HEURISTIC AND EXACT METHODS FOR THE LARGE-SCALE DISCRETE

TIME-COST TRADE-OFF PROBLEMS

A THESIS SUBMITTED TO
THE GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES
OF
MIDDLE EAST TECHNICAL UNIVERSITY

BY

SAMAN AMINBAKHSH

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR
THE DEGREE OF DOCTOR OF PHILOSOPHY
IN
CIVIL ENGINEERING

MAY 2018
 Approval of the thesis:

HEURISTIC AND EXACT METHODS FOR THE LARGE-SCALE

DISCRETE TIME-COST TRADE-OFF PROBLEMS

submitted by SAMAN AMINBAKHSH in partial fulfillment of the requirements

for the degree of Doctor of Philosophy in Civil Engineering Department, Middle
East Technical University by,

Prof. Dr. Halil Kalipçilar

Dean, Graduate School of Natural and Applied Sciences

Prof. Dr. İsmail Özgür Yaman

Head of Department, Civil Engineering

Prof. Dr. Rifat Sönmez

Supervisor, Civil Engineering Dept., METU

Assoc. Prof. Dr. S. Tankut Atan

Co-Supervisor, Industrial Engineering Dept.,
Işık University

Examining Committee Members:

Prof. Dr. M. Talat Birgönül

Civil Engineering Dept., METU

Prof. Dr. Rifat Sönmez

Civil Engineering Dept., METU

Assoc. Prof. Dr. Selçuk Kürşat İşleyen

Industrial Engineering Dept., Gazi University

Asst. Prof. Dr. Aslı Akçamete Güngör

Civil Engineering Dept., METU

Asst. Prof. Dr. Saeid Kazemzadeh Azad

Civil Engineering Dept., Atılım University

Date: 24.05.2018
I hereby declare that all information in this document has been obtained and
presented in accordance with academic rules and ethical conduct. I also
declare that, as required by these rules and conduct, I have fully cited and
referenced all material and results that are not original to this work.

Name, Last Name: Saman Aminbakhsh

Signature :

iv
 ABSTRACT

HEURISTIC AND EXACT METHODS FOR THE LARGE-SCALE DISCRETE

TIME-COST TRADE-OFF PROBLEMS

Aminbakhsh, Saman
Ph.D., Department of Civil Engineering
Supervisor: Prof. Dr. Rifat Sönmez
Co-Supervisor: Assoc. Prof. S. Tankut Atan

May 2018, 231 Pages

Construction industry necessitates formulating impeccable plans by decision

makers for securing optimal outcomes. Managers often face the challenge of
compromising between diverse and usually conflicting objectives. Particularly,
accurate decisions on the time and cost must be made in every construction project
since project success is chiefly related to these objectives. This is realized by
addressing the time-cost trade-off problem (TCTP) which is an optimization
problem and its objective is to identify the set of time-cost alternatives that provide
the optimal schedule(s). Due to discreteness of many resources in realistic projects,
discrete version of this problem (DTCTP) is of great practical relevance. The Pareto
front extension of DTCTP is a multi-objective optimization problem that facilities
preference articulation of decision makers by providing them with a set of mutually
non-dominated solutions of same quality. Due to the complex nature of DTCTP,
the literature on large-scale problems is virtually void; besides, most of the existing
methods do not suit actual practices and popular commercial planning software lack
tools for solution of DTCTP.

v
The main focus of this thesis relates to providing means for optimization of real-
life-scale Pareto oriented DTCTPs and it aims to contribute to both researchers and
practitioners by tightening the gap between the literature and the real-world
requirements of the projects. The results of the comparative studies reveal that the
proposed methods are successful for solving large-scale DTCTPs and provide the
management with a quantitative basis for decisions on selection of the proper
alternatives for the real-life-scale construction projects.

Keywords: Discrete Time-Cost Trade-Off Problem, Heuristic, Mixed-Integer

Linear Programming, Pareto Front, Particle Swarm Optimization

vi
 ÖZ

BÜYÜK ÖLÇEKLİ KESİKLİ ZAMAN-MALİYET ÖDÜNLEŞİM

PROBLEMLERİ İÇİN SEZGİSEL VE KESİN YÖNTEMLER

Aminbakhsh, Saman
Doktora, İnşaat Mühendisliği Bölümü
Tez Yöneticisi: Prof. Dr. Rifat Sönmez
Ortak Tez Yöneticisi: Doç. Dr. S. Tankut Atan

Mayıs 2018, 231 Sayfa

İnşaat sektöründeki artan rekabet koşulları, verimli ve başarılı sonuçların sağlaması

için karar vericilerin kusursuz planları oluşturmasını gerekli kılmaktadır. Proje
yöneticileri, proje hedeflerini sağlamak doğrultusunda çelişkili olabilen hedefler
arasında tercih yapmanın zorluğu ile yüzleşebilmektedir. Özellikle, proje başarısı
yüksek orantıda bu hedeflere bağlı olduğundan dolayı, her inşaat projesinde zaman
ve maliyet konusunda doğru kararlar alınması gerekmektedir. Bir optimizasyon
problemi olan zaman-maliyet ödünleşim problemi (ZMÖP)’nin amacı optimal
proje program(lar)’ı sağlayan zaman-maliyet alternatiflerini bulmaktır. Yapım
projelerinde birçok kaynağın kesikli olması nedeni ile bu problemin kesikli
versiyonu (KZMÖP) pratik açıdan büyük önem taşımaktadır. Çok amaçlı
optimizasyon problemi olan KZMÖP Pareto eğrisi, aynı kalitede olan ve domine
edilmeyen bir dizi çözüm sunarak karar vericilerin tercihlerinin belirlenmesine
olanak sağlamaktadır. KZMÖP'nin karmaşık yapısından dolayı, literatürde büyük
ölçekli problemler ile ilgili önemli boşluklar bulunmaktadır; ayrıca, mevcut
yöntemlerin birçoğu büyük ölçekli uygulamalar için uygun değildir ve yaygın

vii
olarak kullanılan paket planlama programları KZMÖP’n çözümüne yönelik hiçbir
alternatif sunmamaktadır.

Bu tezin temel odak noktası gerçek hayat ölçeğindeki KZMÖP'lerin çözümü için
Pareto optimizasyonuna yönelik yöntemler geliştirmek, ve literatür ile inşaat
projeleri gereksinimleri arasındaki boşluğu azaltarak araştırmacılara ve
uygulamacılara katkıda bulunmaktır. Karşılaştırmalı sonuçlar, bu tezde önerilen
yöntemlerin büyük ölçekli KZMÖP'lerin çözümünde başarılı olduklarını
göstermektedir. Geliştirilen yöntemler proje yöneticilerine büyük ölçekli projelerde
uygun alternatiflerin seçimine ilişkin niceliksel bir temel yöntem sağlamaktadır.

Anahtar Kelimeler: Karışık Doğrusal Tamsayılı Programlama, Kesitli Zaman-

Maliyet Ödünleşim Problemi, Sezgisel, Kuş Sürüsü Algoritması, Pareto Eğrisi

viii
 In ever loving memory of my dear father
In dedication to my mother and my brother with love and eternal appreciation

ix
 ACKNOWLEDGEMENTS

Moments of my Ph.D. journey have been shared with several special people without
the support and encouragement of whom the completion of this thesis would not
have been possible. Although it is impossible to thank every single person involved
in this journey, I would like to take this opportunity to show my gratitude to those
who have assisted me in a myriad of ways and to whom I am greatly indebted.

It has been a great privilege to spend several years in the civil engineering
department of the Middle East Technical University; members of which will always
remain dear to me. I would first like to express my heartfelt thanks to my supervisor,
Prof. Dr. Rifat Sönmez. A more supportive and considerate supervisor I could not
have asked for. His willingness to offer me so much of his time and intellect is the
major reason this thesis was completed; thus, I attribute the level of my Ph.D.
degree to his invaluable encouragement and guidance. Moreover, I would like to
extend my sincere thanks to my co-supervisor, Assoc. Prof. Dr. S. Tankut Atan, for
his continuous support and valuable comments. I would also like to pay my deepest
gratitude to my thesis monitoring committee members, Prof. Dr. M. Talat Birgönül
and Assoc. Prof. Dr. Selçuk Kürşat İşleyen, who made my research successful and
assisted me at every point to cherish my goal. Also, I would like to present my
special thanks to the examining committee members including Asst. Prof. Dr. Aslı
Akçamete Güngör and Asst. Prof. Dr. Saeid Kazemzadeh Azad for their valuable
feedbacks.

This thesis study was partially funded by a grant from the Scientific and
Technological Research Council of Turkey (TÜBİTAK), Grant No. 213M253.
TÜBİTAK’s support is gratefully acknowledged.

x
Words cannot express how grateful I am to my family for their relentless
encouragement, support, and patience during the years it took me to finish my Ph.D.
degree. I salute them all for the selfless love, care, pain and sacrifice they did to
shape my life. This thesis is heartily dedicated to my late father, Prof. Dr.
Mohammad Aminbakhsh, a brilliant scholar, a noble person, and a wonderful role
model for me who took the lead to heaven before the completion of this thesis. To
my dearest darling mother, Dr. Simin Nahaei, who always showed me the value of
education and discipline and to whom I owe my life for her constant love and
blessings. It is impossible to put into words everything I appreciate about her and
to acknowledge all the sacrifices that she has made on my behalf. To my gorgeous
friend and to whom I randomly call my brother, Sina Aminbakhsh. Words do not
suffice to express all the wisdom, love, and support my one and only brother has
given me at every stage of my life. I take pride in acknowledging his unconditional
and continuous support, both spiritually and materially, that has always been a great
source of motivation. To my angelic late grandmother whose thoughts for me have
resulted in this achievement and without her affection, loving upbringing, and
nurturing my dreams of excelling in education would have remained mere dreams.
My acknowledgement would be incomplete without thanking my uncles, Dr.
Mehrdad Nahaei and Prof. Dr. Mohammadreza Nahaei, who have kept me going
on my path to success in whatever manner possible and for ensuring that good times
kept flowing.

I take this opportunity to show my greatest appreciation and to offer my sincerest

gratitude with all my heart to Dr. Leili Nabdel for helping me get through the
difficult times, for all the fun times and for all the crazy and good memories. I can
barely find the words to thank her for all the emotional support, camaraderie, and
caring she provided. I am also especially grateful to two special ladies, Dr. Haleh
Mortazavi and Dr. Sahra Shakouri; they have cherished with me every great
moment and supported me whenever I needed it.

xi
I have great pleasure in acknowledging my gratitude to my wonderful colleagues
and fellow research scholars at METU, Ali Can Tatar, Dr. Babak Rahmani, Emad
Rezvankhah, Dr. Mahdi Abbasi-Iranagh, Mert Bilir, Dr. Saeed Kamali, and Dr.
Sahra Mohammadi, in propelling me on the course of this thesis and for sharing
their experiences.

Last but not least, I would like to express my thankfulness to my dearest friends,
Dr. Amir Fadaei, Dr. Arsham Atashi, Dr. Mahdi Mahyar, and Shima Ebrahimi, for
their precious friendship, and for all the academic and non-academic supports they
provided through all these years. I would also like to show my greatest appreciation
to the members of our so-called Joint-Venture, Dr. Handan Gündoğan, Dr. Murat
Ayhan, and Özgür Dedekargınoğlu, who incented me to strive towards my goal.

xii
 TABLE OF CONTENTS

ABSTRACT ............................................................................................................ v 
ÖZ ......................................................................................................................... vii 
ACKNOWLEDGEMENTS .................................................................................... x 
TABLE OF CONTENTS ..................................................................................... xiii 
LIST OF TABLES ............................................................................................... xvi 
LIST OF FIGURES ............................................................................................. xix 
LIST OF ABBREVIATIONS .............................................................................. xxi 
CHAPTERS ............................................................................................................ 1 
1. INTRODUCTION ........................................................................................... 1 

1.1. Problem Description and Mathematical Model ........................................ 6 

1.1.1. Assumptions ....................................................................................... 6 

1.1.2. Mathematical Model .......................................................................... 7 

1.2. Scope and Objectives ................................................................................ 8 

1.2.1. Scope .................................................................................................. 8 

1.2.2. Objectives........................................................................................... 8 

1.3. Research Methodology ............................................................................. 9 

2. LITERATURE REVIEW .............................................................................. 13 

2.1. Critical Path Method (CPM) ................................................................... 13 

2.2. Time-Cost Trade-off Problem (TCTP) ................................................... 15 
2.3. Exact methods for TCTP ........................................................................ 22 
2.4. Heuristic methods for TCTP ................................................................... 27 
2.5. Meta-heuristic methods for TCTP .......................................................... 30

xiii
3. DISCRETE PARTICLE SWARM OPTIMIZATION METHOD FOR DTCTP
........................................................................................................................... 59 

3.1. Particle Swarm Optimization (PSO) ....................................................... 59 

3.2. Siemens Approximation Method (SAM) ................................................ 60 
3.3. Discrete Particle Swarm Optimization Method (DPSO)......................... 61 

3.3.1. Case Example ................................................................................... 65 

3.4. Computational Experiments of DPSO .................................................... 70 

3.4.1. Parameter Configuration of DPSO ................................................... 70 

3.4.2. Small-Scale Benchmark Problems ................................................... 71 
3.4.3. Medium-Scale Benchmark Problems ............................................... 74 
3.4.4. Large-Scale Benchmark Problems ................................................... 75 
3.4.5. New Sets of Instances ....................................................................... 76 

4. PARETO FRONT PARTICLE SWARM OPTIMIZATION METHOD FOR

DTCTP ............................................................................................................... 83 

4.1. Pareto Optimality .................................................................................... 84 

4.2. Simplified Heuristic ................................................................................ 85 

4.2.1. Case Example ................................................................................... 86 

4.3. Pareto front Particle Swarm Optimizer (PFPSO).................................... 88 

4.4. Computational Experiments of PFPSO ................................................... 96 

4.4.1. Parameter Configuration of PFPSO ................................................. 96 

4.4.2. Small-Scale Benchmark Problems ................................................... 97 
4.4.3. Medium-Scale Benchmark Problem............................................... 105 
4.4.4. Large-Scale Benchmark Problems ................................................. 108 

5. COST-SLOPE HEURISTIC METHOD FOR DTCTP ............................... 113 

5.1. Cost-Slope Heuristic ............................................................................. 115 

5.1.1. Network Reduction Techniques ..................................................... 116 

5.1.1.1. Serial Merging Technique....................................................... 116

xiv
 5.1.1.2. Parallel Merging Technique ................................................... 119

5.1.2. Partial-CPM Calculator .................................................................. 121 

5.1.3. CS-Heuristic for Deadline DTCTP ................................................ 124 
5.1.4. CS-Heuristic for Pareto front DTCTP ........................................... 133 

5.2. Computational Experiments of CS-Heuristic ....................................... 142 

5.2.1. Generation of New Sets of Instances ............................................. 142 

5.2.2. Performance Indices ....................................................................... 147 
5.2.3. Mixed-Integer Linear Programming Technique ............................ 153 
5.2.4. Cost Minimization and Deadline DTCTPs .................................... 161 

5.2.4.1. Small-Scale Benchmark Problems ......................................... 161

5.2.4.2. Medium-Scale Benchmark Problems ..................................... 162
5.2.4.3. Large-Scale Benchmark Problems ......................................... 163

5.2.5. Pareto front DTCTP ....................................................................... 167 

5.2.5.1. Small-Scale Benchmark Problems ......................................... 167

5.2.5.2. Medium-Scale Benchmark Problem....................................... 169
5.2.5.3. Large-Scale Benchmark Problems ......................................... 170
5.2.5.4. New Sets of Instances ............................................................. 172
5.2.5.5. Case-Problems ........................................................................ 188

6. INTEGRATION OF THE PROPOSED METHODS INTO MICROSOFT

PROJECT ........................................................................................................ 191 
7. CONCLUSIONS ......................................................................................... 199 
REFERENCES.................................................................................................... 209 
CURRICULUM VITAE ..................................................................................... 227 

xv
 LIST OF TABLES

TABLES

Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of
TCTP. ........................................................................................................... 43 
Table 3.1 – Parameter configuration of the DPSO. ............................................... 70 
Table 3.2 – Performance of DPSO for problem 18a. ............................................ 72 
Table 3.3 – Performance of DPSO for problem 18b. ............................................ 72 
Table 3.4 – Performance of DPSO for problem 18c. ............................................ 73 
Table 3.5 – Performance of DPSO for problems 63a and 63b. ............................. 75 
Table 3.6 – Performance of DPSO for problems 630a and 630b. ......................... 76 
Table 3.7 – Performance of DPSO for 200-activity instances. ............................. 79 
Table 3.8 – Performance of DPSO for 500-activity instances. ............................. 80 
Table 4.1 – Candidate solutions found by simplified heuristic. ............................ 87 
Table 4.2 – Cost-slopes of crash modes. ............................................................... 87 
Table 4.3 – Parameter configuration of the PFPSO. ............................................. 97 
Table 4.4 – Performance comparison of PFPSO for small-scale problems. ......... 99 
Table 4.5 – Comparison of 19 non-dominated solutions for problem 18d............ 99 
Table 4.6 – Comparison of four non-dominated solutions for problem 18g. ...... 101 
Table 4.7 - Performance of PFPSO for problem 18d. ......................................... 102 
Table 4.8 – Performance of PFPSO for problem 18e. ......................................... 103 
Table 4.9 – Performance of PFPSO for problem 18f. ......................................... 104 
Table 4.10 – Performance of PFPSO for problem 18g. ...................................... 105 
Table 4.11 – Comparison of the results for 180-activity problem....................... 107 
Table 4.12 – Performance of PFPSO for 180-activity problem. ......................... 108

xvi
Table 4.13 – Comparison of the results for 360-activity problem. ..................... 109 
Table 4.14 – Non-dominated solutions of PFPSO for 360-activity problem. ..... 109 
Table 4.15 – Comparison of the results for 360-activity problem. ..................... 110 
Table 4.16 – Non-dominated solutions of PFPSO for 720-activity problem. ..... 110 
Table 5.1 – Candidate solutions found by CS-Heuristic for deadline DTCTP. .. 131 
Table 5.2 – Cost-slopes, DDiff s and CDiff s of crash modes........................... 132 
Table 5.3 – Candidate solutions found by CS-Heuristic for deadline DTCTP. .. 140 
Table 5.4 – Complexity of the generated instances. ........................................... 147 
Table 5.5 – Hypothetical Pareto fronts. .............................................................. 150 
Table 5.6 – Percentage of RanGen2 instances optimally solved for deadline
DTCTP....................................................................................................... 157 
Table 5.7 – Average CPU time of MILP for RanGen2 instances solved for deadline
DTCTP....................................................................................................... 157 
Table 5.8 – Percentage of RanGen2 instances optimally solved for Pareto front
DTCTP....................................................................................................... 158 
Table 5.9 – Average CPU time of MILP for RanGen2 instances solved for Pareto
front DTCTP. ............................................................................................. 158 
Table 5.10 – Performance of CS-Heuristic for problem 18a. ............................. 162 
Table 5.11 – Performance of CS-Heuristic for problem 18b. ............................. 162 
Table 5.12 – Performance of CS-Heuristic for problem 18c. ............................. 162 
Table 5.13 – Performance of CS-Heuristic for problems 63a and 63b. .............. 163 
Table 5.14 – Performance of CS-Heuristic for problems 630a and 630b. .......... 164 
Table 5.15 – Performance of CS-Heuristic for problems 1800a and 1800b. ...... 165 
Table 5.16 – Performance of CS-Heuristic for problems 3150a and 3150b. ...... 165 
Table 5.17 – Performance of CS-Heuristic for problems 6300a and 6300b. ...... 166 
Table 5.18 – Performance comparison of CS-Heuristic for small-scale problems.
................................................................................................................... 168 
Table 5.19 – Comparison of the results for 180-activity problem. ..................... 169 
Table 5.20 – Comparison of the results for 360-activity problem. ..................... 170 
Table 5.21 – Comparison of the results for 720-activity problem. ..................... 171 

xvii
Table 5.22 – Comparison of ONVG values. ....................................................... 174 
Table 5.23 – Comparison of ND pct values......................................................... 175 

Table 5.24 – Comparison of APD values. ........................................................... 176 

Table 5.25 – Comparison of APD bin values. ...................................................... 177 
Table 5.26 – Comparison of HR values. ............................................................. 178 
Table 5.27 – Comparison of CPU times. ............................................................. 179 
Table 5.28 – Comparison of the results for the first case problem. ..................... 188 
Table 5.29 – Comparison of the results for the second case problem. ................ 189 

xviii
 LIST OF FIGURES

FIGURES

Figure 2.1 – Nonlinear progression of direct cost resulted from schedule

acceleration. ................................................................................................. 18 
Figure 2.2 – Linear decline of indirect cost resulted from schedule acceleration. 19 
Figure 2.3 – Variation of total cost resulted from schedule acceleration.............. 19 
Figure 3.1 – Case Example. .................................................................................. 66 
Figure 3.2 – Probability matrix. ............................................................................ 66 
Figure 3.3 – Position matrix. ................................................................................. 67 
Figure 3.4 – Flowchart of the proposed discrete PSO algorithm. ......................... 68 
Figure 3.5 – Pseudo-code of the proposed discrete PSO algorithm. ..................... 69 
Figure 3.6 – ProGen/Max Interface. ..................................................................... 77 
Figure 4.1 – Schematic diagram of PFPSO at time-step t . .................................. 92 
Figure 4.2 – Schematic diagram of PFPSO at time-step t +1. .............................. 92 
Figure 4.3 – Flowchart of the proposed Pareto front PSO algorithm. .................. 94 
Figure 4.4 – Pseudo-code of the proposed Pareto front PSO algorithm. .............. 95 
Figure 4.5 - Network diagram of the core problem for medium-scale and large-scale
problems. ................................................................................................... 106 
Figure 4.6 – Pareto fronts located by PFPSO for 180, 360, and 720-activity
problems. ................................................................................................... 111 
Figure 5.1 – Serial merge applied to the Case Example. .................................... 118 
Figure 5.2 – Parallel merge applied to the Case Example. ................................. 120 
Figure 5.3 – PERT chart of all-crashed schedule for the Case Example. ........... 123 
Figure 5.4 – PERT chart of the updated schedule for the Case Example. .......... 124

xix
Figure 5.5 – Flowchart of the proposed parallel merging technique. .................. 128
Figure 5.6 – Flowchart of the proposed serial merging technique. ..................... 128 
Figure 5.7 – Flowchart of the uncrash free-float module of the proposed CS-
Heuristic. .................................................................................................... 129 
Figure 5.8 – Pseudo-code of the proposed CS-Heuristic for deadline DTCTP... 129 
Figure 5.9 – Flowchart of the proposed CS-Heuristic for Pareto front DTCTP. 137 
Figure 5.10 – Pseudo-code of the proposed CS-Heuristic for Pareto front DTCTP.
.................................................................................................................... 138 
Figure 5.11 – RanGen2 Interface. ....................................................................... 143 
Figure 5.12 – Hypothetical Hypervolume comparison for PF a . ........................ 151 
Figure 5.13 – Hypothetical Hypervolume comparison for PFb . ........................ 152 
Figure 5.14 – Serial merge applied to the Case Example. ................................... 154 
Figure 6.1 – User Interface of the Microsoft Project Add-in. ............................. 192 
Figure 6.2 – Defining time-cost alternatives for Microsoft Project Add-in. ....... 193 
Figure 6.3 – Project Details window of Microsoft Project Add-in. .................... 194 
Figure 6.4 – Optimal solution window of Microsoft Project Add-in. ................. 194 
Figure 6.5 – Optimal schedule generated by Microsoft Project Add-in. ............. 195 
Figure 6.6 – Pareto front solutions window of Microsoft Project Add-in. .......... 196 
Figure 6.7 – Selection of a non-dominated solution achieved by Microsoft Project
Add-in. ....................................................................................................... 196 
Figure 6.8 – Schedule for the selected Pareto solution generated by Microsoft
Project Add-in. ........................................................................................... 197 

