See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/227420674

Predicting the development cost of TFT-LCD

manufacturing equipment with artificial
intelligence models
Article in International Journal of Production Economics November 2010
DOI: 10.1016/j.ijpe.2010.07.031 Source: RePEc

CITATIONS

READS

14

231

3 authors, including:
Jui-Sheng Chou

Yian Tai

National Taiwan University of Science and

National Taiwan University of Science and

104 PUBLICATIONS 881 CITATIONS

61 PUBLICATIONS 776 CITATIONS

SEE PROFILE

All in-text references underlined in blue are linked to publications on ResearchGate,

letting you access and read them immediately.

SEE PROFILE

Available from: Jui-Sheng Chou

Retrieved on: 20 October 2016

This article appeared in a journal published by Elsevier. The attached

copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elseviers archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright

Author's personal copy

Int. J. Production Economics 128 (2010) 339350

Contents lists available at ScienceDirect

Int. J. Production Economics

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

Predicting the development cost of TFT-LCD manufacturing equipment with

articial intelligence models
Jui-Sheng Chou a,n, Yian Tai a, Lian-Ji Chang b
a
b

National Taiwan University of Science and Technology, 43 Section 4, Keelung Road, Taipei 106, Taiwan
Contrel Technology Corporation, Taiwan

a r t i c l e in fo

abstract

Article history:
Received 17 March 2009
Accepted 14 July 2010
Available online 23 July 2010

Accurately and timely estimating product costs is extremely benecial to corporate survival. This study
assesses the reliability of multiple regression analysis (MRA), articial neural networks (ANNs), casebased reasoning (CBR), and hybrid intelligence (Hi) to forecast costs of thin-lm transistor liquid-crystal
display (TFT-LCD) equipment. Newly completed equipment-development projects are provided by
departments in a Taiwanese high-tech company, which is a top global producer of TFT-LCD equipment.
The cross-fold validation method is applied to measure model performance, reliability, and prediction
ease. Through comparison of various performance indices, the Hi method outperforms MRA, ANNs and
CBR when used for cost prediction during conceptual stages. Although it is well developed in academia,
articial intelligence (AI) is rarely applied in practical project management. This study successfully
describes an actionable knowledge-discovery process using a real-world data mining approach for the
high-tech equipment manufacturing industry. Project managers (PMs) can benet from applying the Hi
approach to establish latent non-linear cost estimation relationships. The Hi approach is empirically
proven an effective prediction technique for PMs considering overall evaluation criteria when
determining the best selling prices of TFT-LCD manufacturing equipment to clients.
& 2010 Elsevier B.V. All rights reserved.

Keywords:
TFT-LCD industry
Manufacturing
Project management
Forecasting
Articial intelligence

1. Introduction
The rst impression of a display device is based on its design,
quality and presentation in an advertisement. The span of devices
ranges from simple displays in watches to complex technologies
in computers. More than 500 million displays for televisions and
computers were produced in 2008. Roughly 90% of these displays
were cathode ray tubes (CRTs) in 1999. However, more than 60%
of displays are now at panels (e.g., liquid crystal displays (LCDs),
plasma screens). The global market for panels, which was worth
over US $70 billion in 2006, is predicted to reach US $92 billion in
2011 (ITIS, 2008). Therefore, thin-lm transistor-LCDs (TFT-LCDs)
now play a leading role in various at-panel display devices
(Menozzi et al., 2001), as TFT-LCDs have excellent features such as
low proles, low operating voltage, low power consumption, fullcolor capabilities, large visible area, and high-quality resolution.
Prices of TFT-LCDs are strongly inuenced by the high degree
of uncertainty in research and development (R&D). In practice, a
at panels functions and specications cannot be completely
determined during the early R&D stages; thus, initial cost
estimation (ICE) is generally based on the subjective judgment
of experienced engineers. This ICE is only a quoted price for

Corresponding author. Tel.: +886 2 2737 6321; fax: + 886 2 2737 6606.
E-mail address: jschou@mail.ntust.edu.tw (J.-S. Chou).

0925-5273/$ - see front matter & 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.ijpe.2010.07.031

potential buyers and a comparative estimate for potential

providers. Although knowledgeable sales managers or estimators
may generate a good cost assessment via teamwork, professionals
are difcult to train and are very mobile in the small- and
medium-sized enterprises in Taiwan. Thus, this experience and
project knowledge is hard to retain, resulting in such problems as
loss of company credibility and potential customers.
The TFT-LCD industry, promoted by Taiwans government in
the Two Trillion Twin Stars plan, has developed rapidly in recent
years. The TFT-LCD industry is second only to the semiconductor
industry in driving Taiwans economic growth. Notably, LCD
production involves hundreds of complex processes. Each panel
manufacturer typically has unique patented processes and
production lines requiring diverse manufacturing equipment with
a high degree of customization. The production of diverse panel
sizes is generally in demand as the need for next-generation
panels increases.
The technological environment has become increasingly
competitive due to the rapid speed of globalization. Product cost
estimation is a critical and important task for sales managers at
high-tech equipment development companies to gain competitive
advantage. As the time a panel manufacturer devotes to investment and production is critical, fast and accurate estimations
are essential to successful delivery of devices in a timely manner,
as is quality to customers in the increasingly competitive market
environment.

Author's personal copy

340

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

Despite signicant advances in tools and techniques that assist

in project management (PMI, 2004), project managers (PMs) in
the engineer-to-order manufacturing industry still face problems
associated with guesswork and subjective judgment that result
frequently in inaccurate estimates. No concrete functional
relationship exists between nal cost and the basic cost drivers
for equipment for manufacturing TFT-LCD panels. Thus, companies can lose market share and orders during the early marketing
phase of attracting customers.
Although many studies of the TFT-LCD industry have analyzed
raw materials, processes, supply chains, and marketing, TFT-LCD
equipment cost forecasting has seldom been examined. Hence,
this study reviews relevant literature and collected past projects
with equipment specications. Additionally, this study presents
novel articial intelligence (AI)-based prediction models for the
TFT-LCD manufacturing equipment industry. The optimal estimation model must transform implicit corporation assets into
explicit knowledge, provide a basis for bargaining between a
manufacturer and buyer, and, consequently, avoid delays to
production plans by reducing the time required by tedious price
negotiation processes.
Data mining (DM) is a business intelligence tool used for
exploring and analyzing data to identify meaningful patterns and
rules. Several widely used DM techniques, namely, multiple
regression analysis (MRA), articial neural networks (ANNs),
and case-based reasoning (CBR), have been applied to diverse
cost management domains such as textiles and apparel (Camargo
et al., 2003; Metaxiotis, 2004), machinery centers (Ciurana et al.,
2008; Gang, 2005), software engineering (Berlin et al., 2009;
Finnie et al., 1997; Marban et al., 2008; Mendes et al., 2002;
Thomasson et al., 2006; Toussaint and Cheng, 2006), process
vessels (Caputo and Pelagagge, 2008), building construction
(Al-Sakran, 2006; An et al., 2007a, 2007b; Emsley et al., 2002;
Kim et al., 2004; Marir and Watson, 1995; Nassif et al., 2007;
Staub-French et al., 2003), infrastructure projects (Chou, 2009a,
b), die manufacturing (Bouaziz et al., 2006), the automotive
industry (Cavalieri et al., 2004), and product design (Asiedu and
Gu, 1998; Bode, 2000; Seo et al., 2002; Tornberg et al., 2002).
However, these techniques have not been applied to cost
estimation of TFT-LCD equipment development. This study
assesses the ability of MRA, ANNs, CBR and proposed hybrid
intelligent techniques to estimate cost with merely adequate
knowledge of the TFT-LCD manufacturing process. By applying
the AI model in practice, PMs can estimate equipment

development costs early in the price quotation and negotiation

stage, and thereby gain competitive advantage.
The remainder of this paper is organized as follows. Literature
is reviewed in Section 2 to identify the techniques utilized and
challenges in project cost estimation in the engineering industry.
Sections 3 and 4 discuss the TFT-LCD manufacturing process and
employ the research methodology, respectively. Section 5 examines the case of a Taiwanese high-tech company and presents
analytical and model validation results. Section 6 gives
conclusions, and presents research ndings and managerial
implications.

