chen et. al. (2008) states that classification
obtained from traditional ABC analysis sometimes should undergo further
adjustments. For example, annual dollar usage of inventory may not be significant,
but inventory’s stock-out penalty cost may be sufficiently high and vice versa. In
such cases, inventory manager has to pay attention to other significant criteria, which
will influence his/her decision on how much inventory to keep in stock or which
inventory should gain close stock control. These criteria for inventory control,
suggested in different literature, may be ordering cost, demand distribution,
obsolescence, criticality of part, lead time, number of requests, substitutability,
scarcity, perish ability, durability, stock ability, reparability, criticality,
commonality, etc. (Flores and Whybark, 1986, 1987; Cohen & Ernst, 1988; Partovi

45
and Anandarajan, 2002; Ramanathan, 2006; Zhou & Fan, 2007; Chen et al., 2008;
Hadi-Vencheh, 2010 ).

Evaluation attributes suggested by Partovi and Anandarajan (2002) for evaluation of

spare parts model by ANN are unit price, ordering cost, demand and lead time.
These criteria suggested by Partovi and Anandarajan (2002) may be sufficient for the
evaluation, however, criteria used by Çakır abd Canbolat (2008) given in Table 5.1
seems more efficient to be chosen for the evaluation for spare part inventory.

Table 5.1: Subject criteria definitions (Çakır and Canbolat, 2008).

Criterion name Definition

Price/cost Price of the inventory item in case it

is outsourced and cost in case it is
manufactured in the company
Annual demand Annual demand of the inventory
item in the production process
Blockade effect in Critical threat on the stability of
production process when the item is
case of stockout not available for
production right on time
Availability of the substitute Availability and conformance of
material substitute materials
Lead time Time required to get the item
available on hand for production.
Delivery time in case the item is
outsourced and production time in
case the item is manufactured
Common use Common usage of the item in the
product family

According to the nature of inventory items and industry, the criteria by which
inventory is classified will have different weights. In the real world, prioritizing of
the weights of criteria is always subjective that is depending on the conditions
governing on industry and market, inventory managers usually assign different
weights to the criteria of their inventory (Hadi-Venchen and Mohamadghasemi,
2011). For instance, when items are guaranteed to be supplied at certain time, the
weight of lead time criteria is set lower than other criteria during evaluation.
Moreover, it is stated that importance of criteria might differ among several
industries (Hadi-Venchen and Mohamadghasemi, 2011). In hospitals, inventory’s
expiration date criteria will be important criterion; however, for manufacturing
industry like cement this criterion will be insignificant. Similar is applicable for

46
maintenance spare parts inventory in railway systems, there are criteria that are more
important than several others. Çelebi et al. (2008) have evaluated spare parts
inventory in light railway systems according to following criteria: criticality, value
usage, unit cost and lead time.

Value usage criterion, similar as in classical ABC classification, is known as annual

value usage where items are evaluated through their annual monetary usage values
which are the products of annual usage quantities and the average unit prices of the
items. The value of an item represents the common control characteristic of items
and it is avoided to stock large amounts of high value items. Usually, a high value
item favors controlling of inventories of that item more carefully. Thus, this criterion
is positively related to the importance of the item because items with the highest
annual monetary value receive the most attention where low monetary value items
are controlled routinely (Çelebi et al., 2008).

Lead time criterion plays an important role among several factors that can affect the
management of the inventory. Lead time of spare part item includes the time between
placing order at the supplier of the item and the moment the item is available for
usage. Both the length of the lead time and its variability might be crucial in
maintaining adequate supply of an item without excessive costs (Flores and
Whybark, 1985). The period of lead time is important since it directly establishes the
stock levels of items with unknown demand and states the response time to a crisis.
The variability affects the amount of safety stock necessary to provide the level of
desired service. For instance, items with long lead times may incur financial losses as
a result of possible interruption of maintenance operations or large inventory stocks.
Usage of time as a measure basis of lead time criterion constitutes a common basis
for all items in the product catalog and also provides both the user and the supplier
with a common means of setting the goals and for controlling the performance of
operations.

Unit cost is another criterion and is important in control of items on strategic and
tactical level. High value items necessitate decisions like setting up the incentives,
developing coordination with suppliers and creating opportunity for high negotiation
power. On the other hand, the replenishment arrangements with low price items have
to be efficient so that the administrative costs do not increase unreasonably in
proportion to the value of the items themselves (Çelebi et al., 2008).

47
The criticality analysis includes a substantive amount of subjective criteria used in
evaluating the criticality of parts practice (Cohen et al., 1997). It is related to the
results of damage of a failure and shortage of inventory and the possibilities to
control the situation. The criticality was divided into three sub-criteria: commonality,
substitutability and stock-out penalty which covers the essentials of railway
maintenance systems (Çelebi et al., 2008).

a. Stock-out penalty criteria is the costs associated with the spare part being out
of stock. The criticality of a spare part item is related to the consequences
resulted from the lack of the item when it is necessary. In theory, stock-out
penalty can be evaluated by estimation of downtime expenses caused by
failure to be corrected by the use of the part. Yet, it is usually difficult to
determine the degree for criticality in practice and it might be sufficient to
define a few degrees of criticality for practical purposes. One known
approach is to relate criticality to the time in which the failure is required to
be corrected. Huiskonen (2001) has mentioned three degrees of criticality in
regards with consequences might be estimated as the following bases:
• High: The failure needs to be repaired and the spare parts have to
be provided immediately,
• Moderate: The failure can be tolerated with temporary
adjustments for a short time while spare part can be provided,
• Low: The failure is not critical for the operation and it can be
corrected after a long period of time when spare is supplied.

The quantification of the given criticality degrees is done through assigning a

penalty index ( @) for each item n and setting it 5 for a high level item, to 1
for a low level and to 3 for a moderately critical item (Sönmeztütk et al.,
2008).

b. Substitutability criterion is another important aspect of criticality. The

substitution potential provides flexibility in response to problems, reducing
the importance of the item in the spare parts inventory. If the item has a close
substitute, more flexibility and reduced response time is possible, both of
which reduce the criticality of the part.

48
Substitutability of spare part has a direct influence on the purchasing decisions
which may affect both efficiency and effectiveness of maintenance operations.
One dimension of substitutability is technical. Technical similarities between
spare parts inventory can allow mutual substitution avoiding any loss of function
or suitability. Higher level of substitutability might enable lower levels of stock-
out risks that when there is lack of one type of item, it can easily be replaced with
interchangeable item. However, in practice there are some situations where
replaceable item is suitable as an alternative for the item in question and yet it
may cost so much higher that it becomes a poor substitute (Çelebi et.al, 2008).

Other dimension of substitutability is connected to the availability of different

suppliers who can provide the similar product with very little or zero quality and
cost difference. Among the wide range of maintenance spare parts there are
typically two type, standard type parts and special parts. Standard parts are
widely used by many users and are readily available from many suppliers; on the
other hand, special parts are specifically tailored for and used by a particular use
only. The availability for standard parts is always good and there are stocks of
these parts at different levels of the supply chain and the suppliers are eager to
cooperate with the users as volumes are high and they offer economies of scale.
However, for user-specific parts quite the opposite is valid: suppliers are not
interested to keep low volume special items in stock and user is responsible from
the availability and control of the special items himself. Similar to stock-out
penalty index, a substitutability index might be a method of determining the
substitution availability of an item. A substitutability index of G = 5 implies that
item i is fully differentiated, G = 1 implies that it has perfectly substitutable
products and G = 3 implies that item i is moderately substitutable.

c. Commonality criterion is another aspect of criticality which is the measure of

how many uses there are for a spare part item. If the item is used in many
different vehicles or maintenance types, it might be important to devote extra
attention to it and for management purposes it could be classified in a group
of A items. The usage of common parts can be beneficial in terms of risk
sharing and substantial savings up to can be achieved by use shared stocks
compared to using separate stocks (Kranenburg and Van Houtum, 2007).
Moreover, when the same spare part item can be used with several

49
 maintenance types or vehicles, the system allows the possibility of economies
of scale since a common component can be supplied in large volumes (Çelebi
et al., 2008). However, a situation when common components are out of
stock will have higher influence on the maintenance system and as a result
the maintenance schedules that have a use of shared component will be
delayed or changed. Measurement of commonality is not so easy. Though
there are several types of commonality indexes available in literature (Lyly-
Yrjanainen et al., 2004), a simple and useful measure is the number of
different maintenance types that the use of the item is needed.

