Conceptual Cost Estimation Model for EngineeringServices in Public Construction Projects

Khaled Hesham Hyari1; Ahmad Al-Daraiseh2; and Mohammad El-Mashaleh3

Abstract: Cost estimation for public projects includes, but is not limited to, construction costs and engineering services costs. Theavailable cost estimation models for these projects focus on the construction phase, with little or no consideration given to engineeringservices. This paper presents an artificial neural network model for the conceptual cost estimation of engineering services for publicconstruction projects that considers both design costs and construction supervision costs. In developing the model, the authors first identifythe factors that influence the cost of engineering services, and then apply a suitable artificial neural network for a cost estimation model.The model predicts the cost of engineering services as a percentage of construction cost based on project type, engineering servicescategory, project location, and project scope. The model is trained on a data set obtained from the Governmental Tenders Departmentin Jordan, and then tested on some core data samples that had not been seen by the network during training. The optimal network for themodel was selected by conducting a systematic search among a large number of networks with different network architectures and param-eter values. The findings demonstrate that the model is able to predict the cost of engineering services with an acceptable performance forconceptual cost estimation models. The present model complements existing models that focus on construction cost estimating by adding inthe cost of engineering services. The authors expect that this work will contribute to supporting public construction planners in the earlydevelopment of total cost estimates for public construction projects. DOI: 10.1061/(ASCE)ME.1943-5479.0000381. © 2015 AmericanSociety of Civil Engineers.

Author keywords: Design fees; Cost estimating; Public projects; Consultancy fees; Conceptual cost.

Introduction

Cost estimating is one of the most important processes in construc-tion project management. The estimation process is usually per-formed during different phases of the construction project andwith various levels of detail and accuracy, depending on the objec-tive of the estimation task. Costs are estimated, on average, seventimes during a project for each building component (Laitinen1998). As the project evolves, different types of estimates are usu-ally required. In the early stages of project development (i.e., assess-ing project feasibility and evaluating budgetary requirements),conceptual and preliminary estimates are made based on historicaldata and cost forecasting models prior to the completion of theplans and specifications for the project. Once the project has ad-vanced to the procurement stage, after completion of the detailedproject design, detailed estimates are usually made.

The objectives of conceptual cost estimating for public projectsare usually the following: (1) to enable investment decision makingat the conceptual stage (i.e., a go/no go decision); (2) to establish

a construction budget; and (3) to forecast the probable actualcost. The total cost of a public project includes, but is not limitedto, the construction costs and engineering services costs. Thelatter comprise design costs and construction supervision costs(i.e., construction phase monitoring). Of course, constructionplanners need to estimate both construction costs and engineeringservices fees, in order to produce a budget proposal for a project.

Numerous models have been developed to support plannersin the challenging task of estimating construction project cost(Petroutsatou et al. 2012; El-Sawy et al. 2011; Attal 2010; Ugwuand Kumaraswamy 2004; Gunaydin and Dogan 2004; Elhag andBoussabaine 1999; Hegazy and Ayed 1998). Many of the modelsare based on artificial neural networks (Petroutsatou et al. 2012;Attal 2010; Feng et al. 2010; Xin-Zheng et al. 2010; Bouabazand Hamami 2008; Emsley et al. 2002; Al-Tabtabai et al. 1999;Ayed 1997). However, all these models, except the Emsley et al.(2002) model, focus on construction phase costs, with little orno consideration of the cost of engineering services. Emsley et al.(2002) developed a model for estimating the total cost of construc-tion, which includes not only the tender price, which is the con-struction cost to the owner, but also other costs incurred by theowner, such as professional fees.

There is a need for a comprehensive model that identifiesthe drivers of engineering services costs, and includes not only thedesign cost, but also construction supervision costs. To meet thisneed, this paper proposes an intelligent neural network-basedinformation system that will predict the cost of both engineeringdesign and construction project supervision. To the best of the au-thors’ knowledge, this is the first study to propose a comprehensivemodel for the conceptual cost estimation of engineering servicesthat incorporates construction supervision costs. The proposedmodel complements existing models for construction cost estimat-ing, and is expected to help construction planners to develop early

1Associate Professor, Construction Engineering and Management,Dept. of Civil Engineering, Hashemite Univ., P.O. Box 330127, Zarqa13115, Jordan (corresponding author). E-mail: hyari@hu.edu.jo; hyari1@gmail.com

2Assistant Professor, Dept. of Information Systems, College of Compu-ter and Information Sciences, King Saud Univ., P.O. Box 51178, Riyadh,Kingdom of Saudi Arabia 11543.

3Associate Professor, Construction Engineering and Management,Dept. of Civil Engineering, Hashemite Univ., P.O. Box 1580, Amman11821, Jordan.

Note. This manuscript was submitted on April 20, 2014; approved onMarch 4, 2015; published online on April 24, 2015. Discussion period openuntil September 24, 2015; separate discussions must be submitted forindividual papers. This paper is part of the Journal of Management inEngineering, © ASCE, ISSN 0742-597X/04015021(9)/$25.00.

© ASCE 04015021-1 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

estimates of the total cost of completing public construction proj-ects. The remaining sections of the paper illustrate the developmentof this intelligent information system and include a literaturereview, identification of the factors that influence the cost ofengineering services, data collection and preparation, and the train-ing and testing of various architectural representations of artificialneural networks in order to select the architecture that best repre-sents the available bidding data.

