Expert Systems with Applications 38 (2011) 54565462

Contents lists available at ScienceDirect

Expert Systems with Applications

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

Project delivery system selection of construction projects in China

Yong Qiang Chen a, Jun Ying Liu a,, Bingguang Li b, Binshan Lin c
a

School of Management, Tianjin University, Tianjin 300072, China Harry F. Byrd, Jr. School of Business, Shenandoah University, 1460 University Drive, Winchester, VA 22601, USA c Department of Management and Marketing, Louisiana State University in Shreveport, Shreveport, LA 71115, USA
b

a r t i c l e

i n f o

a b s t r a c t
The suitability of the project delivery system (PDS) selected for a project greatly inuences the efciency to conduct a project. It is not an easy task to select an appropriate PDS as a large amount of ambiguous information exists. The paper aims to develop a PDS selection model to help owners to make a decision. The similar projects are identied through the similarity metrics between the target project to be predicted and those in the database. In addition, some of the indicator values are examined and modied through DEA-BND model, and then they are trained by ANN model to predict an appropriate PDS for the target project. A survey was conducted by postal questionnaire to empirically validate the reliability of the model in China. The indicator system for the selection of an appropriate PDS is established. Through the comparison of results predicted by different models, it is found that the PDS selection model developed in this paper can predict PDS precisely, and shows higher reliability than the ANN model. A PDS selection model is developed by inputting project-specic data, which proves to be more accurate and less dependent on experts judgment. Its practical application will benet the owners decision making in the selection of the PDS. 2010 Elsevier Ltd. All rights reserved.

Keywords: Construction project Project delivery system Data envelopment analysis Articial neural network China

1. Introduction Project delivery system (PDS) describes how the project participants are organized to interact, transforming the owners goals and objectives into nished facilities (ASCE, 2000). There are several PDSs available in Chinas construction industry, such as Design-Bid-Build (DBB), Design-Build (DB), Construction Management (CM) and Engineering Procurement and Construction (EPC) (Chen, Lu, & Zhang, 2005). PDS, as one of the critical factors of project success, affects project schedule, cost, quality and contract management to a great extent (Al Khalil, 2002; Chan, Yung, Lam, Tam, & Cheung, 2001; Kumaraswamy & Dissanayaka, 2001). Selecting an appropriate PDS can improve the project performance effectively (Hong, Kim, Kim, & Leem, 2008; Ojiako, Johansen, & Greenwood, 2008; Oyetunji & Anderson, 2006). Project performance under different PDSs can be predicted to serve as a basis of PDS selection. Konchar and Sanvido (1998) empirically compared cost, schedule and quality performance of CM at risk, DB and DBB, using project-specic data collected from 351 US building projects. However, the author did not illustrate

Corresponding author. Tel.: +(86)13802136518.

E-mail addresses: chenyongqiang@tju.edu.cn (Y.Q. Chen), liujunying@tju.edu.cn (J.Y. Liu), bli@su.edu (B. Li), blin@lsus.edu, Binshan.Lin@lsus.edu (B. Lin). 0957-4174/$ - see front matter 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.10.008

how to apply the research into PDS selection. Ling and Liu (2004) developed models to predict performance of DBB and DB projects in Singapore through multivariate analysis. Some other methods like AHP were also applied in this eld. The priority of PDSs can be determined through the pairwise comparison matrix (Al Khalil, 2002; Mahdi & Alreshaid, 2005). The accuracy of AHP is interfered by the experts uncertain and subjective judgments. Mafakheri (2007) utilized the Interval AHP to determine the interval priorities for alternative PDSs and Rough Set theory to fully rank the alternatives. But the full ranking depends on a higher risk, which increases the inaccuracy. Moreover, AHP tends to require a set of established indicators, including project participants, project characteristics and external environment (Alhazmi & McCaffer, 2000; Mafakheri, Dai, Slezak, & Nasiri, 2007; Mahdi & Alreshaid, 2005). It is very complex if a large number of indicators are used. Careful selection of indicators is needed to reduce the number as well as their correlation. Multi-attribute utility can be used for PDS selection (Chan et al., 2001; Love, Skitmore, & Earl, 1998). The overall utility is calculated by multiply the weights by the utility of indicators. Speed, certainty, exibility, quality, complexity, risk allocation, responsibility, arbitration and dispute, and price competition are often identied as its common indicators. The simplicity makes this model easy for practical use. However, the utility values of indicators always fail to reect the actual status and the project may not achieve the specied objectives as initially expected.