In our model criticality will be analyzed under above three sub-criteria: namely
stock-out penalty, commonality and substitutability and other criteria: value usage,
lead time and unit cost as was suggested by Celebi et al. (2008) in their evaluation of
criticality of spare parts inventory model. The dimensions of criteria are summarized
in Table 5.2 as chosen to be as was suggested by Çelebi et al. (2008).

Table 5.2: Dimension of evaluation criteria for classification of spare parts.

Criteria Dimension
Unit cost Turkish lira (TL)
Lead time Weeks
Value usage Unit
Commonality Number of different
usage
Substitutability 1. Low
3. Medium
5. High
Stock out penalty 1. Low
3. Medium
5. High

5.2 Network Building

Designing a neural network requires the determination of several parameters and

variables that define the network and that influence the learning performance of the
network.

The most important choice is the determination of the necessary architecture, the
number of layers, the number of processing elements in every layer and the number

50
of synapses that are connected to each node. Moreover, it is also important to choose
the appropriate activation function of the hidden and output nodes, the training
algorithm, data transformation and normalization methods, training and testing sets
and the performance criteria (Zhang et al, 1998).

5.2.1 Network structure

Artificial neural network structure may differ according how nodes and their
connections are linked (Elmas, 2003). A network is said to be feedforward network if
it contains no directed cycles back, otherwise it is called as recurrent network.

Most research models use a three layer network constituting one input, one hidden
and one output layer (Partovi and Anandarajan, 2002; Çelebi and Bayraktar 2008;
Turer et al., 2009). Hidden nodes with their activation function are required to
introduce nonlinearity in the neural network. Most popular network model used for
multi criteria inventory classification is multi layer perception network (Partovi and
Anandarajan, 2002; Çelebi and Bayraktar 2008; Turer et al., 2009).

5.2.2 Multi layer feedforward networks

Networks that consist of a set of neurons that are logically arranged into two or more
layer are known as multi layer feedforward neural networks (Masters, 1993). The
architecture of the multi layer network constitutes of one input and one output layer
and additional one or more hidden layer which have computation nodes that are
called hidden neuron or processing elements.

Hidden units with their activation function are used to capture the nonlinear
structures in the system observed (Qi and Zhang, 2001). The ability of hidden
neurons to extract nonlinear structures is particularly important when the systems
under consideration is complex, if the size of the input layer is large and the
underlying relationship between the input and output units are undeterminable.

A Multi layer feedforward neural network, especially the one hidden layer
feedforward type is the mostly used model form in literature (Masters, 1993). For
the classification of spare part invetory in railway system Multi-layer perception
(MLP) model was chosen because it has simple architecture and algorithm (Turer et
al., 2008). Moreover, it has been successful in solving approximation and
classification problems (Çelebi and Bayraktar, 2008). Generally, MLP consists of a
group of nodes which form an input layer, one hidden layer and last, an output layer.

51
The popularity of the one hidden layer feedforward networks come from the fact that
they are universal approximators (Zhang, 2003). Regarding the two hidden layer
networks, it has been stated in the literature that two hidden layer network slows the
learning rate and affect the overall performance of the network (Şahin, 2002;
Masters, 1993).

5.2.3 Learning rule in neural network

All knowledge in an artificial neural network is encoded in its interconnection
weights among the neurons calculated by the learning process. A weight represents
the strength of associated among connected features, concepts propositions or events
during a training phase. A neural network learns by adjusting according to learning
method by which the interconnection weights are changed appropriately. As a result,
learning can be explained as the modification in weights in the layers in order to
perform the required performance criteria (Haykin, 1999).

Learning method for the artificial neural network model was taken as supervised
learning, specifically back-propagation algorithm as was suggested by Partovi and
Anandarajan (2002) for ABC classification in their network building method. The
multilayer perception back propagation model suggested by Partovi and Anandarajan
(2002) is given below Figure 5.1:

Figure 5.1: Artificial neural network structure (Partovi and Anandarajan, 2002).

Hidden layers are connected to previous layer by means of synapses and also to
output layer. In the hidden layer, activation function is generally hyperbolic tangent
or logistic sigmoid (Çelebi and Bayraktar, 2008; Turer et al. 2009; Partovi and

52
Anandarajan, 2002; Aksungur and Kavlak, 2009). MLP performs approximation or
classification; as a result the output of neurons activation function can be linear or
nonlinear.

The model will contain one input layer, one hidden layer and one output layer. In our
model there will be six criterion nodes, which are stock out penalty, commonality,
substitutability, value usage, unit cost and lead times, in the input layer and three
nodes for the output layer, namely inventory items A, B and C. Total data contained
in the evaluation consist a values of 71 data set of different spare parts. Complete
data set has been divided into two subsets, the training and the test set, where 85 %
(70% training and 15% cross validation data) of the total data is used as training set
and the rest 15 % of the total data used as testing set to evaluate the accuracy and
generalization ability of the network. In the evaluation data set of 60 data of the spare
parts was used as a training set and for validation of the developed neural network
sample of 11 data of spare parts was used as a testing set.

5.2.4 Choice of transfer function

The transfer function is important and it depends on the nature of the output of the
network. There are a number of choices including the step function, sigmoid
function, hyperbolic tangent function and linear function among others. Turer et.al.
(2009) have preferred to utilize hyperbolic tangent function for their ANN for
supplier evaluation process. Furthermore, it has been specified in several
examination of different modeling of artificial neural network approaches with MLP
for classification problems that hyperbolic tangent function, in equation 5.1, results
in more efficient and reliable solution because of its nonlinear behavior which allows
networks to model nonlinear mappings. Also, it has a larger range [1,+1] in contrast
to [0,1] as in logistic sigmoid that corresponds to a greater weight range (Celebi and
Bayraktar, 2008).

JKℎ  =
M N 4M ON
M N 0M ON
(5.1)

 = ∑@
  (5.2)

On the other hand, sigmoid tangent function is also widely used in classification
models. For example, Partovi and Andarajan (2002) have used sigmoid transfer
function in their ANN model as well as Aksungur and Kavlak (2009). The sigmoid
transfer function is illustrated in eq. 5.3

53
P#QR  =

0M ON
b (5.3)

 = ∑@
  (5.4)

5.2.5 Determination of hidden layer number and nodes in hidden layer

The aim of neural network is to detect the hidden relationship in the data and to
perform nonlinear mapping between input and output variables which is know as
training of network. As have been mentioned before, the neural network structure
chosen for classification model includes one input, one hidden and one output layers,
which was chosen for its simplicity. The number of nodes in the input and output
layers are predefined according to the evaluation criteria and classification model.
Yet, the number of nodes in the hidden layer has to be determined by trial and error.
The number of neurons in the hidden layer was calculated based on five-fold-cross
validation method which was suggested by Çelebi and Bayraktar (2008).
In five-fold cross validation method, training data is divided into five disjoint subsets
and one of these subsets is hold as validation data and the rest four subsets are
utilized as training data set. The network is run five times by 1000 iterations and by
changing the number of nodes from two to twenty in the hidden layer. The results of
training gave the minimum of MSE values for both training and cross validation data
which were noted. Then, the first subset input back into training set and second
subset is chosen as validation set and training is repeated again by noting the
minimum MSE for the second subset. This procedure is performed till all five
subsets are used as validation sets and we result with data of minimum MSE for five
subset sets for different number of nodes in hidden layer (Çelebi and Bayraktar,
2008). MSE, mean squared error function calculates the accuracy rate of the
configuration model by determining the difference between the output of the model
and the desired output as can be seen in equation 5.5.