Literature Review

Although engineering and consulting services are major contribu-tors to the delivery of construction projects, relatively little researchhas addressed the cost of these services relative to that of construc-tion activities. A review of the available literature reveals that pre-vious research efforts that have tackled professional engineeringservices have focused on (1) the relationship between design costand design quality (Bubshait et al. 1998; Hoxley 2000; Ling 2004;Shrestha and Mani 2012), and (2) bid-level cost estimation for de-sign services (Bajaj et al. 2002; Sturts and Griffis 2005). Bubshaitet al. (1998) study the relationship between design fee and designdeficiency, and conclude that a decrease in design fee is generallyassociated with an increase in design deficiency. Hoxley (2000) in-vestigates the impact of the use of competitive tendering for pro-fessional services on the quality of the services provided. Hoxleybased this study on a survey he conducted of clients’ perceptions ofthe service quality they obtained relative to the professional feesthey paid, and concludes that competitive tendering has no impacton the quality of the engineering services provided. Ling (2004)highlights the conflict in determining design fees in design-buildprojects, because consultants look for high fees to maximize theirprofits, while contractors want to keep their fees low to improvetheir overall competitiveness. Ling (2004) reports that owners sup-port the designers, as they wish to hire quality consultants and mo-tivate them to improve design quality. Ling indicates that ownerssupport designers because the design cost is very small relative tothe construction cost. Shrestha and Mani (2012) study the relation-ship between design cost and the performance of projects that usethe traditional project delivery approach (i.e., design-bid-build),and conclude that higher design costs are correlated with betterconstruction performance in terms of construction duration andcost overruns.

Bajaj et al. (2002) propose the use of parametric estimating topredict design cost, similar to the way it is used to forecast con-struction cost. The objective of their proposed methodology isto help design firms to prepare their bids for design projects. How-ever, this model depends on the availability of extensive databasesof data from previous projects, which include the effort needed tocomplete each design element. Sturts and Griffis (2005) developeda value bidding model for engineering services, which is designedto help engineers to maximize the probability of winning a bid,maximizing profit, and optimizing price.

Carr and Beyor (2005) investigate the design fee schedules usedby government agencies to provide guidance in design fee nego-tiations, and suggest that these schedules should be adjusted overtime at the same rate of increase in construction costs, in order toarrive at fair compensation for design services. The only study thatspecifically addresses the conceptual cost estimation of engineeringservices is reported by Feldmann et al. (2008). They developed amultiple regression model to identify variables that affect engineer-ing services costs for construction and renovation projects carriedout at higher education facilities. However, their model is limited,in that they focus only on typical projects carried out at such

institutions, like research facilities, office/classroom buildings, andresidence/dining facilities.

Cost Attributes of Engineering Services

To identify attributes that affect engineering services’ costs, inter-views with engineering consultants and a comprehensive literaturereview were conducted. Interviews were held with senior engineersfrom seven large engineering consultant firms in Amman, Jordan.The interviewed engineers are involved in preparing bidding offersfor engineering services’ tenders. They were asked to list the factorsthat affect their bid prices for engineering services costs and howsuch factors can be divided into subcategories. Also, the availabledata were presented to the interviewed engineers with the limita-tions imposed by data availability. For example, the interviewedengineers mentioned that the complexity of the design warrantshigher cost. Design complexity may result from site difficultiesor uniqueness of the requested design. However, this factor cannotbe quantified based on the available data. As such, the authors haveidentified the following factors that could influence engineeringservices’ bids for construction projects:1. Project type, divided into the following four types: (1) build-

ings; (2) transportation projects; (3) water and sewage treat-ment and distribution; and (4) land development projects;

2. Engineering services category, including: (1) design services;(2) construction phase monitoring and supervision; and(3) both design and construction phase supervision;

3. Project location: construction site relative to the office locationof the engineering consultant; it is expected that projects inremote areas, far from the consultant’s office, will entail higherfees. This factor is quantified as the distance in kilometersbetween the consultant’s office and the location of the project;

4. Construction costs: the total cost to build the project; pub-lished guidelines available for engineering service fees arein the form of curves that show service fees as a percentageof the estimated construction cost [Feldmann et al. 2008;Carr and Beyor 2005; The Association of Consulting Engi-neers NZ (ACENZ), and The Institution of Professional En-gineers NZ (IPENZ) 2004; Hudgins and Lavelle 1995]; and

5. Project scope: a new construction or a maintenance project.

Data Collection

Engineering sections in public works departments in Jordan, suchas municipalities and infrastructure authorities, usually handle thedesign and supervision of few construction projects, delegatingmost of the engineering services to engineering consulting firmsin the private sector (GTD 2012). Procurement of these servicesis handled via competitive bidding, where engineering consultingfirms are invited to submit sealed bids for design services and/orconstruction phase monitoring services. Selection of the successfulbidder is often based on qualifications, especially for large projects.The Government Tenders Department (GTD) at the JordanianMinistry of Public Works and Housing is responsible for adminis-tering the bidding process for public projects. The GTD has adatabase containing all the bids awarded since 1991, including con-struction bids, design bids, and engineering supervision bids. These2,926 bids, worth a total of 2.9 billion Jordan dinars (JD), which isequivalent to around $4 billion USD, are available online as a pub-lic record. These bids are of two types: (1) construction bidsawarded to contractors and (2) engineering services bids awardedto engineering consultants. The engineering services bids are eitherdesign bids or supervision bids, or both.

© ASCE 04015021-2 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

To make this research possible, the GTD database was searchedfor construction projects in which there are both types of bid data(i.e., a construction bid and an engineering services bid). Thesearch process revealed the existence of 224 construction projectsthat include the construction bid data contained in contractsawarded to contractors, and engineering services bid data con-tained in contracts awarded to engineering consultants. The con-struction costs of the projects considered a range from 68,974 to55,061,658 JD, for a total of 947,405,142 JD. The engineeringservices bids range from 6,640 to 3,229,439 JD, for a total of35,225,988 JD paid for engineering services. The projects usedto develop the model include a wide range of public constructionprojects consisting of (1) building projects such as schools,government administrative buildings, warehouses, and hospitals;(2) constructing transportation facilities such as highways, bridges,and interchanges as well as maintenance and rehabilitation ofexisting highways; (3) water projects such as water distribution net-works, irrigation projects, sewer disposal, dams, and waste watertreatment plants; and (4) infrastructure projects for housing newdevelopment projects. The data pertaining to these projects wereused to develop an artificial neural network model for estimatingthe costs of engineering services in public construction projects, asillustrated in the following section.