Y.Q. Chen et al. / Expert Systems with Applications 38 (2011) 54565462

5457

Case-based reasoning (CBR) is a name given to a reasoning method that uses specic experiences rather than a corpus of general knowledge (Ribeiro, 2001). Luu, Ng, and Chen (2003a, 2003b, 2005, 2006) conducted a few researches of its application in PDS selection. The cases (i.e. projects), similar with the target one, are retrieved by calculating the similarity metrics between target and exemplar ones. Luu et al. (2005) proposed indicators for CBR from the perspectives of owners characteristics and objectives, project characteristic and external environment. However, the CBR oversimplied the selection process. The overall similarity metric is measured by linearly calculating similarity of respective indicators. Articial neural network (ANN) is extensively introduced as an attempt to overcome the existing defects (Zaghloul, Lee, & Trimi, 2009). For example, over 12% of papers published from 1995 to 2005 (54 out of 445) in the ASCE Journal of Computing have used the term of neural in title (Flood, 2008). Ling, Chan, Chong, and Ee (2004) developed a nonlinear prediction method using ANN to predict the performance of DB projects, which could be used as a basis of PDS selection. But it is not easy to directly apply ANN models to the PDS selection because of the difculty in acquiring the real-world project data. Data envelopment analysis (DEA) is used to measure the efciency of DMU (Decision Making Unit) (Charnes, Cooper, & Rhodes, 1978). Lo and Chao (2007) applied DEA to calculate the efciency of DBB, DB, CM and DBM in road projects. It is necessary to improve/enhance the accuracy and effectiveness in prediction of the current existing methods because the PDS plays an important role in Chinas rapidly developing construction industry. Therefore, the aim of this paper is to develop an accurate and reliable PDS selection model to overcome the disadvantages of the current existing problems. 2. Indicators of the PDS selection 2.1. Articial neural network Articial neural network (ANN) does not fully rely on existing experiences and knowledge, but extract the internal rules from the training samples using specied learning rules, through which
Table 1 List of the indicator of the PDS selection. Indicators Literature A Project objectives Delivery speed Schedule delay Cost growth Cost certainty Quality performance project characteristics Project type Project scale Complexity Ability to dene the project scope Flexibility Disputes Characteristics of owner and contractor Owners willingness to be involved Owners willingness to take risks Owners available human resources Contractors capability External environment Market competitiveness Regulatory feasibility Technology availability B C D E

the internal relationship between inputs and outputs can be built. Compared with the multiple regression method, ANN is not need to consider the weights and has stronger self-adaptation and fault-tolerant abilities. In this paper, the PDS is dened as the output of ANN, while indicators of the PDS selection are dened as its inputs. 2.2. Data envelopment analysis Data envelopment analysis (DEA) does not need pre-estimated parameters. It has superiority in avoiding subjective inuence, simplifying calculation and reducing deviation (Chen, 2008). In this research, DEA-BND (Bound Variable) model, one of the DEA models, is selected to examine the validity of project data, and then the indicator values are correspondingly modied. Its most significant characteristic is to allow the input and output indicators to uctuate between the specied bounded ranges. Those data with signicant deviation shall be modied by DEA-BND model for two reasons: rstly, the average Production Possibility Frontier (PPF) can be generated, and the signicant deviation of individual DMU is reduced by adjusting input indicators; Secondly, the efciency value of a project is changed into 1 (maximal efciency) by adjusting input indicators into projection values, which replaces its original values to achieve the modication of indicator values. 2.3. Indicators of the PDS selection The indicator system is an integral part of the prediction model. Establishment of a universal set of indicators for PDS selection has become a black hole to attract researchers on PDS selection. It is not necessary to hold up all the technical development to wait for establishing such indicators (Chan, 2007). Based on the literature review mainly shown in Table 1, the indicators can be categorized into four groups, namely project objectives, project characteristics, characteristics of owner and contractor and external environment. As to indicators with the same or similar meanings but with different expressions, only one expression form is selected here for purpose of clarication and unication. For example, exibility