× ∑@
 −  

@
MSE= (5.5)

Finally, the average of minimum MSE for five subsets is calculated and the network
model with hidden nodes which has the least average of minimum MSE for cross
validation data is chosen as the optimum configuration (Çelebi and Bayraktar, 2008).
To sum up, by five fold cross validation method the number of nodes the hidden
layer is calculated by a stepwise training of five subsets from training data in order to
select the model which minimizes the generalization error.

54
For the determination of hidden nodes in the hidden layer in classification of spare
part inventory, evaluation data for the five fold cross validation method consists of
60 spare parts data, of which 48 of data is taken as training and 12 data as validation
set. The structure of neural network which will have the minimum of Mean Squared
Error (MSE) at the end of training process will be chosen for the evaluation of the
testing data. Data for 11 spare parts is left out for the testing process to check the
validity of the trained network.

5.2.6 The stopping criteria, learning rate and momentum

The learning process should be stopped when network reaches a predefined criterion.
Generally mentioned in literature, the training is stopped either when training
performance reaches a target performance or when a specific epoch number is
reached. In our multi layer perception network model, the stopping criterion for the
training process is taken to be the minimum of mean square error (MSE) or a
network reached certain number of epochs. In their evaluation , Çelebi and
Bayraktar (2008) have stopped training of the neural network as epoch number has
reached 1000 iterations or the MSE for cross validation set is decreased below the
threshold value of 0,01.

The learning rate of the network is dependent on the step size and momentum in each
layer. Ghate and Dudul (2010) have evaluated optimal parameters for the multi layer
perception model with three layers and have found that optimal step sizes should to
be 1.0 and 0.1 for hidden and output layers respectively and moreover, for the
momentum for hidden and output layers the optimal values to be 0.7 as can be seen
in the below Figures 5.2 and 5.3:

55
Figure 5.2: Variation of average minimum mse on training dataset with step size
of hidden layer (Ghate and Dudul, 2010).

Figure 5.3: Variationof average minimum mse on training dataset with momentum
output layer. (Ghate and Dudul, 2010)

To sum up, in the evaluation of spare parts inventory by ABC classification with
neural networks, the network parameters for the neural network was chosen as
mentioned below in Table 5.3.

56
Table 5.3: Network parameteres for the ann model
Learning Method Supervised
Learning model and type Multi layer perception with back
propagation algorithm
Performance Function Mean Square Error (MSE)< 0,01
Number of Hidden Layer 1
Number of Neurons in Hidden Layer To be determined by Five Fold Cross
Validation Method
Transfer Function 1. Hyperbolic Tangent
2. Sigmoid
Number of Input Neurons 6 (Lead time, Value Usage, Unit Cost,
Stock-out Penalty, Substitutability and
Commonality
Number of Output Neurons 3 (A,B,C)
Hidden Layer Step Size 1.0
Output layer Step Size 0.1
Momentum for Hidden and Output 0.7
Layers
Number of Maximum Epoch 1000 iterations
Learning mode Sequential (on-line)

5.3 Neurosolutions 5.0 Software Used for Classification Model

Neurosolutions 5.0 software was used to run the neural network model for railway
spare part inventory classification using ABC analysis. Neurosolutions 5.0 is
incorporated package in Excel and it has very convenient tools such as
Neurosolutions for Excel, Neurobuilder and Neuroexpert. The data set of 71 spare
parts was entered in Excel file as can be seen in Table A.1 in the App. A.1

For the modeling the data has been randomized before the building neural network.
As was mentioned in previous sections, 85% (70% training and 15% cross
validation)of data was selected as training and 15% of data was used as testing data
as was suggested by Ghate and Dudul (2010). Evaluation data for 48 spare parts was
chosen as training, 12 data as cross validation and 11 data for spare parts was taken
as testing to validate to results of the training process.

57
5.4 Training Process by Neurosolutions

As was mentioned in previous sections, the number of hidden neurons in the hidden
layer is determined by five-fold cross validation method.

5.4.1 Five-fold cross validation method

The training data set is divided into 5 disjoint subsets; one of these subsets is retained
as cross validation data and other are selected of training data subset. The training
data has been divided into 5 subsets as can be seen in following Table A.2 in App.
A.2.

For a different number of hidden neurons, the network is run 5 times by 1000
iterations by choosing one subset as cross validation data from the divided 5 subsets
till all subsets has been used as validation data at once. The minimum of MSE data
for training and cross validation data is recorded for each subset. At the end, we end
up with minimum MSE data of training and validation data for each five subsets for
nodes from two to twenty in the hidden layer. The results of training process for five
fold cross validation method can be seen in Table B.1 and Table B.2 in App. B.1

The average of minimum MSE of five subsets is calculated for each number hidden
neuron in the hidden layer and results are presented in Table 5.4 and Table 5.5. The
example of calculaion of average minimum MSE can be seen in App. B.2.

The neural network was run for two different activation functions, hyperbolic tangent
and sigmoid functions.

As can be seen from Table 5.3, the neural network with 8 nodes in hidden layer gives
the least MSE data as 0,03736, in the five-fold cross-validation method for
hyperbolic tangent function.

58
Table 5.4: Average minimum mse for training and cross validation data with
hyperbolic tangent function.
AVERAGE OF MINIMUM FOR MSE
# Of Training Data Cross Validation Data
Nodes
2 0,085196608 0,060057349
3 0,104065011 0,080212494
4 0,098453339 0,063394639
5 0,103797196 0,058799254
6 0,108533403 0,051304857
7 0,124705773 0,078014818
8 0,117554536 0,037361614
9 0,119401166 0,070503936
10 0,102225101 0,050416581
11 0,122653927 0,067211223
12 0,121354667 0,054275056
13 0,150951974 0,099301261
14 0,1308181 0,087902196
15 0,14452272 0,092461572
16 0,131054975 0,109547602
17 0,136192376 0,091467151
18 0,163597183 0,08304966
19 0,138571933 0,079926288
20 0,130147528 0,088493876

Similarly, the neural network with sigmoid transfer function has least minimum
average MSE at 5 nodes in the hidden layer as 0,029434 as a results of five fold cross
valdation method.