2. Literature review of cost estimation

Cost estimation methods can generally be categorized as
analogous cost estimation (ACE), bottom-up estimation techniques, and computing technology combined with AI (Ben-Arieh
and Qian, 2003; Bode, 2000; Camargo et al., 2003; Caputo and
Pelagagge, 2008; Cavalieri et al., 2004; Duverlie and Castelain,
1999; Kwak and Watson, 2005; Marban et al., 2008; Metaxiotis,
2004; PMI, 2004; Tornberg et al., 2002). Table 1 lists the tools and
techniques used by these methods in project phases.
2.1. Analogical and analytical techniques
Analogous cost estimating, also referred to as top-down
estimating, uses values of such parameters as project length, size,
duration, cost, and design complexity from a similar project to
estimate the cost of a new project. This technique, which is
commonly used to estimate cost early in a projects lifecycle when
little information is available, requires sufcient knowledge and
expert judgment (PMI, 2004). Notably, ACE is generally less costly
and time consuming than other methods, but is also less accurate.
Although project activities in two projects may be similar,
differences in project background and environment, such as
technology usage (Rogozhin et al., 2010), number of patents used
in product development (Kee, 2010), clients (Hwang et al., 2008),
new materials or equipment (Lin et al., 2009), can result in
signicantly different estimates.
The analytical technique of bottom-up estimation determines
project cost by decomposing a process into manageable tasks,
operations, or activities (PMI, 2004). Activity-based costing (ABC),
a generic technique, calculates costs incurred as performing

Table 1
Cost estimation techniques at various product development phases.
Estimating techniques

1. Analogous cost estimating

Engineering experience & expert judgment (PMI, 2004)
2. Bottom-up estimates
Activity-based costing; ABC
(Baykasoglu and Kaplanoglu, 2008; Ben-Arieh and Qian, 2003; Tornberg et al., 2002)
3. Articial intelligence
Articial neural networks
(Bode, 2000; Caputo and Pelagagge, 2008; Cavalieri et al., 2004)
Case-based reasoning (Al-Sakran, 2006; An et al., 2007a; Duverlie and Castelain, 1999;
Mendes et al., 2005; Toussaint and Cheng, 2006)
4. Multiple regression and parametric techniques
Multiple regression and parametric cost estimating (Camargo et al., 2003; Caputo and
Pelagagge, 2008; Cavalieri et al., 2004; Chou, 2009a; Duverlie and Castelain, 1999; Kwak
and Watson, 2005; Marban et al., 2008)
Feature-based estimation (Bouaziz et al., 2006; Ou-Yang and Lin, 1997)

Product life cycle

Initiation stage

Planning phase

Detailed design

Contracting
& mass production

Author's personal copy

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

activities when manufacturing a product (Niazi et al., 2006).

Ben-Arieh and Qian (2003) presented a methodology that uses
ABC to determine the costs of design and development activities
in manufacturing machined parts in a controlled facility (BenArieh and Qian, 2003). Tornberg et al. (2002) investigated the use
of ABC and modeling design, purchasing and manufacturing
processes to generate useful cost information for product
designers (Tornberg et al., 2002). An application of ABC to a land
transportation company demonstrated the effectiveness of ABC
combined with business process modeling and an analytical
hierarchy approach (Baykasoglu and Kaplanoglu, 2008).

2.2. Computing technology and articial intelligence in cost

estimation
The MRA estimates product cost based on product features,
functions, and characteristics. The estimation technique uses
validated relationships between a projects known technical and
cost characteristics given a subset of factual data. In some cases,
one must obtain the detailed designs for equipment when using
the FBM.
Smith and Mason (1997) noted that professional cost estimators in diverse domains regularly utilized regression to construct
cost models (Smith and Mason, 1997). Many studies have
examined the relationships between cost and project characteristics (Berlin et al., 2009; Bouaziz et al., 2006; Camargo et al.,
2003; Caputo and Pelagagge, 2008; Cavalieri et al., 2004; Chou,
2009a; Duverlie and Castelain, 1999; Kwak and Watson, 2005;
Marban et al., 2008; Ou-Yang and Lin, 1997), and generated high
levels predictability depending on the quality of the underlying
data source and sophisticated statistical techniques employed to
build the model in diverse disciplines.
Although AI has an innovative role in the cost estimation
process, considerable effort is required to determine the estimation accuracy by validating computing outcomes. Notably, AI used
for cost estimation has been widely applied in manufacturing
sectors, computer and software engineering, the process industry,
concurrent engineering, the construction industry, R&D management, and product development (Bashir and Thomson, 2001;
Bode, 1998, 2000; Camargo et al., 2003; Caputo and Pelagagge,
2008; Cavalieri et al., 2004; Duverlie and Castelain, 1999; Marban
et al., 2008)
As a typical AI technique, ANNs are feasible alternatives for
early cost prediction as ANNs can model complex systems given a
minimal amount of data. Many studies have veried the
predictive power and excellent accuracy of ANNs in process
vessels (Caputo and Pelagagge, 2008), the automotive industry
(Cavalieri et al., 2004), sheet metal parts (Verlinden et al., 2008),
TFT-LCD yield (Hsieh and Lu, 2008), machine-tool selection
(Ciurana et al., 2008), software development (De Barcelos Tronto
et al., 2007), product lifecycle cost (Seo et al., 2002), vertical high
speed machining centers (Ciurana et al., 2008), and building
construction (Emsley et al., 2002).
Case-based reasoning, another common AI technique, uses the
nearest-neighbor algorithm, which is based on the rationale of
characterizing projects by a number of essential attributes. These
attributes are then weighted such that they match similar cases.
Case-based reasoning requires a series of systematic procedures
to extract relevant knowledge from experience, integrate a case
into an established knowledge structure, index the case for
subsequent matching, compare a new case with similar cases,
and save knowledge (Toussaint and Cheng, 2006). Numerous
studies have identied CBR as particularly appropriate for
application to the domain knowledge at conceptual stage

341

(Chantaraskul and Cuthbert, 2005; Chou, 2009b; Gang, 2005;