U U U

U U U U U

U U

U U

U U U U U

U U U

U U U

U U U

U U

U U U

U U U U

U U U U U

U U

U U

U U

U U U

U U U U U U U U U

U U

U U U U U

U U

U U U

U U U U U U U U U

U U U U U U U U U

U U U

U U U

U U

U U

U U

U U

U U U U U U

U U

U U U

Note: A = Kumaraswamy and Dissanayaka (1996); B = Konchar and Sanvido (1998); C = Molenaar and Songer (1998); D = Gransberg et al. (1999); E = Alhazmi and McCaffer (2000); F = Kumaraswamy and Dissanayaka (2001); G = Cheung et al. (2001); H = Ng et al. (2002); I = Al Khalil (2002); J = Luu (2003); K = Ling and Liu (2004); L = Mahdi and Alreshaid (2005); M = Luu et al. (2005); N = Oyetunji and Anderson (2006); O = Mafakheri et al. (2007)

5458

Y.Q. Chen et al. / Expert Systems with Applications 38 (2011) 54565462

is selected to represent the indicator Potential for design changes during construction. Project schedule, cost and quality are commonly used as project objectives and they are also identied for the PDS selection. Schedule delay is the most frequently used indicator for project schedule. Delivery speed is eliminated because it is mainly used for building projects and not comparable among different project types. Therefore, as shown in Table 2, only Schedule delay is selected to represent the project schedule. As to project cost, Cost growth and Cost certainty are both frequently used. Cost certainty means the certainty over the cost for completion of the project. As seen from Table 1, most researchers selected only one of them, because a high correlation exists between them. Only Cost growth is selected in this paper, shown in Table 2. Project quality is measured with varied indicators by different researchers. In this paper, the formula developed by Ng, Luu, Chen, and Lam (2002) is adopted to represent the level of Project quality, which is divided into Material quality, Workmanship quality and Design quality, shown in Table 2. For Project characteristics, six indicators listed in Table 1 are selected, which is shown in Table 3. The formulas of Project
Table 2 Description of indicators of project objectives. Indicators Description
durationNormal anticipated duration Normal anticipated duration Actual contract priceOriginal contract price Original contract price Material qualityWorkmanship qualityDesign quality 3 of money paid for extra material fees for better quality than the normal Total material fee of the normal money paid for extra workmanship fees for better quality than the normal Total workmanship fee of the normal of money paied for extra design fees for better quality than the normal Total design fee of the normal