Model Development

The model is developed based on an artificial neural network(ANN). An ANN is an information processing system inspired bythe biological neural networks in the human nervous system. AnANN consists of a group of neurons, or processing elements, sim-ilar to nerve cells, interconnected to form a network. NNs are usedin software engineering to model complex relationships betweeninputs and outputs through adaptive learning from training exam-ples. After training, the model can be used to predict the outputbased on given input data.

The network is designed for a specific set of inputs and outputs.In the input layer are found the factors that influence a problem,while the solution to the problem (e.g., prediction, classification,etc.) resides in the output layer. The input information that will ul-timately produce the solution is processed in the hidden layer(s), aseach neuron receives input, processes it, and delivers output. Theneuron computes a weighted sum, SðxÞ, of its input signals, Xi(i ¼ 1 to m), and their corresponding weights, Wi, as illustrated

in Fig. 1. Then, the neuron generates output through an activationfunction, FðsÞ.

In the present model, a back-propagation network (BPN) is usedto estimate the engineering services costs for public constructionprojects. Many cost estimating models adopt the BPN becauseof its simplicity and good generalization capability (Arafa andAlqedra 2011; Xin-Zheng et al. 2010; Sodikov 2005; Kim et al.2004; Bhokha and Ogunlana 1999). A BPN is usually composedof three layers, the input layer, the hidden layer, and the output layer,although it is sometimes necessary to incorporate more than one hid-den layer (two, three, or more). The number of neurons contained inthe input layer is equal to the dimensionality of the input patterns(i.e., the number of attributes in every input pattern). In fact, the inputlayer in our proposed model contains five neurons, as shown inFig. 1. The output layer contains one neuron since the model hasonly one output, which is the estimated cost of engineering services.Unfortunately, there is no precise rule as to how many neurons thehidden layer should contain, which is why the authors use the opti-mization technique to select this number, in addition to other net-work architecture parameters, as described in the following sections.

The present model is implemented in three major phases:(1) the data preparation phase, in which the historical bid data tobe used to develop the model are organized and formatted; (2) thenetwork training phase, in which the values of the network param-eters are determined; and (3) the testing phase, in which theperformance of the network is tested on new cases by comparingthe engineering fee estimated by the ANN model with the actualengineering services fees charged. The detailed computation pro-cedure for these three phases is explained in the following sections.

Phase 1: Data Preparation

The main objective in this phase is to modify the original formatof the data, so that it can be used in theMATLAB environment. Thedata pertaining to the input parameters for all the projects was en-tered into a spreadsheet program (Microsoft Excel), and then trans-formed into numerical values, as depicted in Fig. 1. The values forthe “project type” parameter, for example, have been transformedinto numbers from 1 to 4 for buildings, transportation, water, andland development projects, respectively. It should be noted that thedivision of project type into four types is intended to match theareas of specialization of engineering consultants in Jordan. A sep-arate licensing procedure is required from any consultant in order topractice in each one of those project types. The model output,

Input

I1 Project Type(1= Buildings, 2= Transportation, 3= Water, 4= Land Development)

I2 Engineering Service Category (1= Design, 2= Construction Phase Monitoring and Supervision, 3= Design and Construction Phase Supervision)

I3 Project Location

I4 Project Cost ($)

I5 Project Scope(1= New Construction, 2= Maintenance)

OutputEngineering Services Cost (% of Construction Cost)

I2

O

I1

I3

I4

I5

Input

I1 Project Type(1= Buildings, 2= Transportation, 3= Water, 4= Land Development)

I2 Engineering Service Category (1= Design, 2= Construction Phase Monitoring and Supervision, 3= Design and Construction Phase Supervision)

I3 Project Location

I4 Project Cost ($)

I5 Project Scope(1= New Construction, 2= Maintenance)

OutputEngineering Services Cost (% of Construction Cost)

I2

O

I1

I3

I4

I5

I2

O

I1I1

I3

I4

I5

Fig. 1. Engineering services cost estimation model

© ASCE 04015021-3 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

which is the cost of the engineering services of each project, wastransformed into a percentage of the construction cost for theproject type.

Next, the data were examined to identify and remove outliers.This step was performed manually because the data includes a man-ageable number of patterns. As a result, some patterns were re-moved based on personal experience and the use of basicgraphing techniques. Subsequently, the cost data collected was ran-domly split into two sets, one for training and the other for testing.The ratio of training to testing patterns used in developing themodel was about 2.5∶1. Since the amount of data used is relativelysmall in this case (i.e., a total of 207 patterns), no validation set wascreated.

Phase 2: Network Training

The training phase comprises two modules: training and networkoptimization. The training module trains the network with a num-ber of cases with known output values, while the optimization mod-ule searches for the best network architecture and parameters thatyield the lowest error in the predicted results.

Network Training ModuleThe training set trains the network in order to choose the weights ofthe interconnections between network nodes. The training moduleuses the back-propagation learning (training) algorithms for adjust-ing the weights of the interconnections, so that the estimates gen-erated by all the training patterns have minimal error relative to the

known output values of all the input patterns. The ANN is trained inthis module using the cycle of steps shown in Fig. 2:1. Initialize the weights randomly;2. Apply the input vector to the input neurons;3. Calculate the net values of the input to the neurons in the hid-

den layer. Each hidden neuron calculates the weighted sum ofthe inputs received from the input nodes (i.e., the values of theindependent variables);