Thomasson et al., 2006).
2.3. Challenges in determining the costs of TFT-LCD manufacturing
equipment
Estimation accuracy is closely related to the amount and
quality of information available at estimation time. Thus,
generating accurate preliminary cost estimates is very challenging. (Berlin et al., 2009; Cavalieri et al., 2004; Chou, 2009a). Fig. 1
illustrates the price quotation and engineer-to-order process. The
primary challenge is that when a sales manager actively or
passively contacts clients make-to-order quotes, the estimated
costs are determined entirely by known product attributes,
whereas while the vast majority of product costs are implicitly
determined in the early stages of product development for almost
all manufactured goods.
The analogous technique is highly subjective and reliant on
personal judgment; thus, it can easily result in inaccurate
estimates during an analytical estimation process. For example,
ABC is extremely dependent on the availability of detailed design
information, which is useful in general, but not ideally suited to
early estimation in practice. The bottom-up estimation process is
relatively straightforward and can yield good results when
applied properly. However, without detailed product design
specications, this process leads to inaccurate estimates. In such
cases, manual estimates are unsatisfactory and time-consuming.
Additionally, high-tech companies typically encounter high
rates of personnel loss in sales departments and, thus, have
difculty keeping experience and implicit corporate knowledge.
Conversely, only incomplete and uncertain equipment design
specications are available during the early quotation and
negotiation processes. Rapidly responding to customer enquiries
with little available information is also one of the critical factors
to secure the orders.
Fairly accurate predictive tools for determining costs before an
actual device is manufactured are needed for engineered-to-order
suppliers to retain an advantageous position in the global market
(Tsai and Wang, 2009). Therefore, this study uses computing
technology and AI to estimate the cost of TFT-LCD manufacturing
equipment. An early accurate estimate during the negotiation
stage with clients can reduce unnecessary bargaining and cost
changes in the subsequent manufacturing process.
Further, this study examined the ability of a series of
congurations of MRA, ANNs, CBR, and hybrid models to estimate
cost during early quotation stages in a make-to-order TFT-LCD
equipment development industry to overcome challenges inherent in the process equipment industry. The implicit knowledge in
historical project data can be uncovered using AI techniques for
sales managers or decision-makers who may have little knowledge of the detailed components in TFT-LCD manufacturing
equipment.

3. TFT-LCD manufacturing process

A TFT-LCD device has three main layersa liquid-crystal layer
encapsulated (sandwiched) between two glass substrate layers,
namely, the TFT and color-lter plates (Clark, 2001). Fig. 2 shows a
typical TFT-LCD fabrication process scheme, which can be divided
into three stages: the array process, LCD process, and LCM process
(Clark and Robert, 2001). The array process stage has two
manufacturing processes, namely, fabrication of the TFT and
fabrication of the color lter (CF).
The CF is made with dyes. Several methods can be utilized to
deposit red, green, and blue (RGB) color dyes onto glass substrate,

Author's personal copy

342

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

1. Routine visits
2. Observe the dynamic global
market
3. Acquire conceptual product
attributes

Marketing department
Seek & identify for potential
customers & market places

Lose order

Visit customers &

understand their needs
1. Conventional manual estimation
2. Retrieve past similar projects or
start guess work
3. Inputs from other departments
(Design, R&D, Accounting Office)

Price quotation and

negotiation process

Offer accepted

After-sale
service

AI models to facilitate the

estimating process

Procurement/legal affairs/design/
production departments
Make to order by conducting
detailed design

Product delivery

Deliver equipment with quality, as

scheduled, and within budgetary
estimates

Project close-out

Fig. 1. The price quotation and make-to-order processes.

Array Process

In

Process

LCM & Assemble

Process

Cell Process

Inspection

TFT Array
Mount

LC fill

Seal

Panel
assembly

Module
assembly

Final
Inspection

CF Array
In

Process

Inspection

Fig. 2. Flowchart of the TFT-LCD fabrication process.

including diffusion, electro-deposition, and printing. Among these

many combinations of CF fabrication methods and congurations,
the color-resist method with a stripe-type RGB arrangement is the
most popular (Lee et al., 2008; Sabnis, 1999). Between the blocks
of color in the CF is a black matrix made of a dense metal, such as
chromium, to prevent a-Si TFTs from stray light, and avoid light
leakage between pixels.
The LCD panel applications vary depending on what TFT-LCD
products are being produced (e.g., monitors, TVs, portable PCs,

e-books, and communication products) (Fig. 3). During the LCD

manufacturing process, the TFT array and CF substrates are then
made into an LCD panel by joining the two substrates with
a sealant; the cell gap is maintained by spacers. The two
substrates are laminated together and then the panel is lled
with liquid crystals. In the nal manufacturing stage, driver
circuits, the back-light unit and other components are
integrated into the panel to complete the LCD module (LCM)
(Smith, 2005).

Author's personal copy

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

343

Fig. 3. Categories of TFT-LCD panel applications.

4. Estimating models applied to TFT-LCD manufacturing

equipment
Comparative analyses of prediction models use two common
DM algorithms, ANNs and the nearest-neighbor algorithm, along
with a most commonly used multiple regression technique. These
two algorithms are implemented using NeuralTools (Palisade,
2008) and Excel macros, and the multiple regression technique is
applied using SPSS Statistics.

4.1. Multiple regression model

Multiple regression model provides data analysts and decision-makers with a rapid technique for examining functional
relationship between input and output variables based on
historical cases. Stepwise MRA was applied in this study to
identify the signicant factors and cost prediction model, as well
as a compare basis with other AI-based techniques.
As the functional relationship between response and predictor
variables is imposed, the general form of the multiple regression
formulation can be expressed as
YMRA b0

N
X

bi xi ep

i1

where YMRA is the response variable for equipment costs using

stepwise MRA; xi are predictorseither numerical or dummy
variables; bi are coefcients related to the problem; N is the
number of covariate variables (predictors); and b0 is the intercept.
The generalized linear model reduces to the familiar multiple
linear regression model plus the random component, ep, which is
believed to have a normal distribution with mean zero and
variance of N(0,s2). With collected data, including explanatory
parameters and associated known values of the response variable,
the multivariate regression model can determine which bi
coefcients best t historical data and explain the greatest
amount of variance.
To identify the variables that should be included in the
regression estimation model, stepwise variable selection is
adopted to construct the model with the optimal set of project
attributes. The stepwise procedure inputs variables into the
discriminant equation by successively adding or deleting variables according to a pre-determined criterion (F- or p-value). The

iterative process terminates when no variables meet the criterion

(Albright et al., 2006). The aim of this procedure is to generate the
least standard error and largest goodness-of-t (R2) of the
estimation model. Additionally, interaction terms were added to
the regression models to determine whether multicollinearities
exist between predictor variables and interactions terms.
4.2. Articial neural networks model
A neural network is a system that mimics the human brain
when transforming inputs into one or more outputs with a set of
neurons. Multi-layer feed-forward nets (MLFN) and generalized
regression neural networks (GRNN) were utilized in this study.
Specically, a hyperbolic tangent function is used as the
activation function in hidden layer neurons for MLFN. The output
neuron uses identity as the activation function that it simply
returns the weighted sum of its inputs (Palisade, 2008). The GRNN
is originally proposed by Specht (1991) and approximates any
arbitrary function between input and output vectors with
drawing the function estimate directly from the training data
(Cigizoglu, 2005; Specht, 1991). Additionally, the GRNN is related
to the radial basis function network and is based on a standard
statistical technique, namely kernel regression (Specht, 1991).
Fig. 4 shows the adopted MLFN topology, which consists of
three layersinput, hidden, and output layer, and GRNN
topology, which consists of four layersinputs, the pattern
layer (one neuron per training case), summation layer
(numerator and denominator nodes), and output layer. Since a
single hidden layer is sufcient for most problems (Hegazy et al.,
1994; Palisade, 2008), the MLFN topology was congured as a
single hidden layer starting with 0.75m to 2m+ 1 neurons to
generate the lowest error and prevent over-training, where m is
the number of processing elements in the input layer (Hegazy
et al., 1994). Moreover, a set of features selected in advance via
MRA was used as input variables in GRNN models to evaluate the
impact of dimension reductions.
Several heuristic techniques have been developed to accelerate
computational convergence. To reduce the risk of nding the local
rather than global minimum, MLFN were used with the Conjugate
Gradient Decent algorithm (Bishop, 1995) and Simulated Annealing method (Masters, 1995) rather than back propagation to
achieve the global optimum. Here, the hyperbolic tangent
function was used as the activation function in hidden layer

Author's personal copy

344

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

neurons; this function simply returns the weighted sum of inputs.

