An ERP software selection process with using artificial ...

Expert Systems with Applications 36 (2009) 9214–9222

Contents lists available at ScienceDirect

Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

An ERP software selection process with using artificial neural network based on analytic network process approach Harun Resit Yazgan a, Semra Boran a, Kerim Goztepe b,* a b

Department of Industrial Engineering, Sakarya University, Sakarya, Turkey Institute of Science, Sakarya University, Sakarya, Turkey

a r t i c l e

i n f o

Keywords: ERP software selection Analytic network process (ANP) Artificial neural network (ANN)

a b s t r a c t An enterprise resource planning (ERP) software selection is known to be multi attribute decision making (MADM) problem. This problem has been modeled according with analytic network process (ANP) method due to fact that it considers criteria and sub criteria relations and interrelations in selecting the software. Opinions of many experts are obtained while building ANP model for the selection ERP then opinions are reduced to one single value by methods like geometric means so as to get desired results. To use ANP model for the selection of ERP for a new organization, a new group of expert’s opinions are needed. In this case the same problem will be in counter. In the proposed model, when ANP and ANN models are setup, an ERP software selection can be made easily by the opinions of one single expert. In that case calculation of geometric mean of answers that obtained from many experts will be unnecessary. Additionally the effect of subjective opinion of one single decision maker will be avoided. In terms of difficulty, ANP has some difficulties due to eigenvalue and their limit value calculation. An ANN model has been designed and trained with using ANP results in order to calculate ERP software priority. The artificial neural network (ANN) model is trained by results obtained from ANP. It seems that there is no any major difficulty in order to predict software priorities with trained ANN model. By this results ANN model has been come suitable for using in the selection of ERP for another new decision. Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction ERP is a integrated, consulate enterprise wide information system that combines all necessary business functions like production planning, purchase, inventory control, sales, finance, human resource. Organizations require ERP implementation for the purposes of customer-order integration, standardization of production process, reduction of inventory level and order preparation time, standardization human resources information. Today organizations operate in an economic environment where customer demands are continuously changing and increasing. In today, markets a great number of competitors are in places and competition is so fierce. Quality and cost do not suffice in competition and therefore new competition parameters are needed like delivery date in right time and customize product (Yusuf, Gunasekaran, & Wu, 2006). These organizations strive to reduce total cost through supply chain, production cycle, and inventory. Additionally, they request increasing diversity of product, more accurate delivery dates and * Corresponding author. E-mail addresses: [email protected] (H.R. Yazgan), [email protected] (S. Boran), [email protected] (K. Goztepe). 0957-4174/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.12.022

coordinating the supply and production effectively (Liao, Li, & Lu, 2007; Xiuwu et al., 2007). ERP software automates and integrates business processes and allows information sharing in different business functions. In addition that ERP software supports the finance, human resources, operations and logistic, sale and market in functions by through more effected and productive business process. At the same time it improves the performance of organization’s functions by controlling those (Hallikainen, Kimpimaki, & Kivijarvi, 2006). Although organizations can develop their own ERP software, other ones may prefer ready systems to shorten application cycle. The vendors sale ERP software that is developed on different operating system and database in the market. When the organizations prefer to buy ready systems, it is going to be very height cost (Verville & Halingten, 2003). Therefore ERP selection is an important decision making problem of organization and effects directly the performance. The ERP selection is tiresome and time consuming in terms of complexity of business environment, resource shortages. There are a lot of ERP alternatives in market (Wei & Wang, 2004). The best suitable ERP system selection yields positive results like increasing productivity, timely delivery, reduction of setup time, reduction of purchasing cost. The failure in selection of ERP system leads to the

