ICES5 proceeding-PP. 1-3 Gaza, 4-5 november 2014 JOURNAL OF ENGINEERING RESEARCH AND TECHNOLOGY, VOLUME 1, ISSUE 4, DECEMBER 2014 24 Influencing Cost Factors in Road Projects in Gaza Strip Using ANN Hasan KH. Abujamous 1 , Rifat N. Rustom 2 , and Mahmoud Y. Abukmail 3 1H. Abujamous, Civil Engineering Department, Islamic University-Gaza, Palestine, e-mail mhmod_85@hotmail.com 2R. Rustom, Ph.D. in Civil Engineering from Drexel University in the U.S.A., Palestine, e-mail rustom@iugaza.edu.ps 3M. Abukmail, Civil Engineering Department, Islamic University-Gaza, Palestine, e-mail. mhmod_85@hotmail.com Abstract— Conceptual cost estimate can serve the owners’ feasibility estimate and assists in the establishment of the owner's funding which aids the engineers in designing to a specific budget. Conceptual estimating exhibits low accuracy level due to the lack of project information and the high level of uncertainty at early stage of project de- velopment. The purpose of this paper is to determine the most influencing cost factors in road projects using Delphi technique and Artificial Neural Networks. These factors were employed in a neural network (NN) for building a multi-layer perceptron (MLP) model to estimate the road project cost. Historical data of Gaza strip road projects were used to train and test the MLP model. The model developed showed a reduced error rate of 6.3% which demonstrates the ability to estimate the cost of road projects at early stage with higher accuracy. Index Terms— Cost factors, Conceptual cost estimate; Artificial neural networks. I. INTRODUCTION Early stage cost estimate plays a significant role in any initial road projects decisions; despite the pro- ject scope has not yet been finalized. Major prob- lems faced are lack of preliminary information, database of road works costs and appropriate cost estimation methods [2, 3]. Project managers in Gaza Strip, often need to estimate the cost of road projects at early stage quickly and approximately to provide funding or to obtain the adoption of the budget from decision-makers. Therefore, it is im- portant to know the cost of the road projects in a short time with acceptable accuracy. Artificial Neural Networks (ANN) is well suited to model complex problems where the relationship between the model variables is unknown [4]. The main objective of this research is to develop ANN model to estimate the cost of road projects in Gaza strip at early stage to reduce the error in es- timation. To achieve this, the factors that affect the cost of road projects that can be available at early stage were identified and modeled. II. LITERATURE REVIEW Cost is on the mind of every business. Every busi- ness is expected to do more with less. The objec- tive is to minimize cost, maximize profit, and maintain the competitive edge. Methods for cost estimation vary as the project evolves from the early stages of conception to the construction phase. Conceptual cost estimate is made at the early stage of the project where the budgets are to be decided, and available information is limited. It is conduct- ed without working drawings or detailed specifica- tions. The estimator may have to make such an estimate from rough design sketches, without di- mensions or details and from an outline specifica- tion and schedule of the owner's space require- ments [5]. The conceptual cost estimate can serve several purposes, such as:  It supplements or serves as the owners fea- sibility estimate.  It aids the Architect/Engineer in designing to a specific budget.  It assists in the establishment of the owner's funding. A. Parameters Affecting Cost of Road Projects There are a lot of parameters that affected the cost of road projects, Hegazy and Ayed [2] relied on ten parameters to determine highway construction cost in Canada. Those parameters were project type, project scope, year, construction season, location, duration, size, capacity, water body, and soil condi- tion. Wilmot and Mei [6] took on price of labor, price of material, price of equipment, pay item quantity, contract duration, contract location, quar- ter in which contract was let, annual bid volume, bid volume variance, number of plan changes, and changes in standards or specifications for estimat- ing the escalation of highway construction costs over time in Louisiana state. Whilst for estimating the cost at conceptual phase of highway projects in poland and thailand, sodikov [7] picked out con- mailto:mhmod_85@hotmail.com