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

net/publication/337021489

Artificial Intelligence and Parametric Construction Cost Estimate Modeling:

State-of-The-Art Review

Article  in  Journal of Construction Engineering and Management · November 2019

DOI: 10.1061/(ASCE)CO.1943-7862.0001678

CITATIONS READS
71 3,688

1 author:

Haytham H. Elmousalami
Zagazig University
22 PUBLICATIONS   285 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

PREDICTION OF CONSTRUCTION COST FOR FIELD CANALS IMPROVEMENT PROJECTS IN EGYPT View project

All content following this page was uploaded by Haytham H. Elmousalami on 25 January 2021.

The user has requested enhancement of the downloaded file.

 State-of-the-Art Review

Artificial Intelligence and Parametric Construction Cost

Estimate Modeling: State-of-the-Art Review
Haytham H. Elmousalami 1

Abstract: This study reviews the common practices and procedures conducted to identify the cost drivers that the past literature has classified
into two main categories: qualitative and quantitative procedures. In addition, the study reviews different computational intelligence (CI)
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

techniques and ensemble methods conducted to develop practical cost prediction models. This study discusses the hybridization of these
modeling techniques and the future trends for cost model development, limitations, and recommendations. The study focuses on reviewing
the most common artificial intelligence (AI) techniques for cost modeling such as fuzzy logic (FL) models, artificial neural networks (ANNs),
regression models, case-based reasoning (CBR), hybrid models, diction tree (DT), random forest (RF), supportive vector machine (SVM),
AdaBoost, scalable boosting trees (XGBoost), and evolutionary computing (EC) such as genetic algorithm (GA). Moreover, this paper pro-
vides the comprehensive knowledge needed to develop a reliable parametric cost model at the conceptual stage of the project. Additionally,
field canals improvement projects (FCIPs) are used as an actual case study to analyze the performance of the ML models. Out of 20 AI
techniques, the results showed that the most accurate and suitable method is XGBoost with 9.091% and 0.929 based on mean absolute
percentage error (MAPE) and adjusted R2 , respectively. Nonlinear adaptability, handling missing values and outliers, model interpretation,
and uncertainty are discussed for the 20 developed AI models. In addition, this study presents a publicly open data set for FCIPs to be used for
future model validation and analysis. DOI: 10.1061/(ASCE)CO.1943-7862.0001678. © 2019 American Society of Civil Engineers.
Author keywords: Artificial intelligence; Feature engineering; Ensemble methods; Hybrid intelligent systems; Fuzzy analytic hierarchy
process; Genetic algorithm; Factor analysis; Fuzzy logic; XGBoost; Project cost modeling.

Introduction definitive cost estimate at 30% of project design completion to con-

Cost estimating is a primary part of construction projects, in which
cost is considered as one of the major criteria in the projects’
feasibility studies and decision making at the early stages of the
project. The accuracy of estimation is a critical factor in the success
of any construction project, for which cost overruns are a major
problem, especially with the current emphasis on tight budgets.
Indeed, cost overruns can lead to the cancellation of a project
(Zhu et al. 2010; AACE 2004).
Subsequently, the cost of a construction project needs to be es-
timated within a specified accuracy range, but the largest obstacles
standing in front of a cost estimate, particularly in the early stage,
are lack of preliminary information and larger uncertainties as a
result of engineering solutions. As such, to overcome this lack of
detailed information, cost estimation techniques are used to ap-
proximate the cost within an acceptable accuracy range (AACE
2004; Chen 2002; Waty et al. 2018).
The American Association of Cost Engineers (AACE) defines
five classes of cost estimates (AACE 2004). The first class is the
detailed cost estimate at 65% of project design completion to check
estimate of bid or tender based on detailed unit cost, for which ac-
curacy ranges from −10% to þ15%. The second class is the
1
Project Engineer and Project Management Professional (PMP),
General Petroleum Company, St. 8, Ash Sharekat, Nasr City, Cairo
Governorate, Egypt; Fellow, Dept. of Energy Engineering, Technical
Univ. Berlin, Campus El Gouna, Governorate 84513, Egypt. Email:
Haythamelmousalami2014@gmail.com
Note. This manuscript was published online on October 19, 2019.
Discussion period open until March 19, 2020; separate discussions must
be submitted for individual papers. This paper is part of the Journal
of Construction Engineering and Management, © ASCE, ISSN
0733-9364.
trol of bid or tender based on detailed unit cost, for which accuracy
ranges from −15% to þ20%. The third class is the preliminary cost
estimate at 10% of project design completion to budget authoriza-
tion based on semi-detailed unit cost, for which accuracy ranges
from −20% to þ30%. The fourth class is the studying cost estimate
at 1% of project design completion for feasibility studying based
on the parametric model, for which accuracy ranges from −30% to
þ50%. The fifth class is the conceptual cost estimate at 0% of
project design completion for conceptual screening based on capac-
ity factored, judgment, analog, and parametric models, for which
accuracy ranges from −50% to þ100%. Conceptual estimating
works as the main stage of project planning in which limited project
information is available and high levels of uncertainty and risk ex-
ist. Moreover, the estimating should be completed within a limited
time period. Therefore, the accurate conceptual cost estimate is a
challenging task and a crucial process for cost engineers, project
managers, and decision makers (Jrade 2000).
This research aims to review the common computational intelli-
gence (CI) techniques used for parametric cost models and to high-
light the future trends. An accurate cost estimate is a critical aspect of
the project’s success (AACE 2004; Hegazy 2014). At the conceptual
stage, cost prediction models can be based on numerous techniques,
including probabilistic techniques such as Monte Carlo simulation,
and machine learning based techniques such as support vector ma-
chines (SVM) and artificial neural networks (ANNs) (Elfaki et al.
2014; Friedman et al. 2001). The object of selecting the optimal tech-
nique for cost modeling is to provide accurate results, minimize
prediction errors, and provide a more reliable model.
Research Methodology
The objective of the study is to review and analyze the most
common techniques and practices to develop a reliable parametric

© ASCE 03119008-1 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Literature Survey
Model Development
AI
Key parameters identification Regression Analysis
Qualitative procedure
Fuzzy Logic
Quantitative procedure ANNs
CBR
SVM
DT
Ensemble methods
Hybrid intelligent system
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
Analysis and Discussion
Remarks and future research
Fig. 1. Research methodology.
cost model. In addition, the study emphasizes the important role of
machine learning and computational intelligence that can effec-
tively improve and automate cost modeling predictions. As shown
in Fig. 1, the research methodology consists of three main phases.
The first phase is reviewing the past literature. Past literature has
been surveyed to identify the practices for cost drivers identification
and cost prediction modeling in the construction industry. Many
high-impact research journals have been reviewed, such as the
Journal of Construction Engineering and Management, the Journal
of Civil Engineering and Management, and Construction Manage-
ment and Economics. More than 100 studies have been reviewed
and analyzed, of which most papers were published from 2002 to
2018, to draw the studies’ recommendations and conclusions.
According to the first phase, developing the parametric cost
model can be divided into two main steps: cost drivers identifica-
tions and computational intelligence modeling. Therefore, the scope
of this study is reviewing the past practices used to identify the
key parameters for parametric cost models and then reviewing the
different intelligent models for cost prediction. The first part of
the literature survey is reviewing the different approaches used
for key parameters identification. The second part of the literature
survey is reviewing the different intelligent models for cost predic-
tion. The second phase of the research methodology is an analysis
and discussion of the reviewed papers, whereas the third phase is
developing the concluding remarks and future research trends. Thus,
a methodology can be generalized to review the similar and relevant
studies as follows. Content analysis is one of the research method-
ologies to analyze the content of the selected texts. However, con-
tent analysis suffers from several disadvantages; it can be extremely
time-consuming, information can be lost because of poor selection
of categorization, and it may be prone to bias. Content analysis can
be difficult to automate or computerize in cases that computationally
need text mining techniques. Therefore, content analysis is subject
to increased error. With complex texts, content analysis may provide
poor performance in text interpretation and analysis. Quantitatively,
content analysis tends often to simply consist of word counts.
However, it needs a qualitative approach to correctly interpret the
collected texts. This study has conducted a narrative review method
(Green et al. 2006) to obtain a broad qualitative perspective on AI
applications for construction cost prediction.
Poor identification of parameters means poor performance and
accuracy of the parametric model. On the other hand, the optimal
© ASCE
J. Constr. Eng. Manage., 2020, 146(1): 03119008
Key cost drivers' identification
Qualitative procedure Quantitative procedure
Traditional delphi method(TDM) Exploratory factor analysis(EFA)
Likert scale Regression methods
Fuzzy delphi method Correlation matrix
Fuzzy analytic hierarchy process(FAHP) Genetic algorithm(GAs)
Boruta feature selection algorithm
Fig. 2. Qualitative and quantitative procedure.
set of parameters produces the optimal performance of the devel-
oped model with less computation effort and fewer parameters
needed to run the model (Kan 2002). As illustrated in Fig. 2, such
procedures can be categorized into two main types: qualitative
and quantitative. The qualitative procedure includes all practices
that depend on the experts’ questionnaire and gathering opinions.
On the other hand, the quantitative procedure depends on the col-
lected data where statistical techniques are required to discover and
learn the patterns of data to extract the knowledge based on the
collected data (Hastie et al. 2009).
Key Cost Drivers Identification by Qualitative
Procedures
Conceptual cost estimation mainly depends on the conceptual
parameters of the project. Therefore, defining such parameters is
the first and the most significant step in the cost model develop-
ment. Identifying cost model parameters is the critical stage in
model development in which poor identification of cost parameters
decreases the model accuracy. Cost drivers can be determined
based on a questionnaire survey or the collected data. The quali-
tative procedures for key parameters identification are dependent
on experts’ interviews and field surveys. Many approaches such as
the traditional Delphi method, the Likert scale, the fuzzy Delphi
method (FDM), and the fuzzy analytic hierarchy process (FAHP)
have been conducted to select and evaluate the key cost drivers
based on the viewpoints of experts.
Traditional Delphi Method and Likert Scale
Traditional Delphi method (TDM) is conducted to collect experts’
opinions about a certain case. Based on the collected experts’ opin-
ions, all parameters affecting a system can be identified. Delphi
technique consists of several rounds for collecting, ranking, and
revising the collected parameters. Therefore, experts are also asked
to give their feedback and revise their opinions to enhance the qual-
ity of the survey. The Delphi rounds continue until no other opin-
ions remain and a consensus is reached (Hsu and Sandford 2007).
Therefore, the first step is to select the experts to be asked based on
their experience. The second step is to prepare a list of questions to
discover the knowledge and parameters of the proposed case study.
The third step is to apply Delphi rounds, where all experts should
be asked through interviews or their answers can be collected via
emails. The fourth step is to collect all experts’ answers and make a
list of all the collected parameters. The fifth step is to ask experts
again to assess and evaluate the parameters. Finally, the experts can
revise their parameters and state the reasons for their rating (Hsu
and Sandford 2007).
The Likert scale is a rating scale to represent the opinions of
experts that can consist of three points, five points, or seven points.
For example, a five-point Likert scale may include ratings of
03119008-2 J. Constr. Eng. Manage.
 “Extremely Important,” “Important,” “Moderately Important,” point Lij , and the maximum as point U ij . This is illustrated in
“Unimportant,” and “Extremely Unimportant,” that the experts will Eqs. (3)–(6) (Klir and Yuan 1995; Elmousalami et al. 2018a):
select to answer the questions received (Bertram 2017).
Based on the completed survey forms, statistical indices can be Lj ¼ Minðlij Þ; i ¼ 1; 2; : : : n; j ¼ 1; 2; : : : m ð3Þ
calculated to gather the final rank for each question or criterion by
experts’ ratings. Mean score is (MS) used to gather the final rating !1=n
Y
n;m
for each criterion of the survey as Eq. (1), whereas the standard Mj ¼ mij ; i ¼ 1; 2; : : : n; j ¼ 1; 2; : : : m ð4Þ
error (SE) is calculated to check the sample size of experts as i¼1;j¼1
in Eq. (2)
P
ðf × sÞ U j ¼ Maxðuij Þ; i ¼ 1; 2; : : : n; j ¼ 1; 2; : : : m ð5Þ
MS ¼ ð1Þ
n
where MS = mean score to represent the impact of each parameter ðW ij Þ ¼ ðLj ; M j ; U j Þ ð6Þ
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

based on the respondents’ answers; s = a score set to each param-

eter by the respondents; f = frequency of responses to each rating
for each parameter; and n = total number of participants
σ
SE ¼ pffiffiffi
n the experts’ common consensuses (Lij ); Mj = opinions mean of the
where SE = standard error; σ = standard deviation among partic-
ipants’ opinions for each cost parameter; and n = total number of
participants. Thus, all parameters will be collected and ranked
based on the experts’ opinions.
Fuzzy Delphi Method
Fuzzy Delphi Method (FDM) consists of the traditional Delphi
method and fuzzy theory (Ishikawa et al. 1993). Maintaining the
fuzziness and uncertainty in participants’ opinions is the advantage
of this method over the traditional Delphi method. Instead of ap-
plying the experts’ opinions as deterministic values, this method
uses membership functions such as triangular, trapezoidal, or
Gaussian functions to map the deterministic numbers to fuzzy num-
bers. Accordingly, the reliability and quality of the Delphi method
are improved (Liu 2013). The objective of the FDM is to avoid
misunderstanding of the experts’ opinions and to make a good
generalization to the experts’ opinions.
The first step of FDM is collecting initial parameters affecting a
proposed system like the first round of TDM. The second step is to
assess each parameter by fuzzy terms in which each linguistic term
consists of three fuzzy values ðl; m; uÞ as shown in Fig. 3. For ex-
ample, an unimportant term would be (0.00, 0.25, 0.50), and an
important term would be (0.50, 0.75, 1.00). The third step applies
triangular fuzzy numbers to handle fuzziness of the experts’ opin-
ions where the minimum of the experts’ common consensuses as
where i = an individual expert; j = the cost parameter; Lij = mini-
mum of the experts’ common consensuses; M ij = average of the
experts’ common consensuses; U ij = maximum of the experts’
ð2Þ common consensuses; Lj = opinions mean of the minimum of
average of the experts’ common consensuses (Mij ); U j = opinions
mean of the maximum of the experts’ common consensuses (U ij );
W ij = fuzzy number of all experts’ opinions; n = number of experts;
and m = number of cost parameters.
The fourth step is using a simple center of gravity method to
defuzzify the fuzzy weight wj of each parameter to develop value
Sj by Eq. (7)
Lj þ Mj þ U j
Sj ¼ ð7Þ
3
where Sj = crisp number after defuzzification process. Finally, the
fifth step is that the experts provided a threshold to select or delete
the collected parameters as follows:
If Sj ≥ α; then the parameter should be selected
If Sj < α; then the parameter should be deleted
where α = defined threshold. The FDM can be summarized as the
following steps:
1. Identify all possible variables affecting a proposed system.
2. Assess evaluation score for each parameter by fuzzy terms.
3. Aggregate fuzzy numbers (W ij ).
4. Apply defuzzification (S).
5. Define a threshold (α).

Fuzzy Analytic Hierarchy Process

The analytic hierarchy process (AHP) is a decision-making ap-
µ(X) proach to evaluate and rank the priorities among different alter-
A natives and criteria (Saaty 1980; Vaidya and Kumar 2006). The
1.0 conventional AHP cannot deal with the vague or imprecise nature
of linguistic terms. Accordingly, Laarhoven and Pedrycz (1983)
combined fuzzy theory and AHP to develop FAHP. In the tradi-
tional FAHP method, the deterministic values of AHP could be
expressed by fuzzy values to apply uncertainty during making
decisions. The aim is to assess the most critical cost parameters
determined by FDM.
In FAHP, linguistic terms have been applied in pairwise com-
0.0 X
l m u parison that could be expressed by triangular fuzzy numbers
(Srichetta and Thurachon 2012; Erensal et al. 2006). Triple triangu-
Support Set of A lar fuzzy set numbers ðl; m; uÞ are used as fuzzy values where
l ≤ m ≤ u. Ma et al. (2010) and Elmousalami et al (2018a) have
Fig. 3. Triangular fuzzy number.
applied the following steps:

© ASCE 03119008-3 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 1. Identify criteria and construct the hierarchical structure. Optimization
2. Set up pairwise comparative matrices and transfer linguistic Filtering or Ranking Regression Regression
terms of positive triangular fuzzy numbers by a linguistic scale
of importance.
3. Generate group integration by Eq. (8).
4. Estimate the fuzzy weight. Selected Features Selected Features Selected Features

5. Defuzzify the triangular fuzzy number into a crisp number.

6. Rank defuzzified numbers. (a) (b) (c)
Experts’ opinions are used to construct the fuzzy pairwise com-
Fig. 4. (a) Filter; (b) Wrapper; and (c) Embedded algorithms.
parison matrix to construct a fuzzy judgment matrix. After collect-
ing the fuzzy judgment matrices from all experts using Eq. (8),
these matrices can be aggregated by using the fuzzy geometric
mean (Buckley 1985). The aggregated triangular fuzzy numbers of computation, and to provide high accuracy by removing noisy
(n) decision makers’ judgment in a certain case W ij ¼ ðLij ; M ij ; and redundant features. Moreover, Kohavi and John (1997) con-
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

U ij Þ where, for example, civil (C), mechanical works (M), and cluded that too many features higher than optimal diminishes
electrical works (E) refer to three different criteria, respectively the model performance. As shown in Fig. 4, a feature selection al-
(Elmousalami et al. 2018a) gorithm can be categorized into three main categories, filter, wrap-
per, and embedded, according to their selection manners (Guyon
Y
n 1 Yn 1 Y
n 1  and Elisseeff 2003). Filter algorithms rank features by evaluating
n n n
W ijn ¼ lijn ; mijn ; uijn ð8Þ correlation with independent variables such as Boruta feature
n¼1 n¼1 n¼1 analysis and principal component analysis. They do not optimize
the prediction accuracy. Therefore, filter algorithms are often more
where i = a criterion such as C, M, or E; j = screened cost param- computationally efficient than wrapper algorithms. However, the
eter for a defined case study; n = number of experts; Lij = minimum key limitation of the filter algorithms is not taking into consider-
of the experts’ common consensuses; M ij = average of the experts’ ation the effects of the selected features. Wrapper algorithms
common consensuses; U ij = maximum of the experts’ common depend on an inductive algorithm for sample regression in an iter-
consensuses; Lj = opinions mean of the minimum of the experts’ ative manner. Wrapper algorithms include optimization based on
common consensuses (Lij ); M j = opinions mean of the average of genetic algorithm (GA) or greedy approaches such as forward se-
the experts’ common consensuses (M ij ); U j = opinions mean of the lection and backward elimination. However, wrapper algorithms
maximum of the experts’ common consensuses (U ij ); and W ijn = can be prone to overfitting and high-cost computation. The em-
aggregated triangular fuzzy numbers of the nth expert’s view. bedded approach is a hybrid approach between filter algorithms
Based on the aggregated pairwise comparison matrix, the value and wrapper algorithms. The embedded approach depends on cer-
of fuzzy synthetic extent Si with respect to the ith criterion can be tain types of an inductive algorithm such as SVM-based recursive
computed by Eq. (9) by algebraic operations on triangular fuzzy feature elimination (RFE) (Lin et al. 2012).
numbers (Saaty 1994; Srichetta and Thurachon 2012) The objective of variables identification is to increase the model
X m Xn Xm −1 prediction accuracy and provide a better understanding of collected
Si ¼ W ij × W ij ð9Þ data (Guyon and Elisseeff 2003). Accurate cost drivers identification
j¼1 i¼1 j¼1 leads to the optimal performance of the developed the cost model.
Quantitative methods depend on the collected data such as factor
where i = a criterion; j = screened parameter; W ij = aggregated analysis, regression methods, and correlation methods in which ma-
triangular fuzzy numbers of the nth expert’s view; and Si = value chine learning models can be applied and conducted to figure out
of fuzzy synthetic extent. Based on the fuzzy synthetic extent val- patterns and relations in the collected data. Therefore, quantitative
ues, this study used Chang’s method (Saaty 1980) to determine techniques can automatically identify the key cost drivers.
the degree of possibility by Eq. (10). Accordingly, the degree of
possibility can assess and evaluate the system alternatives
8 9 Factor Analysis
>
> 1; if mm ≥ mc >>
>
< >
= Factor analysis (FA) is a machine learning method to cluster correlated
0; if lm ≥ uc variables to a lower number of factors that is used to filter data and
VðSm ≥ Sc Þ ¼ ð10Þ
>
> lc − um >
> determine key parameters. Many types of factoring exist such as prin-
>
: ; Otherwise ; >
ðlm − uc Þ − ðlm − uc Þ cipal component analysis (PCA), canonical factor analysis, and image
factoring (Polit and Beck 2012). The advantage of exploratory factor
where VðSm ≥ Sc Þ = degree of possibility between (C) criterion analysis (EFA) is to combine two or more variables into a single factor
and (M) criterion; ðlc ; mc ; uc Þ = fuzzy synthetic extent of (C) that reduces the number of variables. However, factor analysis cannot
criterion; and ðlm ; mm ; um Þ = fuzzy synthetic extent of (M) provide results’ causality to interpret the factored data.
criterion. EFA is conducted by PCA to reduce the number of variables,
as well as to understand the structure of a set of variables (Field
2009). The following questions should be answered before con-
Key Cost Drivers Identification by Quantitative ducting EFA:
Procedures 1. How large does the sample need to be?
2. Is there multicollinearity or singularity?
One of the key challenges of predictive modeling is the high dimen- 3. What is the method of data extraction?
sionality of data with small data size. Therefore, the dimensionality 4. What is the number of factors to retain?
reduction technique should be equipped in the prediction model 5. What is the method of factor rotation?
to simplify the model to a decision maker, to enable faster 6. Should factor analysis or principal component analysis be used?