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222

failure of project or company performance will get weakened (Liao et al., 2007). The fact that software programs are costly and their adaptation takes too much time so the cost of wrong selection is too high. In past years, the AHP method has been used often for selection of the ERP software, but recently years ANP method has been preferred more than AHP. These methods have been considered as MADM methods. MADM is a methodology that helps decision-makers makes preference decisions selection regarding a finite set of available alternatives courses of action characterized by multiple potentially conflicting attributes (Chang, Wu, Lin, & Lin, 2007). The ANP method used selection problems to do following reasons (Liang & Li, 2007):  With ANP, the criteria priorities may be determined based on pair comparison rates by decision maker’s judgment rather than arbitrary scales.  With ANP, decision-makers can be consider both tangible and intangible factors.  ANP can transform qualitative values into numerical values to make comparative analysis ANP is so simple and intuitive approach that decision-makers can easily understand and apply it without having professional or special knowledge.  ANP can motivate all stakeholders and decision-makers to join the decision making process. The evaluation criteria and alternatives as regards the problems in the ERP software selection presented in this article have been modeled with using ANP. The problem in the model consists of hierarchic order which has goal, evaluation criteria, sub criteria and alternatives. The structure provides holistic approach for multiple criteria decision making problem such as ERP software selection. ANP method can be considered a solution method in order to solve too sophisticated multi criteria decision making problems. While the ANP model for ERP selection is built into reality, many expert opinions are obtained and these opinions come down to one single value through methods like geometric means and thus the results are obtained. If new organization select ANP model to solve ERP selection problem, a new expert group should be established in order to obtain their opinions. Some organization may have some difficulties establishing experts groups. A single decision maker can make ERP selection decision so subjectivity and bayes problem may appear. In order to get rid of difficulties, in this study an ANN model is proposed to use. Every person in the project team complete pair wise comparisons in the ANP model then results of the ANP is going to be use training of ANN model. Consequently, if an organization has a single decision maker, the proposed ANN can be applied to predict the best suitable ERP software. An important characteristic of the proposed model removes requirement of the establishing group. From the point of difficulties in calculation, an ANP model has difficulties due to eigenvalue and their limit matrix value particularly when number of factors, sub factors and alternatives are large. When the model training is completed, there will be no major difficulty for prediction. This study consists of five chapters. The second chapter consists of the literature research. The third chapter introduces the methods (ANP and ANN) which are applied in ERP selection. Chapter four is related with application of the proposed model and the results which have been obtained in the last chapter.

9215

is one of the mathematical programming models, is used together with ANP for information system (IS) project selection and an application is carried out by Lee and Kim (2000). In another study, however Badri and Davis (2001) adopted 0–1 goal programming for IS selection model which they have themselves developed. Karsak and Ozogul (2007) used QFD, fuzzy linear regression and 0–1 goal programming by integrating them in the ERP selection problem. Bernroider and Stix (2006) defined ERP software selection to be multi-purpose decision making system. In their model, they adopted utility ranking and DEA methods. Additionally, they have developed multi attribute utility model and alternative profiles calculated by DEA optimization. In ERP Selection, AHP/ANP methods have been used. Wei, Chien, and Wang (2005) studied on AHP based ERP selection.Ravi, Shankar, and Tiwari (2005) developed ANP model for ERP software selection problem. In recent studies, one can observe that fuzzy sets have been used together with AHP/ANP methods for ERP Software selection. In ERP selection, a study was made by Ayag and Ozdemir (2007) where a fuzzy ANP was adopted. Literature contains studies in which AHP/ANP and artificial neural Networks have been used together. For example, Stam, Minghe, and Haines (1996) made use of ANN in calculation of dual comparison matrix values in AHP. First, Hopfield network was introduced and this trained network was used in the calculation of comparison data whose absolute values were unknown. The trained network by the help of simulation techniques, proved to be affective in the solution of multi-purpose decision problems depended on vague or fuzzy data especially when data was uncertain and fuzzy. In their relevant study, Hu and Tsai (2006) studied the case where the data for the comparison matrix was partly missing and they have proposed to find the missing data by way of back propagation method. The multi-layer back propagation they have proposed estimates the missing data and enables the usage of AHP. Kuo, Chi, and Kao (2002) developed fuzzy AHP structure for the selection of the most convenient store place problem. They also studied the interrelations between factors and store performance by ANN model. Hu and Tsai (2006) developed multi-layer ANN model to calculate dual comparison which have been used in AHP approach. ANN model, which they proposed, estimates the calculation of missing dual comparison data from the present dual comparison and thus completes the missing dual matrix data. Chao and Skibniewski (1995) determined the performance characteristics of each of sample technology by AHP for the problem of the acceptance of new technology in construction industry and estimated the acceptance of new technology within certain performance criteria with ANN model. Matsuda (2006) dealt with the problem in case of missing or no information when the decision is made with ANP based ANN model and tested the validity of his model with simulation. Mikhailov (2004) studied on the determination of group priority in AHP method. He mentioned that the usage of his fuzzy based optimization method in reducing the group decisions to a single one and priorities to a single value is more appropriate than the usage of geometric mean method of Aczel and Saaty (1983). In his study, a fuzzy approach has been implemented to solve the problem in the case of missing information of the group members.