Schedule delay Actual Cost growth Project quality

Where: Material quality Amount Workmanship quality Design quality

complexity and Flexibility are selected from the paper of Ng et al. (2002). Project type includes three types, which are industrial, infrastructure and building projects (Shen, Lu, & Yam, 2006). Industrial projects include power facilities, mining projects, smeltery projects, petrochemical engineering projects, and facilities for communication; infrastructure projects includes road facilities, railway facilities, water transport facilities, hydroelectricity facilities and urban infrastructure projects; building projects includes building and residential facilities, and mechanical and electric installation in building projects. For characteristics of owner and contractor, the ve indicators listed in Table 1 are selected, shown in Table 4. The formulas of Owners willingness to be involved and Owners willingness to take risks are selected from the paper of Ng et al. (2002). For External environment, the indicators are measured by scale 19. As one of the widely used indicators, Market competitiveness means that level of competition in market with regards to this project (Luu et al., 2005) and is subject to the available contractors. 1 means low competitiveness and limited available contractors, 9 means high competitiveness and extensively available contractor. Regulatory feasibility measures the degree to which the regulations inuence the implantation of project. 1 means adverse inuence, 9 means favorable inuence. Technology availability measures the degree of availability of necessary technologies from market. 1 means quite hard to acquire necessary technologies, 9 means quite easy to acquire technologies. The indicator value is expressed in the form of the triangular fuzzy number, which involves more information than the crisp number or expected value. In the survey, respondents are required to give the triangular fuzzy number which includes three numbers, two are the boundary value and one is the most probable value. 3. PDS selection model By combined usage of ANN and DEA, the proposed model is composed of three parts: Selecting Similar Projects, Examining Indicator Values, and Training and Predicting. The framework of the model can be shown in Fig. 1. 3.1. Stage 1 selecting similar projects Similar projects are selected to increase the prediction accuracy of the ANN model. The rst step is to establish a project database including the needed project data. It is assumed that the project database has been established and has enough project data. Then, the similar projects can be selected according to the procedure shown in Fig. 2. As shown in Fig. 1, the rst step is to retrieve projects from the project database. Based on the indicator of project type, projects of the same type are chosen as the target ones. Then, retrieval II is executed to select the most similar projects according to the indicators of project scale, project complexity and external environment. The critical value of the similarity metrics should be specied to determine which projects are selected as similar projects. In order to ensure selecting enough similar projects with high degree of similarity, the project, whose similarity metric exceeds the critical value, would be selected and the others will be eliminated. Some indicator values are difcult to dene because the target project has not started when the PDS selection is made. By inputting different values, the alternative PDSs are predicted, from which the owner can choose the optimal one. The similarity between the target project and projects in database should be calculated when retrieval II is implemented. The

Amount of Amount

Table 3 Description of indicators of project characteristics. Indicators Project type Project scale Description Industrial, infrastructure and building projects Project scale is measured with scale 19 to represent the amount of project cost 1 = much low than the average level 9 = much higher than the medium

working packages resulting need for Project Number of extra total working packages of a from thewith basic extra complexity Number of project complexity complexity Variation caused by the indefinite project scope at contract signing Ability to dene 1 Total v ariation the project scope Flexibility Anticipated v ariations Original contract sum

Disputes

Loss

directly caused by disputes Original contract sum

Table 4 Description of indicators of characteristics of owner and contractor. Indicators Owners willingness to be involved Owners available personnel Owners willingness to take risks Contractors capability Description Total
time of direct inv olv ement of owner in the project Total project time

Working hours expected to be spent by owner Owner0s av ailable working hours Variations incurred to the client only Total v ariations This indicator is dened by scale 19. 1 means low level in specialty and management capabilities while scale 9 means high level in these capabilities.

Y.Q. Chen et al. / Expert Systems with Applications 38 (2011) 54565462

5459

Selecting Similar Projects

Examining Indicator Values

Training and Predicting

Project Databse

Retrieval I Indicators of Target Project Retrieval II Similar Projects Modified Similar Projects

Training of BP ANN Prediction of BP ANN Predicted PDS

External Environment Characteristics of Owner and Contractor Project Characteristics

DEA- BND Model

Project Objectives
Fig. 1. Framework of the PDS selection model.

similarity between the target project T and project R from database can be calculated by the following formula,

ST; R

in X i1

wi sim fiT ; fiR

Thereinto, n means the number of indicator; wi means the Pin weight of each indicator wi P 0; i1 wi 1. The weights are t directly given by the owner. fi and fiR respectively represent the value of indicator i of project T and project R, and the value is either crisp or fuzzy. sim fiT ; fiR = the similarity between project T and project R. Both the range of s(T, R) and sim fiT ; fiR are between 0 and 1. The similarity increases as the value of s(T, R) becomes larger. When it reaches 1, the two projects are completely matched. sim fiT ; fiR can be calculated in two ways considering the following two situations, (1) When the indicator value is crisp number,

Thereinto, fiT and fiR respectively represent the value of indicator i of the target project and projects in the database;a and b respectively represent the upper and lower bound of the value. (2) When the indicator value is fuzzy, and if the indicator i of T T T T T T project T and project R satisfy, T fi1 ; fi2 ; fi3 ; fi1 6 fi2 6 fi3 fi R R R R R f ; f ; f ; f 6 f R 6 f R , then Graded Mean Integraand f i i1 i2 i3 i1 i2 i3 tion-representation Distance (Hsieh & Chen, 1999) method is applicable.