© ASCE 03119008-4 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 1. Survey of sample size for FA
Reference
Nunnally (1978)
Kass and Tinsley (1979)
Tabachnick and Fidell (2007)
Comrey and Lee (1992)
Guadagnoli and Velicer (1988)
MacCallum et al. (1999)
Kaiser (1970, 1974), and
Hutcheson and Sofroniou
(1999)
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
Summary of findings
Sample size is 10 times the number of variables.
Sample size is between 5 and 10 cases per variable.
Sample size is at least 300 cases. 50 observations = very poor; 100 = poor; 200 = fair; 300 = good; 500 = very good;
1,000 or more = excellent.
Sample size can be classified to 300 as a good sample size, 100 as poor, and 1,000 as excellent.
A minimum sample size of 100–200 observations.
The minimum sample size depends on the design of the study where a sample of 300 or more would probably provide a
stable factor solution.
Based on the Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy values (Kaiser 1970), the values greater than
0.5 are barely acceptable (values below this should lead you to either collect more data or rethink which variables
to include). Moreover, values between 0.5 and 0.7 are mediocre, values between 0.7 and 0.8 are good, values between
0.8 and 0.9 are great, and values above 0.9 are superb.

Kline (1999) The absolute minimum sample size required is 100 cases.

Sample Size
Factors obtained from small data sets cannot generalize as well as
those derived from larger samples. Some researchers have sug-
gested using the ratio of sample size to the number of variables.
As illustrated in Table 1, this ratio can be 10 times the number of
all variables (Nunnally 1978) or between 5 and 10 cases per each
variable (Kass and Tinsley 1979). For example, if the number of
variables is ten variables, the sample size will be at least 100 ob-
servations (Nunnally 1978) or between 50 and 100 observations
(Kass and Tinsley 1979).
Multicollinearity or Singularity
The first step is to check the correlation among variables and avoid
multicollinearity and singularity (Tabachnick and Fidell 2007;
Hays 1983). Multicollinearity means variables are correlated too
highly, whereas singularity means variables are perfectly corre-
lated. It is used to describe variables that are perfectly correlated
(it means the correlation coefficient is 1 or −1). There are two
methods for assessing multicollinearity or singularity. The first
method depends on the correlation matrix scanning, whereas the
second methods depends on the determinant of the correlation
matrix. The first method is conducted by scanning the correlation
matrix among all independent variables to eliminate variables with
correlation coefficients greater than 0.90 (Field 2009; Hays 1983)
or correlation coefficients greater than 0.80 (Rockwell 1975). The
second method is to scan the determinant of the correlation matrix.
Multicollinearity or singularity may be in existence if the determi-
nant of the correlation matrix is less than 0.00001. One simple heu-
ristic is that the determinant of the correlation matrix should be
greater than 0.00001 (Field 2009; Hays 1983). If the visual inspec-
tion reveals no substantial number of correlations greater than 0.3,
PCA probably is not appropriate. Also, any variables that corre-
late with no others (R ¼ 0) should be eliminated (Field 2009;
Hays 1983).
Bartlett’s test can be used to test the adequacy of the correlation
matrix. It tests the null hypothesis that the correlation matrix is an
identity matrix where all the diagonal values are equal to 1 and
all off-diagonal values are equal to 0. A significant test indicates
that the correlation matrix is not an identity matrix with a signifi-
cance value less than 0.05 and the null hypothesis can be rejected
(Dziuban and Shirkey 1974).
According to factor extraction, factor (component) extraction is
conducting EFA to determine the smallest number of components
that can be used to represent interrelations among a set of variables
(Tabachnick and Fidell 2007). Factors can be retained based on
eigenvalues where a scree plot diagram can be developed to retain
factors (Cattell 1966). All factors that have eigenvalues greater than
1 can be retained (Kaiser 1960). On the other hand, Jolliffe (1972,
1986) recommended retaining factors that have eigenvalues more
than 0.7.
According to factor rotation, two types of rotation exist: orthog-
onal rotation and oblique rotation (Field 2009). Orthogonal rotation
can be varimax, quartimax, and equamax, whereas oblique rota-
tion can be direct oblimin and promax. Accordingly, the resulting
outputs depend on the selected rotation method. For a first analysis,
the varimax rotation should be selected to easily interpret the fac-
tors, and this method can generally be conducted. The objective of
the varimax is to maximize the loadings dispersion within factors
and to load a smaller number of clusters (Field 2009). Stevens
(2002) concludes that no difference between factor analysis and
component analysis exists if there are 30 or more variables and
communalities greater than 0.7 for all variables. On the other hand,
a difference between factor analysis and component analysis exists
if the variables are fewer than 20 variables and there are low
communalities (<0.4).
Regression Methods for Key Cost Drivers Selection
Regression analysis can be used for both cost drivers selection and
cost prediction modeling (Ratner 2010). Regression models can
learn from the given data by adjusting the regression parameters
to map a mathematical relationship based on the given data. The
current study focused on cost drivers selection. Therefore, the for-
ward, backward, and stepwise methods are reviewed as follows.
Forward selection initiates with no variables in the model, where
each added variable is tested by a comparison criterion to improve
the model performance (Wilkinson and Dallal 1981). If the inde-
pendent variable significantly improves the ability of the model to
predict the dependent variable, then this predictor is retained in the
model and the method searches for a second independent variable
(Field 2009; Draper and Smith 1998).
Backward selection is the opposite of the forward method.
In this method, all input independent variables are initially selected,
and then the most unimportant independent variables are eliminated
one-by-one based on the significance value of the t-test for each
variable. The contribution of the remaining variable is then reas-
sessed (Field 2009; Draper and Smith 1998).
Stepwise selection is an extension of the forward selection ap-
proach, in which input variables may be removed at any subsequent
iteration (Field 2009; Draper and Smith 1998). Despite forward se-
lection, stepwise selection tests at each step for variables to be

© ASCE 03119008-5 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 2. Genetic algorithm (GA) processes
Process Element Binary coding
Crossover Chromosome A 1011001
Chromosome B 1111111
Offspring 1 1011111
Offspring 2 1111001
Mutation Offspring 1 1011111
Offspring 2 1111011
included or excluded where stepwise is a combination of backward
and forward methods (Flom and Cassell 2007).
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
Correlation Method
The relations among all variables are shown in the correlation ma-
trix; the aim is to screen variables based only on the correlation
matrix. Therefore, all independent variables that are highly corre-
lated with each other are eliminated (R >¼ 0.8), and all de-
pendent variables that are weakly correlated with the dependent
variable (R <¼ 0.3) are eliminated. Such an approach is dependent
on the hypothesis that the relevant input independent variable is
highly correlated with the output dependent variables and less
correlated with the other input independent variables in the input
subset (Ozdemir et al. 2001).
Pearson correlation is a measure of the linear correlation be-
tween two variables, giving a value between þ1 and −1, where 1,
0, and −1 mean positive correlation, no correlation, and negative
correlation, respectively. It was developed by Karl Pearson as a
measure of the degree of linear dependence between two variables
(Field 2009). On the other hand, Spearman correlation is a nonpara-
metric measure of statistical dependence between two variables
using a monotonic function. A perfect Spearman correlation of þ1
or −1 occurs when each of the variables is a perfect monotone func-
tion of the other (Field 2009).
Genetic Algorithm for Key Cost Drivers Selection
Evolutionary computing (EC) is a set of natural selection–inspired
methods based on evolutionary theory (Darwin 1859), such as the
genetic algorithm (GA) (Holland 1975). The genetic algorithm
(GA) is an evolutionary algorithm (EA) used for search and opti-
mization based on a fitness function (Siddique and Adeli 2013).
GA can be applied to select the input parameters of a prediction
machine learning model such as an artificial neural network
(ANN). All irrelevant, redundant, and useless parameters can be
removed to reduce the size of the ANN. The chromosome can
be represented in a binary-coded system in which the bits num-
ber in the chromosome string equals the input variables number.
This approach was proposed by Kohavi and John (1997) and by
T
X1 1 0 0
X2 1 0 1 Si
P=
Xm 0 1 1
Feature selection
Fig. 5. GA for cost driver identification.
© ASCE
J. Constr. Eng. Manage., 2020, 146(1): 03119008
Siedlecki and Sklansky (1988, 1989) and is called the wrapper
approach to screen input variables (features).
The genetic information can be represented as chromosomes;
it is a powerful tool for optimization and search problems. Chromo-
some representation is the first design step of the EC, in which a
chromosome is the possible candidate solution for searching and
optimization purposes. The gene is the functional unit of inherit-
ance, and any chromosome is expressed by a number of genes.
A chromosome can be represented as a vector (C) consisting of (n)
genes denoted by (cn ) as follows: C ¼ fc1 ; c2 ; c3 : : : cn g. Each
chromosome (C) represents a point in the n-dimensional search
space.
The first task in chromosome construction is to encode the
genetic information. Binary coding is commonly one of the most
used chromosome representations in which (bi ) is a binary value
equal to zero or one and X is a search space of chromosomes as
follows:
X ¼ fðb1 ; b2 : : : bl Þðb1 ; b2 ; : : : ; bl Þ; : : : g; bi ∈ f0; 1g
A fitness function (FCi ) is a problem-dependent function that
guides the search model to converge and get the optimal solution.
FCi is applied to evaluate the fitness of each chromosome to select
the best subset of chromosomes for crossover and mutation proc-
esses. The crossover process is intended to produce new genera-
tions and solutions, whereas the mutation process is intended to
randomly escape from converging to a local minimum solution.
As illustrated in the following example, where two chromosomes
A and B are composed of seven genes, Offspring 1 and Offspring 2
are produced by a crossover of the first chromosome and the second
chromosome where the one-point crossover is applied at the third
generation of the chromosomes. The next process is mutation,
in which the sixth gene of the second offspring is mutated to
value one. An example is given in Table 2.
Relative fitness is the fitness function for each chromosome in
which relative fitness is the criterion to select the next generation
of chromosomes. Genetic operators are selection processes of the
fittest chromosomes and then conducting crossover and mutation
processes subsequently. Many selection approaches exist such as
random selection, proportional selection (roulette wheel selection),
tournament selection, and rank-based selection. Five main steps are
required to develop an optimization problem by EC:
1. chromosome representation,
2. an initial population representation,
3. definition of the fitness function as a chromosome selection
criterion, and
4. EA parameter values determination such as population size and
the maximum number of generations.
As shown in Fig. 5, data can be screened to key cost drivers in
which each chromosome represents a possible solution for input
P Yfinal
EA ANNs
F
system
03119008-6 J. Constr. Eng. Manage.
 parameters. The chromosome consists of a binary gene for which
one represents the existence of a parameter and zero represents the
absence of the parameter. Each gene in the chromosome is asso-
ciated with an input feature for which the value of one represents
the input feature existence and the value of zero represents the
elimination of this variable. Thus, the number of ones in a chromo-
some is the number of the screened variables by GA. Chromosomes
build a population of a set of possible solutions (Si ). The objective
of EA is to select the best subset of parameters (P) based on fitness
function (F) that inherently minimizes the total system error. The
fitness function is minimizing ANNs’ prediction error. The main
disadvantage of this methodology is a high computational effort
(Siddique and Adeli 2013).
The fitness function is the guide to EA to converge, and wrong
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
fitness function formulation means false search and inaccurate
optimization. The fitness function can be formulated in terms of
a minimum number of selected ANNs features, maximum accu-
racy, and minimized computational cost (Yang and Honavar 1998;
Ozdemir et al. 2001).
Boruta Feature Selection Algorithm
Cao et al. (2018) have conducted a Boruta feature analysis to rank
57 related variables and filter them to 20 features. The Borura
algorithm is an ensemble based on multiple decision trees voting
for which the computed votes are applied to rank the importance
of the variables. Boruta feature selection algorithm selects the
greatest potential for increasing the prediction accuracy (Kursa and
Rudnicki 2010). Unimportant features are identified based on a
two-tailed hypothesis. The procedure of the Boruta feature selec-
tion algorithm is as follows:
1. Boruta feature selection algorithm identifies shadow variables
that have no correlation with the target variable.
2. Each explanatory variable has shadow variables in which a new
data set is formulated based on both the original explanatory
variable and the shadow variables.
3. A random forest algorithm is conducted for fitting the resulting
data set.
4. The z-score of each feature (the original and the shadow) are
computed in which the z-score is the average loss in accuracy
divided by the standard deviation.
5. The original attributes that have higher z-scores than the shadow
attributes are retained as important features of the given data set.
6. Steps 1–5 are repeated iteratively several times until all original
attributes that are significantly higher than the maximum the
shadow variables are retained.
7. Features that are significantly below the maximum z-score of
the shadow variables are recorded as unimportant.
Review Survey and Discussion Key Cost Drivers
Identification
Past literature has been surveyed to identify the practices for
cost drivers identification in the construction industry. Table 3
presents 37 papers to review the key cost drivers techniques;
most papers were published from 2002 to 2018. For example,
Petroutsatou et al. (2012) has applied a questionnaire survey to
determine the basic parameters affecting the final construction cost.
The variables were then classified into independent variables and
dependent variables. Accordingly, the data have been collected
and analyzed for ANN model development. Similarly, ElSawy et al.
(2011) have conducted a questionnaire survey to select the 10 most
influential key cost drivers out of 52 factors of building con-
struction in Egypt based on the experts’ opinions. Consequently,
© ASCE 03119008-7
J. Constr. Eng. Manage., 2020, 146(1): 03119008
ElSawy et al. (2011) have used the selected 10 factors to develop
an ANN in which the selected 10 factors represent the 10 neurons
in the first layer of the ANN model. Table 3 provides a structured
review of 37 papers that can be used effectively in future research in
which the reviewed paper can be categorized by the method used
and its year of publication. In addition, the tables summarize the
paper contributions in a high-level description. Based on Table 3,
the questionnaire survey approach is the most common approach
conducted to identify and assess the cost drivers of a certain case
study. Therefore, the qualitative approach is more common than
quantitative methods. Such a claim can be a result of unavailability
of data for the studied cases. Moreover, asking experts is a simple
approach for which no deep statistical knowledge is needed. On the
other hand, the data-driven procedures require understanding and
conducting statistical methods.
As shown in Fig. 6, based on the literature survey, FAHP is the
most commonly used technique. However, traditional techniques
have many advantages such as identifying divergence of opinions
among participants and share of knowledge and reasoning among
participants. In addition, rounds enable participants to review, re-
evaluate, and revise all their previous opinions. Moreover, these
methods require simple calculations. However, the major disad-
vantage of the traditional techniques is their inability to maintain
uncertainty among different participants’ opinions. Accordingly,
information extracted from a selected group of experts may not
be representative. Alternatively, the advanced techniques such as
FDM and FAHP have applied fuzzy methods to maintain uncer-
tainty among participants’ opinions, and this feature is the major
advantage of the advanced techniques. However, the advanced
techniques require more calculations and statistical forms to be
conducted. Generally, qualitative techniques such as the Delphi
method, FDM, and FAHP are time-consuming because they require
collecting experts’ opinions, interviews, and meetings. In contrast,
quantitative techniques need only the collected data to automati-
cally explore the key cost drivers.
On the one hand, the most common qualitative methods are
questionnaire surveys and FAHP. On the other hand, the most
common quantitative methods are regression methods and factor
analysis. The key advantage of the quantitative methods is that they
depend on machine learning models that can learn from iterations
and automatically identify accurate key cost drivers based on the
given data. The key disadvantage of the quantitative methods is
related to data availability, which controls the applicability of
machine learning.
Cost Estimating Methods
According to estimating methods, top-down and bottom-up ap-
proaches are the main two approaches to the cost estimate. On the
one hand, the top-down approach occurs in the conceptual phase
and depends on the historical cost data, from which the data from
similar projects are retrieved to estimate the current project. On the
other hand, the bottom-up approach requires detailed information
about the studied project. First, all projects are divided into items to
create a cost breakdown structure (CBS). The main items of a CBS
are dependent on the number of resources (labor, equipment,
materials, and subcontractors). The next step is to calculate the
cost of each broken item and sum up the total construction cost
(AACE 2004).
Cost estimating methods have been classified into four types
(Dell’Isola 2002): single-unit rate methods, parametric cost mod-
eling, elemental cost analysis, and quantity survey.
J. Constr. Eng. Manage.
 Table 3. Review of cost drivers identification
Reference Method Key findings
ElSawy et al. (2011) Questionnaire survey Based on 52 factors of building construction in Egypt, 10 cost drivers have been selected by a
questionnaire survey of experts for an ANN cost model.
Emsley et al. (2002) Questionnaire survey þ factor Based on 300 building projects, FL is investigated to select the key cost drivers to be used by
analysis ANNs and regression models.
El-Sawah and Trial and error approach Based on trial and error approach and combination of input variables, ANN models have been
Moselhi (2014) built for cost prediction of steel building and timber bridge.
Moselhi and Questionnaire survey A questionnaire survey has been conducted to discover the input variables for ANNs for
Hegazy (1993) markup estimation.
Attalla and Hegazy Questionnaire survey A questionnaire survey has been operated to identify the input variables for ANNs for cost
(2003) deviation; 36 factors have been identified.
El-Sawalhi and Questionnaire survey Eighty questionnaires have been conducted to determine significant variables for cost
Shehatto (2014) prediction for the building project.
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

Petroutsatou et al. Questionnaire survey A questionnaire survey has been conducted to determine significant parameters for an ANN
(2012) cost prediction model for tunnel construction in Greece.
Williams (2002) Regression methods Based on bidding data, the stepwise regression method has been utilized to check the
significance of each parameter and select the key cost drivers for the regression model.
El Sawalhi (2012) Questionnaire survey Both a questionnaire survey and relative index ranking technique have been conducted
to investigate and rank the factors affecting the cost of building construction for the fuzzy
logic model.
Knight and Fayek Questionnaire survey Based on past related literature and interview surveys, all parameters affecting cost overruns
(2002) for building projects have been identified and ranked for the fuzzy logic model.
Choi et al (2014) Questionnaire survey Based on a questionnaire survey, attributes of the road construction project have been
identified.
Alroomi et al. Questionnaire survey þ factor Based on 228 completed questionnaires, all relevant cost data of competencies have been
(2012) analysis collected by experts, whereas the factor analysis has been conducted to investigate the
correlation effects of the estimating competencies.
Kim (2013) Questionnaire survey þ factor Based on a questionnaire survey and factor analysis, all parameters affecting best practices of
analysis infrastructure projects have been identified and ranked.
Manoliadis et al. FAHP Based on qualifications survey, FDM is conducted to assess bidders’ suitability for improving
(2009) bidder selection.
Pan (2008) FAHP Fuzzy AHP is conducted to provide the vagueness and uncertainty for selecting a
bridge construction method. Therefore, FAHP obtains more reliable results than the
conventional AHP.
Marzouk and Questionnaire survey A questionnaire survey has been conducted to identify and evaluate 14 parameters affecting
Ahmed (2011) the construction costs of pump station projects.
Liu (2013) FDM þ FAHP Both FDM and FAHP were conducted to evaluate and filter all factors affecting indicators of
managerial competence.
Saaty (2008) AHP For a countless application, AHP has been conducted as a powerful decision-making
procedure among different criteria and alternatives.
Laarhoven and FAHP AHP and fuzzy theory were combined to produce FAHP in which the objective was to evaluate
Pedrycz (1983) the most important cost parameters.
Ma et al. (2010) FAHP FAHP was conducted for pile-type selection based on the collected field factors where the
fuzzy AHP approach produces an efficient performance for pile-type selection.
Erensal et al. (2006) FAHP FAHP was conducted for evaluating key parameters in technology management.
Srichetta and FAHP FAHP was conducted for evaluating notebook computer products.
Thurachon (2012)
Hsu et al. (2010) FDM þ FAHP This study utilized two processes of selection and decision making. FDM is the first process to
identify the most important factors, whereas the second process is FAHP to identify the
importance of each factor.
Marzouk and Elkadi Questionnaire survey þ FA EFA was conducted to select the cost drivers of water treatment plants in which a total of
(2016) 33 variables were reduced to four components. Such components are used as inputs to the
ANN model.
Woldesenbet and FA Based on the collected roadway project data, factor analysis of a covariance and correlation
Jeong (2012) matrix have been investigated to identify critical factors of the project.
Park and Kwon Questionnaire survey þ FA Both questionnaire and FA have been investigated to discover the critical success factors for
(2011) infrastructure projects in Korea.
Akintoye (2000) FA Seven factors out of 24 factors influencing contractors’ cost estimating have been selected
by FA.
Stoy et al. (2012) Regression method Backward regression method has been computed to determine key cost drivers based on a total
of 75 residential projects.
Lowe et al. (2006) Regression method Based on 286 sets of data collected in the United Kingdom, both forward and backward
stepwise regression have been used to develop six parametric cost models.
Yang (2005) Correlation method Correlation matrix should be scanned to reduce variable and to detect redundant variables.
Ranasinghe (2000) Correlation method This study presents induced correlation concept to analyze input cost variables for residential
building projects in Germany.