3. ANP and ANN methodology 3.1. ANP method

2. Literature survey An ERP software selection is considered to be one of the multicriteria decision making problems. 0–1 goal programming, which

Saaty (1980) firstly developed AHP method to solve the problems of MCDM. But, in AHP method, interdependence and feedback that may be seen in evaluation criteria has not been defined. To make up for these deficiencies, an ANP method was developed

9216

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222

by Satty (2001). Strictly relation should be established among criteria in AHP. However in ANP, criteria in the lower level may provide feedback to the criteria in the higher level, and the interdependence among the criteria in the same level is permitted (Liang & Li, 2007; Ravi et al., 2005). Another difference between AHP and ANP in calculation process is that a new concept ‘‘supermatrix” is introduced in ANP (Liang & Li, 2007). Fundamental of the ANP could be found in Saaty (1999). Advantages of the ANP are follows (Ravi et al., 2005; Saaty, 1999):  ANP is a comprehensive technique that allows for the inclusion of all the relevant criteria: tangible as well as intangible, which have some bearing on decision making process (Saaty, 1999).  AHP model is based on decision making framework that assumes unidirectional hierarchical relationship among decision levels, whereas ANP allows for more complex relationship among the decision levels and attributes as it does not require a strict hierarchical structure.  In decision making problems, it is very important to consider the interdependent relationship among criteria because of the characteristics of interdependence that exist in real life problems. The ANP methodology allows for the consideration of interdependencies among levels of criteria. Thus, it is a more desirable multi-criteria decision making tool. This feature makes it superior from AHP which fails to capture interdependencies among different enablers, criteria and sub-criteria (Agarwal & Shankar, 2003).  ANP is suitable in considering both qualitative as well as quantitative characteristics which need to be considered ( Boran, Goztepe, & Yavuz, 2008), as well as taking non-linear interdependent relationship among the attributes into consideration (Meade & Sarkis, 1999). ANP is unique in the sense that it provides synthetic scores, which is an indicator of the relative ranking of different alternatives available to the decision maker.  Because of the increased number of pairwise comparison, the calculation process for ANP is more time consuming. Nevertheless, ANP is more closely to real situation which has considered the feedback and interdependence among criteria. Therefore it gives more flexibility for constructing the decision model, ANP is considered to be a promising method.  Generally speaking, the application steps of ANP are as follows. Building a model: Starts with determining evaluation criteria, sub criteria and alternatives. Relations are shown between them in a network structure. Network is formed based on relationship among clusters and within elements in each cluster. Clusters consist of main criteria and their sub criteria. In addition that alternative can also be considered as a cluster. There are different relationships in model such as in clusters, within cluster among themselves and a cluster with other clusters. Unidirect relationship which means there is relationship from one cluster to another one. Indirect relationship represents that there is no direct relation between two clusters but one cluster can be affected by through third cluster. The another relationship is a self-interaction that there is a relationship among sub criteria in same cluster. Last one is called as an interrelationship among criteria. To bring out the value judgment through wise comparison: Decision elements are compared according to their priorities with pairwise comparison and the values are assigned by using 9 point scale defined by Satty (2001). The values of pairwise comparisons are allocated in comparison matrix and local priority vector is obtained from eigenvector which is calculated from this equation (Promentilla et al., 2006):