4. Calculate the output of the hidden layer nodes (neurons) usingthe tan-sigmoid transfer function (tansig) that squeezes the va-lues of the weighted sum computed in the previous step into alimited range, between –1 and þ1, as shown in the followingequation:

tansigðxÞ ¼ 2

1þ e−2x− 1 ð1Þ

5. Calculate the output from the output node using the lineartransfer function. The input to the output nodes is first calcu-lated as the weighted sum of the input received from thehidden layer nodes, and then the output is calculated by mod-ifying that input using the linear transfer function (purelin).The purelin transfer function is simply a linear function thatproduces the same output as its input: the output produced isthe estimated value of the dependent variable, which is repre-sented by the output node

purelinðxÞ ¼ x ð2Þ

Fig. 2. ANN training and optimization

© ASCE 04015021-4 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

6. Calculate the error terms for the output layer by comparing thenet output with the desired, or target, value;

7. Calculate the error terms for the hidden layer, based on theoutput error; and

8. Update the weights on the output and hidden layers, based onthe calculated error.

The aforementioned training steps (i.e., which present the train-ing patterns to the NN) are repeated until the stopping criteria havebeen satisfied, which is either: (1) when the number of iterationsexceeds the number of epochs or (2) when the mean standard error(MSE) drops below target. The number of epochs in this model isset to 15,000, and the MSE target is set to 0.1. The selection of arelatively large value for the MSE target is intended to avoid over-fitting of the network to the training patterns owing to the smallnumber of training patterns available.

Network Optimization ModuleThe design of the ANN involves a number of decisions that shouldbe taken, with respect to (1) the number of neurons in the hiddenlayer, (2) the training algorithm that will be used to train the net-work, and (3) the parameter settings for the training algorithm se-lected. The network architecture and parameter settings selectedaffect the performance of the network. However, no theory has beenclearly defined for selecting the ideal network architecture and set-ting its parameters (El-Sawy et al. 2011). The available ANN mod-els for cost estimation use either trial and error (El-Sawy et al. 2011;Sodikov 2005; Kim et al. 2005; Bhokha and Ogunlana 1999) or anoptimization technique to choose these decision variables (Fenget al. 2010; Kim et al. 2004).

The proposed model takes a systematic search approach to op-timizing the network topology and parameters in order to obtain thebest ANN for the engineering services cost data. During the search,various ANN training algorithms and parameters are used to create

a large number of ANNs. The best ANN is then chosen based onthe least mean-square error (LMSE), and used on the test data.

The computations in this module are performed in the cyclicalsteps, as shown in Fig. 2:1. Select a training algorithm. The algorithms used in this module

are (1) Gradient Descent—traingd; (2) Gradient Descent withMomentum—traingdm; (3) Variable Learning Rate—traingdx;(4) Resilient Back-propagation—trainrp; (5) Levenberg–Marquardt—trainlm; and (6) Powell–Beale Restarts—traincgb. A brief description of these training algorithms isprovided in Table 1.

2. For each training algorithm, select a value for the first para-meter. As shown in Table 2, the first parameter for the firstfour training algorithms is the learning rate. This parameteraffects how quickly/slowly weights are updated, and mayplay an important role in the convergence of the network.(Thirty possible values for the learning rate were tried in thismodule.)

3. For each training algorithm, select a value for the second para-meter. For example, the second parameter that should be cho-sen for the Resilient Back-propagation learning algorithm isthe maximum weight change. (Eleven possible values for thisparameter were tried in this module, as shown in Table 2.)

4. Select a value for the third parameter if the particular learningalgorithm takes a third parameter. As shown in Table 2, thethird parameter is the number of neurons in the hidden layer.(Twelve possible values for this parameter were tried, from 4neurons up to 15 neurons.)

5. Train the network using each training algorithm and itsparameter values, using the aforementioned network trainingmodule.

6. Save the network’s objects and calculate the mean standarderror for the network.

Table 1. Description of Training Algorithms

Algorithm Description

Gradient descent (traingd) Traingd is probably the oldest training algorithm in use for training NNs. It uses the gradient descent method to adjust theweights of the nodes, and works well with any transfer function, as long as it has a derivative (Hagan et al. 1996)

Gradient descent withmomentum (traingdm)

This algorithm uses the same parameters as traingd, in addition to a momentum coefficient, mc, to consider not only the localgradient, but also the recent trends of the error surface. This enhances the speed of network training, and allows a globalminimum to be found, rather than becoming stuck in a local minimum (Hagan et al. 1996). The mc takes values between 0and 1, where 0 represents no momentum and 1 represents a high value of the momentum

Variable learningrate (traingdx)

This training algorithm adds adaptive learning to the functionality of traingd and traingdm. It uses the gradient descent withmomentum algorithm and changes the learning rate at the same time, resulting in faster convergence. The idea is toincrementally increase the learning rate by a small amount determined by lr_inc (learning rate increment), if the performancemoves towards the goal and as long as the increment does not exceed a specific value, max_perf_inc (maximum performanceincrement). If it does, a value of lr_dec (learning rate decrement) is subtracted from the learning rate (Hagan et al. 1996)

Resilient backpropagation (trainrp)

This algorithm, described in Riedmiller and Braun (1993), is used to update the network’s weights. The weights and biases areadjusted by either incrementing them or decrementing them, based on whether or not the gradient of this variable changed signs(i.e., from þve to −ve, or vice versa). The increment and decrement is determined by the parameters delta_inc and delta_dec.The trainrp algorithm usually converges faster than any of the above three training algorithms

Levenberg–Marquardt(trainlm)

The trainlm algorithm optimizes its weight and bias values according to the Levenberg–Marquardt optimization techniqueLevenberg (1944) and Marquardt (1963). It uses the second-order training speed without having to calculate the hessian matrix,but rather approximates this matrix by calculating and using a Jacobian matrix. The approximated matrix is then used to updateweights and biases. This is probably the fastest back-propagation training function, its only drawback being the amount ofmemory it needs. The following are the new parameters that are used in this algorithm: (1) the adaptive value (Mu), which isdirectly used in calculating the weight or bias change; and (2) Mu_inc and Mu_dec, which make Mu an adaptive parameter

Powell–Bealerestarts (traincgb)