© ASCE 03119008-8 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 3. (Continued.)
Reference Method Key findings
Stoy et al. (2008); Questionnaire survey þ regression Based on 70 residential properties in Germany, Stoy et al. used a regression method to select
Stoy and Schalcher method cost drivers.
(2007)
Kim et al. (2005) Gas for parameter selection This study has built three cost ANN models by back-propagation (BP) algorithm, GA for
optimizing ANNs weights, and GA for parameters optimization of the BP algorithm.
Optimizing parameters of the BP algorithm produces better results.
Xu et al. (2015) GA þ correlation method A correlation method was used to rank model features, whereas GA was used for selecting the
optimal subset of features for the model.
Elmousalami et al. FDM þ FAHP þ traditional This study has compared the traditional Delphi method, FDM, and FAHP to evaluate and
(2018a) Delphi method select the key cost drivers of field canals improvement projects in Egypt.
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

FDM +
FAHP
8%
Questionnaire
survey +
Literature survey
FAHP 24%
24%
Trial and error
approach
4%
AHP
4%
GA Regression
8% methods Factor analysis
12% 8%
Correlation method
8%
Fig. 6. Cost drivers identification methods.
Single-unit rate methods calculate the total cost of the project
based on a unit such as the area of the building or an accommo-
dation method such as cost per bed for hotels or hospitals. Para-
metric cost modeling is to develop a model based on statistical
relations of the key parameters extracted by conducting qualitative
techniques (Elmousalami et al. 2018a) or statistical analyses such
as regression models, ANNs, and FL models. Elemental cost analy-
sis is dividing the project into the main elements and estimating the
cost of each element based on historical data. A quantity survey is a
detailed cost estimate based on quantities surveyed and contract
unit costs rates in which such quantities include the resources used
such as materials, labor, and equipment for each activity. Therefore,
estimators usually apply single-unit rate method or parametric cost
estimate at the conceptual stage where no detailed information of
the project is available.
The estimating process consists of six main elements (Kan
2002): project information, historical data, current data, estimating
methodology, cost estimator, and estimates. Project information in-
cludes the project characteristics that can be used as inputs to the
cost model. Historical data are the collected data of the previous
projects to statistically develop the cost model. Current data are
the data extracted from the project information such as unit cost
rates of material, labor, and equipment. Estimating methodology
is the method used for cost estimate such as parametric cost model.
The cost estimator is the user who uses the cost model and enters
the input parameters or data to obtain the cost estimate. The esti-
mates are the outputs of the cost model.
Computational Intelligence
Computational intelligence (CI) techniques are aspects of human
knowledge and make computations adaptively to become more
vigorous in system modeling than classical mathematical modeling
(Bezdek 1994). Based on CI, an intelligent system can be devel-
oped to produce consequent outputs and actions depending on
the observed inputs and outputs of the system (Siddique and
Adeli 2013).
The objective is to solve complex real-world problems based on
data analytics such as classification, regression, prediction, and
optimization in an uncertain environment. The core advantage of
the intelligent systems is their human-like capability to make de-
cisions that depend on information with uncertainty and impreci-
sion. The basic approaches to computational intelligence are fuzzy
logic (FL), artificial neural networks (ANNs), and evolutionary
computing (EC). Accordingly, CI is a combination of FL, neuro-
computing, and EC (Engelbrecht 2002). The scope of this study
focuses on the three methodologies of computational intelligence—
FL, ANNs, and EC—and their fusion.
Multiple Regression Analysis
Multiple regression analysis (MRA) is a statistical analysis that
uses given data for prediction applications. Based on historical
cases, regression analysis develops a mathematical form to fit the
given data (Field 2009; Walker 1989). This mathematical form can
be formulated as Eq. (11)
Y ¼ B0 þ B 1 X 1 þ B 2 X 2 þ : : : : : : : : : B n X n ð11Þ
where Y = dependent variable; B0 = constant; Bi = variable coef-
ficient; and X i = independent variables. The change by one unit of
the independent variable X 1 causes a change by B1 in the dependent
variable Y. Similarly, the change by one unit of the independent
variable X 2 causes a change by B2 in the dependent variable Y.
In addition, the sign of B1 and B2 determines the decrease or in-
crease in the dependent variable Y. The objective of the regression
model is to mathematically represent data with minimal prediction
error. Therefore, regression analysis is applied in cost estimate
modeling to represent the cost-estimate relationships where the
cost prediction is represented as the dependent variable and the cost
drivers are represented as the independent variables.
Multiple regression analysis includes a model called polynomial
regression. Polynomial regression regresses a dependent variable

© ASCE 03119008-9 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 based on powers of the independent variables. Mathematically, the
polynomial regression model can be formulated as Eq. (12)
Y ¼ B0 þ B1 X i þ B2 X 2i þ : : : Bn X Ki þ ei ; for i ¼ 1; 2; : : : ; n
ð12Þ
where K = degree of the polynomial.
According to the sample size, (50 þ 8k) may be the minimum
sample size, where k is the number of predictors (Green 1991).
According to deleting outliers, Cook’s distance detects the impact
of a certain case on the regression model (Cook and Weisberg
1982). If the Cook’s values are <1, there is no need to delete that
case (Stevens 2002). Otherwise, if the Cook’s values are >1, there
is a need to delete that case. Variables are highly correlated where
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
the coefficient of determination is higher than 0.8 (R2 > 0.8)
(Rockwell 1975). The variance inflation factor (VIF) examines
the linear relationship with the other variable (Field 2009), and if
the average VIF is greater than 1, then multicollinearity occurs and
can be detected (Bowerman and O’Connell 1990; Myers 1990).
Homoskedasticity occurs when the residual terms vary constantly
and the residual variance should be constant to avoid a biased re-
gression model (Field 2009). The Durbin–Watson test is conducted
to check the correlations among errors where the test values range
between 0 and 4. The value of two denotes that residuals are un-
correlated (Durbin and Watson 1951). Accordingly, regression
models can be summarized in the following steps:
1. Collect and prepare the historical cases.
2. Divide the collected cases into a training set and a validation set.
3. Check the sample size of the collected training data (Green
1991; Stevens 2002).
4. Define key independent parameters (cost drivers) and dependent
parameter (cost variable).
5. Develop a regression model and check the significance
(P-value) of each coefficient (Field 2009).
6. Check outliers (Cook and Weisberg 1982).
7. Check the variance inflation factor (VIF) (Bowerman and
O’Connell 1990; Myers 1990).
8. Check homoskedasticity (Durbin and Watson 1951).
9. Calculate the resulting error such as mean absolute percentage
error (MAPE).
Fuzzy Logic
Fuzzy logic (FL) is the modeling of human decision making by
representing uncertainty, incompleteness, and randomness of the
real-world system (Zadeh 1965, 1973). In addition, FL represents
the experts’ experience and knowledge by developing fuzzy rules.
Such knowledge is represented in fuzzy systems by membership
functions (MFs), which range from zero to one. MFs can be tri-
angular, trapezoidal, Gaussian, and bell-shaped functions in which
the selection of the MF function is problem-dependent. Fig. 7
illustrates a trapezoidal MF that consists of a core set fa2 ; a3 g and
a support set fa1 ; a2 ; a3 ; a4 g. The shape of MF significantly influ-
ences the performance of a fuzzy model (Wang 1997; Chi et al.
1996). Therefore, many methods, such as clustering approaches
and genetic algorithms, are applied to develop MFs automatically
to select the optimal shape of MFs.
Once MFs can be identified for each dependent and independent
parameter, a set of operations on fuzzy sets can be conducted. Such
operations are the union of fuzzy sets, the intersection of fuzzy sets,
and complement of fuzzy set and α-cut of a fuzzy set. Linguistic
terms are used to approximately represent the system features
where such terms cannot be represented as quantitative terms
(Zadeh 1976). Once MFs and linguistic terms have been defined,
© ASCE 03119008-10
J. Constr. Eng. Manage., 2020, 146(1): 03119008
µ (X)
A B
1.0
Core Core
0.5
0.0 X
a1 a2 a3 a4 b3 b4
b1 b2
Support Set of A Cross point of A and B
Fig. 7. Fuzzy trapezoidal membership function (MF).
if-then rules can be developed to establish rule-based systems.
Each rule represents human logic and experience in which all rules
represent the brain of the fuzzy system. A single fuzzy if-then rule
can be represented by the following:
If hfuzzy proposition ðx is A1 Þi Then hfuzzy proposition ðy is B2 Þi
where x = input parameter; A1 = MF of x; y = output parameter; and
B2 = MF of y. Rule-based systems are systems that have more than
one rule to represent human logic and experience to the developed
system. Aggregation of rules is the process of developing the over-
all consequent from the individual consequents added by each rule
(Siddique and Adeli 2013).
As shown in Fig. 8, there are two parameters X 1 and X 2 , where
μ X 1 ¼ fa1 ; b1 ; c1 ; d1 g, μ X 2 ¼ fa2 ; b2 ; c2 ; d2 g, μ Y ¼ fay ; by ;
cy ; dy g and the fuzzy system consists of two rules as follows:
Rule 1: IF X 1 is a1 AND X 2 is c2 THEN y is ay
Rule 2: IF X 1 is b1 AND X 2 is d2 THEN y is by
where two inputs are used fX 1 ¼ 4; X 2 ¼ 6g. Such two inputs
intersect with the antecedents MF of the two rules where two con-
sequent rules are produced {R1 and R2 } based on minimum inter-
sections. The consequent rules are aggregated based on maximum
intersections where the final crisp value is 3. The aggregated output
for Ri rules is given by
Rule 1: μ R1 ¼ min½μ a1 ðX 1 Þ and μ c2 ðX 2 Þ
Rule 2: μ R2 ¼ min½μ b1 ðX 1 Þ and μ d2 ðX 2 Þ
Y: Fuzzification½max½R1 ; R2 
Fuzzification is transforming crisp values into fuzzy inputs.
Conversely, defuzzification is transforming a fuzzy quantity into a
crisp output. Many different methods of defuzzification exist such
as max-membership, center of gravity, weighted average, mean-
max, and center of sums (Runker 1997). An inference mechanism
is the process of converting input space to output space such as
Mamdani fuzzy inference, Sugeno fuzzy inference, and Tsukamoto
fuzzy inference (Mamdani and Assilian 1974; Takagi and Sugeno
1985; Sugeno and Kang 1988; Tsukamoto 1979).
Fuzzy modeling identification includes two phases: structure
identification and parameter identification (Emami et al. 1998).
Structure identification is to define input and output variables
J. Constr. Eng. Manage.
 X1 X2 Y

a1 ay
c2
X2R1 X2R1
X1R1 AND Rule 1
min X1R1
X1R1

one ten one ten one ten

b1 X1R2 d2 X1R2 by
Rule 2
X2R2 X2R2 min X2R2
AND

one ten one ten one ten

X1 = 4 fuzzification X2 = 6 Second: consequent rules
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

First: antecedent rules

one ten
YC = 3
Third: aggregation and defuzzification

Fig. 8. Fuzzy rules firing.

and to develop input and output relations through if-then rules. The
following points summarize the structure identification of fuzzy
system:
1. Determine relevant inputs and outputs.
2. Select fuzzy inference system, e.g., Mamdani, Sugeno, or
Tsukamoto.
3. Define the linguistic terms associated with each input and output
variable.
4. Develop a set of fuzzy if-then rules to represent the relation
between the inputs and outputs.
On the other hand, the parameters identification is an optimiza-
tion problem in which the objective is to maximize the performance
of the developed system. Defining MFs such as the shape of
MF (triangular, trapezoidal, Gaussian, and bell-shaped functions)
and its corresponding values can significantly optimize the system
performance.
Artificial Neural Networks
ANNs are biologically inspired models to mimic the human neural
system for information-processing and computation purposes.
ANN is a machine learning (ML) technique that can learn from
past data. Learning forms can be supervised, unsupervised, and
reinforcement learning. Contrary to traditional modeling tech-
niques such as linear regression analysis, ANN models have the
ability to approximate nonlinear functions to a specified accuracy.
The first model of artificial neural networks came in 1943 when
Warren McCulloch, a neurophysiologist, and Walter Pitts, a young
mathematician, outlined the first formal model of an elementary
computing neuron (McCulloch and Pitts 1943). The first model
of ANNs proposed by Warren McCulloch to mimic the human neu-
ral system, the model is based on the concept of electrical circuits,
in which the output is zero or one. This is called a perceptron or
neuron, and such a neuron is the unit of the ANN (McCulloch and
Pitts 1943). Hopfield connected these neurons and developed a
network to create ANNs (Hopfield 1982). Generally, ANNs can be
categorized into two main categories: feedforward networks and
recurrent networks.
In a feedforward network, all neurons are connected together.
The feedforward network consists of the input vector (x), a weight
matrix (W), a bias vector (b), and an output vector (Y) that can be
formulated as Eq. (13)
Y ¼ fðW · x þ bÞ ð13Þ
where fðÞ includes a nonlinear activation function. Different types
of activation functions exist such as linear function, step function,
ramp function, and tan sigmoid function. Selection of ANN param-
eters such as the number of neurons, connections transfer functions,
and hidden layers mainly depend on the ANN’s application.
Several types of feedforward neural network architectures exist
such as multilayer perceptron networks (MLP), radial basis
function networks, generalized regression neural networks, prob-
abilistic neural networks, belief networks, Hamming networks,
and stochastic networks in which each architecture is problem-
dependent (Siddique and Adeli 2013). In this study, multilayer
perceptron networks (MLP) are explained in some detail.
As shown in Fig. 9, an MLP network is a network with several
layers of perceptrons in which each layer has a weight matrix (W),
a bias vector (b), and an output vector (Y). The input vector X ¼
fX 1 ; X 2 ; X 3 ; : : : g feeds forward to n neurons in the hidden layer
with a transfer function fðÞ where weights w ¼ fw1 ; w2 ; w3 ; : : : g
combined to produce the output. The outputs of each layer are
computed as Y kn ¼ fðW n;m;k · X m þ bi;k Þ, where k is the number
of layers, d is the number of inputs, i is the number of bias nodes,
n is the number of neurons, m is the number of weight for each
sending neuron, and fðÞ is the activation function (e.g., sigmoid
and tan sigmoid functions). No exact rule exists to determine the
number of hidden layers and neurons in the hidden layer. Huang
and Huang (1991) and Choi et al. (2001) stated that one hidden
layer MLP needs at least (P − 1) hidden neurons to classify P
patterns. A standard rectified linear unit (ReLU) is an activation
function that can enhance the computing performance of ANNs
(Nair and Hinton 2010). Mathematically, ReLU is defined as
follows:

© ASCE 03119008-11 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 neuron
first hidden layer second hidden layer output layer
input layer w1,1,k
w1,1,k N net11 Y11 net21 Y21
11 f() N21 f()
X1 w1,1,k
b i,k b i,k
net12 Y12 net 22 Y22 net31 Yfinal
X2 N12 f() N22 f() f()

Xd b i,k wn,m,kb i,k wn,m,k

net kn Ykn netkn Ykn
wn,m,k Nkn f() Nkn f()
b i,k b i,k

Fig. 9. Multilayer perceptron network (MLP).

Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.


Xi ; if X i ≥ 0 yi ðW × X i þ bÞ ≥ 0 − ξ; i ¼ 1; 2; 3; : : : m ð15Þ
A¼
0; Xi < 0
Accordingly, the objective function for SVM optimization is as
For big data and high dimensionality of data, deep neural net- expressed in Eq. (16)
works (DNNs) can be conducted with a ReLU activation function
for high performance computing (LeCun et al. 2015). X
i¼m
1 X
i¼m

A three-layer network (input layer, hidden layer, and output
layer) can solve a wide range of prediction, approximation, and
classification problems. Moreover, to avoid overfitting problems
and enhance the generalization capability, the number of training
cases should be more than the size of the network (Rutkowski
2005). The learning mechanism of ANNs is modifying the weights
and biases of the network to minimize the in-sample error. Devel-
oping ANNs can be summarized in the following steps:
1. Collect and prepare the historical cases.
2. Divide the collected cases into a training set and a valida-
tion set.
3. Determine relevant inputs and outputs.
4. Select the number of hidden layers.
5. Select the number of neurons in each hidden layer.
6. Select the transfer function.
7. Set initial weights.
8. Select the learning algorithm to develop the ANNs’ weights.
9. Train the model for several iterations to get minimal predic-
tion error.
10. Calculate the resulting error such as mean absolute percentage
error (MAPE).
Support Vector Machines
A support vector machine (SVM) is a nonparametric supervised
ML algorithm that can be applied for regression and classification
problems (Vapnik 1979). The objective is to minimize misclassifi-
cations cases through optimizing the margin and hyperplanes
distance as shown in Fig. 10.
Slack variables are added to solve the inseparability problem
(Cortes and Vapnik 1995). According to the linear separable case,
the objective is to maximize the hyperplane distance between the
two class boundaries
 main processes of CBR are retrieving, reusing, revising, and retain-
W × Xi þ b ≥ 1; if yi ≥ 0
Linear SVM ¼ ð14Þ
W × X i þ b < −1; if yi < 0
for i ¼ 1; 2; 3; : : : ; m.
Eq. (14) can be generalized as yi ðW × X i þ bÞ ≥ 0. For opti-
mum separation, the distance between Pi¼m the two marginal hyper-
1
planes should be minimized as i¼0 2 w × wT . For nonlinear
data, a positive slack variable (ξ) is added to handle the nonlinearity
of the data as shown in Eq. (15)
Min w × wT þ C ξi ð16Þ
i¼0
2 i¼0
Decision Trees
Decision trees (DT) is a supervised ML model that divides the
given data into hierarchical rules on each tree node by a repetitive
splitting algorithm. Three of the most commonly applied algo-
rithms for decision trees modeling are the chi-square automatic in-
teraction detector (CHAID), classification and regression trees
(CART), and C4.5/C5.0 algorithms; CHAID is for categorical var-
iables, and CART and C4.5/C5.0 are for both continuous variables
and categorical variables (Berry and Linoff 1997; Breiman et al.
1984). CART is a tree learning model that can be applied for both
regression (continuous variables) and classification (categorical
variables) applications (Quinlan 2014). An alternative to the
black box nature of ANNs, DT generates logic statements and inter-
pretable rules that can be used for identifying the importance of
data features (Perner et al. 2001). Another advantage of DT is
avoiding the curse of dimensionality and providing high per-
formance computing efficiency through its splitting procedure
(Prasad et al. 2006). However, DT produces unsatisfactory perfor-
mance in time series, noisy, or nonlinear data (Curram and
Mingers 2017).
Case-Based Reasoning
Case-based reasoning (CBR) is a sustained learning and incremen-
tal approach that solves problems by searching the most similar
past case and reusing it for the new problem situation (Aamodt
and Plaza 1994). Therefore, CBR mimics human problem solving
(Ross 1989; Kolodner 1992). As illustrated in Fig. 11, CBR is a
cyclic process of learning from past cases to solve a new case. The
ing. The retrieving process solves a new case by retrieving the past
cases. The case can be defined by key attributes. Such attributes are
used to retrieve the most similar case, whereas the reusing process
uses the new case information to solve the problem. The revising
process evaluates the suggested solution to the problem. Finally,
retaining process is to update the stored past cases with such a
new case by incorporating the new case to the existing case-base
(Aamodt and Plaza 1994). A CBR model can be developed to pre-
dict the conceptual cost based on similar attributes of the entered

© ASCE 03119008-12 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Tree_1 Tree_2
SVM hyperplane
wx+b = 0 Predictor_1 Predictor_2
-1 in
+b= ar g Is male ?
M age 15 ?
wx
1
Y N Y N
W = default default
1/ W +b
1/ wx X1 X2 X1 X2

+ 20 - 20 +5 +5
F(X1) = +20 +5 = + 25
Misclassified cases
F(X2) = -20 +5 = - 15
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

Supportive vectors
Fig. 12. Additive function concept.
Fig. 10. Linear support vector machine. (Data from Burges 1998.)

Ensemble Methods
Problem
Case attributes Ensemble methods (fusion learning) are elegant data mining tech-
1 niques to combine multiple learning algorithms to enhance the
New case
Retrieve overall performance (Hansen and Salamon 1990). Ensemble meth-
Learned case ods can apply any ML algorithm such as ANN, decision tree, and
SVM, which are called base model or base learner, as inputs for
4 Retain ensemble methods. The concept behind ensemble methods can be
Cases-base Retrieved case
illustrated as in Fig. 12 and mathematically as Eq. (19) (Chen and
Confirmed case
2 Guestrin 2016).
3 Reuse For a given data set (D) with n examples and m features
Revise Suggested case D ¼ fðxi ; yi Þgðxi ∈ Rm ; yi ∈ RÞ where R is the real numbers set.
K is an additive function to predict the output as Eq. (19)
Fig. 11. CBR processes. (Data from Aamodt and Plaza 1994.)
X
K
y^ i ¼ fk ðX i Þ; fk ∈ F ð19Þ
k¼1

case comparable with the stored cases. Once attributes are entered,
attributes similarities (AS) can be computed based on Eq. (17)
(Kim et al. 2004)
MinðAV N ; AV R Þ
AS ¼
MaxðAV N ; AV R Þ
where AS = attribute similarity; AVN = attribute value of the newly
entered case; and AVR = attribute value of the retrieved case.
Depending on AS and attribute weights (AW), case similarity (CS)
can be computed by Eq. (18) (Perera and Watson 1998). AW are
selected by an expert to emphasize the existence and importance of
the case attributes
Pn
ðAS × AW i Þ
CS ¼ i¼1 Pn i ð18Þ
i¼1 ðAW i Þ
where CS = case similarity; AS = attribute similarity; AW = attrib-
ute weight; and i = number of the attributes (key cost drivers).
The advantage of CBR is that it deals with a vast amount of data
where all past cases and new cases are stored in database techniques
(Kim et al. 2004). Developing CBR methods can be summarized in
the following steps:
1. Collect and prepare the historical cases.
2. Divide the collected cases into a training set and a validation set.
3. Determine relevant input attributes and outputs.
4. Identify the similarity function and conduct CBR processes.
5. Calculate the resulting error such as mean absolute percentage
error (MAPE).
where F ¼ ffðxÞ ¼ wqðxÞ g (q: Rm → T, T ∈ Rm ). q is the struc-
ture of each tree that maps an example to the corresponding. T cor-
responds to the number of leaves in the tree. y^ i is the predicted
dependent variable. Each fk represents an independent tree struc-
ð17Þ ture q and leaf weights w. xi represents independent variables.
F represents the regression trees space. Fig. 12 shows a tree ensem-
ble model. wi corresponds to the score on ith leaf where each leaf
has a continuous score unlike decision trees.
Ensemble methods compromise different methods such as
voting, stacking, bagging, and boosting. Voting and averaging
are two of the simplest ensemble methods, in which averaging is
used for regression and voting is used for classification (Opitz
and Maclin 1999). The ensemble learning methods can effec-
tively deal with the problems of high-dimension data, complex
data structures, and small sample size (Dietterich 2000). Bagging
algorithms [Fig. 13(a)] can increase generalization by decreasing
variance (Breiman 1998), whereas boosting [Fig. 13(b)] can im-
prove generalization by decreasing bias error (Schapire et al.
1998).
In addition, ensemble models can be classified into two main
types: homogeneous and heterogeneous. The homogeneous model
applies the same base algorithm on different training data sets,
whereas the heterogeneous model uses different base algorithms
on the same training data (Reid 2007). Ensemble methods can
effectively handle continuous, categorical, and dummy features
with missing values. However, ensemble methods may increase
model complexity, which decreases the model interpretability
(Kuncheva 2004).