 1 sim fiT ; fiR 1 p fiT p fiR 

Thereinto,
T T f T 4fi2 fi3 p fiT i1 6 R R f R 4fi2 fi3 p fiR i1 6

4 5

 T  f f R  T R i i sim fi ; fi 1 ba

3.2. Stage 2 examining indicator values Part of the indicator values of similar projects are examined and modied in order to be effectively applied into the ANN model. A DEA-BND (Bound Variable) model is selected to evaluate the efciency of owners management during the construction stage (owners management during this stage is DMU here). In order to apply DEA-BND model, the value of each input indicator is expressed by triangular fuzzy number, which includes two boundary number and one probable number. Their boundary ranges can be obtained by questionnaire. By measuring the efciency of projects using the DEA-BND model, the projection values of each input indicator can be calculated and then substitute the original values as amendments to modify the input indicators. (1) Input indicators. Input indicators are chosen from three aspects, time, personnel and expense. The three indicators, Owner participation, owners available personnel and project scale, respectively represent time, personnel and expense inputs. Since the owners management is affected by contractors capability, the indicator contractors capability is also selected here.

Target Project

Project Database

Retrieval I

Project Type

Retrieval II Similarity Calculation Selecting Critical Value

Project Scale/ Project Complexity External Environment

Similar Projects
Fig. 2. Selection procedure of similar projects.

5460

Y.Q. Chen et al. / Expert Systems with Applications 38 (2011) 54565462

(2) Output indicators. As the owners management is to be evaluated by DEA-BND model, it is necessary to convert the project performance into indicators that can represent the output of owners management. This paper proposes the indicator of Project output satisfaction as the output indicator of DEA. It can measure the owners satisfaction to the outputs of management activities. Project output satisfaction can be generated from three aspects which are schedule, cost and quality. The formula is as follows,

Sa

aT bCcQ abc

where, S means project outputs satisfaction, T means owners schedule management satisfaction, C means owners cost management satisfaction and Q means owners quality management satisfaction. a, b and c, respectively represent the importance weights of T, Q and C, which can be specied by owner. As satisfaction is quite difcult to be measured accurately, in this paper it is acquired from the value of schedule, cost, and quality management satisfaction, which can be respectively obtained by converting schedule delay, cost growth and project quality into values between from 1 and 100 using linear calculation, convex function or concave function. 3.3. Stage 3 training and predicting Applying the ANN into training the modied similar projects, and formulating the nonlinear function between indicators and PDS to predicate which PDS is suitable for the target project. Back Propagation (BP) ANN is a multi-layer feed-forward neural network based on error back-propagation algorithm. In the application of ANN model, BP ANN is widely used. Therefore BP ANN is applied here to train and predict the similar projects. The correlation of input indicators needs to be tested before the application of BP ANN model. Factor analysis can be used to reduce or eliminate the correlation among the input indicators. A potential precondition of factor analysis is that there should be strong correlation among indicators. By KMO (KaiserMeyerOlkin) and Bartletts test of sphericity, the correlation coefcient should exceed the commonly used value 0.3. For BP ANN, rstly set the random number between (1, 1) as the original weights and design one hidden layer. The transfer function from the input layer to the hidden layer is single-stage Sigmoid function and the transfer function from the hidden layer to the output layer is linear function. Secondly the minimum error is set as 0.01. Then the PDS can be predicted by inputting the indicators of the target project into the trained BP ANN model.