Aw ¼ kenb w

ð1Þ

In this equation, A,w and kenb stands for the pairwise comparison matrix, eigenvector and eigenvalue, respectively. Saaty (1980) has proposed normalization algorithm for approximate solution for w. The matrix which shows the pairwise comparison between factors is obtained as follows:

A ¼ ½aij nxn

i ¼ 1; . . . ; n j ¼ 1; . . . ; n

ð2Þ

Significance distribution of factors as percentage is obtained as follows:

Bi ¼ ½bij nx1

i ¼ 1; . . . ; n

aij

bij ¼ Pn

ð4Þ

i¼1 aij

C ¼ ½bij nxn Pn wi ¼

j¼1 c ij

n

ð3Þ

i ¼ 1; . . . ; n j ¼ 1; . . . ; n W ¼ ½wi nx1

ð5Þ ð6Þ

Consistency of pairwise matrix is checked by consistency index (CI). The consistency of elements comparisons are calculated as follows:

D ¼ baij cnxn x½wi nx1 ¼ ½di nx1 di Ei ¼ i ¼ 1; . . . ; n wi Pn Ei k ¼ i¼1 n CI CR ¼ RI

ð7Þ ð8Þ ð9Þ ð10Þ

In the equations above, CI, RI and CR represent consistency indicator, random indicator and consistency ratio, respectively. For accepted consistency, CI must be smaller than 0.10. (Saaty, 1980). Forming supermatrix: A supermatrix, known as partition matrix, is formed by setting the local priority vectors on suitable columns. Local priority vectors are classified and occupied in suitable places based on effect flow from one component to another. Supermatrix may consist of zero value. In general, there exists interdependence between clusters, the sum of one column in the supermatrix is mostly bigger than 1. In case, the supermatrix is not stochastic, the cluster is weighted and column is normalized to transform into a stochastic matrix where the sum of columns are 1. This matrix can be called as a weighted supermatrix. Forming limit supermatrix: In case k displays a great random number, the supermatrix is increased to power 2k + 1 and thus it approximates to limit namely importance weight. The new matrix which is called limit supermatrix, displays the effects of elements on each other in the long run. The limit supermatrix represents the same structure as the weighted supermatrix. All columns of limit supermatrix are alike. Selection of the best alternative: The final priorities of the all elements in the matrix may be determined by normalizing each column in the supermatrix. Therefore, the priorities of alternatives may be seen in the column of alternatives in the normalized supermatrix. 3.2. Artificial neural network (ANN) ANN methodology has been used in many applications and research areas. One of the important benefits of using ANN is to be ability of generalizing variables which are gained from a real world problem. An ANN is defined by Rumelhart and McClelland (1986) as ‘‘massively parallel interconnected network of simple (usually adaptive) elements and their hierarchical organizations which are intended to interact with objects of the real world in same way as biological nervous systems do”. An ANN can be constructed on the basis of a number of simple elements. The elements are organized into a sequence of layers,

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222

each linked by weights, which are adapted in supervise learning. A neural network’s structure can be characterized by the connection pattern among elements, the transfer function for transforming input to output in elements, and the learning strategy. Among the several well-known supervised learning neural models are back propagation, learning vector quantization, and counter propagation network. Of which, the back propagation model is most extensively used and therefore, selected herein. A back propagation (BP) neural network consists of three or more layers, including an input layer, one or more hidden layers, and an output layer. Training data set is initially collected to develop a back propagation neural network model. Through a supervised learning rule, the data set comprises an input and an actual output (target). Two phases are available for computation: forward and backward. First, the propagation network receives the input data pattern and directly passes it onto the hidden layer. Each element of the hidden layer calculates an activation value by summing up the weighted inputs and then, transforming the weighted input into an activity level by using a transfer function. Each element of the output layer is used to calculate an activation value by summing up the weighted inputs attributed to the hidden layer. Next, the actual network output is compared with the target value. If a difference arises, i.e. an error term, the gradient-descent algorithm is used to adjust the connected weights. If no difference arises, no learning is preceded with (Su & Hsieh, 1998).

