See discussions, stats, and author profiles for this publication at: [Link]

net/publication/245283775

Dynamic Prediction of Project Success Using Artificial Intelligence

Article in Journal of Construction Engineering and Management · April 2007

DOI: 10.1061/(ASCE)0733-9364(2007)133:4(316)

CITATIONS READS
47 2,681

2 authors:

Chien-Ho Ko Min-Yuan Cheng

Clemson University National Taiwan University of Science and Technology
69 PUBLICATIONS 902 CITATIONS 241 PUBLICATIONS 5,175 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Min-Yuan Cheng on 12 May 2014.

The user has requested enhancement of the downloaded file.

 Dynamic Prediction of Project Success Using
Artificial Intelligence
Chien-Ho Ko, Ph.D.1; and Min-Yuan Cheng2

Abstract: The purpose of construction management is to successfully accomplish projects, which requires a continuous monitoring and
control procedure. To dynamically predict project success, this research proposes an evolutionary project success prediction model
共EPSPM兲. The model is developed based on a hybrid approach that fuses genetic algorithms 共GAs兲, fuzzy logic 共FL兲, and neural networks
共NNs兲. In EPSPM, GAs are primarily used for optimization, FL for approximate reasoning, and NNs for input-output mapping. Further-
more, the model integrates the process of continuous assessment of project performance to dynamically select factors that influence
project success. The validation results show that the proposed EPSPM, driven by a hybrid artificial intelligence technique, could be used
as an intelligent decision support system, for project managers, to control projects in a real time base.
DOI: 10.1061/共ASCE兲0733-9364共2007兲133:4共316兲
CE Database subject headings: Project management; Predictions; Fuzzy sets; Neural networks; Evolutionary computation; Artificial
intelligence.

Introduction Several studies have also endeavored to reveal factors that

influence project success. Lists of variables have abounded in the
Success on a project means that certain expectations for a given literature. Chan et al. 共2004a, 2005兲 developed a conceptual
participant are met, whether owner, planner, engineer, contractor, framework on critical success factors. Five major groups of
or operator 共Sanvido et al. 1992兲. The measurements of project independent variables: namely, project-related factors, project
success in the construction industry are: Cost, schedule, perfor- procedures, project management actions, human-related factors,
mance, and safety 共Hughes et al. 2004兲. Hughes et al. 共2004兲 and external environment, are identified as crucial to project suc-
developed a Construction Project Success Survey instrument to cess. Nguyen et al. 共2004兲 uncovered success factors for large
identify important success metrics before the start of a project, projects, using factor analysis. A survey questionnaire was used to
and to evaluate the level of success achieved at project com- collect data from practitioners. These factors were grouped under
four categories: Comfort, competence, commitment, and commu-
pletion. The measuring metrics include the objective 共such as,
nication. Through a postal questionnaire survey, Chan et al.
cost, schedule, performance, and safety兲 and subjective consider-
共2004b兲 analyzed a set of success factors using factor analysis and
ations. Griffith et al. 共1999兲 developed an objective metrics that
multiple regression. Chua et al. 共1997兲 identified key project man-
comprised of four variables: Budget achievement, schedule
agement attributes associated with achieving successful budget
achievement, design capacity, and plant utilization. The authors
performance, using a neural network approach. Eight key man-
discovered that despite the complexities involved in measuring
agement factors were identified in the research: 共1兲 number of
project success, a measurement can be developed based on objec-
organizational levels between the project manager and craft work-
tive project performance. Shields et al. 共2003兲 established a
ers, 共2兲 amount of detailed design completed at the start of con-
metric for measuring the success of the construction phase of struction, 共3兲 number of control meetings during the construction
projects. The metric is quantitatively derived from the Construc- phase, 共4兲 number of budget updates, 共5兲 implementation of a
tion Industry Institute’s Benchmarking and Metrics 共BM and M兲 constructability program, 共6兲 team turnover, 共7兲 amount of money
database. Scoring ranges from 0–10 based on statistical expended on controlling the project, and 共8兲 the project manager’s
distributions—for construction cost growth, construction schedule technical experience.
growth, lost workday case incident rates, and rework factor—are Identification of the success factors could formulate effective
presented. strategies for minimizing construction conflicts; whereas, incor-
porating success factors into developing a research model for
1
Dept. of Construction Engineering, National Taiwan Univ. of Science predicting the success levels improves project performance.
and Technology, Taiwan. E-mail: D8805102@[Link] Parfitt and Sanvido 共1993兲 proposed a checklist that can be used
2
Professor, Dept. of Construction Engineering, National Taiwan Univ. by building professionals, as a guideline in predicting the success
of Science and Technology, Taiwan. of a project. Items represented in the checklist include questions
Note. Discussion open until September 1, 2007. Separate discussions similar to the questions and categories initially used to gather
must be submitted for individual papers. To extend the closing date by
information for the identification of the critical project success
one month, a written request must be filed with the ASCE Managing
Editor. The manuscript for this paper was submitted for review and pos- factors. Chua et al. 共1997兲 developed a predictive model of
sible publication on October 1, 2004; approved on July 31, 2006. This project success, using neural networks to forecast budget perfor-
paper is part of the Journal of Construction Engineering and Manage- mance of a construction project.
ment, Vol. 133, No. 4, April 1, 2007. ©ASCE, ISSN 0733-9364/2007/4- Previous results in measuring and predicting project success
316–324/$25.00. show that the measurements of project success are crucial for