When using GRNN, one does not need to investigate the structure
of a net and execute an iterative training procedure, which are
required by the back-propagation method. The net conguration
always has two hidden layers of neurons.
In either MLFN or GRNN, identied input variables (described
in the following section) were employed to develop cost output
model with the aid of NeuralTools, a neural network add-in for
Microsoft Excel. The following describes the overview of this
process: (1) data preparationdene the data sets with the Data
Set Manager function; (2) traininga neural network is generated
from the dened data set comprised of cases with known cost

Input layer

Hidden layer

Output layer

x1

x2
Y

xp

Input
layer

Pattern layer (one neuron

per training case)

Summation
layer

Output
layer

y1

x1

y2
N
x2

xp

Y
D

yn

Fig. 4. MLFN and GRNN topology: (a) MLFN and (b) GRNN.

output values; (3) testinga subset of the historical data is

reserved for testing to see how well it does at predicting known
cost values; (4) predictiononce trained and tested, the networks
can be used to predict outputs for reserved or newly collected
cases. The chosen topology and network training procedure were
generated and conducted based on the aforementioned algorithms with an iterative and time-intensive process. NeuralTools
supports MLFN and GRNN congurations to give the best possible
predictions. Once the best net is reached, the user can quickly use
it for predicting.

4.3. Case-based reasoning model

Case-based reasoning has been demonstrated as an appropriate technique for preparing early estimates from subsequent
evidential support and observations. (1) Many studies demonstrated that CBR provides rapid and accurate solutions despite
some information lacking precision (Duverlie and Castelain,
1999), (2) case-based reasoning has been used in problem-solving
as cognitive learning and adaptive intuitive solutions to new cases
(Han et al., 2005), (3) notably, CBR can be applied to both
qualitative and quantitative data, reecting the dataset types in
the real world (Mendes et al., 2005), and (4) case-based reasoning
provides a simple means of measuring project cost given that
most studies have identied the existence of non-linear relationships between cost and inuential factors (Caputo and Pelagagge,
2008; Cavalieri et al., 2004; Lowe et al., 2006).
Thus, CBR is based on the method of characterizing projects
using a number of essential attributes. Additionally, CBR is a
cognitive model that incrementally and sustainably learns from
particular experiences. In summary, CBR requires a series of
systematic procedures to extract relevant knowledge from
experience, integrate a case into an established knowledge
structure, index the case for subsequent matching, adapt a new
case to similar cases, and save cases in a knowledge base
(Toussaint and Cheng, 2006). Fig. 5 presents a generic CBR
owchart.
Eq. (2) shows the similarity measure between the new case
and previously stored cases. Similarity between the cases
increased as this value approached one:
Pn
I R
j 1 Wj  Simfj ,fij
Similarityf I ,fiR
2
Pn
j 1 Wj

Problem
statement

Target case entry

Historical
case input
Case
indexing

Case
saving

Case
retrieval
Similarity
measure

Case base
Case
mapping

Case
adaptation

Determination
of
attribute weights

Case
testing
Fig. 5. CBR owchart.

Author's personal copy

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

where Similarityf I ,fiR represents the case similarity measure

between the input case, fI, and the retrieved ith case of the casebase, fiR . Furthermore, Wj is assumed to be equal weight of the
Pn
input attributes or determined through GRNN model.
j 1 Wj
represents the summation weights of n attributes in each case,
I
and here n equals 14. Simfj ,fijR is calculated as a ratio
representing the jth attribute similarity measure of the input
case, fjI , and the retrieved ith case from the case library. Eq. (3)
provides the measurement formula for the above attribute
similarity. The closer the value is to one, the greater the similarity
of the two attributes:
SimfjI ,fijR

min9fjI 9,9fijR 9
max9fjI 9,9fijR 9

345

industry report indicates there will be 1.2 billion LCD TVs in 2015
(PIDA, 2008). The study company fabricates more than 15 types of
manufacturing equipment for TFT-LCD production. The laser
repair (Repair), light-on-text (LOT), and macro inspection systems
(Macro) account for 48% of total revenue over the last three years.
Since obtaining a complete collection of data for all previous
projects is unlikely due to the lack of centralized project
management information system (PMIS), the primary goal is to
generalize prediction capability of AI techniques using available
project data. Based on the principle of priority management, these
three systems were treated as the target samples and were
thereafter utilized in prediction model development.

where min (  ) represents the minimum value of the new input

case, fjI , and the retrieved ith case for the jth attribute and max (  )
represents the higher of the above two values.
There are two variants of the above CBR model proposed in
this study. One model is to assume the equal weight for each
attribute (CBREW) and the other is to delete the worst reasoning
case and repeat the CBR ow (Fig. 5). The purpose of the latter CBR
model is to adapt the equal-weight CBR procedure by leaving the
worst case out in the reasoning system and is thereafter named as
CBREWA model. All the similarity measure and accuracy calculations can be accomplished with the usage of spreadsheet and
Macro language.
4.4. Hybrid intelligence model
This study uses a novel Hi procedure, which incorporates
previously described stepwise MRA with GRNN (SMRAGRNN),
and GRNN with CBR (GRNNCBR) to determine whether Hi
achieves better modeling performance compared with that when
using individual techniques. Further, rather than incorporating
subjective opinions about the relative importance of multiple
attributes, impact analysis of attributes using the GRNN model
was performed to acquire average weighted coefcients for
similarity computations in CBR modeling (GRNNCBR).
The third variant, CBREWAGRNNCBR, is to establish the Hi
model by performing the three estimating techniques in series,
rst CBREWA (delete the worse case rst through equal-weight
attribute CBR), then GRNN (determine the attribute weights from
the resulting historical cases), and nally the CBR system with the
attribute weights from GRNN output.
The last variant Hi model, GRNNCBR, is to determine the
attribute weights through GRNN rst, the attributes weights are
subsequently input to Eq. (2) to perform the CBR, and nally the
worst reasoning case is removed and the typical CBR procedure is
repeated as the nal estimating model.
Real-world project data were acquired from a leading vendor
of TFT-LCD optoelectronic equipment located in the Tainan
Science Park, southern Taiwan. The same number of predictor
variables was used such that performance comparisons of
response variables were based on the same input conguration
across various AI models.