xx
 LIST OF ABBREVIATIONS

ACO Ant Colony Optimization

Acts. Activities
AHP Analytic Hierarchy Process
AoA Activity on Arrow
AoN Activity on Node
APD Average Percent Deviation
Avg. Average
B&B Branch and Bound
BD Benders Decomposition
BT Budget
CA Cash Availability
CDiff Cost Difference
CPM Critical Path Method
CPU Central Processing Unit
CS Cost-Slope
CS-H Cost-Slope Heuristic
DDiff Duration Difference
DDR3 Double Data Rate type Three
DL Deadline
DP Dynamic Programming
DPSO Discrete Particle Swarm Optimization
DTCTP Discrete Time-Cost Trade-Off Problem
Dur. Duration
EM Electromagnetic
ER Error Ratio

xxi
GA Genetic Algorithm
GD Generational Distance
GHz Gigahertz
HR Hyperarea Ratio
HS Harmony Search
HV Hypervolume
IC Indirect Cost
ID Identity
IGD Inverted Generational Distance
IP Integer Programming
KZMOP Kesikli Zaman-Maliyet Ödünleşim Problemi
LP Linear Programming
MA Memetic Algorithm
MAWA Modified Adaptive Weight Approach
Max. Maximum
MHz Megahertz
Min. Minimum
MIP Mixed-Integer Programming
MILP Mixed-Integer Linear Programming
NP-Hard Non-deterministic Polynomial-time Hard
NPV Net Present Value
ONVG Overall Non-dominated Vector Generation
PERT Program Evaluation and Review Technique
PF Pareto Front
PFPSO Pareto Front Particle Swarm Optimizer
PM Parallel Merge
Pred. Predecessor
PSO Particle Swarm Optimization
QSA Quantum Simulated Annealing
RAM Random Access Memory

xxii
RV Range Variance
SA Simulated Annealing
SAM Siemens Approximation Method
SFL Shuffled Frog Leaping
SM Serial Merge
Sol. Solution
St. Stochastic
Succ. Successor
TCT Time-Cost Trade-Off
TCTP Time-Cost Trade-Off Problem
TP True Pareto
UCS Uncrash Cost-Slope
UD Uniform Distribution
UF Unified Front
UFF Uncrash Free-Float
UI User Interface
USD United States Dollar
VBA Visual Basic for Applications
ZMÖP Zaman-Maliyet Ödünleşim Problemi

xxiii
xxiv
 CHAPTER 1

INTRODUCTION

Construction industry necessitates formulating impeccable plans by decision

makers and construction planners for securing optimal outcomes. Managers often
face the challenge of compromising between diverse and usually conflicting
objectives of each project. Particularly, accurate decisions on the time, cost, quality,
and resource utilization of a project are essential prerequisites of an exhaustive plan.
Focusing on the noted aspects coupled with consideration of other impartible
components of prosperous project deliveries – such as provision of safety and
productivity upkeeps – further narrow the field for the project management team.
Of the specified aspects, time and cost are regarded as two of the most significant,
yet counteracting factors that need to be considered in every construction project
since project success is chiefly related to these objectives. A key process for
achieving the anticipated resolutions is preparation of a schedule. Logical
relationships – known as the precedence constraints, lag times, working calendars,
resource requirements, and contingency plans are some of the many essential
concerns for preparing a decent schedule. Idle labor and equipment, delays,
dissatisfied customers, disputes, bad reputation are some of the adverse outcomes
that might arise in the absence of an adequate schedule. More importantly, in capital
intensive construction projects, major financial losses might occur in the light of a
suboptimal schedule, or because of small deviations from an optimal solution to the
scheduling problem. In contrast, a flawless schedule can obviate occurrence of such
problems which ultimately results in completion of a project on or ahead of time.

1
In project scheduling, critical path method (CPM) is used to determine the
completion time of a project by calculating the longest sequence of the activities in
the project network which is known as the critical path. Critical path plays a crucial
role in planning of a project since any delay in realization of an activity on this path
results in overall project delay. Classical network analyses like CPM merely
incorporates the time aspect of the projects. Such methods attempt to minimize
duration of the project without taking into account the availability of resources (both
nonrenewable and physical resources). Generally, it is desirable for the involved
parties to minimize the duration in order to finish a construction project ahead of a
prescribed completion deadline. Delay penalties can be avoided by finishing a
project earlier; besides, managers accelerate a schedule due to many factors like
improving cash-flow, avoiding unfavorable climate conditions, early
commissioning, earning early-completion incentive, starting another project earlier,
and mainly in the interest of increasing profit margins. Known as crashing, any
reduction in project duration is facilitated by compression or acceleration of the
project schedule. Decision makers speed up the project by optimally crashing
selected critical activities that levy least additional cost. Crashing a project schedule
is usually facilitated by provisioning resource overloads – i.e. allocating additional
manpower and machinery resources or recruiting subcontractors – or by
implementing alternative speedy construction techniques. Although shortening the
project duration below its normal level enables reducing indirect cost and avoiding
potential delay penalties, all the schedule expedition techniques add to the total cost
of a project. Obviously, it is because of the fact that the aforementioned crashing
approaches require greater financial expenditures known as the direct costs. This
trade-off between time and non-renewable resources (e.g., money) of the project is
known as the time-cost trade-off problem which is abridged as TCTP and is one of
the most important and applicable research areas in project management; especially
because of the prevailing emphasis on time-based completion of the construction
firms. The objective of general time-cost trade-off problem is to identify the set of
time-cost modes (alternatives) that will provide the optimal schedule under certain

2
conditions. It is a problem solving and decision-making science which provides the
management with a quantitative basis for decisions on selection of the proper
alternatives. The importance of TCTP has been recognized since development of
the CPM (De, Dunne, Ghosh, and Wells, 1995). Starting since early 1960’s, several
different researches have been conducted to address this problem in the literature.

In the literature, three types of TCTP have been commonly studied; the deadline
problem, the budget problem, and the Pareto front problem. The objective of the
deadline problem is to determine the set of time-cost alternatives that will minimize
the total cost – including direct and indirect costs, penalties and bonuses – for a
given project deadline. The budget problem aims to identify the time-cost
alternatives to minimize the project duration without exceeding the budget. The
Pareto front problem which is also the chief focus of this thesis, is a multi-objective
optimization problem and involves determination of the non-dominated time-cost
profile over the set of feasible project durations to generate Pareto fronts of the
problems (Vanhoucke and Debels, 2007). Pareto front optimization of TCT
problems is recognized to be the ultimate resolution of TCTP analyses (e.g., Zheng,
Ng, and Kumaraswamy, 2005; Yang, 2007b; Eshtehardian, Afshar, and Abbasnia,
2008; Aminbakhsh and Sonmez, 2017). The importance of Pareto front is
emphasized since preferences of decision makers can be articulated by providing
them with a set of mutually non-dominated solutions of same quality instead of a
single optimal solution. This way decision makers can choose the best solutions
based on their own concerns.

Over the years, numerous studies have been conducted to model the time-cost
relationship of the project. Early research on TCTP assumed the relation between
time and cost to be continues (Kelley and Walker, 1959; Fulkerson, 1961, Siemens,
1971; Goyal, 1975). In recent years there has been increased attention toward the
discrete version of the problem due to its great practical relevance. This
consideration is imperative to TCT analyses since in practice many resources (e.g.,

3
workforce, equipment) are available in discrete units; in addition, time-cost
function of any type can be estimated by discrete functions. Being the main focus
of this thesis, the discrete version of TCTP considers discrete sets of time-cost
options for the activities and it is known as the discrete time-cost trade-off problem
– abridged as DTCTP – in the literature.

All of the three extensions of DTCTP that were mentioned earlier are Non-
deterministic polynomial-time hard (NP-hard) problems in the strong sense (De,
Dunne, Ghosh, and Wells, 1997). That is to say, solution of DTCTP requires
concurrent searches over the solution space and that any escalation in the size of
the project (any growth in the number of activities, modes, or both simultaneously)
contributes to the significantly higher computational burden. Due to this fact,
exhaustive enumeration is incapable of providing an efficient and convenient mean
for DTCT analyses; therefore, researchers have come up with numerous
optimization techniques for solution of DTCT problems. The methods proposed for
DTCTP could be categorized into three areas: exact methods, heuristics, and meta-
heuristics. Traditionally, solution of DTCTP has been modelled by mathematical
programming – known as the exact methods – such as linear programming (Kelley,
1961), dynamic programming (Butcher, 1967), hybrid LP/IP programming (Liu,
Burns, and Feng, 1995), and branch-and-bound algorithm (Demeulemeester, De
Reyck, Foubert, Herroelen, and Vanhoucke, 1998). The literature on heuristics for
DTCTP is limited to the methods proposed by Fondahl (1961), Siemens (1971),
Goyal (1975), Moselhi (1993), and Bettemir and Birgonul (2017). Evolutionary
algorithms are among the meta-heuristics practiced in solution of DTCT problems.
Meta-heuristics include genetic algorithm (GA) (Feng, Liu, and Burns, 1997;
Zheng, Ng, and Kumaraswamy, 2004), ant colony optimization (ACO) (Xiong and
Kuang, 2008), and particle swarm optimization (PSO) (Yang, 2007b).

Generally, exact methods are incapable of obtaining optimal solutions for large-
scale problems efficiently. The proposed exact algorithms, while requiring massive

4
computational resources, are more difficult to implement and are prone to being
stagnated in local optima in non-convex solution spaces (De et al., 1995; Feng et
al., 1997; Eshtehardian et al., 2008; Afshar, Ziaraty, Kaveh, and Sharifi, 2009). Due
to the alleged drawbacks of the mathematical programming models, researchers
have turned their interests toward alternative optimization methods. Of the
alternative optimization techniques, existing studies providing heuristic algorithms
acknowledge that they are problem dependent and cannot handle large-scale
problems efficiently (Siemens, 1971). Most of the present heuristics assume merely
linear time-cost functions and they fail to solve the Pareto front problem (Feng et
al., 1997; Zheng et al., 2005). In addition, the existing meta-heuristics inability to
escape from local optima is observed as their main deficiency (Zheng et al., 2005;
Sonmez and Bettemir, 2012). In fact, it is worth mentioning that although a large
body of the literature has hitherto been dedicated to development of optimization
methods for DTCTP, only some of these methods are used in real-life practices.
That is largely resulting from the fact that they do not suit actual practices and that
major domain of the literature focus on proving applicability of various
optimization models rather than providing means for optimization of real
construction projects. Accordingly, it might be perceived that there exists a gap
between the theoretical achievements of scholars and practical applications of
professionals in the field of construction. To expand on this, majority of the past
research have used problems including up to only eighteen activities in
computational experiments. Very few of the existing methods can be applied to
optimization of real-life construction projects which typically comprise more than
300 activities (Liberatore, Pollack-Johnson, and Smith, 2001). Furthermore, a few
methods that are tested for real-life-size large-scale problems, require enormous
computation time and resources thanks to the inherent complexity of solving
DTCTPs. Last but not least, despite the fact that any scientific decision support tool
would have a pivotal role in the decision-making process, none of the commercial
scheduling software packages (e.g., Microsoft Project, Primavera) include tools or
modules for TCT analyses of the scheduling problems.

5
As mentioned earlier, DTCTP has been unraveled for relatively small instances and
the literature on large-scale DTCTP is virtually void thanks to the complex nature
of the problem. It was also stated that the ultimate resolution and the most complex
extension of DTCTP is the Pareto front problem. Further, despite the dramatic
increase in computing speed of the modern computers, they might prove to be
insufficient for practical applications. Accordingly, an efficient optimization model
for tackling Pareto oriented optimization of real world DTCT problem is long
overdue. Recognizing this, the main focus of this thesis is set as Pareto oriented
optimization of large-scale DTCT problems and it aims to contribute to both
researchers and practitioners by tightening the gap between the literature and the
real-world requirements of the projects. In this respect, new models applicable in
real projects are developed which are believed that will suit the actual practices of
construction managers.

1.1. Problem Description and Mathematical Model

The objective of general DTCTP is to determine the optimal set of time-cost modes
that will minimize the sum of direct and indirect costs by taking into account the
liquidated damages and bonuses for a project.

1.1.1. Assumptions

The assumptions of the general DTCTP are as follows:

- There is a linear relation between the indirect costs and the project duration;
- The duration of project activities is a discrete, non-increasing function of the
amount of resources assigned to them (Vanhoucke and Debels, 2007);
- An activity can start after all of its predecessors are finished;
- Activities cannot be interrupted. Each activity is executed without interruption
from its start time to its finish time.

6
1.1.2. Mathematical Model

The general DTCTP can be formulated by modifying the formulation of De et al.

(1995) to include the indirect costs and the delay penalty as follows:

S m( j )
minimize   (dc jk x jk )  D  ic  T  dp (1.1)
j 1 k 1

subject to:

m( j )
 x jk  1 ,  j  {1,..., S} and  k  {1,..., m( j)} (1.2)
k 1

T  D  D dl (1.3)

m( j )
 d jk x jk  St j  St l ,  l Sc j and  j  {1,..., S} (1.4)
k 1

D  Ft S (1.5)

St 0  0 (1.6)

where; dc jk is the direct cost of mode k for activity j ; x jk is a binary (0/1)

variable and is set to 1 when activity j is undertaken with mode k ; ic is the daily
indirect cost; D is the project duration; D dl is the project deadline; T is the delay
amount of the project; dp is the daily delay penalty; db is the daily bonus; d jk is

the duration of mode k for activity j ; St j and Ft j are the start time and the finish

time for activity j , respectively; and Sc j is the set of immediate successors for

activity j .

7
Objective function of a general DTCTP is defined as Eq. (1.1) which attempts to
minimize the summation of the direct and the indirect costs, i.e., the total cost of
the project. The constraint defined as Eq. (1.2) secures selection of only a single
time-cost alternative for each of the activities. The project delay is calculated using
Eq. (1.3) by subtracting the imposed completion deadline from the total duration.
The precedence relationships are maintained by means of Eq. (1.4). Completion
time of the project is determined using Eq. (1.5) which ensures total duration not to
be smaller than the finish time of the final activity. Eq. (1.6) reflects the start time
of the initial activity which is set to be equal to zero.

1.2. Scope and Objectives

1.2.1. Scope

This study concerns development of an exact, a heuristic, and a meta-heuristic for

the deadline and Pareto front classes of DTCTP. The main focus of the study is the
Pareto front optimization of large-scale projects.

1.2.2. Objectives

The primary objective of this thesis is to design and develop state-of-the-art

methods that can achieve successful results for discrete time-cost optimization of
large-scale projects. For Pareto oriented optimization of large-scale projects, the
proposed method aims to provide a large number of non-dominated solutions with
minimal or no deviations from the optimal solutions in a short amount of
computation time.

8
1.3. Research Methodology

New models applicable in real projects are designed and developed to achieve the
research objectives. The new proposed models include Mixed-Integer Linear
Programming techniques that use Gurobi solver version 6.0.5, new Particle Swarm
Optimizers, and new Cost-Slope Heuristics. For all the proposed methods, two
variants have been designed and developed to address both the deadline and the
Pareto front classes of DTCT problems. Since, exact procedures are the only
methods guaranteeing optimality of the solutions and that heuristics and meta-
heuristics are incapable of securing the optimality of the solutions, the proposed
Mixed-Integer Linear technique is used in performance evaluation of the developed
heuristic and meta-heuristic methods. Exact solutions are used to validate the
accuracy of the results obtained by the proposed algorithms. The developed Mixed-
Integer Linear Programming technique incorporates an efficient Upper-bound
calculation for reducing the size of the solution space and a new merging technique
is also applied which exponentially decreases the scale of the practiced problems.
Resultantly, computation time of DTCTP analyses is significantly reduced in the
light of the implemented procedures. The proposed Particle Swarm Optimizer is
equipped with a unique semi-deterministic initialization technique and uses new
principles for Presentation and Position-updating of the particles. Regarded as the
chief contribution of this thesis, the novel Cost-Slope Heuristic engages unique
scientific and programmable rules which enjoys the fastness for solving DTCTP.
For the proposed Cost-Slope Heuristic, an original method for CPM calculations is
designed to accelerate the solutions process. Furthermore, similar to the Mixed-
Integer Linear Programming method, a merging technique is implemented to the
Cost-Slope Heuristic for the sake of reducing scale and computation cost of the
practiced problems.

This thesis study also presents integration of DTCTP optimization modules into
Microsoft Project – a widely used commercial planning software in the construction

9
industry – by means of an add-in which is capable of solving two variants of
DTCTP, namely, deadline and Pareto problems. The integrated modules include
both the proposed Particle Swarm Optimizer and the Cost-Slope Heuristic.
Benefiting from the presented add-in, users of Microsoft Project will readily be able
to visualize the optimized schedules for the practiced projects; hence, the proposed
methods are supposed to be more readily accepted and used by the parties to
construction projects. It is expected that these approaches might prove to be an
efficient and effective base for exerting this highly challenging problem.

In order for better evaluation of the capabilities of the proposed optimization models
and methods, new sets of DTCT problems have been generated by means of a
project instance generator ProGen/Max that includes multi-mode problems with up
to 500 activities; in addition, a random network generator RanGen2 is also used to
generate new sets of multi-mode DTCTPs including up to 990 activities, with more
complex networks and more realistic sets of time-cost alternatives. Along with the
systematically generated large-scale instances, benchmark problems – acquired
from the literature – and case projects are also simulated and solved using the
proposed methods and their performances are evaluated and compared against
existing optimization approaches according to a set of performance comparison
indices. All the proposed algorithms and performance evaluation procedures have
been implemented in C++ and C# programming languages using Microsoft Visual
Studio 2013. The proposed approaches engage CPM calculations for logical
relationships of type finish-to-start, considering no lags in between. The results of
the comparative studies are promising and reveal that the obtained solutions not
only are comparable, but also are better than previous approaches. Experiments also
attest to the efficacy of the proposed methods for successful solution of real-life-
size large-scale DTCT problems within only moderate computational effort.

The sequel of the thesis is structured as follows. Chapter 2 starts with an

introduction on CPM and TCTP, followed by a detailed review of the existing

10
research on TCTP in the construction management domain. Chapter 3 presents the
background and theoretical properties of PSO and SAM approaches; furthermore,
the details of the proposed DPSO method for cost minimization and deadline
extensions of DTCTP, as well as comparative studies on DPSO’s performance are
described in this chapter. In Chapter 4, followed by an introduction on the Pareto
optimality, the proposed PFPSO model for Pareto front extension of DTCTP is
explained along with the results of the computational experiments and performance
comparisons conducted for this method. Chapter 5 includes the major body of this
thesis study, which illustrates new techniques used in development of different
variants of CS-Heuristic and MILP methods. This chapter also covers results of the
inclusive performance measurements and comparative studies carried out over
benchmark problems and new sets of complex instances, for validation of all the
proposed approaches. Chapter 6 explains the methodology implemented to develop
a Microsoft Project add-in. Chapter 7 includes the concluding remarks and points
out some directions for the future research.

11
12
 CHAPTER 2

LITERATURE REVIEW

In this chapter, the principles of project scheduling are outlined. Insight is given
into one of the major classical network analyses, the Critical path method (CPM).
Different types of construction project expenses are also described in this chapter.
The time-cost trade-off problem (TCTP) is elaborated in addition to the existing
solution techniques proposed for this problem. Prospects of exact, heuristic, and
meta-heuristic methods for practical solution of large-scale TCT problem are
presented.

2.1. Critical Path Method (CPM)

Scheduling is defined by Mubarak (2010) as making judgments on timing and

sequence of the work-packages, by means of which, the overall completion date of
a project is determined. Preparing schedules for projects is facilitated by conducting
network analysis (Lock, 2007). CPM is widely used for network analysis and
scheduling of construction projects. This method is capable of calculating the
overall completion date of a project by computing the longest path in the project
network. In CPM method the longest path in the network show the shortest amount
of time required to complete the project (Kerzner, 2009).

CPM calculations take place based on the duration and precedence constraints of
the activities. Network illustrations – known as logic diagrams – are used as
medium of tracking the critical paths of a project network. In general, the activity

13
networks of the projects are illustrated either by using the activity on arrow (AoA)
or the activity on node (AoN) notation systems. As the names imply, in the first
system (AoA) the activities are represented by arrows intersecting nodes which
resemble events; while, in the second system (AoN) activities are represented by
nodes and the logical relationships are traced by arrows. In this thesis, the activity
on node representation system is preferred due to the following reasons:

- They are more flexible and simple to construct;

- Most of the modern project management software support this network chart;
- It is easy to build Gantt-Chart based on the AoN network;
- This representation does not require dummy activities as the arrows represent
only the dependencies;
- It is easier to track parallel and serial paths in the network for merging purposes.

In CPM scheduling technique, there might be more than a single critical path of the
same length. The remainder of the paths are shorter in length, which contributes to
a certain degree of freedom in starting and finishing of the non-critical activities. In
contrast, a critical activity without a degree of freedom that its allowable delay is
equal to zero, must be started as soon as its predecessors are finished. This degree
of freedom, which is hailed as the float or slack time in the project scheduling
domain, can be calculated by evaluating the difference between an activity’s earliest
and latest dates (either start or finish dates). Slack time can be interpreted as the
amount an activity can be delayed without affecting the overall completion of the
project. Earliest and latest dates of the activities can be determined by implementing
forward-pass and backward-pass, respectively.

Consideration of a completion deadline is imperative for construction projects,

since, in practice majority of projects include a target date which is stipulated in the
contract by the client. Completion deadline of a project is generally the earliest
possible date decided by dint of the network analyses such as CPM; nevertheless,

14
the definite completion date might also be targeted on a later date. The main
resolution in time-limited projects is to secure completion of the project on a date,
no later than the prescribed deadline. In such projects, any projected resource over-
allocation might be dealt with hiring subcontract workforce or by making
alternative short-term resource provisions.

In this thesis, due to the practical relevance of the completion deadline in projects,
all the systematically generated instances are tackled by assuming a target
completion date. In addition, all the proposed methods are equipped with CPM
technique for calculation of the early and late dates and floats of the work-packages,
as well as the overall completion date of the practiced instances. Original to this
thesis, the proposed Cost-Slope Heuristic is also complemented with a unique
CPM-esque approach for faster network analyses.

2.2. Time-Cost Trade-off Problem (TCTP)

Of the specified aspects, time and cost are regarded as two of the most significant,
yet counteracting factors that need to be considered in every construction project
since project success is chiefly related to these objectives in today’s market-driven
economy. Generally, it is a desirable resolution for the involved parties to minimize
the duration in order to finish a construction project ahead of a prescribed
completion deadline. Unambiguously, finishing on or under the specified budget is
another favorable achievement; hence, simultaneous realization of the noted
objectives is sought after undeniably. For this purpose, decision makers evaluate
the cost per unit time (cost-Slope) as well as a feasible budget (cash-flow) region
for the project. Hence, one of the dominant prospects of the network analysis can
be concluded as finding a solution that not only satisfies the completion deadline,
but also has the lowest feasible total cost that resides within the feasible budget
boundaries. In network analysis, normal duration is defined as the time required to
complete the project under ordinary conditions without deliberate delay or

15
acceleration; further, normal cost is the amount that is required for completion of
the project within the normal duration. Despite the fact that it is possible to avoid
delay penalties by finishing a project earlier, managers accelerate a schedule due to
many other factors like improving cash-flow, avoiding unfavorable climate
conditions, early commissioning, earning early-completion incentive, starting
another project earlier, and mainly in the interest of increasing profit margins.
Usually, the contractor knows the expected mobilization date to another project; in
this regard, the contractor might esteem to accelerate the current project in favor of
supplying the required resources so that they can be allocated to the new project.

Accelerating a project schedule might be profitable only up to a certain level. This

is because of the fact that in practice alternatives of the activities are mutually
incomparable with convex relationships; that is, the faster it gets to execute an
activity, the more expensive it gets. The act of accelerating (compressing) or
crashing a project schedule literally means reducing the duration of a project.
Though, there is a fine line between these two approaches. Whilst both the
techniques aim at advancing the completion date of a project, accelerating does not
necessarily mean targeting to reach the least possible duration. It must be also
clarified that the possibility of schedule acceleration or availability of the crashing
alternatives for the activities of a project is highly affiliated with its typology. Those
with complex precedence constraints usually offer less flexibility regarding the
schedule acceleration compared with those with less strict logical relationships.