316 / JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT © ASCE / APRIL 2007

Downloaded 14 Nov 2008 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 Fig. 1. EFNIM architecture

controlling projects, so as to achieve project objectives. Besides,

statistical methods provide a theoretical basis for analyzing fac-
tors of project success from questionnaire surveys or collected Fig. 2. Predictor analysis of owner expenditures 共adapted from
data. However, previous research in predicting project success Russell et al. 1997兲
either adopted fixed factors at various points of time, or used an
inference method that has a few difficulties and was inadequate.
From an owner’s perspective, the definition of a successful
project is one that meets or exceeds budgetary and schedule ex-
GA is a stochastic searching process based on the mechanism of
pectations. On the other hand, a less-than-successful project fails
natural selection and natural genetics. It combines the survival-
to meet budgetary and/or schedule expectations 共CII 1996兲.
of-the-fittest principles among strings with a randomized informa-
Project managers are responsible for project success. In order to
tion exchange to form an evolutionary algorithm 共Goldberg
ensure that the projects can be accomplished successfully, project
1989兲. Under GA terminology, an individual represents a solution
managers have to continuously monitor the project performance,
to a problem. The group of individuals at each generation is a
so that they can take proper corrective actions to control the
population. A generation denotes that a new population of indi-
project.
viduals is created. In artificial genetic systems, each individual is
Project outcomes are affected by different factors at diverse
represented by a string-named chromosome.
time points. During the course of a project, predicting project
outcomes at different stages requires analysis of dissimilar factors
共Russel et al. 1997; Griffith et al. 1999兲. A dynamic prediction Fuzzy Logic
methodology is, thus, required for project managers to continu- FL was first developed by Zadeh, in the 1960s, for representing
ously monitor project performance. However, every time point uncertain and imprecise information 共Zadeh 1965兲. In a wide
has numerous time-dependent variables affecting the project’s sense, FL is synonymous with fuzzy set theory; that is, the theory
outcomes. In addition, due to the nature of the construction in- of classes with unclear boundaries. In a narrow sense, FL is a
dustry, those variables are uncertain 共Barraza et al. 2000兲. Dy- logic system that is intended to serve as logic of approximate
namically predicting the project’s outcome under such complex reasoning 共Zadeh 1994兲. Fuzzy logic systems 共FLSs兲 simulate the
and uncertain circumstances is never easy. Human experts can high level human decision making process, which aims at mod-
judge a project’s outcome according to their knowledge; however, eling the imprecise modes of reasoning to make rational decisions
the significance of these judgments is restricted by their subjec- in an environment of uncertainty and imprecision. In general,
tive cognitions and/or limited knowledge. FLSs contain four major components: Fuzzifier, inference engine,
Artificial intelligence 共AI兲 is concerned with building com- rule base, and defuzzifier 共Klir and Yuan 1995兲.
puter systems that solve the problem intelligently by emulating
the human brain. AI technology provides techniques for the com-
puter programs to carry out a variety of tasks, at which humans Neural Networks
are currently better 共Haykin 1999兲. Consequently, AI paradigms The concept of artificial neural networks originated from model-
are appropriate for solving project management problems 共Ko ing the brain in the 1950s. Artificial neural networks, commonly
2002兲. The most popular AI paradigms are genetic algorithms referred to as “neural networks,” are massively parallel-
共GAs兲, fuzzy logic 共FL兲, and neural networks 共NNs兲. The combi- distributed processors made up of simple processing units 共neu-
nation of GAs, FL, and NNs offsets the demerits of one paradigm rons兲, which perform computations and store knowledge 共Haykin
by the merits of another 共Martin and Jain 1999兲. In the last few 1999兲. Since the brain is robust, fault tolerant, highly parallel, and
years, several articles have been devoted to the study of fusing able to deal with noisy information, modeling the functions of the
GAs, FL, and NNs to derive a better model performance than brain provides an alternative approach to conventional methods
those using a single conventional method 共Linkens and Nyongesa 共Aggarwal and Song 1997兲. NNs model the brain using networks
1996, Cheng and Ko 2003兲. composed of neurons.