mailto:mhmod_85@hotmail.com Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 25 struction factors which are predominant work ac- tivity (asphalt or concrete), work duration, pave- ment width, shoulder width, ground rise fall, aver- age site clear/grub, earthwork volume, surface class (asphalt or concrete), and base material (crushed stone or cement stab). likewise in the West Bank – Palestine Mahamid and Bruland [8] adopted construction factors. which are road length, pavement width, pavement thickness after compaction, asphalt hauls distance, pavement area, base coarse thickness after compaction, base coarse haul distance, the base coarse area, terrain condi- tion (semi even, hilly), soil drillability (good, poor), and soil suitability (good, poor) to predict the cost of road construction activities. Pewdum et al. [9] used traffic volume, topography, weather condition, evaluating date, construction budget, contract duration, % of as planned completion, and % of actual completion to forecast final budget and duration of a highway construction project during construction stage. Attal [10] employed norm cost estimate, location of the project, area location (ru- ral, urban, etc.), loops and ramps, new signal counts, construction length, sidewalks, curb and gutter, crossover count, average daily traffic, and geometric design standard for predicting highway construction cost in Virginia. As shown above the factors varied widely. This is referred to several reasons like the location of study and the tools, which was used to determine the parameters. A number of techniques were available to determine the influencing factors on road projects costs. Delphi technique is one of the- se methods. It provided the opportunity to evaluate knowledge based on the experience of individual practitioners and it is suitable for this research [11]. While this research focused on the "implementa- tion factors" that affected the budget of road pro- jects in the Gaza Strip. The research adopted the most nine influential factors in roads budgets, which are determined by using Delphi technique. B. Artificial Neural Networks (ANN) Neural networks are the preferred tool for many predictive data mining applications because of their power, flexibility, and ease of use [12]. A neural network is an adaptable system that can learn relationships through repeated presentation of data and is capable of generalizing to new, previ- ously unseen data [13]. During training, both the inputs (representing problem parameters) and out- puts (representing the solutions) are presented to the network normally for thousands of cycles. At the end of each cycle, or iteration, the network evaluates the error between the desired output and actual output. Then use this error to modify the connection weights according to the training algo- rithms used [14]. Over the last few years, the use of artificial neural networks (ANN) has increased in many areas of engineering. Many research in construction man- agement have been carried out to use ANN in vari- ous topics. Moselhi et al [15] are among the first scholars to research ANN as a promising management tool in construction. Following on from their work, Moselhi and Hegazy [16] used neural network methodology to markup estimation. Then ANN became widely spread and, it is used in construc- tion management. Recently Hola and Schabowicz [17] estimated earthworks execution time cost by means of artifi- cial neural networks. Wang and Gibson [18] stud- ied pre project planning and project success by using ANNs and regression models. Chen [19] de- veloped hybrid ANN-Case Based Reasoning (CBR) model for disputed change orders in con- struction projects and Oral et al [20] predicted productivity of the construction crew by using NN with supervised versus unsupervised learning. Many researchers focused on predicting the cost of construction by using NN, like Arafa and Alqedra [4] who developed ANN model to predict the early stage cost of buildings. The analysis of the training data revealed that there are seven key parameters. Kim et al [21] developed a hybrid conceptual mod- el for estimating cost of large building projects. Gunaydın and Dogan, [22] also, built a neural net- work model to estimate the cost in early phases of building design process. Cost and design data were used for training and testing the neural network methodology with eight design parameters utilized in estimating the square meter cost of reinforced concrete structural systems of 4–8 