Decision makers speed up the project by optimally crashing selected critical

activities that levy least additional cost. Crashing a project schedule is usually
facilitated by provisioning resource overloads – i.e. allocating additional manpower
and machinery resources or recruiting subcontractors – or by implementing
alternative speedy construction techniques. Although shortening the project
duration below its normal level enables reducing indirect cost and avoiding
potential delay penalties, all the schedule expedition techniques add to the total cost

16
of a project. Obviously, it is because of the fact that the aforementioned crashing
approaches require greater financial expenditures known as the direct costs.

According to the Construction Industry Institute (1988), acceleration of a project

schedule can be facilitated by using more than 90 techniques. Mubarak (2010) has
summarized some of the more significant schedule acceleration methods as:

- Reviewing or evaluating the schedule for errors or imperfect precedence

relationships;
- Applying fast-tracking;
- Studying constructability and value engineering;
- Assigning over-time schedule or using shift-works;
- Setting incentives for the more productive work-forces;
- Increasing the size of the allocated work-forces;
- Employing more efficient construction techniques;
- Using materials with faster installation processes;
- Enhancing project management and supervision;
- Preventing communications breakdowns.

Since schedule acceleration remarkably influence both the direct and indirect costs
of a project, shedding some light on their differences is inevitable and necessary.
The main principle for distinguishing the direct expenses from the indirect costs
can be depicted as a direct cost item is directly associated with an explicit work
item; whereas, the indirect costs cannot be related to any specific task or a particular
project. Direct expenses may include labor, material, equipment, subcontractor,
machinery, and costs related to fees and permits. On the other hand, indirect costs
might include project overhead and general overhead expenditures that contain
salaries of the guard, cook, and office personnel as well as the energy costs.

17
It was already mentioned that in practice, the time-cost relation is convex among
the alternatives of an activity. Therefore, the more an activity accelerates, the more
the daily cost of acceleration increases. This is because of the fact that over-staffing
or working over-time notably decreases the productivity which in turn, increases
the ratio of unit cost per unit output. A realistic nonlinear time-direct-cost profile is
shown in Figure 2.1.

Direct Cost ($)

Duration (days)

Figure 2.1 – Nonlinear progression of direct cost resulted from schedule acceleration.

The indirect costs, on the other hand, consists of time-dependent and time-
independent types. Besides, they usually tend to increase faster at the initial stages
of the project and then to remain constant. Nevertheless, they are usually assumed
to be linearly comparative to the duration of the project for schedule acceleration
purposes. Likewise, for the sake of simplified cost computations, second order cost
components (e.g., insurance and bond payments) are usually not included in the
daily indirect expenses. As shown in Figure 2.2, the indirect cost is assumed to
decrease linearly in case of schedule acceleration.

18
 Indirect Cost ($)
Duration (days)

Figure 2.2 – Linear decline of indirect cost resulted from schedule acceleration.

As shown in Figure 2.3, the total cost profile of a project can be obtained by adding
up the direct and indirect costs together. Starting from the normal schedule – which
consists of least direct cost/largest duration alternatives – the total cost of the project
decreases as a result of acceleration till reaching an optimal point; after this point,
as the acceleration gets closer to the crash schedule – which consists of largest direct
cost/shortest duration alternatives – the total cost surges up at an increasing rate.
Total Cost ($)

Duration (days)

Figure 2.3 – Variation of total cost resulted from schedule acceleration.

This trade-off between time and cost of the project is known as the time-cost trade-
off problem which is abridged as TCTP and is one of the most important and
applicable research areas in project management. The objective of general time-
cost trade-off problem is to identify the set of time-cost modes (alternatives) that
will provide the optimal schedule under certain conditions. TCTP mainly attempt
to speed up the critical activities while relaxing non-critical ones (Siemens, 1971).

19
It is a problem solving and decision-making science which provides the
management with a quantitative basis for decisions on selection of the proper
alternatives. The importance of TCTP has been recognized since development of
the CPM (De et al., 1995).

In the literature, three types of TCTP has been commonly studied; the deadline
problem, the budget problem, and the Pareto front problem. The objective of the
deadline problem is to determine the set of time-cost alternatives that will minimize
the total cost – including direct and indirect costs, penalties and bonuses – for a
given project deadline. The budget problem aims to identify the time-cost
alternatives to minimize the project duration without exceeding the budget. The
Pareto front problem which is also the chief focus of this thesis, is a multi-objective
optimization problem and involves determination of the non-dominated time-cost
profile over the set of feasible project durations to generate Pareto fronts of the
problems (Vanhoucke and Debels, 2007). Pareto front optimization of TCT
problems is recognized to be the ultimate resolution of TCTP analyses (e.g., Zheng
et al., 2005; Yang, 2007b; Eshtehardian et al., 2008; Aminbakhsh and Sonmez,
2017). The importance of Pareto front is emphasized since preferences of decision
makers can be articulated by providing them with a set of mutually non-dominated
solutions of same quality instead of a single optimal solution. This way decision
makers can choose the best solutions based on their own concerns.

Over the years, numerous studies have been conducted to model the time-cost
relationship of the project. Early research on TCTP assumed the relation between
time and cost to be linear (Kelley and Walker, 1959; Fulkerson, 1961; Siemens,
1971; Goyal, 1975). Subsequently, the assumption of linearity was relaxed allowing
for consideration of other types of the objective function, namely, concave function
(Falk and Horowitz, 1972), convex function (Foldes and Soumis, 1993), a hybrid
of concave and convex functions (Moder, Phillips, and Davis, 1983), and quadratic
function (Deckro, Hebert, Verdini, Grimsrud, and Venkateshwar, 1995). In recent

20
years there has been increased attention toward the discrete version of the problem
(Skutella, 1998; Zheng et al., 2004) due to its great practical relevance. This
consideration is imperative to TCT analyses since in practice many resources (e.g.,
workforce, equipment) are available in discrete units; in addition, time-cost
function of any type can be estimated by discrete functions. Being the main focus
of this thesis, the discrete version of TCTP considers discrete sets of time-cost
options for the activities and it is known as the discrete time-cost trade-off problem
– abridged as DTCTP – in the literature.

All the three extensions of DTCTP that was mentioned earlier are Non-
deterministic polynomial-time hard (NP-hard) problems in the strong sense (De et
al., 1997). That is to say, solution of DTCTP requires concurrent searches over the
solution space and that any escalation in the size of the project (any growth in the
number of activities, modes, or both simultaneously) contributes to the significantly
higher computational burden. Moreover, any variation in selection of the
alternatives modifies the project schedule which requires rescheduling the project
for potential changes in its total cost and total duration amounts. It is obvious that
any rescheduling process necessitates reanalyzing the network using the critical
path method. Exhaustive enumeration even with the modern computers is,
therefore, not a convenient and economically feasible method for solution of
DTCTP. Therefore, since the early 1960’s – concurrent with introduction of project
analysis techniques by Fulkerson (1961) and Kelley (1961) – researchers have come
up with several different optimization techniques to address this problem in the
literature. Heuristic method of Nicolai Siemens (1971) remains the first notable
approach developed for solution of TCTP. The methods proposed for DTCTP can
be categorized into three areas: exact methods, heuristics, and meta-heuristics.
Sections 2.3, 2.4, and 2.5 present the existing literature on the exact, heuristic, and
meta-heuristic methods, respectively.

21
2.3. Exact methods for TCTP

Traditionally, solution of DTCTP has been modelled by mathematical

programming, known as the exact methods. These approaches attempt to explore
the entire solution space to find the exact optimal solution. Generally, exact
methods are incapable of obtaining optimal solutions for large-scale problems
efficiently. The proposed exact algorithms, while requiring massive computational
resources, are more difficult to code and implement and are prone to being stagnated
in local optima in non-convex solution spaces (De et al., 1995; Feng et al., 1997;
Eshtehardian et al., 2008; Afshar et al., 2009). Due to the alleged drawbacks of the
mathematical programming models, researchers have turned their focus toward
alternative optimization methods. Nevertheless, exact procedures are the only
methods guaranteeing optimality of the solutions. Owing to this very reason, they
play a crucial role in performance evaluation of alternative optimization methods,
viz., heuristics and meta-heuristics. In the absence of an exact solution, the accuracy
of the results obtained by other methods cannot be calculated. Some of the most
popular variants of the exact algorithms include linear programming, dynamic
programming, mixed-integer programming, branch-and-bound algorithm, hybrid
LP/IP programming, and Benders decomposition method.

Of the exact methods, mixed-integer programming (MIP) consists of decision

variables constrained to include integer values, with objective function and other
variables allowed to select non-integer (continuous) values. MIP is a subset of the
broader linear programming (LP) in which all the variables and objective functions
are linear. TCT problems of any type can be converted into LP/MIP and solved by
implementing a commercial optimization solver. LP/MIP approaches suit the nature
of real-life TCTPs with convex time-cost relationships. Though, as the number of
activities of a project increases, the number of parameters to define for LP/MIP
approaches grows significantly.

22
The MIP approach of Meyer and Shaffer (1965), remains the pioneering attempt for
exact solution of TCTP. Another early MIP for TCTP analysis includes the method
proposed by Crowston and Thompson (1967). A rather small sample problem with
eight activities is used in their study. Liu et al. (1995) proposed a hybrid approach
combining linear and integer programming (IP) together. The hybrid approach uses
LP to set lower-bounds for cost and uses IP to find the exact solutions. This model
is implemented as a macro in Microsoft Excel environment and tested using a small-
scale seven activity problem. Moussourakis and Haksever (2004) present a flexible
MIP model that tackles deadline and budget TCT problems with linear, piecewise
linear, or discrete objective functions. Piecewise linearity of nonlinear continuous
functions is the only assumption required for this approach. This method is studied
using a seven-activity sample problem. Another MIP is proposed by Chassiakos
and Sakellaropoulos (2005) for Pareto front TCT problem, objective function of
which includes delay penalty and incentive payments. A small-scale 29-activity
sample problem – including finish-to-start with lag/lead time relationship – is fed
into the model. A more recent study on MIP is carried out by Szmerekovsky and
Venkateshan (2012). This method is capable of handling problems with time-cost
alternatives defined as different cost functions. This model is able to minimize Net
Present Value (NPV) of costs and maximize Cash Availability (CA) with applying
adjustments to the objective function. Sample problems of size 30 to 90 are used
and the results are compared with three other MIP methods. This method is shown
to obtain better results for denser networks in the light of its tight LP relaxation,
sparse constraint matrix, and small number of binary constraints. Bilir (2015)
propose a MIP approach for large-scale TCTPs. This method is tested using a set
of new samples generated using ProGen/Max random network generator. This
model is able to obtain solutions for the deadline TCTP with 1000 activities and 20
time-cost alternatives. The Pareto front problem is also solved for problems with
200 activities and five time-cost modes. A more recent mixed-integer linear
programming (MILP) approach is introduced by Zou, Fang, Huang, and Zhang
(2016) for repetitive deadline TCTPs with fixed logic by considering multiple crews

23
for each activity. In addition to MILP, an alternative approximate model with less
number of constraints and variables is also presented for unraveling larger
problems. Test instances are generated by means of ProGen random network
generator including 30, 40, 50, 80, 90, and 100 activities and CPLEX MIP Solver
is used for solving these instances. MILP is shown to provide solutions for problems
with up to 50 activities within the implemented computation time-frame of one
hour. It is also revealed that this MILP cannot solve problems that include more
than 50 activities even in three hours. Therefore, the approximate method is used
for the solution of problems with up to 100 activities within the enforced maximum
CPU-time of one hour. Dragović, Vulević, Todosijević, Kostadinov, and Zlatić
(2017) presents an LP model for direct cost minimization of deadline TCTP. This
approach is implemented using MATLAB toolbox and is applied for direct cost
minimization of torrent-control projects including 10, 12, and 13 activities. The
effect of daily indirect cost is not considered in this optimization procedure results
of which are revealed as direct cost reductions ranging from 5.58% to 9.19%.

Dynamic programming (DP) is another category of exact procedures which aims at

reducing the size of a practiced network by merging its activities. DP reduces the
network size by decomposing it into a sequence of smaller sub-problems. Though,
this reduction cannot be applied to a network with complex logical relationships
among its activities. DP imposes a reasonable computation burden for small and
simple problems; however, it might not be practical to use DP for larger and more
complex networks.

Early study on DP includes Butcher’s (1967) method. This approach is capable of

unraveling budget TCTPs with plain series or plain parallel networks. Another DP
for budget TCTP is developed by Robinson (1975). This method is also based on
network decomposition. Specific assumptions and conditions are set for reducing
the network into a one-dimensional problem. However, taking a multidimensional
optimization approach is suggested for problems with more complex networks, due

24
to their difficult decomposition processes. DP method of Panagiotakopoulos (1977)
implements problem simplification to solve the TCT problems by means of
enumeration. De et al. (1995), giving an inclusive literature review on previous
approaches, discuss the scant and sparse attention toward discrete version of TCTP.
By implementing modular decomposition and incremental reduction, this thesis
study introduces a centralized DP model for Pareto front TCT problems with
parallel modules. Demeulemeester, Herroelen, and Elmaghraby (1996) present two
DP methods for Pareto front DTCTP. The first method uses node-reduction to
convert the problem network into a series-parallel network and the second model
reduces the number of possible time-cost alternative combinations. Both the
methods are coded in C programming language and a sample problem with 45
activities is practiced.

Other major exact methods which are classified as branch-and-bound (B&B)

algorithms were first introduced by Land and Doig (1960). B&Bs are generally used
for combinatorial optimization problems. B&B consists of systematic enumeration
of all possible combinations of time-cost alternatives in the solution space. B&B
implicitly searches only certain portions of the solution space by implementing
upper and lower estimated bounds on the optimal solution. It portions the problem
into subsets and is able to identify and discard those schedules that will not lead to
any improvements in the objective function value.

Early B&B methods for Pareto front TCTP include horizon-varying approach of
Demeulemeester et al. (1998). Lower boundaries are calculated by setting convex
piecewise linear underestimations of activity time-cost curves. The quality of the
underestimations is assessed using a vertical distance calculation. This method
applies branching to time-cost alternatives by identifying the activity with the
largest vertical distance. Branching divides time-cost options into two subset
groups. This approach is implemented in Visual C++ platform and instances with
10, 20, 30, 40, and 50 activities with up to 11 time-cost alternatives are practiced.

25
The results are compared with Demeulemeester et al.’s (1996) method which reveal
capability of this approach in providing solutions for instances with up to 30
activities and 4 time-cost options. It is also shown that nearly half of the problems
including 50 activities with six or more alternatives cannot be solved using this
B&B algorithm. Vanhoucke, Demeulemeester, and Herroelen (2002) present a
B&B algorithm for deadline TCTP incorporating the time-switch constraints of
Yang and Chen (2000) and the lower-bound calculation procedure of
Demeulemeester et al. (1998). Contrary to CPM, time-switch constraints impose
specific intervals for execution of the activities which are designed to deal with day,
night, and weekend shifts. This algorithm is coded in Visual C++ environment and
tested using a sample problem with 20 activities. Another B&B algorithm is
developed by Vanhoucke (2005) for deadline TCTP with time-switch constraints.
The branching process creates three child nodes as it divides the start time of
activities into three sections. Branching ignores time-switch constraints of those
activities that cause exceeding the project deadline. This method is implemented in
Visual C++ platform and tested using instances with 10, 20, and 30 activities having
up to seven time-cost alternatives. On average, this newer B&B requires less than
seven seconds to unravel 30-activity sample with seven time-cost options which is
four times faster than Vanhoucke et al.’s (2002) approach. A more recent B&B is
generated by Degirmenci and Azizoglu (2013) for budget TCTP. This method aims
to find the solution with the least cost among the solutions having the minimum
completion time. This method embeds a mode elimination procedure and a network
size reduction technique. Time-cost alternatives not leading to a feasible or an
optimal solution are eliminated. Lower and upper-bounds for total cost are
calculated using linear programming relaxation. Two procedures of LPR-based and
naïve are proposed for calculation of lower-bounds. This approach is implemented
in C# environment used for solution of 360 problems including up to 35 activities
with up to 20 time-cost alternatives. The optimal solutions are found by means of
CPLEX 10.1.

26
Benders decomposition (BD) is another exact procedure which was first introduced
by Benders (1962). It is generally used for solution of large-scale stochastic/linear
programming problems which partitions the practiced problem into multiple
simpler sub-problems. Benders decomposition consists of two stages, in first of
which a lower-bound is set and the master problem is solved for a subset of
variables and in the second stage, an upper-bound is set and the remaining variables
are calculated for a sub-problem using the values determined in the preceding stage.
If an infeasibility of first stage decision is identified in the latter stage, additional
constraints are used to solve the master problem.

Hazır, Haouari, and Erel (2010) propose a BD method for budget TCTP by
modifying the original method of Benders. It includes an improved decomposition
approach and as well as a branch-and-cut procedure. Problems networks with 85 to
136 activities are solve using this procedure. It is shown that this BD can outperform
both Benders and IBM’s CPLEX 9.1 methods. Another BD was presented by Hazır,
Erel, and Gunalay (2011) for deadline TCTP with uncertain cost parameters. Three
alternative optimization models are proposed that assume interval uncertainty for
unknown parameters. This model assumes certain start and finish times for
activities as their time-cost alternatives are associated with fixed durations. All the
uncertainty is reflected on the cost of an activity using probabilistic intervals. This
method is tested using problems with 85 to 136 activities.

2.4. Heuristic methods for TCTP

In contrast to the exact procedures, heuristic methods are convenient tools of

optimization, requiring non-substantial computation time and resources. Some of
the heuristics can be implemented even short of a computer’s assist. Derived from
the Greek word “Heuriskein” meaning “to find”, heuristics involve simple rules for
finding solutions to difficult optimization problems. Nonetheless, the optimality is
not guaranteed in the heuristic methods. The solutions obtained by means of

27
heuristics are rather satisfactory since they are either optimal or near-optimal. The
constructive and the improvement heuristics are the most revered variants of the
heuristic algorithms. The former uses a stepwise procedure to generate solutions,
generating them one at a time until a feasible solution is met. Generally, a feasible
solution is not obtained in the course of the construction heuristics unless the
conclusion of the procedure is reached. The latter type of the heuristic algorithms
i.e. the improvement heuristics, initiate with a feasible solution and successively
improve it via a series of modifications. In the course of this procedure, usually a
feasible solution is preserved regardless of the progression of the process
(Aminbakhsh, 2013). Existing studies providing heuristic algorithms acknowledge
that they are problem dependent and cannot handle large-scale problems efficiently
(Siemens, 1971). Most of the present heuristics assume merely linear time-cost
functions and they fail to solve the Pareto front problem (Feng et al., 1997; Zheng
et al., 2005).

Early heuristics in the literature of TCTP include the procedure proposed by

Siemens (1971). Known as SAM (Siemens Approximation Method), a logical
systematic procedure is developed to minimize the overall project cost which is
suited for both manual and computer aided calculations. It is capable of obtaining
solution for convex nonlinear TCTPs by making multiple piecewise linear curve
approximations. The procedure initiates with the construction of the project
network, thereby, crashing critical activities one at a time based on some specific
rules. Of these rules, the most important is selection of a critical activity with
minimum amount of cost-slope. The act of crashing continues until either all the
activities are crashed, or those with normal duration have cost-slopes greater than
the daily indirect cost. The major drawback related to this approach is that it
requires determination of all the critical paths in the network which might prove
expensive to solve complex problems. A sample problem with eight activities is
fitted into the model and it is shown that the results compare well with the results
of a linear programming approach. Goyal (1975) proposes a modified version of

28
SAM which uses effective cost-slopes instead of the plain cost-slopes and is capable
of obtaining solutions for TCTPs with convex nonlinear cost-slopes. This method
also introduces a de-shortening technique which is applied to the excessively
shortened paths. The same eight activity network of Siemens (1971) is used to test
the proposed heuristic. An alternative heuristic is proposed by Moselhi (1993) to
minimize the overall project cost which involves replacing CPM network with a
corresponding structure. Using the analogy between the support settlement of
structural analysis and the completion deadline of TCTPs, the overall cost of
schedule compression is calculated using the sum of all member forces.
Aminbakhsh (2013) and Aminbakhsh and Sonmez (2016, and 2017) develop a new
cost-slope-based heuristic approach for their hybrid optimization method which is
regarded as one of the first TCTP heuristics operating in discrete search space.
Named as modified-SAM, this method is designed for problems with discrete time-
cost relationships. Modified-SAM, unmatched by any other previous heuristic for
TCTP, is capable of obtaining multiple non-dominated solutions in a single run for
Pareto front problem. Bettemir and Birgonul (2017) explain another discrete
heuristic method called Network Analysis Algorithm (NAA) for deadline and cost
minimization TCTP. Least cost-slope activities are crashed under three conditions
of single critical path, multi-critical path, and necessity of crashing non-critical
activities. This approach uses an elimination algorithm to reduce the number of
crashing options by excluding those critical activities that are shared with all the
critical paths. 18-activity problem of Feng et al. (1997) and 63-activity problem of
Sonmez and Bettemir (2012) are used for performance measurement of NAA.
Compared with hybrid-GA of Sonmez and Bettemir (2012), NAA is concluded to
require one percent of meta-heuristics computation burden and contributing to
higher accuracy. Su, Qi, and Wei (2017) propose an equivalent simplification-based
approach for solution of nonlinear continuous deadline TCTP. This algorithm
transforms TCTP into a simple activity float problem then uses a polynomial
algorithm to unravel the transformed problem. Both the simplification and
polynomial algorithms are implemented using LINDO optimization software. This

29
method narrows the domain of direct cost for each activity duration to achieve
higher efficiency and accuracy. Redundant activities are also discarded to reduce
the scale of the TCTP. This algorithm is experimented on a 116-activity problem
which requires 30 minutes of processing time in the absence of the simplification
method, and takes 2 seconds when the problem is simplified to a 33-activity
network.

2.5. Meta-heuristic methods for TCTP

Inspired by the stochastic occurrences of the nature, evolutionary algorithms are

among the meta-heuristics practiced in solution of DTCT problems. “Meta”,
meaning “beyond” is an indication of higher-level algorithms when compared with
the heuristics. This is because of the fact that heuristics are problem dependent
while meta-heuristics are independent of the nature of the problem. Existing meta-
heuristics unravel an optimization problem by randomly searching the solution
space. Contrary to heuristic methods, meta-heuristics are designed to explore a
problem in an iterative fashion. This is mainly performed in view of prevention of
meta-heuristic from getting stuck into the local optima. Ironically, the main
deficiency of the existing meta-heuristics is observed as their inability to escape
from local optima (Zheng et al., 2005; Sonmez and Bettemir, 2012). In addition,
similar to the heuristics, meta-heuristics are incapable of securing the optimality of
the solutions; rather, they can provide near-optimal solutions within only moderate
computational effort. This category of optimization methods is well associated with
the modern studies, including genetic algorithms (GA), ant colony optimization
(ACO), particle swarm optimization (PSO), shuffled frog leaping (SFL), and
simulated annealing (SA).

Of the meta-heuristic methods, Genetic algorithms (GA) are the most popular and
highly practiced approaches. GA was first introduced by Holland (1975) which as
the name implies, is inspired by the natural selection and genetic reproduction

30
process. Feng et al. (1997) present a GA for solution of Pareto front TCT problem.
Addressing the lack of sound methods for coping with large-scale TCTPs, a GA is
generated in which normal and crash modes of an activity are defined as two
chromosomes. This method develops new solutions by iterative cross-overs and
mutations, fitness values of which are evaluated using their minimal distances to a
convex hull. Genetic drift (raised by Goldberg and Segrest, 1987) is avoided by
retaining each string for the next generation. An 18–activity problem is used to test
the proposed model, different variants of which has since been used extensively
within the TCTP literature. It is shown that this method is capable of locating more
than 95% of the non-dominated solutions for the benchmark problem. Li and Love
(1997) introduce a GA based approach for deadline TCTP with improved mutation
and crossover processes. A sample problem of 10 activities is fitted into both the
proposed and the original GAs and the results are compared. It is indicated that the
presented method outperforms the original GA due to its higher search efficiency.