Literature Review Evolutionary Fuzzy Neural Inference Model

One of the pioneering models hybridizing GAs, FL, and NNs for
Genetic Algorithms
solving construction management problems was developed by Ko
GAs were first proposed by Holland, in the 1970s, for mimicking 共2002兲. Architecture of the model is shown in Fig. 1. In Fig. 1, the
some observed natural evolution processes 共Holland 1975兲. The combination of GAs, FL, and NNs offset the demerits of one

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT © ASCE / APRIL 2007 / 317

Downloaded 14 Nov 2008 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 Fig. 3. EPSPM architecture

paradigm by the merits of another. FL is primarily concerned with

imprecision and approximate reasoning; NN with fuzzy input-
output mapping; and GAs with optimization.
The model was developed based on FL that mimics the high Fig. 5. Model application process
level of human inference process. However, FL encounters diffi-
culties in determining fuzzy rules and distributions of member-
ship functions 共MFs兲, while problem complexity increases. NNs Continuous Assessment of Project Success
are adaptive to changes in the surrounding environment through
learning. The nonlinearity of networks captures complex interac- To monitor the progress of a project, continuous assessment of
tions between input and output variables, which provides a prom- project success 共CAPP兲 共Russell et al. 1996兲 was developed to
ising direction to represent fuzzy rules. The combination, of these identify continuous variables that have the ability for predicting
two methods into an integrated system, appears to be a promising project outcome. A continuous variable is a time-dependent quan-
path toward the development of intelligent systems capable of tity whose value can be collected at several points of time during
capturing qualities characterizing the human brain 共Canuto et al. the course of a project 共e.g., contractor expenditure, invoice paid
1999; Rajasekaran and Vijayalakshmi Pai 2000兲. Although, the by the contractor, owner project commitment, and project cost兲
integration of FL and NN is more reasonable than traditional FL 共Russell et al. 1996兲. In the model, a continuous variable is con-
to simulate the characteristics and process of human inference, sidered as a predictor of project outcome, while its difference
the NN has demonstrated the difficulty in selecting an appropriate between the s-curves for “successful” and “less-than-successful”
topology for learning different tasks, as well as appropriate pa- projects is significant. The significance between the two projects
rameters, for a network. In addition, the determination of suitable is analyzed through hypothesis tests conducted using the student’s
distribution for MFs, for solving disparate problems is time con- t distribution. The significant difference exists, if the value of
suming, and the difficulties increase with the problem complexity. alpha 共Type I error兲 is less than or equal to 0.10. An alpha equal
To conquer the remaining difficulties, the evolutionary fuzzy neu- to 0.10 is frequently used as the critical level for rejecting the null
ral inference model 共EFNIM兲 employed GA to simultaneously hypothesis. Fig. 2 displays an example analyzing whether or not a
search for the fittest shapes of MFs, optimum NN topology, and variable—owner expenditure—is a predictor at different points in
optimum parameters of NN, including interconnections status, time. In the figure, the x-axis represents time from 0 共beginning of
synaptic weights, bias values, and slopes of activation functions. design and procumbent兲 to 100% completion 共end of construc-
tion兲 during a project life cycle; while the y axis represents the
normalized value of the continuous variable. The vertical bars,
denoting the significance level, below 0.10 represents a statistical
difference between successful and less-than-successful projects.
Observing Fig. 2, variable owner expenditures is a predictor of
project outcome from about 26 to 50% and from 69 to 92%
project complete.