The conjugate gradient back propagation with Powell–Beale restarts (traincgb) searches the error surface in the conjugatedirection, rather than in the direction of the steepest descent (Powell 1977). This modification produces a quicker trainingfunction. Instead of resetting the search direction periodically (as is the case with all conjugate algorithms), this algorithm resetsthe search direction only if there is little orthogonality between the current gradient and the previous one. This algorithm usesmany parameters, such as: (1) the initial step size in the interval location step (Delta); and (2) the maximum step size (Bmax)

© ASCE 04015021-5 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

7. Repeat steps 4 to 6 until all possible values for the thirdparameter have been selected.

8. Repeat steps 3 to 7 until all possible values for the secondparameter have been selected.

9. Repeat steps 2 to 8 until all possible values for the firstparameter have been selected.

10. Repeat steps 1 to 9 until all the learning algorithms havebeen used.

11. Choose the network and parameter values that return the mini-mum mean standard error on the training data.

The procedure works as follows: Gradient Descent (traingd) isselected as a training algorithm; 0.01 is selected as a learning rate;four hidden nodes are selected, since there is no third parameter,and the NN is trained and stored; five hidden nodes are selected,and the NN is trained and stored. This process continues until all 15nodes in the hidden layer have been used. The second value for thelearning rate (i.e., 0.02) is selected, and so on.

The outcome of this procedure is an ANN that shows the bestperformance on the training data. This ANN is then used to predictthe engineering fee for the unseen previous cases from the test set.This module investigated 18,552 different ANNs using six differenttraining algorithms and the associated parameter setting options.The best ANN obtained uses the Resilient Back-propagation train-ing algorithm (trainrp), a learning rate of 0.05, and a maximum

weight change of 16, and includes 15 neurons in the hidden layer.The performance of the model is evaluated using the mean absoluteerror rate (MAER), which is the average of the absolute values ofthe differences between each predicted output and the correspond-ing target output, as shown in Eq. (3)

MAER ¼ðP j ESCe−ESCa

ESCa× 100jÞ

nð3Þ

where ESCe represents the estimated engineering services costscalculated by the NN model, ESCa represents the actual engineer-ing services costs, and n is the number of sets of test data. Fig. 3 andTable 3 show the performance of the model for the training patternsby comparing actual and predicted results for the engineering serv-ices costs. An average percentage accuracy of 26.3% was obtained.The error percentage ranges from 0 to 130% for the minimum andmaximum, respectively. The standard deviation is 0.279.

Phase 3: Network Testing

The objective of this phase is to evaluate network performance(i.e., the prediction accuracy of the network) by using new patternsthat were not used in the training phase. This phase simulates thenetwork response to new inputs. The test set is used to measure

Table 2. Systematic Search Parameters

Training algorithm First parameter Second parameter Third parameter

Gradient descent (traingd) Learning rate(0.01:0.01:0.3)a (30)b

Number of neurons in hidden layer(4:1:15)a (11)b

—

Gradient descent with momentum (traingdm) Momentum coefficient (mc)(0.05:0.05:0.95)a (19)b

Number of neurons inhidden layer (4:1:15)a (11)b

Variable learning rate (traingdx) Inc/Dec factors Dec (0.65:0.01:0.75)a

Inc (1.01:0.01:1.11)a (11)b

Resilient back propagation (trainrp) Maximum weightchange (1:5:51)a (11)b

Levenberg-Marquardt (trainlm) Adaptive value (Mu)(0.001:0.005:0.05)a (11)b

Inc/Dec factors Mu-Dec (0.05:0.01:0.15)a

Mu-Inc (10:2:30)a (11)b

Powell-Beale restarts (traincgb) Delta (search parameter)(0.01:0.01:0.15)a (15)b

Maximum step size (15:2:35)a (11)b

aRange of values for each parameter (minimum: step size: maximum).bNumber of possible values for each parameter.

Fig. 3. Actual versus predicted results for the training patterns

© ASCE 04015021-6 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

the performance of the selected ANN model. The performance ofthe model was evaluated with 54 sets of test data using the MAER,which is used in the training phase. Fig. 4 and Table 4 illustrate theperformance of the model using the test patterns. The MAER ob-tained is 28.2%. The error percentage ranges from 0.8 to 86% forthe minimum and maximum, respectively. The standard deviationis 0.194.

Although these results reveal a relatively high error rate, itshould be noted that conceptual cost estimation is inherently asso-ciated with a high percentage error. Traditional methods for costestimation report a range of error rates between 20.8 and 27.9%

(Emsley et al. 2002). Harding et al. (1999) report that the actualerror percentage obtained in early cost estimation was greater than20%. Also, previous conceptual estimation models that used an NNreport comparable results. For example, El-Sawy et al. (2011) usedan NN to estimate the site overheads of a construction project, andreport a MAER value of 27.6%.

The proposed model was developed using a relatively smallsample of 224 construction projects, and, since the NN techniqueis data intensive, the reliability of the model is expected to improveas larger data sets are obtained and used to train and test thenetwork.

Table 3. Actual versus Predicted Results in the Training Phase

PatternActualfee (%)

NNpredictedfee (%) Error (%) Pattern

Actualfee (%)

NNpredictedfee (%) Error (%) Pattern

Actualfee (%)

NNpredictedfee (%)

Error(%)