Hegazy (1999) comes up with another GA for deadline TCTP. This method requires
shorter computation time as it employs the same techniques proposed by Li and
Love (1997). This GA is implemented as a macro in Microsoft Project 1995 and is
tested using instances with 18, 36, 108, and 360 activities. It is concluded that
although this method is capable of obtaining solutions for all the practiced
instances, the computation time increases as the problem becomes more
complicated. Meanwhile, the core 18-activity problem is actually the same instance
introduced by Feng et al. (1997) with a slight difference in one of its activities.
Zheng et al. (2005) introduce a GA model for Pareto front TCTP. A modified
adaptive weight approach (MAWA) is used to adjust the priority of each objectives
of time and cost with regard to performance of preceding generation. As the
generations evolve, MAWA administers a decreasing pattern for the mutation rate
to prevent premature convergence. Pareto ranking and niche formation are also
implemented to this model with the former serving as a selection criterion and the
latter as a population diversifier. This approach is implemented as a macro in

31
Microsoft Project 2000 and is practiced using Feng et al.’s (1997) 18-activity
problem. The results indicate robustness of this method and it is shown that the
results keep improving as generation count is increased beyond 300.

Kandil and El-Rayes (2006) present a GA with two parallel processing techniques.
Global parallel and coarse-grained GAs are used for solution of Pareto front TCTPs.
Instances with 180, 360, and 720 activities are generated by using the time-cost
alternatives of Feng et al. (1997) and by copying a slightly modified version of the
18-activity network of Feng et al. (1997) in serial several times. Global parallel GA
is able to find 267 solutions for the 180-activity problem in 42.6 minutes. The
coarse-grained GA, on the other hand, achieves 68 Pareto solutions in 11.4 minutes.
Similarly, global parallel GA is able to find 232 solutions for the 360-activity
problem in 173.4 minutes. The coarse-grained GA, on the other hand, achieves 94
Pareto solutions in 30.6 minutes. 136.5 hours on a single processor, and 15.4 hours
over a supercomputing cluster of 50 processors is required for global parallel GA
to achieve 303 Pareto front solutions for the 720-activity problem and the coarse-
grained GA achieves 132 non-dominated solutions in 118.18 minutes. Eshtehardian
et al. (2008) discuss a GA which can reflect the uncertainties in time and cost of
project activities using the fuzzy sets theory of Zadeh (1965). This method can
obtain solution for stochastic Pareto front TCTP by assigning triangular fuzzy
numbers to time and cost of the work-packages. Hamming-distance and Euclidian-
distance techniques are implemented to GA and α cut approach is used in ranking
the non-dominated solutions in accord with the decision maker’s acceptance of risk
level. The GA based approach is tested using the 18-activity problem of Feng et al.
(1997).

Eshtehardian, Afshar, and Abbasnia (2009) demonstrate another fuzzy-based multi-

objective GA for Pareto front optimization of TCTPs with uncertainties in their
activity time-cost pairs. This model addresses risk acceptance level and degree of
optimism of a decision maker employing α cut approach and optimism index,

32
respectively. Fuzzy set theory is explicitly embedded into this optimization process
which ranks fuzzy numbers using an approximate method in reference to the left
and right dominance. This GA model uses the general concepts of NSGA-II
combined with single cut cross over, uniform mutation, and tournament selection
procedures. Time-cost options of 7-activity problem of Zheng et al. (2005) and 18-
activity problem of Feng et al. (1997) are defined as triangular fuzzy numbers.
These instances are fed into the GA, deterministic results of which are compared
with the original studies mentioned. Sonmez and Bettemir (2012) present a hybrid-
GA for both the deadline and cost minimization TCTPs. This method combines the
complementary capabilities of simulated annealing (SA) and quantum simulated
annealing (QSA) together with the GA approach. This method is coded in Visual
C++ environment, SA module of which aims to improve hill-climbing, while its
QSA module aims to improve local search capabilities of the hybrid model.
Problem sizes ranging from 18 to 630 activities are fitted into the model and the
results are compared with the optimal solutions achieved by mixed integer
programming technique implemented in AIMMS optimization software. Test
instances are based on the 18-activity problem of Feng et al. (1997), 29-activity
problem of Chassiakos and Sakellaropoulos (2005), and a hypothetical 63-activity
problem. Outputs of paired t-test is given which verifies improved convergence
capability of this model against a typical GA.

Mungle, Benyoucef, Son, and Tiwari (2013) demonstrate a fuzzy clustering-based

GA for Pareto TCTPs with uncertainties in their time-cost amounts. The anticipated
qualities of execution alternatives are measured by means of analytical hierarchy
process (AHP). This approach combines NSGA-II with external repository concept
and linkage-based hierarchical clustering technique for storing and managing
diversity of the Pareto solutions, respectively. Implemented in MATLAB, this GA
incorporates crowded-comparison, crowding distance, two-point crossover, and bit-
flip mutation operators. This model is used for analyzing 7-activity problem of
Burns, Liu, and Feng (1996), a hypothetical 12-activity problem, and 18-activity

33
problem of Feng et al. (1997). Findings of this approach is compared with HS of
Geem (2010), SPEA-II of Zitzler, Laumanns, and Thiele (2001), and exact method
of Burns et al. (1996). Generational distance (GD) and Range Variance (RV)
metrics are used for measuring convergence and diversity of the obtained Pareto
solutions, respectively. These metrics are then measured for PSO of Zhang and Li
(2010) and GA of Feng et al. (1997) to carry out comparisons. Zhang, Zou, and Qi
(2015) introduce a GA for repetitive projects with soft logic. A mixed-integer
nonlinear programming approach is used to combine TCTP and the soft logic
concept which allows for changes to be made in the logical relationships, in view
of a faster and cheaper project realization. This approach is tested using a sample
problem with five activities including a single resource for execution of each of its
activities.

Koo, Hong, and Kim (2015) demonstrate GA-based integrated multi-objective

optimization (iMOO) model for Pareto front TCTP. This model is able to carry out
optimization using four different fitness functions including trade-offs between
time-cost, cost-quality, sustainability-cost, and productivity-safety. Prior to GA
phase, heuristic module of this model defines the minimum and maximum extreme
points in the solution space. OptQuest tool of Crystal Ball is used to implement the
GA procedure. Capabilities of this approach for the time-cost objective function is
tested using the 7-activity problem of Zheng et al. (2005). It is declared that this
model is able to find the Pareto front identical to the previous studies. Huang, Zou,
and Zhang (2016) develop another GA-based model for repetitive deadline TCTPs
with soft logic by considering a single crew for each activity. They aim to reduce
the overall cost of project by implementing a proper sequencing. They use linear
programming for reducing a repetitive construction projects to a classical TCTP,
afterwards, they use a GA to tackle the reduced problem which is programmed in
MATLAB. This approach is tested using a 5-activity problem solved considering
two sequencings as well as a 6-activity project solved considering three sequencing
strategies. El-Abbasy, Elazouni, and Zayed (2016) develop an elitist non-dominated

34
sorting GA for Pareto front TCTP that considers multiple projects with multiple
objectives. This method unravels TCTP by taking into account the financing cost,
maximum required credit, profit, and resource levelling of the project. Of the three
models implemented, scheduling model is validated using the 18-activity problem
of Feng et al. (1997). This model which is also able to tackle pure TCTPs is shown
to locate 31 non-dominated solutions for this instance which are compared visually
to the results of Feng et al. (1997).

A more recent and notable GA-based approach is introduced by Agdas, Warne,

Osio-Norgaard, and Masters (2018) for large-scale cost minimization and deadline
TCTPs. Parallel GA is developed by applying the Scalable Concurrent Operations
method to the open source Distributed Evolutionary Algorithms framework of
Python programming language. A custom GA is developed employing a new
problem encoding, implementing weighted graph method instead of iterative CPM
calculations, a parallel fitness evaluation technique, and a new stagnation criterion.
This model is shown to run four times faster in the light of parallel computing on a
high-performance computing facility with eight CPU cores. Combined with their
modifications, this GA is proved to be 100 times faster than a typical GA running
on the same computer on a single CPU core. Large-scale problems with 630, 1800,
3150, and 6300 activities are practiced and the results are compared with Sonmez
and Bettemeir (2012). The 1800-activity problem is based on the 18-activity
network of Feng et al. (1997) and the remaining instances are based on the 63-
activity problem of Sonmez and Bettemir (2012) which are generated by copying
the same network in serial several times. Performance of the parallel GA is
evaluated with respect to its accuracy and processing time in two stages; viz., before
and after applying the modifications. It is contended that the results of unmodified
parallel GA exceed the accuracy of Sonmez and Bettemir’s (2012) hybrid-GA
approach for the 630-activity problem; though, it requires more than four hours to
unravel this project. The accuracy of the modified version however is demonstrated
to be on par with hybrid-GA of Sonmez and Bettemir (2012) with much less runtime

35
requirements. CPU time for the larger problems is reported to vary from 5.82 to
16.76 hours with average deviations ranging from 4.73% to 14.72%.

An alternative meta-heuristic approach covers Ant Colony Optimization (ACO)

methods. ACO was first proposed by Colorni, Dorigo, and Meniezzo (1992) which
is inspired by the coordinated interactions of ant colonies in search for sources of
food. Ng and Zhang (2008) present an ACO based model for Pareto front TCTP.
This model is equipped with modified adaptive weight approach of Zheng et al.
(2004) for fitness function evaluations. This model is programmed in Visual Basic
platform and its results on the 18-activity problem of Feng et al. (1997) are
compared with Elbeltagi, Hegazy, and Grierson’s (2005) typical ACO method. This
method is shown to decrease the computation resource requirements of the original
ACO. Xiong and Kuang (2008) develop another ACO employing the modified
adaptive weight approach of Zheng et al. (2004) for Pareto front TCTP. This
method incorporates two selection criteria for the alternatives. The first selection is
made regarding a maximization criterion and the second involves a probability
distribution function. The performance of this method is evaluated using 7-activity
problem of Zheng et al. (2004) and the 18-activity of Feng et al. (1997). This
method is capable of obtaining solutions by exploring rather smaller portion of the
solution space.

Afshar et al. (2009) develop a multi-colony non-dominated archiving ACO for

Pareto front TCTP. Separate ant colonies are assigned to time and cost objectives.
These separate colonies are designed to share solutions to be evaluated with regard
to the counteracting objective. Performance of this method is measured using the
18-activity problem of Feng et al. (1997) under three assumptions regarding the
amount of daily indirect cost. For all the cases, the results prove to be better than
the solutions achieved by ACO-based and GA-based methods of Zheng et al.
(2005). Kalhor, Khanzadi, Eshtehardian, and Afshar (2011) explain a fuzzy-based
non-dominated archiving ACO for Pareto front optimization of TCTPs with

36
uncertainties in their activity time-cost pairs. In this model α cut approach accounts
for the risk acceptance level of the decision maker. Degree of optimism of a
decision maker is also embedded to this method using an optimism index. Left and
right dominance rule is employed for conducting comparison in the non-dominated
sorting process. Coded in MATLAB, this ACO uses one colony for minimization
of fuzzy cost and another for minimization of fuzzy duration. The iteration best
solution inside the colonies and the best-so-far solution are used for pheromone trail
updates. 18-activity network of Feng et al. (1997) with fuzzy time-cost pairs of
Eshtehardian et al. (2009) is practiced by assuming non-symmetric triangular shape
for each activity execution alternative. The extent of convergence and spread of the
obtained Pareto front is assessed using Euclidean distance and delta (Δ) metrics,
respectively. Findings of this model are demonstrated to be fully compatible with
outcomes of Eshtehardian et al. (2009). Zhang and Ng (2012) propose another ACO
approach for cost minimization and Pareto front TCTPs. This method is
implemented in VBA and is integrated into Microsoft Project. 18-activity problem
of Feng et al. (1997) is practiced and the results are compared with GA-based
method of Zheng et al. (2005).

One other domain of meta-heuristics includes the Particle Swarm Optimization

(PSO) algorithms. Developed by Kennedy and Eberhart (1995), PSO is a
population-based algorithm imitating the choreography of bird flocks that
communicate together as they fly. Elbeltagi et al. (2005) studies performance of
five different evolutionary algorithms for cost minimization TCTP, one of which is
a PSO-based model. The five meta-heuristics include GA, memetic algorithm
(MA), PSO, ACO, and shuffle frog leaping (SFL) all of which are coded in Visual
Basic platform. The practiced test instances also contain the 18-activity network of
Feng et al. (1997). The results of this comparison reveal PSO outperforming other
techniques based on the solution quality and placing second regarding the
computation time requirements. Yang (2007b) presents a modified PSO for Pareto
front TCTPs. Pbest is updated in case the solution is strongly dominating and gbest

37
accepts only the particle that dominates fewest solutions. Non-dominated solutions
are stored and updated in a separate archive of elite solutions. Indirect cost is not
included in the optimization process; however, it is added to the final Pareto front
solutions. Performance of this algorithm is measured for solution of a 14-activity
problem. Yang (2007a) proposes another PSO method for Pareto front TCTP. This
model is coded in MATLAB programming environment and the indirect cost is
implemented exogenously to the final Pareto front solutions. A hypothetical
example with 8 activities and a case problem with 28 activities is practiced to test
the algorithm. Small average percent deviation (APD) amounts from the optima is
reported for this approach.

Zhang and Li (2010) introduce a PSO for Pareto front TCTP. Sparse-degree and
roulette-wheel selection are implemented to improve diversity of particles and
convergence capabilities of the model. Gbest is determined using the combined
scheme used in this model which is coded in Visual C++ environment. Results of
this model for a 7-activity and an 18-activity (of Feng et al., 1997) problem is shown
to best GA of Feng et al. (1997). Fallah-Mehdipour, Bozorg Haddad, Rezapour
Tabari, and Mariño (2012) explain revised multi-objective PSO and GA models for
Pareto front TCTP. Both methods are modified for problems with discrete domains,
e.g., PSO is provided with a classification technique which divides a feasible
continuous space into discrete units and the continuous decision is rounded to the
closest integer value. Capabilities of these two approached are assessed and
compared using the 18-activity problem of Feng et al. (1997). PSO and GA are
compared with each other with respect to Generational Distance (GD) of their
results also using each method’s share of non-dominated solutions in the final
Pareto front; GA is revealed to be more successful for solution of the practiced
instance. Aminbakhsh and Sonmez (2016) propose a discrete PSO for cost
minimization and deadline TCTP. This method is based on the novel principles for
representation, initialization and position-updating of the particles which facilitate
adequate representation of the discrete search space and enhance accuracy due to

38
improved quality of the initial swarm. This method uses semi-deterministic
initialization scheme by employing the modified-SAM (Aminbakhsh, 2013)
heuristic. This approach is implemented in C++ and tested using benchmark
problems and large-scale instances generated using ProGen/Max network
generator. Different variants of 18-activity problem of Feng et al. (1997), 63-
activity and 630-activity problems of Sonmez and Bettemir (2012), and large-scale
problems including 200 and 500 activities with two to five time-cost options are
used for performance measurement of this procedure. Findings of this approach is
compared with outcomes of numerous previous studies and the deviations from the
optimal solutions are calculated using exact solutions of a MIP method.

In a more recent study, Aminbakhsh and Sonmez (2017) present a Pareto front PSO
for Pareto oriented optimization of large-scale TCTPs. This model is treated with
unique particle representation, initialization, and position-updating techniques. This
multi-objective PSO is also embedded with a modified version of SAM which is
named simplified-SAM; though, unlike any other variant of SAM, instead of
determining the least-cost-slope critical activities for all of the multiple parallel
critical paths, one critical activity at a time is crashed to circumvent the bottleneck
of finding all the critical activities of the network. This PSO incorporates the
external repository concept and redefines position update and gbest calculation
routines. C++ is used for coding of this approach. This model is experimented on
different variants of 18-activity problem of Feng et al. (1997), 180-activity, 360-
activity, and 720-activity problems of Kandil and El-Rayes (2006). Pareto fronts
obtained by this approach are compared with the results of existing methods and
the deviations from the true Pareto fronts are calculated for some of the practiced
instances by means of a MIP method.

Another meta-heuristic approach is categorized as Shuffle Frog Leaping (SFL)

algorithm. SFL imitates the behavior of a group of frogs locating a source with the
maximum amount of available food which was originally developed by Eusuff and

39
Lansey (2003). Elbeltagi, Hegazy, and Grierson (2007) explain an SFL for cost
minimization TCTP. The original SFL is modified by implementing a search-
acceleration parameter to avoid getting stuck into the local optima. This parameter
is designed to sustain the balance between the local and the global search of the
algorithm. The proposed SFL is implemented as a VBA code in Microsoft Project.
The performance of this model is validated using benchmark problems as well as
instances with 18, 90, and 180 activities all of which are based on the 18-activity
network of Feng et al. (1997). Results reveal less computation time requirement and
high success rate of this model compared with the original SFL and a GA algorithm.

Of the meta-heuristic algorithms, Simulated Annealing (SA), being first proposed

by Kirkpatrick, Gellat, and Vecchi (1983), is inspired by the heating and controlled
cooling of alloys subject to tempering. At higher temperatures, the molecules move
freely in any direction which can be restricted by lowering the temperature.
Anagnostopoulos and Kotsikas (2010) analyze five variants of SA for cost
minimization of TCTP. Visual Basic is used to implement SA and problems with
100, 200, and 300 activities are practiced for performance evaluation. All the
instances are generated by means of RanGen random network generator. Analysis
of variance (ANOVA) and Duncan Multiple Range Test are used to measure the
performance and to rank the five SA algorithms.

An alternative meta-heuristic algorithm is developed by Vanhoucke and Debels

(2007). TCTP is exerted considering time/switch constraints, work continuity
constraints, and net present value maximization. This method is coded in Visual
C++ platform and is a hybrid of a heuristic and a truncated dynamic programming
technique. The first module undertakes neighborhood search and diversification
steps and the second module relaxes non-critical activities. The quality of the
solutions obtained by this method is measured using the B&B algorithm of
Demeulemeester et al. (1998).

40
Geem (2010) introduce Harmony Search (HS) meta-heuristic algorithm for Pareto
front TCTP which is based upon the analogy between musician’s experience and
the optimization process. This method involves a phenomenon-mimicking
approach which is programmed in Microsoft Excel. Unlike GA that uses two
solutions vectors, this method considers all the solutions in generating a new one.
Performance of this model is measured using the 7-activity problem of Zheng et al.
(2005) and the 18-activity network of Feng et al. (1997). The provided results stack
up well against GA of Zheng et al. (2004) and ACO of Xiong and Kuang (2008).

Abdel-Raheem and Khalafallah (2011) come up with another meta-heuristic

method hailed as Electimize for cost minimization TCTP. This approach simulates
the behavior of electrons as they tend to move through a multi-branch electric
circuit via the branch with the least resistance. Each solution is represented by a
wire, segments of which represents the activity options. Coded as a macro in
Microsoft Project, Electimize uses Ohm’s law and Kirchhoff’s rule through its
simulation process which assess quality of each solution using their global
resistance value. This method is applied for solution of the 18-activity problem of
Feng et al. (1997).

In a similar fashion, Vanhoucke (2015) develops Electromagnetic (EM) model for

four different variants of deadline TCTP. Time/switch constraints, work continuity
constraints, and net present value maximization are also considered in optimization
of the deadline problem. EM relies on the law of Coulomb and uses the principles
of Birbil and Fang (2003) to calculate charges and forces. This method incorporates
two local searches; one for crashing and one for relaxing activities. Projects with
10, 20, 30, 40, and 50 activities with up to 11 time-cost options are fitted into the
model which is coded in Visual C++ platform. Results are compared with the exact
solutions obtained using the procedures described in Demeulemeester et al. (1998)
and Vanhoucke (2005) and by using LINDO optimization software. Results of this
method is contended to best Vanhoucke and Debels’s (2007) solutions.

41
The major existing exact, heuristic, and meta-heuristic approaches for TCTP are
summarized in Table 2.1. Entries are arranged in a chronological order in this table,
including the practiced problem types, implemented platforms, size of the instances
used with their daily indirect cost rate (IC), computation time of the proposed
methods (CPU time), their number of consecutive runs, and their associated average
percent deviation from optima (APD). The last column of this table gives a brief
explanation and highlights remarkable points of each study. Unreported details are
tabulated as ‘na’ in Table 2.1. Different TCT problem types are abridged as follows
in this table:

- TCT: Cost minimization problem;

- TCT-BT: Budget problem;
- TCT-DL: Deadline problem;
- TCT-PF: Pareto front problem;
- St. TCT-DL: Stochastic deadline problem;
- St. TCT-PF: Stochastic Pareto front problem.

42
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP.