59
Table 5.5: Average minimum mse for training and cross validation data with
sigmoid function.
Average Of Minimum Mse
# Of
Nodes Training Data Cross Validation Data
2 0,012802124 0,039772375
3 0,010029668 0,036783332
4 0,010363869 0,037398745
5 0,008184525 0,029434182
6 0,007971996 0,034192718
7 0,007298897 0,030355518
8 0,007575996 0,034793708
9 0,007062787 0,036389971
10 0,00716824 0,033420092
11 0,006585697 0,030453519
12 0,006237987 0,037203207
13 0,007348075 0,036478183
14 0,007059592 0,031441525
15 0,006556774 0,034703476
16 0,0068532 0,036517449
17 0,006869229 0,036944207
18 0,006975038 0,035741524
19 0,006835798 0,035016715
20 0,006801946 0,036829793

5.4.2 Training proposed artificial neural network model

After the number of hidden neurons in hidden layer has been determined to be 8 and
5 for hyperbolic tangent and sigmoid function models respectively, networks have
been trained with 60 spare part data (48 training and 12 cross validation data) 2 times
each by 10 runs and 1000 iterations with randomized initial weights in order to avoid
network to be stuck at local minima. As can be seen from the results of training
network with hyperbolic tangent function has minimum mean squared error is of
0,003122 and sigmoid function network has minimum mean squared error of
0,022120 in Table 5.6.

60
Table 5.6: Results of training for hyperbolic tangent and sigmoid transfer
functions

Best Hyperbolic Sigmoid

Networks Tangent Function
Run # 1 2
Epoch # 490 1000
Minimum
MSE 0,003122 0,022120

For network with hyperbolic tangent function, training has been stopped at 1st run
and 490th iteration. Similarly, training for network with sigmoid function has been
stopped at 2nd run and 1000th iteration.

As a result, we can see that neural network with hyperbolic tangent function has
minimum MSE of 0,003122 compared to 0,022120 of the sigmoid function and it is
chosen to be most optimal network model for the spare part multi criteria ABC
classification. The neural network model with hyperbolic tangent activation function
has minimum MSE when hidden layer has eight nodes. The architecture of the
proposed artificial neural network model is given in Figure 5.4.

Input layer Hidden layer Out put layer

Figure 5.4: Proposed artificial neural network model.

The network chosen consists of 6 nodes in input layer, 8 nodes in hidden layer and 3
nodes in output layer.

5.5 Testing of the Neural Network

In order to validate the training results, 11 spare parts data, given in Table 5.7,
previously not introduced to network were used as testing.

61
Table 5.7: Data for 11 spare part used for testing of the network.

Stock-
ID out Value Lead
no. Penalty Substitutability usage Times Unit Cost Commonality A B C
61 3 3 2 5 1484,71 8 0 1 0
7 2 5 850 1 0,465 21 0 0 1
71 1 5 1 1,5 105 8 0 0 1
1 5 5 10 3 22 12 0 1 0
3 5 5 6 1 64,53 11 0 1 0
56 5 5 14 4 32,21714 6 0 1 0
62 1 5 3 1 1,22 14 0 0 1
40 1 5 211 2 1,1 3 0 0 1
58 2 5 5 1 5,85 11 0 0 1
53 5 3 2 5 2963,1 5 0 1 0
22 5 1 9 24 4174,138 1 1 0 0

The results of overall testing of the network are given in Table 5.8.

Table 5.8: Results of testing process for proposed neural network.

Results for overal network after testing

Hyperbolic tangent
MSE 0,006483
NMSE 0,01037
r 0,9999
% ERROR 1,038739

The results of neural network model for multi criteria ABC classification of 11
testing data can be viewed in Table 5.9. The calculation of the MSE, NMSE, R and
% Error values can be in App. B.2.

62
Table 5.9: Output matrix of the neural network for 11 testing data.
Output /
Desired A B C
A 1 0 0
B 0 5 0
C 0 0 5

The output matrix gives the ratio of output of the model and the actual desired data.

The performance of the network over the 3 output desired parameters A, B and C set
can be seen in Table 5.10:

Table 5.10: Performance of classification of network for 11 testing data according to

3 desired output sets, a, b and c.

Performance A B C
MSE 0,002802827 0,001439141 0,001760844
NMSE 0,033914208 0,005804534 0,007102071
MAE 0,051057237 0,03160245 0,036232916
Min Abs Error 0,006789058 0,000959325 0,013045936
Max Abs Error 0,055555555 0,055553772 0,055555538
r 0,999999898 0,999969019 1
Percent Correct 100 100 100

As can be seen from performance measurements the correlation, r, for A set is

0,9999, for B set 0, 99996 and for C set is 1,00. The overall correlation ( r) of the
testing is 0,9999 for the testing data. The mean absolude error for the overall testing
network is obtained to be 1,038 % which shows that developed neural network has
results in acceptable solution.

The comparison of neural network output and desired data are given in Table 5.11.

63
Table 5.11: Desired data and output of neural network
A B C A Output B Output C Output

0 1 0 -0,0555551 1,05555374 -0,0555548

0 0 1 -0,0555556 0,01518085 0,98695406
0 0 1 -0,0555556 0,01518085 0,98695406
0 1 0 -0,0555551 1,05555377 -0,0555548
0 1 0 -0,0555551 1,05555377 -0,0555548
0 1 0 -0,0551998 1,05206318 -0,0555555
0 0 1 -0,0555556 0,01518085 0,98695406
0 0 1 -0,0555556 0,01518094 0,98695402
0 0 1 -0,0555556 0,01518005 0,98695279
0 1 0 -0,0551977 1,05203963 -0,0555555
1 0 0 0,99321094 0,00095933 -0,0555555

64
6. CONCLUSION AND RECOMMENDATIONS

Spare parts inventory management has always been a very important part in almost
all organizations. In railway systems management, main attention is given to
maintenance operations, because it directly influences the performance of railway
vehicles and play a crutial role in railway services to provide uninterrupted and high
quality service to passengers. Except from preventative maintenance, the demand of
spare part for maintenance tasks is generally random; thus, quick and reliable
management of the spare parts inventory is an important factor for the successful
implementation of the maintenance process. The purpose of this thesis is to
investigate multi criteria ABC inventory classification of spare parts in railway
systems, based on one method of artificial intelligence, which is an artificial neural
network.

The traditional ABC classification method uses only the unit price and the annual
usage of inventory items in ranking. Yet, in some cases, the classification done using
only these two criteria turns out to be insufficient. On the other hand, the new
method enables the integration of multi criteria into ABC classification. Firstly,
multiple criteria have been determined from a various criteria mentioned in various
literatures and most important ones have been chosen as evaluation criteria for the
ABC classification model of spare parts inventory. The classification model on
railway systems spare parts have been performed based on the following criteria:
annual unit cost, lead time, value usage, substitutability, commonality and stock-out
penalty. Then, in order to estimate the priorities among the criteria, the optimum
network architecture and network parameters is determined and neural network is
built. The chosen network constitutes of one input layer, one hidden layer and one
output layer. The suitable network algorithm is chosen as supervised learning
network type which is multi-layer perception based on back-propagation algorithm.
The number of nodes in the hidden layer have been estimated by five-fold cross
validation method, which is a good estimation method widely used in literature. The
network have been trained with two different activation functions, which are
hyperbolic tangent and sigmoid functions, in order to compare classification

65
accuracies and choose more suitable network model. The number of nodes in the
hidden layer has been found to be eight for hyperbolic tangent neural network and
five for neural network with sigmoid transfer function. It was found that hyperbolic
tangent function gives more precise results with MSE of 0,003122 compared to the
MSE of 0,022120 of the sigmoid function and it is chosen to be most optimal
network model for the spare part multi criteria ABC classification. The developed
artificial neural network architecture has 6-8-3 structure, which constitutes of six
nodes in the input layer, eight nodes in the hidden layer and three nodes in the output
layer. Six nodes in input layer represents the evaluation criteria of classification
model which are annual unit cost, lead time, value usage, subctitutability,
commonality and stock-out penalty. The three nodes in output layer are the three
group of A,B and C item classes.

