168
University of Calgary PRISM: University of Calgary's Digital Repository Graduate Studies Legacy Theses 1998 An artificial neural network approach to assess project cost and time risks at the front end of projects Liu, Xiaoying Liu, X. (1998). An artificial neural network approach to assess project cost and time risks at the front end of projects (Unpublished master's thesis). University of Calgary, Calgary, AB. doi:10.11575/PRISM/23290 http://hdl.handle.net/1880/42597 master thesis University of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission. Downloaded from PRISM: https://prism.ucalgary.ca

An artificial neural network approach to assess project

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: An artificial neural network approach to assess project

University of Calgary

PRISM: University of Calgary's Digital Repository

Graduate Studies Legacy Theses

1998

An artificial neural network approach to assess

project cost and time risks at the front end of

projects

Liu, Xiaoying

Liu, X. (1998). An artificial neural network approach to assess project cost and time risks at the

front end of projects (Unpublished master's thesis). University of Calgary, Calgary, AB.

doi:10.11575/PRISM/23290

http://hdl.handle.net/1880/42597

master thesis

University of Calgary graduate students retain copyright ownership and moral rights for their

thesis. You may use this material in any way that is permitted by the Copyright Act or through

licensing that has been assigned to the document. For uses that are not allowable under

copyright legislation or licensing, you are required to seek permission.

Downloaded from PRISM: https://prism.ucalgary.ca

Page 2: An artificial neural network approach to assess project
Page 3: An artificial neural network approach to assess project

THE UNIVERSITY OF CALGARY

An Artificial Neural Network Approach

to Assess Project Cost and Tirne Risks at The Front End of Projects

Xiaoying Liu

A THESIS

SUBMITTIED TO THE FACULTY OF GRADUATE STUDIES

IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE

DEGREE OF MASTER OF SCIENCE

DEPARTMENT OF CIVIL ENGINEERING

CALGARY,ALBERTA

APRIL, 1998

1998 O Xiaoying Liu

Page 4: An artificial neural network approach to assess project

National Cibrary of Canada

BiMiothèque nationale du Canada

Acquisitions and Acquisitions et Bibliographie Services seMces bibliographiques 385 Wetihgtm Street 395, ru8 Wellington OriawaON K 1 A W -WB ON K1A ON4 Canada canada

The author has granted a non- exciusive licence allowing the National Libmry of Caoada to reproduce, loan, distn'bute or sell copies of this thesis in rnicroform, paper or electronic formais.

The author retains ownefship of the copyright in this thesis. Neither the thesis nor substantial extracts fiom it may be printed or otherwise reproduced without the author's ~ermission.

L'auteur a accordé une licence non exclusive permettant à la Bibliothèque nationale du Canada de reproduire, prêter, distniuer ou vendre des copies de cette thèse sous la fonne de microfiche/nlm, de reproduction sur papier ou sur format électronique.

L'auteur conserve la propriété du droit d'auteur qui protège cette thèse. Ni la thèse ni des extraits substantiels de celle-ci ne doivent être imprimés ou autrement reproduits sans son

Page 5: An artificial neural network approach to assess project

The objectives of this research were to explore the applications of artificial neural

network technology in capital project risk analysis and to develop intelligent cornputer

models to predict project cost and Ume variations at the front end stages of projects. The

models were used to evaluate the potential effects of nsks and project decisions on

outcomes. The application of this technology will help decision makea improve the

effectiveness of the decision making process at the front-end stage of projects in the oil

and gas industry.

Results indicated that artificial neural networks have the capability to capture general

patterns by learning from samples of similar past projects. Artificial neural network

models were superior to multiple linear regression models in the prediction of project cost

and thne variations. Artificial neural networks with stepwise regression provide more

accurate estimations than stand alone neural networks. The research showed that

artificial neural network technology has promising potential in the domain of project risk

analysis.

The procedures for developing artincial neural network models to predict project cost and

time variances were proposed. This research provides valuable results and

recommendations for fiiture research work in this area.

Page 6: An artificial neural network approach to assess project

ACKNOWLEDGEMENT

This thesis was made possible through the contributions of many people in a variety of

ways. First, the author would like to thank al1 survey respondents for their cooperation.

As well, great thanks to Keith Pedwell, Wynne Chin, Ben Magnusson, Bob McTague,

Richard Balfour, Peter Bourque, Doug Rowan, Tony Goldsmith and Lome Berg for their

advice and suggestions.

With sincerity, the author would like to thank her s u p e ~ s o r , George lergeas, for his

ongoing patience and support during this entire process. Special thanks also to Francis

Hartman for his valuable suggestions, guidance with and sponsorship of this work.

The author wishes to extend her gratitude to Jackie Wilson, Daji Gong, Erin Williamson,

and Greg Skulmoski for helping to edit and review the thesis.

Page 7: An artificial neural network approach to assess project

To my father, a great father and teacher, and my mother

Page 8: An artificial neural network approach to assess project

TABLE OF CONTENTS

TABLE OF CONTENTS .................................................................................................. V I

......................................................................................... CHAPTER 1 INTRODUCTION 1

.............................................................................................................. BACKGROUND 1

.................................................................................... ARTIFICIAL NEURAL NETWORK 3

.................................................. RISK ANALYSIS AND ARTIFICIAL NEURAL NETWORK 4

P ~ C I P A L OBJECTIVES OF THE RESEARCH ................................................................... 5

.......................................................................................... RESEARCH METHODOLOGY 6

....................................................................................................................... PROCESS 9

PRINCIPAL ACHIEVEMENTS .......................................................................................... 9

................................................................................................. GUIDE TO THE THESIS 12

CEAPTER 2 REVIEW OF PROJECT RISK ANALYSIS AND TECEINIQUES ....... 14

Page 9: An artificial neural network approach to assess project

.......................... 2.3. CU- USAGE AND BENEFITS OF ~ S K ANALYSE M INDUSTRY 16

............................................... 2.4. PROJEC~ ESTIMATE AND CONTMGENCY ALLOWANCE 19

................................................................... 2.5. REVIEW OF BSK ANALYSIS TECHNIQUES 21

.................................................. 2.6. DATA REQUIRED FOR PERFORMMG RISK ANALYSE 33

........................................ . 3.1 INTRODUCTION TO ARTIFICIAL NEURAL NEWORK (ANN) 35

.................... ........................ 3.2. ARTIFICIAL NEURAL NETWORK TECHNOLOGY ......... 3 5

....................................... 3.2. 1 . Hktorical Background of Arttpcial Neural Network 35

...................................................................... 3.2.2. Artifcial Neural Network Model 3 7

3.3. COMPARISON BETWEEN ANN's AND OTHER TECHNIQUES ......................................... 42

.................................................. . R3.1. Amifcial Neural N e ~ k vs Expert System 4 2

....................................... 3.3.2. Neural Networks vs . Conventional Stamic AnaiysrS 46

.................................................... 3.4. ISSUES RELATED TO NEURAL NETWORK TRAMMG 47

3.5. APPLICATIONS OF ARTIFICML NEURAL NETWORK M BUSINESS AND NAGEME MENT.^^

3.6. POTENTIAL APPLICATIONS OF NEURAL NETWORK M PROECT WSK ANALYSIS ......... 51

................................................................................................. 4.0. LITERATURE SEARCH 53

4.1. INDUSTRY SURVEY ..................................................................................................... 54

....................................................................................................... 4.1.0. Introduction 5 4

Page 10: An artificial neural network approach to assess project

.................................................... Research Advkory Commilee in Indusfry 5 5

.................................................................. Ethics Approvol on Human Studies 55

...................................................................................... Expert Panel Interview 5 6

................................................... Design and Testing of Survey Questionnaire 5 7

.................................................................................................... Samp le Design 59

............................................................................ Administration of the Survey 6 1

......................................................................................... Raw Data Anabsis 6 2

................. 4.2. DEVELOPMENT OF ARTIFICIAL NEURAL NETWORK MODELS 62

.............................................................................. 4.2.1. Introduction to ANN Model 6 2

............................................ 4.2.1.1. Artificial Neural Network Learning Algorithm 63

............................................................................ 4.2.1.2. Layers and Neuron Nodes -69

....................................................................................... 4.2.1 . 3 . Transfer Functions -71

............................................... 4.2.2. Artricial Neural Network Training Procedures 74

........................................................................ 4.2.3. Issues Related to AMV Training 7.5

4.2.3.1. Leaming Rate, Momentum Constant and Training Tolerance ................... -75

4.2.3 .2 . Training and Testing Examples ............................................................. 7 7

4.2.4. Perjiormance Measures of Neural Network ModeLF ........................................ 78

............................................................... 4.2.5. Generalization and Cross- Validation 80

........................................ 4.3. NEURAL NETWORK COMPUTATION SOFTWARE PACKAGES 81

CHAPTER 5 STUDY RESULTS ~ ~ a ~ ~ ~ a ~ ~ a m ~ ~ m a a a ~ m ~ ~ ~ m ~ ~ ~ ~ ~ a ~ ~ m ~ o ~ m a a m m m m m ~ a a m m ~ ~ ~ ~ ~ ~ ~ ~ m ~ m m ~ o a ~ ~ a ~ m ~ a ~ e a ~ a ~ a m m a m m a ~ ~ ~ 8 3

Page 11: An artificial neural network approach to assess project

5.1. RAW DATA ANALYSIS ............................. ,., ................................................................ 83

........................................ 5 .2 . ARTIFICIAL NEURAL NETWORK TRAMMG AND RESULTS 8 5

.............................................................. 5.2. I . Formatring of Input Data in Training 87

........................................................................................ S . 2.2. Preliininary Training 8 7

. . ............................................................................................. 5.2.3. Phase I Training 8 9

..................................................... 5.2.3.1. Training and Results with 60% Samples -90

...................................................... 5.2.3.2. Training and Results with 75% Samples 92

..................................... 5.2.3.3. Training and Results with 90% Samples

...................................................... 5.2.4. Comparison arnung Three Group Samples 9 6

5.2. 5 . Defining Critical Input Variables Using Multiple Linear Regression ............. 97

..................................................................................... . 5.2.5.1 Forward Regression -99

5.2.5.2. Forward Regression with Dummy Variables ......................................... 100

..................................... 5.2.5.3. Forward Regression without Durnmy Variables 1 05

....................................... 5.2.6. Phase II Training - Using Critical Input Variables 107

5.3. PERFORMANCE COMPARISON BETWEEN NEURAL NETWORK AND MULTIPLE LMEAR~ 1 1

REGRESSION ANALYSIS .................................................................................................... 1 11

CHAPTER 6 CONCLUSIONS & RECOMMENDATIONS ................... .................... 1 12

Page 12: An artificial neural network approach to assess project

APPENDIX V, A SAMPLE OF A BATCH FILE FOR ANN TRAINING ..... , ......... 151

Page 13: An artificial neural network approach to assess project

CHAPTER 1 INTRODUCTION

1.1. Background

At the early stages of a project, decision making relies on experienced practitioners.

These early decisions are made mostly at the time when there are a great amount of

unceaainties involved. The quaiity of decisions cannot be venfied until the project is

completed. Artificial intelligence based decision support models are one option to help

practitioners screen alternative decisions and evaluate their impact on project outcornes in

ternis of project fuial cost and duration, thereby improving the quality of early decisions

and the effectiveness of decision making.

To test this concept, this thesis describes the development of some artificial intelligence

models using artificial neural networks to predict project cost and tirne variances.

Decisions at the fiont-end stage of a project are inherent with uncertainties and risks.

New technology, increasing complexity, political involvement, a changing economy and

regulatory environment, weatherfnatural conditions are some of the major sources of

uncertainty and risk. A contingency is usually allocated to allow for the uncertainties in a

project. Based on i n t e ~ e w s with practitioners, the researcher observed that the

conventional methods of contingency provision, for instance, the use of a flat 'percentage

Page 14: An artificial neural network approach to assess project

rate' and 'classes of estimate', are popularly used. Since the contingency allocation is

largely a matter of judgrnent and, therefore, arbitrary, estimators often Fud their estimates

difficult to justify or defend [Dey, Tabucanon and Ogunlana, 19941. The needs for

quantitative risk analysis to provide a rational bais for estimating and contingency

allocation have been addressed in the literature [Cooper and Chapman, 19871. Personal

interviews with practitioners in industry led to the conclusion that the need for better

techniques is increasing in industry as the project environment is becoming more

complex.

Conventionai risk analysis techniques, such as Monte Car10 analysis, provide tools to

help practitioners to assess impacts of uncertainties, to support the detemination or the

assessrnent of the risk level of a project, and to allocate a contingency associated with the

possibility of success. Unfortunately, the effectiveness of using this technique is heavily

dependent upon experts' opinions and judgments. These experts usually represent the

key funftions required to support a project, such as marketing, finance, estimating,

planning, engineering/design, procurement, construction and operation. To capture the

experts' knowledge and opinions is tirne consuming, laborious and complex.

hdeed some risk analysis techniques have other weaknesses, an important one being the

inability to quant@ correlations and interactions between risk factors and to mode1 a

complex system effectively.

Page 15: An artificial neural network approach to assess project

1.2. Artificial Neural Network

In cornparison to conventional risk analysis techniques, an Artificial Neural Network

(ANN) is an approach that is fiee of mathematical models. It requires less expert opinion

and judgments than do other techniques. Neural networks represent an attempt to

simulate the human brain's learning process through massive training. A neural network

is able to leam fiom samples. Knowledge learned is stored within the network. This

technology provides a powerful and robust means to assess uncertainty through learning

and capturing general patterns in available data.

Amficial neural network technology is used in many areas ranging from engineering to

business management, especially in finance and banking management. One of the most

significant feanires of a multiple-layer neural network is its generalization and

classification capabilities. These capabilities are utilized to develop predictive models in

real hancing and banking applications.

The application of neural network technology is rare in project cost and time risk

analysis. The one example found in the literatue is McKim's [1993] development of a

neural network model to predict project cost o v e m . The predictive error of the model

was less than half compared to using a conventional method - arithmetical mean.

McKim's study showed that the ne& network approach outperformed the conventional

Page 16: An artificial neural network approach to assess project

4

method. The initial research work in this area considered only four risk factors and only

twenty project samples were used to develop the ANN model.

Cross-validation (e.g., R2 in testing data which are unknown to the model) is an important

performance measure of predictive models for both traditionai method models and ANN

models. Unfortunately, it was not performed by the authors (McKim [1993], Denton

[1995], Sohl [1995]) in the development of their ANN models. This may Ieaves the

models developed both unreliable and invalidated.

This thesis attempts to consider more risk factors than other researchea in this area in the

development of ANN models to predict both cost variations and t h e variations. Cross-

validation is applied as a major performance measure in evduating the ANN models.

This thesis is intended to make some progress in this research area.

1.3. Risk Analysis and Artifical Neural Network

Project risk analysis basically involves the identification of uncertainties and the

assessrnent of the overali effects of these uncertainties. In this thesis, risk impacts on the

overall project cost and duration were assessed using artifîcial neural network technology.

Page 17: An artificial neural network approach to assess project

The sources causing project cost and t h e variations were identified by the author as

falling under two major categories. The h t category has to do with issues that relate to

the nature of the project, such as, the type of project, location, and complexity. n i e

second category identifies project decision related issues such as fast-tracking approach,

contract strategy, etc.

Al1 these attributes have different degrees of efEects on project cost and time relative to

the original estimates. In other words, these factors have certain correlations with project

cost and time variances. Artificiai neural network models are able to map the

relationships between these factors and the van-ances of cost and duration by 1e-g

extensively fiom real project samples and capturing general patterns within these

samples. Trained neural networks can therefore be used as intelligent predictive models.

7.4. Principal Objectives of the Research

The main objectives of the research are to explore artincial neural network applications in

project risk anaiysis, and to develop intelligent predictive models for project cost and

time variances at early stages in the development of projects. The models will form the

basis for the development of an estimation decision support system. The decision support

system wiil aiso help ownea evaluate impacts of rkks and decisions on project cost and

Page 18: An artificial neural network approach to assess project

6

tirne in the early stage of project, thus enabling improved evaluation of alternatives when

making critical formative decisions.

The secondary objective of the research was to identiQ major risk factors that have

significant impacts on project cost and time variations. This information was then to be

used to develop efficient and effective artificial neural network models.

1.5. Research Methodology

The stages of the research methodology followed were:

1. Research Proposal

An intensive search and review of the available literature was perforrned in the areas of

risk analysis and artificid neural network applications in project and business

management. Knowledge of the limitations of existing risk anaiysis techniques and the

unique advantages of artificial neural network technology was gained and the research

objectives were identified. (See Chapter 2 & 3)

Page 19: An artificial neural network approach to assess project

2. Data Collection and Suweys

Having identified the research objectives, the author carried out the following activities in

the second stage:

1. Established a industrial research cornmittee consisting of four people fiom the oil and

gas industry to direct the research, to help the researcher identify project critical risk

factors and to provide expertise and references. (See Chapter 4)

2. Applied for ethics approval on human studies. (See Chapter 4)

3. Conducted an exploratory study to identiQ major project risk factors by interviewhg

eight experts who had rich and extensive experience in nsk and project management.

Results fiorn the exploratory study were used to form a structured questionnaire. (See

Chapter 4)

4. Conducted a s w e y of thirty-one practitioners in project management to investigate

the current usage and benefits of using risk analysis techniques by industry. The

s w e y was completed by mail or through personal interviews wherever possible.

(See Chapter 2)

Page 20: An artificial neural network approach to assess project

8

5. Designed a structured questionnaire and tested it with colleagues and practitioners.

(See Chapter 4)

6. Carried out an industry survey through personal i n t e ~ e w and mailing within twenty-

four organizations including owner organizations and EPC consulting companies.

Data from one hundred and six completed projects were coilected. The types of

projects primarily included pipeline, refmery, and cornpressor station projects. Al1 of

the data collected referred to projects that were completed within the last 5 years and

had a value of over $1 million. (See Chapter 4)

7. Performed a descriptive analysis of the raw data gathered fiom the survey.

Information collected from the sample projects included the standard deviation and

average values of cost and tirne overnins. (See Chapter 5)

3. Development of ANN Models

ANN technology was applied to develop intelligent models using the survey data. A

three-layer neural network with back-propagation Ieaming algorithm was used. Transfer

fiinctions of the hidden and output layers, the learning rate, the number of hidden nodes,

and the training tolermce were determined through training processes for the effective

performance of the models. (See Chapter 4)

Page 21: An artificial neural network approach to assess project

9

The ANN mode1 development consisted of three phases. These are preliminary training,

phase 1 training and phase II training. In the preliminary training, a training d e and a

learning rate were selected; transfer fûnctions on hidden and output layers were initialiy

screened. In phase 1, the h?iasfer fiinctions were fmally determined. These transfer

functions allow ANN models to reach the best performance in generalization capability.

In phase II, ANN models were developed using cntical input variables and three groups

of project data. A forward regression technique was used to define the critical input

variables before moving to the phase II. The samples used in training and testing in these

three phases were randornly selected in order to eliminate bias. (See Chapter 5)

Multiple linear regression analysis (MLRA) was used to develop MLRA models for the

purpose of providing cornparisons to the ANN models. The results fiom the comparison

were then discussed. (See Chapter 5)

The research process is illustrated in Figure 1.1.