storey residen- tial buildings in Turkey, an average cost estimation accuracy of 93% was achieved. Likewise, in Korea Kim et al. [23] used the construction cost data for residential buildings constructed between 1997 and 2000. The back-propagation network (BPN) model incorporating genetic algorithms (GAs) were used to improve the accuracy of construction cost esti- mation. ANN technique was also used in highway pro- jects. Pewdum, Rujirayanyong and Sooksatra [9] presented a study of back-propagation neural net- works for predicting final budget and duration of highway construction projects by using the actual data collected from project progress reports of 51 http://www.sciencedirect.com/science/article/pii/095605219390039Y http://www.sciencedirect.com/science/article/pii/095605219390039Y Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 26 highway construction projects in Thailand between 2002 and 2007. Sodikov [7] focused on the devel- opment of a more accurate estimation technique for highway projects in developing countries at the conceptual phase using artificial neural networks. He used database of road works cost data from two developing countries Poland and Thailand, which have a relatively large number of projects. There- fore, they investigated the relationship between project cost and other variables such as work activ- ity, terrain type, road parameters, etc. The ANN model was developed by multilayer perception (MLP) with back-propagation algorithm. Wilmot and Mei [6] developed an artificial neural network model, which relates overall highway construction costs to improve a procedure that estimates the escalation of highway construction costs over time. The model was able to replicate past highway con- struction cost trends in Louisiana with reasonable accuracy. The multilayer feed-forward network structure for ANNs was chosen for this study, and for train- ing, the backpropagation learning algorithm was used. This research used multi-layer perceptron architec- ture of ANN applications to introduce a model for estimating the cost of road projects in Gaza strip at the early stage. III. METHODOLOGY The research was carried out to achieve the objec- tive of the study. In the first step, Delphi technique was used for determining the influencing imple- mentation factors on the cost of roads. Secondly, historical data of road projects implemented be- tween 2011 and 2012 in Gaza was collected. In the third step, this data was used in developing the neural networks proposed models. The model was tested on separate data for best-possible architec- ture. These steps will be explained in the following sections. IV. INFLUENCING FACTORS AND DATA COLLECTION To obtain the factors, which have the most effect on the cost of road projects, the Delphi method was utilized and the following steps were followed[1]:  Seven exploratory interviews were done with experts. The experts worked in various posi- tions: cost engineers, managers and site engi- neers. They worked with consulting offices, municipalities and contractors.  The factors that affect the cost of road projects, which have been drawn from previous studies were presented to them.  Then the experts' opinions were showed con- sensus on nine factors, which have the greatest impact on the road project cost in Gaza strip. Neural networks models require many data. There- fore, Eighty-six (86) projects were collected from municipalities, Ministry of Public Works and Housing, contractors and consultants. As a result for using Delphi technique, the follow- ing nine factors were recommended which cover the parameters of the cost of road projects influenc- ing: Project scope: The collected data includes 26 pro- jects that have a "new project with a good soil" scope, 27 projects that have a "new project with a poor soil" scope and 33 projects that have a "reha- bilitation" scope. This is a clear indication that the collected projects are distributed as per the project scope so it was considered a representative sample and can be used in modeling. Pavement type: The sample was representative of two types; 48 projects with asphalt pavement and 38 with interlock pavement. Pavement area: There is no data for projects, which have pavement area less than 2,000 m 2 , while 65% of projects in the data set have pave- ment area between 2000 to 10,000 m 2 . The low rate of projects that have an area more than 20,000 square meters may lead to reducing the accuracy of the model in estimating the