The training of network has been done with 60 data of different spare part items and
classification efficiency of the developed ANN was tested by 11 data spare part
items. The evaluation data was taken from company performing maintenance
operations of light railway system.

The comparison of actual data and output of the network is given in Table 5.9.

The testing of developed neural network for classification spare parts items gave
mean squared error of the overall network as 0,006483 and overall error as 1,038 %.
The results also indicate that neural network has been trained enough which can be
seen from the classification accuracy of 99% of 11 data used for testing the
developed artificial neural network.

Inventory management in Turkish Railway Systems is not widely investigated area in

the literature. The previous researches on multi criteria inventory management have
been done using analytical hierarchy process and data envelopment analysis (Çelebi
et al., 2008; Sönmeztürk et.al., 2008). The main contribution of this thesis is that
different technique has been applied, which is artificial neural network, for
classification of spare parts inventory of light railway systems based on multi criteria
ABC analysis. The developed network has shown good classification accuracy and
this network can be used for classification of other real-world inventory data in
railway systems maintenance.

66
REFERENCES

Aksungur, S. and Kavlak, K., 2009: "Realization Of The Collision-Free Motion Of

A Scara Robot In An Environment With Obstacles Using Artifical
Neural Networks And Genetic Algorithm", Journal of Technical-
Online, Vol. 8, No. 2, pp. 112-126.
Apaydın, E., 2004: Introduction To Machine Learning, The MIT Press, London,
England.
Braglia, M., Grassi, A. and Montanar, R., 2004: "Multi-attribute Classification
Method for Inventory Management”, Journal of Quality in
Maintenance Engineering, Vol. 10, No.1, pp. 55-65.
Bonnett, Clifford F., 2005: Practical Railway Engineering.London,GBR: Imperial
College Press.
Cakir, O. and Canbolat, M.S., 2008: “A web-based decision support system for
multi-criteria inventory classification using fuzzy AHP methodology”,
Expert Systems with Applications, 35, 1367–1378
Chen, Y., Li, K.W., Kilgour, D.M. and Hipel, K.W., 2008: "A case-based
distance model for multiple criteria ABC analysis", Computers &
OperationsResearch, 35, 776 – 796.
Chen, F.L., Chen, Y.C. and Kuo, J.Y., 2009: “Applying Moving Back
PropagationNeural Network and Moving Fuzzy Neuron Network to
PredicttheRequirement of Critical Spare Parts”, Expert Systems
with Applications, Vol. 36, No. 7, pp. 4358– 4367.
Chen J-X., 2011: "Peer-estimation for multiple criteria ABC inventory
classification, Computers and Operations Research, Vol. 38, No. 12,
pp. 1784-1791.
Chu, C.W, Liang, G.S and Liao, C.T., 2008: Controlling Inventory by combining
ABC analysis and fuzzy classification, Computers and Industrial
Engineering, 55, 841-851.
Çelebi, D. and Bayraktar, D., 2008: "An integrated neural network and data
envelopment analysis for supplier evaluation under incomplete
information”, Expert Systems with Applications, 35, 1698–1710.
Çelebi, D., Bayraktar, D. and Aykaç, D.S., 2008: Multi Criteria Classification for
Spare Parts Inventory, CIE.
Çınar, D., 2007:Hidroelektrik enerji üretiminin hibrid bir model ile tahmini, Master
Thesis, ITU Institute of Science and Technology, Istanbul.
Cohen, M. A and Ernst, R., 1988: "Multi-item classification and genetic inventory
stock control policies", Production and Inventory Management
Journal, Vol. 29, No.3, pp.6-8.

67
Cover, T. M., and Hart, P. E., 1967: "Nearest neighbor pattern classification",
IEEE Transactions on Information Theory, 13, pp. 21–27.
Diakoulaki, D., Zopounidis, C., Mavrotas, G. And Doumpos, M., 1999: "The use
of a preference disaggregation method in energy analysis and policy
making", Energy–The International Journal, Vol. 24, No. 2,
pp. 157-166.
Doumpos, M., 2002: Multicriteria Decision Aid ClassificationMethods. Secaucus,
NJ, USA: Kluwer Academic Publishers, p. 4.
Elmas, Ç., 2003: Yapay Sinir Ağları, Seçkin Yayınevi, Ankara.
Ernst, R. and Cohen, M.A., 1990: "Operations Related Groups (ORGs): A
Clustering Procedure for Production/inventory Systems”, Journal of
Operations Management, Vol. 9, No.4, pp. 574-598.
Fisher, R.A., 1936: "The use of multiple measurements in taxonomic problems."
Annals of Eugenics,7, pp. 179–188.
Flores, B.E. and Whybark, D.C., 1985: “Multiple Criteria ABC Analysis”,
Journal of Operations and Production Management, Vol. 6, No. 3, pp.
38-46.
Flores, B. E., Olson, D. L., and Dorai, V. K., 1992: "Management of multicriteria
inventory classification", Mathematical and Computer Modeling,
Vol. 16, No. 12, pp. 71–82.
Guvenir, H.A. and Erel, E., 1998: “Multi-criteria inventory classification using a
genetic algorithm”, European Journal of Operational Research, 105,
pp. 29–37.
Ghate, V.N. and Dudul, S.V., 2010: “Optimal MLP neural network classifier for
faultdetection of three phase induction motor”, Expert Systems with
Applications, 37, pp. 3468–3481.
Haykin, S., 1999: Neural Networks: A Comprehensive Foundation, Prentice Hall,
New Jersey, USA.
Hadi-Venchen, A., 2010: “An improvement to multi criteria ABC inventory
classification”, European Journal of Operational Research, 201,
pp. 962-965.
Hadi- Venchen, A. and Mohamadghasemi, A., 2011: “A fuzy AHP-DEA approach
for multiple criteria ABC inventory classification”, Expert Systems
with applications, 38, pp. 3346-3352.
Huiskonen, J., 2001: “Maintenance spare parts logistics: Special characteristics and
strategic choices”, International Journal of Production Economics,
71, pp. 125-133.
Jacquet-Lagreze, E. and Siskos, Y., 1982: “Assessing a set of additive utility
functions for multi-criteria decision making: the UTA method,
European Journal of Operations Research, Vol. 10, No.2,
pp. 151-164.
Kranenburg, A.A. andVan Houtum, G.J., 2007: “Effect of commonality on spare
parts provisioning costs for capital goods”, International Journal of
Production Economics, Vol. 108, No. 1-2, pp. 221-227.