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
1967 Crowston and MIP TCT-DL na 8 na 5 na na A mathematical model for Decision
Thopson CPM is presented to carry out path
reductions.
1971 Siemens Heuristic TCT na 8 na na na na Logical systematic approach involving
a number of rules for expediting the
activities that incur least additional
costs. Since this method only considers
minimum cost-slope for crashing
activities, it might shorten the duration
beyond required amount.
1975 Robinson DP TCT-BT na na na na na na Based on network decomposition, this
method sets specific assumptions and
conditions for reducing the network into

43
a one-dimensional problem.
1995 De, Dunne, DP TCT-PF na 5 na na na na A centralized approach for deadline
Ghosh, and 10 problem with no parallel modules and a
Wells combination of modular decomposition
with incremental reduction approaches
for problems with parallel modules.
This method is only effective for
networks with reasonably low values of
certain parameters.
1995 Liu, Burns, and Hybrid TCT-DL VBA 7 na 1800 na na The hybrid approach uses LP to set
Feng LP/IP Macro lower-bounds for cost and uses IP to
find the exact solutions.
1996 Demeulemeester, DP TCT-PF C 45 na 530.4 na na Two DPs are presented first of which
Herroelen, and uses node-reduction to convert the
Elmaghraby network into a series-parallel and the
second method reduces the number of
possible time-cost combinations.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
1997 Feng, Liu, and GA TCT-PF na 18 $0 na na na A GA that calculates fitness values
Burns $1500 exploiting minimal distance to convex
hull and retains each string for next
generation to avoid genetic drift. This
approach only tackles Finish-to-Start
relationships neglecting possible
resource constraints.
1997 Li and Love GA TCT-DL na 10 na na na na A modified GA algorithm with
improved mutation and crossover
processes.
1998 Demeulemeester, B&B TCT-PF Visual 10 na 0.34 na na Horizon-varying approach is embedded
Reyck, Foubert, C++ 20 19.17 into branch and bound method and
Herroelen, and qualities of lower boundary

44
Vanhoucke 30 58 underestimations are assessed by
40 105.4 vertical distance computations.
50 127.56 Effectiveness and efficiency of this
model decreases significantly for larger
networks with multiple modes.
1999 Hegazy GA TCT-DL VBA 18 $200 390 na 8.139 A premier GA model that minimizes
Macro total cost while taking into account the
project-specific constraints on time.
Random nature of model requires long
processing time in VBA environment.
2002 Vanhoucke, B&B TCT-DL Visual 20 na na na na A B&B with time-switch constraints is
Demeulemeester, C++ presented which is designed to deal with
and Herroelen day, night, and weekend shifts.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2004 Moussourakis MIP TCT-DL na 7 na na na na A method able to solve deadline and
and Haksever budget TCT problems with linear,
piecewise linear, or discrete objective
functions. Requires no network notation
system and makes minimal assumptions
regarding the type of TCT functions.
Since this method requires substantial
computational resources, it suits small
to medium instances.
2005 Chassiakos and MIP TCT-PF LINDO 29 €1200 <1 na na Exact and approximate models are
Sakellaropoulos developed and compared for solution of
a hypothetical problem.
2005 Elbeltagi, GA TCT Visual 18 $500 16 20 2.171 Five meta-heuristic algorithms (GA,

45
Hegazy, and MA Basic 21 0.759 MA, PSO, ACO, and SFL) are
Grierson compared, revealing poor performance
PSO 15 0.415 of GA as well as robustness of PSO
ACO 10 3.351 methods.
SFL 15 2.96
2005 Vanhoucke B&B TCT-DL Visual 10 na 0.003 na 0 Branching process of this B&B
C++ 20 0.010 to 0 algorithm ignores time-switch
0.089 constraints of those activities that cause
30 0.089 to 0.015 exceeding the project deadline. Lower-
6.535 bound branching is applied to activities
having greater than duration lead times.
2005 Zheng, Ng, and MAWA- TCT-PF na 18 $1500 na na 1.324 A GA that recruits MAWA, Pareto
Kumaraswamy GA with (Based ranking, and Niche formation to avoid
Niche on 4 genetic drift, administer selection
Formation Obs.) pattern, and exert diversifier,
respectively.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2006 Kandil and El- GP-GA TCT-PF na 180 $0 2,556 to na na A GA with two parallel processing
Rayes 14,688 techniques of global parallel and coarse
360 10,404 grained is presented and used for Pareto
to oriented optimization of problems with
75,096 up to 720 activities. Despite parallel
720 55,296 computing, there remains the
to computation time problem since both
491,400 variants require significant processing
CG-GA 180 684 times.
360 1,836
720 7,092
2007 Elbeltagi, MSFL TCT VBA 18 $500 8 20 0 An SFL algorithm that incorporates a

46
Hegazy, and Macro 90 810 na na time-variant parameter to avoid falling
Grierson into local optima.
180 na
2007 Vanhoucke and Meta- TCT Visual 10 na 0.008 na 0.037 Heuristic portion involves
Debels heuristic C++ 20 0.096 0.05 neighborhood search and diversification
steps. The second portion of algorithm
30 0.337 0.044 uses truncated dynamic programming to
40 0.811 0.098 relax non-critical activities.
50 1.605 0.114
2007a Yang PSO TCT-PF MATLAB 8 $500 48 na 0.406 A PSO algorithm capable of handling
28 na 600 na any function type which requires
manual calculations for subsequent
"what if" analysis. Indirect costs are not
provisioned throughout the
optimization process.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2007b Yang PSO TCT-PF na 14 $600 na na na Selection process of PSO method is
modified in favor of strongly dominant
solutions as pbest, and solutions of scant
areas as gbest. Indirect costs are not
included explicitly in the optimization
process.
2008 Eshtehardian, Fuzzy-GA St. TCT- na 18 $0 na na 0 Fuzzy set theory enables GA to handle
Afshar, and PF (Based stochastic TCTPs.
Abbasnia on 4

47
Obs.)
2008 Ng and Zhang ACS-TCO TCT-PF Visual 18 $500 na na 0 MAWA is integrated into ACO. This
Basic $1500 0.018 model is susceptible to premature
convergence with higher iterations and
it is too sensitive to selection of
parameters.
2008 Xiong and MAWA- TCT-PF na 7 na na na na MAWA is embedded into ACO and
Kuang ACO 18 $1500 selection of the options are made
according to membership of a random
variable, the first selection involving a
maximization criterion, and the other
incorporating a probability distribution
function.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2009 Afshar, Ziaraty, NA-ACO TCT-PF na 18 $0 na na 0 Multi-colony non-dominated archiving
Kaveh, and (Based ACO that assigns separate ant colonies
Sharifi on 44 to each objective and evaluates the found
Obs.) solutions respecting the competing
$200 0 objective within the next colony.
(Based
on 18
Obs.)
$1500 0
(Based
on 4
Obs.)

48
2009 Eshtehardian, Fuzzy- St. TCT- na 7 $500 na na na Fuzzy numbers comparison and GA are
Afshar, and MOGA PF 18 $200 900 0 utilized to locate PF based on the project
Abbasnia (Based manager's risk acceptance level and
on 20 degree of optimism. This model is
Obs.) operational only for project managers
familiar with the concepts of uncertainty
and risk management.
2010 Anagnostopoulos SA TCT Visual 100 1%, 0.18 to 30 na Performance of five variants of SA
and Kotsikas Basic 2%, 0.45 algorithm are analyzed and compared
200 and 0.99 to with each other.
3% of 1.92
300 Total 1.20 to
Cost 3.29
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2010 Geem HS TCT-PF MS Excel 7 $1500 na na 0 Compared with GA, all solution vectors
(Based are considered in generating a new
on 5 vector rather than two; it does not
Obs.) require binary-to-decimal conversion.
18 $0 5 2.27
(Based
on 19
Obs.)
2010 Hazir, Haourai, BD TCT-BT na 136 na na na na Original BD is modified with an
and Erel improved decomposition approach and
as well as a branch-and-cut procedure.
2010 Zhang and Li CSMO- TCT-PF Visual 7 $0 63.66 50 na A multi-objective PSO that combines
PSO C++ 18 205.08 sparse-degree and roulette-wheel

49
selection for determining the gbest of
each particle.
2011 Abdel-Raheem Electimize TCT VBA 18 $500 na 20 0 This method simulates the behavior of
and Khalafallah Macro electrons moving through electric
circuit branches with the least
resistance.
2011 Hazir, Erel, and BD St. TCT- na 136 na 19139.6 na na Three optimization models that assume
Gunalay DL 1 interval uncertainty for unknown
parameters are presented. All the
uncertainty is reflected on the cost of an
activity using probabilistic intervals.
2011 Kalhor, NA-ACO St. TCT- MATLAB 18 $200 29.42 10 0 α-cut approach accounts for accepted
Khanzadi, PF (Based risk level and β concept is used for
Eshtehardian, on 18 ranking solutions; left and right
and Afshar Obs.) dominance ranking method is used for
non-dominated sorting.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2012 Fallah- MOPSO TCT-PF na 18 $0 na na 0.944 Discrete versions of MOPSO and
Mehdipour, (Based NSGA-II are compared for solution of
Bozorg Haddad, on 1 Pareto front TCTP. NSGA-II is
Rezapour Tabari, Obs.) concluded to be more successful.
and Mariño
2012 Sonmez and HA TCT-DL Visual 18 $200 na 10 0 A hybrid GA combining potencies of
Bettemir C++ 29 €1200 0 SA along with quantum simulated
annealing (QSA).
290 €1200 0.43
TCT 18 $1500 0
29 €1200 0

50
63 $2300 2.61
63 $3500 2.5
290 €1200 0.74
630 $2300 4,380 2.41
630 $3500 4,380 2.47
2012 Szmerekovsky MIP TCT na 90 na 206 na na An MIP method capable of obtaining
and Venkateshan solutions for TCTPs with irregular cost
functions. This method can be adapted
for Net Present Value (NPV)
minimization of costs and Cash
Availability (CA) maximization.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2012 Zhang and Ng ACS TCT VBA 18 $500 na na 0 MAWA is applied to combine the time
TCT-PF Macro $1500 0.004 and cost objectives; ACS-SGPU
(Based guarantees diversity while ACS
on 4 provides quality solutions.
Obs.)
ACS- TCT-PF 0.586
SGPU (Based
on 5
Obs.)
2013 Aminbakhsh SAM- TCT-DL C++ 18 $200 0.08 10 0 Siemens method is modified and
PSO TCT 18 $0 0.08 0 embedded to a custom multi-objective
PSO for Pareto oriented optimization of
18 $200 0.08 0 discrete TCTP. Applicability of this

51
18 $500 0.08 0 approach is not documented for large-
18 $1,500 0.08 0 scale problems with more than 63
activities.
63 $2,300 45 0.02
63 $3,500 45 0.02
TCT-PF 18 $0 8.18 0.1
18 $200 7.98 0.08
18 $1,500 7.97 0.03
63 $2,300 111.51 0.31
63 $3,500 108.02 0.27
2013 Degirmenci and B&B TCT-BT C# 29 to na 494.88 na na A mode eliminator and a lower and
Azizoglu 35 to upper-bound calculator is embedded to
5,094.54 a B&B for network reduction purposes.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2013 Mungle, FCGA TCT-PF MATLAB 7 $0 17.22 50 0.128 Fuzzy clustering technique embeds
Benyoucef, Son, 12 130.86 na NSGA-II with crowded-comparison
abd Tiwari operator together with crowding
18 847.44 0.531 distance for selection and generation of
(Based population.
on 18
Obs.)
2015 Bilir MIP TCT-DL C# 50 $250 1.03 na na MIP approach for large-scale TCTPs
100 $500 15.3 which is tested using samples generated
by means of ProGen/Max.
200 $500 57.44
500 $500 96.39
1000 $500 174.04

52
TCT-PF 50 $250 302.67
100 $500 260.23
200 $500 487.27
2015 Koo, Hong, and iMOO TCT-PF OptQuest 7 $500 11 na na This GA-based approach is designed for
Kim (GA) optimization of four fitness functions
including trade-offs between time-cost,
cost-quality, sustainability-cost, and
productivity-safety.
2015 Vanhoucke Electro- TCT-DL Visual 10 na 0.049 to na 0.02 to This model is used for four different
magnetic C++ 20 2.869 98.6 variants of deadline TCTP with
Time/switch constraints, work
30 continuity constraints, and NPV
40 maximization considerations. Two local
50 searches are implemented for crashing
and relaxing activities.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2016 Aminbakhsh and DPSO TCT-DL C++ 18 $200 0.04 10 0 A PSO with novel principles for
Sonmez TCT 18 $500 0.04 0 representation, initialization and
position-updating of the particles which
18 $1,500 0.04 0 operates in discrete space and
63 $2,300 1.3 0.02 incorporates a semi-deterministic
initialization scheme.
63 $3,500 1.3 0.05
630 $2,300 14.6 0.33
630 $3,500 14.6 0.34
200 $500 4.6 0.19
500 $500 17.2 0.21

53
2016 Zou, Fang, MILP TCT-DL na 30 na 1.63 to na na A MILP and an alternative
Huang, and 44.08 approximation approach for larger
Zhang 40 2.55 to instances is proposed for repetitive
175.43 problems with fixed logic and multiple
50 3.95 to crews. This method assumes typical
546.19 projects with a single mode.
TCTP- 30 0.32 0.06
Approx. 40 0.74 0.04
50 1.78 0.06
80 56.05 na
90 104.42
100 643.4
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2016 El-Abbasy, MO- TCT-PF na 18 $0 na na na A GA-based model that considers
Elazouni, and SCOP-EA financial and resource aspects. Pure
Zayed TCTP results of GA is validated only by
visual comparison of TCT curves.
2017 Aminbakhsh and PFPSO TCT-PF C++ 18 $0 2 10 0 PSO is embedded with a simplified-
Sonmez 18 $200 2 0 SAM which unlike any previous
extension of SAM, finds the least-cost-
18 $1,500 2 0 slope critical activity by considering
180 $0 21 0.1 only a single critical path at a time. This
360 $0 43 na PSO employs the external repository
concept and redefines position update

54
720 $0 92
and global best calculation routines.
2017 Bettemir and Heuristic TCT-DL na 18 $200 na na 0 Elimination technique of the new
Birgonul TCT 18 $200 0 Heuristic contributes to faster
computations. Paths are considered
63 $2,300 0.046 critical if they do not include critical and
63 $3,500 0.077 non-critical activities simultaneously. A
project duration predictor is also
employed to reduce the number of
schedule evaluations. This is done by
only considering the crashed activities
and those with floats greater than
critical total float amount.
 Table 2.1 – Existing exact, heuristic, and meta-heuristic methods for solution of TCTP (Continued).

Year Author(s) Method Problem Platform # of IC CPU # of APD Remarks

Acts. Time (s) Runs (%)
2017 Dragović, LP TCT-DL MATLAB 10 to na na na na An LP model that is used for direct cost
Vulević, 13 minimization of river regulation and
Todosijević, check dam construction projects. This
Kostadinov, and model neglects daily indirect cost in the
Zlatić optimization process.
2017 Su, Qi, and Wei Heuristic TCT-DL LINDO 116 na 2 na na Equivalent simplification-based
approach which transforms nonlinear
continuous TCTP into a simple activity
float problem. It uses a complex
polynomial algorithm to solve the
transformed problems.
2018 Agdas, Warne, Parallel- TCT-DL Python 1800 $200 20,952 5 7.05 A parallel GA which is modified using

55
Osio-Norgaard, GA TCT 63 $2,300 111 0.26 new techniques including problem
and Masters encoding, weighted graph, parallel
63 $3,500 88.8 1.24 fitness evaluation, and stagnation
630 $2,300 364.2 2.76 criteria for solutions of large-scale
630 $3,500 396.6 2.29 problems.
1800 $1,500 21,024 14.72
3150 $2,300 32,940 6.5
3150 $3,500 33,876 4.73
6300 $2,300 59,112 7.66
6300 $3,500 60,336 6.96
Respecting the state of the existing research on TCTP in the construction industry,
it can be observed that GAs are the most prominent methods used in a multitude of
studies. Above all, it can be concluded that the majority of the articles address
instances with small problem networks and that detailed performance evaluation on
accuracy and efficiency of most of the presented approaches appear to be lacking.
In fact, performances of the algorithms are often only reported for the small
benchmark network of Feng et al. (1997) that includes 18 activities with up to five
time-cost alternatives. Kalhor et al. (2011), Fallah-Mehdipour et al. (2012) and
Mungle et al. (2013) are among the few studies reporting on different performance
metrics of their proposed models. Besides, it is worth mentioning that although a
large body of the literature has hitherto been dedicated to development of
optimization methods for DTCTP, only some of these methods are used in real-life
practices. That is largely resulting from the fact that they do not suit actual practices
and that major domain of the literature focus on proving applicability of various
optimization models rather than providing means for optimization of real
construction projects. To expand on this, very few of the existing methods can be
applied to optimization of real-life construction projects which typically comprise
more than 300 activities (Liberatore et al., 2001). Furthermore, a few methods that
are tested for real-life-size large-scale problems, require enormous computation
time and resources thanks to the inherent complexity of solving DTCTPs. The few
efforts that handle large-scale problems include parallel GA of Kandil and El-Rayes
(2006), constrained programming of Menesi, Golzarpoor, and Hegazy (2013), and
parallel GA of Agdas et al. (2018) employing problems including up to 720, 2000,
and 6300 activities, respectively. However, all the practiced large-scale problems
are based on small networks which are generated by copying the base problem in
serial several times; hence, these problems are believed to have limitations in
reflecting the complexity of the real-life construction projects. In addition, the few
studies practicing large-scale instances require an unreasonably large computation
time; for instance, GA-based approach of Kandil and El-Rayes (2006) require 15.4
hours over a supercomputing cluster of 50 processors to solve a 720-activity

56
problem. Likewise, GA-based approach of Agdas et al. (2018) require 16.7 hours
on a high-performance computing facility with eight CPU cores to unravel a 6300-
activity problem. Nevertheless, for any solution method to be practically viable,
accuracy needs to be accompanied with the efficiency. The main reasons behind the
gap between the theoretical achievements of researchers and practical applications
of professionals is not only related to the inherent computational complexity of
large-scale networks, but also due to the dearth of real-life-scale problems.
Therefore, presenting performance of large-scale, realistic TCTP instances is one
of the most important contributions of this thesis. In order for better evaluation of
the capabilities of the proposed optimization models, new sets of multi-mode large-
scale DTCT problems have been generated by means of random project instance
generators. Unlike a large body of the existing literature, performance
measurements are carried out by taking a more holistic approach that involves a set
of performance comparison indices. Hence, the presented findings are highly
relevant for real-life applications.

With regard to the summary of the past research given in Table 2.1, it can also be
observed that the literature is not rich with the more complex Pareto oriented
optimization in the domain of TCTP, especially for the discrete version of this
problem. This is despite the fact that obtaining non-dominated sets of solutions is
widely acknowledged as the ultimate resolution of TCTP studies (e.g. Zheng et al.,
2005; Yang, 2007b; Eshtehardian et al., 2008; Aminbakhsh and Sonmez, 2017).
This thesis study emphasizes the importance of discrete and Pareto front TCTPs
due to their practical relevance and because they facilitate expressing decision
makers preferences so that they can select the best solution based on their own
concerns. To the respect of this, new models with exceptional accuracy and
efficiency, applicable in real-life-scale projects are designed and developed in this
thesis. This research study aims to contribute to both researchers and practitioners
by tightening the gap between the literature and the real-world requirements of the
projects. The new proposed models include Mixed-Integer Linear Programming

57
techniques that use Gurobi solver version 6.0.5, new Particle Swarm Optimizers,
and new Cost-Slope Heuristics. For all the proposed methods, two variants have
been designed and developed to address both the deadline and the Pareto front
classes of DTCT problems. Since, exact procedures are the only methods
guaranteeing optimality of the solutions and that heuristics and meta-heuristics are
incapable of securing the optimality of the solutions, the proposed Mixed-Integer
Linear technique is used in performance evaluation of the developed heuristic and
meta-heuristic methods.

Last but not least, despite the fact that any scientific decision support tool would
have a pivotal role in the decision-making process, none of the commercial
scheduling software packages (e.g., Microsoft Project, Primavera) include tools or
modules for TCT analyses of the scheduling problems. Therefore, this thesis study
also presents integration of DTCTP optimization modules into Microsoft Project –
a widely used commercial planning software in the construction industry – by
means of an add-in which is capable of solving two variants of DTCTP, namely,
cost minimization/deadline and Pareto problems. The integrated modules include
both the proposed Particle Swarm Optimizer and the Cost-Slope Heuristic. By
means of this, the new models are envisioned to be applicable in real projects and
to suit the actual practices of construction managers. In the ensuing chapters,
characteristics of the proposed particle swarm optimizers, Cost-Slope Heuristics,
and Mixed-Integer Linear Programming methods developed for solution of
different extensions of discrete time-cost trade-off problems are going to be
presented. Meanwhile, since the approaches presented in this thesis are non-
domain-specific, they can easily be used for solution of similar optimization
problems.

58
 CHAPTER 3

DISCRETE PARTICLE SWARM OPTIMIZATION METHOD

FOR DTCTP

As is clear, there is a lack of Particle Swarm Optimization (PSO) exemplars in the

existing literature with the capacity to tackle realistic large-scale construction
DTCT problems efficiently and effectively to make them practicable in real-life
projects. Accordingly, this chapter includes the background and theoretical
properties of PSO method. This chapter also covers a background of Siemens
Approximation Method (SAM), a modified variant of which is used for
development of a new PSO model. Following the definitions and principles of the
base models, the proposed discrete particle swarm optimization method for cost
minimization and deadline extensions of DTCTP is described. Results of inclusive
comparative studies are presented which prove the outstanding performance of the
proposed approach with solid convergence capabilities.

3.1. Particle Swarm Optimization (PSO)

Studies on biological evolution and collective behavior extant in natural systems

such as animal herds, schools of fish, and flocks of bird established the primitive
initiatives for development of methods based on swarm intelligence. The earliest
precursor for swarm intelligence-based optimization approach encompass the first
paradigm of the PSO developed by Kennedy and Eberhart (Eberhart and Kennedy,
1995; Kennedy and Eberhart, 1995) who later introduced the binary version of this
algorithm for problems with discrete search spaces (Kennedy and Eberhart, 1997).

59
PSO is rooted upon imitating the choreography of bird flocks that communicate
together as they fly; therefore, the population is called “swarm”, while, the potential
solutions are named as “particles”. This algorithm conceptually resembles
evolutionary algorithms and vastly relies on stochastic procedures.

The system initializes with a population of random potential solutions. Particles

iteratively fly over the search space in explicit directions and are attracted to self-
attained historical best position (personal best; pbest ), as well as the best position
among the entire swarm (global best; gbest ). Each particle records the coordinates
associated with the best location it has visited so far. At each time step, particles
evaluate their own positions with respect to fitness criteria, then by comparing the
fitness values, they communicate to identify the particle located at the best position.
Each particle moves towards the best position using a velocity that incorporates
coordination of the personal best location as well; then, it evaluates the domain from
its new location, and the process reiterates until either the swarm reaches to a
predefined target, or a computational limit.

3.2. Siemens Approximation Method (SAM)

Siemens approximation method (SAM) (Siemens, 1971) is one of the first heuristics
developed for TCTP. SAM involves a logical and systematic procedure to minimize
the overall project cost. It was originally developed for the continuous TCT problem
and is capable of obtaining solution for convex nonlinear TCTPs by making
multiple piecewise linear curve approximations. The procedure initiates with the
construction of the project network, thereby, crashing critical activities one at a time
based on some specific rules. Of these rules, the most important is selection of a
critical activity with minimum amount of cost-slope. The act of crashing continues
until either all the activities are crashed, or those with normal duration have cost-
slopes greater than the daily indirect cost rate. Nevertheless, SAM accelerates the
project based on the minimum cost-slope calculated at each iteration with respect

60
to the all-normal (uncrashed) and all crashed options. However, in discrete
problems, it is possible to have more than two options for each of the activities.
Thus, m  1 number of cost-slopes can be calculated for an activity with m number
of alternatives. This study extends the use of SAM to discrete TCT problems by
modifying some of the major steps of this heuristic method. This modified variant,
hereafter called modified-SAM, is later embedded into the discrete PSO in order to
improve the quality of the initial swarm. The new SAM-based heuristic – similar to
the original method – is initiated with the construction of the project network using
the normal modes. The activity with minimum cost-slope is identified and crashed
according to Eq. (3.1). In case of a tie, the activity leading to a shorter project
duration is selected, if the tie is not broken, the activity with the smaller activity
number is selected.

CS j  (dc jk  dc j (k 1) )(d jk  d j (k 1) ) 1

(3.1)
 j  {1,..., S} ,  k  {1,..., m( j)}

where; S is the number of activities, m( j) is the number of available time-cost

options for activity j , dc jk is the direct cost of k th time-cost option of activity j ,

d jk is the duration of the k th time-cost option of activity j . The cost-slopes of the

first network are evaluated by setting k  m( j) initially; and then, decreasing the
value of option k , one at a time, as the j th activity is crashed.

3.3. Discrete Particle Swarm Optimization Method (DPSO)

Few researches have presented PSO algorithms for the DTCTP (Yang, 2007a;
Bettemir, 2009; Zhang and Xing, 2010; Fallah-Mehdipour et al., 2012) which have
operated in continuous space. In this thesis, a discrete particle swarm optimization
(DPSO) algorithm is designed to achieve an improved particle swarm
representation for cost minimization and deadline DTCTPs. Previous researches

61
employ random initialization schemes to produce the first generation of particles.
However, the proposed DPSO method is equipped with a semi-random pattern for
generation of the initial particles. To this end, DPSO is designed to create a certain
percentage, pct , of the initial population of size N by using the modified-SAM
method and generating the rest of the swarm randomly. The solutions obtained by
using modified-SAM are fed into the particle swarm optimizer of DPSO. Positions
for the remaining 1  pct percent of the population are set randomly. That is, for an
S -dimensional problem, PSO generates S number of random values for each
variable j which are selected from the feasible range {1, m( j)} . In the case of
TCTP, S resembles the number of activities and the feasible range {1, m( j)}
represents the available time-cost alternatives for the j th activity. Solutions are
denoted by x ijk which represents a binary value that holds i th particle’s position for

the j th activity, which can only have one k value equal to one while all the
remaining positions for the j th activity of the same particle are set to zero (

k 1
m( j )
x (jkt )  1 ). Logical sequencing of the activities for the generated particles are

implemented according to Eq. (3.2).

ES (jt )  d (jt )  ES l t   0 ,  l Sc j (3.2)

where; t is the generation number; d j is activity j ’s duration; ES j is the early

start time of j th activity; and Sc j are the immediate successors of the j th activity.

In the first iteration, bound by the feasible region [v max , v max ] , random velocity
(1)
vectors, v ijk , are determined for all the initial seeds (i.e., deterministic and random

particles). Velocities of the individuals are clamped to avoid swarm divergence.