1 1.01 1.11 10.0 52 5.56 4.63 16.8 103 5.31 3.82 28.12 5.4 3.89 28.0 53 4.38 3.89 11.2 104 4.54 4.85 6.93 7.23 5.53 23.5 54 2.31 5.02 117.3 105 2.58 3.97 53.84 2.35 5.43 130.9 55 5.3 4.73 10.7 106 10.26 5.43 47.15 2.78 4.73 70.2 56 2.84 4.14 45.7 107 4.31 4.49 4.36 4.6 5.53 20.3 57 5.44 5.62 3.3 108 4.36 4.36 0.07 5.93 3.93 33.7 58 2.78 4.35 56.5 109 4.29 6.47 50.88 9.37 4.63 50.6 59 6.92 6.59 4.7 110 3.75 5.62 49.99 3.97 5.43 36.7 60 2.81 3.43 22.2 111 5.13 7.58 47.710 9.03 5.53 38.7 61 7.71 4.73 38.6 112 6.57 4.98 24.111 4.29 3.98 7.2 62 1.74 3.66 110.6 113 4.29 3.98 7.212 1.62 3.01 85.8 63 2.21 4.49 103.4 114 3.66 3.98 8.813 4.04 4.26 5.4 64 8.41 4.00 52.4 115 5.16 4.73 8.314 4.18 4.64 11.1 65 3.52 3.98 13.1 116 8.84 4.73 46.515 4.5 5.96 32.5 66 6.77 6.47 4.4 117 5.03 4.73 6.016 4.6 5.53 20.3 67 6.19 5.96 3.7 118 2.05 3.98 94.317 9.52 6.47 32.0 68 4.12 3.79 8.1 119 3.52 3.98 13.118 3.9 4.43 13.7 69 2.99 4.29 43.5 120 7.54 7.70 2.219 7.31 7.00 4.2 70 8.27 7.75 6.3 121 2.53 4.73 87.020 7.52 6.93 7.8 71 5.26 5.62 6.9 122 5.29 4.73 10.621 4.4 3.01 31.6 72 4.41 4.26 3.4 123 3.09 3.72 20.522 7.3 5.04 31.0 73 5.35 6.22 16.3 124 2.78 4.73 70.223 9.55 9.40 1.6 74 4.22 4.73 12.1 125 7.21 4.60 36.124 8.03 5.62 30.0 75 9.6 9.36 2.5 126 3.5 4.73 35.225 3.51 3.32 5.5 76 7.36 7.96 8.2 127 5.14 6.41 24.626 5.58 4.29 23.1 77 9.64 7.47 22.5 128 3.59 4.46 24.127 10.69 5.02 53.0 78 3.68 3.75 1.8 129 3.59 3.98 10.928 8.81 9.02 2.4 79 4.21 4.23 0.5 130 4.29 4.26 0.729 5.02 4.73 5.8 80 4.77 4.73 0.8 131 2.33 4.34 86.430 1.93 3.93 103.7 81 3.35 4.98 48.8 132 9.5 7.55 20.631 3.3 4.55 37.8 82 4.98 4.73 5.0 133 5.34 4.63 13.432 6.9 7.58 9.8 83 4.61 4.73 2.6 134 4.17 3.72 10.733 4.62 4.47 3.3 84 3.66 4.93 34.7 135 7.26 3.98 45.134 8.33 4.49 46.0 85 3.71 5.04 35.7 136 2.98 3.98 33.635 3.99 4.73 18.6 86 5.96 4.35 27.0 137 5.96 6.04 1.336 1.61 2.54 58.0 87 4.79 4.00 16.4 138 8.2 6.41 21.937 6.81 5.20 23.6 88 5.88 2.54 56.8 139 4.62 5.04 9.038 8.67 7.58 12.6 89 5.29 4.73 10.6 140 4.8 3.70 22.939 9.6 9.36 2.5 90 5.89 7.47 26.8 141 9.88 9.02 8.740 5.56 4.63 16.8 91 4.58 3.98 13.1 142 6.73 7.07 5.041 3.34 3.72 11.2 92 3.81 3.98 4.5 143 1.77 4.00 126.342 1.22 1.29 5.9 93 3.35 3.98 18.9 144 4.93 4.83 2.043 4.31 4.35 0.9 94 2.88 5.20 80.7 145 3.1 4.55 46.744 4.63 4.97 7.3 95 7.74 4.73 38.9 146 4.93 7.00 42.045 6.38 6.93 8.7 96 5.91 5.01 15.3 147 3.15 5.05 60.446 1.71 2.54 48.7 97 1.84 1.62 12.0 148 4.8 3.72 22.647 3.95 4.73 19.8 98 4.99 4.00 19.7 149 4.9 4.73 3.548 7.05 4.63 34.4 99 7.31 7.00 4.2 150 2.71 5.53 104.249 3.16 4.00 26.7 100 9.19 9.02 1.9 151 6.15 5.15 16.350 1.47 1.53 3.8 101 5.88 3.70 37.0 152 3.69 4.98 35.151 4.68 6.47 38.3 102 1.4 1.62 15.6 153 5.95 5.05 15.1

© ASCE 04015021-7 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

Conclusions

The cost of public projects includes, but is not limited to, construc-tion costs and engineering services costs, but relatively few re-search papers have addressed the cost of engineering services,which includes design and construction supervision costs, in com-parison the number examining construction activity costs. This pa-per presented the development of a conceptual cost estimationmodel for engineering services in public construction projects.The factors influencing cost were identified by interviewing indus-try experts and through an extensive review of the literature.The model was developed based on an ANN that was trainedand tested using a sample of 224 construction projects obtainedfrom the General Tenders government department in Jordan.The projects used to develop the model include a wide range ofpublic construction projects that includes buildings, transportationfacilities, water projects, and land development projects. The ANNarchitecture and parameters were optimized using a systematicsearch approach. The search led to the choice of an optimal networktopology consisting of a five-neuron input layer, a 15-neuron hid-den layer that use nonlinear sigmoid transfer functions, and a linearsingle-neuron output layer. The findings showed that the model wasable to map the underlying relationship between the input cost fac-tors and the cost of engineering services during the training stage,and maintained an average accuracy percentage of 26.3%. In con-trast, in the testing phase, acceptable generalization capability wasobtained and average accuracy percentage of 28.2% was achieved.It is recognized that the ANN technique is data intensive, and so itis recommended that more projects be exploited for future develop-ment to enhance the reliability of the model. Furthermore, themethodology presented in this paper can be utilized to develop sub-models that focus on a specific type of construction projects such asbridges or schools. Such specialization is expected to enhance theprediction accuracy of the model by limiting the variability in thescope of engineering services required for various types of con-struction projects.