5. Case study of the Taiwanese high-tech equipment

development company
The study company is a world-leading equipment manufacturer that provides midstream TFT-LCD panel manufacturers in
Taiwan, including Chi Mei Optoelectronics, AU Optronics, Innolux
Display Corp., Chunghwa Picture Tubes, Ltd., Allied Material
Technology Corp., and TPO Displays Corp. Taiwans TFT-LCD

5.1. TFT-LCD manufacturing inspection and repairing equipment

In TFT-LCD manufacturing, especially such large products as
TVs, display panels must have high resolutions and improved
picture quality, in particular, high luminance, high contrast ratio,
wide viewing angle and rapid refresh rates (Pardo et al., 2004).
Therefore, to ensure a high yield rate for products that meet these
criteria and reduce costs, strict quality-assurance and qualitycontrol measures must be applied by the manufacturer, meaning
that careful inspection of panels has a vital role in the
manufacturing process (Lin et al., 2009). Products with defects
or imperfections can then be identied and repaired or recycled.
Panel inspections include various examination methods that use,
say, optical instruments, electrical instruments, and even the human
eye (Gaillet et al., 2007; Park and Yoo, 2009). For instance, during
LCM process, many spot-type defects can occur on its surface. These
defects are classied into two types, i.e., macro-defects and microdefects, which can be detected by macro- and micro-inspection
systems, respectively. Inspections for such defects are very important in determining whether an LCM is defect-free.
The LOT is also an important quality-control measurement
system used during the panel and module assembly processes
when manufacturing TFT-LCDs. Generally, the LOT process
includes inspection of the panel surface, modules, and panel
picture display quality.
Defects can be repaired when identied. A typical repair
system has a laser repair unit, tape grinding unit, color repair unit,
and dry lm repair unit (Ishikawa et al., 2005; Kim, 2007; Ning
et al., 2003). For example, a laser repair unit uses a solid-state
355/532 nm-wavelength laser to repair devices.
5.2. Data description and feature selection
The study company transformed itself into a high-tech
company in the 1990s specializing in developing TFT-LCD
manufacturing and testing equipment. Despite rapid business
growth, the management pays little attention to the PMIS and
project les were archived in various departments in paper-based
or electronic formats. Although PMIS are important to PMs, some
problems exist (i.e., the extent to which a PMIS should be
customized, software efcacy, reliability, effort spent and benet/
cost ratio), for most manufacturing-oriented companies using
PMIS. Additionally, experienced personnel are often reluctant to
change from a familiar working environment to prevent making
additional mistakes. The data preprocessing stage consumed
considerable time as the research team needed to acquire cost
data and related parameters from manual documents, project
les, and accounting information.
For instance, geometrical and physical features of equipment
were collected from the R&D department; manufacturing process
data were collected from the product management department;
client and order data were obtained from the business and

Author's personal copy

346

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

marketing department; receivable and payment data were

acquired from the nance and accounting department; contract
and patent adoption issues were obtained from legal affairs; and
equipment layouts were acquired from the engineering & design
department. Accordingly, the company should keep operational
project data in a centralized data warehouse, not only to archive
project knowledge as corporate assets but also facilitate business
intelligence discovery.
Developing a model for early estimation of TFT-LCD manufacturing equipment during the early price quotation stage requires
careful selection of features that have predictive relationships in the
manufacturing process. These features must be identied based on
equipment functions and characteristics (e.g., size, weight, shape,
process, and equipment type); these features can be accessed easily
from the client during the conceptual stage. In particular, quoted
prices were not traceable in the study company, and only actual
close-out cost data after equipment delivery were obtained. A pilot
analysis shows that the explanatory power (R2) and standard error
of time-adjusted cost data compared to those of non-adjusted cost
data is minor. Equipment development costs were therefore not
converted based on year.
Due to the accessibility and availability of project data, tradeoffs were made while selecting factors for subsequent model
construction. After consulting experienced engineers of the hightech company and extracting project attributes from documents,
the following candidate factors (Table 2) were determined as
potential predictors (inputs) for preliminary cost estimates of
TFT-LCD manufacturing equipment.
The project sample originally consisted of 519 TFT-LCD
equipment development projects that were delivered in 2005
2007. After aggregation of multiple orders from the same client,
sample size was reduced to 129 and a derived variable, number of
orders (NOOs), was created to account for the impact of NOOs on
quoted price. The geometrical features comprise the physical
properties of fabrication equipment. The size of the glass substrate
in product lines ranges from 316  538 to 2200  2500 mm. The
production characteristics are fabrication equipment typesRepair, Macro, and LOTand patents employed to manufacture the
machines. The manufacturing processes include TFT, CF, LCD, and
LCM. A nominal variable stands for customers who place orders.
Some variables, such as NOOs, down payment, and payment
conditions, were adopted to represent operations planning.

5.3. Model validation and measures of prediction performance

Data sets are frequently referred to as training sets, validation sets,
and test sets in modeling processes; however, some confusion exists
in literature regarding the denitions of these terms. To further dene
the data sets, modeling performance and reliability were veried
using k-fold cross validation, also known as rotational estimation.
Extensive trials on numerous data sets with different learning
techniques suggest that a k value of 10 is an optimal number to
generate the smallest estimation error (Kohavi, 1995; Witten and
Frank, 2005). Therefore, cases were randomly separated into ten
mutually exclusive subsets of roughly equal size to minimize bias
associated with random sampling of training and holdout data
samples. All models were trained and validated ten times, and each
time each model was trained on all except for one fold, which was left
for validation purposes. Neter et al. (1996) indicated the best means
of model testing is through collecting new data, which enables users
to assess the applicability of a model to data beyond those on which
the model is based (Neter et al., 1996). Thus, data for newly
completed projects were collected upon the verication of model
building for test purposes..
In this study, three accuracy measures, mean absolute error
rate (MAER), root mean square error (RMSE) and complementary
correlation coefcient R0 , were used as comparison criteria. The
MAER, RMSE and R0 are derived as follows:
Pn
1 i 1 9Yi,predicted Yi,observed 9
 100%
4
MAER
n
Yi,observed
v
u n
u1 X
RMSE t
Y
Y
2
n i 1 i,predicted i,observed

Pn
i 1 Yi,predicted Y predicted Yi,observed Y observed
Ru 1R 1 q
Pn
2 Pn
2
i 1 Yi,predicted Y predicted
i 1 Yi,observed Y observed

6
where n is the number of validation data, Yi is the equipment
development cost and the bar denotes the variable mean.
As the prediction model is trained and tested k times, the
overall performance through cross-validation can be denoted as

Table 2
List of model variables and their Descriptive statistics.

Input variables

Output variable

Parameter

Min

Max

Mean

MRA

ANNs

CBR

Geometrical features
1. Length of fabrication equipment (mm)
2. Width of fabrication equipment (mm)
3. Height of fabrication equipment (mm)
4. Weight of fabrication equipment (ton)
5. Length of glass substrate (mm)
6. Width of glass substrate (mm)

1650
1000
1475
0.3
538
316

5100
4200
29071
15
2500
2200

3233.1
2220.7
2979.0
3.7
1266.6
953.8

V1
V2
V3
V4
V5
V6

X1
X2
X3
X4
X5
X6

Numerical
Numerical
Numerical
Numerical
Numerical
Numerical

Production characteristics
7. Equipment development type
8. Manufacturing process
9. Customer
10. Order type
11. No. of patent

(Repair, Macro, LOT)

(TFT, CF, LCD, LCM)
(CMO, AMTC, CPT, INNOLUX, TPO, AUO)
(old, new)
0
7
2

Dummy
Dummy
Dummy
Dummy
V18

X7
X8
X9
X10
X11

String
String
String
String
Numerical

Operations planning
12. No. of orders (NOOs)
13. Down payment
14. Payment conditions

1
26
4.6
(60%, 70%, 80%)
(1-month check, 2-month check)

V19
Dummy (D20)
Dummy (D21)

X12
X13
X14

Numerical
String
String

Predicted cost (10,000 TWD)

198

YMRA

YANNs

YCBR

1700

691.7

(D7, D8)
(D9, D10, D11)
(D12, D13, D14, D15, D16)
(D17)

Author's personal copy

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

the average of k individual measures:

M

k
X

1
M
ki1 i

in which M is the cross-validation mean performance measure,

k is the number of folds used, and Mi is the performance measure
of each fold.
To compound the effect of accuracy measures, a dimensionless
minmax normalization score (MNS) is proposed as
Pn
Ac minAc =maxAc minAc 
8
MNS i 1
NOM
where Ac represents the three different accuracy measures, and
NOM is the number of measurements. The MNS range is 01; the
size of the coefcient is inversely related to the effectiveness of
the overall evaluation measures.
5.4. Results and discussions
In total, nine models were created utilizing resource data. All
models utilized the same set of input variables described in
Table 2. Data sets were split into ten cross-fold subsets for each
model. As described in Section 4, proposed modeling techniques
and procedures were stepwise MRA (SMRA), MLFN, GRNN, CBR
using equal weights of attributes (CBREW), CBR using equal
weights and the leave-the-worst-one-out adaption method
(CBREWA), SMRA to determine the signicant attributes as inputs
for the GRNN (SMRAGRNN), GRNN to determine the weights of
attributes for CBR modeling (GRNNCBR), CBREWA rst and then
GRNN to determine the weights of attributes for subsequent CBR
modeling (CBREWAGRNNCBR), GRNN to determine the weights
of attributes for subsequent CBR, and the leave-the-worst-one-out
adaptation method (GRNNCBR).
Accuracy measures were calculated using Eqs. (2)(5) for all
cross-fold models (Table 3). Variability in accuracy measures was
also derived using their standard deviations. For accuracy
performance via the MAER, the SMRA (21.78%) was better than
MLFN (30.88%), but still worse than GRNN (13.74%). Generally,
CBR (14.73%) was superior to MRA (21.78%) and ANNs (22.31) in
terms of single techniques and MAER.
Combining AI techniques increased performance accuracy by
up to 46.62% compared to with that of MRA, ANNs, and CBR used
alone (Table 3). Of the various hybrid models, the GRNNCBR
model was best with an MAER of 8.83%, which is satisfactory
during the quotation stage given the scarcity of project information for this industrial sector. Generally, the proposed hybrid
models generate MAER of 8.8314.29% with a variability range of
3.616.19%. Such estimation techniques as GRNN or CBREWA
alone were better than some hybrid models.
Statistical analyses indicate that the RMSE and R0 have patterns
similar to that of MAER (Fig. 6), revealing that using MAER only

347

may be sufcient for evaluating prediction performance in term of

accuracy. To compound the effects of accuracy measures and their
estimation reliability, Eq. (6) was applied to calculate the MNS to
assess the overall evaluation measure. The hybrid GRNNCBR
method has the lowest error rate and MNS. From a comparative
perspective, combining GRNN with CBR and the adaptive
technique is the best at estimating development cost for the
TFT-LCD equipment.
In some cases, accuracy may not be the only criterion used to
choose the best estimation model, and computation complexity is
usually considered the time spent modeling, consistency in
precise model training and testing over time, ability to clearly
explain modeling results, ease of modeling, complexity encoding
a problem, and the availability of off-the-shelf software, which are
commonly considered when estimating. Analyses show that the
GRNNCBR model had better accuracy, shorter training time, and
easier topological specications of numbers of hidden layers and
nodes to predict development costs of TFT-LCD equipment
To test the proposed models, the best top two estimation
models, the hybrid GRNNCBR and CBREWA models, were
utilized to predict costs of the newly collected ten projects from
the study company. Moreover, the conventional approach (the
analogous approach) was also used to assess the inuence of
using AI. Fig. 7 shows the calculated prediction errors of the
proposed articial models versus the conventional approach
currently utilized in this industrial sector. The GRNNCBR and
CBREWA methods generate MAERs of 10.4% and 13.3%,
respectively, outperforming the analogous approach (16.1%),
which veries prediction validity. For the best Hi model (GRNN
CBR), the stand deviation of prediction errors is 9.9% which is
slightly higher than the manual method (9.4%) while achieving
35.4% accuracy improvement.

6. Conclusions and managerial implications

The lack of a rapid and accurate estimation model warrants the
application of AI-based methods. This study presents and
compares MRA, ANNs, CBR, and hybrid AI models to estimate
costs for equipment used to manufacture TFT-LCDs. Field interviews indicate that AI-based approaches are not commonly
adopted in practice for project management. Although the
primary task of sales managers is to sell a product to satised
customers, possessing some knowledge of intelligent cost estimation techniques is benecial to provide timely quotes without
jeopardizing a companys nancial plan.
Furthermore, project data were not stored in a centralized data
warehouse or database management system and were archived in
various departments. Thus the research team expended considerable effort to clean, merge, append and transform raw data into a
at le. This experience suggests that the study company, though

Table 3
Cross-fold modeling performance with three evaluation measures.
Technique Method
MRA
ANNs
CBR
Hi

19

SMRA
MLFN9
GRNN4
CBREW8
CBREWA2
SMRA-GRNN6
GRNN-CBR3
CBREWA-GRNN-CBR5
GRNN-CBRA1

MAER (%)

sMAER (%)

RMSE (104TWD)

sRMSE (104TWD)

R0

sRu

MNS

21.78
30.88
13.74
18.89
10.56
14.29
12.87
11.66
8.83

5.33
9.35
3.20
8.75
4.10
3.61
6.19
4.97
5.07

165.25
212.48
132.00
174.68
139.64
148.55
126.02
135.87
101.99

63.37
63.90
43.98
56.58
79.15
60.00
64.15
79.34
51.48

0.10
0.15
0.07
0.13
0.06
0.09
0.07
0.09
0.05

0.07
0.07
0.05
0.10
0.04
0.07
0.07
0.11
0.05

0.55
1.00
0.23
0.64
0.17
0.36
0.20
0.28
0.00

denotes performance ranking based on the MNS

Author's personal copy

348

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

250.00
RMSE
MAER
(1-R)

Cross-fold Performance Measure

200.00

150.00

100.00
0.40
0.30
0.20
0.10
0.00
MR

ML

MRA

FN

NN
GR

RE

R
R
RA
NN
-CB
-CB
CB
GR
NN
NNN
AA
N
A
R
A
M
WA
RE
CB

WA

CB

CB

RE

CBR

ANNs

Hi

Modeling Technique
Fig. 6. Cross-fold performance measures for various modeling techniques.

60
GRNN-CBRA
CBREWA
Conventional Approach

Absolute Error Rate (%)

50

40

Estimating Method

MAER

GRNN-CBRA

10.4%

CBREWA

13.3%

Conventional Approach 16.1%

30

20

10

5
6
Test Case

10

Fig. 7. Prediction error of articial intelligence models versus the conventional approach.

providing cutting-edge devices, lacks an appropriate centralized

information system for warehousing information. Some project
records, of course, may have no benets or learned knowledge for
subsequent reuse. However, many projects with valuable nuggets,
such as implicit corporate experience, unknown patterns and
interesting rules, may be lost when not kept by a sustainable
management system for subsequent knowledge discovery in a
timely manner, thereby leading to organizational losses.
Notably, ANNs, unlike MRA and CBR, are relatively opaque to
human interpretation and frequently require extended training

time. In this study, most time was spent identifying the optimal
ANN topologies and training to improve the error rate. Nevertheless, an improved ANN technique, namely, GRNN integrated
with CBR to calculate the similarity measure ts data better than
traditional ANNs alone in terms of trade-offs such as time spent,
learning curve and explanatory power. Case-based reasoning
alone and hybrid ANN-CBR models are applicable and superior
AI-based techniques for estimating cost of developing TFT-LCD
equipment. In particular, the accuracy of the proposed hybrid
AI-based technique was up to 35.4% better than that to the

