Sunteți pe pagina 1din 16

The current issue and full text archive of this journal is available on Emerald Insight at:

www.emeraldinsight.com/1463-5771.htm

Project
A hybrid grey based artificial portfolio
neural network and C&R tree selection

for project portfolio selection


Farshad Faezy Razi 651
Department of Industrial Management, Semnan Branch,
Received 13 June 2016
Islamic Azad University, Semnan, Iran, and Revised 24 August 2016
Seyed Hooman Shariat Accepted 4 October 2016

Department of Industrial Management, Karaj Branch,


Islamic Azad University, Karaj, Iran

Abstract
Purpose – The purpose of this paper is twofold: the selection of project portfolios through hybrid artificial
neural network algorithms, feature selection based on grey relational analysis, decision tree and regression;
and the identification of the features affecting project portfolio selection using the artificial neural network
algorithm, decision tree and regression. The authors also aim to classify the available options using the
decision tree algorithm.
Design/methodology/approach – In order to achieve the research goals, a project-oriented organization
was selected and studied. In all, 49 project management indicators were chosen from A Guide to the Project
Management Body of Knowledge (PMBOK Guide), and the most important indicators were identified using a
feature selection algorithm and decision tree. After the extraction of rules, decision rule-based multi-criteria
decision making matrices were produced. Each matrix was ranked through grey relational analysis,
similarity to ideal solution method and multi-criteria optimization. Finally, a model for choosing the best
ranking method was designed and implemented using the genetic algorithm. To analyze the responses,
stability of the classes was investigated.
Findings – The results showed that projects ranked based on neural network weights by the grey relational
analysis method prove to be better options for the selection of a project portfolio. The process of identification
of the features affecting project portfolio selection resulted in the following factors: scope management,
project charter, project management plan, stakeholders and risk.
Originality/value – This study presents the most effective features affecting project portfolio selection
which is highly impressive in organizational decision making and must be considered seriously. Deploying
sensitivity analysis, which is an innovation in such studies, played a constructive role in examining the
accuracy and reliability of the proposed models, and it can be firmly argued that the results have had an
important role in validating the findings of this study.
Keywords Grey relational analysis, Genetic algorithm, Feature selection, Artificial neural network,
Project portfolio selection
Paper type Research paper

1. Introduction
A project is a temporary effort to create a product, service or a unique result (Project
Management Institute, 2013). Project selection is an important topic in the discussions about
industrial management, industrial engineering, governmental, non-profit and business
organizations (Hall and Nauda, 1990). Selection of the best project to the satisfaction of the
whole organization has always been a popular topic in previous studies (Wang et al., 2009).
Project selection is a complex process in terms of decision making, because different factors
such as market position, access to raw materials, advantages and technical and
technological capabilities, and finally governmental laws and regulations have to be
Benchmarking: An International
included in the process (Bard, 1988). The complexity of the portfolio selection process is Journal
increased by the collective nature of decision making and the inclusion of high-level decision Vol. 24 No. 3, 2017
pp. 651-665
makers and project management personnel in, according to their professional, cultural and © Emerald Publishing Limited
1463-5771
social backgrounds ( Jiang and Klein, 1999). The process of project selection can be defined DOI 10.1108/BIJ-06-2016-0087
BIJ as such: this process begins with continuously collecting, analyzing and judging the
24,3 available information on the project, and ultimately ends by selecting a project based on the
important criteria. Regarding the multiple decision-making nature of the portfolio selection
process, Gabriel argues that first, many decision-making matrices can assess potential
projects. For example, minimizing the costs can be an assessment criterion. However, with
such an approach, other aspects will be neglected and the project will not have a chance to
652 succeed. Therefore, project selection is essentially a multi-criteria decision-making process.
Second, for project selection it is necessary to pay attention to issues such as costs, human
resources and raw material suppliers. It has to be noted that estimation of the
above-mentioned criteria in the planning stage is very difficult. Therefore, project managers
are forced to make a decision without thorough knowledge (Gabriel et al., 2006). Such
contingent aspects of project selection encourage the use of multi-criteria decision-making
methods and contingency planning approaches. In the multi-criteria decision-making model,
the model’s answer may be known in advance but the goal is to choose the best answer in a set
of available options. These types of decision-making problems are referred to as
multi-criteria decision-making models. Multi-criteria decision-making models are based on
the utility theory and human pressures in dealing with the maximum seeking behavior
(Rabbani et al., 2010). A lot of research has been carried out on this subject before, but what
makes this study different from other similar studies is the assessment of validity and
reliability of the implemented model. To assess the validity of the utilized models, we used the
consistency analysis method. Another point of difference from other similar studies is that the
analysis of the weight of the indicators has been performed using an artificial neural network.