Evolutionary Project Success Prediction Model

Model Architecture
Project success can be predicted at different points in time with
their corresponding factors while it is progressing. In the begin-
ning of a project, the procurement of material and equipment
effects the outcome of the project. By contrast, in the construction
phase, project success can be predicted by the factor of owner
expenditures. Previous projects provide patterns for analyzing
those factors and project success. The architecture of EPSPM
shown in Fig. 3 contains four components: The project database,
factors of project success, adaptation engine, and predictive solu-
tion. The project database contains previous cases of projects
that serve as patterns for analysis. The element of the factors
of a project’s success determines which factors are effective for
Fig. 4. EPSPM adaptation structure predicting the project outcome at a specific point in time. Com-

318 / JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT © ASCE / APRIL 2007

Downloaded 14 Nov 2008 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
Table 1. Time-Dependant Variables Analyzed by CAPP Initialize Population
Sign. The first step of the adaptation process is to randomly generate a
Variable name level set of initial solutions. Each solution encodes model variables into
Actual design % complete 0.03 a binary string to simulate a natural chromosome. Every string is
Actual owner expenditures 0.00 comprised of two segments: MF substring and NN substring. Two
Cost of contractor project commitments 0.01 codification methods—summit and width representation method
Cost of owner project commitments 0.04 共SWRM兲 and block representation method 共BRM兲—proposed by
Actual owner effort hours 0.03 Ko 共2002兲 are employed to encode MFs and NNs into substrings.
Recordable incident rate 共by period兲 0.01
Actual overtime work 0.02 Evaluate Individuals
Cost of change orders 0.08
The aim of the adaptation process is to obtain a predictive model
Quantity of change orders 0.02 with high accuracy and good generalization properties. The model
Days lost to weather 共gross working days兲 0.08 accuracy on input patterns, acquired from the project’s database,
Note: Significance level below 0.10 represents a statistical difference can be improved by increasing the network complexity. However,
between successful and less-than-successful. Variable names are defined an accurate model fit to input patterns does not mean that the
in Russel et al. 共1996兲. overall problem behaviors are captured well. A large network size
has higher computational cost. Also, in general, it suffers from an
overfitting of the data in the input patterns and deterioration of
generalization properties 共Maier et al. 2000兲. Thus, the objective
plex relationships between the factors and the project outcome are of the adaptation engine is to preserve the acceptable prediction
automatically identified through previous patterns, using an adap- accuracy—using the fittest shapes of MFs with the minimum NN
tation engine. Finally, the derived solution can be used as a pre- topology and optimum NN parameters, which is posed as an op-
dictive model for forecasting a project’s outcome in a real-time timization problem. The objective function of the model, f ob, is a
base. combination of the model accuracy and model complexity, as
given in Eq. 共1兲
Model Adaptation Process f ob = caw ⫻ ser + ccw ⫻ mc 共1兲
The adaptation engine is driven by EFNIM through GAs. The where c ⫽accuracy weighting coefficient; s ⫽prediction error
aw er

adaptation process is shown in Fig. 4. In the process, P共t兲 denotes between actual output and desired output; ccw⫽complexity
a population at generation t, Po共t兲 is an offspring population at weighting coefficient; and mc is the model complexity, which is
generation t, and Pm共t兲 indicates a mutation population at genera- simply formulated by the number of active connections in the
tion t. Each procedure is defined in the next sections. network.