© ASCE 03119008-13 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Classification Classification Classification
OR OR OR
D Bootstrap D Feature Regression Regression Regression
Randomization I1 I2
Sampling Ik

Ik Ik
D1 D2 Dk D1 D2 Dk
D D D

Averaging Averaging Averaging

OR OR OR
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

Voting Voting Voting

(a) (b) (c)

Fig. 13. (a) Bagging; (b) RF; and (c) Boosting.

Bagging
Bagging is a variance reduction algorithm to train several classifiers
based on bootstrap aggregation. A bagging algorithm randomly
draws replicas of a training data set with replacement to train each
classifier (Breiman 1996). As a result, diversity is obtained by re-
sampling several data subsets. On average, each bootstrap sample
contains 63.2% of the original training data set. The following three
steps summarize the algorithm execution (Breiman 1999):
1. T bootstrap samples BS1 ; BS2 ; : : : ; BST are generated.
2. A classifier Ci is developed based on each bootstrap sample Bsi.
3. An optimal classifier C is selected from C1 ; C2 ; : : : ; CT whose
output is the class predicted most often by its subclassifiers, with
ties broken arbitrarily.
Random Forest
The random forest (RF) is a bagging ensemble learning model
that can produce accurate performance without overfitting issues
(Breiman 2001). RF algorithms draw bootstrap samples to develop
a forest of trees based on random subsets of features. Therefore,
some features may be selected more than once, whereas others
might never be selected (Breiman 2001). RF is more robust against
noisy data or big data than the DT algorithm (Breiman 1996;
Dietterich 2000). The key limitation of the RF algorithm is that
it cannot interpret the importance of features or the mechanism
of producing the results. RF does not search for the best split var-
iables that diminish the correlation among the developed trees and
the strength of every single tree. As a result, RF decreases the gen-
eralization error (Breiman 2001). However, RF cannot interpret the
produced predictions. An extremely randomized tree (ERT) algo-
rithm merges the randomization of random subspace to a random
selection of the cut-point during the tree node splitting process.
Extremely randomized tree mainly controls the attribute randomi-
zation and smoothing parameters (Geurts et al. 2006).
Boosting and Adaptive Boosting (AdaBoost)
Schapire has presented a boosting procedure (also known as adap-
tive resampling) as an algorithm that boosts the performance of
weak learning algorithms (Schapire 1990). Bagging generates clas-
sifiers in parallel, whereas boosting develops the classifiers sequen-
tially as shown in Fig. 13(c). Thus, boosting converts weak models
to strong ones. Freund and Schapire (1997) have presented an adap-
tive boosting algorithm (AdaBoost). AdaBoost is selected as one of
the top ten data mining algorithms, and the authors of AdaBoost
won the Gödel Prize for their work (Wu et al. 2008). AdaBoost
serially manipulates the given data for each base learner.
AdaBoost assigns equal weights for all instances in which larger
weights are assigned to the misclassified cases. The objective is
to make a greater focus on the misclassified cases to be corrected
in the consequent iteration. In addition, the AdaBoost algorithm
uses other weights to rank each individual base learning algorithm
based on its accuracy (Bauer and Kohavi 1999).
Extreme Gradient Boosting (XGBoost)
Extreme Gradient Boosting (XGBoost) is a large-scale machine
learning system that can build a highly scalable end-to-end ensem-
ble tree boosting system for big data processing. The unique
advantage of the XGBoost is its scalability, so it can process noisy
data. XGBoost applies parallel computing to effectively reduce
computational complexity and learn faster (Chen and Guestrin
2016). The unique advantage of XGBoost is its scalability to fit
high-dimension data without overfitting based on the following
Eq. (20):
X
n X
K
LðÞ ¼ ðx þ aÞn ¼ lð^yi ; yi Þ þ Ωðfk Þ;
k¼0 k¼1
1
where ΩðfÞ ¼ γT þ λkwk2 ð20Þ
2
where L represents a differentiable convex cost function that deter-
mines the difference between the predicted output y^ i and the actual
output yi . Ω is to avoid overfitting and smooth the learnt weights
(W i ), The second term penalizes the complexity of the regression
tree functions.
XGBoost is a traditional gradient boosting tree algorithm
with a regularization parameter. Once the regularization term
is removed, the XGBoost is converted back to the traditional
gradient tree boosting. A differentiable convex cost function
can be replaced by Taylor series as a second-order approxima-
tion for fasting optimization (Friedman et al. 2000). Another key
advantage of XGBoost is handling the missing values for which
defaults direction is identified as shown in Fig. 12. Accordingly,
no effort is needed for cleaning the collected data (Fan et al.
2008).

© ASCE 03119008-14 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Stochastic Gradient Boosting
The performance of gradient boosting can iteratively be improved
as a stochastic gradient boosting (SGB) algorithm by injecting ran-
domization into the selected data subsets. Adding randomization to
a boosting algorithm can substantially boost both the fitting accu-
racy and computational cost of the gradient boosting algorithm
(Breiman 1996; Freund and Schapire 1996). The training data is
randomly drawn at each iteration without replacement from the
data set. Stochastic gradient boosting can be viewed in this sense
as a boosting–bagging hybrid.
Hybrid Intelligent System
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
The fusion of the CI methodologies is called a hybrid intelligent
system, and Zadeh (1994) has predicted that the hybrid intelli-
gent systems will be the way of the future. FL is an approximate
reasoning technique. However, it does not have any adaptive capac-
ity or learning ability. On the other hand, ANNs provide an efficient
mechanism for learning from given data and accounting for uncer-
tainty that naturally exists. EC enables an optimization structure
for the developed system. Combining these methodologies can
enhance the computational model so that the limitations of any
single method can be compensated by other methods (Siddique and
Adeli 2013). Fig. 14 illustrates a fusion of three basic models:
FL, ANNs, and EC. Based on such models, many hybrid models
can be evolved such as neurofuzzy models and evolutionary neural
networks.
Moreover, the objective of data transformation is to address the
normality assumption of data distribution in which the probability
distribution shape has an important role in statistical modeling
to convert error terms for linear models (Tabachnick and Fidell
2007). Data transformation may produce more accurate results.
Stoy et al. (2008, 2012) have developed a semilog model to pre-
dict the cost of residential construction in which the MAPE for the
semilog model (9.6%) was better than the linear regression model
(9.7%). The previous result proved that semilog models may pro-
duce a more accurate model than a plain regression model. How-
ever, this is not a rule; in other words, plain regression models
may produce more accurate and simpler models than transformed
models.
Lowe et al. (2006) have established a predictive model based on
286 historical cases in which three alternatives—cost=m2 , the log
of cost variable, and the log of cost=m2 —have been developed
Case-based reasoning
Fuzzy Logic Evolutionary Computing
Fuzzy Logic
Evolutionary Computing Hybrid intelligent models
Neural Networks
Fig. 14. Hybrid intelligent systems.
© ASCE 03119008-15
J. Constr. Eng. Manage., 2020, 146(1): 03119008
instead of raw cost data model in which such a data transformation
approach has more accurate results than an untransformed data
model. Love et al. (2005) have represented the project time–cost
relationship using a logarithmic regression model. Wheaton and
Simonton (2007) have performed a semilog regression model to
assess a building cost index. Thus, a transformation of raw data
can help to produce a more reliable cost model.
Genetic-Fuzzy Model
Many approaches exist for evolutionary fuzzy hybridization
(Angelov 2002; Pedrycz 1997)]. Traditionally, an expert is con-
sulted to define such fuzzy rules, or the fuzzy designer can use
the trial and error approach to map the fuzzy rules and MFs. How-
ever, such an approach is time-consuming and does not guarantee
the optimal set of fuzzy rules. Moreover, the number of fuzzy
if-then inference rules increases exponentially by increasing the
number of inputs, linguistic variables, or outputs. In addition, the
experts cannot easily define all required fuzzy rules and the asso-
ciated MFs. In many engineering problems, the evolutionary algo-
rithm (EA) has been conducted to automatically develop fuzzy
rules and MFs to improve system performance (Chou 2006; Loop
et al. 2010). The genetic-fuzzy model has been developed to
optimally generate fuzzy inference rules. The formulation of the
genetic algorithm model depends mainly on defining two core
terms: a chromosome representation and an objective function.
Based on the Michigan approach, the chromosomes represents the
fuzzy rules where the number of chromosomes is the number of
fuzzy rules. Each chromosome contains a number of genes. The
process of the developed model consists of five main steps:
1. An initial population of chromosomes has been identified to
represent the initial state of the fuzzy rules. The four key cost
drivers have been fed to the fuzzy system.
2. The fuzzy system produces the final predicted output of the
system, y^ i .
3. The predicted cost (^yi has been fed to fitness function (F) to
evaluate the model performance where fitness function (F) is
model evaluation function.
4. GA uses the fitness function (F) to evaluate the search process in
which crossover probability and mutation probability have been
set at 0.7 and 0.01, respectively.
5. The new population of fuzzy rules has been produced based
on the crossover and mutation processes to form the optimal
fuzzy rules.
Neural Networks
Fuzzy Logic Evolutionary Computing
Neural Networks
Fuzzy Logic
Evolutionary Computing
J. Constr. Eng. Manage.
 Evaluation Techniques
The whole collected data set is divided into two main sets: a train-
ing set and validation set. The validation set is ranges from 10% to
30% of the whole set to evaluate the performance of the developed
model. Moreover, the training set could be divided into training and
cross-validation data to apply K-fold cross-validation technique.
K-fold cross-validation data are used to set the optimal hyper-
parameters. Evaluation techniques for predictive regression model
can be mean absolute percentage error (MAPE), the mean squared
error (MSE), the root-mean squared error (RMSE), the coefficient
of determination (R2 ) or adjusted R2 .
MAPE compares the predicted and actual outcomes
(Makridakis et al. 1998). MAPE can be classified as an excellent
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
prediction if MAPE is less than 10%, and if between 10% and
20% it can provide good prediction. Between 20% and 50%,
MAPE generates acceptable forecasting, and more than 50%
gives inaccurate prediction (Lewis 1982). MAPE can be ex-
pressed as follows:
1X n
jyi − y^ i j
MAPE ¼ × 100 ð21Þ
n i¼1 y^ i
where n = number of cases; i = number of the case; Y i ðhatÞ =
outcome of model; and Y i = actual outcome. The MSE
measures of how well the regression fits the data as following
Eq. (22) (Aczel 1989):
1X n
MSE ¼ ½y − y^ i 2 ð22Þ
n i¼1 i
where yi = observed values; and y^ i = predicted values of the
dependent variable Y for the ith case. The RMSE equals the
square root of MSE.
The R2 is expressed as Eq. (23):
Pn
2 SSE ðyi − y^ i Þ2
R ¼1− ¼ 1 − Pi¼1n ðy − ȳ Þ2
ð23Þ
SST i¼1 i i
where SSE = sum of squares of the residuals; SST = total sum of
squares; y = arithmetic mean of the variable Y; and R2 measures
the percentage of the variation percentage of the predictor Y ex-
plained by the dependent variable X. Thus, R2 indicates how well
the regression model fits the data. R2 ranges from zero to one,
−1 ≤ R2 ≤ 1. If the R2 value is 0.9 or above, it is classified as very
good, above 0.8 is good, above 0.5 is satisfactory, and below 0.5 is
poor (Aczel 1989; Ostertagová 2011). Adjusted R-squared (R2 )
is computed by Eq. (24):
ð1 − R2 ÞK
R2 ¼ R2 − ð24Þ
n − ðK þ 1Þ
where R2 is adjusted for the number of variables included in the
regression equation where R2 is lower than the R2 value. For
model evaluation, R2 is always preferred to R2 to avoid the over-
fitting problem (Aczel 1989; Ostertagová 2011).
Cost Modeling Review
The objective of the cost modeling review is to provide an overview
of the recent and future trends in construction cost model develop-
ment. The study has reviewed the past practices of parametric
cost estimation at the conceptual stage for construction projects.
Recently, many international journals have been reviewed such as
© ASCE 03119008-16
J. Constr. Eng. Manage., 2020, 146(1): 03119008
the Journal of Construction Engineering and Management, the
Journal of Computing in Civil Engineering, and Automation in
Construction, and Construction Management and Economics.
Such journals represent the most common and high-ranked journals
for construction cost modeling.
Table 4 presents 70 papers to review the machine learning
techniques for construction cost prediction, with most papers hav-
ing been published from 2002 to 2018. The study focuses on
surveying the most common techniques used to build a reliable
parametric cost model based on the collected data. Many machine
learning (ML) and statistical techniques have been conducted
such as a regression model, ANNs, CBR, and SVM. Moreover,
hybrid models and fuzzy models have been reviewed to provide
an overall perspective of the cost models developments as shown
in Table 4.
Many previous studies have applied AI techniques and ML
models. Building information modeling (BIM) can feed data for
cost estimation, whereas a predictive ML model such as a regres-
sion model or ANN can predict the project’s cost on a macro level
(Juszczyk 2017). ANN has been applied for cost estimation of
sports fields where the general applicability of ANNs model has
been investigated (Juszczyk et al. 2018). ANNs have conducted
the early cost estimation of building projects for reinforced con-
crete buildings with acceptable performance (Ambrule and Bhirud
2017). CBR has been proposed for estimation the preliminary costs
of sports field construction based on 16 predictors using 143 con-
struction projects. Different calculations were conducted to formu-
late the case similarity based on quantitative and qualitative data for
which the final total error was 14% at the early stage (Leśniak and
Zima 2018). Prediction performance of a cost prediction model has
been improved by 17.23% and 4.39% for business facilities and
multifamily housing, respectively.
Based on more than 1,400 projects, a multilayer ensemble of
methods has been developed for forecasting the unit price bids of
resurfacing highway projects (Cao et al. 2018). Wang and Ashuri
(2016) have applied a random tree model for construction cost
index prediction. Williams and Gong (2014) have built a stacking
ensemble learning and text mining model to estimate the cost over-
run using the project contract document for which the accuracy was
44%. Building information modeling (BIM) can automate cost
estimation processes and improve inaccuracies where New Rules
of Measurement (NRM) for cost estimation can be mined for au-
tomatic cost estimation based on a four-dimensional BIM modeling
software (Kim et al. 2019).
Arabzadeh et al. (2018) have developed ANN, regression, and
hybrid models for cost estimation of spherical storage tanks. The
results indicated that ANNs were more accurate than a hybrid re-
gression model and hybrid ANNs were more accurate than single
ANNs. Linear and multiple regression models have been counted to
predict the preliminary estimate of road projects in Nigeria at the
early stage (Ogungbile et al. 2018). However, the whole collected
data set was only 50 for seven predictors, which is not a sufficient
data size to train regression models. Zhang et al. (2018) have con-
verted the time series model into a graph to forecast the construc-
tion cost index, and the application showed its ability to provide
more accurate estimations.
A parametric model mainly depends on parameters to simu-
late and describe the case studied (AACE 2004; Elfaki et al. 2014).
Parametric modeling builds a mathematical relationship between
dependent and independent variables. CI, machine learning, and
data science are the disciplines that map the relationships among
these variables and figure out such patterns. The most common
techniques for those disciplines have been reviewed and discussed
in the paper, including MRA, FL, ANNs, CBR, and hybrid systems
J. Constr. Eng. Manage.
 Table 4. Review of past practices of cost model development
Algorithm Application Findings and characteristics Reference
ANNs Mark up estimation In 1992, This study showed that ANNs produce better performance than the Moselhi and
hierarchical model for markup estimation. Moreover, GA is used for optimizing Hegazy (1993)
ANN weights for markup estimation. Such a model has displayed good
generalization results.
ANNs Highway construction In 1998, this study conducted a regularization neural networks model that produced a Adeli and Wu
more predictable and reliable model for highway construction projects. (1998)
ANNs Building projects Based on 300 examples, three cost models composed of 5, 9, and 15 input parameters Emsley et al.
have been developed to predict the cost per m2 and the log of cost per m2 in the (2002)
United Kingdom in 2002. The cost per m2 model produces higher R2 , whereas the
log model produces lower values of MAPE. For the selected model, the R2 value is
0.789 and the MAPE is 16.6%
ANNs Structural systems of Based on 30 examples, a cost model composed of eight design parameters was Günaydın and
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

residential buildings developed to predict the cost per m2 of reinforced concrete for 4- to 8-story Doğan (2004)
residential buildings in Turkey in 2004. The cost estimation accuracy is 93%
ANNs Highway construction In 2005, an ANN model was built for highway construction costs in which the index Wilmot and Mei
of highway construction cost reflected the change in overall cost over time. (2005)
ANNs Building projects Based on 286 past cases of data collected in the United Kingdom, linear regression Lowe et al. (2006)
models and ANN models have been established to assess the cost of buildings. Three
alternatives—cost=m2 , the log of cost variable, and the log of cost=m2 —have been
conducted instead of raw cost data when such data transformation approaches have
better results than an untransformed data model. A total of six models have been
developed based on forward and backward stepwise regression analyses. The best
regression model was the log of cost backward model.
ANNs Building projects Based on 169 examples, an ANN cost model has been developed for building El-Sawalhi and
construction projects with acceptable prediction error. Shehatto (2014)
ANNs Highway construction A prediction model has been developed with an MAPE of 1.4% for the unit cost of Elbeltagi et al.
the highway project in Libya by changing ANNs structure, training functions, and (2014)
training algorithms until the optimum model was reached.
ANNs Public construction Based on 232 public construction projects in Turkey, a multilayer perceptron Bayram et al.
projects (MLP) model and radial basis function (RBF) model were developed to (2015)
estimate construction cost. RBF shows superior performance to MAPE, with
approximately 0.7%.
ANNs Building projects Based on 657 building projects in Germany, a multistep ahead approach is conducted Dursun and Stoy
to increase the accuracy of the model’s prediction. (2016)
ANNs Water treatment plant First, cost drivers that influence construction costs of water treatment plants have Marzouk and
costs been identified. Cost drivers have been determined through descriptive statistics Elkadi (2016)
ranking (DSR) and exploratory factor analysis. Principal component analysis (PCA)
with varimax rotation through five iterations has been used to minimize the
multicollinearity problem. Kaiser criterion was used so that a total of 33 variables
were reduced to eight components, whereas using Cattell’s scree test reduced
variables to four components.
ANNs and Highway projects Radial basis neural networks and regression models were developed for completed Williams (2002)
regression project cost estimation. The regression model produced better performance than the
ANN model. Moreover, a hybrid model was developed and produced reliable results.
A natural log transformation helped to improve the linear relationship between
variables.
ANNs and Cost deviation in Based on 41 examples, this study compared an ANN model with the regression Attalla and
regression reconstruction projects model for cost deviation in reconstruction projects in 2003. Hegazy (2003)
ANNs and Tunnel construction Based on 33 constructed tunnels, both ANNs and regression models have been Petroutsatou et al.
regression developed for tunnel construction in which the developed models were fitted for their (2012)
purpose and were reliable for cost prediction.
ANNs and Structural steel Based on 35 examples, a cost model consisted of three input parameters was El-Sawah and
regression buildings developed to predict the preliminary cost of structural steel buildings in 2014. ANNs Moselhi (2014)
produced better performance than regression models in which the ANNs model had
improved the MAPE by approximately 4% compared to the regression model.
ANNs and Field canals This paper developed a quadratic regression model and ANNs that can predict ElMousalami
regression improvement projects the conceptual cost at 9% MAPE. Data transformation plays important role in et al. (2018b)
prediction accuracy.
CBR Building projects This study incorporated the decision tree into CBR to identify attribute weights of Doğan et al.
CBR. Such an approach shows more reliable results for residential building projects (2008)
cost assessment.
CBR Pavement Based on the library of past cases, this study developed a CBR model for pavement Chou (2009)
maintenance operation maintenance operations costs based on computing case similarity.
CBR Pump stations A parametric cost model was presented in which a questionnaire survey was Marzouk and
organized to analyze the most critical factors affecting the final cost of pump Ahmed (2011)
stations. Using a Likert scale, these factors were screened to determine the key
factors. A case-based reasoning was built and tested to develop the proposed model.