2. Literature review
It is a very common action among researchers to use rough set theory and also C5.0, CHAID,
QUEST and C&R Tree (CART) in their researches in order to classify data and removal or
reduction of irrelevant information which are surplus to requirements of a database in
project portfolio selection issue.
There are various research studies in which such techniques have been used, but the
combination of different techniques and methods which have been used in the papers of
researchers around the world about portfolio selection is missing the combination of
CART and GRA.
By investigating the available literature, the authors of this paper were fascinated to the
subject of the lack of using the CART algorithm along with the GRA algorithm in case of
project portfolio selection.
The following studies are few samples of using the mentioned techniques in highly cited
articles: Zhang et al. in their 2007 research “A new multi-criteria quadratic-programming
linear classification model for VIP e-mail analysis”, in order to compare the efficiency of
multi-criteria quadratic-programming linear classification and C5.0, found that the proposed
MQLC model performs better than decision tree on small samples (Zhang et al., 2007, May).
Haughton and Oulabi in their research “Direct marketing modeling with CART and CHAID”
used CART and CHAID to check the response. They found that response lifts are very close
for both types of models (Haughton and Oulabi, 1997). Even though Haughton and
Oulabi compared the outcome, they did not investigate the robustness and optimization of
each technique. One of the researches which had the most significant effect on selecting the
CART algorithm in this study was the outcome and result of Hamadani and
Khademolqorani’s research in 2015 titled “Development of a decision support system for
handling health insurance deduction”. As a result of Khademolqorani and Hamadanis’
(2015) research, the CART algorithm has the highest overall accuracy among C5.0, QUEST,
CHAID and CART. The subject of portfolio selection reveals its importance in several
different aspects. The importance of portfolio selection is highlighted primarily in modeling,
orientation and the extent of application. Several researchers have tried to optimally choose Project
portfolios in their investigated organizations. Faezi Razi et al. have proposed a hybrid portfolio
multi-purpose fuzzy model using grey relational analysis, C-means and genetic algorithm. In selection
this study, the researchers have designed a decision support system to help managers in
portfolio selection. Since the fuzzy C-means method is unable to thoroughly rank the
projects, a combination of the fuzzy C-means method and grey relational analysis based on
the intuitive trapezoidal fuzzy number has been utilized to solve the project portfolio 653
selection problem (Razi et al., 2015). In another study, a hybrid model for project portfolio
selection has been proposed. In this study, projects are ranked by the grey relational
analysis and then the project risks are identified and analyzed by the fuzzy inference
system. Finally, rank and risk objectives have been optimized using a two-objective, zero-
one mathematical planning model, taking into account the constraints of the decision-
making problem (Razi, 2014). In another study, a hybrid model for R&D project portfolio
selection has been proposed that utilizes the K-means approach and genetic algorithm. This
model has the advantage of using the K-means algorithm for project selection and choosing
the most appropriate choice under conditions of uncertainty using the grey system
(Eshlaghy and Razi, 2015). In another study, a non-rating-based model for the selection of
project portfolios has been presented (Faezy, 2015). In one study, project selection is
examined in the framework of information and communication technologies. This study
uses an analytical hierarchy process (AHP) for weighting (Kim and Chang, 2013). In another
study, a hybrid decision model has been designed to select the project information.
The researchers have proposed a hybrid measurement model to assess the impacts of
project information. This model considers quantitative, qualitative and critical indicators
and helps managers to identify the advantages and disadvantages of operating conditions.
The model not only accelerates the evaluation and project selection process, but also makes
it possible for organizations to optimize project selection, thus leading to sustainable
development in the activities and operations of the organization (Yang et al., 2014). Project
selection using Yager’s fuzzy decision-making approach and solving the decision-making
problem regarding the revival of mining operations using AHP have been the subjects of
another study (Yavuz and Altay, 2015). Compliance of project management methods and
strategies with neuro-fuzzy inference systems for selecting variables has also been studied.
In this study, the existing parameters have been subject to analysis in order to increase the
efficiency of the energy produced by wind turbines in an area. The study also utilizes the
regression function approximation method for modeling complex systems based on neuro-
fuzzy inference systems (Petković et al., 2014). In another study, a multi-stage multi-attribute
model for project portfolio management is provided. The researchers presented a goal-
oriented programming model based on the current understanding of future values of the
existing project portfolios and provided a model for optimal portfolio selection according to
the investors’ preferences (Arasteh et al., 2014a, b). The application of grey system and fuzzy
set theory combined with real options theory in project portfolio management has also been
studied. In this research, the necessary approaches for portfolio selection and evaluation of
investment in the portfolio have been provided using linear planning and grey relational
analysis (Arasteh et al., 2014a, b). The hybrid model based on the least-squares support
vector machines method has also received attention. Among the traditional weighting
models, artificial intelligence has been compared to traditional methods for better
performance. This study introduces an effective intelligence-system-based model based on
neural networks which enhances decision making for managers (Vahdani et al., 2014).
Another study investigated control and integrated scheduling for multi-mode projects. This
study attempts to develop statistical models and algorithms that combine the Tabu Search
strategy and linear programming with scheduling and project control. Table I summarizes a
number of studies about the project portfolio selection process.
BIJ No. Decision model Decision problem Author/researcher
24,3
1 Multi-criteria quadratic-programming VIP e-mail analysis Zhang, Shi
linear classification
2 CART, CHAID Direct marketing modeling Haughton, Oulabi
3 CART, CHAID, C5.0, QUEST Handling health insurance Khademolqorani,
deduction Hamadan
654 4 GRA & C-means & GA Project portfolio selection Faezy Razi
5 GRA Project portfolio Selection Faezy Razi
6 K-means R&D project selection Eshlaghy
7 Fuzzy GRA Project selection Faezy Razi
8 APH Project framework selection Kim
9 Hybrid decision model Project information selection Yang
10 Yager fuzzy decision-making Revival of mining operations Yavuz
11 ANFIS & PM Variable selection Petković
12 Multi-criteria multi-mode model Project portfolio management Arasteh
Table I. 13 Fuzzy set & real option theory Project portfolio management Arasteh
Studies on the subject 14 Least squares support vector machine Enhancing manager’s decision Vahdani
of project selection hybrid model making