1.7. Principal Achievements

The main achievements of this research were:

Page 22: An artificial neural network approach to assess project

Identified eighteen critical risk factors of project cost and time variations.

Concluded that risk factors such as project type, location, complexity, design

completeness, nurnber of key organizations, and weather are some of the major causes

of project cost and time variations. These factors have a mutual impact on project

cost and time outcomes.

Demonstrated that ANN technology has promising potential as an application in

project risk analysis.

Shown that ANN technology, combined with stepwise regression, is superior to stand

alone ANN or stand alone multiple linear regression analysis.

Suggested development procedures for ANN models based on this study and the

results fiom the research.

Created intelligent predictive models as a basis for developing decision support

systems in the future. This will help project managers develop project contingency

plans, screen alternatives and options and make better decisions.

Page 23: An artificial neural network approach to assess project

Research Proposal

Perform Literature rn

Review h Establish Research Advisory Cornmittee

I Explore Significant Risk Factors 1

1 Design Survey Questionnaire 1

Test Questionnaire ?l

1 Analyze Raw Data 1

Prepare for Training 1 Develop Statistical Mode1 1 1 (Multiple Linear Regression) 1

Result Andysis r - l 1 Analyze Result 1

I Compare ANN and Statistical Model I

1 CONCLUSION (

Figure 1 .1 Research Process Flow Chart

Page 24: An artificial neural network approach to assess project

1.8. Guide fo the Thesr's

Chapter two presents a background on project risks, risk analysis and relevant

techniques. The investigation of current usage and benefits of nsk analysis in industry is

presented. Existing nsk analysis techniques are reviewed and their strengths and

weaknesses are discussed. Finally, the data required for performing risk analysis is

discussed.

Chapter three descnbes ANN technology and provides a histoncal background of its

development. An ANN model is reviewed. Cornparisons between ANN vs. expert

system and between ANN vs. statistic analysis are presented. Current applications of

ANN technology in business and management and potentid applications of this

technology are also discussed.

Chapter four presents the detailed research methodology . The research method

undertaken includes an industry survey to gather data fiom completed projects in

industry. The details of the ANN model and learning algorithm and relevant issues are

then presented.

Chapter W e presents the development of the ANN models in three stages: preliminary

phase, phase I and phase II. The objectives of each of the three phases are presented in

detail. Two standard (ANN and multiple linear regression) methods and one hybrid

Page 25: An artificial neural network approach to assess project

13

(ANN with multiple linear regression) approach are proposed and used to develop the

models to predict project cost and tirne variations. ANN models were aiso developed

using grouped data such as pipeline projects and refinery projects.

Chapter six provides conclusions to the research and recommendations for future

researchers and industry practitioners. Finally, fbrther research and contributions of this

study to the body of knowledge are discussed.

Page 26: An artificial neural network approach to assess project

CHAPTER 2 REVIEW OF PROJECT RlSK ANALYSIS AND TECHNIQUES

2.f. Risk

Several definitions of risk appear in the literature. Papageorge [1988] defines risk as "any

exposure to a loss or damage". Doherty [1985] defines risk as "the lack of predictability

of outcornes". Cooper and Chapman [1987] defme risk as "exposure to the possibility of

economic or financial loss or gain, physical damage or injury, or delay, as a consequence

of the uncertainty associated with pursuing a particular course of action." No matter the

diEerences in words, the fùndamentals are the same. For the purpose of this thesis, nsk

refea to "the volatility of the outcorne" and is measured by "the deviation fiom expected

values." It can be either positive or negative.

The significance of risk sources varies thmugh project life cycles. Sources of risk often

cited in the literature and identified by industry practitioners include environmental,

regulatory, political, financial, scope, engineering, technology, complexity, changes,

project management skills and experience, contract and weather. These are ofien

considered critical risk factors in the early phase of a project and contribute signincantly

to project cost and t h e variances.

Page 27: An artificial neural network approach to assess project

2.2. Risk Analysis

Cooper and Chapman [1987] state in their book, Risk Anuiysis for Large Projects, that

risk analysis cm involve a nurnber of approaches to dealing with the problems created by

uncertainty, including the identification, evaluation, control and management of risk.

Project risk analysis generally involves the assessment of the overall effects of these

factors, either implicitly or explicitly. This is frequently overlooked by owners who

evaluate nsks separately and qualitatively. The need for overall quantitative risk

assessment has increased with time largely due to global cornpetition, increases in

technological complexity, options and innovations, public involvement and regulatory

change.

Risk anaiysis may be appropriate in many circumstances throughout the life cycle of a

project, especially in the early stage in which uncertainty is significant and is thereby a

major factor. In the preliminary appraisal of a proposed project, a decision may have to

be made, often based on the owner's minimal amount of available information, to discard

the project, to postpone it, or to proceed with more detailed feasibility studies. For

example, a decision may be required to determine whether the project wilI be profitable.

This is determined by calculating the rate of r e m using the best estimates of capital

requirements and cash flows generated by the project If the resulting rate of return is

Page 28: An artificial neural network approach to assess project

16

equd to or greater than the opportunity cost of capital, or the net present value is greater

than zero, the project should be undertaken.

Overall, there are many formai requirements for nsk analysis: economic viability,

financial feasibility, insurance purposes, project managers' accountability, contrachial

purposes, regulatory purposes, and communication purposes.

2.3. Cunent Usage and Beneflts of Risk Analysis in lndustry

An industry survey was carried out in late 1996 by the researcher. The objectives of the

survey were to investigate the current usage and benefits of nsk anaiysis techniques and

to gather state-of-the-art information on risk analysis practices in the oil and gas industry,

mainly in the Calgary area.

A formai questionnaire (see Appendix 1) was used and sent to fifty-five individuals, of

whom thirty-two responded. The response rate was 58%. Al1 the responses excepting

one fiom Montreal were used in data analysis since the study was focused in the Calgary

area.

Twenty-six respondents (84%) were fiom the oil and gas industry sector. On average the

thirty-one respondents had 18.4 years of project management experience. These

Page 29: An artificial neural network approach to assess project

17

respondents' roles in project management included being project leaders, project

managers, project estimators, project engineers, project control managers and engineers

and project economists.

The survey results (see Appendix I for details) showed that thcre was a very strong

consensus of opinion (100% of respondents) that risk analysis must be used to assess risk

impacts on major projects (over $20 Million). Sixty-six percent of respondents stated

that risk anaiysis should be performed on projects with a value fiom $1 million to $20

million. Some of the respondents commented that conducting risk analysis was

dependent upon not only the size of a project but on the number of uncertainties

identified in a project.

The survey results dso showed that the selection of risk analysis technique was highly

dependent upon the nature of the problem. For instance, checklists and Monte Car10

simulation were the two techniques most fiequently used to assess project cost risks.

Checklists and CPMRERT were perceived as the two most useful techniques in the

assesment of risk in project scheduling.

Simister [1994] identified eight benefits of perfomiing risk analysis in his study.

Simister's eight items were incorporated in the m e y conducted by the researcher to

study the raak order of these items and to compare these results with Simister's hdings.

Page 30: An artificial neural network approach to assess project

18

Participants were asked to identiQ what they thought were the most beneficial reasons for

conducting risk analyses. The eight benefits were ranked fiom the s w e y in the

following order:

1. Gives an increased understanding of the risks in a project

2. Allows the assessment of contingencies that actually reflect the risks

3. Allows the formulation of more realistic plans in terms of cost estimates and

timescales

4. Facilitates greater, but more rational, risk taking, thus increasing the benefits that can

be gained fiom taking risks

5. Leads to the use of the most suitable form of procurement/contract

6. Builds up statistical information about historical nsks that help mode1 future project

risks

7. Helps distinguish between good luck/good management and bad luckhad

management

8. Identifies the party best able to handle risks

There was a consistency between Simister's research and this study. The fhst three items

were considered the most important reasons in both studies. However, the rank order in

his hdings was slightly different compared to this study. The rank order in Simister's

study was #3, #1, #2, #4, #8, #5, #6, and #7. His study was conducted in the UK and the

Page 31: An artificial neural network approach to assess project

19

industries he studied included defense, telecommunication, system-based information

technology and management consulting. Different countries and different industries may

cause the findings to be slightly different between the two studies.

In addition to the eight benefits listed above, participants to this s w e y identified the

following additional benefits to performing risk analysis of a project.

1. Assists in making the right decisions

2. Identifies golno go points and critical issues

3. 1s a part of the management of change process and assesses risk in order to set

priorities

4. Identifies critical activities that require risk management

5. Helps in the bidding process

6. Helps to determine the scope of projects during the conceptual phase

7. Assists in the selection of contractor(s)

2.4. Project Estimate and Contingency Allowance

Order of magnitude estimates are d y prepared in the early phases of a project when

project information is sparse and uncertainty is high. These estimates are often used to

assess the economic viability of projects or to compare dii3erent alternatives.

Page 32: An artificial neural network approach to assess project

20

Project estimating should always deal with a range of outcomes. To cover uncertainties

and risks, practitioners usually build contingencies into the estimates. A large

contingency may make a proposed project unattractive. A low contingency may produce

inadequate hding. Therefore, realistic estimates are important to ensure the accuracy

and usefulness of: (1) economic assessments and (2) project cost and schedule control

baselines. Thus, realistic estimates indirectly influence project overall success.

Single point estimates are still popular in industry. Unfortunately, single point estimates

are not realistic as they miss a key piece of information: the probability of success. To

quote an expected cost of a project at $100 million means potentially different things. It

could mean the likely cost is $100 million, or it may mean that the estimate represents

absolute certainty (100% probability of success) or it could be an absolute minimum

required budget (100% probability of being over budget). A solution to this 'hizy'

situation is to apply a scientific approach, such as Monte Carlo simulation, to calculate

the probability of success associated with the single point estirnate.

A scientific approach u d y involves assessing the project outcomes' probability density

function and accepting certain levels of nsk. This approach relates a probability to the

docation of a contingency and uses the engineea' estimate as the basis for this analysis.

Doherty [1985] suggests using a contingency-allocation mode1 that includes the

following six logical steps:

Page 33: An artificial neural network approach to assess project

1. Organizing and analyzing the estimating parameters;

2. Computing the estimator's base estimate;

3. Assessing the level of risk;

4. Assigning the classes of risk;

5. Establishing the even-chance estimate and the design contingency allowance;

6. Establishing the probability of success, the management contingency, and the cost

target.

The first two steps are based on traditional estimating methods. Risk analysis techniques

are brought in at the third step to assess overall project risk and variances. The

probability of achieving cost and duration targets associated with each contingency can

then be identifïed using the remaining three steps.

2.5. Review of Risk Analysis Techniques

The most difficult aspect of risk management is risk quantification. There are many risk

analysis techniques available to the project manager to assess project cost and schedule

risks. Some are qualitative methods (e.g., checklists), while others are quantitative (e.g.,

Monte Car10 simulations, regression). Some techniques can be performed either

qualitatively or both qualitatively and quantitatively (e.g., influence diagrams and

decision tree analyses).

Page 34: An artificial neural network approach to assess project

22

Different risk analysis techniques can be used in different phases of a project. For

example, qualitative techniques are generally employed in the early stages of a project

where few or no precise measures or numeric data are available. Some existing

techniques and their strengths and limitations will be discussed in the following

~ a r w w h s +

Checkiist based techniques are among the simplest and most commonly used in industry

either formally (pre-stnictured) or informally (non-stnictured). Checklists are lists of

items that can be questions or categories used to gather information from a group of

experts (e.g., a project management team) to determine the identity of potential nsks, the

probability of occurrences and the potentiai severity of financialhime losses. Based on

the information gathered, project managers know what significant nsks have a high

probability of occurrence or can trigger potentially severe losses and what the risk level

of the overall project is. A contingency plan can then be allocated to cover the project

nsks. Risk mitigation plans c m be made to avoid or reduce the probability and/or

potentiai severity of those losses from occurring.

The advantages of checklist based techniques are their simplicity and ease of use. The

process is easily understood by project teams who have domain expertise but not risk-

analysis training. The success of the process is heavily dependent upon experience

combined with intuition and personal perspectives to nsks.

Page 35: An artificial neural network approach to assess project

23

Monte Carlo simulation is a well-known quantitative tool. This form of analysis

typically starts with an established equation. For example, it is common to perform a

Monte Car10 simulation on a project cost estimate. The estimate is usuaily a

mathematical formulation of the following form:

N Estimated Total Cost = C Ci

i=l

where Ci is the cost of the i cost component of a project and N the total number of

individual cost elements. Every cost component with a potential for variabilit. is

modeled as a randorn variable. Others are treated as constant costs (those cost items

believed not to be expected to have any variations). Statistical distributions for each of

the random variables m u t be established before perfomiing Monte Car10 simulations. In

Monte Carlo simulations, randorn numbers are generated to produce values for random

variables based on pre-established distributions. The values produced and the constant

cost figure are added up to compute a value for the total cost of the project. This

procedure is repeated hundreds of times and a distribution is obtained for the total project

cost. The distribution can then be used to estirnate the project cost associated with a

probability of complethg the project on budget.

Monte Carlo simulation c m also be applied in the nsk aadysis of project schedules to

evaluate activities of uncertain durations identiflied within the Critical Path Method

Page 36: An artificial neural network approach to assess project

24

(CPM) schedule. User-specified distributions for each uncertain activity are essential to

obtain in order to perform the Monte Carlo simulation. The probability density

distribution of the total project duration is presented after sufficient iterations. The

critical activities are those that appeared on the critical path in the largest percentage of

iterations during the simulation.

The distribution function of each random variable is initially established and is a matter

of judgment. It is often complex and difficult to obtain real probability distributions that

accurately reflect underlying uncertainties.

Monte Carlo Andysis can provide project managea with a range of estirnates and a

probability for each outcorne. However, this technique has some weaknesses identified

by several authors [Touran and Wiser, 19921 [Pedwell, Liu and Hartman, 19961

[Diekmann, 1 9921 :

Inability to hande situations where no explicit mathematical model is appropnate.

Inability to model or quantify actual distributions and correlations between individual

components.

Inaccuracies created in circumstances where there are important variables which are

not included in the mathematical model.

Page 37: An artificial neural network approach to assess project

25

These disadvantages result in the limited use of Monte Carlo simulation in very complex

and unsûuctured situations. Care is recomrnended when using this simulation technique.

A study by Pike [1991] noted that although many managers are familiar with Monte

Carlo nsk analysis, cornparatively few use it.

PERT (Program Evaluation and Review Technique) is a well known but often

misunderstood approach to project scheduling. PERT was developed in 1958 for the US

Navy's Polaris missile/submarine project. PERT is performed in a similar fashion to

CPM (the Critical Path Method) which starts by establishing logic diagrams of projects.

Instead of using deterrnined durations for uncertain activities in CPM, optimistic

durations (a), most likely durations (b), and pessixnistic durations (c) of each activity are

used in PERT. The duration of each uncertain activity is calculated using the following

formula:

Therefore, project duration can be calculated and critical paths can then be identifïed.

This approach is an easy to use tool that can be used in project scheduiing involving

activities with uncertain durations. There are, however, two major shortcomings

associated with this approach. One is that it does not take into account the interactions

Page 38: An artificial neural network approach to assess project

26

between activities because it assumes the duration variation of each activity is

independent. The other is that the determination of the three durations (a, b, and c) of

each activity is heavily dependent upon experts' experience and personal judgments.

Annlytic Hierarchy Process (AHP) is a risk anaiysis technique that provides a flexible

and easily understood approach to andyzing project risi<. AHP is a multi-critena

decision-making method that allows objective or subjective factors to be considered in

project risk analysis. This approach initially formulates the hierarchical structure of the

decision problem. Then, the relative importance of each element at different levels has to

be determined by decision makers. The total risk likelihood is detennined by aggregating

the relative weights through the developed hierarchy.

The AHP approach was intmduced and applied to the construction risk assessrnent of the

Jarnuna Multipurpose Bridge by Mustafa and Al-Bahar [1991]. The AHP mode1 is

illustrated in Figure 2.1.

Page 39: An artificial neural network approach to assess project

GOAL

FACTORS

SUBFACTORS

LEVEL OF RISK

Constnrcting a Bridge Project

Total Risk TotaI Risk Total Risk

Figure 2.1. The AHP Risk Assessrnent Mode1

The approach provides valuable support for contractors in the decision making process,

but there are several weaknesses associated with this technique. The outcome of the AHP

is highly dependent upon the hierarchy established by decision malcers and the relative

judgments made about the various elements of the problem. The AHP approach aiso

lacks rigor. The effectiveness of AHP relies heavily on experts' personal intuition and

experience. This can lead to the inclusion of biases and manipulation, especially when

expert opinions conflict.

The Influence Diagram Technique is a method for representing the relationships of

decisions and uncertainties in a decision problem. The Decision Tree is a simple form of

infhence diagram. Basicaily, an infiuence diagram is constructed with a series of nodes

that are comected by lines (arcs). The nodes represent the uncertain variables in the

problem and the h e s represent the connections (influences) that exist between the

variables. By using influence diagramming, decision malcers seek to explain and

Page 40: An artificial neural network approach to assess project

28

understand the complex causal relationships between decisions and outcornes. Influence

relationships (i.e. risk dependencies) are usually quantified using conditional

probabilities.

An influence diagram must be solved by the propagation of the influence of input

variables after an influence diagram has been constmcted. An example of influence

diagrarnrning with conditional probabilities is presented in Figure 2.2.

Material Unavailability a

P(W 1 IA 1): Probability S=S I given that A=A I

Adverse Weather Schedule Delay

(9

S2: Schedule delayed

M 1 : Material unavailable

M2: Materiai available

AI A2 A3 A 1 : Bad Weather: temperature <-20°C

Figure 2.2. An Example of Muence Diagramming with Three Variables

A2: Moderate Weather: 1 SOC >temperattue >-20°C

A3: Good Weather: temperature> 15OC

S 1: Schedule not delayed

SI

S2

0.7

0.3

0.2

0.8

1.0

0.0

Page 41: An artificial neural network approach to assess project

29

The influence diagram technique is a common tool in the developrnent of expert systems.

The use of conditional probabilities to propagate influences is subject to one serious

shortcoming that depends on the topology of the influence diagram. It will become more

complex and subjective when a variable has more than one predecessor variable. The use

of influence diagram approach may be limited when the numbers of variables is so large

that the identification of dependent relationships between them becomes impossible.

Fuzy-Set Theory provides an unique approach to deal with decision problems described

with fupiness, vagueness and imprecision. Fuzzy sets are an attempt to capture the

richness of linguistic descriptions in a mathematical function Piekmann, 19921. This

approach is used

In risk analysis,

2.1).

to mesure qualitative subjects such as humanistic and societal factors.

risk interdependencies are measured using linguistic variables (Table

Table 2.1. Linguistic Variables for Determining Severity of Weather Risk

Linguistic Variables Definition

Low Less than 5% activities afTected by weather

Moderate 5% - 15% activities affected by weather

larger than 15% activities affected by weather

Page 42: An artificial neural network approach to assess project

30

F u z y set fiinctions are then used to descnbe the linguistic variables quantitatively. For

example, low, moderate and high are defined as the foilowing funy set functions with

degrees of membenhip chosen so as to produce rectangular fuzzy sets.