68
Kurt, A., 2003: “Flexible manufacturing system design using simulation artificial
neural network”, J. Fac. Eng Arch. Gazi Univ., Vol. 18, No. 2,
pp. 31-38.
Leenders, M.R., Fearon, H.E. and England, W.B., 1989: Purchasing and
MaterialManagement, Irwin Inc.
Li, S.G. and Kuo, X., 2008: “The Inventory Management System for Automobile
Spare Parts in a Central Warehouse”, Expert Systems with
Applications, 34, pp. 1144-1153.
Lyly-Yrjanainen, J., Lahikeinen, T. and Paranko, J., 2004: “Cost effect of
component commonality: Manufacturing cost perspective.
NordDesing 2004, Product design in Changing Environment 18-20
August 2004, Tampere, Finland.
Masters, T., 1993: Practical Neural Network recipes in C++, Academic Press, Inc.
Malakooti, B. and Raman, V., 2000: "Clustering ans Selection of Multiple Criteria
Alternatives Using Unsupervised and Supervised Neural Networks",
Journal of Intelligent Manufacturing, 11, pp. 435-451.
McCulloch, W. and Pitts, W., 1943: "A logical calculus of ideas immanent in
neural activity", Bulletin of Mathematical Biophysics, 5, pp. 115-133.
Ng, W.L., 2007: "A Simple Classifier for Multiple Criteria ABC Analysis",
European Journal of Operational research, 177, pp. 344-353.
Nieddu, L. and Patrizi, G., 2000: "Formalmethods in patternrecognition:A
review", European Journal of Operational Research, 120,
pp. 459-495.
Paliwal, M., and Kumar, U.A., 2009: "Neural networks and statistical techniques:
A review of applications", Expert Systems with Applications, 36,
pp. 2–17.
Partovi, F.Y. and Anandarajan, M., 2002: "Classifying inventory using an
artificialneural network appoach”, Computer and Industrial
Engineering, 41, pp. 389-404.
Partovi, F. Y., and Burton, J., 1993: "Using the analytic hierarchy process for ABC
analysis.", International Journal of Operations and Production
Management, Vol. 13, No. 9, pp. 29–44.
Petrovic, D. and Petrovic R. 1992: "Sparta ii: Further development in an expert
system for advising on stocks of spare parts", International Journal of
of Production Economics, Vol. 24, No. 3, pp. 291-300.
Principle, J.C., Euliano, N.R. and Lefebvre, W.C., 1999: Neural and Adaptive
Systems. Fundamentals Through Simulations, John Wiley&Sons,
New York, USA.
Qi, M. and Zhang, G.P., 2001: "An investigation of model selection criteria for
neural network time series forecasting", European Journal of
Operational Research, Vol. 132, No. 3, pp. 666-680.
Ramanathan, R., 2006: “ABC inventory classification with multiple-criteria using
weighted linear optimization”, Computers & Operations Research, 33,
pp. 695–700.

69
Reid, R. A., 1987: "The ABC method in hospital inventory management: A practical
approach", Production and Inventory Management Journal,
Vol. 28, No. 4, pp. 67–70.
Roy, B., 1996: Multi Criteria Methodology for decision aiding. Dordrecht: Kluwer.
Saaty, T.L., 1980: “The Analytic Hierarchy Process, New York, NY:McGraw-Hill.
Siskos, J., Lochard, J. and Lombardo, J., 1984: “A multi criteria decision-making
methodology under fuzziness: Application to the evaluation of
radiological protection in nuclear power plants”, North-Holland,
Amsterdam, pp. 261-283.
Schalkoff, R., 1997: Artificial neural networks, McGraw-Hill International series,
New York.
Sharaf, M.A. and Helmy, H.A., 2001: “A classification model for inventory
management of spare parts”, Proceedings of 7th International
Conference on Production, Industrial Engineering, Design and
Control, Alexandria, Egypt, 13-15 February, 1, pp. 375-382.
Shen, L., Tay, F.E.H, Qu, L. and Shen, Y., 2000: “Fault diagnosis using rough
sets theory”, Computers in Industry, 43, pp. 61-72.
Sherbrooke, C.C., 1968: “METRIC: a multi-echelon technique for recoverableitem
control”, Operation Research, 16, pp. 122-141.
Simunovic, K., Draganjac, T. and Simunovic, G., 2008: “Application of Different
Quantitative Techniques to Inventory Classification”, Technical
Gazatte, Vol. 15, No. 4, pp. 41-47.
Slowinski, R., 1995: “Rough set approach to decision analysis”, AI Expert
Magazine, Vol. 10, No. 3, pp. 18-25.
Sönmeztürk, G.I., Şafak, M. E., Celebi, D. and Bayraktar, D., 2008: “Bulanık
AHY ile Yedek Parça Stoklarının Sınıflandırılması”, Yöneylem
Araştırması ve Endüstri Mühendisliği 29.Ulusal Kongresi.
Şahin, Ş.Ö., 2002: Yapay sinir aglari yardimi ile dinamik bir senaryo analizi,
Master Thesis, ITU, Institute of Science and Technology, Istanbul.
Tosun, S., 2007: Classification with artificial neural networks and data envelopment
analysis: application with student performance. Master Thesis, ITU
Institute of Science and Technology, Istanbul.
Turer, S., Ayvaz, B., Bayraktar, D. and Bolat, B., 2009: “Tedarikçi Değerlendirme
Süreci için bir Yapay Sinir Ağı Yaklaşımı: Gıda Sektöründe bir
uygulama”,Endütsri Mühendisliği Dergisi, Vol. 20, No. 2, pp. 31-40.
Tsumoto, S., 1998: “Automated extraction of medical expert system rules from
clinical databases based on rough set theory”, Information Sciences,
112, pp. 67-84.
Vapnik, V. N., 1995: The nature of statistical learning theory. New York:
Springer-Verlag.
Viale, J. D., 1996: Inventory Management : From Warehouse to Distribution
Center.Menlo Park, CA, USA: Course Technology Crisp.

70
Wei, S., Zhang, J. and Li, Z., 1997: “A Supplier-Selecting System Using a Neural
Network”, IEEE International Conference on Intelligent Processing
Systems, pp. 468-471.
Yu, M.C., 2011: “Multi-criteria ABC analysis using artificial-intelligence based
classification techniques”, Expert Systems with Application, 38,
pp. 3416-3421.
ZhangG.P., Patuwo, B.E. and Hu, M.Y., 1998: "Forecasting with artificial
neural network networks: The state of the art", International Journal
of Forecasting, 14, pp. 35-62.
Zhang, G.P., 2003: "Time series forecasting using a hybrid ARIMA and neural
network model", Neurocomputing, 50, pp. 159–175.
Zhang, Y. and Wan, X., 2007: "Statistical fuzzy interval neural networks for
currency exchange rate time series prediction", Applied Soft
Computing, 7, pp. 1149-1156.
Zhou, P. and Fan, L., 2007: “A note on multi-criteria ABC inventory classification
using weighted linear optimization”, European Journal of
Operational Research, 182, pp. 1488–1491.
Zopounidis, C., 1998: Operational Tools in the Management of Financial Risks,
KluwerAcademicPublishers, Dordrecht.

71
72
APPENDICES

APPENDIX A.1 : Evaluation data for Spare parts in Excel file

APPENDIX A.2 : Randomized training data divided into 5 disjoint subsets
APPENDIX B1 : Results of five fold cross validation method
APPENDIX B.2: Formulas for calculation of network performance measurements
APPENDIX B.3: Network building with Neurosolutions software.

73
74
APPENDIX A.1 Evaluation data for Spare parts in Excel file

Table A.1: Evaluation data for spare parts in excel file.