152 questionnaires in total were sent out to the managers of owners and other participants of projects undertaken by the top 51 Chinese Contractors ranked by ENR in 2008. After 3-month questionnaire survey from October 2008 to January 2009, 132 questionnaires were returned, and 81 questionnaires were valid for other 52 questionnaires had not lled out the form. Among the projects referred in the questionnaires, the mount of up to 90% of them were more than RMB 10 million; over 70% were invested more than RMB 100 million; 53% of them were DB projects (DB projects in this paper contains EPC projects); 47% of them are DBB projects. The three project types of the DB projects, i.e. industrial, infrastructure, and building projects, respectively account for 60%, 28% and 12%; while as to DBB projects, the three project types respectively account for 26%, 26% and 48%. 5. Data analysis 5.1. PDS prediction (1) The Project #003 is randomly selected as the target project, and other projects are treated as projects in a project database. Considering the number of questionnaire is small, it is impossible to take account of all the project types. Thus the Retrieval I is omitted. When calculating the similarity metrics in the part of Retrieval II, this research assumes the weight of each indicator is equal. So, it is calculated that the maximum value is 0.89, and the minimum value is 0.35. If set critical value as 0.7, there will be 31 projects exceeding 0.7, including 7 DB and 14 DBB projects. (2) Before applying the DEA-BND model, the boundary ranges of input indicators can be found from the received questionnaires. The output indicator is converted by indicators of project objectives using linear function. Therefore, the input indicators are renewed by using projection values. (3) By calculating KMO (KaiserMeyerOlkin) and Bartletts test of sphericity, the correlation coefcient turns out to be 0.51, which satises the precondition of factor analysis. As to determining the number of factors, it is required that either the accumulation contribution rate exceeds 85% or its eigenvalue exceeds the mean value. Factor analysis is carried out through SPSS (Statistical Package for the Social Science). The 16 factors are extracted through factor rotation and among them the 10 factors with largest contribution rate are selected. Their accumulative contribution rate is 93.4%, and the analysis shows that there are very low correlation coefcients among them. The indicators processed by factor analysis are regarded as input indicators of the BP ANN model. The output indicator is PDS, among which the value of DBB is set as 1 and the value of DB is set as 2. The indicator values of the target project are shown in Table 5. BP ANN is constructed to train and predict the modied similar projects. With the tool of MATLAB, the predicted result of Project #003 is 1.3115 (the actual value is 1). 5.2. Reliability of model The predicted result of target Project #003 is very close to the actual value. To validate the predicting reliability of the developed model, a comparative analysis has been carried out. Firstly, select a group of control projects, whose codes shown in questionnaires are in order and starting from #001. Meanwhile, the amount should be the same as that of projects similar to Project #3, namely 31 projects with 14 DB projects and 17 DBB

4. Method Questionnaire was adopted to obtain the necessary data and was designed according to the established indicator system. Values of the indicators can be obtained in two ways. (1) For indicators whose variables in the formula can be obtained from respondents, values of them are calculated by formulas; (2) values of them can be either dened by scale 19 or directly estimated by respondents. All the indicator values obtained by questionnaires should be normalized into the scale 19. The respondents are required to provide the maximum and minimum of the indicator values. If they fail, then the uctuation boundary will be 1, i.e. the indicator values will be changing between (value 1, value + 1).

Y.Q. Chen et al. / Expert Systems with Applications 38 (2011) 54565462 Table 5 Indicator values of target project. Indicator Schedule delay Cost growth Project quality Project scale Project complexity Ability to dene the project scope Flexibilty Disputes Value 9 9 5 7 3 2 9 3 Indicator Owners willingness to involve Owners willingness to take risks Owners available personnel Contractors capability. Market competitiveness Regulatory feasibility Technology availability Value 5 1 1 4.5 9 1 3