cost for this category. Length of road: Gives an indication for the size of works in the project. The shortest length in the data set is 250m and the number of projects that have road lengths more than 2,250m is less than five percent. While 77% of the projects in the data set, have road lengths between 250 to 1,250 m. Sewage, water and lighting networks: The cost of projects in the data set contains the cost of im- plementing one or more of the networks listed in Table 1. The data represented all feasible possibili- ties in the presence or absence of the three net- works. This supports the possibility of the adequa- cy of this data to build the model. Curbstone length: A quarter of the projects in the data set do not include the cost of curbstones, which means it was well represented. The projects, which include curbstones with more than 5000 meter are few. This may reduce the accuracy of the model in estimating the cost of this category of projects. Pavement area of (side walk+ island): Forty-two percent of the projects in the data set did not in- clude the cost of paving sidewalks and islands. Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 27 Table 1: Number of projects including different networks Projects Budget for the gathered cases are present- ed in Figure 2. It is very clear that more than 67 percent of the projects in the data set have a budget less than 400,000 dollars. This means that the accu- racy of the model will be good for projects that have cost within this range. V. MODEL DEVELOPMENT To obtain the best models with minimum error, the research followed the procedures explained in Fig- ure 1. A. Model Structure Design The choice of ANN architecture depends on a number of factors such as the nature of the prob- lem, data characteristics and complexity, the num- ber of sample data, etc.[7]. The models designed to include an input layer of nine processing elements (neurons) corresponding to the nine input parame- ters and an output layer of one processing element (neuron) as the target. In this research, the data is textual and numeric, so it is encoded to be only numeric or integer according to Table 2. Table 2: Inputs/Output encoding. The design of the neural network architecture is a complex and dynamic process that requires the determination of the internal structure and rules (i.e., the number of hidden layers and neurons, up- date weights method, and the type of activation Networks No. of projects % Sewage only 7 %8.1 Water only 3 %3.5 Lighting only 14 %16.3 No Networks 35 %40.7 Sewage &Water 6 %7.0 Sewage &Lighting 4 %4.7 Water &Lighting 6 %7.0 Sewage, Water &Lighting 11 %12.8 No Input Parameters Code 1. Project scope New with a good soil = 1 New with a bad soil = 2 Rehabilitation = 3 2. Pavement type Interlock = 1 Asphalt = 2 3. Pavement area in m 2 4. Sidewalk & Island pavement area in m 2 5. Road length in meters length 6. Curbstone length in meters length 7. Water networks Exist = 1 Not exist = 0 8. Lighting networks Exist = 1 Not exist = 0 9. Sewage networks Exist = 1 Not exist = 0 Output Parameter Code 1 Project budget in thousand dollar's Figure 2: Number of projects according to their budget Train the model Does the performance of the testing set is acceptable? No Implement the model Design the model structure Varying the size of the network and/or learning parameters Save the model that has the best performance Has acceptable performance been reached? Test the model End Save the model Start No Yes Yes Figure 1: Modelling Procedures Flowchart [1]. Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 28 function) [22]. This research depended on the backpropagation algorithm, which is a type of supervised learning algorithms that is mostly used in civil engineering applications. Also, Levenberg-Marquardt learning rule is selected. The choice of ANN in this study is based on optimum design and prediction using multilayer perceptron neural network architectures. NeuroSolution 6.07 application and Microsoft Ex- cel 2007 were selected to build the models. There are many types of activation functions, which are used to transform an input signal into output. The hyperbolic tangent (Tanh) was used. B. Model Implementation The problem at hand needs to identify and tag the data as input or as output. So the data was orga- nized in the preliminary stage to neural network modeling. Then three processes were followed in modeling: Data Sets: Any model selection strategy requires validation by the process