4. Proposed model In this study, proposed model consists of integration of ANP and ANN. At the first, the ERP selection problem is modeled with using an ANP. Every factor’s weighted values and ERP software’s priority values are determined. Then, an ANN model is designed and values in previous ANP model are going to be used in training stage. The trained ANN consists of project team judgments and can be used for prediction of the best ERP software for a new organization. Detailed framework can be found in the following.

9217

4.1. Project team establishment In this study, a project team is composed of 35 persons such as managers who have ERP implementation responsibility, experts from consulting companies and universities. Defining criteria, sub criteria and alternatives: There are number of evaluation criteria in order to select ERP software among many which exist in the software market. In the literature, different evaluation criteria, sub criteria and alternative ERP softwares have been examined by Karsak and Ozogul (2007), Liang and Li (2007), Wei and Wang (2004), Wei et al. (2005). In this study, the project team has decided five main evaluation criteria and seventeen sub criteria based on their experiences, studies in the literature and reports about software. The main criteria are financial analyzes, general characterizes, system control and software design, production planning, data and knowledge properties. Financial analyzes (FA): An ERP software must meet the financial analysis needs of a customer. It has five sub criteria: debt and assets (DA), invoice and receipt (IR), cost analyzing (CA), customer procedure (CP) and taxes (TX). General characteristics (GC): GC consists of most common features of ERP software. These are varieties of the program (VP), production policy (PP) and manufacturing structure (MS). System control and software design (SCSD): The criteria represent all system security and values security. Sub criteria are records security (RS), quick and effective report capabilities (RC) and all the system security (SS). Production planning (PPL): It is related with obtaining raw material to finalizing production. It consists of capacity investment (CI), raw material procurement (RM), material resource planning (MRP) as sub factors. Data and Knowledge properties (DK): In this criteria variety of data and data size can be consider an important distinction of ERP software. Machine and equipment data (ME), customer information (CIN) and purchase and planning information (PPI) are sub criteria.

Fig. 1. ANP model in Superdecision software.

9218

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222

Fig. 2. Hierarchical model structure.

Alternatives: Although there are many ERP software which are developed by the different companies in the market, among only four of them have been chosen with PARETO analyzing which covers 85% of market which was reported the latest surveys. Establishing ANP model: In summary, five criteria and one alternatives represent six cluster and shown at Fig. 1. The network has been built with establishing relation within and among clusters. For example, FA cluster has three type of relationship. First one is interrelationship between FA and DK and it is shown as a two sided arrow. Second one is relationship with unidirect which exists from FA to GC. The last one is relationship itself which can be seen at cluster FA and shown as an arc symbol. The overall of objective of the ANP is to select best suitable ERP software which meets a company’s expectations. The Network consists of four hierarchical levels and the overall objective is located at the top of the hierarchy. Five critical factors are located at the second level. Next level, seventeen sub factors are settled according to their relevant factors. At the bottom of the hierarchy, four alternatives ERP software are located. Relationship among subcriteria and hierarchical relation between subcriteria and alternatives are shown in detail manner at Fig. 2. Pairwise comparison matrix: A special questionnaire form for ANP is used to complete pairwise comparison matrix. Thirty-five project team members have completed the questionnaire form. An example of pair wise comparison is given as follows (Table 1). In this step factors, sub factors and alternatives are compared with each other and among themselves with using pair wise comparisons questionnaire sheet which are required in establish-

Table 1 Pairwise comparison. Financial anlaysis

98765432123456789

Production planning

Table 2 Scale of relative importance (Adapted from Saaty (1980)). Intensity of importance

Definition

1 3 5 7 9 2,4,6,8

Equal importance Moderate importance Essential or strong importance Very strong importance Extreme importance Intermediate value between adjacent scale values

An example of weighted matrix for clusters is given at Table 3.