3. Tools
The main purpose of this research is to manage the industries which meet the challenge of
project portfolio selection. Regarding this issue it is mandatory to identify and utilize the tools
which can optimize ranking and weight the criteria. There are several techniques for ranking
and also for weighting criteria; in this study the authors have selected three techniques of
ranking including grey relational analysis, VIKOR and TOPSIS to study. For weighting they
have selected artificial neural network methods and also entropy Shannon. The reason behind
using these techniques in this paper is to identify the most robust hybrid system in the case of
project portfolio selection. Because of frequent use of researchers in their papers from three
ranking GRA, TOPSIS and VIKOR techniques, the authors decided to select these three
popular techniques for studying and investigating. It is also needed to mention that weighting
methods which are used in this study are from the most popular ones among researchers.
In this section, the primary tools used for project portfolio selections in this study are
introduced. Because the obtained results are based on the grey relational analysis algorithm
and artificial neural network weighting system, the literature about VIKOR and TOPSIS in
this article is refused to prevent an increase in the volume of paper.

3.1 The concept of a feature selection process


Data mining is a problem with hundreds or thousands of dimensions that can potentially be
used as predictors. As a result, realizing which variables or dimensions are significant
enough to be included in the modeling takes a lot of time and effort. To limit variable
selection, it is possible to use a feature selection algorithm to identify the dimensions that are
most important. The feature selection process includes three steps:
(1) screening: removing the problematic and insignificant predictors that have many
variances;
(2) ranking: classification of the remaining predictors from the screening process and
ranking them based on importance; and
(3) selection: identifying a subset of features to include in the model (Saeys et al., 2007).
At a time when organizations are overwhelmed with data, feature selection has the benefit of
simplifying and speeding up the modeling process. By focusing attention on issues that
matter the most, we can significantly reduce the time required for modeling calculations. Project
The feature selection algorithm is a combination of search methods for providing a new portfolio
subset of features, along with a measurement tool that rates the feature subsets. The simplest selection
algorithm is testing each feature subset in order to find the subset with the minimum error.
Two main types of feature selection algorithms include filters and wrappers.
3.1.1 The filter method. In this method, each feature receives a rating. Then the rates are
arranged and the features with the lowest ratings are omitted. The resulting feature subsets 655
are submitted to a classification system as input.
3.1.2 The wrappers method. This method is known as the black box. This method uses a
classification function to assess the competency of feature subsets. The method utilized a
feedback from the learning algorithm applied. A genetic algorithm is also used for finding
the valid features (Guyon and Elisseeff, 2006).

3.2 The concept of an artificial neural network


Neural network is an information processing system consisting of a large number of processing
elements known as neurons. Each neural network consists of a number of nodes that are
essentially neurons and communication weights that connect the nodes together. The input
data are multiplied by their corresponding weight and the sum enters the neurons. Each neuron
possesses an activation function. This input passes through the activation function which
determines the output. The number of neurons and neural network layers appropriate to the
issue under consideration is obtained based on trial and error. Network behavior also depends
on the communication between members. Each two layers of a network communicate each
other based on weight and through connections. The input data for each layer are calculated
based on the provided weights in a process of repeated adjustment which is known as the step
learning stage in a neural network (Ghafari and Mohammadzadeh, 2013). The structure of a
multi-layer perceptron neural network consists of three layers: input layer, output layer and
hidden layer. In the network architecture, each layer includes a number of neurons. The
number of neurons in the input and output layers is determined by the nature of the problem
being analyzed. The number of neurons in hidden layers as well as the number of these hidden
layers is determined by trial and error in order to reduce the model’s errors (Sogh et al., 2010).
3.2.1 Data preparation and classification. Since the numerical values of the important
indicators in each project in the organization and the average importance of each project for
the organization have the same units and measures, the artificial neural network training
process is performed well. It is important to note that the difference in the size of the
numbers has a great impact on the correction of network weights, and therefore, we need a
very large number of experimental data to balance the weights and achieve desired results,
but since in the current study, numerical values all have the same units and sizes, the
23 data obtained from the limited research population were deemed sufficient. For example,
if an input vector has input numbers that are much bigger than the other inputs, after
weight correction based on the large numbers, other inputs have to be applied to the
network numerously to be able to affect the network and balance the impact of bigger data
to an extent. Therefore, before neural network training, the weights need to be balanced out
to solve this issue. To balance the data, using Equations (1) and (2), data input and output
divided by the maximum amounts (Dashtbayaz and Ghanbarian, 2015):
p
Pn ¼ (1)
pmax

T
Tn ¼ (2)
T max
BIJ It should be noted that in Equations (1) and (2), Pn is the adjusted input vector, Tn is the
24,3 adjusted output vector, P and T are the input and output vectors, and Pmax and Tmax are the
maximum input and output vectors, respectively.
A schematic view of a neural network structure is shown in Figure 1.