5461

higher prediction reliability than ANN model with Chinese data and may be widely used in selection of PDSs. Project-specic data are inputted to improve the accuracy of prediction and reduce the deviation of experts subjective judgments. Project database is the basis of model prediction, so it is only by establishing the project database that the model can be more likely applied into practice, which is also the future work of this research. Besides, the model is compared with ANN model in this paper. The reliability of the model can be further proved through comparison with other PDS selection models. Acknowledgement The authors would like to acknowledge the support received from the National Natural Science Foundation of China funded project (under Project Number 70772057) 71072156 and all the respondents of questionnaire. References
Alhazmi, T., & McCaffer, R. (2000). Project procurement system selection model. Journal of Construction Engineering & Management, 126(3), 176. Al Khalil, M. I. (2002). Selecting the appropriate project delivery method using AHP. International Journal of Project Management, 20(6), 464469. ASCE (2000). Quality in the constructed project: A guide for owners, designers and constructors (2nd ed.). Reston, Virginia: American Society of Civil Engineers. Chan, C. T. W. (2007). Fuzzy procurement selection model for construction projects. Construction Management and Economics, 25(6), 611618. Chan, A. P. C., Yung, E. H. K., Lam, P. T. I., Tam, C. M., & Cheung, S. O. (2001). Application of Delphi method in selection of procurement systems for construction projects. Construction Management and Economics, 19(7), 699718. Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring the efciency of decision making units. European Journal of Operational Research, 2(6), 429444. Chen, H. H. (2008). Stock selection using data envelopment analysis. Industrial Management & Data Systems, 108(9), 12551268. Chen, Y. Q., Lu, W. X., & Zhang, S. B. (2005). Developing an integrated management system of engineering projects. China Civil Engineering Journal, 38(5), 111115. Cheung, S. O., Lam, T. I., Wan, Y. W., & Lam, K. C. (2001). Improving objectivity in procurement selection. Journal of Management in Engineering, 17(3), 132139. Flood, I. (2008). Towards the next generation of articial neural networks for civil engineering. Advanced Engineering Informatics, 414. Gransberg, D. D., Dillon, W. D., Reynolds, L., & Boyd, J. (1999). Quantitative analysis of partnered project performance. Journal of Construction Engineering and Management, 125(3), 161166. Hong, H. K., Kim, J. S., Kim, T., & Leem, B. H. (2008). The effect of knowledge on system integration project performance. Industrial Management & Data Systems, 108(3), 385404. Hsieh, C. H., & Chen, S. H. (1999). Similarity of generalized fuzzy numbers with graded mean integration representation. In Proceedings of the eighth international fuzzy systems association world congress, Taipei (Vol. 2, pp. 551 555). Konchar, M., & Sanvido, V. (1998). Comparison of US project delivery systems. Journal of Construction Engineering and Management, 124(6), 435444. Kumaraswamy, M. M., & Dissanayaka, S. M. (1996). Procurement by objectives. Journal of Construction Procurement, 2(2), 3851. Kumaraswamy, M. M., & Dissanayaka, S. M. (2001). Developing a decision support system for building project procurement. Building and Environment, 36(3), 337349. Ling, F. Y. Y., Chan, S. L., Chong, E., & Ee, L. P. (2004). Predicting performance of design-build and design-bid-build projects. Journal of Construction Engineering and Management, 130(1), 7583. Ling, F. Y. Y., & Liu, M. (2004). Using neural network to predict performance of design-build projects in Singapore. Building and Environment, 39(10), 12631274. Lo, S. C., & Chao, Y. (2007). Efciency assessment of road project delivery models. In Computation in modern science and engineering: Proceedings of the international conference on computational methods in science and engineering, volume 2, Part B (Vol. 963, pp. 10161019). Love, P. E. D., Skitmore, M., & Earl, G. (1998). Selecting a suitable procurement method for a building project. Construction Management and Economics, 16(2), 221. Luu, D. T., Ng, S. T., & Chen, S. E. (2003a). Parameters governing the selection of procurement system An empirical survey. Engineering, Construction and Architectural Management, 10(3), 209218. Luu, D. T., Ng, S. T., & Chen, S. E. (2003b). A case-based procurement advisory system for construction. Advances in Engineering Software, 34(7), 429438. Luu, D. T., Ng, S. T., & Chen, S. E. (2005). Formulating procurement selection criteria through case-based reasoning approach. Journal of Computing in Civil Engineering, 19(3), 269276.