Table 2. Patterns for Dynamic Prediction of Project Success 共Shown and Reused CAPP Database with CII’s Kind Permission兲
Input patterns
Input Output
Pattern
number 共1兲 共2兲 共3兲 共4兲 共5兲 共6兲 共7兲 共8兲 共9兲 共10兲 共11兲
1 0.0000 0.9700 0.5322 0.9253 0.0000 0.7555 0.0000 0.0000 0.7877 0.7719 1.0000
2 0.0000 1.0000 0.5625 0.0000 0.9372 0.6954 0.3103 0.0000 0.0000 0.0000 1.0000
3 0.3125 1.0000 0.0000 0.7581 0.0000 0.0000 0.0000 0.0000 1.0000 1.0000 0.0000
4 0.0000 0.9950 0.0000 0.9555 0.0000 0.0000 0.0000 0.0000 0.8665 0.6304 1.0000
5 0.0000 0.0000 0.4694 0.6044 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.6667
6 0.0000 0.0000 0.0000 0.7098 0.0000 0.0000 0.0000 0.3102 0.0000 0.0000 0.0000
7 0.5897 0.9767 0.0000 0.0000 0.9598 0.7708 0.0000 0.0000 0.0000 0.0000 1.0000
8 0.0000 0.9592 0.0000 0.8629 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000
9 0.0000 0.9150 0.8931 0.9170 0.9960 0.0000 0.0000 0.0000 0.0000 0.0000 0.6667
10 0.0000 1.0000 0.3808 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000
11 0.7647 0.0000 0.4137 0.0000 0.5522 0.5869 0.0000 0.1507 0.0000 0.0000 1.0000
12 0.0000 0.0000 0.8654 0.0000 0.9813 0.7909 0.0000 0.0000 0.0000 0.0000 0.3333
Test patterns
13 0.0000 1.0000 0.0000 0.7472 1.0000 0.0000 1.0000 0.0000 0.9612 0.9348 0.0000
14 0.0000 1.0000 0.0000 0.9537 1.0000 0.0000 0.0000 0.1448 0.0000 0.0000 0.6667
15 0.0000 0.0000 0.0000 0.8574 0.8396 0.0000 0.0000 0.0000 0.5000 0.5000 1.0000
Note: The captions of the above columns are: 共1兲 actual design % complete; 共2兲 actual owner expenditures; 共3兲 cost of contractor project commitments;
共4兲 cost of owner project commitments; 共5兲 actual owner effort hours; 共6兲 recordable incident rate; 共7兲 actual overtime work; 共8兲 cost of change orders; 共9兲
quantity of change orders; 共10兲 day lost to weather; and 共11兲 project performance.

JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT © ASCE / APRIL 2007 / 319

Downloaded 14 Nov 2008 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
 Fig. 6. EPSPM management module

Evaluate Fitness Function Select Individuals

Fitness is a major index to evaluate the status of the chromo- The selection process emulates the survival-of-the-fittest mecha-
somes. A larger fitness value indicates the result more closely nism in nature. It selects a new population, with respect to the
achieves the model objective. In this research, fitness function is probability distribution, based on fitness for survival. Since ge-
the reciprocal of objective function and is given by Eq. 共2兲 netic operators are blind in nature, to avoid fitter chromosomes
from being lost in the evolutionary process, the adaptation engine
1 uses a modified enlarged sampling space that contains all of the
vfik = 共2兲 population, offspring population, and their mutation.
vobk