Low = [O1 1 .O, 1 ~0.9,2~0.8,3~0.7,4~0.6,5~0.5,6~0.4,7~0.3,80.2,90.1,10~0.0]

Moderates [0~0.0,1~0.2,2~0.4,3~0.6,4~0.8,5~ 1 .0,6~0.8,710.6,8~0.4,9~0.2, 1010.01

High = [0~0.0,1~0.1,2~0.2,3~0.3,4~0.4,5~,6~.6,7~0.7,8~0.8,90.9,10~ 1 .O]

Then, fuzzy sets are manipuiated using fuPy algebra to propagate impacts of initial nsk

conditions throughout the influence diagram. The results then are defupified and

converted to linguistic terms. Figure 2.3 illustrates the steps of how fuPy-set theory is

applied.

1 IdentiQ the problem using influence diagrams 1

1 Assign linguistic variables I r I

Convert linguistic variables to fuzzy sets

Defutzify results 77 Convert fuzy sets to linguistic variables

1 Solutions to the problem I Figure 2.3. Steps of Applying Fuzy Set Theory

Page 43: An artificial neural network approach to assess project

3 1

By using linguistic variables, fuzzy set theory can overcome the problem of complex

conditional probabilities used in influence diagrams. This modification allows experts to

express relationships in more natural, linguistic tems which are easily undentood.

Linguistic tems are subjective and need to be clearly defined and consistently understood

by experts.

Some authors [Wirba, Tah and Howes 19961 have demonstrated an application of fuay-

set theory in risk analysis in construction management. Linguistic variables through

funy sets were used to assess the likelihood of occurrence of each risk. The

effectiveness of the fuPy set approach to risk analysis was highly dependent upon the

d e f ~ t i o n s of risk dependencies and fiipv sets.

Regression AnaIysis is a generai statistical technique cornmonly used to solve important

research problems. Its uses range fiom the most general problems to the most specific.

This technique analyzes the relationships between a single dependent variable and severai

independent variables. This method relies on historic data and seeks to obtain patterns or

trends fkom these data. These trends are used to forecast future outcornes. An example

of a linear-multiple regression mode1 for a cost ovemm with two independent variables is

expressed as follows.

Actual Cost/Estimated Cost = a + bXi+cXi+dXj

Page 44: An artificial neural network approach to assess project

Where: a is a constant

b,c,d is the coefficient of variable XI, X2, and X1, respectively

XI is the value of project type

Xz is the value of project complexity

X3 is an interaction variable between Xi and Xz

Merrow and Yarossi [1990] applied multiple linear regression analysis in the

development of a project evaiuation system which included a database and many

mathematical models. The system involved the assessrnent of cost, schedule,

performance, safety and market factors which posed risks to project success. Models

were built based on historic project data.

Federle and Pigneri [1993] introduced and developed a predictive model of project cost

ovemuis for the Iowa Department of Transportation (IDOT). The mode1 was built to

predict the amount of cost under/overrun based on several cost factors such as project

location, project type, project duration, etc. Multiple linear regression analysis was

applied aud seventy-aine projects were used to develop the model. The model helped

DOT engineers predict the potential for cost o v e m . The limitations of the model

were that it assumed observations were independent and it neglected possible interactions

between the cost factors.

Page 45: An artificial neural network approach to assess project

33

Regression analysis is subject to a number of limitations. The use of historic data and

variables may not always be appropriate. The variable coefficients are obtained by

running the complete set of data. To incorporate new data, the complete set must be

reanalyzed. The mode1 mut also be specified in advance. If non-linear regression is

used, determinhg the exact nature of the non-linearity rnay be a burdensome task.

In summary, there are a numbers of risk assessment techniques in use today including

Monte Carlo simulation, analytic hierarchy process, fupy theory, etc. While some

positive results are available, they al1 suffer fiom one or more of the following

limitations:

assume uncertain variables independent

expert opinion and judgement

simplified models

simplified distributions of nsk variables

2.6. Data Required For Perfonning Risk Analysis

Numerou risk analysis techniques are available for the quantitative analysis of project

risk, but without quality data they are wortidess. Risk analysis software packages will

experience "garbage in - garbage out" if there is no quality data fed into them. Data may

Page 46: An artificial neural network approach to assess project

34

corne from a variety of sources, representing the expenence of the project team, the

organization and the outside world [Bowers 19941:

Corporate: knowledge gained in past projects is dispersed throughout the

organization. Information may be stored as personal mernories, diverse reports

and databases that compare project plans and outcomes. Any organization is

encouraged to set up information databases to store past projects' data in a

consistent manner that may be valuable for future projects.

Project tearn: specific project experience is possessed by individuais within

project tearns. Such knowledge is probably relevant, but quite possibly limited

and biased.

Extemai: usefbl data may be obtained fiom other relevant organizations and

data warehouses,

Both academics and industrial practitioners have paid more and more attention to the

applications of quantitative risk analyses. Being aware of the strengths and weaknesses

of various approaches of risk analysis is necessary. Experts' expenence, personal insight

and judgment are heavily utilwd in most of the risk analysis models discussed above. A

recommendation can be made that the development of continuously updated knowledge

databases (project information, personal insight and experience, and organi;rational

information) is essentiai for an effective risk analysis.

Page 47: An artificial neural network approach to assess project

CHAPTER 3 ARTlFlClAL NEURAL NETWORK TECHNOLOGY (ANN) AND APPLICATIONS

3.1. Introduction fo Aldificial Neurrl Nehvork (AIVN)

Artificial neural networks (ANN) are one of the fastest growing and most innovative

areas of computing. Neural networks represent an attempt to simulate hurnan thinking

processes through massively parallel, highly-interconnected processing systems.

Artificial neural network technology has been developing for several decades but has

ody found solid applications in the last decade. In recent years, ANN has been moving

fiom research laboratories into the business world. ANN applications are currently being

applied in many disciplines ranging from engineering to business management, especially

in fmancial and banking management.

3.2. Adficial Neurel Network Technology

3.2.1. Historical Background of Artificial Neural Network

Despite the curent attention of both researchers and practitioners, d c i a l neural

systems are not a new concept. According to Eberhart and Dobbins [1990], the history of

ANN can be divided into four periods of tirne: initial penod (1890-1969); depression

Page 48: An artificial neural network approach to assess project

36

period (1 969- 1982); recovery period (1 982- 1986); heightened interest period (1 986-

present).

In 1943, McCulloch and Pitts suggested that a network of simple binary neurons could

perform highly cornplex cornputations. The neuron mode1 could perform logical

processing but people did not understand how information was stored in the model or

how intelligent behaviours were learned.

In 1949, Donaia O. Hebb postulated that information was stored in the connections

between the neurons, and that "leaming" consisted of modiwng these connections and

altering the excitory and inhibitor effects of the various inputs. Today, the major leamhg

paradigms for ANN are based on modifications of Hebb's original concepts.

In 1958, Frank Rosenblatt made a major contribution to neural network research with the

development of the perceptron which was the first real artificiai neural network. The

perceptron provided a simple model permitting extensive mathematical analysis of neural

networks and was simulated on a digital computer iBM704.

In 1960, Bernard Widrow and Marcian Hoff developed the Widrow-Hoff algorithm

which improved the speed of leaming and the accuracy of results. The algorithm was

widely used in back-propagation and other signal processing systems.

Page 49: An artificial neural network approach to assess project

37

In 1969, MaMn Minsky and Seymour Pappert conducted an in-depth mathematical

anaiysis of the perceptron. Using simple examples, they showed that only a few

functions are guaranteed to be learned by the perceptron. In the case of the well-known

exclusive or (XOR) function, they showed that the function could not be learned by a

two-layer network. MaMn and Pappert's work had a devastating effect on neural

network research.

In 1982, John J. Hopfield introduced entirely connected networks called "Hopfield"

networks. He was one of the most important persons in the history of ANN development.

He re-triggered ANN researchers' interests after his several papers on the applications of

ANN were published. His work, and that of others, laid the foundation for significant

advances in neural network theory that has overcome the objections presented by Minsky

and Pappert.

3.2.2. Artificial Neural Network Mode1

An artificial neural network's paradigm allows it to mimic the fùnctions of a hurnan

brain. Information is distnbuted and processed in parallel within an ANN. ANNs

exhibit certain feahires such as the ability to lem, recognize trends, and simulate human

thinking processes. Today ANNs can be trained to solve problems that are difncult for

conventionai cornputers or even for h u . beings.

Page 50: An artificial neural network approach to assess project

38

An artificial neural network consists of many simple-processing elements called neurons.

Each neuron has a small amount of local memory and some elementary computing

power. Neurons are c o ~ e c t e d to each other. A neuron receives input fiom other neurons

on incoming pathways and has only one output which can be directed to many other

neurons. In Figure 3.1 a single neuron with r inputs is show. The individuai inputs Po),

weighted by elements W(1 j) of the matrix W, are summed to form the weighted inputs to

the transfer fùnction F. The neuron has a bias b and an output A.

Input Neuron

7 0 k

Figure 3.1. A Multiple Input Neuron Mode1

Neurons communicate only through these input/output pathways. A neuron cannot

access the memory of other neurons. Each n e m n has a transfer h c t i o n which is used to

compute the output fiom the inputs. Different neurons in a network can have different

transfer bctions .

A neural network typicaily consists of six primary components: neurons, connections

be~reen neurons, weights, transfer fiinctions, leaming d e s , and the overail

Page 51: An artificial neural network approach to assess project

transformation process. These components determine the artificial neural network

paradip. The details on these topics are discussed in Chapter 4.

There are two basic types of connections: feed-forward in which neurons in one layer are

connected only to those in higher layea; feed-back in which any neuron can be connected

to any other.

There are three possible learning methods: supervised, unsupervised, and reinforcement

learning. With supetvised leamllig, the desired output for a set of training set is provided

to the network; thus it lems by example. UnsupeMsed learning is conducted when there

is no desired output provided to the network. The network defines its own output set

based on relationships derived kom the training set and learns by self-organization.

Reinforcement leanillig is a hybnd method, the network is given a scalar evaluation

signal instead of behg told the desired output, and evaluation can be made intemiinently

instead of with every nainllig input set.

Many different artîficial neural network models have been proposed that address many

distinct and diverse problem domains [Wilson 19901. One such usefùl neural network

model, especiaily for business and management applications, is the multi-layer, feed-

forward model - backpropagation. The model typicaily has one input layer, several

hidden layers and one output layer. Neurons on one layer are comected only to those in

Page 52: An artificial neural network approach to assess project

40

higher layers and the output of one neuron on one layer becomes the input of the neurons

on the next higher layer. A neural network mode1 with three layers is shown in Figure

3.2.

1 Xr - - - - - Input Amy

Hidden Layers

Output Layer

Figure 3.2. Neural Network Mode1 with Three Layers

The back-propagation algorithm attempts to mlnimize the sum of the squares of errors at

the output layer during the training process. A training set is comprised of pairs of input

values and the desired output values (used for a s u p e ~ s e d learning process). A training

set is presented to the neural network and the activations are fed forward through the

network, resulting in output at the output layer. This output provided by the network is

compared to the desired output for the particular input set. Network weights are adjusted

such that the ciifference between the network output and the desired output is minimhed.

Adjustments due to the output error are propagated backward through the network,

starthg at the output layer and moving back toward the input layer. The procedure is

Page 53: An artificial neural network approach to assess project

41

repeated over the training set until the network converges and produces the desired

responses within certain error cnteria.

ANN can be represented mathematicdly in terms of vector and matrix algebraic systems,

with the input and output values described in tems of vectors, the topology in terms of a

comectivity matrix consisting of the weights of various connections, and the learning

d e as a differential equation relating the output-target value pair, and the weights.

Training the network consists of solving this differential equation to determine the next

set of weights to be used to generate an output. Successful ANN paradigrns allow the

network to converge on stable solutions.

The training of a neural network takes place in the following conditions: a significantiy

sized training set is available; a multi-layered network is initialized with random

interconnection weights. The size of the training set and the structure of the network are

important when training a neural network.

Page 54: An artificial neural network approach to assess project

3.3. Compdson between ANN's and Ofher Techniques

3 . 3 Artifhial Neural Network vs. Expert Systems

Neural networks offer a novel approach to the decision making problems where there is

no information available regarding the assumption of data distributions or relationships in

the categorization dilemrna. Thus, problem domains such as unstructured, fuzy, or

nonlinear are ones where neural networks may represent substantial information

processing improvement.

Expert systems represent one of the most important developments in information

technology. Based on Artificial Intelligence (Ai) concepts, expert systems attempt to

incorporate the reasoning process and knowledge of experts into cornputer prograrns.

Expert systems were first developed in the 1960s and became commercially available in

the early 1980s. Commercial applications can be found in many industries that include

aerospace, rnilitary, banking and financing, manufacturing, retail, personnel management,

marketing planning, etc.

In contrat to expert systems, neural networks represent an innovative technology that

uses parailel information processing. The parallel architecture of neural networks makes

them particularly adept at analyzing problems with many variables by simultaneously

considering numerous factors. The pattern recognîtion and prediction capabilities of

Page 55: An artificial neural network approach to assess project

43

neural networks make them suitable for a number of business and management

applications.

Expert systems and neural networks differ in a nurnber of ways (refer to Table 3.1). One

key ciifference is their foundation. Expert systems are based on logical sequential

processing, and neural networks use parallel processing to attempt to simulate

intelligence. In expert systems leaming usuaiiy takes place outside the system.

Knowledge is obtained outside and coded into the knowledge base. In contrast,

knowledge in neural networks is stored as the weights of the connections between the

neurons. The leaming process is interna1 to the networks and can be dynamic. The logic

for processing knowledge in the two systems is different. Expert systems use deduction,

and neural networks use induction.

Table 3.1. Features of Expert Systems and Neural Networks

Expert Systems Neural Networks

Macro scope Micro scope

Sequential processing

Learning takes place outside systems

Deductive

System buiit through knowledge extraction

Mathematical logic origin

Exact rnatching

People Oriented

Knowledge is explicit

Parallel Processing

Learning takes place within systems

Inductive

System B d t through training

Statisticai and stochastic ongin

Approximate matching

Data oriented

Knowledge is implicit

Page 56: An artificial neural network approach to assess project

44

The underlying theones of the two technologies differ. Expert systems are based on

mathematicai logic "IF-THEN", following an objectsriented approach. Neural networks

are statistical and stochastic in origin. In an expert system the user cm query the system

to determine why and how output was derived. In contrast, the knowledge base in a

newal network remains a "black-box" to users.

Table 3.2 shows the strengths and weaknesses of expert systems and neural networks.

Table 3.2. Strengths and Weaknesses of Expert Systems(ES) and Neural Networks(ANN)

ES ANN

P

Exphnation No Explmation Capaciîy Capacity

Many turnkey Most ANN m u t be shells are customized per available application

Strong user Weak user interface interface

Software well Hardware and developed software are in

development stages.

Easy to validate Difficult to validate

Requires an articulate expert to develop ES

Requires a significant number of exarnples

Average developrnent time is 12 to 18 months

Leaming is static and extemal

Data should be complete and error fiee

Knowledge engineering is difficult and time- consuming.

Development tirne is as littie as a few weeks or months

Leaming is dynamic and intemal

Data can be incomplete, error ridden and noisy.

Knowledge engineering is data driven and simple.

Page 57: An artificial neural network approach to assess project

45

Neural networks have many benefits in terms of knowledge acquisition compared to

expert systerns. In many unstmctured decision environments, particularly those that

involve classification, associative memory, or clustering, neural networks offer distinct

advantages over expert systems. One of the most significant strengths of neural neh~orks

is their fast computation and the possibility to convert a neural network to an electronic

chip using VLSI technology. Expert systems have some advantages that neural networks

do not have. They have user-fiiendly interfaces that provide explmations to the choices

made by the system and have an interactive, dynamic, and visual problem-solving

capability [Osyk and Vijayaraman 1 9951.

Several authoa have attempted the integration of expert systems and neural networks.

Expert systems can be employed to train neural networks, control information flow

through several neural networks, and anaiyze the responses provided by the neural

networks [Wong and Monaco 19951 . The greatest potential in AI technology may lie in

combining neural network hardware with expert system software [Duggai and Popovich

199249931. integration of fuzy logic and neural networks is another potential to deal

with unstmctured, fuzzy and large information problems [Bataheh 1 995- 1 9961.

Page 58: An artificial neural network approach to assess project

3.3.2. Neural Networks vr. Conventional Statistic Analysis

ANNs are distributed, parallel information-processing systems that exhibit certain

features such as the ability to learn, recognize trends, and capture patterns. ANN's ability

to build the relationships between input data and output data can be used to tackle

problems that have been conventionally handled by statistical methods such as regression

anal y sis, discriminant analy sis and cluster anal y sis. Cornparison of neural networks with

more traditional statistical techniques has been the focus of many recent studies. Some of

the main differences between regression and neural networks are addressed by authors

[Bansal, Kaufian and Weitz 19931. These are as following:

Neural networks consistently improve during training. Regression

techniques, on the other hand, process al1 training data simultaneously, before

using new data;

in theory, neural networks may be more robust than nonlinear classification

models;

Regression equations repuire mode1 specification in advance. In nonlinear

regression, specifjhg the exact nature of nonlinearity may be a burdensome

task. Neural networks avoid mode1 specification in the regression sense but

require specification of neural network architecture.

Page 59: An artificial neural network approach to assess project

47

Neural networks perform relatively well with missing or incomplete data, whereas

regression does not. Neural networks also typically have been shown to produce more

accurate predictions with good-quality data than regression models [Dutta and Hekhar

19881. De Groot and Wurts [1991] concluded that neural networks are the best

approach when data exhibit non-linear characteristics. However, neural networks are not

able to explain how they solve problems due to their "black-box" nature; no decision

d e s are generated for a decision maker's reference and inspection.

3.4. Issues Relafed fo Neural Network Taining

Numerous studies have unquestionably shown the utility of ANN models. However,

some issues must be addressed when training an ANN. These issues include:

The learning rate is a parameter that affects how neural connection weights are

updated. If the learning rate is large, it may cause rapid error correction in the

network, but it may potentially lead the aaining to non-convergent solutions. A

lower learning rate may be appropriate in allowing the nework to gently

converge, but it may cause the network to require more iterations through the

training set.

Training tolerance values represent the aliowable variation between the acnial

output set and the desued output of the training set.

Page 60: An artificial neural network approach to assess project

48

The stopping cnteria for training a neural network can be that the network

reaches either its traùillig tolerance or its maximum number of iterations.

Composition of the training set is an important consideration in neural network

training.

The size of the training set will have impacts on network training. It is intuitive

that the larger a training set, the more rich that set is in describing the problem

domain and might therefore result in a more accurate network than with a

smaller training set.

The parameterization of a neural network is an important issue. It has been

postdated that the configuration of the network may affect the accuracy and the

generalizability of the trained network. Network pararneterization includes the

number of input neurons, the number of output neurons, the number of hidden

layers and neurons, and the transfer function. The numbers of input and output

neurons are determined by the nature of the problem to be solved. The number

of hidden layers and neurons on each of them are established during the training

process for efficient performance.

There are few d e s or guidelines to denve a network configuration for a specific

problem. The appropriate number of hidden neurons and layers remallis an unsolved

research question.

Page 61: An artificial neural network approach to assess project

3.5. Applications of Arüficial Neural Network in Business and M8nagement

The field of neural networks has a five-decade history but has found solid applications

only in the 1st decade, and this field is still developing rapidly. Neural network

technology has been applied in a wide range of areas, such as image recognition, signal

processing, control systems, forecasting, text retrieval, and optimization.