Generation of the initial particles is completed with the determination of the pbest

, Pi t 0 and the gbest , Pgt 0 positions. The indices of the particle’s best and best

62
particle among the entire swarm are represented by i and g , respectively. Each
particle acquires the “best” positions, using Eq. (3.3) as follows:

1
Pijkt 0  Pgjk
t0
 x ijk (3.3)

Each particle i ’s fitness is evaluated with respect to the Eq. (3.4) and Eq. (3.5),
which minimize the sum of direct and indirect costs. Incentive and disincentive
payments are explicitly considered in project cost calculations.

S m 
D i  max  d jkt  x jkt   (3.4)
 j 1k 1 

 S m ( j ) (t ) (t )
  dc jk x jk  D i  ic  T  dp if T  0
 j 1 k 1
Ci  
 S m ( j ) (t ) (t )
  dc jk x jk  D i  ic  T  db otherwise
 j 1 k 1
(3.5)
 j  1, , S  ,  k  1, , m( j)

where; d jk and dc jk represent duration and direct cost of the k th options of the j

th activity, respectively; ic denotes the daily indirect cost rate; D i and C i represent
duration and cost of the project for the i th particle; D dl is the project deadline; dp
is the daily delay penalty; db denotes the daily bonus.

The qualities of the solutions are compared with each other according to Eq. (3.6):

uv if Cu  Cv (3.6)

63
which makes the discrimination in favor of decision vector u in case the total cost
of that particle is less than decision vector v ; while, in case of equality ( C u  C v )
discrimination is made in favor of the particle with smaller duration by Eq. (3.7).

uv if CD  CD
u
u
v
v
(3.7)

For the occasion in which both particles u and v have the same total costs and
durations, discrimination is made randomly. Pi ’s and Pg ’s are updated after

identification of the better fitted individuals. Particles are flown to their new
positions using the velocity vector given in Eq. (3.8), which incorporates Kennedy
and Eberhart’s (1997) linearly decreasing time-varying inertia weight ( w ) (Eq.
(3.9)). This parameter provides the PSO with more exploration ability at the initial
stages followed by more exploitation ability at the closing cycles. According to Eq.
(3.8), each particle updates its velocity using current velocity, the distance to
personal best experience, and the distance to best position of the entire swarm.

 t 1
v ijk t 
 w (t ) v ijk 
 c 1r1 Pijkt   x ijk
t 
 t 
 c 2r2 Pgjk t 
 x ijk  (3.8)

 t t 
w  w min   w max  w min   max  (3.9)
 t max  1 

where; r1 and r2 are random vectors with uniformly distributed components within
the range [0,1]; the constants c 1 and c 2 are the cognitive and social parameters,
respectively; w min and w max denote lower and upper-bounds of inertia weight,
respectively; and t max is the total number of iterations.

The values of c 1 and c 2 parameters control the convergence capabilities of a

particle by biasing its movement toward pbest or gbest positions, respectively.

64
Velocity vectors are transformed into probabilities (Aminbakhsh, 2013) and are
normalized to the range [0,1] using a logistic transformation function given in Eq.
(3.10). Each particle is then migrated to a new position subject to the probabilistic
condition according to Eq. (3.11).

 t 
sig v ijk 
1
t 
1  exp (v ijk )
(3.10)

(t 1)
1

if  t 1
sig (v ijk   t 1
)  max sig (v ijk 
)
x ijk  (3.11)
0 otherwise


t  (t 1)
where each sig (v ijk ) represents the probability that x ijk would be selected.

Eq. (3.11) differs from the position update equation of the binary PSO proposed by
Kennedy and Eberhart (1997), in that, it involves determination of the alternative(s)
associated with the maximum amount of probability for every activity. This
condition indicates that in each row of position matrix, a single alternative with the
largest probability will have a value of one. If the value of max sig (v ijk
(t 1)
) is same

for more than one alternative, then, discrimination is made randomly.

3.3.1. Case Example

Given in Figure 3.1, a hypothetical case example including six non-dummy and two
dummy activities is introduced herein which will be used to elucidate various
concepts, definitions, and applicability of the approaches that are proposed within
the context of this thesis study. This case example is used in this section to illustrate
the binary representation for particles’ positions.

65
 3

1. (16, $52,000)
2. (20, $45,000)
1 3. (28, $33,000)
1. (12, $21,600)
2. (14, $17,000)
3. (26, $14,000) 4 6

Start 1. (5, $52,000) 1. (3, $36,000) Finish

2. (8, $34,000) 2. (5, $24,000)
2

1. (7, $38,000)
2. (10, $26,000)
3. (12, $20,000) 5 ID

1. (10, $40,000) 1. (Dur., Cost)

2. (15, $24,000) m. (Dur., Cost)
3. (16, $22,000)

Figure 3.1 – Case Example.

Figure 3.2 illustrates the probability matrix for the 5th particle in the 2nd iteration for
the case example. The velocity matrix v 5(2) , which is calculated by Eq. (3.8), is

 
transformed to the probability matrix sig v 5 2 using the sigmoid function given in

Eq. (3.10).

Modes
Activities k1 k2 k3
j1 0.71 0.11 0.73
j2 0.94 0.94 0.97
j3 0.82 0.12 0.27
j4 0.75 0.96
j5 0.28 0.97 0.31
j6 0.69 0.51

Figure 3.2 – Probability matrix.

The 5th particle’s new position matrix ( x 5(3) ) in the 3rd iteration is determined by
using the Eq. (3.11) as shown in Figure 3.3.

66
 Modes
Activities k1 k2 k3
j1 0 0 1
j2 0 0 1
j3 1 0 0
j4 0 1
j5 0 1 0
j6 1 0

Figure 3.3 – Position matrix.

The discrete position matrix in Figure 3.3, assigns a value of “one” to the mode
selected, and a value of “zero” to the remaining modes, to make a clear
discrimination of the selected modes. For example, for the first activity the third
mode is selected, hence the position x 13 is assigned a value of “one”, and the
remaining positions of the first activity ( x 11 and x 12 ) are assigned a value of “zero”.
If a continuous position matrix was used instead, the probability of x 11 (0.71) would
be very close to that of x 13 (0.73), hence a clear discrimination between the first
mode and the third mode would not be made as the two probabilities are very close
to each other.

The optimization process is reiterated until the preset number of iterations is

reached. DPSO returns the final gbest particle as the solution for the DTCT
problem when the optimization process is terminated. The proposed DPSO method
is graphically explained as a flowchart in Figure 3.4 and the pseudo-code of this
algorithm is illustrated in Figure 3.5.

67
 Start

Construct Determine Calculate

Store Solution
Project Network Duration Total Cost

Feed Initial Crash Least All Critical

Evaluate No
Solutions to Cost Slope in Activities
Cost Slopes
PSO Critical Path Crashed
Yes

Randomly
Evaluate
Generate Determine
Duration versus
Remaining Duration
Deadline
Particles

Yes pbest Has Compare Calculate Total

Update pbest
Changed Fitness Cost

No

gbest Has Yes

Update gbest
Changed

No

Transform
Evaluate
Update Position Velocity to
Velocity
Probability

No

Yes Output Update Inertia

Stop Iteration
gbest Weight

End

Figure 3.4 – Flowchart of the proposed discrete PSO algorithm.

68
Begin
For ∀ j ∈ [1, S]
For ∀ k ∈ [1, m]
Retrieve Values;
End;
End;
Construct network;
While ∃ j ∈ Critical-Acts: Fully-Crashed True ∧ i  pct.N
Calculate CPM;
Calculate Dur/Cost;
Store Solution;
Calculate CS;
Crash Min-CS;
Set x i as pbest and gbest ;
Set random Velocity;
Break;
For ∀ i ∈ (pct.N, N]
For ∀ j ∈ Acts
Set random Position;
Set random Velocity;
Set xi as pbest and gbest ;
End;
End;
While t tmax
For ∀ i ∈ [1, N]
Calculate CPM;
Calculate Dur/Cost;
For ∀ j ∈ Acts
Set w;
If x i  Pi
Set x i as pbest ;
If Pi  Pg
Set Pi as gbest ;
End;
End;
Calculate Velocity;
Transform Velocity to Probability;
Update Position;
End;
Update w;
End;
Break;
Return gbest ;
End;

Figure 3.5 – Pseudo-code of the proposed discrete PSO algorithm.

69
3.4. Computational Experiments of DPSO

Computational experiments are conducted to evaluate the performance of the

proposed DPSO method for the DTCTP using a set of benchmark instances
acquired from the literature as well as a set of large-scale problems introduced in
this section. The proposed optimization algorithm is coded in C++ and compiled
within Visual Studio 2013 on a 64-bit platform. All of the tests are carried out on a
desktop computer with a P9X79 Chipset motherboard, 16 GB 667 MHz DDR3
random-access memory (RAM), Intel Core i7-3.40 GHz CPU, and 64-bit Windows
10 operating system. DPSO is executed solely (no other software is ran
simultaneously) on a single processor and overclocking is not performed.

3.4.1. Parameter Configuration of DPSO

It is broadly acknowledged that evolutionary algorithms are very sensitive to

configuration of their parameters and the proposed DPSO is not an exception to
this. Hence, pilot experiments were conducted to determine an adequate set of
parameter values for the DPSO. The pilot experiments revealed that the set of
parameters that are summarized in Table 3.1 provided an adequate combination for
the DPSO.

Table 3.1 – Parameter configuration of the DPSO.

Parameter Description Value

i # of Birds 250
pct % of deterministic swarm 0.8
c1 Cognitive Parameter 5
c2 Social Parameter 1
w max Max. Inertia Weight 1.2
w min Min. Inertia Weight 0.0
v max Max. Velocity 3.7

70
50,000 schedules (objective function evaluations) is used as the stopping criterion
in all of the experiments (Kolisch and Hartmann, 2006; Sonmez and Bettemir,
2012). Since PSO is a stochastic search algorithm, performance of DPSO is
evaluated for ten consecutive runs for each instance, and average percent deviation
(APD) from the global optimal solution is reported.

3.4.2. Small-Scale Benchmark Problems

The performance of the proposed DPSO method was first tested using the small-
scale DTCTP test instances which are commonly used in the literature (Elbeltagi et
al., 2005; Zheng et al., 2005; El-Rayes and Kandil, 2005; Kandil and El-Rayes,
2006; Elbeltagi et al., 2007; Ng and Zhang, 2008; Xiong and Kuang, 2008; Afshar
et al., 2009; Fallah-Mehdipour et al., 2012; Sonmez and Bettemir, 2012; Zhang and
Ng, 2012; Monghasemi, Nikoo, Khaksar Fasaee, and Adamowski, 2015) for
performance evaluations. The small-scale test instances consisted of an 18-activity
network (Feng et al., 1997) with the time-cost alternatives defined in Hegazy
(1999). Although it has not been pointed out by any other preceding study, it is
worth mentioning that this problem is flawed since the cost of third time-cost
alternative of the eighth activity must have been selected from the interval
DU(208, 215) ; however, the assigned cost is $200. Nevertheless, in order to
conduct a fair comparison with the previous studies, the benchmark problem is used
without applying any corrections. The activity on node (AoN) representation of this
instance can be obtained from Feng et al. (1997) and the time-cost data can be
attained from Hegazy (1999). This problem includes one activity with two modes,
ten activities having three modes, two activities with four modes, and five activities
with five modes; accounting for a total of 5.9 10 9 possible schedules. This
benchmark problem is examined under three different conditions. In problem 18a,
the indirect cost figure is $200/day, the delay penalty is set as $20,000/day, the
incentive payment is assumed as $1,000/day, and the completion deadline is
assigned as 110 days. In problem 18b, the indirect cost rate is $1,500/day, and in

71
problem 18c, the indirect cost is $500/day. The optimal solutions for the practiced
small-scale problems are determined by applying the mixed-integer programming
formulation given in Eqs. (1.1)-(1.6) using AIMMS 4.2 optimization software. The
optimal results are also verified by means of a mixed-integer linear programming
technique that employs Gurobi solver 6.0.5. This method is discussed in detail in
Section 5.2.3. The optimal solutions for problems 18a, 18b, and 18c are calculated
as $128,270, $271,270, and $161,270. Snapshots of the performance of the results
for problems 18a, 18b, and 18c are given in Table 3.2, Table 3.3, and Table 3.4,
respectively. The average percent deviations from global optima of DPSO compare
favorably with the existing methods which are summarized in the following tables.

Table 3.2 – Performance of DPSO for problem 18a.

Algorithm # of runs APD (%)

GA (Hegazy, 1999) 1 8.139
GA (Sonmez and Bettemir, 2012) 10 2.17
HA (Sonmez and Bettemir, 2012) 10 0.00
DPSO (Section 3.3) 10 0.00

Table 3.3 – Performance of DPSO for problem 18b.

Algorithm # of runs APD (%)

MAWA-GA (Zheng et al., 2005) 1 0.903
ACS-TCO (Ng and Zhang, 2008) 1 0.018
NA-ACO (Afshar et al., 2009) 1 0.00
ACS-SGPU (Zhang and Ng, 2012) 1 0.698
ACS (Zhang and Ng, 2012) 1 0.018
GA (Sonmez and Bettemir, 2012) 10 1.29
HA (Sonmez and Bettemir, 2012) 10 0.00
DPSO (Section 3.3) 10 0.00

72
 Table 3.4 – Performance of DPSO for problem 18c.

Algorithm # of runs APD (%)

GA (Elbeltagi et al., 2005) 20 2.171
MA (Elbeltagi et al., 2005) 20 0.759
PSO (Elbeltagi et al., 2005) 20 0.415
ACO (Elbeltagi et al., 2005) 20 3.351
SFL (Elbeltagi et al., 2005) 20 2.960
MSFL (Elbeltagi et al., 2007) 20 0.00
ACS-TCO (Ng and Zhang, 2008) 1 0.00
Electimize (Abdel-Raheem and Khalafallah, 2011) 20 0.00
ACS-SGPU (Zhang and Ng, 2012) 1 0.00
ACS (Zhang and Ng, 2012) 1 0.00
DPSO (Section 3.3) 10 0.00

In all of the ten trials for the three problems, DPSO was able to obtain the global
optimal results within 50,000 schedules by searching only 8.47 10 4 percent of
the entire solution space. As a result, DPSO was able to solve all the small-scale
test instances in 0.4 seconds on average.

The comparison of DPSO with the state-of-the-art methods proves that proposed
DPSO is among the top performing algorithms for the small-scale DTCTP. The
DPSO has outperformed the genetic algorithms (Hegazy, 1999; Sonmez and
Bettemir, 2012) for instance 18a (Table 3.2). Similarly, for instance 18b, DPSO was
better than the genetic algorithms (Zheng et al., 2005; Sonmez and Bettemir, 2012),
ant colony system time-cost optimization algorithm (ACS-TCO) of Ng and Zhang
(2008), and ant colony system (ACS) and ant colony system with global updating
strategy (ACS-SGPU) algorithms of Zhang and Ng (2012) for obtaining the optimal
solution. The proposed DPSO also outperformed the genetic, memetic, PSO, ant
colony optimization (ACO), and shuffled frog-leaping optimization (SFL)
algorithms of Elbeltagi et al. (2005) for the problem 18c.

The modified shuffled frog-leaping optimization algorithm (MSFL) (Elbeltagi et

al., 2007), the nondominated archiving ACO (NA-ACO) (Afshar et al., 2009), the

73
Electimize algorithm (Abdel-Raheem and Khalafallah, 2011), and the hybrid
genetic algorithm with simulated annealing (HA) (Sonmez and Bettemir, 2012) are
among the state-of-the-art methods that are competitive with the proposed DPSO
for obtaining high quality solutions for the small instances having no more than 18
activities. Though, the elapsed CPU time for MSFL was eight seconds compared
with 0.4 seconds of the DPSO. Besides, except for GA and HA (Sonmez and
Bettemir, 2012), none of the existing state-of-the-art methods were tested for larger
instances.

3.4.3. Medium-Scale Benchmark Problems

The medium-scale instances used for performance evaluation of DPSO consisted

of the 63-activity problem of Sonmez and Bettemir (2012). The activity on node
(AoN) topological representation and the time-cost data of this instance can be
obtained from Sonmez and Bettemir (2012). This benchmark problem contains two
activities with three modes, fifteen activities with four modes, and forty-six
activities having five modes, totaling 1.37 10 42 different schedules. This test
instance is fitted into the model by assuming two indirect cost figures. In the first
problem, 63a, the indirect cost is set as $2,300/day, and in the second problem, 63b,
the indirect expense is assumed as $3,500/day. The optimal solutions for the
practiced medium-scale problems are determined and verified by the methods
pointed out in Section 3.4.2. The MIP/AIMMS and the MILP/Gurobi approaches
provided the optimal solutions for problems 63a and 63b as $5,421,120 and
$6,176,170, respectively. By evaluating only 3.64 10 36 percent of the solution
space, DPSO achieved APD values of 0.02% and 0.05% for instances 63a and 63b,
respectively. The proposed DPSO algorithm was very successful for obtaining high
quality solutions for the medium-scale test instances and outperformed the sole-
genetic algorithm and hybrid-genetic algorithm (Sonmez and Bettemir, 2012) as
shown in Table 3.5.

74
 Table 3.5 – Performance of DPSO for problems 63a and 63b.

63a 63b
Algorithm
# of runs APD (%) # of runs APD (%)
GA (Sonmez and Bettemir, 2012) 10 5.86 10 5.16
HA (Sonmez and Bettemir, 2012) 10 2.61 10 2.50
DPSO (Section 3.3) 10 0.02 10 0.05

The contents of Table 3.5 shed some light on exceptional performance of DPSO for
solution of medium-scale problems. By searching the same amount of 50,000
solutions, HA was able to achieve APD values of 2.61% and 2.50%. DPSO was
able to determine very high-quality solutions, with very marginal deviations from
the optimal in just 1.3 seconds on average.

3.4.4. Large-Scale Benchmark Problems

The large-scale problem used for performance measurement of the proposed DPSO
comprise 630 activities. In fact, this problem is generated by copying the 63-activity
problem of Sonmez and Bettemir (2012) ten times in serial. The approach of
creating multiple copies of the base problem has the benefit of knowing the
expected optimal solution. This instance incorporates 20 activities with three
modes, 150 activities with four modes, and 460 activities having five modes,
totaling 2.38  10 421 possible realizations. This instance represents the size of a
realistic construction project and is studied under two assumptions regarding the
amount of the daily indirect cost. In the first problem, 630a, the indirect cost is set
as $2,300/day, and in the second problem, 630b, the indirect expense is assumed to
be $3,500/day. The optimal solutions for the practiced large-scale problems are
calculated as the multiples of the optimal solutions of 63-activity instance used in
Section 3.4.3. Accordingly, the optimal solutions for problems 630a and 630b are
determined as $54,211,200 and $61,761,700, respectively. The performance of the
DPSO for large-scale instances is summarized in Table 3.6. DPSO achieved very

75
successful results and outperformed the hybrid genetic algorithm (HA) (Sonmez
and Bettemir, 2012) for large-scale instances.

Table 3.6 – Performance of DPSO for problems 630a and 630b.

630a 630b
Algorithm
# of runs APD (%) # of runs APD (%)
GA (Sonmez and Bettemir, 2012) 10 8.83 10 7.50
HA (Sonmez and Bettemir, 2012) 10 2.41 10 2.47
DPSO (Section 3.3) 10 0.33 10 0.34

The APD values for instances 630a and 630b were 0.33% and 0.34%, respectively;
and the processing time of DPSO was 14.6 seconds on average. By searching only
50,000 solutions out of 2.38 10 421 potential solutions, DPSO was able to achieve
high quality solutions for the large-scale instances. HA was able to provide
solutions with APD values of 2.41% and 2.47% by evaluating the same amount of
50,000 schedules. It was also revealed by Bettemir (2009) that HA was able to
achieve an APD value of 2% by solving the 630-activity problem in 73 minutes
through 1,000,000 schedule evaluations. Compared with HA, DPSO was able to
achieve significantly better solutions in a much shorter computational time.

3.4.5. New Sets of Instances

As is clear, the network and the data for the large-scale benchmark problem
included in the literature are generated by copying a simpler problem several times
that might not fully reflect the complexity of some real-life-scale construction
projects. To this end, 80 new large-scale instances with more complex activity
networks are generated to further evaluate the performance of the proposed DPSO.
ProGen/Max network generator developed by Schwindt (1995) (Figure 3.6) is used
to generate the networks for the new sets of problems.

76
 Figure 3.6 – ProGen/Max Interface.

Project networks are developed using four different complexity indices which are
represented by Thesen restrictiveness coefficient in ProGen/Max. For each Thesen
restrictiveness coefficient value of 0.2, 0.4, 0.6, and 0.8, ten networks are generated
totaling 40 networks of size 200 and 40 networks with 500 activities. The
parameters of ProGen/Max are configured so as to let “Minimal number of Initial
acivities” and “Minimal number of terminal activities” be one and “Maximal
number of initial activities”, “Maximal number of terminal activities”, “Maximal
number of predecessors”, and “Maximal number of successors” be 20. Microsoft
Excel is used to generate time-cost alternatives for each network. First, the available
number of time-cost alternatives is decided by randomly selecting a value from the
interval DU(2,5) (i.e., discrete uniform distribution with parameters 2 and 5) for
each of the activities. Thereafter, using the procedure described in Akkan, Drexl,
and Kimms (2005), duration of each time-cost alternative is selected from the range
obtained by dividing the interval DU(3,123) into the number of alternatives. Cost
of time-cost alternatives are generated using convex cost functions as proposed by
Akkan et al. (2005). Cost for the normal (uncrashed) mode is randomly selected

77
from the range DU(100,50000) . Cost of the remaining alternatives are calculated
using Eq. (3.12). In addition, the indirect cost rate is set as $500/day for both the
networks sizes.

dc jk  dc j (k 1)  CS jk  (d j(k 1)  d jk )

(3.12)
 j  {1,..., S} ,  k  {2,..., m( j)}

where dc j is the direct cost of the k th alternative for the j th activity; CS jk is the

cost-slope which is generated randomly between the range U(0.1,1) ; d jk is the

duration of the k th options of the j th activity.

The new instances were solved to optimality by means of the mixed integer
programming model presented in Eqs. (1.1)-(1.6) and using Gurobi solver version
6.0.5; this method will later be discussed in detail in Section 5.2.3. Optimal
solutions were achieved for all the 40 instances with 200 activities, and for 36 of
the instances with 500 activities within a CPU time limit of five hours for each. The
results of the proposed DPSO at the end of the 50,000 schedules are compared with
the optimal solutions for the 76 instances that are solved by the mixed integer
programming method. As shown in Table 3.7, DPSO achieved an average 0.19%
deviation from the optimal solutions within an average CPU time of 4.61 seconds
for 40 instances with 200 activities. As illustrated in Table 3.8, it also succeeded an
average 0.21% deviation from the optimal solutions within an average CPU time of
17.24 seconds for 36 instances with 500 activities. Unavailable values are tabulated
as ‘na’ in Table 3.8.