© ASCE 03119008-17 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 4. (Continued.)
Algorithm Application Findings and characteristics Reference
CBR Military barrack Based on 129 military barrack projects, a CBR model has been developed for cost Ji et al. (2012)
projects estimation in which the model produces reliable results.
CBR Building projects This study introduced an improved CBR model based on multiple regression analysis Jin et al. (2012)
(MRA) technique in which MRA has been applied in the revision stage of the CBR
model. Such a model significantly improved the prediction accuracy; the
performance of the business facilities model was improved by 17.23%.
CBR Storage structures This study built a CBR model to estimate resources and quantities in construction De Soto and Adey
projects. The nearest neighbor technique was conducted to measure the retrieval (2015)
phase similarity of the CBR model. The model has shown reliable MAPE ranging
from 8.16% to 28.40%.
CBR Building projects Ahn et al. (2017) examined the CBR similarity measure based on the weighted Ahn et al. (2017)
Mahalanobis distance to take attributes covariance into consideration.
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

CBR Building projects Ji et al. (2018) have proposed a learning method to handle missing data values Ji et al. (2018)
based on a data mining algorithm to improve the stability and performance of the
CBR model.
CBR and GA Bridges A cost estimation model was developed based on CBR and GA for bridge projects Kim and Kim
which was used for optimizing parameters of CBR. Such methodology improves the (2010)
accuracy compared to the conventional cost model.
CBR and AHP Highway Analytic hierarchy process (AHP) was incorporated into CBR to build a reliable cost Kim (2013)
estimation model for highway projects in South Korea.
Evolutionary neural Highway Based on 18 examples, a reliable NN cost model was developed based on optimizing Hegazy and Ayed
network (NN) NN weights for highway projects. Simplex optimization of neural network weights is (1998)
more accurate than trial and error and GA optimization for which the MAPE was 1%.
Evolutionary NN Residential buildings Based on 498 cases, a reliable NN cost model was developed based on optimizing Kim et al. (2005)
NN weights for residential buildings. GA optimization of NN parameters was more
accurate than trial and error model for which the MAPE was 4.63%.
Evolutionary fuzzy Building projects This study incorporated computation intelligence models such as ANNs, FL, and EA Cheng et al.
neural inference to make a hybrid model that improves the prediction accuracy in a complex project. (2009)
model (EFNIM), As a result, an evolutionary fuzzy neural model was developed for conceptual cost
estimation for building projects with reliable accuracy.
Evolutionary fuzzy Building projects An evolutionary fuzzy hybrid neural network model was developed for conceptual Cheng and Roy
hybrid neural cost estimation. FL was used for fuzzification and defuzzification for inputs and (2010)
network outputs, respectively. GA was used for optimizing the parameters of models such as
NN layer connections and FL memberships.
Evolutionary fuzzy Building projects Hybrid AI system based on SVM, FL, and GA were built for decision making for Cheng and Roy
and SVM project construction management. The system used FL to handle uncertainty in the (2010)
system, SVM to map fuzzy inputs and outputs, and GA to optimize the FL and SVM
parameters. The objective of such a system is to produce accurate results with less
human intervention in which MF shapes and distributions can be automatically
mapped.
GA for ANNs Residential buildings This study has built three cost NN models by back-propagation (BP) algorithm, GA Kim et al. (2005)
for optimizing NN weights, and GA for parameters optimization of the BP algorithm.
Optimizing the parameters of the BP algorithm produced the best results.
GA for ANNs Bridge construction GA was used as an optimizing tool for ANNs and CBR cost models in which two Chou et al. (2015)
and CBR projects such models have been developed for bridge projects in Taiwan. Both models have
displayed reliable results.
Fuzzy linear Wastewater treatment Based on 48 wastewater treatment plants, a fuzzy logic model was developed with Chen (2002)
regression plants acceptable error and uncertainty considerations.
Fuzzy logic Design cost overruns Based on the collected building projects in 2002, a fuzzy logic model was developed Knight and Fayek
on building projects for estimating design cost overruns on building projects with acceptable error and (2002)
uncertainty considerations.
Fuzzy sets Cost range estimation This study proposed the use of fuzzy numbers for cost range estimation and claimed Shaheen et al.
the fuzzy numbers for fuzzy scheduling range assessment. (2007)
Fuzzy model Wastewater treatment This study compared a linear regression model with a fuzzy linear regression model Papadopoulos
projects for wastewater treatment plants in Greece. The results of both models are similar and et al. (2007)
reliable.
Fuzzy model Building projects A fuzzy model is built based on four inputs and one output in which a set of if-then Yang and Xu
rules, the center of gravity fuzzification, the product inference engine, and singleton (2010)
fuzzifier are applied. The maximal error is 3.2%.
Fuzzy model Building projects This study applied index values for membership degree and exponential smoothing Shi et al. (2010)
method to develop a construction cost model.
Fuzzy neural Cost estimation An evolutionary fuzzy neural network model was developed for cost estimation Zhu et al. (2010)
network based on 18 examples and 2 examples for training and testing, respectively. GA is
used to avoid sinking in local minimum results.
Fuzzy logic Building projects Based on 106 building projects in the Gaza Strip in 2012, a fuzzy logic model was El Sawalhi (2012)
developed for building projects with acceptable error and good generalization.

© ASCE 03119008-18 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 4. (Continued.)
Algorithm Application Findings and characteristics Reference
Fuzzy model Cost prediction An improved fuzzy system is established based on fuzzy c-means (FCM) to solve the Zhai et al. (2012)
problem of fuzzy rules generation. Such a model has produced better results for
scientific cost prediction.
Fuzzy logic and Construction materials This study developed an ANN model for predicting construction material prices, Marzouk and
neural networks prices whereas a fuzzy logic model was applied to determine the degree of importance of Amin (2013)
each material to use for an ANN model. Such modeling has acceptable accuracy in
training and testing phases.
Fuzzy subtractive Telecommunication Based on 568 cases, a four-input fuzzy clustering model and sensitivity analysis were Marzouk and
clustering towers conducted for estimating telecommunication towers construction cost with Alaraby (2014)
acceptable MAPE.
Fuzzy logic Satellite cost Based on two input parameters, a fuzzy logic model was developed for satellite cost Karatas and Ince
estimation estimation. Such models work as fuzzy expert tools for satellite cost prediction. (2016)
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

Regression Building projects Based on 530 examples, three cost models composed of nine parameters were Kim et al. (2004)
analysis, NN, and developed to predict the building costs in Korea in 2004. The NN model produces
CBR better results than the CBR and regression models. However, CBR produces better
results than NN for long-term use due to updating cases to the CBR system.
Neurogenetic Residential buildings Based on 530 cases of residential buildings, a neurofuzzy cost estimation model was Kim et al. (2005)
built in which GA is applied for optimizing BP algorithm parameters. Such an
approach has more accurate results than a trial and error BP algorithm.
Neurofuzzy Residential This study developed an adaptive neurofuzzy model for cost estimation for Yu and
construction projects residential construction projects. Such a model is the integration of the ratio Skibniewski
estimation method with the adaptive neurofuzzy method to obtain mining assessment (2010)
knowledge that is not available in traditional approaches.
Neurofuzzy and GA Semiconductor Based on 54 case studies of semiconductor hookup construction, a neurofuzzy cost Hsiao et al. (2012)
hookup construction estimation model was built and optimized by GA. Such a model has an accuracy
better than the conventional cost method by approximately 20%.
Neurofuzzy Water infrastructure Based on 98 examples, a combination of neural networks and fuzzy set theory was Ahiaga-Dagbui
incorporated to develop a more accurate and precise model for water infrastructure et al. (2013)
projects in which MAPE is 0.8%.
Neurofuzzy Water infrastructure Based on 1,600 water infrastructure projects in the United Kingdom, a neurofuzzy Tokede et al.
projects hybrid cost model has been built in which max-product composition produces better (2014)
results than max-min composition.
Regression Building projects A logarithmic regression model has been developed to examine the project time–cost Love et al. (2005)
relationship. Projects in various Australian states have performed a transformed
regression model (semilog) to estimate a building cost index based on historical
construction projects in several markets (Wheaton and Simonton 2007).
Regression Building projects A semilog model has used to predict the cost of residential construction projects in Stoy et al. (2008)
which the MAPE for the semilog model (9.6%) is more than a linear regression
model (9.7%). The previous result proved that semilog models may produce a more
accurate model than plain regression models. However, this is not a rule; in other
words, plain regression models may produce more accurate and simpler models than
transformed models.
Regression Building projects A semilog regression model was performed to develop cost models for residential Stoy et al. (2012)
building projects in Germany. The most significant variables were identified by
backward regression method. For the selected population, the proposed model has a
prediction accuracy of 7.55%.
ANNs and MRA Highway Gardner et al. (2016) built an ANN and MRA for conceptual cost estimating for Gardner et al.
infrastructure Highway infrastructure (2016)
Bayesian regression Masonry retrofit Nasrazadani et al. (2017) have created a Bayesian regression to develop probabilistic Nasrazadani et al.
cost models for retrofit actions based on 167 masonry retrofit projects. (2017)
LASSO regularized Highway construction Zhang et al. (2017) developed a LASSO regularized regression for Forecasting cost Zhang et al.
regression projects of highway construction projects (2017)
ANN and GA Sale prices of real Rafiei and Adeli (2015) have proposed a novel hybrid model of deep belief restricted Rafiei and Adeli
estate units Boltzmann machine and genetic algorithm for estimation of sale prices of real (2015)
estate units
SVM Building construction Based on 62 cases of building construction projects in Korea, the SVM model was An et al. (2007a,
project developed to evaluate conceptual cost estimation. Such a model can help clients to b)
know the quality and accuracy of cost prediction.
SVM Building projects This study utilized the theory of the rough set (RS) with SVM to improve the HongWei (2009)
prediction accuracy. RS was used for attributes reduction.
SVM and ANNs Building projects Based on 92 building projects, ANNs and SVM were used to predicted cost and Wang et al. (2012)
schedule success at the conceptual stage. Such a model has a prediction accuracy of
92% and 80% for cost success and schedule success, respectively.
SVM Commercial building Based on 84 cases of commercial building projects, a principal component analysis Son et al. (2012)
projects method was developed into SVM to predict cost estimate based on project
parameters.

© ASCE 03119008-19 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 4. (Continued.)
Algorithm Application Findings and characteristics Reference
SVM Building projects This study incorporated a least squares support vector machine (LS-SVM), Cheng et al.
differential evolution (DE), and machine learning–based interval estimation (MLIE) (2013)
for interval estimation of construction cost. DE was used for optimizing
cross-validation process to avoid overfitting.
SVM Building projects Based on 122 historical cases, a hybrid intelligence model was developed for Cheng et al.
construction cost index estimation with 1% MAPE. Such a model consists of least (2013)
squares support vector machine (LS-SVM) and differential evolution (DE), and DE
is applied to optimize LS-SVM tuning parameters.
SVM Building projects This study developed a hybrid cost prediction model for the building based on the Cheng and Hoang
machine learning based interval, least squares support vector machine (LS-SVM), (2014)
and estimation (MLIE), and differential evolution (DE).
SVM Predicting bidding Based on 54 tenders, an SVM model was developed with 2.5% MAPE for bidding Petruseva et al.
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

price price prediction. (2016)

SVM, ANN, School building This study presented a comparison for a reliable cost prediction model based on Kim et al. (2013)
regression construction SVM, ANN, and regression models.
SVM, regression Construction costs This study proved the superiority of SVM compared to regression models in Petruseva et al.
prediction accuracy. (2017)
Structural equation Cost of construction This study discussed reducing construction costs based on reducing the Ajayi and Oyedele
modeling wastes construction wastes. (2018)

(Siddique and Adeli 2013). For example, ElMousalami et al.
(2018b) have identified the key cost drivers based on both quali-
tative and quantitative techniques. Consequently, ElMousalami
et al. (2018b) have conducted two machine learning models by uti-
lizing MRA and ANNs. Accordingly, the selected quadratic regres-
sion model produces a prediction accuracy of 9.12% and 7.82% for
training and validation, respectively. Similarly, Marzouk and Elkadi
(2016) have identified the key cost drivers based on qualitative
techniques such as questionnaires and quantitative techniques such
as exploratory factor analysis. Consequently, the next stage is the
model development. Marzouk and Elkadi (2016) have applied
ANNs when the MAPE for test sets was 21.18%.
Fan et al. (2006) have developed a decision tree approach for
investigating the relationship between house prices and housing
characteristics. Moussa et al. (2006) have established a decision
tree model using integrated multilevel stochastic networks. Cao
et al. (2018) have proposed a multilayer ensemble of methods for
prediction of unit price bids of resurfacing highway projects based
on more than 1,400 projects. The ensemble of methods was com-
posed of gradient boosting machine, XGBoost, and RF models, of
which the XGBoost has the best accuracy. Monte Carlo simulation
and a multiple linear regression model have been developed as a
benchmark model to evaluate the model’s performance where the
MAPE was 7.56%. Wang and Ashuri (2016) have developed a
highly accurate model based on random tree ensembles to predict
a construction cost index in which the model’s accuracy has
reached 0.8%. Williams and Gong (2014) have built a stacking en-
semble learning and text mining to estimate the cost overrun using
the project contract document in which the accuracy was 44%.
Chou and Lin (2012) have established an ensemble learning model
of ANNs, SVM, and a decision tree for predicting the potential for
disputes in public–private partnership (PPP) with an accuracy
of 84%.
Review Analysis and Discussion for Cost Modeling
Techniques
Based on the reviewed studies shown in Table 4, the survey has
been classified into six main categories to represent models used
for cost model development. These categories are ANN model,
FL model, regression model, SVM model, CBR model, and hybrid
models where hybrid models represent all combined methods such
as fuzzy neural network and evolutionary fuzzy hybrid neural net-
work. As shown in Fig. 15(a), the percentages of the catego-
ries are 27%, 25%, 14%, 13%, 11%, and 10% for hybrid models,
ANNs, fuzzy models, regression, SVM, and CBR, respectively.
These percentages indicate that hybrid models are the current trend
in parametric cost estimate modeling in which the researchers use
such hybrid models to enhance the performance of the developed
model and the accuracy of the prediction results. In addition, hybrid
models avoid the limitations of a single method. For example, the
hybrid model of ANNs and FL produces a neurofuzzy model that
provides uncertainty for ANNs. On the other hand, the ANN model
provides learning ability to the fuzzy system.
The second percentage of 25% represents the use of ANNs,
which is a powerful ML technique to represent nonlinear data.
The third percentage is 14% that represents fuzzy models. The
fuzzy model should be widely conducted since the fuzzy model
provides vagueness and uncertainty to the results and more reli-
able prediction to the future real-world cases. The fourth percent-
age is 13% that represents the regression model. Generally, the
regression model has been widely conducted because of its sim-
plicity. However, ANNs can provide better results than the regres-
sion model, specifically with nonlinear data. SVM and CBR have
similarly small percentages. However, CBR represents a promis-
ing technique in which a CBR works as an incremental search
engine for similar cases.
Based on the reviewed studies as shown in Table 4 and
Fig. 15(a), the survey has been classified into four main categories
to represent the projects used for the cost estimate. These categories
are buildings, highway, water infrastructure, and other projects. The
buildings category includes residential, industrial, and commercial
building projects, whereas the highway category includes highway,
road, pavement maintenance, and bridge construction projects.
Water infrastructure includes wastewater treatment and water infra-
structure projects. Other projects include tunnel projects, steel proj-
ects, and telecommunication towers, etc.
As shown in Fig. 15(b), the building category represents 48% of
all collected projects, whereas the other projects category repre-
sents 30%, the highway category represents 13% of all collected
projects, and the water infrastructure category represents 9%. Sub-
sequently, building projects and highway projects have the greatest

© ASCE 03119008-20 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Regression Other Modeling Methods
13%
CBR
10% Regression In addition to the common or established models previously dis-
ANNs cussed, multiple simple methods exist from industry practice. Time
Hybrid SVM series is a statistical model that uses the successive time data points
model ANNs to build predictive models (Bishop 2006; Wang and Ashuri 2016).
27% 25% Fuzzy model
Ashuri and Lu (2010) have developed a time series for construction
Hybrid model
SVM
cost index (CCI) prediction with approximately 1% MAPE.
CBR
11% Monte Carlo simulation can model the project cost where no
Fuzzy model independent attributes exist. Monte Carlo simulation uses a ran-
14% domness procedure that maintains the uncertainty to the estimation.
(a) However, the prediction performance is relatively lower than the
ML algorithms. Back et al. (2000) have applied Monte Carlo sim-
ulation to randomly generate an escalation rate at any point in time
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

buildings for project cost likelihood quantification. However, Ilbeigi et al.

Other (2016) pointed out that the limitation of Monte Carlo tech-
Highway nique is ignoring the autocorrelation effects of the collected data.
projects
30% Anderson et al. (2006) have estimated highway projects cost based
buildings water infrastructure
48% on a simple escalation approach where possible changes in the
Other projects future prices of materials have been modeled as an inflation cost
water estimator.
infrastructure
9%
Highway
13% Application to Field Canals Improvement Projects
(b)
AI can automatically develop the relations among cost drivers and
the project costs for which the final prediction error can be mini-
mized. Therefore, AI can diminish human interventions to estimate
above 300 above 300 cases the project cost. Moreover, the automated parametric model needs a
cases few cost drivers as inputs to predict the final cost without quantity
23% above 100 cases survey using low computational time and memory (ElMousalami
less 100 cases
50% less 100 cases
et al. 2018b). In this section, the selected AI techniques are applied
above 100
cases to the conceptual cost prediction of field canals improvement proj-
27% ects (FCIPs) in Egypt as an actual case study.

(c) Case Background

Fig. 15. Percentage breakdown of the reviewed studies: (a) AI models;
(b) project categories; and (c) data sample size.
share of researchers’ interest, whereas the other projects have fewer
research efforts.
Based on the surveyed studies, the collected sample sizes range
from 18 cases to 1,600 cases. The sample size can be divided into
three categories: less than 100 cases, over 100 cases, and over
300 cases. These three categories represent the most frequent values
of the collected sample size. In addition, many sample size studies
stand on these values: 100 and 300 observations (Tabachnick and
Fidell 2007; Comrey and Lee 1992; Guadagnoli and Velicer 1988;
MacCallum et al. 1999). As shown in Fig. 15(c), about 50% of
cases are less than 100 cases, whereas the studies with above 100
and 300 cases have 27% and 23%, respectively. Most of the sample
sizes are less than 100 cases, and that may reflect a model bias and
less model ability for generalization. Most of the studies do not
provide detailed steps for model development such as singularity,
multicollinearity, outliers, and sample size. For example, Hegazy
and Ayed (1998) have applied only 14 cases and 4 cases for training
and validation, respectively. Based on Green (1991), this study may
have concerns about the sample size, and Green (1991) has recom-
mended that 50 þ 8k may be the minimum sample size, where k is
the number of predictors.
FCIPs are one of the main projects in irrigation improvement
projects (IIPs) in Egypt. The strategic aim of these projects is to
save fresh water, facilitating water usage and distribution among
stakeholders and farmers. To finance this project, conceptual cost
models are important to accurately predict preliminary costs at the
early stages of the project (ElMousalami et al. 2018b; Radwan
2013).
Data Collection and Feature Selection
Based on contract information, cost drivers of FCIPs can be rep-
resented through a total of 17 parameters (ElMousalami et al.
2018b). Data preprocessing includes data normalization, cleaning,
and transformation. Once the inputs (key cost drivers) and the out-
put (conceptual construction cost of FCIP) are identified, relevant
data are collected to build the parametric cost model. The quantity
and quality of the collected instances are significant for the concep-
tual estimate that affects the accuracy of the developed model
(Bode 2000).
These 17 parameters can be filtered to be entered to ML models.
Therefore, Elmousalami et al. (2018a) have conducted qualitative
approaches such as the fuzzy Delphi method and fuzzy analytical
hierarchy process to rank the cost drivers. Moreover, ElMousalami
et al. (2018b) have developed a quantitative hybrid approach based
on both Pearson correlation and stepwise regression to filter the
key cost drivers. Accordingly, the final key cost drivers were area
served (P1 ), pipeline total length (P2 ), the number of irrigation

© ASCE 03119008-21 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 5. Accuracy of the developed algorithms
MAPE (%)
Notation Algorithm/model Algorithm type MAPE (%) categorization R2 R2
M1 XGBoost Ensemble methods 9.091 Below 10 0.931 0.929
M2 Quadratic regressiona MRA 9.120 Below 10 0.857 0.851
M3 Plain regressiona MRA 9.130 Below 10 0.803 0.796
M4 Quadratic MLPa ANNs 9.200 Below 10 0.904 0.902
M5 Plain MLPa ANNs 9.270 Below 10 0.913 0.912
M6 Semilog regressiona MRA 9.300 Below 10 0.915 0.910
M7 Extra trees Ensemble methods 9.714 Below 10 0.948 0.947
M8 Natural log MLPa ANNs 10.230 Below 20 0.905 0.910
M9 Bagging Ensemble methods 10.246 Below 20 0.914 0.911
M10 RF Ensemble methods 10.503 Below 20 0.916 0.913
M11 AdaBoost Ensemble methods 10.679 Below 20 0.875 0.871
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

M12 SGB Ensemble methods 11.008 Below 20 0.926 0.924

M13 Reciprocal regressiona MRA 11.200 Below 20 0.814 0.801
M14 Power (2) regressiona MRA 11.790 Below 20 0.937 0.931
M15 DNNs ANNs 12.059 Below 20 0.785 0.779
M16 DT Tree model 12.488 Below 20 0.886 0.883
M17 Genetic fuzzy Hybrid model 14.700 Below 20 0.863 0.857
M18 CBR Case-based 17.300 Below 20 0.859 0.852
M19 SVM Kernel-based 21.217 Unacceptable 0.136 0.133
M20 Fuzzy Fuzzy theory 26.300 Unacceptable 0.857 0.851
a
ElMousalami et al. (2018b).

valves (P3 ), and construction year (P4 ). Accordingly, a total of
144 FCIPs during 2010 and 2015 have been collected.
For validation purposes, this collected sample has randomly
branched into a training sample (111 instances) and a testing sam-
ple (33 instances). The training sample in the present case study of
111 instances would be sufficiently acceptable to train reliable ML
models (Green 1991).
Comparison and Analysis
MAPE and R2 have been validated for the 20 developed models as
displayed in Table 5. The whole developed models have been
sorted in descending order from M1 to M20 based on MAPE.
ElMousalami et al. (2018b) have presented a quadratic regression
model (M2) as the most accurate for FCIPs among the developed
regression and ANN models (M3, M4, M5, M6, M8, M13, and
M14) with 9.120 and 0.851 for MAPE and R2 , respectively. How-
ever, this study presents that XGBoost (M1) is more accurate than
quadratic regression (M2). XGBoost (M1) has obtained the first
place slightly higher than M2 with 9.091% and 0.929 for MAPE
and R2 , respectively. Moreover, the unique advantage of the
XGBoost is its high scalability; it can process noisy data and fit
high dimension data without overfitting. XGBoost applies parallel
computing to effectively reduce computational complexity and
learn faster (Chen and Guestrin 2016). Another key advantage of
XGBoost is handling the missing values where defaults direction is
identified. Accordingly, no effort is needed for cleaning the col-
lected data.
Ensemble methods such as Extra Trees (M7), bagging (M9),
RF (M10), AdaBoost (M11), and SGB (M12) have produced high,
acceptable performance with accuracy ranging from 9.714% to
11.008%. The ensemble learning methods can effectively deal with
the problems of high-dimension data, complex data structures, and
small sample size. Bagging algorithms can increase generalization
by decreasing variance error (Breiman 1998), whereas boosting can
improve generalization by decreasing bias error (Schapire et al.
1998). Ensemble methods can effectively handle continuous, cat-
egorical, and dummy features with missing values. However, en-
semble methods increase the model complexity, which decreases
the model interpretability (Kuncheva 2004). RF (M10) is more a
robust algorithm against noisy data or big data than the DT (M16)
algorithm (Breiman 1996; Dietterich 2000). However, the RF algo-
rithm is unable to interpret the importance of features or the mecha-
nism of producing the results.
DNNs (M15) produce 12.059% MAPE, which is less accurate
than all the developed MLP models (M4, M5, and M8). Accord-
ingly, DNNs provide bad performances with a small data set.
Conversely, deep learning and DNNs can produce the most accu-
rate performance with high-dimension data (LeCun et al. 2015).
An alternative to the black box nature of ANNs and DNNs, DTs
generate logic statements and interpretable rules that can be used
for identifying the importance of data features (Perner et al. 2001).
Another advantage of DT is avoiding the curse of dimensionality
and providing a high performance computing efficiency through its
splitting procedure (Prasad et al. 2006). However, DT produces
unsatisfactory performance for time series, noisy, or nonlinear data
(Curram and Mingers 2017). Although DT (CART) is inherently
used as a base learner for the ensemble methods, DT (M16) pro-
duces 12.488% MAPE, which is less accurate than all developed
ensemble methods (M1, M7, M9, M10, M11, and M12). Therefore,
ensemble methods produce better performance than a single learn-
ing algorithm. Moreover, ensemble methods can effectively handle
missing values and noisy data because of their scalability.
Ensemble methods and data transformation play an important
role in prediction accuracy. However, the main gap of the previous
models is the lack of uncertainty modeling in the prediction cost
model. Therefore, fuzzy logic theory has been conducted to main-
tain uncertainty concept through the fuzzy logic model (M17) as
shown in Fig. 16 and hybrid fuzzy model (M20) as shown in
Fig. 17. The number of generated rules by the fuzzy genetic model
(M17) is 63 rules, and the MAPE is 14.7%. On the other hand, a
traditional fuzzy logic model (M20) has been built based on the
experts’ experience in which a total of 190 rules were generated
to cover all the possible combinations of the fuzzy system, and
the MAPE is 26.3%. Moreover, the fuzzy rules (if-then rules) gen-
erated by experts have redundant rules that can be deleted to im-
prove the model computation and performance. Moreover, the
experts’ knowledge cannot cover all combinations to represent