The presented model is expected to support construction plan-ners in the challenging task of providing cost estimates in the earlydevelopment stages of public construction projects. Providing rea-sonable estimates for costs in the conceptual phase of projects iscrucial for taking informed decisions in the initiation stage of proj-ects. Finally, although the present model was developed based ondata from the Jordanian construction industry, the methodologywould suggest a much broader geographical applicability of theANN technique to estimating engineering services costs. As themodel complements existing models that focus on constructioncosts, this research may be extended to develop a comprehensivemodel that incorporates both construction and engineering servicescosts. Also, future research is needed to evaluate the performance

of the model and examine the drivers of engineering services costsusing different data sets from various places worldwide.

Acknowledgments

This work was supported by Deanship of Scientific Research andResearch Center of College of Computer and Information Sciences,King Saud University. The authors are grateful for this support.

Fig. 4. Actual versus predicted results for the test patterns

Table 4. Actual versus Predicted Results in the Testing Phase

PatternProjecttype

Servicescategory

Projectlocation

Projectscope

Actualfee (%)

NNpredictedfee (%)

Error(%)

1 1 2 47 1 5.95 5.05 15.12 1 2 23 1 3.77 5.16 36.83 1 2 70 1 3.15 4.63 46.94 4 2 47 1 6.37 5.96 6.45 1 2 10 1 3.24 4.73 46.06 3 2 51 2 3.62 4.78 32.17 3 2 10 1 2.51 4.00 59.68 1 2 89 1 3 3.98 32.79 1 2 103 1 4.07 5.62 38.110 1 2 94 1 6.97 3.79 45.611 1 2 23 1 2.92 5.16 76.612 1 2 25 1 5.04 7.27 44.313 3 2 328 1 3.17 4.93 55.614 1 2 35 1 6.31 6.24 1.115 3 2 10 1 3.27 4.00 22.516 3 2 32 2 5.41 4.29 20.717 3 2 230 2 10.91 9.44 13.518 4 2 210 1 6.46 4.00 38.019 3 2 10 1 4.09 4.93 20.620 3 2 328 1 5.24 7.02 34.021 1 2 32 1 4.77 4.73 0.822 1 2 10 1 2.58 3.75 45.223 3 2 103 1 5.36 6.24 16.424 1 2 35 1 5.07 4.63 8.825 1 2 70 1 6.12 7.47 22.126 1 2 30 1 4.28 4.98 16.527 1 2 51 1 5.09 4.73 7.128 1 2 10 1 5.2 5.51 5.929 4 2 10 1 3.49 3.98 14.130 1 2 89 1 1.47 1.62 10.131 1 1 70 1 4.84 6.34 30.932 2 1 105 1 6.87 5.16 24.933 1 2 23 1 6.61 4.46 32.634 2 2 328 1 7.82 5.02 35.835 3 2 230 1 4.05 4.73 16.836 1 2 10 1 3.66 3.72 1.537 3 2 89 1 4.94 4.00 18.938 3 2 10 1 5.87 3.98 32.239 1 2 89 1 3.64 4.73 30.040 1 2 10 1 6.84 4.78 30.141 3 2 51 2 5.44 6.94 27.542 4 2 32 1 6.69 8.70 30.043 1 2 230 1 7.17 8.70 21.344 1 2 230 1 3.24 5.96 84.045 1 2 10 1 4.11 4.73 15.146 1 2 10 1 2.54 4.73 86.247 1 1 10 1 2.2 2.54 15.648 1 2 89 1 6.27 3.98 36.549 4 2 70 1 7.05 4.61 34.650 1 2 328 1 3.27 5.20 59.151 1 2 15 1 3.91 4.71 20.552 1 2 10 1 5.47 4.73 13.553 1 1 10 1 2.25 2.54 13.054 4 2 212 1 7.06 9.36 32.6

© ASCE 04015021-8 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.

References

ACENZ, and IPENZ (The Association of Consulting Engineers NZ, andThe Institution of Professional Engineers NZ). (2004). “Guidelineon the briefing and engagement for consulting engineering services.”Wellington, NZ, ⟨http://www.ipenz.org.nz/ipenz/practicesupport/endorsedinfo/BE_Guildine.pdf⟩ (Oct. 26, 2013).

Al-Tabtabai, H., Alex, P. A., and Tantash, M. (1999). “Preliminary costestimation of highway construction using neural networks.” CostEng., 41(3), 19–24.

Arafa, M., and Alqedra, M. (2011). “Early stage cost estimation of buildingconstruction projects using artificial neural networks.” J. Artif. Intell.,4(1), 63–75.

Attal, A. (2010). “Development of neural network models for prediction ofhighway construction cost and project duration.” Master’s thesis, RussCollege of Engineering and Technology of Ohio Univ, Athens, OH.

Ayed, A. (1997). “Parametric cost estimating of highway projects usingneural networks.” Master’s thesis, Memorial Univ. of Newfoundland,Newfoundland, Canada.

Bajaj, A., Gransberg, D., and Grenz, M. (2002). “Parametric estimatingfor design costs.” 2002 AACE Int. Transactions, EST.08, Associationfor the Advancement of Cost Engineering, Morgantown, WV,EST.08.01–EST.08.06.

Bhokha, S., and Ogunlana, S. (1999). “Application of artificial neural net-work to forecast construction cost of buildings at the pre-design stage.”Proc., 7th East Asia-Pacific Conf. on Structural Engineering andConstruction (EASEC-7), Asian Institute of Technology, Bangkok,Thailand, 7–12.

Bouabaz, M., and Hamami, M. (2008). “A cost estimation model for repairbridges based on artificial neural networks.” Am. J. Appl. Sci., 5(4),334–339.

Bubshait, A., Al-Said, F., and Abolnour, M. (1998). “Design fee versusdesign deficiency.” J. Archit. Eng., 10.1061/(ASCE)1076-0431(1998)4:2(44), 44–46.

Carr, P., and Beyor, P. (2005). “Design fees, the state of the profession, anda time for corrective action.” J. Manage. Eng.J. Manage. Eng., 10.1061/(ASCE)0742-597X(2005)21:3(110), 110–117.