In recent years, neural network technology has been moving fiom research laboratories

into the business world. A number of practical applications of neural networks have

emerged in a variety of business and management fùnctions fiom banking and finance to

quality control and natural resource exploration.

Sorne figures can give an indication of the intense investment of neural networks. In

1986 the U.S. govemment was advised to commit over $400 million for research over an

eight-year period. in 1988 the private sector invested an estimated $20 million in the

purchase of neural network systems. About 80 percent of the Fortune 500 companies

have an investment in neural networks [Johnston 199 11.

Although the fastest growing sector for neural network application development is

defense, in private industry, the hancial services industry was one of the earliest

adopters of neural network technology. As we know, many highly stnictured decision

problems c m be solved effectively by conventional digital computer systems. However,

Page 62: An artificial neural network approach to assess project

50

most top-level decision making problems faced by financial managers are highly

unstnictured, very complex, and not easily adapted to conventionai approaches of

cornputer-aided analysis and decision support. Neural networks offer distinct advantages

to help solve these problerns.

In recent years, academic researchers have explored the use of neurai network technology

for bankniptcy prediction [Koster, Sondak and Bourbia 1 990- 199 11, bank failure

predictions [Tarn and Kiang 19921, optimum markup estimation Noselhi, Hegazy and

Fazio 19911, cost engineering pck im 19931, portfolio selection [Suh and LaBarre 19951,

strategic management [Slicher, Vakalis and Singh 19951, and environmental engineering

[Basheer and Najjar, 19961. But more companies are exploring the commercial use of

neural network technology for tasks such as credit card h u d detection, signature

verification, and evaluating the financial audit process.

AVCO Financial Services, a division of Teatron, Inc., has successfidly implemented a

neural network for a n a l m g loan applications. Banc Tec, Inc. plans to introduce a neural

system to read handwritten numbers on checks, now a human chore. Adaptive System

Inc. has developed neural networks for assessing mortgage applications. Chase

Manhattan Bank, the second largest issuer of bank credit cards, announced the installation

of a neural network to detect credit card hud. Citicorp's Quotron Systems is using

neural network software c d e d Braincel to predict stock-index movements five minutes

Page 63: An artificial neural network approach to assess project

51

ahead of t h e . Neural Systems Inc. makes use of a supervised network to mimic the

recommendations of money managers on the optimal allocation of assets among Treasury

instruments. Ward Systems Group, hc . created an example showing how one might set

up an ANN application to predict stock market behavior. Companies such as Ford,

Morgan Stanley, AT&T, and Raytheon are already exploring their potential in

applications as diverse as faultfinding, equities dealing, and sonar detection systerns. In

1994, the Canadian Imperia1 Bank of Commerce (CIBC) replaced its index-based

Leading Indictors with a neural network-based system [Tal and Nazareth, 19951. These

systems are frequently used for tracking general economic direction and their

performance to date has been very encouraging.

3.6. Potential Applications of Neural Network in Pmject Risk Analysis

The early decision making system for project definition and selection is quite a complex

process due to risk and uncertainty. The system can be defhed as a nonlinear and

unsû-uctured probiem. Risks, project decisions and project outcornes are interrelated.

The relationships between process variables cannot be explicitly represented in a

mathematical model.

In this study, project decisions c m inciude, for instance, fast-tracking considerations and

contractuai strategies. Project risks can be, for example, weather and technology factors.

Page 64: An artificial neural network approach to assess project

Project outcornes refer to cost and t h e variances. These are correlated with each other

and compose a complex and uncertain system during the early stage of a project.

Project and risk anaiysis techniques are currently used in project cost and time estimating

and evaluation of risk impacts. The use of conventional risk analysis techniques is

usually based on well-defined and stnictured mathematical models. These techniques are

not suitable to solve complex and unstructured problems.

Mathematical fiee methods, such as aaificiai neural networks, provide a powerful and

potentiaily robust means to assess uncertainties through learning and capturing general

patterns in available data. ANN are usually used where problems cannot be readily

represented by explicit analytic models, and where the process of reasoning to obtain a

result may not be logical, and where signifiant data exists. ANN technology promises to

provide an innovative approach to project nsk analysis, especially in the early stage of a

project.

Page 65: An artificial neural network approach to assess project

CHAPTER 4 RESEARCH METHODOLOGY

4.0. Literatum Search

A broad study in the areas of project management, risk analysis and management,

information systems and technology, artificial neural network technology and its

applications was conducted. This began in early 1996 and continued throughout the

research period to ensure that new information, research ideas, and applications were

captured.

The literature search included books, technical and research reports, newsletten, theses

and dissertations, proceedings and journal articles. About thirty journals and proceedings

were reviewed for relevant information. Hundreds of related papers and articles were

reviewed and backed up for references and M e r research.

Sources for references included:

Mackimmie Library, the University of Calgary (stacks, CD-Rom)

The Scurfield Management Library, the University of Calgary

I The Internet

Relevant conferences

Page 66: An artificial neural network approach to assess project

The City of Calgary Main Library, Calgary

Foster Research Center, Calgary

Project Management Body of Knowledge, PMI Standards Cornmittee, PMI USA,

1996

Research Reports by the Construction Industry Institute, The University of Texas at

Austin

The review concentrated on sources fiom the 1990s and the late 1980s because ANN

applications in business and engineering only began in the 1980s. Reference lists in

journal articles were scanned and relevant sources were located for further review.

Relevant idormation was noted for closer review and copied for backup. Al1 backup

information and articles were sorted into foldee fustly by subject area and secondly by

publication year. The research interest and cornrnon themes in the each area can be

identified year by year in this system.

4 . lndusfry Survey

4.1.0. Introduction

A large amount of completed project information was needed to fullill the objectives of

the research. An industry survey was designed for this purpose. Although there were

Page 67: An artificial neural network approach to assess project

55

many industrial sectors of interest, the oil and gas industry was the focus of this study

because it is a major industry in the Calgary area. Tirne constraints on performing the

research were a secondary reason for limiting the scope to one local industry.

4.1.1. Research Advisory Cornmittee in Industry

M e r the research proposai was approved, a research advisory cornmittee comprising four

people was established to direct and guide the research and to provide support, especially

for the industry survey. Each cormnittee member had significant project management

experience and was a recognized leader in the oil and gas industry.

Each committee member committed his tirne, effort, and insight throughout the research.

Each member was updated with feedback in the fonn of research progress reports.

Suggestions from cornmittee members were considered and incorporated into the

research.

4.1.2. Ethics Approval on Human Studies

According to the univeaity policy regarding the ethics of human studies, conducting

research involving human subjects without formal ethics approvaVclearance breaches the

university's polîcy on Integity In Scholarly Activity. Prior to carrying out the research,

Page 68: An artificial neural network approach to assess project

approval from the Committee on

form and the research proposal

56

the EWcs of Human Studies was obtained. A consent

were attached to the application form for review by

cornmittee members. Certification of Institutionai Ethics Review was issued after the

cornmittee examined and approved the research proposai. The approved consent fom is

presented in Appendk II.

The consent form was sent to each potential survey participant with the survey

questionnaire. Each participant was asked to read, agree with, and then sign the consent

forrn and retuni it to the researcher.

4.13. Expert Panel Intewiew

An expert panel interview was conducted to gather information and identiQ indu-

practices related to project risk analysis and management in a relatively short time frame.

The objectives of the expert panel survey were to:

Ensure indusüy interest and support of this research;

0 Observe curent practices with respect to risk identification, risk analysis and

management;

Idente the critical factors that signiticantly affect project cost and time variances.

Page 69: An artificial neural network approach to assess project

57

Expert panel interviews were used to gather information that enhance and address the

gaps in knowledge in the research area between literature and professional practices.

An expert panel of eight people, who had rich and extensive experience in nsk and

project management, was selected from both academia and industry. These experts were

fiom owner organizations, enginee~g consulting companies, construction f m s and the

universities. An open-ended questionnaire was used. Critical factors that impact project

outcorne (cost and tirne) variances were identified by the expert panel and used to design

a draft survey questionnaire.

4.1.4. Design and Testing of Survey Questionnaire

A draft structured questionnaire form was designed by compiling the information and

knowledge fkoin the literature supplemented by the expert panel's knowledge, insight and

practices.

A bnef introduction of the purpose and objectives of the study was covered on the fk t

page of the questionnaire form. Confidentiality of survey data and approaches to the

management of the raw data were also addressed. In addition, instructions were provided

to potential participants on how to m e r the survey questions properly.

Page 70: An artificial neural network approach to assess project

The survey questionnaire was compnsed of the following four sections:

Project details. In this section, participants were asked to provide the estimated cost

and actual cos4 estimated and actual duration, and the year of completion, of the

surveyed project.

Project critical factors that impact time and cost variances. Critical factors for

projects were identified as project type, location, complexity, design completeness,

level of scope definition, project management experience, contract type, project

priorities, number of regdatory permits required, level of technological innovation,

weather, cost spent on the fiont end planning and cost vent on detailed

engineering/design.

Top ten signifcant factors injluencing cost and time variances for the speczpc

project. This section asked participants to identi@ the top ten factors from section 2

above that had significant impacts on the specific project.

Additional information. This section asked participants to provide additional

information not included in the survey questionnaire form that they felt was critical to

the project. Participants' information, such as current position, name, mailing

address, and their interest in receiving a brief copy of the study results, was dso

reauested so the researcher could provide feedback to s w e y participants.

Page 71: An artificial neural network approach to assess project

59

Language, phrashg and temiinology can pose sigaificant problems in a survey. Testing

is necessary to detect weaknesses in the questionnaire. Testing was undertaken through t

persona1 in te~ewing of colleagues, practitionen, some potentid s w e y participants, and

academics. The purpose of the testing was mainly to:

discover respondents' reactions to the questions;

check question interpretation;

use easily understood terminology;

check continuity and 80w.

A variety of valuable comrnents and suggestions fkom these tests were generated. Al1

comments and suggestions were taken into consideration and, where appropriate,

incorporated into the survey. The fmal version of the survey questionnaire is attached in

Appendix m.

4.1.5. Sample Design

The sample population was identifïed as oil and gas hdustry companies including owner

organi;rations and EPC consulting companies. The snowball sampling technique was

used for the following reasons:

Page 72: An artificial neural network approach to assess project

60

The survey required that participants not only have sufiicient knowledge and

experience in project management but that they were also directly involved in the

projects surveyed.

Respondents were difficdt to identifjc

The major concem was to gather information fiorn as many as possible completed

projects within a designated time frame.

The snowball sampling design has found a niche in recent years in applications where

respondents are difficult to identify and are best located through referral networks

[Cooper and Emory, 19951. In the initial stage of snowball sampling, individuals are

selected and then used to locate others who posses similar characteristics and who, in

turn, identifi othen.

The snowball technique allows the researcher to staa with a list of referrals and then

identiQ more potential participants fiom names provided by the previous respondents.

After each interview, conducted over the phone or in person, each participant was asked

to provide names of other potential participants that the researcher could contact and

suwey.

Page 73: An artificial neural network approach to assess project

4.1.6. Administration of the Survey

To emure a high response rate among participants, al1 surveys were sent out following

initial contact. Each potential survey participant was contacted individually by phone or

fax or email to get hislher commitment to participate in the research.

Surveys were sent out only to people who made a commitment to participate. Each

s w e y included a deadline that was approximately five months after the initiai contact. If

no response was received after one month, one or two follow-up telephone calls were

made to remind participants and to check whether they were having problems with the

sumey.

Many diEculties were encountered during the survey, even though most of participants

were willing and committed to participate in the swey. These difficulties were caused

by several reasons:

No hands-on data was readily available and participants needed to look back into the

specific completed project fiie(s);

Organhtion dowosizing;

Lose of key people;

People were too busy to be able to respond.

Page 74: An artificial neural network approach to assess project

62

Completed sweys were mailed or faxed back to the researcher. AU survey data were

coded and entered into a spreadsheet for raw data analysis. No third party was allowed to

access the data file.

4.1.7. Raw Data Analysis

A simple descriptive analysis was carried out for gaining general information on the

overall collected projects. The standard deviations and the average values of cost

overruns and t h e overruns were calculated. Collected projects were categorized to

define the distribution of the project type in the oil and gas industry. Results fiom and

discussion on the descriptive analysis are in Chapter 5.

4.2. DEVELOPMENT OF ARTIFICIAL NEURAL NETWORK MOOELS

4.2.1. Introduction to ANN Mode1

The development of neural network models is highly problem-dependent. Developing an

ANN mode1 requires the determination of an ANN paradigm. This includes ANN

learning algorithm, the numbea of layers and neurons, transfer bction, etc. An ideal

paradigm must optimally describe the nature of the study's system and perform well

Page 75: An artificial neural network approach to assess project

63

according to certain criteria. This results fkom a cornplex and the-consurning training

process. The following subsections address these related issues.

4.2.1.1. Artificial Neural Network Leaming Algorithm

The first step in the development of the ANN mode1 was to select a type of neural

network. In this study, a feed-forward network with the error back-propagation algorithm

was selected and used to develop predictive models of project cost and time risks. This

type of network wiih back-propagation algorithm provides a theoretically sound method

for the supervised training of the networks and requires a continuous transfer fùnction

wasserman and Schwartz 19871. The several reasons for choosing this type of learning

algo rithm included the following :

i It was able to train a multiple-Iayer network;

H It was capable of exploithg the regularities and exceptions in the training samples;

i It guaranteed stability of the network;

Figure 4.1 presents a sample of an ANN feed-forward network with three layers and ten

hidden nodes. Each of the nodes in the network is a neuron. Each of the links is called a

comection. As we can see, neurons in one layer are connected only to those in a higher

layer in the feed-forward type network.

Page 76: An artificial neural network approach to assess project

Cost Time Variance Variance (%) (%)

Output Layer

Hidden Layer

Input Layer

Figure 4.1. An ANN Mode1 With Three Layers and Ten Hidden Nodes

Back-propagation algorithm was based on the error-correction l e h g rule and the

neural network was trained in a supervised manner in order to map the relationships

between the input samples and the output sampies. Basically, the error back-propagation

process consisted of two passes through the different layers of the network: a forward

pass and a backward p a s .

1. Forward Pass

In the forward process (Figure 4.2), an input vector (Pi) was introduced to the input nodes

of the network. Each connection between the input layer and the hidden Iayer had an

Page 77: An artificial neural network approach to assess project

65

associated weight (IVji). The net signal (Ij) to an individual hidden node (j) was expressed

as the sum of al1 the connections between the input layer and that particular hidden node

plus a bias (Bj), as in equation (4.1).

Where R is the nurnber of nodes in the input layer.

The signal fiom the hidden layer was then processed with a tmnsfer function (F,) and the

output signal Oj was generated pnor to being sent to the output layer. The procedure was

performed according to equation (4.2).

Input Nodes

Hidden Node j

Oj Pi

Forward Pass

Figure 4.2. Signal-flow Graph: Two-Layer Network

Page 78: An artificial neural network approach to assess project

66

Each connection between the hidden layer and the output layer had an associated weight

) The net signal (Pd to the output node (k) was the sum of al1 the connections

between the hidden layer nodes and the output node plus a bias (Br&, expressed as:

where Q is the number of nodes in the hidden layer.

The net signal utk) was again transferred to the final output value (0'3 with a &ansfer

function (F3, expressed as:

Finally, a set of outputs (0') was produced as the actual response from the network.

During the forward pass the weights of the network were al1 fïxed. The initial weights

and biases in the network were usually randomly genenited.

2. Backward Pass

During the backward pass the weights were ail adjusted in accordance with the error-

correction rule. The weights were adjusted so as to make the acnial response of the

network move closer to the desired response. The backward process is shown in Figure

4.3.

Page 79: An artificial neural network approach to assess project

Bac kward Pass -1

Figure 4.3. Signal-flow Graph: Two-Layer Network

At the output layer, the net signal (0'3, which estimated outputs fiom the network, was

compared to the desired output (TJ to produce an error signal (E,) which was propagated

back through the network and expressed as:

Ek= Tk - O',

Between the output layer and the hidden layer, the correction A W M and applied to

Wkj and Brk, respectively, were defined as:

Page 80: An artificial neural network approach to assess project

where q is the rate of ieaming; and 6'k is cailed local gradient, defined as:

where F', is the derivative of the transfer function F,.

Between the hidden layer and the input layer, the correction AWji and ABj applied to Wji

and Bj, respectively, were defined as:

where 6j is defined as:

where F', is the derivative of the transfer fiuiction F,; S is the number of output nodes.

3. Iteration

The computation through forward and backward passes by presenting a set of training

examples to the network is called an iteration. The set of training examples was fed

Page 81: An artificial neural network approach to assess project

69

repeatedly to the network until the network converged to a stable solution and the sum of

the squared error (SSE) of the network computed over the entire training set was at a

minimum or acceptably small value, expressed as:

where N is the number of training examples; n is the nfh training example; s is

predefined error criterion, also called training tolerance.

4.2.1.2. Layers and Neuron Nodes

The second issue in the ANN modeling was to defme layea and the number of neuron

nodes on each layer. Three-layer neural networks were w d in this study. The first layer

was called the input layer which consisted of eighteen nodes. Each of the input nodes

represents an independent variable. The input variables are listed in Table 4.1 (See

Appendix IV for details). The M d layer was the output layer which consisted of two

nodes representing cost variances (%) and time variances (%).

Page 82: An artificial neural network approach to assess project

Table 4.1. Input Variables

1 I n ~ u f Variables 1 Label 1 Type of Project 1

Location of project

Complexity

No's of key organization

safeîy I v9

Design completeness

% of cost spent in front end

% of cost spent in detailed engi.

QditY

Input Variables 1 Label 1

V5 V6

V7 V8

Sc hedule I I

Years of experience I VI5 I

Cost

No.s of permitslapprovals

Scope defïned

Previous experience

Type of contract 1 VI6 1

VI1

VI2 V13 V14

m

Weather 1 V18 1 Level of technology

The number of hidden layers was determined through the training process for the

effective performance of the model. Typicaily, the more neurons in the hidden layer the

more powerful the network. The typical result of not using enough neurons in hidden

layers is that no set of weights and biases such that the network can produce outputs that

reasonably close to the targets are produced. This eEect is called underfitting. Too many

neurons in the hidden layer cause speciaiization effects that miss key points and cause

errors [Koster, Sondak, and Bourbia 1990-199 11. It may lead to fitting (or memorizing)

the training sets too weii resulting in poor generalization capabilities. To overcome these

problems, a range of hidden neurons was used in the training process.

V17

Hecht-Nielsen [1989] and Caudill Cl9911 recommend that any continuous fiinction cm be

Unplemented with a muiti-layer network using 2n+l hidden nodes, where n represents the

number of input nodes. Fletcher and Goss [1993] stated that in practice, the number of

Page 83: An artificial neural network approach to assess project

71

hidden nodes for optimal generaiization should be tested in a range fiom approximately

2*sqrt (n)+m to the value 2n+l, where m represents the number of output nodes and n

represents the nurnber of input nodes. In this study, the range of hidden nods was nom

2*sqrt (n)+m to 2n+l.

4.2.1.3. Transfer Functions

Each neuron had a transfer function that was used to compute the output signal fiom the

input signals. Three of the most commonly used transfer Functions for neurons being

trained with back-propagation are Linear, Tan-Sigmoid and Log-Sigrnoid.