3.3 Grey relational analysis


656 The term “grey” or “grey degree” refers to a state of understanding the system. When the
system is totally unknown, the system is black. On the contrary, if the system is known and
there is enough information, then the system is white. To assess this state of understanding
and knowledge, a system with a set of effective factors, each of which determines a degree of
information uncertainty related to their relation in the system, is selected. In the grey relational
analysis, the evolving process of similarity and dissimilarity between these factors is assessed
in the system (Sun et al., 2015). Grey relational analysis is an effective measurement method
which analyzes the uncertain relationships between the main factors (Pandey and Panda, 2015).
The main advantage of grey relational analysis is that using grey relational analysis allows the
decision maker to examine qualitative and quantitative relationships with relatively low data
or high variability (Kadier et al., 2015). The main steps of grey relational analysis include.
(1) Forming the decision matrix:
2 3
x1 ð1Þ x1 ð2Þ ... x1 ðnÞ
6 7
6 x2 ð1Þ x2 ð2Þ ... x2 ðnÞ 7
Xi ¼ 6
6 ^
7 (3)
4 ^ & ^ 7 5
xn ð1Þ xn ð2Þ . . . xn ðnÞ

(2) Normalization:
To normalize data, when the index is the type for which the bigger value is better,
we use Equation (4), if the index is the type for which the smaller the value the better,
we use Equation (5), and when the index has to be closer to the nominal scale, we use
Equation (6):
xi ðjÞmin xi ðjÞ
j
xni ¼ (4)
xi ðjÞmin xi ðjÞmin xi ðjÞ;
j j

Input Layer

Hidden Layer

Figure 1.
The general structure
of an artificial neural
Output Layer
network
max xi ðjÞxi ðjÞ Project
j
xni ðjÞ ¼ (5) portfolio
min xi ðjÞmin xi ðjÞ
j j selection

xi ðjÞxob ðjÞ
n
xi ðjÞ ¼ (6)
max xi ðjÞxob ðjÞ 657
j

Therefore, the normalized matrix is presented as follows:


2 n 3
x1 ð1Þ xn1 ð2Þ ... xn1 ðnÞ
6 n 7
6 x2 ð1Þ xn2 ð2Þ ... xn2 ðnÞ 7
X ni ¼ 6
6 ^
7 (7)
4 ^ & ^ 7 5
n n
xn ð1Þ xn ð2Þ . . . xnn ðnÞ

(3) Distance calculation


The distance to the reference solution is calculated based on the following
equation:
2 3
Wo1 ð1Þ Wo1 ð2Þ . . . Wo1 ðnÞ
6 7
  6 Wo2 ð1Þ Wo2 ð2Þ . . . Wo2 ðnÞ 7
Woi ðjÞ ¼ xno ðjÞxni ðjÞ ¼ 6
6
7
7 (8)
4 ^ ^ & ^ 5
Wom ð1Þ Wom ð2Þ . . . Wom ðnÞ
(4) Grey relational analysis coefficient calculation:
To calculate the grey relational coefficient γoi( j), the following equation is used:
D min þ xD max
goi ðjÞ ¼ (9)
Doi ðjÞþ xD max
In Equation (9), Doi ðjÞ ¼ max max D max, Doi ðjÞ ¼ min min and ‫ ﻭ‬xA ½0:1:
i j i j

3.4 The concept of a genetic algorithm


In the late 1970s, a scientist named John Holland proposed the idea of using genetic
algorithms for optimizing engineering problems. The main idea of this algorithm is that
inherited characteristics are transmitted by genes. If we assume that the set of human
feature are transmitted to the next generation through chromosomes, each gene in a
chromosome is representative of a feature. However, if the entire chromosome is transferred
to the next generation, the next generation will have characteristics similar to the previous
generation. Obviously, this does not happen in practice. Genetic exchange occurs commonly;
in other words, the probability of its occurrence is usually high. Genetic exchange between
chromosomes from parents occurs at a point called the crossover point. This is what
prevents a child from being totally similar to one of his/her parents. In addition to genetic
exchange between parents, another occurrence is genetic mutation. Genetic mutation is the
case wherein some of the genes accidentally undergo change. Of course, the number of such
occurrences is very low. However, this random change is very important, because it
BIJ prevents the checker to be caught in a local optimum point (Momeni, 2013). The main steps
24,3 of the genetic algorithm are as follows:
(1) Defining parameters and cost function: cost function provides the costs of the input
parameter set. These parameter sets are called chromosomes in the genetic
algorithm literature.
• Parameters and cost function selection: chromosomes are defined as the
658 following equation in the genetic algorithm. In this equation, pi (i ¼ 1, 2, 3,…,n)
are the parameters:
 
Chromosome ¼ p1 ; p2 ; p3 ; :::; pN par (10)
The cost of each chromosome in the genetic algorithm is calculated using the
following equation:
 
Cost ¼ f ðchromosomeÞ ¼ f p1 ; p2 ; p3 ; :::; pN par (11)
• Introduction of parameters into the genetic algorithm: the first step in this process
is normalization using Equation (12). The next step is coding based on Equation
(13). The final step is the coding of binary chromosomes using Equation (14):
pn  p10
pnorm ¼ ) 0 ppnorm p1 (12)
phi  p10
( )
X
m1
m r
gene½m ¼ round pnorm 2  gene½m2 (13)
r¼1

NX
gene
pquant ¼ gene½m2m þ2ðM þ 1Þ (14)
m¼1

(2) Introducing limits and constraints: optimization can be constrained or unconstrained.