78
 Table 3.7 – Performance of DPSO for 200-activity instances.

Optimal DPSO
Dur. Dur. CPU APD
Problem Cost ($) Cost ($)
(day) (day) Time (s) (%)
T200_1 1120 6,379,332 1141 6,393,871 3.58 0.23
T200_2 1006 5,809,665 875 5,847,638 3.58 0.65
T200_3 1534 6,124,407 1534 6,128,020 4.00 0.06
T200_4 1064 5,957,758 1066 5,979,991 3.75 0.37
T200_5 1264 5,726,445 1264 5,730,567 3.59 0.07
T200_6 1085 6,182,999 1087 6,184,392 3.61 0.02
T200_7 1316 6,430,572 1316 6,446,318 3.88 0.24
T200_8 1008 5,812,144 1005 5,832,495 3.72 0.35
T200_9 1186 5,863,689 1190 5,878,431 3.84 0.25
T200_10 1341 6,006,200 1344 6,020,848 3.66 0.24
T200_11 1690 6,364,254 1686 6,368,641 4.25 0.07
T200_12 1965 6,728,040 1966 6,734,137 4.16 0.09
T200_13 1744 5,955,149 1737 5,978,204 4.61 0.39
T200_14 2143 6,567,311 2143 6,569,344 4.81 0.03
T200_15 1420 5,939,672 1420 5,952,793 4.27 0.22
T200_16 2121 6,996,398 2121 7,030,999 4.80 0.49
T200_17 1534 6,075,423 1543 6,085,157 4.13 0.16
T200_18 1484 6,271,532 1484 6,285,156 4.20 0.22
T200_19 1398 6,114,428 1398 6,127,591 4.33 0.22
T200_20 1421 6,476,044 1422 6,481,519 4.41 0.08
T200_21 2194 6,671,598 2194 6,677,859 5.09 0.09
T200_22 1494 6,258,513 1519 6,281,031 4.81 0.36
T200_23 1625 6,422,509 1626 6,430,370 4.55 0.12
T200_24 2110 6,047,807 2110 6,056,395 4.97 0.14
T200_25 2170 6,544,127 2171 6,546,307 4.97 0.03
T200_26 2114 6,908,248 2114 6,915,186 5.24 0.10
T200_27 1988 6,728,935 1987 6,745,455 4.86 0.25
T200_28 1668 6,404,964 1668 6,429,866 4.70 0.39
T200_29 1849 6,792,867 1850 6,806,045 4.63 0.19
T200_30 1302 6,159,162 1292 6,178,588 4.55 0.32
T200_31 2139 6,440,172 2121 6,445,376 5.24 0.08
T200_32 1913 6,361,151 1913 6,372,306 5.44 0.18
T200_33 1733 6,254,230 1731 6,271,461 5.33 0.28
T200_34 1820 6,541,587 1810 6,547,852 5.39 0.10
T200_35 2446 6,966,652 2450 6,974,364 5.53 0.11
T200_36 1496 6,400,397 1496 6,408,169 5.28 0.12
T200_37 3235 7,404,801 3235 7,405,375 6.36 0.01
T200_38 2009 6,496,720 1999 6,507,846 5.27 0.17
T200_39 2380 6,504,650 2394 6,509,194 5.44 0.07
T200_40 2604 6,396,572 2599 6,399,542 5.69 0.05
Total 4.61 0.19

79
 Table 3.8 – Performance of DPSO for 500-activity instances.

Optimal DPSO
Dur. Dur. CPU APD
Problem Cost ($) Cost ($)
(day) (day) Time (s) (%)
T500_1 3155 14,960,582 3132 14,997,226 12.68 0.24
T500_2 2601 15,204,223 2597 15,234,850 11.39 0.20
T500_3 2586 15,449,875 2518 15,477,452 11.97 0.18
T500_4 2328 15,365,411 2255 15,410,429 11.39 0.29
T500_5 2774 15,120,770 2676 15,189,132 12.01 0.45
T500_6 na na 1934 15,401,892 11.45 na
T500_7 2999 15,524,026 3018 15,572,852 12.27 0.31
T500_8 2858 15,599,044 2870 15,642,713 11.99 0.28
T500_9 2910 15,059,136 2932 15,109,956 11.97 0.34
T500_10 2990 15,568,991 2990 15,601,912 12.48 0.21
T500_11 3763 15,726,437 3768 15,751,034 15.49 0.16
T500_12 2962 14,609,048 2962 14,686,697 13.93 0.53
T500_13 3968 15,234,093 3968 15,259,241 15.75 0.17
T500_14 3102 15,644,352 3094 15,689,067 14.66 0.29
T500_15 4425 16,040,457 4424 16,058,224 16.13 0.11
T500_16 5248 16,037,675 5252 16,052,318 16.84 0.09
T500_17 3140 16,250,370 3142 16,282,414 14.55 0.20
T500_18 3710 15,189,213 3709 15,207,328 15.60 0.12
T500_19 4017 15,707,877 4017 15,746,844 15.87 0.25
T500_20 3655 15,544,444 3657 15,556,715 15.49 0.08
T500_21 na na 2963 15,613,104 16.91 na
T500_22 3396 15,789,889 3406 15,830,346 17.26 0.26
T500_23 4633 16,926,601 4651 16,959,989 18.66 0.20
T500_24 5014 16,267,849 5014 16,330,184 18.94 0.38
T500_25 6379 18,519,422 6379 18,555,704 21.52 0.20
T500_26 5135 16,939,792 5100 16,972,161 19.72 0.19
T500_27 3521 16,046,011 3467 16,102,094 17.13 0.35
T500_28 3498 15,170,906 3502 15,208,054 16.94 0.24
T500_29 3087 15,361,432 2994 15,432,948 16.50 0.47
T500_30 na na 3050 15,074,188 15.97 na
T500_31 6518 18,446,098 6518 18,462,904 24.13 0.09
T500_32 8273 17,863,876 8273 17,864,920 25.69 0.01
T500_33 na na 3210 15,519,551 18.48 na
T500_34 6592 17,611,269 6588 17,641,231 21.67 0.17
T500_35 5643 16,910,284 5641 16,929,501 21.20 0.11
T500_36 4606 16,475,275 4603 16,486,769 20.08 0.07
T500_37 8824 19,047,488 8824 19,052,000 26.27 0.02
T500_38 7101 17,258,066 7111 17,263,832 24.25 0.03
T500_39 6706 16,986,421 6706 17,004,162 22.16 0.10
T500_40 7545 18,633,690 7545 18,641,215 25.98 0.04
Total 17.08 0.21

80
Computational results demonstrate the effectiveness and accuracy of DPSO for the
new instances that were more complex than the problems practiced by a major body
of the literature. The performance of DPSO was consistent for the large-scale
problems and was able to produce good feasible solutions in an acceptable
timeframe. The obtained results for more complex problems are considered
practically reasonable, since, an average deviation of 0.21% from the optima of
500-activity problems is not high. To the best of author’s knowledge, the proposed
DPSO is one of the first methods capable of obtaining high quality solutions for the
large-scale DTCTP within seconds.

A more comprehensive study on the performance of DPSO is given in Section 5.2.4

using a set of benchmark instances acquired from the literature. Based on
performance indices demonstrated in Section 5.2.2, effectiveness and efficiency of
this method is measured and compared with a new heuristic algorithm which is
presented in Section 5.1.3.

81
82
 CHAPTER 4

PARETO FRONT PARTICLE SWARM OPTIMIZATION

METHOD FOR DTCTP

Discussion on the contents of Table 2.1 revealed that the literature is not rich with
the more complex Pareto oriented optimization in the domain of TCTP, especially
for the discrete version of this problem. This is despite the fact that obtaining non-
dominated sets of solutions is widely acknowledged as the ultimate resolution of
TCTP analyses. This thesis study emphasizes the importance of discrete and Pareto
front TCTPs due to their practical relevance and because they facilitate expressing
decision makers preferences so that they can select the best solution based on their
own preferences. It was also discovered that there is a lack of Particle Swarm
Optimization (PSO) exemplars in the existing literature with the capacity to tackle
real-life-size Pareto front DTCT problems. Therefore, followed by an introduction
on the Pareto optimality, an efficient and effective multi-objective particle swarm
optimization (PSO) model is presented in this chapter. The Pareto front particle
swarm optimizer, hereafter called PFPSO, is treated with a simplified heuristic
method in order to improve the quality of the initial swarm for an accelerated
optimization process. PFPSO is equipped with high capacity to solve large-scale
Pareto front problems using unique principles for initialization, representation, and
position updating of the particles. The descriptions on background and theoretical
properties of PSO and Siemens Approximation Method (SAM) can be found in
Sections 3.1 and 3.2, respectively. The mutual aspects of DPSO (Section 3.3) and
PFPSO will not be repeated for sake of brevity and only the major modifications

83
will be pointed out in this chapter. Results of the comparative studies are presented
which reveal a performance unmatched by the existing meta-heuristic algorithms.

4.1. Pareto Optimality

Multi-objective problems, such as Pareto front TCTP, often do not have a single
optimum solution which makes all the objectives optimal simultaneously. However,
for such problems there exist a series of non-dominated solutions. The non-
dominated solutions cannot be further improved for one of the objectives and cannot
be further worsened for the others (Zhang and Li, 2010). In other words, compared
with dominated solutions, non-dominated ones not only are as good as in all
measures, but also are better in at least one of them (Zheng et al., 2004). Originally
introduced by Vilfredo Pareto, the series consisting of optimal solutions for a multi-
objective problem with conflicting objectives is known as the Pareto front or the
efficient frontier. The components located on this front are mutually non-dominated
with respect to multiple criteria. According to this concept, it is normally not
possible to improve one objective without sacrificing at least one other objective.

As such, Pareto front – also known as time-cost curve – extension of the TCT
problem is a multi-objective decision making problem, and any of its objectives
might reach their optimal amounts at miscellaneous points. Obtaining the Pareto
front for TCT problem in essence involves concurrent optimization of budget and
deadline TCTPs. Within this domain, the Pareto-dominance optimality is defined
according to the following conditions:

- With respect to Eq. (4.1), a decision vector u is said to weakly dominate another
solution v if and only if D u  D v and C u  C v which is denoted as u  v .

u  v iff  y : f y (u)  f y (v) (4.1)

84
- Regarding Eq. (4.2), solution v is considered to be dominated if and only if
there is already another solution u in the efficient frontier such that D u  D v
while C u  C v , and one of these inequalities holds strictly; this dominance rule
is notated as u  v .

u  v iff  y : f y (u)  f y (v) and  y : f y (u)  f y (v) (4.2)

- A shown in Eq. (4.3), decision vector u is said to strongly dominate solution v

if and only if D u  D v and C u  C v ; this dominance condition is denoted as
u  v .

u  v iff  y : f y (u)  f y (v) (4.3)

where; f y represents the y th objective function. Subsequently, solution u is said

to be Pareto optimal as  u, v  O if  v : v  u ; union of all of which forms the

Pareto front denoted by O .

4.2. Simplified Heuristic

Earlier, Aminbakhsh (2013) presented a method that was similar to SAM which
crashed critical activities considering their cost-slopes. Since it is perfectly feasible
for the construction schedules to have more than a single critical path, in the course
of this method, critical activities on all the multiple parallel critical paths were
discovered and the one with the minimum amount of cost-slope was crashed first.
However, particularly for large-scale problems with complex networks, the process
of identifying the activities having the least cost-slope from a massive pool of
critical activities contributes to a substantial computational burden. On the other
hand, any variation in selection of the alternatives modifies the project schedule
which requires rescheduling of the project for potential changes using the CPM

85
procedure. That is, this SAM-based method necessitates completing the topological
sorting and calculation of CPM iteratively for each cycle of crashing.

In this study, inspired by SAM, a simplified heuristic is developed and embedded

into a multi-objective particle swarm optimization method which significantly
improves the quality of the initial population. This heuristic is similar to the
modified-SAM method introduced in Section 3.2. Though, in the simplified
heuristic, instead of determining the complete set of least-cost-slope critical
activities for all of the multiple parallel critical paths, one critical activity is selected
by examining a single critical path at a time. Crashing is performed by selecting the
least-cost-slope activity on the first critical path identified. By doing so, an adequate
number of high quality solutions is achieved in a short amount of computation time.
The cost-slopes are calculated according to Eq. (3.1) given in Section 3.2. In case
there exist multiple least-cost-slope critical activities, this method selects the one
leading to a shorter project duration; if the tie is not broken, the activity with the
smaller activity number is selected. The procedure is repeated until all of the critical
activities in the latest schedule are fully crashed. In order to increase the number of
deterministic solutions of high quality, this method reiterates in a similar fashion,
by selecting the activity leading to a longer project duration in case of multiple
least-cost-slope critical activities. The non-dominated solutions obtained by means
of the simplified heuristic are then fed into the particle swarm optimizer.

4.2.1. Case Example

The same case example given in Figure 3.1 of Section 3.3.1 is used to describe the
simplified heuristic process. The indirect cost for the case example is assumed to
be $1,000/day. For the all-normal (uncrashed) schedule, the project duration is 59
days and the total cost is $206,000 as shown in Table 4.1.

86
 Table 4.1 – Candidate solutions found by simplified heuristic.

# of Dur. Direct Indirect Total

Activity Crash
Schedule (day) Cost ($) Cost ($) Cost ($)
1 Normal 59 147,000 59,000 206,000
2 1 M3 to M2 47 150,000 47,000 197,000
3 3 M3 to M2 39 162,000 39,000 201,000
4 3 M2 to M1 35 169,000 35,000 204,000
5 1 M2 to M1 33 173,000 33,000 206,600
6 5 M3 to M2 33 175,000 33,000 208,600
7 6 M2 to M1 31 187,000 31,000 218,600

In the Schedule-1 which consists of normal modes, activities 1, 3, and 6 are on the
critical path. Among the critical activities, Activity-1 with a cost-slope of $250/day
has the minimum cost-slope as shown in Table 4.2, and is crashed first, to its second
mode (M2).

Table 4.2 – Cost-slopes of crash modes.

Activity Crash Mode Cost-slope ($/Day)

1 M3 to M2 250
1 M2 to M1 2,300
2 M3 to M2 3,000
2 M2 to M1 4,000
3 M3 to M2 1,500
3 M2 to M1 1,750
4 M2 to M1 6,000
5 M3 to M2 2,000
5 M2 to M1 3,200
6 M2 to M1 6,000

The resulting schedule Schedule-2, has a duration of 47 days, and a total cost of
$197,000 (Table 4.1). In Schedule-2, activities 1, 3, and 6 are still on the critical
path, and among the critical activities Activity-3 has the minimum cost-slope and
is crashed to its second mode. In Schedule-3, Activity-3 is crashed again as it has
the minimum cost-slope among critical activities. The resulting schedule Schedule-
4, has a duration of 35 days, and a total cost of $204,000 and the critical paths
include activities 1, 2, 3, 5, and 6, with activities 2, 5, and 6 yet to be crashed.

87
Similarly, crashing is continued with Activity-5 by crashing it to its second mode.
Finally, Activity-6 with the least-cost-slope is crashed to its first mode to obtain a
solution of 31 days and total cost of $218,600 as shown in Table 4.1.

The solution that is included in Schedule-1 is dominated by schedules 2, 3, and 4.

Likewise, the solution obtained in Schedule-6 is dominated by Schedule-5. The
remaining non-dominated solutions that are included in schedules 2, 3, 4, 5, and 7
are recorded as the Pareto front solutions at the end of the first heuristic run. The
procedure is repeated while considering the activity leading to a longer project
duration in case of a tie between the critical activities during crashing. The second
heuristic run also provides exactly the same solutions since in none of the crashing
cycles of the heuristic runs a condition for discrimination arises. Although it is not
encountered in this specific case problem, should there be a critical path consisting
of only Activity-4 and Activity-6, since they both have equal rates of cost-slopes
(6,000) for crashing to their first modes, in the first heuristic run the 4th activity, and
through the second heuristic run the 6th activity would have been selected. At the
end of second heuristic run, the non-dominated solutions are stored, certain
percentage of which is transferred to the particle swarm optimizer.

4.3. Pareto front Particle Swarm Optimizer (PFPSO)

Particle swarm optimization is a stochastic, population-based computational

optimization method which imitates choreography of bird flocks that forage and fly
in unison (for a more comprehensive definition, readers are referred to Section 3.1).
Few researches have presented particle swarm optimization methods for single
objective DTCTP (Elbeltagi et al., 2005), multi-objective time-cost optimization
(Yang, 2007b; Zhang and Li, 2010), times-cost-quality optimization (Fallah-
Mehdipour et al., 2012; Zhang and Xing, 2010), and multi-objective time-cost-
resource optimization (Ashuri and Tavakolan, 2012). A relatively scant work
(Zhang and Li, 2010) is carried out to adopt PSO in locating Pareto front for time-

88
cost trade-off problems. Existing particle swarm optimization methods are designed
to operate in continuous space and are often tested using the small 18-activity
problem of Feng et al. (1997). The original single-objective PSO is designed to
administer a dominance-based approach for selection of the best particles. This
approach cannot guarantee the diversity of the solutions located along the Pareto
front, especially for problems with continuous or dense solution spaces. On the
other hand, the few studies that extend PSO for optimization of multi-objective
problems use aggregating methodology such as weighting or ε-constraint
techniques which transform the problem into a single-objective optimization
problem with different combinations of weighting and constraints. The method
proposed in this chapter aims to achieve diverse non-dominated solutions for Pareto
front DTCTP by introducing a new scheme for selection of the best particle by using
a Pareto-oriented methodology.

Similar to the DPSO which is covered in Section 3.3, the proposed PFPSO operates
in discrete space to present an adequate solution architecture for the DTCTP. While
previous research incorporates random generation of the initial particles, the
proposed PFPSO incorporates a semi-random initialization scheme presented in
Section 3.3. In contrast to DPSO, PFPSO is designed to employ the simplified
heuristic introduced in Section 4.2 for generation of the deterministic portion of the
initial population. Solutions of simplified heuristic are fed into particle swarm
optimizer to improve swarm optimization, by starting the search from high quality
solutions. Just like DPSO, the remaining part of the initial swarm population are
created randomly to achieve diversification. The maximum percentage of the
heuristic solutions that will be included in the initial swarm population is
determined by the parameter, percentage ( pct ). An external archive, O , is
dedicated to the PFPSO, to store all of the non-dominated solutions identified.
Primarily, this repository stores the non-dominated solutions obtained in the
heuristic phase, which is then used to record the new non-dominated particles
identified through the subsequent stages of the PFPSO. The remaining initial

89
population is then generated using a random generation scheme. A controller is
implemented in the optimizer to identify the non-dominated particles to be stored
in the external archive. As given in Eq. (4.4), for any decision vector u , this
controller applies the following criteria based on the measured values of D u and
Cu :

 Accept if Du  Dv

  D u  D v

 or  (4.4)
 C u  C v

 Reject otherwise

where; D v and C v represent duration and cost of solution v – an existing solution

in the archive O . Accordingly, through each iteration the size of the external
archive changes dynamically by storing the non-dominated particles while
discarding the dominated ones.

Logical sequencing of activities, selection of best positions, objective function

evaluations, comparisons, velocity updating, and position updating of the particles
are carried out according to Eqs. (3.2), (3.3), (3.4), (3.5), (3.6), (3.7), (3.8), (3.9),
(3.10), and (3.11), respectively.

Since the optimal solution for a multi-objective problem comprise a set of non-
dominated solutions rather than a single optimum solution, identification of the
global best particle from the archived non-dominated ones is crucial to the Pareto-
oriented PSO. To the respect of this, PFPSO incorporates a multiple global best
approach to determine the Pg position of the particles, in which gbest is selected

randomly from the external repository of non-dominated particles at each iteration.

Hence, in multiple global best approach, equal chance is given to each of the
archived solutions in gbest selection. Since the objective of Pareto front DTCTP

90
is to locate the non-dominated time-cost profile over the set of feasible project
durations, all the archived non-dominated Pareto front solutions are considered to
be of equal quality and are given equal chance to be the gbest . Besides, unlike
some of the previous multi-objective PSO’s that record only parts of the non-
dominated solutions for faster calculations (Coello, Pulido, and Lechuga, 2004),
PFPSO considers all the non-dominated particles within the archive for selection of
the global best particle. The multiple global best approach also enables dynamic
exploitation of the archived solutions to locate additional Pareto front solutions and
prevents converging to a local optimum.

A 3D illustration of PFPSO’s position update scheme for a hypothetical swarm of

eight particles for a 3-activity problem is shown in Figure 4.1 and Figure 4.2, in
which six non-dominated solutions ( Pg ’s) are archived in the external repository.

The first member of the swarm ( i1 ) flies toward a new position ( i 1  ) using a velocity
vector which involves a randomly selected archive member ( Pg 4 ) and the first

particle’s pbest ( P1 ) position, as shown in Figure 4.1.

91
 Figure 4.1 – Schematic diagram of PFPSO at time-step t .

Figure 4.2 – Schematic diagram of PFPSO at time-step t +1.

92
The external archive is updated according to the new position of the first particle.
The positions of the particles i 2 through i 8 and the external archive are updated

similarly. The procedure is repeated for the new position of the first particle ( i1  )
in the next iteration along with the new randomly selected archive member ( Pg 2 ).

The i th particle proceeds to its new position ( i 1  ) at time-step t  1 as shown in

Figure 4.2. The optimization process is repeated until the pre-specified number of
iterations is reached. PFPSO terminates by returning the ultimate archived non-
dominated solutions after discarding the dominated ones according to Eq. (4.5).

u, v  O if  u  v : Remove v (4.5)

The simplified heuristic, initial swarm population creation, and position updating
stages of the proposed PFPSO are graphically explained as a flowchart in Figure
4.3 and the pseudo-code of this method is demonstrated in Figure 4.4. The proposed
PFPSO starts the search with the initial swarm consisting of the deterministic
solutions of the simplified heuristic and random particles of PSO, as shown in
Figure 4.3. Afterwards, PFPSO iteratively improves the entire population and
identifies the non-dominated ones to be stored in the set containing the Pareto
solutions.

93
 Start

Construct Determine Calculate

Non-dominated
Project Network Duration Total Cost
Yes

No

Crash Least All Critical

Evaluate No Store Solution
Cost Slope in Activities
Cost Slopes in Archive O
Critical Path Crashed

Yes

Feed Initial Randomly

Seeds of Generate Determine
Archive O to Remaining Duration
PSO Particles

No

Compare Store Solution Yes Calculate Total

Non-dominated
Fitness in Archive O Cost

No

pbest Has Yes gbest Has Yes

Update pbest Update gbest
Changed Changed

No

Transform
Evaluate
Update Position Velocity to
Velocity
Probability

No

Yes Output Update Inertia

Stop Iteration
Archive O Weight

End

Figure 4.3 – Flowchart of the proposed Pareto front PSO algorithm.

94
Begin;
For ∀ j ∈ [1, S]
For ∀ k ∈ [1, m]
Retrieve Values;
End;
End;
Construct network;
While ∃ j ∈ Critical-Acts: Fully-Crashed True ∧ i  pct.N
Calculate CPM;
Calculate Dur/Cost;
If ∀ Sol ∈ O: (Sol.Dur ≠ Dur) ∨ (Sol.Dur = Dur ∧ Sol.Cost ≤ Cost)
Add Sol to O;
End;
Calculate CS;
Crash Min-CS;
Set x i as pbest and gbest ;
Set random Velocity;
Break;
For ∀ i ∈ (pct.N, N]
For ∀ j ∈ Acts
Set random Position;
Set random Velocity;
Set xi as pbest and gbest ;
End;
Calculate CPM;
Calculate Dur/Cost;
If ∀ Sol ∈ O: (Sol.Dur ≠ Dur) ∨ (Sol.Dur = Dur ∧ Sol.Cost ≤ Cost)
Add Sol to O;
End;
End;
While t tmax
For ∀ i ∈ [1, N]
Calculate CPM;
Calculate Dur/Cost;
For ∀ j ∈ Acts
Set w;
If ∀ Sol ∈ O: (Sol.Dur ≠ Dur) ∨ (Sol.Dur = Dur ∧ Sol.Cost ≤ Cost)
Add Sol to O;
End;
If x i  Pi
Set x i as pbest ;
End;
Set a random Sol ∈ O as gbest ;
Calculate Velocity;

Figure 4.4 – Pseudo-code of the proposed Pareto front PSO algorithm.

95
 Transform Velocity to Probability;
Update Position;
End;
Transform Velocity to Probability;
Update Position;
End;
Update value of w;
End;
Break;
For ∀ Sol ∈ O: Non-dominated ≠ True
Remove Sol;
End;
Return O;
End;

Figure 4.4 – Pseudo-code of the proposed Pareto front PSO algorithm (Continued).