1. Linear (Pureline)

The linear transfer functions shown in Figure 4.4.a and Figure 4.4.b have an output equal

to its input plus the bias. This kind of function is often used with neurons being trained

with the back-propagation.

Figure 4.4.a. Linear without Bias Figure 4.4.b. Linear with Bias

Page 84: An artificial neural network approach to assess project

2. Tan-Sigmoid (Tansig)

Figure 4.5.a and Figure 4.5.b show the tan-sigmoid fûnctions with and without bias. The

tan-sigmoid function is a differentiable hct ion and is used to map a neuron input fiom

the interval (-a, +a) into the interval (-1, +l), i.e., the output is dways between +1 to -1.

Figure 4.5.a. Tan-Sigmoid without Bias Figure 4.5.b. Tan-Sigrnoid with Bias

3. LogSigmoid (Logsig)

The log-sigrnoid tramfer hctions shown in Figure 4.6.a and Figure 4.6.b squash the

input (which may have any value between plus and minus infini@) into the range of O to

1.

Page 85: An artificial neural network approach to assess project

Figure 4.6.a. Log-Sigmoid without Bias Figure 4.6.b. Log-Sigmoid with Bias

The log-sigrnoid is a differentiable function that makes it suitable for neurons being

trained with back-propagation.

The transfer functions on the nght in each of the three cases shown above have a b i s ,

whereas those on the left do not. A bias can be a constant or allowed to change Iike the

weights with an appropriate learning d e .

The three transfer functions discussed above were used in the neuron network training

process. The possible combinations of these functions, such as using the linear fiinction

for ail nodes on the hidden layer and using logsigrnoid for al1 output nodes, were tested

in a training process in order to search out the best combination of those three transfer

fùnctions for this typical problem.

Page 86: An artificial neural network approach to assess project

4.2.2. Artifhial Neural Network Training Procedures

Training an ANN is the task required to arrive at a unique set of weights that are capable

of correctly associating al1 example patterns, used in leamllig, with their desired output

patterns. The algorithm cycles through the training data { [P(n), D(n)]; n = 1,2, . .., N] as

follows, where D(n) is the desired output vector corresponding to the input vector P(n).

1 . Initialization. Start with a reasonable network ~ o ~ g u r a t i o n , and randomly set al1

weights and biases.

2. Presentations of Training Examples. Present the network with a set of training

examples. For each example, perform the following sequence of forward and

backward computations described under step 3 and step 4, respectively.

3. Forward Computation. Let a training example in the epoch be denoted by [P(n),

D(n)], with the input vector P(n) applied to the input nodes and the desired output

vector D(n) presented to the output nodes. Compute the network outputs vector O'

according to equation (4. l), (4.2), (4.3), and (4.4).

4. Backward Computation Compute 6's of the network by preceding backwards, layer

by layer, according to equation (4.8) for output nodes, and equation (4.1 1) for

hidden nodes-

5. Iteration Present examples one by one in the training set to the network and repeat

step 3 and step 4. Repeat the training processes many thousands of times until a

certain preset criterion to stop the leaming session is met One such critenon is to

Page 87: An artificial neural network approach to assess project

75

consider the network to have adequately learned when the error between the output

produced by the network and desired output, accumulated in al1 leaming examples,

is less than a specified limit. The "stopping" point can also be determined by an

acceptable number of training iterations.

4.2.3. Issues Related to Am Training

4.2.3.1. Learning Rate, Momentum Constant and Training Tolerance

Multilayered nonlinear networks trained with back-propagation are sensitive to the

learning rate. The learning rate is a parameter that affects how neuron connection

weights are updated when back-propagating error toward the input layer. The greater the

learning rate, the more radical change an output error in the training process causes in the

connection weights. Thus, a large leaming rate rnay cause rapid error correction in the

network, but potentiaily the network solutions will keep jumping over the error minimum

without converging. A smaller learning rate may cause the network training to require

more iterations through the trainhg samples. So, a suitable leaming rate has to be

obtained by experimentally testing different Ieaming rates. In this study, a leaming rate

of 0.001 was gained after many experiments. It was believed that it was an appropriate

leankg rate to this specific problem and used through aIl of the training processes.

Page 88: An artificial neural network approach to assess project

76

Nonlinear networks introduce a complication because they may have several local

minima. Ideally, the network should fmd the global minimum. But this cannot always be

guaranteed. A network may get "stuck" by rolling into a local minimum [Demuth and

Beale, 19951. The authors suggest a technique called 'momentun' in ANN training to

overcome this problem and to minirnize the chance of getting b'stuck" in a local

minimum. In this study, the momentum constant is set to 0.95 after much experimental

training.

Training tolerance represents the allowable variation when the actual output neuron

values are compared to the desired output of the training case. A training tolerance of 0.0

requires an exact match of desired output to actual output that precludes connection

weight updates. A higher training tolerance will allow more variation in the output

values before errors are propagated back through the network. A smaller training

tolerance will provide more accurate results than one training with hi& tolerance, but at

the cost of more training iterations. A training tolerance used in this study is set to 0.001

f i e r much experimental training.

Page 89: An artificial neural network approach to assess project

42.32. Training and Testing Examples

The composition of training examples was a consideration in the training process. A

wide range of situations in the training set is thought to enhance the possibility of better-

trained networks.

The size of training set was another implementation issue. The larger the training set, the

richer that set becomes in describing the problem domain and the more likely the network

might be accurate.

A set of testing examples was needed to evaluate the network performance and define the

accuracy of the predictive capacity of the network.

Due to the limited number of completed projects collected, different sizes of training sets

and testing sets were used in the training processes. Based on the lhited samples, the

ANN mode1 was developed through intensive training processes.

Several situations that were considered during the training process included:

Training using different sized training data (randody selected), such as 60%, 75%

and 90% of total samples for training, and three Merent compositions of each sized

training data

i Training using oniy pipeline projects

Page 90: An artificial neural network approach to assess project

Training using only refinery projects

4.2.4. Performance Measures of Neural Network Models

The measures of network performance Vary with different applications in the literature.

The ratio of corrective cases in test samples or error ratios are a commody used

measurement used to classify types of neuron network models. Murtaza and Fisher

[ 1994 ] used this measure in a neuron network system for modular construction decision

making. Koster, Sondak and Bourbia [1990 - 199 11 used this measure in a neuron

network to predict banhptcy. Tarn and Kiang [1992] applied misclassification rates as

one of the performance measures for evaluating a network for predicthg bank failure.

The average of operation error, expressed in equation 4.13, is another measurement in

such neuron networks to map the relationships between inputs and outputs. Moselhi,

Hegazy, and Fazio [1991] used this measure in the neuron network for optimum markup

estimation under difFerent bid situations.

(ANN output - Deshed output)

Operation Error (%) = * 100 Desired output

Page 91: An artificial neural network approach to assess project

79

A variation of the cross-validation method known as v-fold cross-validation (CVp) is

another approach to rneasure the performance of an AMJ model. Fletcher and Goss

[1993] used this method to estimate prediction risk of their bankruptcy prediction model.

Three mesures we were concemed with in this study and monitored during the training

process were Surn Squared Errors (SSE), Standard Error of Estimate (SEE), and Squared

Coefficient of Determination (R2) of the predicted outputs over testing samples.

SSE is the mathematical criterion (4.12) for optimality over training sets. A lower

SSE does not necessarily irnply good generalization of the network. SSE indicates

how well neural networks fit training samples.

SEE is a rneasure of predictive error variation. SEE is equal to the square root of the

s u m of the squared residual variance (the dflerence between the predicted value and

desired value) and gives an indication of the average absolute error of prediction.

SEE is more directly interpretable than the residual variance.

R2 is a key performance meamernent through the whole training process. The larger

the R2 over the testing sarnples, the better the generalization capability of the network

model.

Page 92: An artificial neural network approach to assess project

4.2.5. Generaüzation and Cross-Validation

In back-propagation learning, we start with a training set and use the back-propagation

algorithm to compute the weights of a multilayer network by encodhg as many of the

training examples as possible into the network. The hope is that a network designed in

this manner will be able to generalize. A network is said to generalize well when the

inputsutput relationship computed by the network is correct (or nearly so) for

inputloutput (test data) never used in training the network. It is assurned that the test data

are drawn h m the same population used to generate the training data.

When a neural network Iearns too many specific input-output relations (i.e., it is

overtrained), the network may memorize the training data and therefore be less able to

generalize between sllnilar input-output patterns.

To train networks in the sense that they learn enough about the p s t to generalize to the

fbture, a leaming process arnounts to a choice of network configurations for the data set.

More specifically, the learning process may be viewed as choosing, within a set of

candidate mode1 structures (corQurations), the "best" one accordhg to a certain

criterion.

A standard tool in statistics, known as cross-validation, provides an appealing guiding

principle. As usuai, the available data set is randomly partitioned into a training set and a

Page 93: An artificial neural network approach to assess project

81

test set. The training set is used for training a network and the generalization

performance of the network is rneasured on the test set.

4.3. Neural Network Computation SoîYwanr Packages

Since training and tuning an ANN is a computationally intense process, large hi&-speed

cornputer systems are required for the simulations. Mini-workstation cornputer systems

may be suitable for this kind of work. A UNIX system with a speed of about 1 million

cycles per second (36 MHZ) was used to carry out the ANN training job in this study.

There are many commercial software packages available for neurocomputing, ranging in

pnce corn $50 to several thousand dollars and designed for the M X , Macintosh, and

IBM systems. One of those is cailed "Matlab" which is developed by MathWorks Inc.

Matlab typically runs on LTNIX systems and IBM-PCs. The version used in this research

was the 'Matlab' for a U N E system. Braincel is a neurocomputing software package

that works within Microsoft Excel spreadsheets. Other commercial software packages

available for ANN computing include NeuralWorks Professionai II software, developed

by Neural Ware Inc., and NeuroS hel14.1.

Although Matlab performs neuron computing pelfectly for many dserent problems,

customized programming is necessary to configure a neuron network, to initial weights

Page 94: An artificial neural network approach to assess project

82

and bias, to input training and testing samples, and to code a batch file for a training

process. A coded batch file (tralliingl .m) is presented as an exarnple in Appendix V.

Page 95: An artificial neural network approach to assess project

CHAPTER 5 STUDY RESULTS

5.1. Raw Data Analysis

Data on a total of a hundred and six projects were gathered. The researcher contacted

forty-eight participants to collect these data. Eighteen of the participants failed to return a

completed survey. The responding participants represent about twenty-four

organizations. Fifteen of the participants provided more than one project. The r e m rate

of the survey was 62.5% (30148).

Three projects were excluded fiom the raw data pool because of incomplete answers.

One hundred and three projects were used for data analysis and mode1 development.

Collected projects are categorized into five groups. The distribution of project type in the

oil and gas industry is shown in Table 5.1.

Table 5.1. The Frequency of Project Type

Projects Gathered Pipeline

Refinery

Gas Plant Others* Totals

Frequency 1

36

20 15 32 103

Note: * Others Uiclude cornpressor stations, drillhg wells, etc.

Page 96: An artificial neural network approach to assess project

84

A simple descriptive statistic analysis was perfonned. Average score, maximum score,

minimum score, and standard deviation of each of the relative questions were calculated

and the results are listed in Table 5.2.

Table 5.2. Descriptive Analysis Results

---

SURVEY RAW DATA ANALYSIS

1 Outcorne Variances 1 Am.Score Max Min ! S.D. 1

1

-1.1) Rnjecî Estimate Capital Cost (MM) 1 64.95 4600.00 0.53 451.64 lw2) &$kt Acnial Capital Cost 0 1 76.48 5800.00 0.62 569.79

(2) Contingency in capital cost (%) . (3.1) Roject Estimate Duration (Months) (3.21 h iec t Actual Duration CMbnths)

18.2) Numbers of key o ~ o n s involved 1 1 1.71 1'77.00 2.00 24.40

6.34 25.00 0.00 5.33 13.7'7 72.00 200 8.12 -

13.96 84.00 200 8.95

(4) Contingency in scheduie(%)

ûutcome Variance Indicatm (8.1) Complexity

2.75 12.00 0.00 4.22

2.10 3 .00 1 .O0 0.43

ka~ital cost in hn t end as % of estimate caoital cost 1 1.60 2 1.28 0.00 2.95

(9) Piercerrtage of design cwnpleteness ( IO) Capital cost in hnt end (MM)

1 1) Capital cost in Detail k&o 8.68 750.00 0.04 74.51 apital cost in Ml Design as % of estimate capital cost 8.08 32.00 0.55 6.44

84.49 100.00 20.00 22.07 - 0.69 50.00 0.00 5.00

.8) weather &kt (% activity affectedl 1 14.06 - 95.00 , 0.00 17.55

' r

Page 97: An artificial neural network approach to assess project

85

As can be seen nom Table 5.2 there was a wide range of cost variances fiom 57% to

180% (percentage of actuai cost to estirnated cost) and a large range of t h e variances

fiom 83% to 173% (percentage of actual duration to estimated duration) in this study.

The project with a 180% cost ovemm was an extreme case and had a very high risk

profile. The project complexity was hi& and new technology was involved. There were

ten key organizations that directly participated in this project and they were

unaccustomed to working together. The percentage of engineering design completeness

at the start of field construction was 60% and the project had poor scope defuiition. Al1

these factors, together with the project changes that occurred, resulted in the significant

cost overrun.

The average cost variance and tirne variance in these studied projects was 102%. On

average, project cost and t h e deviations from the estimates (Approval For Expenditure)

were not significant (+ 2%). The number of projects with cost overruns was forty-six

which present 45% of the total studied projects. The number of projects with tirne

ovemins was seventeen, 17% of the total projects.

5.2. Artificial Neural Nefwork Training and Results

In the developrnent of ANN models, three phases of A N ' training were carried out The

three phases were prelirninary training, phase 1 training and phase II training. Each phase

Page 98: An artificial neural network approach to assess project

86

of îmining had its specific objectives. The objectives of preliminary training were to

select a training d e , to detemiine a leaming rate, and a training tolerance, and to screen

transfer functions. The objectives of the phase 1 training were to detemine the transfer

functions which ailow networks to have the best performance in terms of high

generalization capability and to determine what sarnple size allows networks to perform

the best. The purpose of the phase II training was to develop ANN models using critical

input variables and grouped project data. The whole process of the ANN training in this

study is presented in Figure 5.1.

1 Format input data 1

Preliminary Phase Determine training rate, tolerance

I and screen transfer functions t

I -

Phase 1 Vary sarnple size, the number of

hidden nodes and transfer firnctions r Phase II

90% sample Pipeline projects Refinery projects

Figure 5.1. The Process of ANN Training

Page 99: An artificial neural network approach to assess project

5.2.1. Fomattîng of Input Data in Training

To make the data comparable, cost variances and t h e variances were normalized. The

percentage of actual cost to estimated cost and the percentage of actual duration to

estirnated duration were used in the model development, instead of using the absolute

actual cost and the actual duration. For exarnple, the estirnated cost of a project was

652MM. The actuai cost of the project was SSOMM. The percentage of actual cost to the

estimated cost (96%) was used in model development.

According to other researchen' experience in ANN training, the magnitudes of each input

data should be as close as possible. In this way, ANN models are easier to train.

Consequently, some input variables, such as the number of key organizations and the

number of permits/approvals, were grouped into five categories. Appendix IV presents

the format of each input variable.

5.2.2. Preliminary Training

The purpose of a prelimlliary experiment was to screen bander fiinctions for fbrther

experimental training. Appropriate trander functions on hidden and output layers should

generate the best performance of the models. As mentioned in Chapter 4, there are three

types of transfer hctions commonly used in back-propagation leaming networks. Any

Page 100: An artificial neural network approach to assess project

88

possible combinations of hem were considered. Clearly, there are nine possible

combinations; shown in Figure 5.2. These nine combinations were used in the

prelimuiary training.

Hidden Layer Output Layer

Figure 5.2. Nine of Combinations of Transfer Functions

The resuits fiom the preliminary training concluded that neuron networks that had an

output layer with Logsig and Tansig transfer functions result in no Iearning occurred.

They ody produced outputs of "1". Networks with Pureline output neurons were capable

of leaming in this shidy. Therefore, three cornbinations of transfer functions were used

for m e r training. These are listed below:

Hidden Layer Output Layer Lable

Logsig - Purehe (L ogs ig-Ehre 1 ine)

Tansig - Pureline (Tansig-Pureline)

Pureline - Purehne (Purdine-Pureline)

Figure 5.3. Three Combinations of Transfer Functions

Page 101: An artificial neural network approach to assess project

5-23. Phase 1 Training

Further to preliminary training, the first phase of ANN training with varied sample sizes

was carried out in order to select the best combination among the three combinations in

Figure 5.3. The criteria for selection was determined by overall network performance as

determined by the transfer functions of the hidden and output layers. In other words, the

networks generating the largest R* with the testing samples among the three cases were

selected.

Three groups of samples with sizes of 60%, 75%, and 90% of the total usable collected

projects were also considered. These were later used in the second phase to determine

what sample size used in training provides the largest RZ. The number of samples used in

the training were 62, 77 and 93; the number of samples used for testing the performance

of networks were 4 1,26, and 10, respectively.

With each group of training samples, three versions of samples were randody generated.

Therefore, there were nine subsets of samples used in the second round training. The

averages of performance panuneters of three versions of samples were calculated for each

group of sized samples.

Page 102: An artificial neural network approach to assess project

533.1. Training and Results with 60% Samples

Three combinations of transfer bct ions were empirically tested through intensive

training. Sixty-two samples were used in training and forty-one for testing the

performance of the model. Various number of hidden nodes (10, 15, ..., 40) were also

used in the training process. Ten thousand was used as the maximum number of

iterations in each training cycle.

Sum Squared Errors (SSE), Standard Deviation (SD), and Coefficient of Determination

(R2) in training and testing were calculated and recorded. The results are shown in Table

5.3, Table 5.4, and Table 5.5.

Table 5.3. SSE (Sum Squared Errors) in Training with 60% Sized Samples

Table 5.3 iilustrates that networks with a Logsig hidden layer and a Pureline output layer

have the srnailest SSEs, regardless of the number of hidden nodes used.

Tansig-Pureline 1 .O4 0.74 0.49

, 0.58 0.6 0.2 1 0.25

Logsig-Pureline 0.52 0.22 0.05 0.13 0.17 0.07 0.03

Hidden Nodes 10 15 20 25 30 35 40

Pureline-Pureline 1.82 1.6 1.91 1.97 2.14 1.53 1.54

Page 103: An artificial neural network approach to assess project

The standard deviations of cost variance and t h e variance with the training set and the

testing set were calculated with varying numben of hidden nodes. The averages of SD of

cost variances and time variances were used in the comparison of three combinations of

transfer bctions, presented in Table 5.4.

Table 5.4. SD of Outputs in Training with 60% Sized Samples

Pureline-Pureline Logsig-Ptueline Tansig-Pureline [ Hidden 1 Training ( Testing ( Training 1 Testing 1 Training ( Testing (

Table 5.4 shows that networks with a Logsig hidden layer and Pureline output layer have

the srnaIlest SD over the training samples and have medium values of SD over the testing

samples, regardless of the number of hidden nodes used.