In unconstrained optimization, the parameters can have any value, so the optimization
checker can freely explore the numbers, but in constrained optimizations, each
parameter has to be introduced to an algorithm. This is done in three ways:
• Introducing strict constraints that come in the form of W, o, ⩽, ⩾.
• Changing the variable forms so that new parameters introduced for the
algorithm inherently possess the required conditions.
• There may be a limited set of numbers, so the algorithm is forced to select the
desired parameters from them. In such a case also, the algorithm is unable to
choose any number.
(3) Initial population: the genetic algorithm begins the search for an optimum cost using
an initial population which consists of a large number of chromosomes. This initial
population has a chromosome or matrix with a random value between 0 and 1,
which is expressed in the following equation:
  
I pop ¼ Round Random N ipop  N bits ;
ðN bits ¼ number of parameters  number of bits in each parameterÞ (15)
(4) Natural selection: selection of chromosomes among the initial population to
participate in the production of a new generation.
(5) Pairing: of the chosen chromosomes waiting for reproduction, two are selected. The Project
reproduction goes on until the vacancy left by the omitted chromosomes has been filled. portfolio
(6) Crossover: crossover and reproduction in the genetic algorithm is defined as the selection
pairing of the two previously selected chromosomes and producing two
new chromosomes from each pair of parent chromosomes. This is the first step in
the search for the genetic algorithm to achieve optimal cost function and is called
cost function exploration. 659
(7) Genetic mutation: genetic mutation is the second method through which genetic
algorithms search for the cost function. Random mutations cause a change in a
portion of binary codes in the chromosomes. These portions are usually small.
Genes undergoing mutation are selected randomly.
(8) New generation formation: the new generation is the result of reproduction and
mutation in the past generation. Therefore, based on the continuous optimization
policy, each generation will have better values compared to the previous ones, and
this process will go on until the desired convergence is achieved.
(9) Testing convergence: to impose convergence we need appropriate and
practical factors, including statistical factors. These factors can be used alone or
in a combination to provide a convergence condition for the algorithm
(Kadier et al., 2015).

4. Case study
The case study of this research is Iran Power Plant Repair Company which is a subset of
Tavanir corporate. Activities of this company are manufacturing, rebuilding, repairs,
installations, designing, researches and monitoring. The main purpose of selecting this
company as a case study was project-based orientation of activities. The population of
this study consists of a group of 25 experts from Iran Power Plants Maintenance Company.
These participants all had five or more years of working experience in project management
in the fields of development, production and maintenance of power plant equipment as well
as project planning and management positions, along with working experience in
companies providing project services. One of the objectives of this study is to identify the
factors affecting successful project portfolio selection. To achieve this goal, 49 components
that were deemed important by the participating experts were extracted from A Guide to the
Project Management Body of Knowledge (PMBOK Guide). The relevant factors have been
presented in Table II.
The identified factors were presented to the participating experts in the form
of questionnaires. The answers were then assessed and analyzed by the feature selection
algorithm. The significance of each factor was determined through the feature selection
algorithm. The researchers then analyzed the top five high ranking factors as the most
important factors in project selection. The results are presented in Table III.
A list of all the projects the organization was involved with in the past five years was
created and a questionnaire was developed and handed to the experts to determine the
importance of each project. They were asked to rate the projects from 1 to 9 based on their
operational and professional experiences in the field. In another questionnaire, we asked the
participants to rate the importance of all the significant factors identified for each project
from a scale of 1 to 9. Data provided by the first questionnaire regarding the importance of
each project were used as input for extraction of Classification and Regression Tree. Data
obtained from the second questionnaire regarding the importance of each factor in the
projects were also used as output. Decision rule extraction and data analysis yielded
BIJ No. Criteria No. Criteria No. Criteria
24,3
1 Business framework 17 Activity list 33 Quality control measurements
2 Environmental impact 18 Milestone list 34 Team performance assessments
3 Work breakdown structure 19 Activity attributes 35 Work performance information
4 Organizational process assets 20 Resource calendars 36 Outputs from planning processes
5 Change request 21 Risk 37 Issue log
660 6
7
Project management plan
Project schedule
22
23
Activity resource requirements
Project charter
38
39
Identify stakeholders
Communication management plan
8 Activity cost estimates 24 Project scope statement 40 Project communication
9 Project deliverables 25 Project staff assignments 41 Risk management plan
10 Scope management plan 26 Cost management plan 42 Project documents
11 Requirements management 27 Human resource management 43 Performance reports
12 Requirements documentation 28 Contracts 44 Procurement management plan
13 Requirements traceability matrix 29 Project funding requirements 45 Source selection criteria
14 Accepted deliverables 30 Quality management plan 46 Seller proposals
Table II. 15 Schedule management plan 31 Develop project management plan 47 Make-or-buy decisions
Factors chosen from 16 Scope baseline 32 Quality metrics 48 Procurement statements of work
the PMBOK Guide 49 Procurement documents