where vfik and vobk⫽fitness and objective values of chromosome Model Application Process
k.
Specific processes and methods used to implement EPSPM are
summarized in Fig. 5. Referring to Fig. 5, the blocks on the left-
Perform Crossover hand side are the procedures used to implement the model. The
The crossover repeatedly exchanges high performance notations blocks on the right-hand side are the detailed methods and at-
in the search for better and better performances. The adaptation tributes concerned with the execution of the tasks on the left-hand
engine uses one-cut-point crossover and exchanges the right parts side.
of their parents. After crossover, the summit positions of MFs,
widths of MFs, hidden layers, hidden neurons, interconnections, Calculate Project Completion
biases, and activation slopes of parents are exchanged. To analyze the factors of success in the next step, progress is
converted to a project completion. Time is normalized such that 0
percent and 100 percent can correspond to a specific date. A ratio
Perform Mutation of to-date to planned is used in this investigation. For example, a
The purpose of mutation is to adjust the value of summits and ratio of 100 days passed the date of the beginning to a duration of
widths of MFs, interconnections, weights, biases, and activation 1000 days is 10% completion.
slopes for better performance. It alters one or more genes with a
probability 共pge兲, which is smaller than or equal to mutation rate Analyze Influencing Factors
共pmu兲. Mutation operation compares the gene’s pge with pmu bit by The factors of success, at a specific completion, are analyzed
bit. If pge ⱕ pmu, then the value of the gene will be altered. using CAPP software. Hypothesis tests with student’s t distribu-

320 / JOURNAL OF CONSTRUCTION ENGINEERING AND MANAGEMENT © ASCE / APRIL 2007

Downloaded 14 Nov 2008 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]
tion are used to identify influencing factors, whose value of alpha 共1996兲兲. Using the CAPP software, 10 of 76 time-dependent vari-
共type I error兲 less than or equal to 0.10 are selected. In statistical ables whose significance level below 0.1 at 67% completion were
research, an alpha equal to 0.1 is frequently used to reject the null identified, as summarized in Table 1.
hypothesis. CAPP software analyzes 76 continuous variables The original patterns are acquired from CAPP system database
based on 54 construction projects. with CII’s kind permission. The database contains 54 projects,
which are the same data used to analyze the significance level.
Collect Training and Test Patterns This study selects 15 of 54 available projects from the database
Training and test patterns are collected from CAPP database that where construction type is process plant projects, designer con-
contains 54 construction projects. These projects are real data tractor type is lump sum construction contract, and the stability of
collected by Russell et al. 共1996兲 from the 16 representative Con- projects is at the completed stage. In AI field, 80% available
struction Industry Institute 共CII兲 member companies. patterns for training purpose, and 20% for test purpose are fre-
quently used. The 12 of them 共80%兲 are, thus, treated as input
Search for Predictive Solution patterns for evolvement, and the other three projects 共20%兲 are
A predictive model 共i.e., fuzzy neural network model兲 is adopted selected as test patterns. Since this validation attempts to predict
using EFNIM according to the identified factors of success with a the project outcomes at 67% completion, values of time-
54 projects database. The model is specifically generated to pre- dependent variables at 67% completion are collected. The nor-
dict project outcomes at current project progress 共project com- malized data of 15 projects are shown in Table 2. In the table,
plete兲. More details about the application of EFNIM can be found four categories of project performance defined by CAPP software,
in Ko and Cheng 共2003兲. namely successful, on time, or on budget, less-than-successful,
and disastrous are linearly represented using 0, 0.3333, 0.6667,
and 1, respectively.
Model Validation EPSPM provides three modules for application: Adaptation
module, inference module, and management module. The 15 se-
Model Implementation lected projects, with 10 identified factors of success, are entered
into EPSPM through management module 共see Fig. 6兲. Adapta-
Performance of the proposed EPSPM is validated using a 67% tion module shown in Fig. 7 is then used to evolve a predictive
completion project. The 67% completion is targeted to coincide model for 67% project completion. To preserve the acceptable
with those examples previously explained in the CAPP related prediction accuracy without suffering overfitting of data, a policy
literatures 共Russell et al. 共1996兲; Russell et al. 共1997兲; CII that reduces model complexity without deteriorating prediction

View publication stats Downloaded 14 Nov 2008 to [Link]. Redistribution subject to ASCE license or copyright; see [Link]