Nodes 10

R2 of COS^ variance and time variance with the training set and the testing set was

calculated with various numbers of hidden nodes. The averages of R2 of COS^ variance

and t h e variance were used in the cornparison of three combinations of transfer

functions, presented in Table 5.5.

Set 0.122

Set 0.163

Set 0.063

Set O. 167

Set 0.090

Set O. 185

Page 104: An artificial neural network approach to assess project

Table 5.5 R2 of Outputs in Training with 60% Sized Samples

Pureline-Pureline Logsig-Pureline 1 Hidden 1 Training 1 Testing ) Training 1 Testing

Nodes 1 Set 10 0.49

Table 5.5 illustrates that networks with a Logsig hidden layer and Pureline output layer

Tansig-Purelhe

have the highest R2 over the training samples, regardless of the number of hidden nodes

Training Set 0.72 0.80 0.87 0.86 0.84

Set O. 1 O

used.

Testing Set 0.04 0.08 O. 15 0.09 0.03

5.2.3.2. Training and Results with 75% Samples

Set 0.85

Seventy-seven samples, representing 75% of total sample of projects, were used in

Set 0 .O6

training and twenty-six samples were used for testing the performance of networks.

Various numbers of hidden nodes (10, 15, 20, ..., 40) were also used in the training

process.

Sum Squared Errors (SSE), Standard Deviation (SD), and Coefficient of Determination

(R2) in training and testing were calcuiated and recorded. The r e d t s are shown in Table

5.6, Table 5.7, and Table 5.8.

Page 105: An artificial neural network approach to assess project

Table 5.6, Table 5.7 and Table 5.8 show that results in the 75% case are sirnilar to those

in the 60% case. Networks with a Logsig hidden layer and a Pureline output layer have

the smallest SSEs and SD over the training samples and the highest R2 over the training

samples.

Table 5.6. SSE ( S m Squared Erroa) in Training with 75% Sized Samples

Table 5.7. SD of Outputs in Training with 75% Sized Samples

Tansig-Purelime 1.908 0.91 1 1.678 0.839 0.554 0.537

0.250

Logsig-Pureline 0.678 0.328 O. 123 0.399 0.084 0.036

0.183

Hidden 10 15 20 25 30

35 40

Pureiine-Pureline Log sig-hireline Tansig-Pureline

Pureline-Pureline 2.335 2.6 19 2,284 2.582 2.304 2.389

, 2.433

Testing Set

0.276

Training Set

0-1 10

Testing Set

0.232

Training Set

0.060

Testing Set

0.151

Hidden Nodes

10

Training Set

0.124

Page 106: An artificial neural network approach to assess project

Table 5.8. R2 of Outputs Ui Training with 75% Sized Samples

5.2.3.3. Training and Results with 90% Samples

Pureline-Fureiine Logsig-Pureline Tansig-Purelhe

Ninety-three samples were used in training and ten samples were used for testing the

performance of networks. Various numbers of hidden nodes were also used in the

training process.

Sum Squared Errors (SSE), Standard Deviation (SD), and Coefficient of Determination

(R2) in training and testing were caiculated and recorded. The results are illustrated in

Table 5.9, Table 5.10, and Table 5.1 1.

Table 5.9, Table 5.10, and 5.1 1 illustrate that networks with a Logsig hidden layer and a

Pureline output layer have the srnaIlest SSE and SD over the training set and have the

highest R2 over both the trainhg set and the testing samples. We also can see that in

average of seven cases of hidden nodes, the R2 over the testing set increase and the R2

Testing Set 0.0 1 0.05 0.08 0.17 0.03 0.15 O. 16

Training Set 0.52 0.79 0.60 0.8 1 0.84 0.88 0.95

Testing Set 0.13 0.04 0.07 0.03 0.02 0.12 0.06

Training Set 0.86 0.9 1 0.97 0.9 1 O .99 O .99

Testing Set 0.05 0.07 0.06 0.09 0.05 O .O4

Hidden . Nodes

10 I S 20 25 30 35

Training Set 0.39 0.37 0.42 0.37 0.4 1 0.41

0.07 1 0.96 40 0.39

Page 107: An artificial neural network approach to assess project

over the training set decrease as the size of training samples increases (see Table 5.5,

Table 5.8, and Table 5.1 1).

Table 5.9. SSE (Surn Squared Errors) in Training with 90% Sized Samples

Table 5.10. SD of Outputs in Training with 90% Sized Samples

Tansig-Pureline 2.545

Logsig-Pureline 1.105

Hidden Nodes 10

Pureline-Pureline Logsig-Pureline Tansig-Pureline

Purelhe-Pureline 2.706

Testing Set

0.16 1 0.160

Testing Set

O. 174

0.167

Hidden Nodes

10

IS

Training Set

0.1 16 0.093

Testing Set

0.095 0.099

Training Set

O. 120

0.133

Training Set

0.074 0.050

Page 108: An artificial neural network approach to assess project

Table 5.1 1. R2 of Outputs in Training with 90% Sized Samples

5.2.4. Cornparison among Three Group Samples

Pureline-Pürebe Logsig-Pureline Tansig-Pureline

In section 5.2.3, the results fiom ANN training with three group samples and three

combinations of ûansfer functions are demonstrated. It is concluded From this

information that networks with a Logsig hidden layer and Pureline output layer

performed the best among the three combinations of transfer function. Networks with

twenty-five hidden nodes had the largest R2 over the testing samples. So, the

combination of Logsig-Pureline transfer function with twenty-five hidden nodes was

selected for fûrther comparison among the three shed samples.

R2 of cost and time variances over testing samples presents the generalization capacity

(cross-validation) of the predictive models. The larger the R2 over testing samples, the

better the network model. In the context of the rest of this chapter, SD and R2 were iisted

Training Set 0.45 0.64

0.69

0.82

0.73 0.84

0.9 1

Testing Set 0.2 1 0.14

I

O .O9 O. 14 O. 10

0.1 1

0.18

Training Set 0.73

0.88 0.86

0.92

0.97 0.94

0.99

Testing Set 0.25

0.20

0.3 5

0.27 0.24

0.22

0.30

Hidden Nodes

t O

15

20

25

30

35

40

Testing Set 0.14

0.17

0.25 0.38 0.25

0.32

0.30

Training Set 0.40

0.36 0.38

0.33

0.36 0.35

0.32

Page 109: An artificial neural network approach to assess project

97

only with testing samples. Table 5.12 presents the sized samples verses SD and R2 of the

testing samples.

Table 5.12. SD and R2 of Outputs VS. Sized Samples

1 Sued Sam~les 1 Standard Devirtion 1 R* 1

It was obvious that 90% sized sarnple had the smallest SD and the largest R2. A

conclusion can be made that the more samples used in training, the better the

performance of neural networks. There is no maximum limit of training set size. It is

important to collect a minimum amount of samples for training in order to reach an

acceptable threshold of prediction accuracy. Recent research [Bode, 19981 showed that

the marginal contribution of additional training data decreases with growing ûaining set

sizes.

5.2.5. Defining Critical Input Variables Using Multiple Linear Regression

Eighteen predictors treated as independent input variables were used in the fist two

phases of ANN training. It was established that these predictors were highly correlated

rather than independent. For example, project type will influence the location of the

project. Pipeline projects are mostly located in remote/suburban areas. Observations

Page 110: An artificial neural network approach to assess project

98

from personal interviews with practitioners in industry indicated that fast-tracking

approaches are rarely applied on projects with brand new technologies. It would be too

risky to do so. There was another learning from the survey that to some degree, design

completeness of a project is related to how familiar the project team is with the

technologies used and their experience with similar projects.

From an practical point of view, the fewer the number of subsets of available good

predictors, the less time needs to be spent gathering information, and hence the more

economical modeling. This holds tme as long as the subset of predictors can explain

nearly as much of the cost and tirne variances as is explained by the entire set of

predictors.

It is necessary to dehe the most critical indicators to project cost and time variances.

This also reduces rnuch of the ANN training time. As known in statistics, a stepwise

solution enables us to include only those predictors that add significantly to predictive

power. A forward regression approach was considered and used in defining the

significaat predictors.

Page 111: An artificial neural network approach to assess project

5.2.5.1. Foiward Regression

In a forward regression, the multiple regression equation is built one step at a time by

sequentially addinp predictors to the equation. The first predictor for entry into the

equation is the one with the largest correlation with the dependent variable (the highest

R2).

To determine whether a variable is entered into the equation, the F value was calculated

and compared to an established critenon. For instance, in this study, PIN (probability of

F-to-entry) with a value of 0.05 was used. In this case, a variable enters the equation only

if the probability associated with the F test (PM) is less than or equal to the default 0.05.

If the criterion is met, the variable is entered into the equation and the procedure is

repeated. The variable with the largea partial correlation is the next candidate. Choosing

the variable with the largest partial correlation in terms of absolute value is equivalent to

selecting the variable with the largest F value. The procedure stops when there are no

variables that meet the entry criterion.

The SPSS software package was used and pefiormed the forward procedures. There were

two dependent variables, cost and time variance. For each dependent variable, a forward

regression was perfomed.

Page 112: An artificial neural network approach to assess project

5.2.52. Forward Regression with Dummy Variables

5.2.5.2.1. Dummy Variable

Because there were categorical variables (discretion variables) involved which had no

quantitative meaning, these variables had to be converted into a series of dichotomously

scored "dummy" variables. The number of dummy variables created was equal to the

number of categones of the original variables and each case was scored O or 1 on each

durnmy variable. Scores of O were used to indicate lack of membership in the category

represented by a dummy variable; Scores of 1 showed that a case was a member of that

category. For example, the original variable 'complexity' was converted to three dummy

variables representing three categories of complexity with 'high', 'medium', and 'low'

respectively. When a case of 'high' occurred, the score of the first dummy variable was

1.

Once independent variables (contûiuous and dummy) were refomed (see Appendix IV

for details), two separate forward regression analyses were carried out using the entire set

of one hundred and three samples. One forward regression analysis was for cost-

variance; the other was for the-variance. Two lists of significant predictors were gained,

shown in Table 5.13 and Table 5.14.

Page 113: An artificial neural network approach to assess project

Table 5.13. "Dummy", Discretion Variables, R2 and SD of Cost Variance

Note: See Appendix IV for the details of the description of durnmy variables.

Dum. V.

' V 3 3

V 16-3

V 1 7-2

v2-1

Figure 5.4. DEerence of R2 VS. SequentiaUy Added Predictors to Cost Variance

Dis. V. V7 V3 V6

V16 V 1 5

7 I

Description

‘‘hi&"

"Unit Price"

"Established in the industry but new to your Org." ,

As can be seen in Table 5.13, nsk factors, such as project cost spent in fiont-end and

detail engineering, hi& complexity, 'Unit prke' contract, years of experience, new

technology and project location with 'Urban', add signincantly predîctive power for cost

variances prediction. Figure 5.4 presents the relative significance of the predictors to the

"Urban"

Description Cost in Detail Design

Compiexity

Cost in Front End

Contract type

Years of Expenence

Level of Technology

v2

R2

0.24

0.3 1 0.40

0.44

0.49

0.5 1 I

SD 0.16 1

0.153 0.143

0.139 0.1 34 0,132

1 Project Location 0.53 0.129

Page 114: An artificial neural network approach to assess project

102

cost-variance. 'Cost spent in the fiont end stage' is the most important variable to cost

variances. 'High Complexity' is the second most important variable to cost variance.

'Unit-price Contract' and 'Years of experience' are right in the middle. 'Level of

technology' and 'Project location ' are less important than othen. Figure 5.4 also

indicates that 'Cost spent in detail design', 'Complexity' and 'Cost spent in fiont-end' are

the top three predicton which rapidly increased the R* of cost-variance.

Table 5.14 "Dummy", Discretion Variables, R2 and SD of Time Variance

Dum. V.

v4-3 v i s

v 16-3 VI-2

V18-3 v4-5

V2-2 v3-3

VI-1

v 12-4

vi2-1 2 - 1

V 1 7-3

Note: See

Dis. V.

v4 VI V16 v1

VI8 v4 V2 v3 VI V5

VI2 v12

v 2 V13 V 1 7

of the

Description "From 15 to 30"

"Othen"

"Unit Price"

"Pipeline"

'5 25% activity affected"

"No's of Key Org. >50"

"Suburban"

"High"

"Gas Plant"

"From 15 to 20"

"From O to 5"

"Urban"

"New to the indutry but used in other industry"

Appendk IV for the details

SD 0.115 0.111 0.108

0.104 0.098 0.095

0.093 0.091

0.090 0.087

0.085 0.083 0.080 0.077 0.075

Description No's of key Orgs.

Project Type

Contract Type

Project Type

Weather

No's of key Orgs.

Project Location

CompIexity

Project Type

Design Completeness

No's of Permits

No's of Permits

Project Location

Scope Defined

Level of Technology

description of dummy

1 R2 0.12 0.19

0.24 0.30 0.39 0.43 0.46 0.49 0.51 0.54

0.57 0.60

0.63 0.65 0.68

variables.

Page 115: An artificial neural network approach to assess project

1 O3

We c m tell from Figure 5.5 that 'Weather' with 'over 25% activity affected' adds the

most to the R-square of time variance. The second most significant variable is the

'Project typeT with 'others'. 'Complexity' ranked lower in tirne variance compared to

cost variance.

.-

Figure 5.5. Difference of R2 vs. Sequentiaily Added Predictors to Time Variance

5.2.5.2.2. Definhg Input Variables

After compiling the two lists of predictors for cost-variance and the-variance

respectively, a List of thirteen variables having sigoificant importance to cost and time

variances was obtained (see Table 5.15).

Table 5.15 illustrates that 'Project Location', 'Complexity', 'Contract Type', and 'Level

of Technology' are signincantly important to both cost variances and time variances.

Page 116: An artificial neural network approach to assess project

1 O4

Table 5.1 5 clearly shows that there are more factors influencing time variances than cost

variances. In other words, time variances rnay result from more complex reasons than

cost variances.

Table 5.1 5. Input Variables in the development of ANN Models

Note: * donates that the variables marked were significant to cost variances or to time variances.

As known, the fast-tracking approach c m speed up project construction but usually

causes increasing cost to get shorter duration. Therefore, it may be argued that 'Design

Completeness' should have significant importance to cost variations. It has to be

understood that each of the factors has a certain level of importance to cost-variance. The

results fiom the regression d y s i s showed 'Design Completeness' did not add

significantiy to the predictive power of cost-variance. Finally, Thirteen varîables listed in

Cost Variance

*

n

*

*

Description

Project Type Project Location

Complexity No'sofkeyorganizations

Design Completeness Cost in Front End

Cost in Detail Design No's of Permits Scope Defined

Years of Experience Contract Type

Level of Technology Weather

ID

1

Time Variance

*

* * n

Variables

v l 2 3 4 5 6 7 8 9

I O Il 12 13

v2 v3 v4 v5 v6 v7 VI 2 VI 3 VI 5 VI 6 VI 7 VI 8

Page 117: An artificial neural network approach to assess project

105

Table 5.15 were selected to be input variables in the second phase of the development of

ANN model.

5.2.5.3. Forward Regression without Dummy Variables

The second way to define the cntical attributes to cost and time variances was the use of

the discretion variables. Artificial neural network approaches can process with both

binary-value and continuous value. it is not necessary to refom discrete variables

(category variables) to dumrny variables (binary-value). The real value of each variable

could be simply used.

Forward regression analysis was performed using SPSS. Two sets of critical indicators to

cost and tirne variances were obtained, shown in Table 5.16 and Table 5.17. The

remaining indicators were not listed.

Table 5.16. Original Variables, R2 and SD of Cost Variance

SD 1

0.161

0.153

V6

R* 0.24

0.3 1

Original Variable

v 7

V3

Description Cost in Detail Design

Complexity I

Cost in Front End 1 0.3 7 0.147

Page 118: An artificial neural network approach to assess project

1 O6

From Table 5.16 and Table 5.17, we can derive the following table (Table 5.18) which

illustrates the significant predictors to codtirne variances or both. These ten predictoa

were selected to be used in th5 second phase of the development of the artificial neural

network models. Table 5.1 8 clearly shows that project complexity is the only variable

that has significant impact on both cost and time variances.

Table 5.17. Original Variables, Et2 and SD of Tirne Variance

Table 5.18. input Variables in The Development of ANN Models

Original Variable

v 4

V18 v1 VI1 V3 v 5

V2 V15

R2 0.12

0.17

0.2 1

0.23 0.25 0.27

0.29 0.30

Description No's of Key Orgs.

Weather

Project Type

Priority: Cost

Comp Iexiq

Design Completeness

Project Location

Years of Experience

SD 0.1 1s

0.1 13

0.1 11

0.1 10

O. 1 09

0.108

0.107

O. 1 06

Time Variance *

t

*

*

Cost Variance ID

1 2 3 4 5 6 7 8 9 10

Variables

v 1 v2 v3 v4 v5 v6 v7 VI 1 v15 v18

Description

Project Type Project Location

Corn plexity No's of key organizations

Design Completeness Cost in Front End

Cost in Detail Design Prïority: Cost

Yearç of Experience Weather

Page 119: An artificial neural network approach to assess project

1 O7

As we see in Table 5.18, only one variable 'Priority: C o d was not in Table 5.15. The

test were in Table 5.15. It indicated that there was a consistency between the two

methods. More interesthg was to see the results fiom ANN training using these two sets

of input variables.

5.2.6. Phase II Training - Using Critical Input Variables

After we determined two sets of input variables, ANN training processes were then

undertaken using grouped project data. The total project data were grouped into three

sets listed as the follows:

W Total projects (Case 1): one hundred and thee projects (ninety-three for training and

ten for testing)

Pipeline projects (Case II): thirty-six pipeline projects (thirty-two for training and four

for testing)

Refinery projects (Case III): twenty refinery projects (sixteen for training and four for

testing)

The configuration of the ANN models is kept the same as in the ANN training ushg

eighteen variables except that the number of nodes on the input layer was changed

because the number of input variables changed. The number of hidden nodes needed to

Page 120: An artificial neural network approach to assess project

108

be determined through a training process. The standard deviations and R2 of cost and

time variance fkom predicting testing data are presented in Table 5.19.

Table 5.19 shows that the performance of ANN models trained using project data fiom

one type of project are better than those using al1 available project data which has a broad

range of project types. The reason is that, in the pool of studied projects, there are several

different types of projects, such as pipeline, gas plant, refinery, etc., each of which has

different risk patterns. Decisions, risk impacts and project outcomes differ significantly

from one project type to another. These differences cause difficulties for the leaniing

processes of neural networks and result in lower performance of ANN models.

The standard deviations of error of predicted cost variance are larger than for time

variance. In the separate data cases for pipelines and refinenes, the R2 of t h e variances

are larger than for cost variances in the three cases of input variables. This means that the

capability of ANN models to predict duration variance is better than for cost variance.

But in case 1, the R2 of COS^ variances are larger than those of t h e variance. The standard

deviations of cost variance of refhery projects are larger than for pipeline projects. This

may be due to the greater complexity of refmery projects than pipeline projects.

Page 121: An artificial neural network approach to assess project

1 O9

We cm see that the pefiomance of ANN models with ten variables is better than those of

other models throughout case 1, case II, and case III. It seems that it is not necessary for

ANN rnodels to use 'dummy' variables in the determination of critical indicaton.