Elhag, T., and Boussabaine, A. (1999). “Tender price estimation:Neural networks versus regression analysis.” Proc., Int. ConstructionResearch Conf. of the Royal Institution of Chartered Surveyors COBRA‘99, The Royal Institution of Chartered Surveyors, London,114–123.

El-Sawy, I., Hosny, H., and Abdel Razek, M. (2011). “A neural networkmodel for construction projects site overhead cost estimating in Egypt.”IJCSI, 3(1), 273–283.

Emsley, M., Lowe, D., Duff, A., Harding, A., and Hickson, A. (2002).“Data modeling and the application of a neural network approachto the prediction of total construction cost.” Constr. Manage. Econ.,20(6), 465–472.

Feldmann, M., et al. (2008). “Architectural and engineering fees fromthe public institutional perspective.” J. Manage. Eng., 10.1061/(ASCE)0742-597X(2008)24:1(2), 2–11.

Feng, W., Zhu, W., and Zhou, Y. (2010). “The application of genetic algo-rithm and neural network in construction cost estimate.” Proc., 3rd Int.Symp. on Electronic Commerce and Security Workshops, AcademyPublisher, U.K., 151–155.

GTD (Government Tenders Department). (2012). “2012 year book.” ⟨http://www.gtd.gov.jo/usersfile/upload_files/11129219191370940537.pdf⟩(Sep. 22, 2013).

Gunaydin, H., and Dogan, S. (2004). “A neural network approach forearly cost estimation of structural systems of buildings.” Int. J. ProjectManage., 22(7), 595–602.

Hagan, M., Demuth, H., and Beale, M. (1996). Neural network design,PWS Publishing, Boston.

Harding, A., Lowe, D., Emsley, M., Hickson, A., and Duff, R. (1999). “Therole of neural networks in early stage cost estimation in the 21st cen-tury.” Proc., Int. Construction Research Conf. of the Royal Institution ofChartered Surveyors COBRA ‘99, The Royal Institution of CharteredSurveyors, London, 161–168.

Hegazy, T., and Ayed, A. (1998). “Neural network model for parametriccost estimation of highway projects.” J. Constr. Eng. Manage.,10.1061/(ASCE)0733-9364(1998)124:3(210), 210–218.

Hoxley, M. (2000). “Are competitive fee tendering and construction pro-fessional service quality mutually exclusive?” Constr. Manage. Econ.,18(5), 599–605.

Hudgins, D., and Lavelle, J. (1995). “Risk management in design engineer-ing bids.” Technical Rep., U.S. Dept. of Energy, Washington, DC.

Kim, G., Yoon, J., An, S., Cho, H., and Kang, K. (2004). “Neural networkmodel incorporating a genetic algorithm in estimating constructioncosts.” Build. Environ., 39(11), 1333–1340.

Kim, S., Choi, J., Kim, G., and Kang, K. (2005). “Comparing cost predic-tion methods for apartment housing projects: CBR versus ANN.”J. Asian Archit. Build. Eng., 4(1), 113–120.

Laitinen, J. (1998). “Model based construction process management.”Ph.D. thesis, Royal Institute of Technology, Stockholm, Sweden.

Levenberg, K. (1944). “A method for the solution of certain non-linearproblems in least-squares.” Q. Appl. Math., 2(2), 164–168.

Ling, F. (2004). “Consultancy fees: Dichotomy between A/E’s needto maximize profit and employers’ need to minimize cost.” J. Prof.Issues Eng. Educ. Pract., 10.1061/(ASCE)1052-3928(2004)130:2(120),120–123.

Marquardt, D. (1963). “An algorithm for least-squares estimation of non-linear parameters.” SIAM J. Appl. Math., 11(2), 431–441.

MATLAB [Computer software]. Natick, MA, MathWorks.Petroutsatou, K., Georgopoulos, E., Lambropoulos, S., and Pantouvakis, J.

(2012). “Early cost estimating of road tunnel construction using neuralnetworks.” J. Constr. Eng. Manage., 10.1061/(ASCE)CO.1943-7862.0000479, 679–687.

Powell, M. J. D. (1977). “Restart procedures for the conjugate gradientmethod.” Math. Program., 12(1), 241–254.

Riedmiller, M., and Braun, H. (1993). “A direct adaptive method for fasterbackpropagation learning: The PROP algorithm.” Proc., IEEE Int.Conf. on Neural Networks (ICNN), IEEE, New York, 586–591.

Shrestha, P., and Mani, N. (2012). “Impact of design cost on design bidbuild project performance.” Proc., 2012 Construction ResearchCongress, ASCE, Reston, VA, 1570–1579.

Sodikov, J. (2005). “Cost estimation of highway projects in developingcountries: Artificial neural network approach.” J. East. Asia Soc.Transp. Stud., 6, 1036–1047.

Sturts, C., and Griffis, F. (2005). “Addressing pricing: Value bidding forengineers and consultants.” J. Constr. Eng. Manage., 10.1061/(ASCE)0733-9364(2005)131:6(621), 621–630.

Ugwu, O., and Kumaraswamy, M. (2004). “Neural network-based decisionsupport for estimating cost of highway bridges—A Hong Kong study.”Proc., Int. Construction Research Conf. of the Royal Institution ofChartered Surveyors COBRA 2004, The Royal Institution of CharteredSurveyors, London.

Xin-Zheng, W., Xiao-Chen, D., and Jing-Yan, L. (2010). “Applicationof neural network in the cost estimation of highway engineering.”J. Comput., 5(11), 1762–1766.

© ASCE 04015021-9 J. Manage. Eng.

J. Manage. Eng.

Dow

nloa

ded

from

asc

elib

rary

.org

by

SDL

Por

tal o

n 05

/05/

15. C

opyr

ight

ASC

E. F

or p

erso

nal u

se o

nly;

all

righ

ts r

eser

ved.