data. Traditionally, avail- able data is divided into three sets [24]; training set (in-sample data), cross-validation set and a test set (out-of-sample). Learning is performed on the training set, which is used for estimating the arc weights while the cross validation set was used for generalization that is to produce better output for unseen examples [7]. However, the test set is used for measuring the generalization ability of the net- work and network performance evaluation [25]. The total available data is 86 exemplars that are divided randomly into three sets:  Training set (includes 60 exemplars ≈ 70%),  Cross validation set (includes 16 exemplars ≈ 18%) and  Test set (includes 10 exemplars ≈ 12%). Normalizing Data: Data is generally normalized for confidentiality and for effective training of the model being developed. The normalization of train- ing data is recognized to improve the performance of trained networks [22]. The input/output data is scaled, zero is the lower bound and the upper bound is one to suit neural networks processing. NeuroSolution 6.07 automat- ically scales input values to {Lower Upper} ac- cording to Equations (1), (2) and (3). (Data ) = Amp × Data + Off ‎0 (1) Where: (Datai)Nor: data represent value for one input for one sample after normalization. Datai: data represent value for one input for one sample. Amp = ( ) ( ) (2) Off = UperBound − Amp × Max (3) Where Maxi and Mini are the maximum and mini- mum values found within channel i, and Upper- bound and Lower-bound are equal 0 and 1 respec- tively. Initial Networks Building: The modeling was started with small networks and increased their size until the performance in the test set is appropriate. This proposed method of growing neural topolo- gies ensures a minimal number of weights, but the training can be fairly long [13]. C. Training Models and Testing Training a NN is an iterative process of feeding the network with the training examples and changing the values of its weights in a manner that is math- ematically guaranteed to reduce consecutively the error between the network's own results and the desired output. Neural networks are able to gener- alize solutions to problems by learning from pairs of input patterns and their associated output pattern [16]. After building a small topology as viewed above, training with cross-validation and testing phase will begin. The optimum architecture is often achieved by trial and error according to the com- plexity of the respective problem, also by testing few proposed designs to select the one that gives the best performance[26]. Figure 3 explains the se- ries of processes to get the best weights, which give the minimum percentage error. VI. PERFORMANCE MEASURES The Performance Measures are important to evalu- ate the models. There are five values that can be used to measure the performance of the network for a particular data set. Mean Square Error (MSE): According to Princi- pe et al [13] the MSE formula is: MSE = ∑ ( ) (4) Where: N= number of exemplars in the data set. yij= network output for exemplar i at PE j. dij= desired output for exemplar i at PE j. Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 29 Correlation Coefficient (r): According to Principe et al [13], the correlation coefficient between a network output x and a desired output d is: 𝑟 = ∑ ( ̅)( ̅) √∑ ( ̅) √∑ ( ̅) (5) Mean Absolute Error (MAE): According to Willmott and Matsuura [27], the MAE is de- fined by the following formula: MAE = ∑ | | (6) Where: N= number of exemplars in the data set. dyij= denormalized network output for exemplar i at PE j. ddij= denormalized desired output for ex- emplar i at PE j. Mean Absolute Percentage Error (MAPE): Ac- cording to Principe et al., (2010) [13], the MAPE is defined by the following formula: MAPE = ∑ | | (7) This research considered Hegazy and Ayed [2] methodology in determining the total MAPE. The training phase was represented by fifty percent of the total MAPE while the test set equals the re- maining fifty percent. Total MAPE can be calculat- ed by the following formula: Total MAPE = ( APE × + APE × ) ( + )⁄ + APE 2 (8) Where: MAPET= MAPE for training data set. NT = number of exemplars in the training data set. MAPEC = MAPE for cross validation data set. NC= number of exemplars in the cross validation data set. MAPES = MAPE for test data set. Total Accuracy Performance (TAP): According to Wilmot and Mei [6], the accuracy performance is defined as (100−MAPE) %. Total Accuracy Per- formance (TAP) can be calculated by the following formula: TAP = 100 − Total MAPE (9) VII. RESULTS AND DISCUSSION From the previous procedure of training and test- ing, many multilayer perceptron topologies were trained for several trials. The best structure has one hidden layer with five neurons although two hidden layers topologies were trained, see Figure 4. The models were trained on sixty exemplars while sixteen exemplars of cross validation set were used for generalization to produce better output for un- seen examples. The models were tested on ten ex- emplars. The results are summarized in Table 3. Determine the total number of epochs Use C.V data Randomize the networks weights Vary the size of the network Save the model Perform the sensitivity analysis End Start Has acceptable performance been reached? n=0 n=n+1 If n=N No Yes No Test the model Does the performance of the testing set is acceptable? Yes No Yes n: is a counter number. N: is a number of the required runs. Figure 3: Training and testing model flowchart [1]. Figure 4: The architecture of the MLP model. Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 30 Table 3: Performance measurements for the model. Training set C.V set Test set MSE 152.2 1446 2819 R 0.999 0.992 0.99 MAE 8.812 30.97 43.3 MAPE 3.49% 9.59% 7.80% AP 96.52% 90.41% 92.20% The back-propagation algorithm involves the grad- ual reduction of the error between model output and the target output. It develops the input to out- put by minimizing a mean square error cost func- tion measured over a set of training examples [28]. The value of the MSE for MLP model has MSE 152, 1446 and 2819 at training, cross validation and testing sets respectively. The size of the mean square error (MSE) can be used to determine how well the network output fits the desired output, but it does not necessarily re- flect whether the two sets of data move in the same direction. For instance, by simply scaling the net- work output, we can change the MSE without changing the directionality of the data. The correla- tion coefficient (r) solves this problem [13]. As show in Table 3 the correlation coefficient (r) for any set of data is not less than 0.989, this means that the fit of the model to the data is reasonably good. Mean absolute error is another factor to measure the models performance. The MLP model has MAE of 8.8, 31 and 43.3 at training, cross valida- tion and testing sets respectively. Note that MAE factor alone is not enough because its value can easily be misleading. For example, say that output data is in the range of 0 to 10. For one exemplar, the desired output is one and the actual output is two. Even though the two values are quite close and the MAE for this exem- plar is one but the mean absolute percentage error is 100. Therefore, this research used the MAPE. The values of the MAPE for the MLP model were 3.5, 9.6 and 7.8 at training, cross validation and testing sets respectively. The accuracy of the best model developed by mul- tilayer perceptron sounds very favorably with data based from the test set. It can be seen from the re- sults that the model performs well and no signifi- cant difference could be discerned between the estimated output and the desired budget value. Re- sults of training, cross validation and test set are shown in Figure 5, Figure 6 and Figure 7 respectively. An average accuracy of 93.7% was achieved, this means that the total MAPE equals 6.3%. The pre- vious results show that MLP model has excellent performance with minor error. As shown in Figure 5, Figure 6 and Figure 7 perfect agreement between the actual and predicted values draws a 45-degree line; this line means that the actual cost values equal the predicted ones. Figure 5, Figure 6 and Figure 7 indicate reasonable con- Figure 5: Desired output and actual network output for training set exemplar. Figure 6: Desired output and actual network output for cross validation set exemplar. Figure 7: Desired output and actual network output for test set exemplar. Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 31 centration of the predicted values around the 45- degree line. The coefficient of determination be- tween the actual and the predicted cost values were 0.998, 0.992 and 0.99 for training, cross validation and test set respectively. VIII. SENSITIVITY ANALYSIS Sensitivity analysis is the method that discovers the cause and effect relationship between input and output variables of the network [13]. The NeuroSolution program provides a useful tool to identify sensitive input variables called ‘‘Sensi- tivity about the Mean’’. The sensitivity analysis was run by batch testing on the MLP model after fixing the best weights then started by varying the first input between the mean ± one standard devia- tion, while all other inputs are fixed at their respec- tive means. The network output was computed for 50 steps above and below the mean. This process was then repeated for each input. Figure8, summa- rizes the variation of output with respect to the variation of each input generated. As shown in Figure 8, the pavement area parame- ter has the value 56.6 that is the greatest effect on the budget output. The second parameter affecting the total budget is pavement type, which has 41.15. These results are logical when compared to actual practice. On the other hand, project scope has a weak impact; likewise, road length has the weakest impact, which may be due to the presence of the pavement area parameter. IX. CONCLUSION This research was achieved the ability to estimate the road projects cost at early stage with high accu- racy and minor error. MLP model had error rate equal 6.3% and MAPE 3.49 %, 9.59% and 7.8% for training, cross validation and test sets respec- tively. In addition, the value of correlation coeffi- cient does not less than 0.989 for any set. This research focused on the "implementation fac- tors" that affected the budget of road projects in the Gaza Strip. The research adopted nine factors, which are determined by using Delphi technique. The remarkable, that the sensitivity analysis results were very logical and showed the impact of each parameter on the cost. Which the pavement area parameter had the greatest effect on the budget output. Nevertheless, project scope and road length had low impact. ANN are well suited to model complex problems where the relationship between the model variables is unknown. Also, ANN does not need any prior knowledge about the nature of the relationship be- tween the input/output variables, which is one of the benefits that ANN has compared with most empirical and statistical methods. Although the ANNs have advantages, on the other hand there are disadvantages. The principal disadvantage being that they give results without being able to explain how they were arrived to their solutions. Their ac- curacy depends on the quality of the trained data and the ability of the developer to choose truly representative sample inputs. In addition to trial and error method is the best solution to obtain the formula to decide what architecture of ANN should be used to solve the given problem and which training algorithm to use. One looking at a problem and decide to start with simple networks or going on to complex ones to get the optimum solution is within the acceptable limits of error. REFERENCES [1] H. K. AbuJamouse, "Parametric Cost Estimation of Road Projects Using Artificial Neural Networks," Thesis for the fulfillment of master of scince Master Thesis, Civil Engineering, Islamic University, Gaza, 2013. [2] T. Hegazy and A. Ayed, "Neural network model for parametric cost estimation of highway projects," Journal of Construction Engineering and Management, vol. 124, pp. 210-218, 1998. [3] Y. A. E.-R. Al-Shanti and K. Sha'at, "A Cost Estimate System for Gaza Strip Construction Contractors," Master, Faculty of Engineering, The Islamic University, Gaza, 2003. Figure 8: sensitivity output Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 32 [4] M. Arafa and M. Alqedra, "Early stage cost estimation of buildings construction projects using artificial neural networks," Journal of Artificial Intelligence, vol. 4, pp. 63-75, 2011. [5] A. S. Ayed, "Parametric Cost Estimating of Highway Projects using Neural Networks," Master, Faculty of Engineering & Applied Sciences, Memorial University, Newfoundland, 1997. [6] C. G. Wilmot and B. Mei, "Neural network modeling of highway construction costs," Journal of construction engineering and management, vol. 131, pp. 765-771, 2005. [7] J. Sodikov, "Cost estimation of highway projects in developing countries: Artificial neural network approach," Journal of the Eastern Asia Society for Transportation Studies, vol. 6, pp. 1036- 1047, 2005. [8] I. MAHAMID and A. BRULAND, "Preliminary Cost Estimating Models for Road Construction Activities," in Facing the Challenges – Building the Capacity, Sydney 2010. [9] W. Pewdum, T. Rujirayanyong, and V. Sooksatra, "Forecasting final budget and duration of