Page 122: An artificial neural network approach to assess project
Page 123: An artificial neural network approach to assess project

5.3. Penonnance Comparison between N e u d Network and Multiple Linear

R e g m i o n Analysis

In this section, performance comparison between ANN rnodels and multiple linear

regression models is presented. The comparison use the Case I sarnple. The same group

data and independent variables were used to build up regression models. Standard

deviation of error and R2 with testing data were calculated and the results are illustrated

in Table 5.20.

Table 5.20 shows that the A m ' s performance measure, R' is much larger than that of

regression models. Predicted values fiom ANN models highly correlate the desired

values. The generdization capability of ANN models is superior to multiple linear

regression models. It means that the predictive accuracy of ANN models is higher than

multiple linear regression models. ANN models are supenor to the multiple linear

regression models in predicting project cost and îirne variances.

Table 5.20. Comparison between ANN Mode1 and Multiple Linear Regression

Page 124: An artificial neural network approach to assess project

CHAPTER 6 CONCLUSIONS & RECOMMENDATIONS

6. f . Conclusions

The effectiveness of the current nsk analysis techniques is heavily dependent upon

experts' personal experience and judgment. The weaknesses of these techniques have

limited their applications in complex situations in which no mathematical models can be

applied and interactions between nsk factors cannot be quantified. This weakness

necessitated the adoption of a new approach to assess project cost and time risks.

This research studied the potential applications of artificial neural networks in the

assessrnent of project risks in the early stages of a project. Neural networks were used to

capture the relationships among risks, project characteristics, decisions, and outcomes.

Intelligent models were then developed and tested. They can be used to predict project

cost and t h e variations.

The research remlts show that artificial neural nehntork technology surpasses

conventional models such as multiple linear regression andysis, which is a cornmonly

used method to build predictive models. The practical application of ANN technology in

project risk analysis is promising, especiaily in the fiont-end stage.

Page 125: An artificial neural network approach to assess project

113

This research made significant progress in the quantitative assessment of project cost and

time risks using artincial neural network technology. More complicated ANN models

with a quality development process and high-level performance measures are applied in

this research.

McKim [1993] studied quantitative assessment of project cost overruns using artificial

neural network technology. Four risk factors (contractor, architect, location and size)

were considered. Twenty projects were used to develop an ANN model. Standard

deviation of error was used to measure the performance of the ANN model. A standard

deviation of error of cost overruns fiom his ANN model was 3.6.

In this research, the largest standard deviation of error of cost ovecruns fiom ANN models

was 0.396. This is about a 90% improvement in standard deviation of error compared to

the results fkom McKim's model. It can be explained that McKim's model was too

simple to explain the variations of cost overruns. More important risk factors should be

considered in the ANN models.

Signifïcant risk factors to cost variances and time variances were identified in this study.

Project type, project location, cornplexity, number of key organizations involved in a

project, design completeness, cost as the number one project priority, years of expenence

in similar projects that a project leader(s) has and weather were identifcd having high

Page 126: An artificial neural network approach to assess project

114

correlations with t h e variation. Cost spent in detail design, complexity and cost spent in

the front end engineering phase were identified to have high correlations with cost

variation. Intuitively, cost as the number one project prionty has important impacts on

cost variance rather than on time variance. But if you have a limited budget you might

spend more time in the engineering phase to fhd a cost-effective approach in your

project. There is a tradesff between money and time.

Recent research [Pedwell, Hartman and Jergeas, 19981 concluded that the fast-tracking

approach produces less project definition at the start of construction and reduces the total

project duration but usually causes cos? increases to achieve this shorter duration.

Therefore, design completeness should have significant impacts on cost variances.

However, in this study, design completeness was not identified as having high correlation

with cost variances.

This study's results show that project complexity is the most significant attribute to cost

variation and tirne variation. In this study, project complexity refers to technical

complexity of a project. This is a very subjective term. Here, project technical

complexity hcludes elements such as the accessibility, technical requirements and

Limitations, numbers of major rotating equipment, and specialty of materials. It can be

concluded that reducing or managing these complexity factors in a project at the fiont end

Page 127: An artificial neural network approach to assess project

115

planning stage may resdt in a greater probability of successful implementation of the

project.

Overall, projects will have a greater chance of success, in tems of 'within budget' and

'on time', when project managers direct more effort into managing the identified

important factors during project planning in the Front end phases.

Other conclusions fiom this research include the following:

An artificial neural network is able to caphue the rkk patterns of projects by learning

fiom histoncal project samples and to generate a reasonable prediction of project cost

and tirne variances.

Artificial neural network technology is superior to conventional multiple linear

regression analysis in building predictive models of project cost and time variance

behavior.

ANN with stepwise regression andysis provides more accurate estimates of project

cost and time variations than ANN on its own. It aiso significantiy reduces the

training t h e and increases the training efficiency of the networks.

Page 128: An artificial neural network approach to assess project

116

0 The larger the sample size used in the development of ANN models, the more

accurate the ANN model is likely to be.

0 ANN model generalization would be improved if similar sarnples were used, for

instance, similar projects and projects collected fiom one organization.

This research built a rational base for developing a decision support system to assist

project managers (decision makers) in better decision making.

The scope of this research is delimited to the project implementation phase and only

project intemal risk factors were modeled. Therefore, the use of the ANN models is

delimited.

The project operation phase is an important phase to a project. The ANN models will be

improved if project operational risk factors are included. Project extemal risks were

ignored in the development of the ANN models. The external factors could be related to

elements such as oil price, labor market, and project revenue cashflow. These factors

have iduences on projects in terms of cost and t h e . Future research in project external

risks in the development of ANN models is recommended.

Page 129: An artificial neural network approach to assess project

117

The resuits fkom this study were positive, but there are some limitations. The major one

is the existence of bias in the data. Since the data was collected from different

organizations and different individuals, the data would be not consistent in subjective

issues. Different individuals would have different expenences and therefore different

opinions about similar occurrences or situations. Inconsistent data could introduce a

certain amount of noise into the ANN models. Solutions to improve data consistency

include the following :

Elirninate subjective issues.

Provide clear explanations on what information is being collected.

Focus on one organization and one group of project personnel if possible.

6.2. Recommendations

ANN technology shows promise for application in the field of project nsk analysis. The

issues of developing of ANN models have to be carefully considered. ANN training is a

time consuming process. The determination of ANN configuration is the most diacult

portion of the work. Recommendations firom this research for the development of ANN

models include the foilowing:

Page 130: An artificial neural network approach to assess project

118

The paradigm and configuration of neural networks have to be congruent to the nature

of the problem. The panuneters of the networks, such as leaming rate, learning

tolerance, and rnomentum constant, have to be defined through an intensive training

process.

ANN training time can be saved if critical input variables are used. Stepwise

regression analysis is an approach to define critical input variables.

During the training process, both over-fitted and under-trained phenornena were

observed. Over-fitting the training samples happens when ANN models are over-

trained (too many iterations allowed or too many hidden nodes). This results in ANN

models losing their generalization capabilities. Not enough iterations of training

could remlt in inadequate Iearning of ANN models. No f o d a exists to calculate

the exact number of iterations and the number of hidden nodes that should be used in

ANN training. Therefore, the number of iterations of ANN training and the number

of hidden nodes have to be adjusted through the extensive training processes.

The performance measures of ANN models have to be set up in advance. The

m e m e s c m be ditferent h m one problem to another. The author strongly

recommends that cross-validation (R2 in the testing samples) be used as a

performance measure of predictive models when developing them.

Page 131: An artificial neural network approach to assess project

119

Sarnple sizes used in training and testing are also important. Data sample size will

significantly affect the performance of ANN models. The Iarger the sample, the

better the performance is likely to be. A minimum size of sample for training is

necessary to reach an acceptable threshold of prediction accuracy.

It is recommended that an organization-wide database, continuously updated be built to

capture al1 records of completed projects in the organization. The database will provide

solid and rich data for performing project risk analysis and predicting the outcornes for

proposed projects. This will definitely benefit the project management tearn.

The recommendations for the practical Mplernentation of the ANN models in industry

would include the following:

r Transfer the ANN models developed (Le., ANN configurations) from the CTMX

system to PCs using IBM-PC version of Matlab software package if necessary.

Design and program the interfaces between users and the ANN models including data

entry, presentation of resuits, and analysis of results.

Page 132: An artificial neural network approach to assess project

6.3. Further Research

This work has opened a wide research potential and made significant progress in the

project risk analysis domain using a novel approach - artificial neural network

technology. More work will need to be done to improve and implement the developed

ANN models.

Further research work should include the:

Use of the backward approach to fmd out the relative importance of the independent

variables to the cost and time variances ushg the developed ANN model.

Consideration of extemal risk factors.

Use of cluster analysis to group projects based on a similarity measure called

"resembtance coefficient". The smailer the value of the coefficient, the more similar

the projects are. Using more similar projects to train neural networks could results in

better performance of the models and improve the accuracy of prediction.

a Performance of the ANN models wodd be much better if the samples used in the

training process are fiom a single organhtion. This is probably because project

Page 133: An artificial neural network approach to assess project

121

performance maintains some similarity within the same organization. Also, risk

factors can be identified more specifically to the organization. This would increase

the predictive accuracy of the models for project cost and time variations.

Use of sensitivity analysis on ANN models to establish how the performance of the

ANN model changes as one of the inputs changes. This will assist project managers

to screen decision alternatives and choose the best one.

Establishment of an industry expert panel compnsing expenenced project mmagers

to test the predictive capability of ANN models against the experts. Experts will be

asked to provide predictions of project outcornes in ternis of cost and time variances.

The projects used in the test are a small sample of the data pool.

6.4. Contribution of the Thesis to the Body of Knowledge

This research has demonstrated the feasibility of applying artificial neural network

technology to the development of predictive models of project cost and tune variances in

the early stages of projects. The artificially intelligent predictive models surpass multiple

linear regression models in the study area This research suggested the development

procedures of the network models and addresses the most sigoificant issues surrounding

neural network model development.

Page 134: An artificial neural network approach to assess project

122

For capital project planning and control, an accurate estimate of capital project costs is

most important, especidly in the front-end stages. This research provides a base to

predict project cost and time variances. This research explores the most significant

factor-project complexity- that correlates to project cost and tirne variances. Other factors

af5ecting cost and tirne variances are also identified. The relationship between factors

and project cost and time deviations are established within the neural networks. More

focus on these factors in the decision making process in the front end stages will increase

the possibility of the success of capital projects in tems of 'within budget' and 'on time'.

This research provides the basis of an ANN model to assess the project cost and time

risks. Based on such an ANN model, a decision making support tool can be developed to

help project managers screen the potentiai options and make better decisions.

The research proposed a different approach to evaluate project risks and to capture project

risk patterns. ANN7s ability to leam would help practitionea capture an organization's

knowledge on projects and to create and irnprove an organization's learning.

There are two practical uses of the ANN models in risk analysis and management. They

are the foiiowing:

Page 135: An artificial neural network approach to assess project

Provide a rational base for a contingency plan when a

for Expenditure (AFE). The ANN models typically

123

project is going for Approval

help owner organizations to

assess not only an individual risk impact but also the impact of a set of risks on

project cost and time at the early stage of a project. The models are used when initial

estimates of cost and duration of the project are established. Project nsk impacts on

project cost and tirne are then evaluated to determine a contingency. By using the

ANN models, project cost and time variances will be quickly determined by assessing

the individual nsk impacts on the project cost and time. The total variances could

form the basis of a contingency plan for the project.

Define cost and time variances resulting from individual risk using the ANN models

and help practitioners to establish inputs for perfonning conventional risk analysis,

such Monte Carlo simulation. It will reduce expert's personal opinion and judgment

on the risk impacts as inputs to Monte Carlo simulation.

Developing a neural network mode1 that simuitaneously considers project cost and time

variances is a previously unexplored area of study. The methodology of, and the results

from, this research will provide valuable references for sirnilar friture research work.

Page 136: An artificial neural network approach to assess project

REFERENCES

Bansal, A., Ka&an, R. J. and Wei@ R. R., Comparing the modeling performance of

regression and neural networks as data quulity varies: a business value approach,

Journal of Management Information Systems, Summer 1993, Vol. 1 O, No. 1, pp. 1 1-32.

Basheer, 1. A. and Nadar, Y. M., Predicting Dynamic Response of Adsorption Columns

with Neural Networks, Jouml of Computing in Civil Engineering, Vol. 10, No. 1,

January, 1996.

Bataineh, S ., An expert cornputer- based system for munuging complex decision-making

systerns, Journal of Cornputer Information Systems, Winter 1995- 1996, pp. 10 1 - 105.

Bode, Jhgen, Neural Networkr for Cost Estimation, Cost Engineering, January 1998.

Bowers, J. A., Data for project risk analyses, International Journal of Project

Management, 1994 12(1), pp.9-16.

Caudill, M., Neural Network Training Tips and Techniques, AI Expert, January 199 1.

Cooper, Dale and Chaprnan, Chns, Risk Analysis for Large Projects, John Wiley & Sons

Ltd., 1987.

Cooper, Donald R. and Emory, C. William, Business Reseurch Methoh, Richard D.

Irwin Inc., 1995.

De Grooî, C. And Wurîz, D., Anaiysis of univariate time series with connectionist nets: a

case study of two clarsical examples, Neuro Computing, 3, pp 1 77- 192, 199 1 .

Page 137: An artificial neural network approach to assess project

125

Demuth, Howard and Beale, Mark, Neural Nemork Toolbox, The MATHWORKS Inc.,

1995.

Denton, J. W., Sayeed, L. and Perkins, N. D., Neural Networks to CZussz~ Employeesfor

T a Purposes, Accting., Mgmt. & Mo. Tech., Vo1.5, No.2, pp. 123- 138, 1995.

Dey, Prasanta, Tabucanon, Mario T. and Ogunlana, Stephen O., Planning for Project

C o n ~ o l Through Risk Analysis: A Peîroleum Pipeline-laying Project, International

Journal of Project Management, 12 (l), pp. 23-33, 1994.

Diekmann, J.E., Risk Analyss: Lessonsfiom Artifcial Intelligence, International Journal

of Project Management, Vol. 1 O, No.2, May 1992.

Doherty, N.A., Corporate Risk Management, McGraw-Hill Inc., New York, NY, 1985.

Duggal, S. M. and Popovich, P. R., Practical applications of neurai network in business.

Journal of Cornputer Information S ystems, winter 1992- 1993.

Dutta, S. and Hekhar, S ., Bond-rating: a non-conservative application of neural

networks, Proceedhgs of the IEEE International Conference on Neural Networks,

Vol.I1(1988), San Diego, CA, pp.443-450.

Eberhart, R. C. and Dobbins, R. W., Neural Network PC Tools, Academic Press, 1990.

Federle, M. 0. and Pigneri, S. C., Predicrnte Mode1 of Cost Chemins, AACE

Transaction, 1993.

Page 138: An artificial neural network approach to assess project

126

Fletcher, D. and Goss, E., Forecasting with neural networkr: an application using

bankruptcy data, Information & Management, 24 (1 993) pp. 159-167.

Hecht-Nielsen, R., Neurocompziting, Addison-Wesley Co., New York, 1989.

Koster, A., Sondak, N. E. and Bourbia, W., A business application of artifcial neural

network qstem, The Journal of Computer Information Systems, winter 1990- 199 1, pp.3-

9.

Johnston, S.J., Neural network are muking inruad: technology has found many

practical applications, Inforworld, July 8, 199 1, pp. 13- 16.

McKim, R. A., Neural networks and identifcation and estimation of risk, AACE

Transactions, 1 993, pp.5.1-5.1 O.

Merrow, W. Edward and Yarossi, Mary Ellen Assessing Project Cost and Schedule Risk,

1990, AACE Transactions.

Morris, P.W.G., and Hough, G.H., The Anatomy of Mq-or Projects, John Wiley and Sons,

New York, 1995.

Moselhi, O., Hegaq, T. and Fazio, P., Neural networks as tools in consîruction, Journal

of Construction Engineering and Management, Vol. 1 17, No.4, December 199 1, pp.606-

625.

Murtaza, Mirza B. and Fisher, Deborah I., Neuromodex - Neural Network System for

Modula7 Construction Decision Making, Journal of Computing in Civil Engineering,

Vol. 8, No. 2, Apnl, 1994.

Page 139: An artificial neural network approach to assess project

127

Mustafa, M. A. and Al-Bahar, J. F., Project Risk Assessrnent Using the Analytic

Hierwchy Process, IEEE Transaction on Engineering Management, Vo 1.3 8, No. 1, Feb.

1991.

Osyk, B. A. and Vijayaraman, B. S., Integrating expert systems and neural networks,

Information System Management, Spring 1995.

Papageorge, Thomas, Risk Management for Building Profesionais, 1 98 8, R. S. Means

Company Inc.

Pedwell, K., Hartman, F. and Jergeas, G., Project Capital Cost Risks and Contracting

Strategies, Cost Engineering, January 1 998.

Pedwell, K., Liu, X. and Hartman, F., Computer Models for Assessing the Probability of

Achieving Time and Cost Targets, Proceedings of The Third Canadian Conference on

Computing in Civil and Building Engineering, Montreal, August 1996.

Pike, R.H. and Ho, S.S.M., Risk Analysis in Capital Budgeting: Barriers and Benefis,

Omega, Vol.19, N0.4, 1991, pp.235-245.

Simister, Steve, Usage and Benefis of Project Risk Analysis and Management,

International Journal of Project Management, 1994 12(1), pp.5-8.

SLicher, A. W. R., Vakalis, P. and Singh, G., An innovative approach to training neural

networks for strategic management of comtruction fim, CML-COMP PRESS,

Edinburgh, UK, 1995.

Page 140: An artificial neural network approach to assess project

128

Sohl, J. E. And Venkatachalam, A. R., A Neural Neîwork Approach to Forecasiing

Mode1 Selection, Information & Management 29, 1 995, pp.297-303.

Suh, Y. H. and LaBane, J. E., An applicaiion of artifcial neural network models to

porfolio selection, Journal of Cornputer Information Systems, Fdl 1995, pp.65-73.

Tal, Benny and Nazareth, L inda, Artifcial Intelligence and Economic Forecast ing

Canadian Business Economics, Spring 1995, pp.69-74.

Tarn, K. Y. and Kiang, M. Y., Managerial applications of neural neiworks: the case of

bank failure predictions, Management Science, Vo1.38, No.7, July 1992, pp.926-947.

Touran, Ali and Wiser, Edward P., Monte Carlo Technique with Correluted Random

Variables, Journal of Construction Engineering and Management, Vol. 1 1 8, No.2, June

1992.

Wasserman, P. D. and Schwartz, T., Neural Nehvorks, Part I. IEEE Expert, 2:4, 1987, pp.

10-13,

Wilson, R.L., The strategic organizational use of neural networks: an exploratory study,

unpublished dissertation, University of Nebraska-Lincoln, 1990.

Wirba, E.N., Tah, J. H. M. and Howes, R., Risk Inierdependencies and Natural Language

Compututions, Engineering, Construction and Architectural Management, Vo1.3, No.4,

Dec. 1996.

Page 141: An artificial neural network approach to assess project

129

Wong, B. K. and Monaco, J. A., &ert qstem applications in business: a review and

analyss of the Ziterature (1 9 77- 1993), Infornation & Management 29( 1 999 , pp. 1 4 1 - 152.