ID Stock- Substitutability Value Lead Unit Cost Commonality Class
no. out usage Times (TL)
Penalty
1 5 5 10 3 22 12 B
2 5 5 230 1 4,92 10 B
3 5 5 6 1 64,53 11 B
4 1 3 1000 3 2,99 16 C
5 5 5 170 3 8,59 4 A
6 2 5 400 1 0,213 20 C
7 2 5 850 1 0,465 21 C
8 5 1 5 28 610,131 2 A
9 3 5 500 1 0,38 22 C
10 2 3 33 18 321,7395 18 C
11 2 5 775 1 0,582 23 C
12 2 3 150 4 1,43568 5 B
13 1 5 1 2 65 19 C
14 1 5 25 3 34,5 14 B
15 5 5 2 5 850 3 A
16 2 5 100 3 10,5 7 B
17 3 1 1 12 101,8708 17 C
18 3 3 10 8 87,02062 6 B
19 3 5 250 4 17 13 B
20 5 5 13 3 24 8 B
21 3 5 22 4 110 15 B
22 5 1 9 24 4174,1376 1 A
23 1 5 80 1 0,75 24 C
24 5 3 37 5 0,3 9 B
25 5 5 235 1,5 28,5 1 A
26 5 1 1 16 1257,783 1 A
27 5 1 3 14 7386,3504 2 A
28 3 5 91 3 57 11 C
29 5 1 4 16 11330,6156 3 A
30 5 1 3 16 23500,5892 4 A
31 5 5 3 12 803,112 5 A
32 5 3 9 18 127,935 6 A
33 5 1 3 12 14660 7 A
34 2 5 3 2 43,488 12 C
35 5 1 1 12 15075,4436 8 A
36 5 1 1 12 836,4884 9 A
37 1 5 14 2 50,92 13 C

75
Table A.1: Evaluation data for spare parts in excel file (contd.)
Stock-
ID out Substitutabil Value Lead Unit Cost Commona Clas
no. Penalty ity usage Times (TL) lity s
38 5 3 2 8 766,774 10 B
39 3 1 54 12 167,22 2 B
40 1 5 211 2 1,1 3 C
41 4 3 5366 10 7,837 1 A
42 5 1 2 30 24255,294 1 A
43 1 5 40 1 7,25 2 C
44 1 5 2100 1 0,97 3 C
45 5 1 1 12 7797,4206 1 A
46 3 1 15 14 2529,271 2 B
47 5 1 20 14 403,705 2 A
48 5 1 20 14 350,406 3 A
49 5 1 20 14 447,741 1 A
50 1 5 4 2 25,5 13 C
51 5 2 140 3,5 15,2 3 A
52 3 3 32 5 4,237 10 C
53 5 3 2 5 2963,1 5 B
54 5 3 3 5 71,972416 1 A
55 5 5 16 3,5 293 2 A
56 5 5 14 4 32,217138 6 B
57 5 3 14 2 196,36 4 B
58 2 5 5 1 5,85 11 C
59 1 3 3 4 175,02 9 C
60 3 1 6 5 418,3 7 B
61 3 3 2 5 1484,71 8 B
62 1 5 3 1 1,22 14 C
63 1 5 20 1 5,6 12 C
64 5 2 157 14 60,211025 1 A
65 1 3 11 10 138,25 2 B
66 1 5 40 1 4,352 7 C
67 3 3 1 8 1337,7 5 B
68 3 5 9 10 350 4 B
69 1 1 337 30 325,325 3 B
70 1 5 217 4 5,5 6 C
71 1 5 1 1,5 105 8 C

76
APPENDIX A.2Randomized training data divided into 5 disjoint subsets.

Table A.2: Randomized training data divided into 5 disjoint subsets.

Stock-
ID out Value Lead
Set no. Penalty Substitutability usage Times Unit Cost Commonality A B C
47 5 1 20 14 403,705 2 1 0 0
44 1 5 2100 1 0,97 3 0 0 1
65 1 3 11 10 138,25 2 0 1 0
70 1 5 217 4 5,5 6 0 0 1
66 1 5 40 1 4,352 7 0 0 1
18 3 3 10 8 87,02062 6 0 1 0
1
43 1 5 40 1 7,25 2 0 0 1
48 5 1 20 14 350,406 3 1 0 0
38 5 3 2 8 766,774 10 0 1 0
20 5 5 13 3 24 8 0 1 0
26 5 1 1 16 1257,783 1 1 0 0
68 3 5 9 10 350 4 0 1 0
32 5 3 9 18 127,935 6 1 0 0
4 1 3 1000 3 2,99 16 0 0 1
10 2 3 33 18 321,7395 18 0 0 1
55 5 5 16 3,5 293 2 1 0 0
17 3 1 1 12 101,8708 17 0 0 1
6 2 5 400 1 0,213 20 0 0 1
2
59 1 3 3 4 175,02 9 0 0 1
15 5 5 2 5 850 3 1 0 0
9 3 5 500 1 0,38 22 0 0 1
60 3 1 6 5 418,3 7 0 1 0
36 5 1 1 12 836,4884 9 1 0 0
64 5 2 157 14 60,211025 1 1 0 0
42 5 1 2 30 24255,294 1 1 0 0
39 3 1 54 12 167,22 2 0 1 0
33 5 1 3 12 14660 7 1 0 0
13 1 5 1 2 65 19 0 0 1
28 3 5 91 3 57 11 0 0 1
50 1 5 4 2 25,5 13 0 0 1
3
2 5 5 230 1 4,92 10 0 1 0
5 5 5 170 3 8,59 4 1 0 0
29 5 1 4 16 11330,616 3 1 0 0
25 5 5 235 1,5 28,5 1 1 0 0
34 2 5 3 2 43,488 12 0 0 1
11 2 5 775 1 0,582 23 0 0 1

77
Table A.2: Randomized training data divided into 5 disjoint subsets (contd.)
Stock-
ID out Value Lead
Set no. Penalty Substitutability usage Times Unit Cost Commonality A B C
54 5 3 3 5 71,972416 1 1 0 0
69 1 1 337 30 325,325 3 0 1 0
51 5 2 140 3,5 15,2 3 1 0 0
30 5 1 3 16 23500,589 4 1 0 0
46 3 1 15 14 2529,271 2 0 1 0
67 3 3 1 8 1337,7 5 0 1 0
4
21 3 5 22 4 110 15 0 1 0
23 1 5 80 1 0,75 24 0 0 1
16 2 5 100 3 10,5 7 0 1 0
14 1 5 25 3 34,5 14 0 1 0
8 5 1 5 28 610,131 2 1 0 0
12 2 3 150 4 1,43568 5 0 1 0
41 4 3 5366 10 7,837 1 1 0 0
57 5 3 14 2 196,36 4 0 1 0
45 5 1 1 12 7797,4206 1 1 0 0
52 3 3 32 5 4,237 10 0 0 1
63 1 5 20 1 5,6 12 0 0 1
27 5 1 3 14 7386,3504 2 1 0 0
5
37 1 5 14 2 50,92 13 0 0 1
49 5 1 20 14 447,741 1 1 0 0
35 5 1 1 12 15075,444 8 1 0 0
24 5 3 37 5 0,3 9 0 1 0
31 5 5 3 12 803,112 5 1 0 0
19 3 5 250 4 17 13 0 1 0

78
APPENDIX B.1 Results of five fold cross validation method

Table B.1: Minimum mse for training data for five subsets with different hidden
node numbers.

No. Node Set No. Min MSE

1 0,08815
2 0,06739
2 3 0,13478
4 0,0445
5 0,09115
1 0,117
2 0,10056
3 3 0,13108
4 0,04473
5 0,12695
1 0,09932
2 0,11548
4 3 0,1527
4 0,04596
5 0,0788
1 0,10625
2 0,14047
5 3 0,14856
4 0,05011
5 0,0736
1 0,14437
2 0,1707
6 3 0,12214
4 0,03823
5 0,06723
1 0,11787
2 0,19249
7 3 0,1443
4 0,07113
5 0,09774

79
Table B.1: Minimum mse for training data for five subsets with different hidden
node numbers (contd.).
No. Node Set No. Min MSE
1 0,07779
2 0,19322
8 3 0,09057
4 0,10379
5 0,12241
1 0,1219
2 0,14422
9 3 0,16835
4 0,07616
5 0,08637
1 0,12905
2 0,14588
10 3 0,14726
4 0,02551
5 0,06343
1 0,15411
2 0,16615
11 3 0,16694
4 0,02693
5 0,09915
1 0,11062
2 0,17151
12 3 0,12117
4 0,11553
5 0,08794
1 0,16306
2 0,20393
13 3 0,14529
4 0,13174
5 0,11075
1 0,15419
2 0,17794
14 3 0,14864
4 0,05335
5 0,11996