Table 3 Comparison matrix of clusters.

Alternatives DK FA GC PPL SCSD

Alternatives

DK

FA

GC

PPL

SCSD

0.25 0.25 0.25 0.25 0 0

0.25 0 0.25 0 0.25 0.25

0.1652 0.0768 0.3714 0.2012 0.0620 0.1231

0.25 0 0 0.25 0.25 0.25

0.20 0 0.20 0.20 0.20 0.20

0.75 0 0 0 0 0.25

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222

ing the ANP. The used scale is proposed by Satty and shown at Table 2. Super matrix and limit super matrix: Supermatrix, which is called unweighted supermatrix, can be obtained with placement eigenvalues which are generated from pair wise comparison within portioned matrix (Appendix A). Then weighted supermatrix is

9219

obtained form normalization of unweighted supermatrix and is given at Appendix A. Limit super matrix is also generated from weighted super matrix and priority values are found with 2k + 1 power of weighting matrix. This process is repeated for all the team members. Therefore, 35 limit super matrixes are generated by following previous steps. An example of a limit super matrix

Fig. 3. ANN Model Structure.

9220

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222

which is generated with using a person’s information is shown at the Appendix B. Establishing ANP mode and calculation of the all matrix values are carried out. 4.2. ANN model At the second stage of the proposed model, an ANN model is designed. The ANN consists of an input layer, a single intermediate layer and an output layer. Priority values of 17 sub factors which are generated from limit supermatrix of ANP are used for input values of the ANN model. Then four alternative ERP software’s priority values which are calculated from the ANP, are chosen as an output of the ANN model. Backpropagation algorithm is selected in order to calculate of weight values of the network and it is known that there is no certain rule to decide number of node in intermediate layer. Therefore, the model has been tested on different number of node in the intermediate layer; the best learning capability and minimum error are found when the number of node is chosen as twice of input values. So, the designed and trained ANN is going to be used for prediction superior ERP software. The ANN model is given at Fig. 3. Designing ANN model, all the required calculations and trained process are performed by using the Matlab software. 4.2.1. The network training As it is mentioned above, the project team consists of 35 persons. 30 of them are chosen for training process. 17 sub criteria’s priority values as an input and priority values of alternatives as an output are used for ANN training. Error rate is decided as a stop-

ping rule. The best error rate is found when learning rate is 0.07 and momentum coefficient is 0.5. A Sigmoid function is chosen as an activation function. During network training phase four different model was constructed in order to reach best ANN structure. For the best one 73 iteration has been carried out and there is no memorization fault. Performance of training processes are shown at the Fig. 4. There is no exact rule deciding on hidden nodes in an ANN model. Four methods of selecting hidden nodes process was executed while constructing ANN models (Kuo et al., 2002). 1. Model 1 (M1): Eight hidden nodes = (the number of input nodes + the number of output nodes)1/2; 2. Model 2 (M2): Eleven hidden nodes = 1/2 (the number of input nodes + the number of output nodes); 3. Model 3 (M3): Sixteen hidden nodes = 1/2 (the number of input nodes + the number of output nodes) + (the number of samples)1/2; 4. Model 4 (M4): Thirty four hidden nodes = 2 (the number of input nodes). Detailed information of ANN models structure can be seen in Table 4. Best of ANN model’s structure is shown at the Fig. 5. 4.2.2. Testing the ANN model Remaining 5 person’s priority values are used to test the ANN model. These data are not used during training of the network. Therefore, the memorizing of the network is going to be

Fig. 4. Performance graphics of tested ANN models.