© ASCE 03119008-22 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 PVC pipeline length Fuzzy system Concluding Remarks and Future Research Trends
Key cost drivers
Command area This study presents a general methodology to develop a robust
IF-THEN rules
parametric cost model at the conceptual stage of the project, and
Number of irrigation valves
the study explores qualitative and quantitative techniques to iden-
Constrution year
FCIPs Cost tify key cost drivers and develop an intelligent predictive model.
All results and recommendations have been extracted based on re-
viewing a total of 100 papers relevant to construction cost modeling
of which most of the papers were published in the period of 2002
µ ( MF) 1.0 1.0 to 2018 as illustrated in Tables 3 and 4. The methodology and the
potential recommendations can be validated by selecting any rel-
evant paper and comparing its methodology and recommendations
0.5 MF1 MF2 MF3 MF4 MF5 MF6 MF7
0.5 to this study’s methodology and recommendations. For example,
Cao et al. (2015) have developed a hybrid computational model
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

for forecasting Taiwan's construction cost index by conducting

0.0
Fig. 16. Fuzzy logic model.
Inputs
P1 P2
CH1 MF1 : MF7 MF1 : MF7 MF1 : MF7 MF1 : MF7 MF1 : MF7
CH2
……..
MF1 : MF7 MF1 : MF7 MF1 : MF7 MF1 : MF7 MF1 : MF7
…….. …….. ……..
CHn MF1 : MF7 MF1 : MF7 MF1 : MF7 MF1 : MF7 MF1 : MF7
5
1 2
4 GA Inputs Fuzzy system
F 3
Fig. 17. Genetic fuzzy model.
all possible rules (2,401 rules). In addition, the generation of the
experts’ rules is a time- and effort-consuming process. Conse-
quently, hybrid fuzzy systems are more effective than the traditional
fuzzy logic system. Although the prediction accuracy of the fuzzy
genetic model (M17) and the fuzzy logic model (M20) is 14.7%
and 26.3%, respectively, the fuzzy model would produce more re-
liable prediction results because it takes uncertainty into account.
However, the traditional fuzzy model gives unacceptable accuracy
of 26.3% MAPE (Peurifoy and Oberlender 2002). Therefore, main-
taining uncertainty decreases the predictive model accuracy.
CBR (M18) produces an acceptably low accuracy of 17.3%
MAPE. The advantage of CBR is dealing with a vast amount of
data in which all past cases and new cases are stored in database
techniques (Kim et al. 2004). Moreover, finding similarities and
similar cases improves the reliability and confidence in the
output. Hybrid models can be incorporated to CBR to enhance the
performance of CBR such as applying GA and decision tree to
optimize attribute weights and by applying regression analysis for
the revision process. SVM can be applied for both regression
and classification tasks. SVM (M19) produces an unacceptable
accuracy of 21.217% MAPE (Peurifoy and Oberlender 2002).
Finally, Table 6 summarizes the strengths and weakness of the de-
veloped models.
0.0 regression, ANN, and evolutionary computing to develop the
hybrid model. This paper is from the literature survey sample
and recommends the hybrid computational model for better predic-
tion accuracy. Therefore, this paper presents a comparison of AI
techniques to develop a reliable conceptual cost prediction model.
Twenty machine learning models are developed utilizing tree-based
models, ensemble methods, fuzzy systems, CBR, ANNs, SVM,
and transformed regression models. The accuracy of the developed
Output
P3 P4 Cost models is tested from two perspectives: MAPE and adjusted R2 .
The results show that the most accurate and suitable method is
…….. …….. XGBoost with 9.091% and 0.929 for MAPE and adjusted R2 ,
respectively.
According to the key cost drivers identification, this study has
reviewed the common practices of cost drivers identification for
1 parametric cost model development. Cost drivers identification
has been classified into two main types of procedures: qualitative
Ymodel
and quantitative. The trends and recommendation of the study are
as follows:
1. This study recommends using fuzzy approaches such as FDM
and FAHP over traditional methods such as TDM and AHP be-
cause the fuzzy approaches produce more reliable performance.
2. This study recommends applying both qualitative and quantita-
tive approaches to obtain the most reliable cost drivers. Such
a procedure can be called a hybrid approach for cost drivers
identification. The limitation of the hybrid approach is that both
experts’ opinions and historical cases are required.
3. This study recommends establishing a database for every con-
struction project and for such projects to be open source to be
used for research development.
4. The genetic algorithm is a powerful tool to select the optimal set
of the cost drivers for which the prediction error is minimized.
5. The paper emphasizes the importance of machine learning
models such as factor analysis and GA-ANNs models to auto-
matically identify the key cost drivers based on the collected
data without human errors and interventions.
Moreover, the core trend of cost modeling is to computerize and
automate the cost model so that less human intervention is required
for operating such models while obtaining higher accuracy and op-
timal results. According to cost modeling techniques, based on the
survey literature (Table 4), this study has performed a survey and
analysis for construction cost modeling development. The study is
presented in two main parts. The first part presents the most
common AI modeling techniques used for cost models, and the
second part presents the review of the current state for cost model
development. The first part explains statistical methods such as
MRA and intelligent methods such as FL, ANNs, EC, CBR,
and hybrid models. The second part reviews the model develop-
ment summarized in Table 4 where modeling techniques, construc-
tion project, parameters used, sample size, and model accuracy

© ASCE 03119008-23 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 Table 6. Characteristics of the developed algorithms
Missing
Algorithm values and
notations Strengths Weaknesses Interpretation Uncertainty noisy data
M1 High scalability, handling missing values, No uncertainty and interpretation No No Yes
high accuracy, low computational cost
M2 More accurate than plain regression, handles Prone to overfitting Yes No No
nonlinearity of data
M3 Works on small size of data set Linear assumptions Yes No No
M4 High accuracy, handling complex patterns Black box nature, need sufficient data for training No No No
M5 High accuracy, handling complex patterns Black box nature No No No
M6 Producing better results than plain regression Unable to capture complex patterns Yes No No
M7 Handling data randomness Black box nature and sufficient data No No Yes
M8 Producing better results than plain MLP Black box nature and sufficient data No No No
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.

M9 Providing higher performance than a single Depending on other algorithms performance No No Yes
algorithm
M10 Accurate and high performance on many No interpretability, need to choose the number No No Yes
problems including nonlinear of trees
M11 High scalability, and high adaptability Depends on other algorithms’ performance No No Yes
M12 Handling difficult examples High sensitivity to noisy data No No Yes
M13 Handling data nonlinearity and training Unable to capture complex patterns Yes No No
small sample size
M14 Handling data nonlinearity and training Unable capture complex patterns Yes No No
small sample size
M15 Capturing complex patterns, processing big Sufficient training data and high cost computation No No No
data and high performance computing
M16 Working on both linear and nonlinear data, Poor results on too small data sets, overfitting can Yes No No
and producing logical expressions easily occur
M17 Handling uncertainty and more accurate than More complex than fuzzy model and needs more Yes Yes No
fuzzy model computational resources
M18 Handling small data sets, simple and needs Poor performance for case in which the optimal Yes No No
less computational time case cannot be retrieved
M19 Easily adaptable, works very well on Compulsory to apply feature scaling, more No No No
nonlinear problems, not biased by outliers difficult to understand
M20 Handling uncertainty Low accuracy Yes Yes No

have been extracted and summarized. The following points sum-
marize the recommendations and future trends:
1. Computational models and information systems have been
applied in business and construction industry to effectively
improve the job efficiency (Davis 1993). Therefore, the hybrid
model represents the current trend of parametric cost model-
ing to improve the model performance and accuracy so that
the limitations of each technique can be avoided. The objective
is to develop computerized automated systems with fewer
interventions of humans to save time and effort, and to
avoid human error for the cost estimate. Moreover, computer
technologies have a great ability to deal with vast data and
complicated computations.
2. AI and CI models such as ANNs, FL, and GA are used widely
for hybrid model development. Moreover, ML techniques can
be efficiently conducted for the parametric cost modeling.
Therefore, the cost modeling researcher should firstly study
ML, CI, and AI techniques.
3. CBR represents the increasing importance of ML tools and
data mining techniques for knowledge acquisition, prediction,
and decision making. Specifically, CBR efficiently deals with
vast data and has the ability to update the case-base for future
problem solving. Moreover, Finding similarities and similar
cases improves the reliability and confidence in the output.
4. Hybrid models can be incorporated to CBR to enhance the
performance of CBR such as by applying GA and decision
tree to optimize attribute weights and by applying regression
analysis for the revision process.
5. Most of the studies focus on building types of construction
projects, so a need exists to apply cost models widely for
different kinds of construction projects to help project man-
agers and cost estimation engineers.
6. Automated cost models are prone to many machine learning
problems such as overfitting issues, and hyper-parameter
selection. It is recommended to develop more than one cost
prediction model such as regression, ANNs, FL, or CBR. As
a result, the researcher can compare the results of the devel-
oped models and set a benchmark to select the most accurate
model. In addition, the comparisons of the developed models
enhance the quality of cost estimate and the decisions based
on it (Amason 1996).
7. There is a need to develop a model that has the ability
to give justification for the model’s results and to give answers
and interpretations for the predicted cost. That may require a
higher level of AI and may represent the future trend of cost
modeling. Moreover, such concept may be generalized for any
prediction model. The objective is to avoid the estimator’s
biases, warn the user to the input parameters of the model,
and avoid the limitation of the black box nature.
8. The conceptual cost estimate is conducted under uncertainty.
Therefore, this study recommends using fuzzy theory such as
FL and to develop a hybrid model based on FL to obtain un-
certainty for the developed model and produce a more reliable
performance as illustrated in Fig. 18 (Elmousalami 2019).
9. One of the key challenges in cost estimation accuracy is
substantial variations over time. Therefore, each cost predic-
tion model should have an input time-related macroeconomic
indicator that represents the change in the market inflation
rate over time because significant variability in material cost
or inflation rate can reduce the prediction performance.

© ASCE 03119008-24 J. Constr. Eng. Manage.