Criteria Weight

Scope management 0.942


Project charter 0.927
Table III. Project management plan 0.92
Names and weights of Stakeholders 0.915
the important factors Risk 0.907

22 classes and 22 multi-criteria decision matrices. Nine repetitious classes were eliminated
and 13 unique classes remained. In order to rank the options available in each class using
the similarity to ideal solution method, the multi-criteria optimization and compromised
solution (VIKOR) and grey relational analysis, the researchers needed to calculate the
weight of each indicator. In order to calculate the weight of each factor, the researchers used
artificial neural network methods, including quick, dynamic, and multiple perceptron
methods. The results are presented in Table IV. In addition to artificial neural network
techniques, Shannon entropy weighting was also used to obtain the weights of the
indicators. Therefore, to classify the available projects into 13 classes, we used the three
methods of similarity to ideal solution, multi-criteria optimization and compromised solution
(VIKOR) and grey relational analysis, and for weighting the factors, the four above-
mentioned techniques were used, namely the quick, dynamic, and multiple perceptron and
Shannon entropy methods. The combination of weighting with artificial neural network and
Shannon entropy with the three above-mentioned methods resulted in 13 weighting-ranking
models. In order to assess the optimality of the combination, each of the methods applied for

Criteria Dynamic Multiple Quick

Scope management 0.2151 0.1776 0.1558


Table IV. Project charter 0.251 0.1441 0.2331
Weights obtained Project management plan 0.0948 0.2081 0.2336
from the artificial Stakeholders 0.2795 0.1708 0.1948
neural network Risk 0.1594 0.2994 0.1829
weighting and ranking, along with all the models used, was modeled through a parametric Project
approach and implemented by the genetic algorithm. Results of the genetic algorithm portfolio
implementation are presented in Table V. As is clear from the results shown in the table, the selection
combination of the quick weighting method and grey relational analysis as the ranking
method makes the best combination of methods in this case study.
In this study, the artificial neural network architecture consists of a multi-layer
perceptron neural network and the neural network is structured with a linear saturation 661
function. Also, in the multi-layer perceptron method, the input layer consisted of
25 neurons; the hidden layer possessed seven neurons after proper network learning, and
finally, the output layer included 26 neurons. In the dynamic method, the input layer
consisted of 11 neurons, while the hidden layer included two neurons after proper network
learning, and the output layer included 26 neurons. In this study the researchers have
utilized a multi-layer perceptron network, a hidden layer with a different number of
neurons and a linear saturation function. Among the various methods of back-propagation
learning, network learning was chosen due to faster convergence. Back-propagation is
based on the error correction learning rule which consists of two main forward and
backward paths. In the forward path, the input vector is applied to the network and its
impacts are spread through the hidden layer to the output layer. The output vector
produces the network’s real answer.
Software and hardware specifications of the computer used for all computations have
been presented in Table VI.
Figure 2 shows the genetic algorithm output obtained by solving a model created by the
grey ranking technique using the weights obtained from the quick method.

5. Sensitivity analysis
In order to analyze the sensitivity of each modeled pattern, the stability of the ratings in
each class has been subject to assessment. To do this, after ranking each class with the

Ranking technic/weighting system Best answer

GRA/quick 119.85
TOPSIS/quick 100.6370
GRA/ Shannon entropy 6.0277
TOPSIS/multiple 5.6651
GRA/multiple 5.46
VIKOR/multiple 5.2257
GRA/dynamic 5.0467
VIKOR/Shannon entropy 4.9041
VIKOR/quick 4.7608 Table V.
VIKOR/dynamic 4.1943 The results of
TOPSIS/ Shannon entropy 3.2188 genetic algorithm
TOPSIS/dynamic 1.5779 in order of size

Name Specification
Table VI.
Operating system Microsoft Windows 7 Ultimate Software and
Processor Intel(R) core(TM) i53210M CPU @ 2.50 GHz hardware
RAM 4.00 GB specifications of
Memory 698.64 GB the computer used
Cache 1,771 for all calculations
BIJ × 108 Best: –119.858 Mean: –119.858
9
24,3 8
Best fitness
Mean fitness

7
6

Fitness value
5

662 4
3
2
1
0
Figure 2.
Genetic algorithm –1
0 1
output 10 10 102 103 104
Generation

relevant methods and calculating the weights of indicators, options for each class were
added to the class separately and the rankings were compared with the original results.
If adding a repetitive option changed the ranking results compared to the original results,
the ranking method failed. Otherwise, the stability of the classes was proven. Based on the
results obtained from stability analysis, it is observed that all classes, except the ones using
the grey relational analysis and Shannon entropy weighting, classes using the similarity to
ideal solution technique and multiple perceptron weighting, and classes using multi-criteria
optimization and compromised solution (VIKOR) with multiple weighting for ranking, show
proper stability. As shown in Table VII, results show that for the examined data in this
study, the combination of the grey relational analysis technique and quick weighting proves
to yield better ranking results for project portfolio selection compared to other techniques
and methods.