80
Table B.1: Minimum mse for training data for five subsets with different hidden
node numbers (contd.).

No. Node Set No. Min MSE

1 0,1487
2 0,19374
15 3 0,16906
4 0,10186
5 0,10925
1 0,13792
2 0,16596
16 3 0,13202
4 0,12651
5 0,09286
1 0,11358
2 0,14936
17 3 0,17715
4 0,1146
5 0,12627
1 0,15
2 0,22412
18 3 0,2271
4 0,10334
5 0,11343
1 0,15252
2 0,15401
19 3 0,201
4 0,12044
5 0,06489
1 0,14244
2 0,17321
20 3 0,12773
4 0,08764
5 0,11971

81
Table B.2: Minimum mse for cross validation data for five subsets with different
hidden node number.

No. Node Set No. Min MSE

1 0,005085
2 0,004325
2 3 0,001183
4 0,140905
5 0,148789
1 0,000588
2 0,090422
3 3 0,001209
4 0,154522
5 0,154322
1 0,001157
2 0,001736
4 3 0,001411
4 0,181418
5 0,131251
1 0,003889
2 0,076077
5 3 0,00256
4 0,112092
5 0,099379
1 0,005218
2 0,012562
6 3 0,002006
4 0,097729
5 0,139008
1 0,003707
2 0,090419
7 3 0,005994
4 0,094414
5 0,195539

82
Table B.2: Minimum mse for cross validation data for five subsets with different
hidden node number (contd.)

No. Node Set No. Min MSE

1 0,003047
2 0,004106
8 3 0,002985
4 0,094213
5 0,082456
1 0,004907
2 0,004451
9 3 0,003486
4 0,105173
5 0,234503
1 0,002915
2 0,054784
10 3 0,003956
4 0,104444
5 0,085984
1 0,002936
2 0,004287
11 3 0,039104
4 0,097593
5 0,192136
1 0,003986
2 0,004837
12 3 0,004139
4 0,087576
5 0,170837
1 0,079069
2 0,054462
13 3 0,003323
4 0,127702
5 0,23195
1 0,003745
2 0,07637
14 3 0,023162
4 0,099809
5 0,236425

83
Table B.2: Minimum mse for cross validation data for five subsets with different
hidden node number (contd.)
No. Node Set No. Min MSE

2 0,054655
3 0,054023
4 0,193675
5 0,156212
1 0,003561
2 0,077017
16 3 0,063359
4 0,216191
5 0,187609
1 0,004463
2 0,093997
17 3 0,004591
4 0,200056
5 0,154228
1 0,054184
2 0,00446
18 3 0,054077
4 0,094209
5 0,208318
1 0,020275
2 0,084552
19 3 0,004016
4 0,154283
5 0,136505
1 0,004325
2 0,054723
20 3 0,004049
4 0,189607
5 0,189766

84
Table B.3: Calculation of average minimum mean squared error for two node of
cross validation data.

NO. SET MIN

NODE NO. MSE
1 0,00508
2 0,00432
2 3 0,00118
4 0,14091
5 0,14879

∑Y>AB TUVW>X
Z
Average minimum MSE for node 2= i=1,2,3,4,5:set number

Average minimum MSE for node 2=

=0,060057349
[,[[Z[\0[,[[70[,[[

\0[,
7[]
0[,
7\^]
Z

Calculation of average minimum MSE for other nodes is done in similar way and the
results are given in Table 5.5 and Teble 5.6 for tangent and sigmoid functions
respectively.

APPENDIX B.2 Formulas for calculation of network performance

measurements.

MSE: The mean squared error is simply two times the average cost (see the access
points of the ErrorCriterion component.) The formula for the mean squared error is:

∑e [ ∑d

 [b −  c
_`a =
f g
Where P=number of output processing elements

N=number of exemplars in the data set

 =network output for exemplar I at processing element j

 =desired output for exemplar I at processing element j

NMSE: The normalized mean squared error is defined by the following formula:

85
 g f _`a
f_`a = C C
∑e [ >Aj >?
d ∑i h 4b∑i
>Aj h>? c
d

Where P=number of output processing elements

N=number of exemplars in the data set

MSE=mean squared error

 =desired output for exemplar I at processing element j

r: the correlation coefficient between a network output x and and a desired output d
is:

∑> F> 4F̅  h> 4hm

k= d

n n
∑> h> 4hmC ∑> F> 4F̅ C
d d

The correlation coefficient is confined to the range [-1,1]. When r =1 there is a

perfect positive linear correlation between x and d, that is, they covary, which means
that they vary by the same amount. When r=-1, there is a perfectly linear negative
correlation between x and d, that is, they vary in opposite ways (when x increases, d
decreases by the same amount). When r =0 there is no correlation between x and d,
i.e. the variables are called uncorrelated. Intermediate values describe partial
correlations. For example a correlation coefficient of 0.88 means that the fit of the
model to the data is reasonably good.

% Error: The percent error is defined by the following formula:

e d
100 r −  r
%akkpk = qq
f g 
 [  [

Where P=number of output processing elements

N=number of exemplars in the data set
 =denormalized network output for exemplar i at processing element j
 = denormalized desired output for exemplar i at processing element j

86
APPENDIX B.3 Network building with Neurosolutions software.

The selection of training, cross validation, testing, input and desired data on Excel
worksheet can seen in the Figure B.1.

Figure B.1: Selection of input and desired parameters in Excel.

After selecting 6 input and 3 output criteria, by similar way 48 data is chosen as
training, 12 of data is selected as cross validation and 11 data for spare parts is
chosen as testing data for the network model. Then Neurobuilder tool from
Neurosolutions menu was chosen to design the network. The starting menu of
Neurobuilder can be seen in Figure B.2.

87
 Figure B.2: NeuroBuilder tool.

In the starting menu the learning model for neural network should be chosen. After

selecting the learning model of neural network sign is ticked and in the
next section number of hidden layer is selected as can be seen in Figure B3:

Figure B.3: Selection of Hidden Layer number in Neurosolutions.

88
After selecting the hidden layer number, the number of hidden neurons, activation
function, step size and momentum rate in the hidden layer is selected as can be seen
in Figure B.4:

Figure B.4: Selection of training parameters for Hidden Layer in Neurosolutions.

In the similar way, the training parameters like activation function, step size and
momentum are selected for the output layer as can be viewed in Figure B.5 below.

89
 Figure B.5: Selection of training parameters for Output Layer in Neurosolutions.

In the least step of network building in Neurosolutions,in Figure B.6, the maximum
number of epochs, stopping performance criteria and weight update methods are
selected.

Figure B.6: Selection of performace criteria in Neurosolutions.

90
After choosing all required parameters with the network structure, probe
configuration, in Figure B.7, is specified is order to view required solutions of the
training.

Figure B.7: Setup for Probe Configuration in Neurosolutions.

Finally, after finishing selection of network structure and learning parameters,

“Build” key is clicked and network is built as can seen in Figure B.8.

91
 Figure B.8: NeuroSolutions Network Structure.

Training of the network is started from NeuroSolutions tool in Excel, the number of
max epochs and runs is specified in the opened icon, which is displayed below in
Figure B.9.

Figure B.9: Training Process in Neurosolutions.

92
CURRICULUM VITAE

Candidate’s full name: JarkynJumabaeva

Place and date of birth: Bishkek-Kyrgyzstan 18.08.1985

Permanent Address: Ogombaeva Street, 179. Bishkek, Kyrgyzstan

Universities and
Colleges attended: Middle East Technical University

93