Page 142: An artificial neural network approach to assess project

APPENDIX 1. SURVEY RESULTS ON CURRENT USAGE AND BENEFIT OF RlSK ANALYSIS TECHNIQUES

1. Suwey Questionnaire on Current Usage and Benefits of Risk

Analysis Techniques

Introduction

Risk analysis and management has received increasing attention fiom project

management practitioners. The purpose of this study is to investigate the:

current use of risk analysis techniques by industry in the Calgary area,

what risk analyses are being perfomed,

what cornputer software packages if any are being used to carry out the

anaiysis, and the reasons for and benefits arising nom its use.

The research results will be released to each participant. The results will provide project

management practitioners with state-of-art information on rîsk analysis and management

in indusîry.

0 Questions

1. Yom Name: Te1 #: Mailing Address:

Page 143: An artificial neural network approach to assess project

2. Willing to participate another main survey: Yes No

1 . Use "4s" to identiQ the size of projects that risk analysis techniques should be used

by project management:

1 Most use 1 Better to use 1 No need to use 1

1 I 1

I Over 20,000,000 I I I

4. If your Company uses risk analysis techniques, what were the reasons:

(Circle one or more items)

(1). Client request

(2). Internai Policy

(3). Persona1 use

(4). Required by other people in own organization

(5). ûther (speciQ):

Page 144: An artificial neural network approach to assess project

Note: I = Less usefull: 2=Neutral; 3 = Usefil; J= Very Useful.

Page 145: An artificial neural network approach to assess project

133

6. Check out one or more cornputer software package(s) in risk analysis used in your

organization:

7. Select one or more of following benefits of using risk anaiysis techniques:

Cornputer Software 1. Ccystal Bal1 (works with Excel, Lotus)

3. Risk + (Works with MS-Project V.4.0)

5. Monte Carlo (Works with Primavera Proiect Planner)

7.Other(speciQ):

(1). Allows the formulation of more realistic plans, in terms of cost estimate and timescale;

(2). Gives an incnased understanding of the risks in a project;

(3). Allows the assessrnent of contingencies that actually reflect the risks;

Used

(4). Facilitates greater, but more rational, risk taking, thus increasing the benefits that can be

gained from risk taking;

(5). Identifies the party best able to handle a nsk;

Computer Software

2. @Risk (for Excel)

4. @ Risk for Project

6. Expert Choice

8.Other(speciQ):

(6). Builds up statistical information about historical N k s that assist in better modeling of firture

project;

Used

Page 146: An artificial neural network approach to assess project

(7). Leads to the use of the most suitable form of procurement/contract;

(8). Assists in distinpishing between good luck/good management and bad luclchad

management.

(9). Other reasons (specify):

8. If your cornpany does not use nsk analysis techniques, which of the following

describes the reason: (Circle one or more items)

(1). Waste of time and efforts to do risk analysis; (2). No tirne to do it;

(3). No money to do it;

(4). No intemal expertise to carry out risk analysis;

(5). No information available to cany out risk analysis;

(6). Other (speciQ):

9. Type and name of company you are in: (Circle one item)

Page 147: An artificial neural network approach to assess project

(2). Engineering Firm

(3). Construction Contractor

(4). Other (speciQ)

Your company's name: (Please specifi Company and division)

10. Your curent position: (Circle one item)

(1 ). Project Estimator (2). Project Engineer (3). Project Manager

(4). Other (specifjQ

1 1. Years of your experience in Uidustry:

Thank you so much for your Pme, efforts, and insights.

Page 148: An artificial neural network approach to assess project

2. Survey Results

(1). Project Size vs. Percentage of Responses:

S i t e of Project

- ~ M u s t Use

Q Better to Use

No Need to Use

(2). Software Packages Used vs. Frequency of Responses:

Softwate Packages Frequency of Responses

Crystal Ball(with Excel, Lotus) 16

@Ris k(for Excel) 21

Risk + (Wh MS Project) 26

@Risk for Project 11

Monte Cario (with PrÎmavera) 37

Expert Choice 11

Otherî (Dyadem Hazop)

D&RA

I n-iiouse Oevelo ped P rograms 26

REPlPC 32

Page 149: An artificial neural network approach to assess project
Page 150: An artificial neural network approach to assess project

(4). Benefits of Performing Risk Analysis vs. Frequency of Responses:

BENEFITS OF RISU ANALYSIS Frequency 01 Res~onses

Gives an increased understanding of the risks in a project

Allows the assessrnent of contingencies that actuaily reflect the 90 risks

Allows the formulation of more reaiistic plans, in terms of cost 71 estimate and timescale

Facilitates greater, but more rational, risk taking, thus increasing 58 the benefits that cm be gained fiom risk taking

L

Leads to the use of the most suitable fom of procurementkontract 29

Builds up statistical information about historical risks that assist in 26 better modeling of future project

Assists in distinguishing between good luck/good management and 26 bad lucklbad management

Identifies the party best able to handle a risk 13

Page 151: An artificial neural network approach to assess project

APPENDIX II. CONSENT FORM

Consent for Industry Survey

Research Project Title: An Artificial Neural Network Approach to Assess Risk Effects and Project Decisions in the Front-End Stage

Researcher: Xiaoying Liu

Research Supervisor: Dr. George Sergeas

Funding: NSERClSSHRC

This consent form, a copy of which has been given to you, is only part of the process of infiormed consent. It should give you the basic idea of what the research is about and what your participation will involve. If you would like more detail about something mentioned here, or information not include here, please ask. Please take the time to read this form carefully and to understand any accompanying information.

n i e purpose of the study is to identiQ project cntical risk factors and to evaluate risk impacts on project performance, capital cost and schedule. You have been contacted because you have expenence in project fiont end planning and development. This consent form is to describe the industry survey to you and to request your consent to conduct one or more surveys.

The s w e y consists of a set of questions I would like to ask you. Al1 questions are related to a project you have worked on. One survey is to be completed for each project. The size of the reviewed project shouid be over S 1 MM in value and the project must have been completed within the past five years.

Participants participating through a personal interview will be asked to provide a time cornmitment of one to two hours. Interviews will be arranged at a time and place that is convenient for you.

Ni data colIected will be pooled in a database and no information fiom any specific individuai or Company will be released without pnor written authorization. The database

Page 152: An artificial neural network approach to assess project

will be stored on diskettes and managed and kept directly by the researcher. Only the researcher's supervisor and the researcher have access to this information. You will have access to the information that you have provided. You will also have the oppomuiity to delete, change and destroy any information you have provided during the study. No othea will be allowed to use the coilected data without your permission. Aggregated (and de-sensitized) data that carmot be separated or identified with a particular contributor will be used to report research W i g s and will therefore be publicly accessible. If you decide to participate, you will need to carefully read and sign this consent form. Thank you for taking the tirne to read this information.

Your signature on this form indicates that you have understood to your satisfaction the information regarding participation in the research project and agree to participate as a subject. In no way does this waive your legal rights nor release the investigators, sponsors, or involved institutions fiom their legal and professional responsibilities. You are fiee to withdraw fiom the study at any t h e . Your continued participation should be as informed as your initial consent, so you should feel fiee to ask for clarification or new information throughout your participation. If you have questions conceming matters related to this research, please contact:

The Researcher: Xiaoyin~ Liu at 220-5970 or email: [email protected]

Researcher's Supervisor: Dr. George lergeas at 22041 85

If you have any questions conceming your participation in this project, you may also contact the Office of the Vice-President (Research) and ask for Karen McDermid, 220- 3381.

Signature of Participant Date

Signature of Researcher Date

Page 153: An artificial neural network approach to assess project

APPENDIX III. SURVEY QUESTIONNAIRE

Industry Survey Oil and Gas Project9s Time and Cost

Overruns Predictive Mode1

Introduction

The purpose of this s w e y is to obtain project information on individual project characteristics, risks, decisions, and economic characteristics fiom the oil and gas industry. This idormation will be used in support of research being undertaken at The University of Calgary. The goal of this study is to develop an intelligent predictive model of project time and cost variances using artificial neural network technology.

The intelligent model will be used as a tool to evaiuate the effects of risks and project decisions on project outcornes. This will assist decision makers in improving the decision making processes at the fiont end stage of projects in the oil and gas industry.

Confidentiality

AN information obtained will be held in strict confidence by the researcher. AI1 data will be pooled in a database and no information fiom any specific individual or Company will be released without prior written authorization.

Guidance

Your answers to al1 questions should be related to a project you have worked on with a value over CD$lMM dollars and must have been completed within the past five years. One questiomaire form is to be completed for each project. There are four different sections in this form relating to: i Project details

Significant project factors affecting t h e and cost oveduder runs I d e n w the top ten significant attributes to time and cost variances Additional idormation you may wish to add

Page 154: An artificial neural network approach to assess project

We would appreciate receiving completed foms for as many project as possible - the larger our database, the more useful the resuiting mode1 will be. We would also appreciate feedback on the most successful projects and the worst projects as measured by traditional project outcomes.

Data Accuracy

Al1 numenc answers need to be given a confidence Level (CL.). Please use the following scale:

Precise Approximate Guess

Note: 1. Precise: m e r based on accounting or other similar corporate data; 2. Approximate: Answer based on knowledge of situation, but no exact data

available; 3. Guess: Best esrimate based solely on experience W o r intuition.

************************************************************************

Questionnaire

Section 1 - Project details:

In this survey, the project's estimated cost refers to the estimated cost on which the initiation of the project detailed design was approved; the project's actual cost refers to the cost incurred when the project started to operate.

1. Estimated and actual cost:

(1.1). The project's estimated cost: ($Millions) . [C.L.: ]

Answer either (12) or (13).

(1.2). The project's actual cost: ($Millions) . [CL.: ]

or (1 3). The project actual cost as a percentage of the project's estimated coa (%): [CL.: 1

Page 155: An artificial neural network approach to assess project

2. The contingency ailocated in the project's estimated cost: (%) . [C.L.: [

3. Estimated and actual duration:

(3.1). The project's estimated duration comsponding with the project's estimated cost(Months): [CL.: 1

Answer either (33) or (33).

(3.2). The actual duration of the project when the project started to operate(Months) . [C.L.: 1

or (3.3). The percentage of the project actual duration to the project's estimated duration corresponding with the project's estimated cos(%): [C.L.: 1

4. The contingency included in the project's estimated schedule identified in your answer to question 3.1.(%): . [C.L.: 1

5. In which year the project was commissioned: 199,

Section 2 -Project critical factors that have impacts on time and cost variances:

6. Type of the project: (Circle one number)

(1) Gas Plant (2) Pipeline (3) Refinery. (4) Offshore development

(5) Other, please specify

7. Project location: (Circle one number)

(1) Urban (2) Suburban (3) Rural (4) Remote

8. Project complexity:

(8.1). Technical cornpiexity of the project: (accessibility, technical requirements and limitations, number of rotating pieces of major equipment, specialty materials, degree of automation, equipment redundancy, ...) (circle one number)

(1) Low (2) Medium (3) Hi&

Page 156: An artificial neural network approach to assess project

(8.2). Number of organizations directly participated in the project: [ CL.: 1 (Including Owners, Engineering Consultants, Contractors, Sub-Contractors, Suppliers)

9. The percentage of engineering design completeness to the entire project engineering design at the start of field construction: [C.L.: 1

10. Capitaüzed cost expended on the front end planning of the project(prior to detailed design)($): . [C.L.: 1

I l . The cost expended on detailed design of the project($): . [C.L.: 1

12. Rank the following project priorities for this project :

( ). Quality ( ). Safety ( ). Schedule ( ). Cost

13. Number of external permits required in total for the project: [C.L.: J (External perniîa including envimnmental, development, land use, construction, rond

use, ...)

14. The extent to which the project scope was firm and elearly defined when the initiation of the project detailed design was approved: (circle one nurnber) (Project scope and definition typically inciude generaf project basis, process design, site information)

-Project scope was vague; -~aven ' t documented project mission staternent; Poor Well

-Haven 't completed project pre-planning and site investigation. t

-Haven? done Design Basis 5 4 3 2 1

Memorandum & Engineering Design Specification

-Project scope was firm; -Have documented project mission statement; -Have completed project pre-planning and site investigation.

-Have done Design Basis Mernorandum & Engineering Design S pecification

15. Project management experience in engineering, procurement, construction and management (EPCIEPCM) projects when the project was proceeded:

(1 5.1). Your organizationldivision had project management experience in similar projects:

Page 157: An artificial neural network approach to assess project

(Similar projects in ternis of similar type, similar technology used )

(1 5.2). Years of project management experience that the project management leader in the owner organization/division has had with EPCEPCM projects . [C.L.: 1

16. Contract strategy and type used on the project by the owner(s):

( 1 ) EPC or EPCM(Sing1e primary contractor)

Contract Type: (Circle one letter)

(a) Stipulated price (b) Cost plus (c) Unit price (d) other please speciQ:

(2) EPC or EPCM(Mu1tiple primary contractors)

Dominant Contract Type: (Circle one letter)

(a) Stipulated price (b) Cost plus (c) Unit price (d) other please specifjc

(3) Design-Bid-Build(0ne primary contractor)

Contract Type: (Circle one letter)

(a) Stipulated price (b) Cost plus (c) Unit price (d) other please specifi:

(4) Design-Bid-Build( Multiple primary contractors)

Dominant Contnrct Type: (Circle one letter)

(a) Stipulated price (b) Cost plus (c) Unit pnce (d) other please specify:

(5) Partnering or alliance

(Note: Design-Bid-Build means project engineering is separated with procurement and construction)

17. The level of technological innovation of the project: (Circle one nurnber)

(1) Established, very familiar to your organization;

(2) Established in the industry but new to you organization;

Page 158: An artificial neural network approach to assess project

(3) New technology to the industry or sector but established and used in other industry or sector;

(4) Brand new technology application to any indusüy.

18. The percentage of construction actMties adverseiy affected by weather during the project constniction phase(%): [C.L.: 1

Section 3 -4dentify the top ten significant factors to cost and time variances from Section 2.

Section 4 - Additional Information:

Please list additional factors you think that are criticai to the project:

1.

Page 159: An artificial neural network approach to assess project

Please indicate the category which best describes your current position:

Senior Executive(VPICE0) Project Engineer

Senior Project Manager Project Manager

- Others, please specQ

Wouid you like to receive a bnef copy of the results? Yes No -

Your narne:

Mailing Address:

Thank you so much for your time, efforts, and insights.

Page 160: An artificial neural network approach to assess project

APPENDIX IV. DEFINITION OF VARIABLES

Eighteen variables and thek descriptions used in this study are listed as the following:

'Dumrny' variables: 1 : Project is a gas plant. V12: Project is a pipeline. V 1-3 : Project is a rehery. V14: Project is an offshore Development. V15: Others

V2: Project Location

'Dumrny' Variables: V2-1: Project is located in an urban area. V2-2: Project is located in a suburban area. V2-3: Project is located in a rurai area. V2-4: Project is located in a remote area.

V3: Project Complexity

'Dummy' Variables: V3-1: Project complexity is low. V32: Project complexity is medium. V33: Project complexity is high.

V4: The number of key organizations directly involved in the project

'Dumrny' Variables: V4-1: The number of key organizations is less thankqual to 5. V4-2: The number of key organizations is greater than 5 but less thanlequa1 to

15. V4-3: The number of key organizations is greater than 15 but less thdequa1 to

30. V4-4: The number of key organi;rations is greater than 30 but less than/equal to

50.

Page 161: An artificial neural network approach to assess project

V4-5: The number of key organizations is greater than 50.

V5: The percentage of engineering design completeness to the entire project engineering design at the start of field construction

V6: The cost expended on the front end planning of the project

V7: The cost expended on the detailed engineering of the project

V8: Project priori@ on quality

V9: Project priority on safety

V10: Project priority on schedule

V 1 1 : Project priority on cost

V12: The numbea of extemal permits/Approvals required in total for the project

'Dummy' Variables: V12-1: The number of permits/approvals is less thadequal to 5. V 12-2: The number of perrnits/approvais is greater than 5 but less

thdequai to 10. V12-3: The number of permitslapprovals is greater than 10 but less

thadequal to 1 5. V12-4: The number of permits/approvals is greater than 15 but less

thanfequal to 20. V12-5: The number of permits/approvals is greater than 20.

V13: The extent to which the project scope was defined when starting detailed engineering

V 14: Having experience in similar projects

'Dummy' Variables: V14-1: The project team has experience in sllnilar projects. Vl4-2: The project team has no experience in similar projects.

VI 5 : Years of experience that the project management leader(s) has had with projects

V16: Type of contract

Page 162: An artificial neural network approach to assess project

'Dummy' Variables: Vl6-1: Contract type is stipulated price. V 16-2: Contract type is cost plus. Vl6-3: Contract type is unit pnce. V 16-4: Contract type is partnering or alliance. Vl6-5: Others.

V 17: Level of technological innovation of the project

' Dummy ' Variables: V17-1: The technology used in the project is established, very familiar to

your organization. V 17-2: The technology used in the project is established in the industry

but new to your organization. Vl7-3: The technology used in the project is new technology to the

industry/sector but established and used in other industry/sector. V 17-4: The technology used in the project is brand new technology

application to any indu-.

V 18: The percentage of construction activities af5ected by weather

'Durnmy' Variables: V18-1: The percentage of construction activities affected by weather is

less thadequal to 5%. V18-2: The percentage of construction activities affected by weather is

greater than 5% but less thadequal to 25%. VI-: The percentage of construction activities affected by weather is

greater than 25%.

Page 163: An artificial neural network approach to assess project

APPENDIX V. A SAMPLE OF A BATCH FILE FOR ANN TRAINING

An exarnple of a batch file for ANN training is listed below:

% "pipeline projects: total 36 samples. 32 samples for training 4 for testing. % 90% training sets=32; testing sew; Logsig-pureline; % Input variables: 10(V 1 ,V2,V3 ,V4,VSYV6,V7,V 1 1 ,V 1 5,V 1 8);

Page 164: An artificial neural network approach to assess project
Page 165: An artificial neural network approach to assess project
Page 166: An artificial neural network approach to assess project

% The foîiowings are for calcdating Standard Deviation. C=pred-T-T-test; C l=C1; STD-test=std(C 1); D=pred-'IT-T; D 1 =Dl; STDTDtrain=std(D 1);

% The foiiowings are for calculating R squre(corre1ation coefficient). %training set: for i=I:32

cost-train(i)=T(l ,i);

Page 167: An artificial neural network approach to assess project

costpred(i)=pred-TT(1 ,i); tirne_train(i)=T(2, i); timegred(i)=pred_TT(2,i);

end Rcost~train=co~coef(cost~train~costpred); Rtime-train=corrcoe f(time-train,timem;

%Testhg set: for i=1:4

cost-t est (i)=T-test ( i,i); costpredT(i)=pred-T(1 ,i); time_test(i)=T_test(2,i); &nejredT(i)=pred-T(2,i);

end Rcost~test=corrcoef(cost~test,cost~redT); Rtime~test=corrcoef(time_test,tirne~redT);

save out corn STD-test STD-train Rcost-train Rcost-test Rtimctest

Page 168: An artificial neural network approach to assess project

l MAGE EVALUATIO N TEST TARGET (QA-3)

APPLIED IWGE . lnc 1653 East Main Street -

=A Rochester. NY 14609 USA ,=-A Phone: 71 6/4l8~93oO -- - Faxr 716t288-5989