6. Conclusion
The primary objective of this study was to identify the factors influencing successful project
portfolio selection. To do this, we extracted the factors provided by A Guide to the Project
Management Body of Knowledge (PMBOK Guide) and presented them to a population of
experts in the industry under investigation. Table II shows the mentioned factors which
are selected for studying. Experts were asked to weight the importance of each factor.
Then the feature selection algorithm was implemented to achieve the research objective.

N Ranking technic Weighting system Robustness status Non-robust classes

1 GRA Dynamic Robust –


2 GRA Multiple Robust –
3 GRA Quick Robust –
4 GRA Shannon entropy Non-robust 1/2/5/11/12/13/16/21
5 TOPSIS Dynamic Robust –
6 TOPSIS Multiple Non-robust 12
7 TOPSIS Quick Robust –
8 TOPSIS Shannon entropy Robust –
9 VIKOR Dynamic Non-robust 22
Table VII. 10 VIKOR Multiple Non-robust 12
Sensitivity analysis 11 VIKOR Quick Robust –
results 12 VIKOR Shannon entropy Robust –
From the 49 studied indicators, the five factors including scope management, project Project
charter, project management, stakeholders and risk gained the most weight and were portfolio
established as the most important indicators. Table III shows the result of implantation of selection
the feature selection algorithm. Another aim of this research included project portfolio
selection with the help of an artificial neural network, grey relational analysis classification
and regression tree. The reason for using neural network was to weight the important
indictors identified by the feature selection algorithm. This was done through three different 663
neural network methods, namely quick learning, multiple perceptron and dynamic methods.
Weight obtained for each criterion from the artificial neural network is presented in
Table IV. The organization projects were classified in 13 classes by the decision tree
algorithm. These classes were then analyzed using three methods of grey relational
analysis, similarity to ideal solution and multi-criteria optimization and compromised
solution (VIKOR). To determine the best ranking weighting technique, we used the genetic
algorithm. The results obtained from the genetic algorithm revealed that the combination of
grey relational analysis and quick learning in neural network provides the best results.
Based on the results obtained from the case study of this research, the best selection
technique is to choose projects ranked highly by the grey relational analysis and quick
weighting. The results of implementing the genetic algorithm are clearly shown in Table V.
The sensitivity analysis test played a constructive role in examining the accuracy
and reliability of the proposed models, and it can be firmly argued that the results have had
an important role in validating the findings of this study. Stability of the designed
models and multiple decision-making matrices confirm the superiority of grey relational
analysis and quick learning methods of the artificial neural network compared to the
other techniques investigated in this research. The results of sensitivity analysis which
shows the robustness and non-robustness of each ranking technic-weighting system are
shown in Table VII.
Since the selection of this research’s case study is real and highly active and the data are
obtained from experts in the relevant industry, the authors can claim that the obtained
results are practical, real and implantable in relevant industries especially in project-based
industries.