highway construction projects," Engineering, Construction and Architectural Management, vol. 16, pp. 544- 557, 2009. [10] A. Attal and M. Tatari, "Development of Neural Network Models for Prediction of Highway Construction Cost and Project Duration," Master, the Department of Civil Engineering, Ohio University, Ohio, 2010. [11] G. D. Creedy, M. Skitmore, and T. Sidwell, "RISK FACTORS LEADING TO COST OVERRUN IN THE DELIVERY OF HIGHWAY CONSTRUCTION PROJECTS," PhD, Faculty of Built Environment and Engineering, Queensland University of Technology, Queensland, 2006. [12] M. J. Noruis, SPSS Neural Networks™ 17.0. Chicago: SPSS Incorporated, 2007. [13] J. Principe, W. Lefebvre, G. Lynn, C. Fancourt, and D. Wooten, "NeuroSolutions-Documentation, the Manual and On-Line Help," 2010. [14] J. L. McClelland and D. E. Rumelhart, "Parallel distributed processing," Psychological and biological models, vol. 2, 1987. [15] O. Moselhi, T. Hegazy, and P. Fazio, "Neural networks as tools in construction," Journal of Construction Engineering and Management, vol. 117, pp. 606-625, 1991. [16] O. Moselhi and T. Hegazy, "Markup estimation using neural network methodology," Computing systems in engineering, vol. 4, pp. 135-145, 1993. [17] B. Hola and K. Schabowicz, "Estimation of earthworks execution time cost by means of artificial neural networks," Automation in Construction, vol. 19, pp. 570-579, 2010. [18] Y.-R. Wang and G. E. Gibson, "A study of preproject planning and project success using ANNs and regression models," Automation in Construction, vol. 19, pp. 341-346, 2010. [19] J.-H. Chen, "Hybrid ANN-CBR model for disputed change orders in construction projects," Automation in Construction, vol. 17, pp. 56-64, 2007. [20] M. Oral, E. L. Oral, and A. Aydın, "Supervised vs. unsupervised learning for construction crew productivity prediction," Automation in Construction, vol. 22, pp. 271-276, 2012. [21] H.-J. Kim, Y.-C. Seo, and C.-T. Hyun, "A hybrid conceptual cost estimating model for large building projects," Automation in Construction, vol. 25, pp. 72-81, 2012. [22] H. Murat Günaydın and S. Zeynep Doğan, "A neural network approach for early cost estimation of structural systems of buildings," International Journal of Project Management, vol. 22, pp. 595-602, 2004. [23] G.-H. Kim, J.-E. Yoon, S.-H. An, H.-H. Cho, and K.-I. Kang, "Neural network model incorporating a genetic algorithm in estimating construction costs," Building and Environment, vol. 39, pp. 1333-1340, 2004. [24] M. Ghiassi, H. Saidane, and D. Zimbra, "A dynamic artificial neural network model for forecasting time series events," Hasan KH. Abujamous, Rifat N. Rustom, and Mahmoud Y. Abukmail (2014) 33 International Journal of Forecasting, vol. 21, pp. 341-362, 2005. [25] G. Zhang, B. Eddy Patuwo, and M. Y Hu, "Forecasting with artificial neural networks:: The state of the art," International journal of forecasting, vol. 14, pp. 35-62, 1998. [26] N. Bakhary, K. Yahya, and N. Ng Chin, "Univariate Artificial Neural Network In Forcasting Demand Of Low Cost House In Petaling Jaya," Jurnal Teknologi B, pp. 67-75, 2004. [27] C. J. Willmott and K. Matsuura, "Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance," Climate Research, vol. 30, p. 79, 2005. [28] M. Bouabaz and M. Hamami, "A cost estimation model for repair bridges based on artificial neural network," American Journal of Applied Sciences, vol. 5, pp. 334- 339, 2008. Hasan Kh. AbuJamous has M.Sc. Civil Engineering from Islamic University of Gaza in Palestine in 2013. He is a project manager in El-Jazeera Construction Compa- ny for Consulting (JCEC). His primary research inter- ests include Neural Networks, Simulation, Construction Management, Construction Analysis and Concrete Tech- nology. Prof. Rifat Rustom has M.Sc. and Ph.D. in Civil Engi- neering from Drexel University in the U.S.A. in 1993. He is the Rector of the University College of Applied Sciences (UCAS). Prof. Rustom is former Vice President for External Affairs and IT at the Islamic University of Gaza (IUG) Prof. Rustom has research interests in Con- struction Management, Institutional Management and Development, Geosynthetics, and Concrete Technology. Mahmoud Y. AbuKmail has M.Sc. Civil Engineering from Islamic University of Gaza in Palestine in 2013. He is a project manager in El-Jazeera Construction Compa- ny for Consulting (JCEC). His primary research interests include Simulation, Neural Networks, Construction Managements, and Computer Applications in Construc- tion.