Table 6 Comparison of predicted values of different models (Project 003#). Prediction number 1 2 3 4 5 6 7 8 9 10 Sample mean value Sample variance Value (similar projects) 1.3115 1.0287 1.2544 1.1669 0.9465 1.1056 1.1059 1.1333 1.1920 0.8823 1.1127 0.0175 Value (control projects) 0.8451 1.7309 1.3474 0.9681 1.6688 1.4830 1.4513 1.1032 1.4067 1.3200 1.3325 0.0819

Table 7 Predicted value comparison for different models (Project 018# and 035#). Questionnare coding #018 #035 Comparison indicators Sample Sample Sample Sample mean value variance value mean value variance value Value (similar projects) 1.2130 0.0300 2.0323 0.0163 Value (control projects) 1.3560 0.0336 2.2124 0.0185

projects. In factor analysis, the 10 rst factors with the highest contribution rate are selected as input indicators of BP ANN model. Because the initial weights of ANN model are determined randomly, the model developed above and the model developed by using control samples are run continuously for 10 times, and the result is shown as Table 6. The third column is the predicted result of BP ANN model which is trained by using the 31 control projects. It can be seen that the mean value of the model developed above is 1.11, and the sample variance is 0.018. Compared with the result of control projects, the developed model is closer to 1, which indicates that the model is reliable. Using the same analysis method, the Project #018 and Project #035 are randomly selected to further validate the reliability of the model. The predicted mean values and variance values of the two projects are shown in Table 7. Similar conclusion can be reached as Project #003. If a large amount of projects are obtained and the retrieval Iis considered, the reliability of the developed model is likely to be higher. 6. Conclusions Sixteen indicators from four categories, namely project objectives, project characteristics, characteristics of owner and contractor and external environment, are identied through literature review. The PDS selection model is developed with ANN and DEA methods. It empirically found that the developed model shows

5462

Y.Q. Chen et al. / Expert Systems with Applications 38 (2011) 54565462 Ojiako, U., Johansen, E., & Greenwood, D. (2008). A qualitative re-construction of project measurement criteria. Industrial Management & Data Systems, 108(3), 405417. Oyetunji, A. A., & Anderson, S. D. (2006). Relative effectiveness of project delivery and contract strategies. Journal of Construction Engineering & Management, 313. Ribeiro, F. L. (2001). Project delivery system selection: A case-based reasoning framework. Logistics Information Management, 14(5), 367375. Shen, L. Y., Lu, W. S., & Yam, M. C. H. (2006). Contractor key competitiveness indicators: A China study. Journal of Construction Engineering and Management, 132(4), 416424. Zaghloul, W., Lee, S. M., & Trimi, S. (2009). Text classication: Neural networks vs support vector machines. Industrial Management & Data Systems, 109(5), 708717.

Luu, D. T., Ng, S. T., Chen, S. E., & Jefferies, M. (2006). A strategy for evaluating a fuzzy case-based construction procurement selection system. Advances in Engineering Software, 37(3), 159171. Mafakheri, F., Dai, L., Slezak, D., & Nasiri, F. (2007). Project delivery system selection under uncertainty Multi-criteria multilevel decision aid model. Journal of Management in Engineering, 23(4), 200206. Mahdi, I. M., & Alreshaid, K. (2005). Decision support system for selecting the proper project delivery method using analytical hierarchy process (AHP). International Journal of Project Management, 23(7), 564572. Molenaar, K. R., & Songer, A. D. (1998). Model for public sector design-build project selection. Journal of Construction Engineering and Management, 124(6), 467479. Ng, S. T., Luu, D. T., Chen, S. E., & Lam, K. C. (2002). Fuzzy membership functions of procurement selection criteria. Construction Management and Economics, 20(3), 285296.