Author's personal copy

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

conventional approach currently used in this sector. Further, the

hybrid AI-based technique outperformed MRA, implying that the
relationship between equipment development cost and project
parameters are most suited to non-linear modeling.
In summary, the proposed models, GRNN, CBR, and the hybrid
method developed in this study were adequate in predicting the
development costs of TFT-LCD equipment at early stages as
prediction error did not exceed an acceptable threshold. Once
operations data from multiple departments can be integrated
within a centralized management information system, subsequent DM can be streamlined and create useful patterns that
become implicit assets of the corporation efciently. A decisionsupport system for decision-makers can then be developed based
on extraction of these non-trivial rules. Future research can focus
on the design and implementation of a user-friendly automated
web service with the proposed hybrid AI-based model embedded
to efciently determine equipment costs.
TFT-LCD panels are currently a global trend and users have
approached the multiple sizes varied from large to economic size
due to diverse purposes make the equipment development in
demand. With experience, good judgment and timely adjustments, systematic knowledge discovery from sales records offers
PMs of TFT-LCD manufacturing equipment an effective and
efcient way of estimating costs in the early stages of development for their stakeholders. This study scope limits to consider
the proposed AI estimating methods and three types of equipment
at conceptual stage. When used to estimating new equipment cost,
the proposed model should be applied with caution and must be
updated regularly once new criteria, such as sustainability and
environmental protection, are adopted, or new manufacturing
techniques/products are developed or costs increase.

Acknowledgements
The authors would like to thank the National Science Council
of the Republic of China, Taiwan, for nancially supporting this
research. Gratitude is also expanded to the study company whose
generosity in data and experience sharing makes the study
possible.
References
Al-Sakran, H., 2006. Software cost estimation model based on integration of multiagent and case-based reasoning. Journal of Computer Science 2, 276282.
Albright, S.C., Winston, W., Zappe, C.J., 2006. Data Analysis & Decision Making with
Microsoft Excel, 3e. Thomson Brooks/Cole, CA.
An, S.-H., Kim, G.-H., Kang, K.-I., 2007a. A case-based reasoning cost estimating
model using experience by analytic hierarchy process. Building and Environment 42, 25732579.
An, S.-H., Park, U.Y., Kang, K.-I., Cho, M.-Y., Cho, H.-H., 2007b. Application of
support vector machines in assessing conceptual cost estimates. Journal of
Computing in Civil Engineering 21, 259264.
Asiedu, Y., Gu, P., 1998. Product life cycle cost analysis: state of the art review.
International Journal of Production Research 36, 883908.
Bashir, H.A., Thomson, V., 2001. Models for estimating design effort and time.
Design Studies 22, 141155.
Baykasoglu, A., Kaplanoglu, V., 2008. Application of activity-based costing to a land
transportation company: a case study. International Journal of Production
Economics 116, 308324.
Ben-Arieh, D., Qian, L., 2003. Activity-based cost management for design
and development stage. International Journal of Production Economics 83,
169183.
Berlin, S., Raz, T., Glezer, C., Zviran, M., 2009. Comparison of estimation methods of
cost and duration in IT projects. Information and Software Technology 51,
738748.
Bishop, C.M., 1995. Neural Networks for Pattern Recognition. Oxford.
Bode, J., 1998. Decision support with neural networks in the management of
research and development: concepts and application to cost estimation.
Information & Management 34, 3340.
Bode, J., 2000. Neural networks for cost estimation: simulations and pilot
application. International Journal of Production Research 38, 12311254.

349

Bouaziz, Z., Younes, J.B., Zghal, A., 2006. Cost estimation system of dies
manufacturing based on the complex machining features. Int J Adv Manuf
Technol 28, 262.
Camargo, M., Rabenasolo, B., Jolly-Desodt, A.-M., Castelain, J.-M., 2003. Application
of the parametric cost estimation in the textile supply chain. Journal of Textile
and Apparel, Technology and Management 3, 112.
Caputo, A.C., Pelagagge, P.M., 2008. Parametric and neural methods for cost
estimation of process vessels. International Journal of Production Economics
112, 934954.
Cavalieri, S., Maccarrone, P., Pinto, R., 2004. Parametric vs. neural network models
for the estimation of production costs: a case study in the automotive industry.
International Journal of Production Economics 91, 165177.
Chantaraskul, S., Cuthbert, L.G., 2005. Congestion pattern matching in case-based
reasoning control for 3G networks. IEEE International Conference on Wireless
and Mobile Computing, Networking and Communications, WiMob2005 2,
134-141.
Chou, J.-S., 2009a. Generalized linear model-based expert system for estimating
the cost of transportation projects. Expert Systems with Applications 36,
42534267.
Chou, J.-S., 2009b. Web-based CBR system applied to early cost budgeting for pavement
maintenance project. Expert Systems with Applications 36, 29472960.
Cigizoglu, H.K., 2005. Generalized regression neural network in monthly ow
forecasting. Civil Engineering and Environmental Systems 22, 7181.
Ciurana, J., Quintana, G., Garcia-Romeu, M.L., 2008. Estimating the cost of vertical
high-speed machining centres, a comparison between multiple regression
analysis and the neural networks approach. International Journal of Production Economics 115, 171178.
Clark, M.G., 2001. Liquid Crystal Devices, in: Robert, A.M. (Ed.), , Encyclopedia of
Physical Science and Technology. Academic Press, New York, pp. 701715.
De Barcelos Tronto, I.F., Simoes Da Silva, J.D., SantAnna, N., 2007. Comparison of
articial neural network and regression models in software effort estimation.
In: IEEE International Conference on Neural NetworksConference Proceedings, pp. 771776.
Duverlie, P., Castelain, J.M., 1999. Cost estimation during design step: parametric
method versus case based reasoning method. International Journal of
Advanced Manufacturing Technology 15, 895906.
Emsley, M.W., Lowe, D.J., Duff, A.R., Harding, A., Hickson, A., 2002. Data
modelling and the application of a neural network approach to the
prediction of total construction costs. Construction Management and
Economics 20, 465.
Finnie, G.R., Wittig, G.E., Desharnais, J.M., 1997. Comparison of software effort
estimation techniques: using function points with neural networks, casedbased reasoning and regression models. Journal of Systems and Software 39,
281289.
Gaillet, M., Yan, L., Teboul, E., 2007. Optical characterizations of complete TFT-LCD
display devices by phase modulated spectroscopic ellipsometry. Thin Solid
Films 516, 170174.
Gang, K.W., 2005. Research planning approach using knowledge map integrated
with case-based reasoning. In: IEEE International Engineering Management
Conference II, pp. 897901.
Han, S.-G., Lee, S.-G., Jo, G.-S., 2005. Case-based tutoring systems for procedural
problem solving on the 29, 573582.
Hegazy, T., Fazio, P., Moselhi, O., 1994. Developing practical neural network
applications using back-propagation. Microcomputers in Civil Engineering 9,
145159.
Hsieh, K.-L., Lu, Y.-S., 2008. Model construction and parameter effect for TFT-LCD
process based on yield analysis by using ANNs and stepwise regression. Expert
Systems with Applications 34, 717724.
Hwang, Y.-D., Lin, Y.-C., Lyu Jr, J., 2008. The performance evaluation of SCOR
sourcing processThe case study of Taiwans TFT-LCD industry. International
Journal of Production Economics 115, 411423.
Ishikawa, T., Miyai, H., Yamazaki, T., Seid, H., Kusunose, H., Yoshimoto, H., 2005. A
novel LCD color lter repair machine with fully dry process, IDW/AD05
Proceedings of the 12th International Display Workshops in Conjunction with
Asia Display 2005, 1 ed, Takamatsu, pp. 335338.
ITIS, 2008. pp. /http://www.itis.org.tw/S, Accessed Dec. 18, 2008.
Kee, R., 2010. The sufciency of target costing for evaluating production-related
decisions. International Journal of Production Economics 126, 204211.
Kim, G.-H., An, S.-H., Kang, K.-I., 2004. Comparison of construction cost estimating
models based on regression analysis. neural networks, and case-based
reasoning 39, 12351242Building and Environment 39, 12351242.
Kim, K.R., 2007. Laser microfabrications for next-generation at panel display.
Korean Journal of Materials Research 17, 352357.
Kohavi, R., 1995. A study of cross-validation and bootstrap for accuracy
estimation and model selection. In: The 14th International Joint Conference
on Articial Intelligence (IJCAI). Morgan Kaufman, San Francisco, CA,
pp. 11371145.
Kwak, Y.H., Watson, R.J., 2005. Conceptual estimating tool for technology-driven
projects: exploring parametric estimating technique. Technovation 25,
14301436.
Lee, T.-M., Choi, Y.-J., Nam, S.-Y., You, C.-W., Na, D.-Y., Choi, H.-C., Shin, D.-Y., Kim,
K.-Y., Jung, K.-I., 2008. Color lter patterned by screen printing. Thin Solid
Films 516, 78757880.
Lin, J.T., Chen, T.-L., Lin, Y.-T., 2009. Critical material planning for TFT-LCD
production industry. International Journal of Production Economics 122,
639655.