J. Constr. Eng. Manage., 2020, 146(1): 03119008

 2. Parametric modeling
Fuzzy
GA
rules
1. Feature selection
Hybrid
Key cost Fuzzy
Data feature
drivers model Prediction
model
Evaluation
Test
Data
Fig. 18. Intelligent methodology for project conceptual cost predic-
tion. [Reprinted from Elmousalami 2019, under Creative Commons
BY-NC-ND 4.0 license (https://creativecommons.org/licenses/by-nc
-nd/4.0/).]
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
ElMousalami et al. (2018b) maintain reliability for future pre-
diction based on the following formula:
Future cost ¼ Predicted cost × ð1 þ IRÞn
where IR = average inflation rate; and n = number of years
from present to the future time.
10. ML gives satisfactory performance and accuracy, but it needs
sufficient features and a sufficient data size to train ML algo-
rithms. Monte Carlo simulation, regression, and time series
analysis are not sufficiently robust for cost prediction with
high variability and nonlinearity of data. Cao et al. (2018) re-
ported that the ensemble learning models provide 37.98% and
26.89% more accurate results than the regression method
and the Monte Carlo simulation method, respectively, using
mean absolute error (MAE) scale.
11. Ensemble methods are promising techniques that can handle
a large number of features, model both numerical and catego-
rical variables, capture nonlinear patterns, and fit data with
missing values.
12. Decision tree algorithms and ensemble methods can provide
an alternative technique to many ML algorithms such as multi-
ple regression analysis and ANNs. The study emphasizes the
importance of ensemble methods for improving the prediction
accuracy and handling noisy and missing data. However, the
key limitation of the ensemble methods in an inability to inter-
pret the produced results.
Most of the studies focus on building types of construction proj-
ects, so a need exists to apply cost models widely for different kinds
of construction projects. In addition, more reviewed studied means
more generalization and better quality of the results. Finally, accu-
rate cost estimate means accurate decisions about the project man-
agement. Therefore, this study has analyzed the past cost modeling
practices to provide a recent direction for construction cost model-
ing. The study shows that the computational intelligence, artificial
intelligence, and machine learning techniques have a powerful abil-
ity to develop applicable and accurate cost predictive models.
Moreover, the cost modeling research area needs more studies
to develop intelligent models that have the ability to interpret
the resulting cost prediction and analyze the input model’s param-
eters. In addition, this study has provided a list of recommendations
and references for cost model developers to build a more practical
parametric cost model.
Data Availability Statement
Data generated by the authors or analyzed during the study are
available at: https://github.com/HaythamElmousalami/Field-canals
-improvement-projects-FCIPs-.
© ASCE
J. Constr. Eng. Manage., 2020, 146(1): 03119008
References
AACE (Association for the Advancement of Cost Engineering). 2004.
AACE International recommended practices. Morgantown, WV:
AACE.
Aamodt, A., and E. Plaza. 1994. “Case-based reasoning: Foundational is-
sues, methodological variations, and system approaches.” AI Commun.
7 (1): 39–59.
Aczel, A. D. 1989. Complete business statistics. Burlington, MA:
Irwin.
Adeli, H., and M. Wu. 1998. “Regularization neural network for construc-
tion cost estimation.” J. Constr. Eng. Manage. 124 (1): 18–24. https://
doi.org/10.1061/(ASCE)0733-9364(1998)124:1(18).
Ahiaga-Dagbui, D. D., O. Tokede, S. D. Smith, and S. Wamuziri.2013. “A
neuro-fuzzy hybrid model for predicting final cost of water infrastruc-
ture projects.” In Proc., 29th Annual ARCOM Conf., 2–4. Reading, UK:
Association of Researchers in Construction Management.
Ahn, J., M. Park, H. S. Lee, S. J. Ahn, S. H. Ji, K. Song, and B. S. Son.
2017. “Covariance effect analysis of similarity measurement methods
for early construction cost estimation using case-based reasoning.”
Autom. Constr. 81 (Sep): 254–266. https://doi.org/10.1016/j.autcon
.2017.04.009.
Ajayi, S. O., and L. O. Oyedele. 2018. “Waste-efficient materials procure-
ment for construction projects: A structural equation modelling of
critical success factors.” Waste Manage. 75: 60–69.
Akintoye, A. 2000. “Analysis of factors influencing project cost estimating
practice.” Constr. Manage. Econ. 18 (1): 77–89. https://doi.org/10.1080
/014461900370979.
Alroomi, A., D. H. S. Jeong, and G. D. Oberlender. 2012. “Analysis of cost-
estimating competencies using criticality matrix and factor analysis.”
J. Constr. Eng. Manage. 138 (11): 1270–1280. https://doi.org/10
.1061/(ASCE)CO.1943-7862.0000351.
Amason, A. 1996. “Distinguishing the effects of functional and dysfunc-
tional conflict on strategic decision making: Resolving a paradox for top
management teams.” Acad. Manage. J. 39 (1): 123–148.
Ambrule, V. R., and A. N. Bhirud. 2017. “Use of artificial neural network
for pre design cost estimation of building projects.” Int. J. Recent
Innovation Trends Comput. Commun. 5 (2): 173–176.
An, S.-H., G.-H. Kim, and K.-I. Kang. 2007a. “A case-based reasoning cost
estimating model using experience by analytic hierarchy process.”
Build. Environ. 42 (7): 2573–2579. https://doi.org/10.1016/j.buildenv
.2006.06.007.
An, S.-H., U.-Y. Park, K.-I. Kang, M.-Y. Cho, and H.-H. Cho. 2007b.
“Application of support vector machines in assessing conceptual cost
estimates.” J. Comput. Civ. Eng. 21 (4): 259–264. https://doi.org/10
.1061/(ASCE)0887-3801(2007)21:4(259).
Anderson, S. D., K. R. Molenaar, and C. J. Schexnayder. 2006. Guidance
for cost estimation and management for highway projects during plan-
ning, programming, and preconstruction. NCHRP Rep. No. 574.
Washington, DC: Transportation Research Board.
Angelov, P. P. 2002. Evolving rule-based models: A tool for design of
flexible adaptive systems. Wurzburg, Germany: Physica-Verlag.
Arabzadeh, V., S. T. A. Niaki, and V. Arabzadeh. 2018. “Construction
cost estimation of spherical storage tanks: Artificial neural networks and
hybrid regression—GA algorithms.” J. Ind. Eng. Int. 14 (4): 747.
https://doi.org/10.1007/s40092-017-0240-8.
Ashuri, B., and J. Lu. 2010. “Time series analysis of ENR construction cost
index.” J. Constr. Eng. Manage. 136 (11): 1227–1237. https://doi.org
/10.1061/(ASCE)CO.1943-7862.0000231.
Attalla, M., and T. Hegazy. 2003. “Predicting cost deviation in reconstruc-
tion projects: Artificial neural networks versus regression.” J. Constr.
Eng. Manage. 129 (4): 405–411. https://doi.org/10.1061/(ASCE)0733
-9364(2003)129:4(405).
Back, W. E., W. W. Boles, and G. T. Fry. 2000. “Defining triangular prob-
ability distributions from historical cost data.” J. Constr. Eng. Manage.
126 (1): 29–37. https://doi.org/10.1061/(ASCE)0733-9364(2000)
126:1(29).
Bauer, E., and R. Kohavi. 1999. “An empirical comparison of voting clas-
sification algorithms: Bagging, boosting, and variants.” Mach. Learn.
36 (1–2): 105–139. https://doi.org/10.1023/A:1007515423169.
03119008-25 J. Constr. Eng. Manage.
 Bayram, S., M. E. Ocal, E. L. Oral, and C. D. Atis. 2015. “Comparison of
multi-layer perceptron (MLP) and radial basis function (RBF) for con-
struction cost estimation: The case of Turkey.” J. Civ. Eng. Manage.
22 (4): 480–490. https://doi.org/10.3846/13923730.2014.897988.
Berry, M. J., and G. Linoff. 1997. Data mining techniques: For marketing,
sales, and customer support. New York: Wiley.
Bertram, D. 2017. “Likert Scales are the meaning of life.” Accessed
February 20, 2017. http://poincare.matf.bg.ac.rs/∼kristina/topic-dane
-likert.pdf.
Bezdek, J. C. 1994. “What is computational intelligence?” In Computa-
tional intelligence imitating life, edited by J. M. Zurada, II, R. J. Marks,
and C. J. Robinson, 1–12. New York: IEEE.
Bishop, C. M. 2006. “Introduction.” In Pattern recognition and machine
learning, 1–58. New York: Springer.
Bode, J. 2000. “Neural networks for cost estimation: Simulations and pilot
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
application.” Int. J. Prod. Res. 38 (6): 1231–1254. https://doi.org/10
.1080/002075400188825.
Bowerman, B. L., and R. T. O’Connell. 1990. Linear statistical models:
An applied approach. 2nd ed. Belmont, CA: Duxbury.
Breiman, L. 1996. “Bagging predictors.” Mach. Learn. 24 (2): 123–140.
Breiman, L. 1998. “Arcing classifier (with discussion and a rejoinder by
the author).” Ann. Stat. 26 (3): 801–849. https://doi.org/10.1214/aos
/1024691079.
Breiman, L. 1999. “Pasting small votes for classification in large databases
and on-line.” Mach. Learn. 36 (1–2): 85–103. https://doi.org/10.1023
/A:1007563306331.
Breiman, L. 2001. “Random forests.” Mach. Learn. 45 (1): 5–32. https://doi
.org/10.1023/A:1010933404324.
Breiman, L., J. H. Friedman, R. Olshen, and C. Stone. 1984. Classification
and regression trees. Belmont, CA: Wadsworth.
Buckley, J. J. 1985. “Fuzzy hierarchical analysis.” Fuzzy Sets Syst. 34 (2):
187–195. https://doi.org/10.1016/0165-0114(90)90158-3.
Burges, C. J. C. 1998. “A tutorial on support vector machines for pattern
recognition.” Data Min. Knowl. Discovery 2 (2): 121–167. https://doi
.org/10.1023/A:1009715923555.
Cao, M.-T., M.-Y. Cheng, and Y.-W. Wu. 2015. “Hybrid computational
model for forecasting Taiwan construction cost index.” J. Constr.
Eng. Manage. 141 (4): 04014089. https://doi.org/10.1061/(ASCE)CO
.1943-7862.0000948.
Cao, Y., B. Ashuri, and M. Baek. 2018. “Prediction of unit price bids
of resurfacing highway projects through ensemble machine learning.”
J. Comput. Civ. Eng. 32 (5): 04018043. https://doi.org/10.1061/(ASCE)
CP.1943-5487.0000788.
Cattell, R. B. 1966. “The scree test for the number of factors.”
Multivariate Behav. Res. 1 (2): 245–276. https://doi.org/10.1207
/s15327906mbr0102_10.
Chen, H. 2002. “A comparative analysis of methods to represent uncer-
tainty in estimating the cost of constructing wastewater treatment
plants.” J. Environ. Manage. 65 (4): 383–409. https://doi.org/10.1016
/S0301-4797(01)90563-8.
Chen, T., and C. Guestrin. 2016. “XGBOOST: A scalable tree boosting
system.” In Proc., 22nd ACM SIGKDD Int. Conf. on Knowledge
Discovery and Data Mining, 785–794. New York: ACM.
Cheng, M.-Y., and N.-D. Hoang. 2014. “Interval estimation of construction
cost at completion using least squares support vector machine.” J. Civ.
Eng. Manage. 20 (2): 223–236. https://doi.org/10.3846/13923730.2013
.801891.
Cheng, M.-Y., N.-D. Hoang, and Y.-W. Wu. 2013. “Hybrid intelligence
approach based on LS-SVM and differential evolution for construction
cost index estimation: A Taiwan case study.” Autom. Constr. 35 (Nov):
306–313. https://doi.org/10.1016/j.autcon.2013.05.018.
Cheng, M.-Y., and A. F. Roy. 2010. “Evolutionary fuzzy decision model for
construction management using support vector machine.” Expert Syst.
Appl. 37 (8): 6061–6069. https://doi.org/10.1016/j.eswa.2010.02.120.
Cheng, M.-Y., H.-C. Tsai, and W.-S. Hsieh. 2009. “Web-based conceptual
cost estimates for construction projects using evolutionary fuzzy neural
inference model.” Autom. Constr. 18 (2): 164–172. https://doi.org/10
.1016/j.autcon.2008.07.001.
Chi, Z., H. Yan, and T. Phan. 1996. Fuzzy algorithms: With applications to
image processing and pattern recognition. Singapore: World Scientific.
© ASCE 03119008-26
J. Constr. Eng. Manage., 2020, 146(1): 03119008
Choi, S., D. Y. Kim, S. H. Han, and Y. H. Kwak. 2014. “Conceptual cost-
prediction model for public road planning via rough set theory and case-
based reasoning.” J. Constr. Eng. Manage. 140 (1): 04013026. https://
doi.org/10.1061/(ASCE)CO.1943-7862.0000743.
Choi, S., K. Ko, and D. Hong. 2001. “A multilayer feedforward neural net-
work having N/4 nodes in two hidden layers.” In Vol. 3 of Proc., IEEE
Int. Joint Conf. on Neural Networks, 1675–1680. New York: IEEE.
Chou, C.-H. 2006. “Genetic algorithm-based optimal fuzzy controller de-
sign in the linguistic space.” IEEE Trans. Fuzzy Syst. 14 (3): 372–385.
https://doi.org/10.1109/TFUZZ.2006.876329.
Chou, J. S., and C. Lin. 2012. “Predicting disputes in public-private part-
nership projects: Classification and ensemble models.” J. Comput.
Civ. Eng. 27 (1): 51–60. https://doi.org/10.1061/(ASCE)CP.1943-5487
.0000197.
Chou, J.-S. 2009. “Web-based CBR system applied to early cost budget-
ing for pavement maintenance project.” Expert Syst. Appl. 36 (2):
2947–2960. https://doi.org/10.1016/j.eswa.2008.01.025.
Chou, J.-S., C.-W. Lin, A.-D. Pham, and J.-Y. Shao. 2015. “Optimized
artificial intelligence models for predicting project award price.” Autom.
Constr. 54 (Jun): 106–115. https://doi.org/10.1016/j.autcon.2015
.02.006.
Comrey, A. L., and H. B. Lee. 1992. A first course in factor analysis.
2nd ed. Hillsdale, NJ: Erlbaum.
Cook, R. D., and S. Weisberg. 1982. Residuals and influence in regression.
New York: Chapman & Hall.
Cortes, C., and V. Vapnik. 1995. “Support-vector networks.” Mach. Learn.
20 (3): 273–297.
Curram, S. P., and J. Mingers. 2017. “Neural networks, decision tree in-
duction and discriminant analysis: An empirical comparison.” J. Oper.
Res. Soc. 45 (4): 440–450. https://doi.org/10.1057/jors.1994.62.
Darwin, C. 1859. The origin of species by means of natural selection or the
preservation of favoured races in the struggle for life. New York:
Mentor.
Davis, F. D. 1993. “User acceptance of information technology: System
characteristics, user perceptions, and behavioral impacts.” Int. J. Man
Mach. Stud. 38 (3): 475–487. https://doi.org/10.1006/imms.1993.1022.
Dell’Isola, M. D. 2002. Architect’s essentials of cost management.
New York: Wiley.
De Soto, B. G., and B. T. Adey. 2015. “Investigation of the case-based
reasoning retrieval process to estimate resources in construction proj-
ects.” Procedia Eng. 123: 169–181. https://doi.org/10.1016/j.proeng
.2015.10.074.
Dietterich, T. G. 2000. “An experimental comparison of three methods for
constructing ensembles of decision trees: Bagging, boosting, and ran-
domization.” Mach. Learn. 40 (2): 139–157. https://doi.org/10.1023
/A:1007607513941.
Doğan, S. Z., D. Arditi, and H. M. Günaydin. 2008. “Using decision trees
for determining attribute weights in a case-based model of early cost
prediction.” J. Constr. Eng. Manage. 134 (2): 146–152. https://doi.org
/10.1061/(ASCE)0733-9364(2008)134:2(146).
Draper, N. R., and H. Smith. 1998. Applied regression analysis. New York:
Wiley.
Durbin, J., and G. S. Watson. 1951. “Testing for serial correlation in least
squares regression, II.” Biometrika 38 (1–2): 159–178. https://doi.org
/10.1093/biomet/38.1-2.159.
Dursun, O., and C. Stoy. 2016. “Conceptual estimation of construction
costs using the multistep ahead approach.” J. Constr. Eng. Manage.
142 (9): 04016038. https://doi.org/10.1061/(ASCE)CO.1943-7862
.0001150.
Dziuban, C. D., and E. C. Shirkey. 1974. “When is a correlation matrix
appropriate for factor analysis? Some decision rules.” Psychol. Bull.
81 (6): 358–361. https://doi.org/10.1037/h0036316.
Elbeltagi, E., O. Hosny, R. Abdel-Razek, and A. El-Fitory. 2014. “Concep-
tual cost estimate of Libyan highway projects using artificial neural
network.” Int. J. Eng. Res. Appl. 4 (8): 56–66.
Elfaki, A. O., S. Alatawi, and E. Abushandi. 2014. “Using intelligent
techniques in construction project cost estimation: 10-year survey.”
Adv. Civ. Eng. 2014: 107926. https://doi.org/10.1155/2014/107926.
J. Constr. Eng. Manage.
 Elmousalami, H. H. 2019. “Intelligent methodology for project conceptual
cost prediction.” Heliyon 5 (5): e01625. https://doi.org/10.1016/j
.heliyon.2019.e01625.
Elmousalami, H. H., A. H. Elyamany, and A. H. Ibrahim. 2018a. “Evalu-
ation of cost drivers for field canals improvement projects.” Water
Resour. Manage. 32 (1): 53–65. https://doi.org/10.1007/s11269-017
-1747-x.
ElMousalami, H. H., A. H. Elyamany, and A. H. Ibrahim. 2018b. “Predict-
ing conceptual cost for field canal improvement projects.” J. Constr.
Eng. Manage. 144 (11): 04018102. https://doi.org/10.1061/(ASCE)CO
.1943-7862.0001561.
El-Sawah, H., and O. Moselhi. 2014. “Comparative study in the use of neu-
ral networks for order of magnitude cost estimating in construction.”
ITcon 19: 462–473.
El Sawalhi, N. I. 2012. “Modeling the parametric construction project cost
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
estimate using fuzzy logic.” Int. J. Emerging Technol. Adv. Eng. 2 (4):
631–636.
El-Sawalhi, N. I., and O. Shehatto. 2014. “A neural network model for
building construction projects cost estimating.” J. Constr. Eng. Project
Manage. 4 (4): 9–16. https://doi.org/10.6106/JCEPM.2014.4.4.009.
ElSawy, I., H. Hosny, and M. Abdel Razek. 2011. “A neural network model
for construction projects site overhead cost estimating in Egypt.” Int. J.
Comput. Sci. Issues 8 (3): 273–283.
Emami, M. R., I. B. Turksen, and A. A. Goldberg. 1998. “Development of a
systematic methodology of fuzzy logic modeling.” IEEE Trans. Fuzzy
Syst. 6 (3): 346–361. https://doi.org/10.1109/91.705501.
Emsley, M. W., D. J. Lowe, A. R. Duff, A. Harding, and A. Hickson. 2002.
“Data modeling and the application of a neural network approach to the
prediction of total construction costs.” Constr. Manage. Econ. 20 (6):
465–472. https://doi.org/10.1080/01446190210151050.
Engelbrecht, A. P. 2002. Computational intelligence: An introduction.
New York: Wiley.
Erensal, Y. C., T. Öncan, and M. L. Demircan. 2006. “Determining key
capabilities in technology management using fuzzy analytic hierarchy
process: A case study of Turkey.” Inf. Sci. 176 (18): 2755–2770. https://
doi.org/10.1016/j.ins.2005.11.004.
Fan, G. Z., S. E. Ong, and H. C. Koh. 2006. “Determinants of house price:
A decision tree approach.” Urban Stud. 43 (12): 2301–2315. https://doi
.org/10.1080/00420980600990928.
Fan, R. E., K. W. Chang, C. J. Hsieh, X. R. Wang, and C. J. Lin. 2008.
“LIBLINEAR: A library for large linear classification.” J. Mach. Learn.
Res. 9 (Aug): 1871–1874.
Field, A. 2009. Discovering statistics using SPSS for windows. London:
Sage Publications.
Flom, P. L., and D. L. Cassell. 2007. “Stopping stepwise: Why stepwise and
similar selection methods are bad, and what you should use.” In Proc.,
2007 Conf. NorthEast SAS Users Group (NESUG): Statistics and Data
Analysis. Portland, OR: NorthEast SAS Users Group.
Freund, Y., and R. E. Schapire. 1996. “Experiments with a new boosting
algorithm.” In Proc., 13th Int. Conf. on Machine Learning. Princeton,
NJ: International Machine Learning Society.
Freund, Y., and R. E. Schapire. 1997. “A decision-theoretic generalization
of on-line learning and an application to boosting.” J. Comput. Syst. Sci.
55 (1): 119–139. https://doi.org/10.1006/jcss.1997.1504.
Friedman, J., T. Hastie, and R. Tibshirani. 2000. “Additive logistic regres-
sion: A statistical view of boosting (with discussion and a rejoinder by
the authors).” Ann. Stat. 28 (2): 337–407. https://doi.org/10.1214/aos
/1016218223.
Friedman, J., T. Hastie, and R. Tibshirani. 2001. Vol. 1 of The elements of
statistical learning. New York: Springer.
Gardner, B. J., D. D. Gransberg, and H. D. Jeong. 2016. “Reducing data-
collection efforts for conceptual cost estimating at a highway agency.”
J. Constr. Eng. Manage. 142 (11): 04016057. https://doi.org/10.1061
/(ASCE)CO.1943-7862.0001174.
Geurts, P., D. Ernst, and L. Wehenkel. 2006. “Extremely randomized
trees.” Mach. Learn. 63 (1): 3–42. https://doi.org/10.1007/s10994-006
-6226-1.
Green, B. N., C. D. Johnson, and A. Adams. 2006. “Writing narrative
literature reviews for peer-reviewed journals: Secrets of the trade.”
© ASCE 03119008-27
J. Constr. Eng. Manage., 2020, 146(1): 03119008
J. Chiropr. Med. 5 (3): 101–117. https://doi.org/10.1016/S0899-3467
(07)60142-6.
Green, S. B. 1991. “How many subjects does it take to do a regression
analysis?” Multivariate Behav. Res. 26 (3): 499–510. https://doi.org/10
.1207/s15327906mbr2603_7.
Guadagnoli, E., and W. F. Velicer. 1988. “Relation of sample size to the
stability of component patterns.” Psychol. Bull. 103 (2): 265–275.
https://doi.org/10.1037/0033-2909.103.2.265.
Günaydın, H. M., and S. Z. Doğan. 2004. “A neural network approach for
early cost estimation of structural systems of buildings.” Int. J. Project
Manage. 22 (7): 595–602. https://doi.org/10.1016/j.ijproman.2004
.04.002.
Guyon, I., and A. Elisseeff. 2003. “An introduction to variable and feature
selection.” J. Mach. Learn. Res. 3 (Mar): 1157–1182.
Hansen, L. K., and P. Salamon. 1990. “Neural network ensembles.” IEEE
Trans. Pattern Anal. Mach. Intell. 12 (10): 993–1001. https://doi.org/10
.1109/34.58871.
Hastie, T., R. Tibsharani, and J. Friedman. 2009. The elements of statistical
learning: Data mining, inference, and prediction. 2nd ed. New York:
Springer. https://doi.org/10.1007/b94608.
Hays, W. L. 1983. “Review of using multivariate statistics. [Review of the
book Using Multivariate Statistics. B. G. Tabachnick & L. S. Fidell].”
Contemp. Psychol. 28 (8): 642. https://doi.org/10.1037/022267.
Hegazy, T. 2014. Computer-based construction project management.
Essex: Pearson Education.
Hegazy, T., and A. Ayed. 1998. “Neural network model for parametric
cost estimation of highway projects.” J. Constr. Eng. Manage. 124 (3):
210–218. https://doi.org/10.1061/(ASCE)0733-9364(1998)124:3(210).
Holland, J. H. 1975. Adaptation in natural and artificial systems. Ann
Arbor, MI: University Michigan Press.
HongWei, M. 2009. “An improved support vector machine based on rough
set for construction cost prediction.” In Vol. 2 of Proc., 2009 Int. Forum
on Computer Science-Technology and Applications. New York: IEEE.
Hopfield, J. J. 1982. “Neural networks and physical systems with emergent
collective computational abilities.” Proc. Natl. Acad. Sci. 79 (8): 2554–
2558. https://doi.org/10.1073/pnas.79.8.2554.
Hsiao, F.-Y., S.-H. Wang, W.-C. Wang, C.-P. Wen, and W.-D. Yu. 2012.
“Neuro-fuzzy cost estimation model enhanced by fast messy genetic
algorithms for semiconductor hookup construction.” Comput.-Aided
Civ. Infrastruct. Eng. 27 (10): 764–781. https://doi.org/10.1111/j.1467
-8667.2012.00786.x.
Hsu, C.-C., and B. A. Sandford. 2007. “The Delphi technique: Making
sense of consensus.” Pract. Assess. Res. Eval. 12 (10): 1–8.
Hsu, Y.-L., C.-H. Lee, and V. Kreng. 2010. “The application of fuzzy
Delphi method and fuzzy AHP in lubricant regenerative technology
selection.” Expert Syst. Appl. 37 (1): 419–425. https://doi.org/10
.1016/j.eswa.2009.05.068.
Huang, S.-C., and Y.-F. Huang. 1991. “Bounds on the number of hidden
neurons in multilayer neurons.” IEEE Trans. Neural Networks 2 (1):
47–55. https://doi.org/10.1109/72.80290.
Hutcheson, G., and N. Sofroniou. 1999. The multivariate social scientist.
London: Sage.
Ilbeigi, M., B. Ashuri, and A. Joukar. 2016. “Time-series analysis for fore-
casting asphalt-cement price.” J. Manage. Eng. 33 (1): 04016030.
https://doi.org/10.1061/(ASCE)ME.1943-5479.0000477.
Ishikawa, A., M. Amagasa, T. Shiga, G. Tomizawa, R. Tatsuta, and
H. Mieno. 1993. “The max-min Delphi method and fuzzy Delphi
method via fuzzy integration.” Fuzzy Sets Syst. 55 (3): 241–253.
https://doi.org/10.1016/0165-0114(93)90251-C.
Ji, S.-H., J. Ahn, E.-B. Lee, and Y. Kim. 2018. “Learning method for
knowledge retention in CBR cost models.” Autom. Constr. 96 (Dec):
65–74. https://doi.org/10.1016/j.autcon.2018.08.019.