9221

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222 Table 4 Data information used in ANN models. Model

M1 M2 M3 M4

Network structure Input nodes

Hidden nodes

Output nodes

17 17 17 17

8 11 16 34

4 4 4 4

Learning rate

Training function

Momentum coefficient

Transform function

Iteration number

Performance

0.07 0.07 0.07 0.07

TRAINGDM TRAINGDM TRAINGDM TRAINGDM

0.5 0.5 0.5 0.5

Sigmoid Sigmoid Sigmoid Sigmoid

71 109 138 73

4.29E22 1.99E22 3.43E22 4.25E23

Fig. 5. Matlab overview of best ANN model.

eliminated. The testing results show that the trained network is capable of predicting priority values in most accuracy. 4.2.3. The ANN prediction The trained ANN can be used for synthesizing judgments of expert people from different organizations. We believe that proposed ANP structure for ERP software selection in this research, which covers general characteristic of the most of organizations. A new organization which is in choosing stage of ERP software can use the proposed ANN unless the proposed ANP structure is not changed. A new organization may decide to establish a project team or enroll a single decision maker for selecting ERP. If there is a project team, the proposed ANN model can easily predict best ERP software with using every team member’s ANP results which are eigenvalues of the sub factors. Although a new project team may generally consist of a few people, trained ANN predicts values which are based on many people experience and knowledge. It may be criticized that a single decision maker may produce a subjective decision and therefore the decision should not be used in decision making process, however in this study the single decision maker’s judgments are passed through from the trained ANN which consists of many expert people’s knowledge and experience. Therefore the ANN predicted decision is going to be a far from subjective decision, but it is most likely going to be reflecting the project team’s experience and knowledge. It seems that the proposed approach which consists of ANP and ANN is capable of to predict selection problems with single or group decision maker. 5. Results Group decision or single decision process has always some difficulties in decision making problems. For a group decision, there can be found some solutions to reduce a from a group decision to a single one in the literature. Satty (2001) has discussed some techniques suchlike geometric mean, arithmetic mean, linear homogeneity, agreement, pareto optimality, condorced paradoks. Although the techniques can be applied in many applications, a new group (i.e. project team) must be established in every practical application. However in this study, when a network is trained with

using group decision it accommodates also group decisions. If there is a single decision maker, in a new decision problem (i.e. ERP software selection), a single person’ subjective decisions is taken as input of the ANN network, but the network predicts a result which is extracted from objectivity (i.e. synthesis of expert people decision). Other important differences of the proposed model in this study are an establishing a new group is not required after completing for once in practical application, saving time and cost and removing difficulties of reducing from a group decision to a single decision. References Aczel, J., & Saaty, T. L. (1983). Procedures for synthesizing ratio judgments. Journal of Mathematical Psychology, 27, 93–102. Agarwal, A., & Shankar, R. (2003). On-line trust building in e-enabled supply chain. Supply Chain Management: An International Journal, 8(4), 324–334. Ayag, Z., & Ozdemir, R. G. (2007). An intelligent approach to ERP software selection through fuzzy ANP. International Journal of Production Research, 45(10), 2169–2194. Badri, M. A., & Davis, D. (2001). A comprehensive 0–1 goal programming model for project selection. International Journal of Project Management, 19(4), 243–252. Bernroider, W. N., & Stix, V. (2006). Profile distance method-A multi-attribute decision making approach for information system investments. Decision Support Systems, 42(2), 988–998. Boran, S., Goztepe, K., & Yavuz, E. (2008). A study on election of personnel based on performance measurement by using analytic network process (ANP). IJCSNS International Journal of Computer Science and Network Security, 8(4). Chang, C.-W., Wu, C.-R., Lin, C.-T., & Lin, H.-L. (2007). Evaluating digital video recorder using analytic hierarchy and analytic network processes. Information Sciences, 177(16), 3383–3396. Chao, L.-C, & Skibniewski, M. J. (1995). Neural network method of estimating construction technology acceptability. Journal of Construction Engineering and Management, 121(1), 130–142. Hallikainen,P., Kimpimaki, H., & Kivijarvi, H. (2006). Supporting the module sequencing decision in the ERP implementation process. In Proceedings of the 39th Hawaii international conference on system sciences, 2006. Hu, Y.-C., & Tsai, J.-F. (2006). Backpropagation multilayer perceptron for incomplete pairwise comparison matrices in analytic hierarchy process. Applied Mathematics and Computers, 180, 53–62. Karsak, E., & Ozogul, C. O. (2007). An integrated decision making for ERP system selection. Expert Systems with Applications. Kuo, R. J., Chi, S. C., & Kao, S. S. (2002). A decision support system for selecting convenience store location through integration of fuzzy AHP and artificial neural network. Computers in Industry, 199–214. Liang, C., & Li, Q. (2007). Enterprise information system project selection with regard to BOCR. International Journal of Project Management.