Author's personal copy

350

J.-S. Chou et al. / Int. J. Production Economics 128 (2010) 339350

Lowe, D.J., Emsley, M.W., Harding, A., 2006. Predicting construction cost using
multiple regression techniques. Journal of Construction Engineering and
Management 132, 750758.
Marban, O., Menasalvas, E., Fernandez-Baizan, C., 2008. A cost model to estimate the
effort of data mining projects (DMCoMo). Information Systems 33, 133150.
Marir, F., Watson, I., 1995. CBRefurb: case-based cost estimation. Case Based
Reasoning: Prospects for Applications, IEE Colloquium on, 5/1-5/3.
Masters, T., 1995. Advanced Algorithms for Neural Networks. Wiley.
Mendes, E., Mosley, N., Counsell, S., 2002. The application of case-based reasoning
to early web project cost estimation. In: ProceedingsIEEE Computer Societys
International Computer Software and Applications Conference, pp. 393398.
Mendes, E., Mosley, N., Counsell, S., 2005. Exploring case-based reasoning for web
hypermedia project cost estimation. International Journal of Web Engineering
and Technology 2, 117143.

Menozzi, M., Lang, F., Napin,

U., Zeller, C., Krueger, H., 2001. CRT versus LCD:
effects of refresh rate, display technology and background luminance in visual
performance. Displays 22, 7985.
Metaxiotis, K., 2004. An expert system for the reduction of environmental cost in the
textile industry. Information Management and Computer Security 12, 218227.
Nassif, L.N., Nogueira, J.M., Karmouch, A., Ahmed, M., De Andrade, F.V., 2007. Job
completion prediction using case-based reasoning for grid computing
environments. Concurrency Computation Practice and Experience 19,
12531269.
Neter, J., Kutner, M.H., Nachtsheim, C.J., Wasserman, W., 1996. Applied Linear
Statistical Models fourth ed. IRWIN.
Niazi, A., Dai, J.S., Balabani, S., Seneviratne, L., 2006. Product cost estimation:
technique classication and methodology review. Journal of Manufacturing
Science and Engineering 128, 563575.
Ning, J.P., Li, J.Q., Lu, X.C., Li, X.G., Yao, J.Q., 2003. Study for application of laser to
repairing ITO lm on LCD board. Guangdian Gongcheng/Opto-Electronic
Engineering 30 15-17 +20.
Ou-Yang, C., Lin, T.S., 1997. Developing an integrated framework for feature-based
early manufacturing cost estimation. International Journal of Advanced
Manufacturing Technology 13, 618629.
Palisade, 2008. NeuralTools: Neural Network Add-In for Microsoft Excel Version
5.0. Palisade Corporation, Ithaca, NY, USA.
Pardo, P.J., Pe!rez, A.L., Suero, M.I., 2004. Validity of TFT-LCD displays for colour
vision deciency research and diagnosis. Displays 25, 159163.
Park, N.K., Yoo, S.I., 2009. Evaluation of TFT-LCD defects based on human visual
perception. Displays 30, 116.

PIDA, 2008. Photonics Industry & Technology Development Association (PIDA),

available online at website /http://www.pida.org.twS, accessed 7/30/2008.
PMI, 2004. A Guide to the Project Management Body of Knowledge (PMBOK
Guide), 3rd edition, Project Management Institute.
Rogozhin, A., Gallaher, M., Helfand, G., McManus, W., 2010. Using indirect cost
multipliers to estimate the total cost of adding new technology in
the automobile industry. International Journal of Production Economics 124,
360368.
Sabnis, R.W., 1999. Color lter technology for liquid crystal displays. Displays 20,
119129.
Seo, K.-K., Park, J.-H., Jang, D.-S., Wallace, D., 2002. Prediction of the life cycle cost
using statistical and articial neural network methods in conceptual product
design. International Journal of Computer Integrated Manufacturing 15,
541554.
Smith, A.E., Mason, A.K., 1997. Cost estimation predictive modeling: regression
versus neural. Engineering Economist 42, 137.
Smith, J.R., 2005. LCD Modules, Programming the PIC Microcontroller with
MBASIC.. Newnes, Burlington, pp. 67-85.
Specht, D., 1991. A general regression neural network. IEEE Transactions on Neural
Networks 2, 568576.
Staub-French, S., Fischer, M., Kunz, J., Ishii, K., Paulson, B., 2003. A feature
ontology to support construction cost estimating. Cambridge Journals 17,
133154.
Thomasson, B.J., Ratcliffe, M.B., Thomas, L.A., 2006. Improving the tutoring of
software design using case-based reasoning. Advanced Engineering Informatics 20, 351362.

Tornberg, K., Jamsen,

M., Paranko, J., 2002. Activity-based costing and process
modeling for cost-conscious product design: a case study in a manufacturing
company. International Journal of Production Economics 79, 7582.
Toussaint, J., Cheng, K., 2006. Web-based CBR (case-based reasoning) as a tool with
the application to tooling selection. International Journal of Advanced
Manufacturing Technology 29, 2434.
Tsai, K.-m., Wang, S.-c., 2009. Multi-site available-to-promise modeling for
assemble-to-order manufacturing: an illustration on TFT-LCD manufacturing.
International Journal of Production Economics 117, 174184.
Verlinden, B., Duou, J.R., Collin, P., Cattrysse, D., 2008. Cost estimation for sheet
metal parts using multiple regression and articial neural networks: A case
study. International Journal of Production Economics 111, 484492.
Witten, I.H., Frank, E., 2005. Data Mining: Practical Machine Learning Tools and
Techniques, Second edition, Morgan Kaufmann.