Ji, S.-H., M. Park, and H.-S. Lee. 2012. “Case adaptation method of case-
based reasoning for construction cost estimation in Korea.” J. Constr.
Eng. Manage. 138 (1): 43–52. https://doi.org/10.1061/(ASCE)CO.1943
-7862.0000409.
Jin, R., K. Cho, C. Hyun, and M. Son. 2012. “MRA-based revised CBR
model for cost prediction in the early stage of construction projects.”
Expert Syst. Appl. 39 (5): 5214–5222. https://doi.org/10.1016/j.eswa
.2011.11.018.
J. Constr. Eng. Manage.
 Jolliffe, I. T. 1972. “Discarding variables in a principal component analysis.
I: Artificial data.” Appl. Stat. 21 (2): 160–173. https://doi.org/10.2307
/2346488.
Jolliffe, I. T. 1986. Principal component analysis. New York: Springer.
Jrade, A. 2000. “A conceptual cost estimating computer system for building
projects.” Masters thesis, Dept. of Building Civil and Environmental
Engineering, Concordia Univ.
Juszczyk, M. 2017. “Studies on the ANN implementation in the macro
BIM cost analyzes.” Przegląd Naukowy. Inżynieria i Kształtowanie
Środowiska 26 (2): 183–192.
Juszczyk, M., A. Leśniak, and K. Zima. 2018. “ANN based approach
for estimation of construction costs of sports fields.” Complexity 2018:
1–11. https://doi.org/10.1155/2018/7952434.
Kaiser, H. F. 1960. “The application of electronic computers to factor
analysis.” Educ. Psychol. Meas. 20 (1): 141–151. https://doi.org/10
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
.1177/001316446002000116.
Kaiser, H. F. 1970. “A second generation little jiffy.” Psychometrika 35 (4):
401–415. https://doi.org/10.1007/BF02291817.
Kaiser, H. F. 1974. “An index of factorial simplicity.” Psychometrika 39 (1):
31–36. https://doi.org/10.1007/BF02291575.
Kan, P. 2002. “Parametric cost estimating model for conceptual cost esti-
mating of building construction projects.” Ph.D. thesis, Faculty of the
Graduate School, Univ. of Texas.
Karatas, Y., and F. Ince. 2016. “Feature article: Fuzzy expert tool for small
satellite cost estimation.” IEEE Aerosp. Electron. Syst. Mag. 31 (5):
28–35. https://doi.org/10.1109/MAES.2016.140210.
Kass, R. A., and H. E. A. Tinsley. 1979. “Factor analysis.” J. Leisure Res.
11 (4): 120–138.
Kim, G. H., D. S. Seo, and K. I. Kang. 2005. “Hybrid models of neural
networks and genetic algorithms for predicting preliminary cost esti-
mates.” J. Comput. Civ. Eng. 19 (2): 208–211. https://doi.org/10.1061
/(ASCE)0887-3801(2005)19:2(208).
Kim, G.-H., S.-H. An, and K.-I. Kang. 2004. “Comparison of construction
cost estimating models based on regression analysis, neural networks,
and case-based reasoning.” Build. Environ. 39 (10): 1235–1242. https://
doi.org/10.1016/j.buildenv.2004.02.013.
Kim, G.-H., J.-M. Shin, S. Kim, and Y. Shin. 2013. “Comparison of school
building construction costs estimation methods using regression analy-
sis, neural network, and support vector machine.” J. Build. Constr.
Plann. Res. 1 (1): 1–7. https://doi.org/10.4236/jbcpr.2013.11001.
Kim, K. J., and K. Kim. 2010. “Preliminary cost estimation model
using case-based reasoning and genetic algorithms.” J. Comput. Civ.
Eng. 24 (6): 499–505. https://doi.org/10.1061/(ASCE)CP.1943-5487
.0000054.
Kim, S. 2013. “Hybrid forecasting system based on case-based reasoning
and analytic hierarchy process for cost estimation.” J. Civ. Eng. Manage.
19 (1): 86–96. https://doi.org/10.3846/13923730.2012.737829.
Kim, S., S. Chin, and S. Kwon. 2019. “A discrepancy analysis of BIM-
based quantity take-off for building interior components.” J. Manage.
Eng. 35 (3): 05019001. https://doi.org/10.1061/(ASCE)ME.1943-5479
.0000684.
Kline, P. 1999. The handbook of psychological testing. 2nd ed. London:
Routledge.
Klir, G. J., and B. Yuan. 1995. Fuzzy sets and fuzzy logic theory and
applications. Upper Saddle River, NJ: Prentice Hall.
Knight, K., and A. R. Fayek. 2002. “Use of fuzzy logic for predicting
design cost overruns on building projects.” J. Constr. Eng. Manage.
128 (6): 503–512. https://doi.org/10.1061/(ASCE)0733-9364(2002)
128:6(503).
Kohavi, R., and G. John. 1997. “Wrappers for feature subset selection.”
Artif. Intell. 97 (1/2): 273–324. https://doi.org/10.1016/S0004-3702
(97)00043-X.
Kolodner, J. L. 1992. “An introduction to case-based reasoning.” Artif.
Intell. Rev. 6 (1): 3–34. https://doi.org/10.1007/BF00155578.
Kuncheva, L. I. 2004. Combining pattern classifiers: Methods and
algorithms. New York: Wiley.
Kursa, M., and W. Rudnicki. 2010. “Feature selection with the Boruta
package.” J. Stat. Software 36 (11): 1–13.
© ASCE 03119008-28
J. Constr. Eng. Manage., 2020, 146(1): 03119008
Laarhoven, P. J. M., and W. Pedrycz. 1983. “A fuzzy extension of Sati’s
priority theory.” Fuzzy Sets Syst. 11 (1–3): 229–241. https://doi.org/10
.1016/S0165-0114(83)80082-7.
LeCun, Y., Y. Bengio, and G. Hinton. 2015. “Deep learning.” Nature
521 (7553): 436. https://doi.org/10.1038/nature14539.
Leśniak, A., and K. Zima. 2018. “Cost calculation of construction proj-
ects including sustainability factors using the case based reasoning
(CBR) method.” Sustainability 10 (5): 1608. https://doi.org/10.3390
/su10051608.
Lewis, C. D. 1982. Industrial and business forecasting methods. London:
Butterworth.
Lin, X., F. Yang, L. Zhou, P. Yin, H. Kong, W. Xing, X. Lu, L. Jia,
Q. Wang, and G. Xu. 2012. “A support vector machine-recursive feature
elimination feature selection method based on artificial contrast varia-
bles and mutual information.” J. Chromatogr. B 910: 149–155. https://
doi.org/10.1016/j.jchromb.2012.05.020.
Liu, W.-K. 2013. “Application of the fuzzy Delphi method and the fuzzy
analytic hierarchy process for the managerial competence of multina-
tional corporation executives.” IJEEEE 3 (4): 313–317. https://doi.org
/10.7763/IJEEEE.2013.V3.248.
Loop, B. P., S. D. Sudhoff, S. H. Zak, and E. L. Zivi. 2010. “Estimating
regions of asymptotic stability of power electronics systems using
genetic algorithms.” IEEE Trans. Control Syst. Technol. 18 (5):
1011–1022. https://doi.org/10.1109/TCST.2009.2031325.
Love, P. E. D., R. Y. C. Tse, and D. J. Edwards. 2005. “Time–cost relation-
ships in Australian building construction projects.” J. Constr. Eng.
Manage. 131 (2): 187–194. https://doi.org/10.1061/(ASCE)0733-9364
(2005)131:2(187).
Lowe, D. J., M. W. Emsley, and A. Harding. 2006. “Predicting construction
cost using multiple regression techniques.” J. Constr. Eng. Manage.
132 (7): 750–758. https://doi.org/10.1061/(ASCE)0733-9364(2006)
132:7(750).
Ma, L., S. Shen, J. Zhang, Y. Huang, and F. Shi. 2010. “Application of
fuzzy analytic hierarchy process model on determination of optimized
pile-type.” Front. Archit. Civ. Eng. China 4 (2): 252–257. https://doi
.org/10.1007/s11709-010-0017-2.
MacCallum, R. C., K. F. Widaman, S. Zhang, and S. Hong. 1999. “Sample
size in factor analysis.” Psychol. Methods 4 (1): 84–99. https://doi.org
/10.1037/1082-989X.4.1.84.
Makridakis, S., S. C. Wheelwright, and R. J. Hyndman. 1998. Forecasting
methods and applications. New York: Wiley.
Mamdani, E. H., and S. Assilian. 1974. “Application of fuzzy algorithms
for control of simple dynamic plant.” Proc., Institution of Electrical
Engineers 121 (12), 1585–1588.
Manoliadis, O. G., J. P. Pantouvakis, and S. E. Christodoulou. 2009.
“Improving qualifications-based selection by use of the fuzzy Delphi
method.” Constr. Manage. Econ. 27 (4): 373–384. https://doi.org/10
.1080/01446190902758993.
Marzouk, M., and M. Alaraby. 2014. “Predicting telecommunication tower
costs using fuzzy subtractive clustering.” J. Civ. Eng. Manage. 21 (1):
67–74. https://doi.org/10.3846/13923730.2013.802736.
Marzouk, M., and A. Amin. 2013. “Predicting construction materials
prices using fuzzy logic and neural networks.” J. Constr. Eng. Manage.
139 (9): 1190–1198. https://doi.org/10.1061/(ASCE)CO.1943-7862
.0000707.
Marzouk, M., and M. Elkadi. 2016. “Estimating water treatment plants
costs using factor analysis and artificial neural networks.” J. Clean.
Prod. 112 (Part 5): 4540–4549. https://doi.org/10.1016/j.jclepro.2015
.09.015.
Marzouk, M. M., and R. M. Ahmed. 2011. “A case-based reasoning ap-
proach for estimating the costs of pump station projects.” J. Adv. Res.
2 (4): 289–295. https://doi.org/10.1016/j.jare.2011.01.007.
McCulloch, W. S., and W. H. Pitts. 1943. “A logical calculus of the ideas
imminent in nervous activity.” Bull. Math. Biophys. 5 (4): 115–133.
https://doi.org/10.1007/BF02478259.
Moselhi, O., and T. Hegazy. 1993. “Markup estimation using neural net-
work methodology.” Comput. Syst. Eng. 4 (2–3): 135–145. https://doi
.org/10.1016/0956-0521(93)90039-Y.
Moussa, M., J. Ruwanpura, and G. Jergeas. 2006. “Decision tree modeling
using integrated multilevel stochastic networks.” J. Constr. Eng.
J. Constr. Eng. Manage.
 Manage. 132 (12): 1254–1266. https://doi.org/10.1061/(ASCE)0733
-9364(2006)132:12(1254).
Myers, R. 1990. Classical and modern regression with applications.
2nd ed. Boston: Duxbury.
Nair, V., and G. E. Hinton. 2010. “Rectified linear units improve restricted
Boltzmann machines.” In Proc., 27th Int. Conf. on Machine Learning
(ICML-10), 807–814. Madison, WI: Omnipress.
Nasrazadani, H., M. Mahsuli, H. Talebiyan, and H. Kashani. 2017. “Prob-
abilistic modeling framework for prediction of seismic retrofit cost of
buildings.” J. Constr. Eng. Manage. 143 (8): 04017055. https://doi.org
/10.1061/(ASCE)CO.1943-7862.0001354.
Nunnally, J. C. 1978. Psychometric theory. New York: McGraw-Hill.
Ogungbile, A. J., A. E. Oke, and K. Rasak. 2018. “Developing cost model
for preliminary estimate of road projects in Nigeria.” Int. J. Sustainable
Real Estate Constr. Econ. 1 (2): 182–199. https://doi.org/10.1504
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
/IJSRECE.2018.092277.
Opitz, D., and R. Maclin. 1999. “Popular ensemble methods: An empirical
study.” J. Artif. Intell. Res. 11: 169–198. https://doi.org/10.1613
/jair.614.
Ostertagová, E. 2011. Applied statistics. [In Slovak.] Sever, Slovakia:
Elfa Košice.
Ozdemir, M., M. J. Embrechts, F. Arciniegas, C. M. Breneman, L.
Lockwood, and K. P. Bennett. 2001. “Feature selection for in-silico
drug design using genetic algorithms and neural networks.” In Proc.,
IEEE Mountain Workshop on Soft Computing in Industrial Applica-
tions, 53–57. New York: IEEE.
Pan, N.-F. 2008. “Fuzzy AHP approach for selecting the suitable bridge
construction method.” Autom. Constr. 17 (8): 958–965. https://doi
.org/10.1016/j.autcon.2008.03.005.
Papadopoulos, B., K. P. Tsagarakis, and A. Yannopoulos. 2007. “Cost and
land functions for wastewater treatment projects: Typical simple linear
regression versus fuzzy linear regression.” J. Environ. Eng. 133 (6):
581–586. https://doi.org/10.1061/(ASCE)0733-9372(2007)133:6(581).
Park, H.-S., and S. Kwon. 2011. “Factor analysis of construction prac-
tices for infrastructure projects in Korea.” KSCE J. Civ. Eng. 15 (3):
439–445. https://doi.org/10.1007/s12205-011-1064-5.
Pedrycz, W., ed. 1997. Fuzzy evolutionary computation. Dordrecht,
Netherlands: Kluwer Academic Publishers.
Perera, S., and I. Watson. 1998. “Collaborative case-based estimating and
design.” Adv. Eng. Software 29 (10): 801–808. https://doi.org/10.1016
/S0965-9978(97)00064-1.
Perner, P., U. Zscherpel, and C. Jacobsen. 2001. “A comparison between
neural networks and decision trees based on data from industrial radio-
graphic testing.” Pattern Recognit. Lett. 22 (1): 47–54. https://doi.org
/10.1016/S0167-8655(00)00098-2.
Petroutsatou, K., E. Georgopoulos, S. Lambropoulos, and J. P. Pantouvakis.
2012. “Early cost estimating of road tunnel construction using neural
networks.” J. Constr. Eng. Manage. 138 (6): 679–687. https://doi.org
/10.1061/(ASCE)CO.1943-7862.0000479.
Petruseva, S., P. Sherrod, V. Z. Pancovska, and A. Petrovski. 2016.
“Predicting bidding price in construction using support vector
machine.” TEM J. 5 (2): 143–151.
Petruseva, S., V. Zileska-Pancovska, V. Žujo, and A. Brkan-Vejzović. 2017.
“Construction costs forecasting: Comparison of the accuracy of linear
regression and support vector machine models.” Tehnicki Vjesnik-
Technical Gaz. 24 (5): 1431–1438.
Peurifoy, R. L., and G. D. Oberlender. 2002. Estimating construction costs.
5th ed. New York: McGraw-Hill.
Polit, D. F., and C. T. Beck. 2012. Nursing research: Generating and
assessing evidence for nursing practice. 9th ed. Philadelphia:
Wolters Kluwer Health, Lippincott Williams & Wilkins.
Prasad, A. M., L. R. Iverson, and A. Liaw. 2006. “Newer classification and
regression tree techniques: Bagging and random forests for ecological
prediction.” Ecosystems 9 (2): 181–199. https://doi.org/10.1007/s10021
-005-0054-1.
Quinlan, J. R. 2014. C4.5: Programs for machine learning. San Mateo,
CA: Morgan Kaufmann.
Radwan, H. G. 2013. “Sensitivity analysis of head loss equations on the
design of improved irrigation on-farm system in Egypt.” Int. J. Adv.
Res. Technol. 2 (1): 1–9.
© ASCE 03119008-29
J. Constr. Eng. Manage., 2020, 146(1): 03119008
Rafiei, M. H., and H. Adeli. 2015. “A novel machine learning model for
estimation of sale prices of real estate units.” J. Constr. Eng. Manage.
142 (2): 04015066. https://doi.org/10.1061/(ASCE)CO.1943-7862
.0001047.
Ranasinghe, M. 2000. “Impact of correlation and induced correlation on the
estimation of project cost of buildings.” Constr. Manage. Econ. 18 (4):
395–406. https://doi.org/10.1080/01446190050024815.
Ratner, B. 2010. “Variable selection methods in regression: Ignorable prob-
lem, outing notable solution.” J. Targeting Meas. Anal. Marketing
18 (1): 65–75. https://doi.org/10.1057/jt.2009.26.
Reid, S. 2007. A review of heterogeneous ensemble methods. Boulder, CO:
Dept. of Computer Science, Univ. of Colorado at Boulder.
Rockwell, R. C. 1975. “Assessment of multicollinearity: The Haitovsky test
of the determinant.” Sociological Methods Res. 3 (3): 308–320. https://
doi.org/10.1177/004912417500300304.
Ross, B. H. 1989. “Some psychological results on case-based reasoning.”
In Proc., Case-Based Reasoning Workshop, 144–147. Burlington, MA:
Morgan Kaufmann.
Runker, T. A. 1997. “Selection of appropriate defuzzification methods
using application specific properties.” IEEE Trans. Fuzzy Syst. 5 (1):
72–79.
Rutkowski, L. 2005. New soft computing techniques for system modeling,
pattern classification and image processing. Heidelberg: Springer.
Saaty, T. L. 1980. The analytic hierarchy process: Planning, priority
setting. New York: McGraw Hill International Book.
Saaty, T. L. 1994. “How to make a decision: The analytic hierarchy pro-
cess.” Interfaces 24 (6): 19–43. https://doi.org/10.1287/inte.24.6.19.
Saaty, T. L. 2008. “Decision making with the analytic hierarchy process.”
Int. J. Serv. Sci. 1 (1): 83. https://doi.org/10.1504/IJSSCI.2008.017590.
Schapire, R. E. 1990. “The strength of weak learnability.” Mach. Learn.
5 (2): 197–227.
Schapire, R. E., Y. Freund, P. Bartlett, and W. S. Lee. 1998. “Boosting
the margin: A new explanation for the effectiveness of voting methods.”
Ann. Stat. 26 (5): 1651–1686. https://doi.org/10.1214/aos/1024691352.
Shaheen, A. A., A. R. Fayek, and S. M. Abourizk. 2007. “Fuzzy numbers
in cost range estimating.” J. Constr. Eng. Manage. 133 (4): 325–334.
https://doi.org/10.1061/(ASCE)0733-9364(2007)133:4(325).
Shi, H., J. Song, and X. Zhang. 2010. “The method and application for
estimating construction project costs.” In 2010 IEEE Int. Conf. on
Advanced Management Science (ICAMS 2010). New York: IEEE.
Siddique, N., and H. Adeli. 2013. Computational intelligence: Synergies of
fuzzy logic, neural networks and evolutionary computing. Chichester,
UK: Wiley.
Siedlecki, W., and J. Sklansky. 1988. “On automatic feature selection.”
Int. J. Pattern Recognit. Artif. Intell. 2 (2): 197–220. https://doi.org/10
.1142/S0218001488000145.
Siedlecki, W., and J. Sklansky. 1989. “A note on genetic algorithms for
large-scale feature selection.” In Handbook of pattern recognition and
computer vision, edited by C. H. Chen, L. F. Pau, and P. S. P. Wang,
88–107. Singapore: World Scientific.
Son, H., C. Kim, and C. Kim. 2012. “Hybrid principal component analysis
and support vector machine model for predicting the cost performance
of commercial building projects using pre-project planning variables.”
Autom. Constr. 27 (Nov): 60–66. https://doi.org/10.1016/j.autcon.2012
.05.013.
Srichetta, P., and W. Thurachon. 2012. “Applying fuzzy analytic hierarchy
process to evaluate and select product of notebook computers.” Int. J.
Model. Optim. 2 (2): 168–173. https://doi.org/10.7763/IJMO.2012
.V2.105.
Stevens, J. P. 2002. Applied multivariate statistics for the social sciences.
4th ed. Hillsdale, NJ: Erlbaum.
Stoy, C., S. Pollalis, and O. Dursun. 2012. “A concept for developing
construction element cost models for German residential building proj-
ects.” Int. J. Project Organ. Manage. 4 (1): 38. https://doi.org/10.1504
/IJPOM.2012.045363.
Stoy, C., S. Pollalis, and H.-R. Schalcher. 2008. “Drivers for cost estimating
in early design: Case study of residential construction.” J. Constr. Eng.
Manage. 134 (1): 32–39. https://doi.org/10.1061/(ASCE)0733-9364
(2008)134:1(32).
J. Constr. Eng. Manage.
 Stoy, C., and H.-R. Schalcher. 2007. “Residential building projects:
Building cost indicators and drivers.” J. Constr. Eng. Manage. 133 (2):
139–145. https://doi.org/10.1061/(ASCE)0733-9364(2007)133:2(139).
Sugeno, M., and G. T. Kang. 1988. “Structure identification of fuzzy
model.” Fuzzy Sets Syst. 28 (1): 15–33. https://doi.org/10.1016/0165
-0114(88)90113-3.
Tabachnick, B. G., and L. S. Fidell. 2007. Using multivariate statistics.
5th ed. Boston: Allyn & Bacon.
Takagi, T., and M. Sugeno. 1985. “Fuzzy identification of systems and its
application to modeling and control.” IEE Trans. Syst. Man Cybern.
15 (1): 116–132. https://doi.org/10.1109/TSMC.1985.6313399.
Tokede, O., D. Ahiaga-Dagbui, S. Smith, and S. Wamuziri. 2014.
“Mapping relational efficiency in neuro-fuzzy hybrid cost models.”
In Construction Research Congress 2014: Construction in a Global
Network, 1458–1467. Reston, VA: ASCE.
Downloaded from ascelibrary.org by haytham elmousalami on 10/19/19. Copyright ASCE. For personal use only; all rights reserved.
Tsukamoto, Y. 1979. “An approach to fuzzy reasoning method.” In Advan-
ces in fuzzy set theory and applications, edited by M. M. Gupta,
R. K. Ragade, and R. Yager, 137–149. Amsterdam, Netherlands:
North-Holland.
Vaidya, O. S., and S. Kumar. 2006. “Analytic hierarchy process: An over-
view of applications.” Eur. J. Oper. Res. 169 (1): 1–29. https://doi.org
/10.1016/j.ejor.2004.04.028.
Vapnik, V. 1979. Estimation of dependences based on empirical data.
[In Russian.] Moscow: Nauka.
Walker, E. 1989. “Applied regression analysis and other multivariable
methods.” Technometrics 31 (1): 117–118. https://doi.org/10.1080
/00401706.1989.10488486.
Wang, J., and B. Ashuri. 2016. “Predicting ENR construction cost index
using machine learning algorithms.” Int. J. Constr. Educ. Res. 13 (1):
47–63. https://doi.org/10.1080/15578771.2016.1235063.
Wang, L.-X. 1997. Adaptive fuzzy systems and control: Design and stability
analysis. Englewood Cliffs, NJ: Prentice-Hall.
Wang, Y.-R., C.-Y. Yu, and H.-H. Chan. 2012. “Predicting construction cost
and schedule success using artificial neural networks ensemble and
support vector machines classification models.” Int. J. Project Manage.
30 (4): 470–478. https://doi.org/10.1016/j.ijproman.2011.09.002.
Waty, M., S. W. Alisjahbana, O. Gondokusumo, H. Sulistio, C. Hasyim,
M. I. Setiawan, D. Harmanto, and A. S. Ahmar. 2018. “Modeling of
waste material costs on road construction projects.” Int. J. Eng. Technol.
7 (24): 474–477. https://doi.org/10.14419/ijet.v7i2.11250.
Wheaton, W. C., and W. E. Simonton. 2007. “The secular and cyclic behav-
ior of true construction costs.” J. Real Estate Res. 29 (1): 1–26.
Wilkinson, L., and G. E. Dallal. 1981. “Tests of significance in forward
selection regression with an F-to-enter stopping rule.” Technometrics
23 (4): 377–380.
Williams, T. P. 2002. “Predicting completed project cost using bidding
data.” Constr. Manage. Econ. 20 (3): 225–235. https://doi.org/10.1080
/01446190110112838.
Williams, T. P., and J. Gong. 2014. “Predicting construction cost overruns
using text mining, numerical data and ensemble classifiers.” Autom.
Constr. 43 (1): 23–29. https://doi.org/10.1016/j.autcon.2014.02.014.
© ASCE 03119008-30
View publication stats J. Constr. Eng. Manage., 2020, 146(1): 03119008
Wilmot, C. G., and B. Mei. 2005. “Neural network modeling of highway
construction costs.” J. Constr. Eng. Manage. 131 (7): 765–771. https://
doi.org/10.1061/(ASCE)0733-9364(2005)131:7(765).
Woldesenbet, A., and D. H. S. Jeong. 2012. “Historical data driven and
component based prediction models for predicting preliminary engi-
neering costs of roadway projects.” In Construction Research Congress
2012. Reston, VA: ASCE.
Wu, X., et al.2008. “Top 10 algorithms in data mining.” Knowl. Inf. Syst.
14 (1): 1–37. https://doi.org/10.1007/s10115-007-0114-2.
Xu, M., B. Xu, L. Zhou, and L. Wu. 2015. “Construction project cost
prediction based on genetic algorithm and least squares support vector
machine.” In Proc., 5th Int. Conf. on Civil Engineering and Transpor-
tation 2015. Paris: Atlantis.
Yang, I.-T. 2005. “Simulation-based estimation for correlated cost ele-
ments.” Int. J. Project Manage. 23 (4): 275–282. https://doi.org/10
.1016/j.ijproman.2004.12.002.
Yang, J., and V. Honavar. 1998. “Feature subset selection using a genetic
algorithm.” IEEE Intell. Syst. 13 (2): 44–49. https://doi.org/10.1109
/5254.671091.
Yang, S.-S., and J. Xu. 2010. “The application of fuzzy system method to
the cost estimation of construction works.” In Proc., 2010 Int. Conf. on
Machine Learning and Cybernetics. New York: IEEE.
Yu, W.-D., and M. J. Skibniewski. 2010. “Integrating neurofuzzy system
with conceptual cost estimation to discover cost-related knowledge
from residential construction projects.” J. Comput. Civ. Eng. 24 (1):
35–44. https://doi.org/10.1061/(ASCE)0887-3801(2010)24:1(35).
Zadeh, L. A. 1965. “Fuzzy sets.” Inf. Control 8 (3): 338–353. https://doi.org
/10.1016/S0019-9958(65)90241-X.
Zadeh, L. A. 1973. “Outline of a new approach to the analysis of com-
plex systems and decision process.” IEEE Trans. Syst. Man Cybern.
SMC-3 (1): 28–44. https://doi.org/10.1109/TSMC.1973.5408575.
Zadeh, L. A. 1976. “The concept of linguistic variable and its application
to approximate reasoning-III.” Inf. Sci. 9 (1): 43–80. https://doi.org/10
.1016/0020-0255(75)90017-1.
Zadeh, L. A. 1994. “Fuzzy logic, neural networks and soft computing.”
Commun. ACM 37 (3): 77–84. https://doi.org/10.1145/175247
.175255.
Zhai, K., N. Jiang, and W. Pedrycz. 2012. “Cost prediction method based
on an improved fuzzy model.” Int. J. Adv. Manuf. Technol. 65 (5–8):
1045–1053.
Zhang, R., B. Ashuri, Y. Shyr, and Y. Deng. 2018. “Forecasting construc-
tion cost index based on visibility graph: A network approach.” Physica
A 493: 239–252. https://doi.org/10.1016/j.physa.2017.10.052.
Zhang, Y., R. E. Minchin, Jr., and D. Agdas. 2017. “Forecasting completed
cost of highway construction projects using LASSO regularized regres-
sion.” J. Constr. Eng. Manage. 143 (10): 04017071. https://doi.org/10
.1061/(ASCE)CO.1943-7862.0001378.
Zhu, W.-J., W.-F. Feng, and Y.-G. Zhou. 2010. “The application of genetic
fuzzy neural network in project cost estimate.” In Proc., 2010 Int. Conf.
on E-Product E-Service and E-Entertainment. New York: IEEE.
J. Constr. Eng. Manage.