9222

H.R. Yazgan et al. / Expert Systems with Applications 36 (2009) 9214–9222

Liao, X., Li, Y., & Lu, B. (2007). A model for selecting an ERP system based on linguistic information processing. Information Systems, 32, 1005–1017. Lee, J. W., & Kim, S. H. (2000). Using analytic network process and goal programming for interdependent information system project selection. Computers and Operations Research(27), 367–382. Matsuda, S. (2006). A neural network model for the decision-making process based on ANP. International Joint Conference on Neural Networks, Canada, July 16–21, 2006. Meade, L. M., & Sarkis, J. (1999). Analyzing organizational project alternatives for agile manufacturing processes: An analytical network approach. International Journal of Production Research, 37(2), 241–261. Mikhailov, L. (2004). A fuzzy approach to deriving priorities from interval pairwise comparison judgement. European Journal of Operational Research, 159(3), 687–704. Promentilla, M. A. B., Furuichi, T., Ishii, K., & Tanikawa, N. (2006). Evaluation of remedial countermeasures using the analytic network process. Waste Management, 26, 1410–1421. Ravi, V., Shankar, R., & Tiwari, M. K. (2005). Analyzing alternatives in reserve logistics for end-of-life computers: ANP and balanced scorecard approach. Computers and Industrial Engineering, 48, 327–356. Rumelhart, D. E., & McClelland, J. L. (1986). Parallel distributed processing: Explorations in the microstructure cognition: Foundations (Vol. 1), MIT Press.

Satty, T. L (2001). Decision making with dependence and feedback: The analytic network process (2nd ed.). 376 p. Saaty, T. L (1980). The analytic hierarchy process. New York: McGraw-Hill. Saaty, T. L (1999). Fundamentals of the analytic network process. Kobe, Japan: ISAHP. Stam, A., Minghe, S., & Haines, M. (1996). Artificial neural network representation for hierarchical preference structures. Computers and Operations Research 23(12), 191–1201; (11) 1996. Su, C.-T., & Hsieh, K.-L. (1998). Applying neural network approach to achieve robust design for dynamic quality characteristics. International Journal of Quality and Reliability Management, 15(5), 509–519. Verville, J., & Halingten, A. (2003). A six stage model of the buying process for ERP software. Industrial Marketing Management, 32, 585–594. Wei, C.-C., & Wang, M.-J. J. (2004). A comprehensive framework for selecting an ERP system. International Journal of Project Management, 22, 161–169. Wei, C.-C., Chien, C.-F., & Wang, M.-J. J. (2005). An AHP based approach to ERP system selection. International Journal of Production Economics, 96, 47–62. Xiuwu, L., Yuan, Li., & Bing, Lu. (2007). A model for selecting an ERP system based on linguistic information processing. Information System, 32(7), 1005–1017. Yusuf, Y., Gunasekaran, A., & Wu, C. (2006). Implementation of enterprise resource planning in China. Technovation.