References
Arasteh, A., Aliahmadi, A. and Omran, M.M. (2014a), “A multi-stage multi criteria model for
portfolio management”, Arabian Journal for Science and Engineering, Vol. 39 No. 5,
pp. 4269-4283.
Arasteh, A., Aliahmadi, A. and Omran, M.M. (2014b), “Application of gray systems and fuzzy sets in
combination with real options theory in project portfolio management”, Arabian Journal for
Science and Engineering, Vol. 39 No. 8, pp. 6489-6506.
Bard, J.F. (1988), “Short-term scheduling of thermal-electric generators using Lagrangian relaxation”,
Operations Research journal, Vol. 36 No. 5, pp. 756-766.
Dashtbayaz, M. and Ghanbarian, M. (2015), “Comparing neural network methods in the process of
turning a polymer matrix composite material modeling”, Amirkabir mechanical engineering
press, Vol. 47 No. 2, pp. 83-98.
Eshlaghy, A.T. and Razi, F.F. (2015), “A hybrid grey-based k-means and genetic algorithm for project
selection”, International Journal of Business Information Systems, Vol. 18 No. 2, pp. 141-159.
Faezy, R.F. (2015), “A grey-based fuzzy electre model for project selection”, Journal of Optimization in
Industrial Engineering, Vol. 8 No. 17, pp. 57-66.
Gabriel, S.A., Kumar, S., Ordonez, J. and Nasserian, A. (2006), “A multi-objective optimization model for
project selection with probabilistic considerations”, Socio-Economic Planning Sciences., Vol. 40
No. 4, pp. 297-313.
BIJ Ghafari, R.M. and Mohammadzadeh, A. (2013), “TEC regional modeling using artificial neural
24,3 network and polynomial model in Iran”, Science and Technology Mapping Journal, Vol. 4 No. 3,
pp. 51-61.
Guyon, I. and Elisseeff, A. (2006), “An introduction to variable and feature selection”, The Journal of
Machine Learning Research, Vol. 3, pp. 1157-1182.
Hall, D.L. and Nauda, A. (1990), “An interactive approach for selecting IR&D projects”, Engineering
Management, IEEE Transactions, Vol. 37 No. 2, pp. 126-133.
664
Haughton, D. and Oulabi, S. (1997), “Direct marketing modeling with CART and CHAID”, Journal of
Interactive Marketing, Vol. 11 No. 4, pp. 42-52.
Jiang, J.J. and Klein, G. (1999), “Project selection criteria by strategic orientation”, Information &
Management journal, Vol. 36 No. 2, pp. 63-75.
Kadier, A., Abdeshahian, P., Simayi, Y., Ismail, M., Hamid, A.A. and Kalil, M.S. (2015), “Grey relational
analysis for comparative assessment of different cathode materials in microbial electrolysis
cells”, Energy International Journal, Vol. 90, pp. 1556-1562.
Khademolqorani, S. and Hamadani, A.Z. (2015), “Development of a decision support system for
handling health insurance deduction”, Development, Vol. 6 No. 2, pp 44-51.
Kim, Y. and Chang, H. (2013), “A study on project selection framework for future ICT technologies”,
Wireless Personal Communications, Vol. 73 No. 4, pp. 1591-1600.
Momeni, M. (2013), New Topics in Operation Research, 5th ed., Mansour Momeni publisher, Tehran.
Pandey, R.K. and Panda, S.S. (2015), “Optimization of multiple quality characteristics in bone drilling
using grey relational analysis”, Journal of Orthopaedics, Vol. 1 No. 12, pp. 39-45.
Petković, D., AbHamid, S.H., Ćojbašić, Ž. and Pavlović, N.T. (2014), “Adapting project management
method and ANFIS strategy for variables selection and analyzing wind turbine wake effect”,
Natural Hazards, Vol. 74 No. 2, pp. 463-475.
Project Management Institute (2013), A Guide to the Project Management Body of Knowledge
(PMBOK), Chapter 4-10 (Trans by Ashtiani), Adineh Publication, Hamedan.
Rabbani, M., Bajestani, M.A. and Khoshkhou, G.B. (2010), “A multi-objective particle swarm
optimization for project selection problem”, Expert Systems with Applications, Vol. 37 No. 1,
pp. 315-321.
Razi, F.F. (2014), “A hybrid grey relational analysis and non-dominated sorting genetic algorithm-ii for
project portfolio selection”, Advances in Operations Research journal, Vol. 2, pp. 1-9.
Razi, F.F., Eshlaghy, A.T., Nazemi, J., Alborzi, M. and Poorebrahimi, A. (2015), “A hybrid grey-based
fuzzy C-means and multiple objective genetic algorithms for project portfolio selection”,
International Journal of Industrial and Systems Engineering, Vol. 21 No. 2, pp. 154-179.
Saeys, Y., Inza, I. and Larrañaga, P. (2007), “A review of feature selection techniques in bioinformatics”,
Bioinformatics Journal, Vol. 23 No. 19, pp. 2507-2517.
Sogh, M.G., Mosaedi, A. and Moosavi, H. (2010), “Investigating the impact of pre-processing neural
network input using regression and gamma test in order to estimate the daily evaporating and
Sweating”, Water and soil Ferdowsi University Press, Vol. 24 No. 3, pp. 610-624.
Sun, B., Jiang, J., Zheng, F., Zhao, W., Liaw, B.Y., Ruan, H. and Zhang, W. (2015), “Practical state of
health estimation of power batteries based on Delphi method and grey relational grade
analysis”, Journal of Power Sources, Vol. 282, pp. 146-157.
Vahdani, B., Mousavi, S.M., Hashemi, H., Mousakhani, M. and Ebrahimnejad, S. (2014), “A new hybrid
model based on least squares support vector machine for project selection problem in construction
industry”, Arabian Journal for Science and Engineering, Vol. 39 No. 5, pp. 4301-4314.
Wang, J., Yujie, X. and Zhun, L. (2009), “Research on project selection system of pre-evaluation of
engineering design project bidding”, International Journal of Project Management, Vol. 27,
pp. 584-599.
Yang, C.L., Chiang, S.J., Huang, R.H. and Lin, Y.A. (2014), “Hybrid decision model for information
project selection”, Journal of Quality & Quantity, Vol. 47 No. 4, pp. 2129-2142.
Yavuz, M. and Altay, B.L. (2015), “Reclamation project selection using fuzzy decision-making Project
methods”, Environmental Earth Sciences, Vol. 73 No. 10, pp. 6167-6179. portfolio
Zhang, P., Zhang, J. and Shi, Y. (2007, May), “A new multi-criteria quadratic-programming linear selection
classification model for VIP e-mail analysis”, International Conference on Computational Science,
Springer Berlin Heidelberg, pp. 499-502.

Further reading
Fatahi, P. (2014), Metaheuristic Algorithms, 3rd ed., Boualisina University Press, Hamedan.
665

Corresponding author
Farshad Faezy Razi can be contacted at: f.faezi@semnaniau.ac.ir

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: permissions@emeraldinsight.com
Reproduced with permission of copyright owner. Further
reproduction prohibited without permission.

S-ar putea să vă placă și