4.4. Computational Experiments of PFPSO

Computational experiments are conducted to evaluate the performance of the

proposed PFPSO model for solution of Pareto front DTCTP using a set of
benchmark instances acquired from the literature. The proposed optimization
algorithm is coded in C++ and compiled within Visual Studio 2013 on a 64-bit
platform. All of the tests are carried out on a desktop computer with a P9X79
Chipset motherboard, 16 GB 667 MHz DDR3 random-access memory (RAM),
Intel Core i7-3.40 GHz CPU, and 64-bit Windows 10 operating system. PFPSO is
executed solely (no other software is ran simultaneously) on a single processor and
overclocking is not performed.

4.4.1. Parameter Configuration of PFPSO

Evolutionary algorithms are broadly recognized to be very sensitive to

configuration of their parameters and the proposed PFPSO is not an exception to
this. Therefore, pilot experiments were conducted to determine an adequate set of
parameter values for the PFPSO. The parameters of PFPSO were configured based
on the experiments which included all combinations of two parameter levels (low

96
and high) for seven parameters. The pilot experiments revealed that the set of tuned
parameters that are summarized in Table 4.3 provided an adequate combination for
the PFPSO.

Table 4.3 – Parameter configuration of the PFPSO.

Factor Levels
Parameter Description Selected Value
Low High
i # of Birds S 3S 3S
pct % of deterministic swarm 0.5 0.8 0.8
c1 Cognitive Parameter 1 2 2
c2 Social Parameter 1 2 2
w max Max. Inertia Weight 1 1.2 1
w min Min. Inertia Weight 0 0.4 0.4
v max Max. Velocity 3 6 3

250,000 schedules (objective function evaluations) is used as the termination

criterion in all of the experiments (Kolisch and Hartmann, 2006; Sonmez and
Bettemir, 2012). The performance of the PFPSO is explored for the Pareto front
optimization of small, medium, and large-scale benchmark DTCTP problems.

4.4.2. Small-Scale Benchmark Problems

The performance of the proposed Pareto front particle swarm optimizer is first
tested with the well-known DTCTP benchmark problems that include 18 activities
and up to five time-cost modes. The activity on node (AoN) representation of this
instance can be obtained from Feng et al. (1997) and the time-cost data can be
attained from Hegazy (1999). As mentioned earlier, even though it has not been
pointed out by any other preceding study, it is worth mentioning that this problem
is flawed since the cost of third time-cost alternative of the eighth activity must have
been selected from the interval DU(208, 215) ; however, the assigned cost is $200.
Nevertheless, in order to conduct a fair comparison with the previous studies, the
benchmark problem is used without applying any corrections. The first test problem

97
(18d) consisted of the DTCTP presented in Feng et al. (1997) which had an indirect
cost of $0/day. Problem 18d includes one activity with a single mode, ten activities
having three modes, two activities with four modes, and five activities with five
modes; accounting for a total of 2.95 10 9 possible schedules. The small-scale
benchmark problems 18e, 18f, and 18g were slightly modified version of the same
problem (Hegazy, 1999) with an indirect cost of $0/day, $200/day, and $1,500/day,
respectively. 18e, 18f, and 18g all include one activity having two modes, ten
activities with three modes, two activities with four modes, and five activities
having five modes; accounting for a total of 5.9 10 9 possible schedules. Although
construction projects would typically include an indirect cost, many of the previous
studies on Pareto front optimization (Feng et al., 1997; Kandil and El-Rayes, 2006;
Afshar et al., 2009; Geem, 2010; Zhang and Li, 2010) used problems with an
indirect cost of $0/day in their performance evaluations. Therefore, the same
problems were also practiced in the performance evaluation of PFPSO. The optimal
solutions for the practiced small-scale problems are determined by applying the
mixed-integer programming formulation given in Eqs. (1.1)-(1.6) using AIMMS
4.2 optimization software. The optimal results are also verified by means of a
mixed-integer linear programming technique that employs Gurobi solver 6.0.5.
This method is explained in detail in Section 5.2.3. Snapshots of the performance
of the PFPSO for the four small-scale DTCTP benchmark problems against those
of other previous multi-objective optimization techniques is given in Table 4.4.
Unavailable values are tabulated as ‘na’ in Table 4.4.

98
 Table 4.4 – Performance comparison of PFPSO for small-scale problems.

CPU Time # of Pareto

Method Problem
(s) Front Solutions
MAWA-GA (Zheng et al., 2005) 18g na 4
ACS-TCO (Ng and Zhang, 2008) 18g na 4
NA-ACO (Afshar et al., 2009) 18e na 44
18f na 18
18g na 4
Fuzzy-MOGA (Eshtehardian et al., 2009) 18f 900 18
Harmony Search (Geem, 2010) 18d 5 19
CSMO-PSO (Zhang and Li, 2010) 18e 205 42
ACS (Zhang and Ng, 2012) 18g na 4
PFPSO (Section 4.3) 18d 2 39
18e 2 44
18f 2 18
18g 2 4

In problem 18d, PFPSO determined 39 non-dominated solutions in 2 seconds which

were significantly more than the 19 non-dominated solutions of the harmony search
method of Geem (2010). The solutions of the PFPSO for the 19 project durations
reported in Geem (2010) are presented along with the solutions of the harmony
search method in Table 4.5.

Table 4.5 – Comparison of 19 non-dominated solutions for problem 18d.

Cost ($)
Duration (day) PFPSO
Geem (2010)
(Section 4.3)
105 127,320 127,270
106 127,100 127,020
107 126,900 126,770
108 119,415 119,270
109 119,070 119,020
110 118,915 118,770
112 118,620 118,470
114 105,270 105,270
115 105,100 105,020

99
 Table 4.5 – Comparison of 19 non-dominated solutions for problem 18d (Continued).

Cost ($)
Duration (day) PFPSO
Geem (2010)
(Section 4.3)
116 104,770 104,770
118 104,470 104,470
119 104,270 104,220
120 104,020 103,970
122 103,850 103,720
124 103,070 103,070
125 102,908 102,820
126 102,708 102,570
128 102,400 102,320
131 102,320 102,170

The results of Table 4.5 can be summarized as follows. The solutions of the PFPSO
in comparison with those of the harmony search are of higher quality. This
comparison shows PFPSO is able to achieve better solutions for 15 project
durations while obtaining the same results for the remaining four durations.

The combined scheme-based multi-objective particle swarm optimization (CSMO-

PSO) method of Zhang and Li (2010) was able to obtain 42 non-dominated
solutions for 18e in 205 seconds. PFPSO located 44 non-dominated solutions in 2
seconds for the same problem. The proposed Pareto front particle swarm optimizer
captured larger number of non-dominated solutions in a significantly less
computation time compared with CSMO-PSO of Zhang and Li (2010). The multi-
colony ant algorithm of Afshar et al. (2009) also achieved 44 non-dominated
solutions for problem 18e, however, the computational time requirement of this
method was not reported.

PFPSO was able to capture 18 non-dominated solutions for 18f in 2 seconds.

Likewise, the multi-objective genetic algorithm (MOGA) of Eshtehardian et al.
(2009) is shown to be able to achieve 18 non-dominated solutions for the same
problem, but it requires 900 seconds to locate the Pareto front. One genetic (Zheng

100
et al., 2005) and three ant colony algorithms (Ng and Zhang, 2008; Afshar et al.,
2009; Zhang and Ng, 2012) reported four non-dominated solutions for 18g.
Computation times of the methods were not included. PFPSO also obtained four
non-dominated solutions for 18g in 2 seconds. All of the four non-dominated
solutions of the PFPSO were better than those of Zheng et al. (2005), and one of
the non-dominated solutions was better than the solution of Ng and Zhang (2008)
and Zhang and Ng (2012) as shown in Table 4.6.

Table 4.6 – Comparison of four non-dominated solutions for problem 18g.

Zheng et Ng and PFPSO

Duration Zhang and Ng
al. Zhang (Section
(day) (2012)
(2005) (2008) 4.3)
100 287,720 283,320 283,320 283,320
101 284,020 279,820 279,820 279,820
104 280,020 276,320 276,320 276,320
110 273,720 271,320 271,320 271,270

The computational experiments reveal that the performance of the PFPSO on the
well-known small-scale benchmark problems is unmatched by the previous meta-
heuristic methods. Results indicate that PFPSO can locate large number of high
quality Pareto front solutions. Although the test configuration details, such as the
CPU speed, are not available for majority of the previous approaches, the results
reveal that the computation time requirement of PFPSO is significantly less than
the time requirement of the existing methods. The computational efficiency of
PFPSO is more prominent in comparison with the existing methods that are capable
of locating large number of Pareto front solutions for the small-scale problems.

In addition to the comparative study with the existing multi-objective methods, the
convergence degree of PFPSO is also investigated by means of the optimal Pareto
front solutions obtained using mixed-integer programming technique. The optimal
solutions prove that the time-cost curves captured by PFPSO for problems 18d, 18e,
18f, and 18g were indeed true Pareto fronts consisting of the optimum non-

101
dominated solutions. True Pareto fronts generated by PFPSO for problems 18d,
18e, 18f, and 18g are illustrated in Table 4.7, Table 4.8, Table 4.9, and Table 4.10,
respectively.

Table 4.7 - Performance of PFPSO for problem 18d.

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
104 132,270 132,270 0 51112311 5 3 1 4 1 1 1 1 3 3
105 127,270 127,270 0 5 1 1 1 1 3 1 1 5 3 1 4 1 1 1 1 3 3
106 127,020 127,020 0 4 1 1 1 1 3 1 1 5 3 1 4 1 1 1 1 3 3
107 126,770 126,770 0 3 1 1 1 1 3 1 1 5 3 1 4 1 1 1 1 3 3
108 119,270 119,270 0 5 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 3 3
109 119,020 119,020 0 4 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 3 3
110 118,770 118,770 0 3 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 3 3
112 118,470 118,470 0 5 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 2 3
113 118,220 118,220 0 4 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 2 3
114 105,270 105,270 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
115 105,020 105,020 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
116 104,770 104,770 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
118 104,470 104,470 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 2 3
119 104,220 104,220 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 2 3
120 103,970 103,970 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 2 3
122 103,720 103,720 0 3 1 1 1 1 1 1 1 5 3 1 3 1 1 1 1 2 3
124 103,070 103,070 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
125 102,820 102,820 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
126 102,570 102,570 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
128 102,320 102,320 0 3 1 1 1 1 1 1 1 5 3 1 3 1 1 1 1 1 3
131 102,170 102,170 0 2 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
132 101,970 101,970 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 2
133 101,820 101,820 0 4 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
134 101,570 101,570 0 3 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
137 101,510 101,510 0 3 1 1 1 1 1 1 1 4 3 1 1 1 1 1 1 1 3
138 101,470 101,470 0 5 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 2
139 101,170 101,170 0 2 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
140 100,970 100,970 0 3 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 2
142 100,870 100,870 0 1 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
143 100,770 100,770 0 3 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 1

102
 Table 4.7 - Performance of PFPSO for problem 18d (Continued).

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
145 100,570 100,570 0 2 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 2
148 100,270 100,270 0 1 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 2
151 100,070 100,070 0 1 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 1
154 100,010 100,010 0 1 1 1 1 1 1 1 1 4 3 1 1 1 1 1 1 1 1
156 99,950 99,950 0 1 1 1 1 1 1 1 1 3 3 1 1 1 1 1 1 1 1
158 99,900 99,900 0 1 1 1 1 1 1 1 1 3 2 1 1 1 1 1 1 1 1
159 99,870 99,870 0 1 1 1 1 1 1 1 1 2 2 1 1 1 1 1 1 1 1
161 99,820 99,820 0 1 1 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 1
169 99,740 99,740 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
APD (%) 0

Table 4.8 – Performance of PFPSO for problem 18e.

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
100 133,320 133,320 0 51112311 5 3 2 4 1 1 2 1 3 3
101 128,320 128,320 0 51111311 5 3 2 4 1 1 2 1 3 3
102 128,070 128,070 0 4 1 1 1 1 3 1 1 5 3 2 4 1 1 2 1 3 3
103 127,820 127,820 0 3 1 1 1 1 3 1 1 5 3 2 4 1 1 2 1 3 3
104 120,320 120,320 0 5 1 1 1 1 2 1 1 5 3 2 4 1 1 2 1 3 3
105 120,070 120,070 0 4 1 1 1 1 2 1 1 5 3 2 4 1 1 2 1 3 3
106 119,820 119,820 0 3 1 1 1 1 2 1 1 5 3 2 4 1 1 2 1 3 3
107 119,770 119,770 0 3 1 1 1 1 2 1 1 5 3 1 4 1 1 2 1 3 3
108 119,270 119,270 0 5 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 3 3
109 119,020 119,020 0 4 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 3 3
110 106,270 106,270 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 3 3
111 106,020 106,020 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 3 3
112 105,770 105,770 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 3 3
114 105,270 105,270 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
115 105,020 105,020 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
116 104,770 104,770 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
118 104,470 104,470 0 51111111 5 3 1 4 1 1 1 1 2 3
119 104,220 104,220 0 41111111 5 3 1 4 1 1 1 1 2 3
120 103,970 103,970 0 31111111 5 3 1 4 1 1 1 1 2 3

103
 Table 4.8 – Performance of PFPSO for problem 18e (Continued).

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
121 103,820 103,820 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 1 3
122 103,570 103,570 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 1 3
124 103,070 103,070 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
125 102,820 102,820 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
126 102,570 102,570 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
128 102,320 102,320 0 3 1 1 1 1 1 1 1 5 3 1 3 1 1 1 1 1 3
131 102,170 102,170 0 2 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
132 101,970 101,970 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 2
133 101,820 101,820 0 4 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
134 101,570 101,570 0 3 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
137 101,510 101,510 0 3 1 1 1 1 1 1 1 4 3 1 1 1 1 1 1 1 3
138 101,470 101,470 0 3 1 1 1 1 1 1 1 5 3 1 2 1 1 1 1 1 2
139 101,170 101,170 0 2 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
140 100,970 100,970 0 3 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 2
142 100,870 100,870 0 1 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 3
143 100,770 100,770 0 3 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 1
145 100,570 100,570 0 2 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 2
148 100,270 100,270 0 1 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 2
151 100,070 100,070 0 1 1 1 1 1 1 1 1 5 3 1 1 1 1 1 1 1 1
154 100,010 100,010 0 1 1 1 1 1 1 1 1 4 3 1 1 1 1 1 1 1 1
156 99,950 99,950 0 1 1 1 1 1 1 1 1 3 3 1 1 1 1 1 1 1 1
158 99,900 99,900 0 1 1 1 1 1 1 1 1 3 2 1 1 1 1 1 1 1 1
159 99,870 99,870 0 1 1 1 1 1 1 1 1 2 2 1 1 1 1 1 1 1 1
161 99,820 99,820 0 11111111 1 2 1 1 1 1 1 1 1 1
169 99,740 99,740 0 11111111 1 1 1 1 1 1 1 1 1 1
APD (%) 0

Table 4.9 – Performance of PFPSO for problem 18f.

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
100 153,320 153,320 0 51112311 5 3 2 4 1 1 2 1 3 3
101 148,520 148,520 0 51111311 5 3 2 4 1 1 2 1 3 3
102 148,470 148,470 0 41111311 5 3 2 4 1 1 2 1 3 3

104
 Table 4.9 – Performance of PFPSO for problem 18f (Continued).

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
103 148,420 148,420 0 3 1 1 1 1 3 1 1 5 3 2 4 1 1 2 1 3 3
104 141,120 141,120 0 5 1 1 1 1 2 1 1 5 3 2 4 1 1 2 1 3 3
105 141,070 141,070 0 4 1 1 1 1 2 1 1 5 3 2 4 1 1 2 1 3 3
106 141,020 141,020 0 3 1 1 1 1 2 1 1 5 3 2 4 1 1 2 1 3 3
108 140,870 140,870 0 5 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 3 3
109 140,820 140,820 0 4 1 1 1 1 2 1 1 5 3 1 4 1 1 1 1 3 3
110 128,270 128,270 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 3 3
111 128,220 128,220 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 3 3
112 128,170 128,170 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 2 1 3 3
114 128,070 128,070 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
115 128,020 128,020 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
116 127,970 127,970 0 3 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 3 3
124 127,870 127,870 0 5 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
125 127,820 127,820 0 4 1 1 1 1 1 1 1 5 3 1 4 1 1 1 1 1 3
126 127,770 127,770 0 31111111 5 3 1 4 1 1 1 1 1 3
APD (%) 0

Table 4.10 – Performance of PFPSO for problem 18g.

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
100 283,320 283,320 0 51112311 5 3 2 4 1 1 2 1 3 3
101 279,820 279,820 0 51111311 5 3 2 4 1 1 2 1 3 3
104 276,320 276,320 0 51111211 5 3 2 4 1 1 2 1 3 3
110 271,270 271,270 0 51111111 5 3 1 4 1 1 2 1 3 3
APD (%) 0

4.4.3. Medium-Scale Benchmark Problem

Kandil and El-Rayes (2006) created medium-scale DTCTP problems by using the
time-cost alternatives of the 18-activity problem of Feng et al. (1997) and by
copying a slightly modified version of the same network ten times in serial. The

105
network, which is used to create medium-scale and large-scale (Section 4.4.4)
problems is demonstrated in Figure 4.5 (Kandil, 2005).

2 10 12 15

4 14

9
Start 6 17 18 Finish

8
1
5 7 11

3 13 16

Figure 4.5 - Network diagram of the core problem for medium-scale and large-scale problems.

The medium-scale problem consists of 180 activities which is practiced with an

indirect cost rate of $0/day. The performance of the proposed PFPSO for 180-
activity problem is compared with the performances of the global parallel and
coarse grained genetic algorithm approaches of Kandil and El-Rayes (2006) in
terms of number of Pareto solutions and computation time requirements. The global
parallel genetic algorithm module was implemented using 50 processors on the
tungsten supercomputing cluster, which was composed of 640 Dell PowerEdge
1750 servers, each with two Intel Xeon 3.2 GHz processors, 1.5 MB of cache
memory, and a total of 3 GB of SDRAM (Kandil and El-Rayes, 2006). The
supercomputing cluster had a peak performance of 6.4 Gflops. As summarized in
Table 4.11, the global parallel GA of Kandil and El-Rayes (2006) obtained 267
Pareto solutions for 180-activity problem running on a single processor (Kandil,
2005) in 14,688 seconds.

106
 Table 4.11 – Comparison of the results for 180-activity problem.

# of CPU # of Pareto
Algorithm
Processor(s) Time (s) Front Solutions
GP-GA (Kandil and El-Rayes, 2006) 1 14,688 267*
GP-GA (Kandil and El-Rayes, 2006) 50 2,556 267*
CG-GA (Kandil and El-Rayes, 2006) 50 684 68*
PFPSO (Section 4.3) 1 21 304
*Kandil (2005)

The computation time requirement of the global parallel GA was reduced to 2,556
seconds by using a cluster of 50 processors. The coarse-grained GA, on the other
hand, located 68 Pareto solutions in 684 seconds over a cluster of 50 processors.
The proposed particle swarm Pareto front optimizer was able to capture 304 Pareto
solutions in 21 seconds on a single processor. PFPSO was able to position a rather
larger set of non-dominated solutions along the efficient frontier than all of the three
GA-based approaches of Kandil and El-Rayes (2006), in a significantly less
computation time.

In addition to the comparative study with the existing parallel GA methods of

Kandil and El-Rayes (2006), the accuracy of PFPSO is also investigated by means
of the optimal costs of Pareto front solutions obtained using mixed-integer
programming technique. The optimal solutions reveal that the proposed PFPSO was
able to capture high quality Pareto front solutions with an average deviation of only
0.1% from the optimal costs. Results of PFPSO for the medium-scale problem is
summarized in Table 4.12. However, for sake of brevity only the first nine solutions
out of a total of 304 showing mode selection of only 17 activities out of 180 are
pointed out in this table.

107
 Table 4.12 – Performance of PFPSO for 180-activity problem.

Pareto Front Mode Selection

Duration Cost ($) Deviation Activities
(day) Optimal PFPSO (%) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 …
1040 1,335,650 1,339,700 0.303 511123 155 3 2 4 1 3 1 3 3
1041 1,330,385 1,334,700 0.324 511113 155 3 2 4 1 3 1 3 3
1042 1,325,120 1,329,700 0.346 511113 155 3 2 4 1 3 1 3 3
1043 1,319,855 1,324,700 0.367 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
1044 1,314,590 1,319,700 0.389 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
1045 1,309,325 1,314,700 0.411 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
1046 1,304,060 1,309,100 0.386 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
1047 1,298,795 1,304,100 0.408 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
1048 1,293,530 1,299,100 0.431 511113 155 3 2 4 1 3 1 3 3
⁝
APD (%) 0.095

4.4.4. Large-Scale Benchmark Problems

Experiments on large-scale problems comprise the 360-activity and 720-activity

benchmark problems created by Kandil and El-Rayes (2006). Similar to their
medium-scale problem, the time-cost alternatives of the 18-activity problem of
Feng et al. (1997) were duplicated for networks of size 360 and 720 which were
generated by copying the same network shown in Figure 4.5 several times in serial.

Both of the large-scale problems are practiced with an indirect cost rate of $0/day.
The performances of the proposed PFPSO for 360-activity and 720-activity
problems are compared with the performances of the global parallel and coarse
grained genetic algorithms of Kandil and El-Rayes (2006) with regard to number
of Pareto solutions and computational time. The global parallel genetic algorithm
module was implemented using a supercomputing cluster which is described in
Section 4.4.3. Performance of the results for the large-scale 360-activity instance is
presented in Table 4.13.

108
 Table 4.13 – Comparison of the results for 360-activity problem.

# of CPU # of Pareto
Algorithm
Processor(s) Time (s) Front Solutions
GP-GA (Kandil and El-Rayes, 2006) 1 75,096 232
GP-GA (Kandil and El-Rayes, 2006) 50 10,404 232
CG-GA (Kandil and El-Rayes, 2006) 50 1,836 94
PFPSO (Section 4.3) 1 43 536

Over a cluster of 50 processors, the global parallel GA was able to obtain 232 non-
dominated solutions in 10,404 seconds, whereas the coarse-grained GA achieved
94 non-dominated solutions in 1,836 seconds for 360-activity problem. The
proposed particle swarm Pareto front optimization method was able to capture 536
non-dominated solutions in 43 seconds for the same problem by running on a single
processor. The exceptional performance of PFPSO was consistent for larger
problems as it captured larger set of non-dominated solutions than all of the three
GA-based approaches of Kandil and El-Rayes (2006), within a considerably less
computational effort. Results of PFPSO for the 360-activity problem is summarized
in Table 4.14. Though, in the consideration of briefness only the first nine solutions
out of a total of 536 showing mode selection of only 17 activities out of 360 are
tabulated as follows.

Table 4.14 – Non-dominated solutions of PFPSO for 360-activity problem.

Pareto Front Mode Selection

Duration Cost Activities
(day) ($) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 …
2080 2,679,345 5 1 1 1 2 3 1 5 5 3 2 4 1 3 1 3 3
2081 2,674,345 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
2082 2,668,145 5 1 1 1 1 3 1 5 5 3 2 4 1 2 1 3 3
2083 2,663,195 5 1 1 1 1 3 1 5 5 3 2 4 1 2 1 3 3
2084 2,658,145 5 1 1 1 1 3 1 5 5 3 2 4 1 2 1 3 3
2085 2,653,145 5 1 1 1 1 3 1 5 5 3 2 4 1 2 1 3 3
2086 2,648,145 5 1 1 1 1 3 1 5 5 3 2 4 1 2 1 3 3
2087 2,643,145 5 1 1 1 1 3 1 5 5 3 2 4 1 2 1 3 3
2088 2,638,195 5 1 1 1 1 3 1 5 5 3 2 4 1 2 1 3 3
⁝

109
The successful performance of the PFPSO is also consistent for the large-scale 720-
activity benchmark problem, as shown in Table 4.15.

Table 4.15 – Comparison of the results for 360-activity problem.

# of CPU # of Pareto
Algorithm
Processor(s) Time (s) Front Solutions
GP-GA (Kandil and El-Rayes, 2006) 1 491,400 303
GP-GA (Kandil and El-Rayes, 2006) 50 55,296 303
CG-GA (Kandil and El-Rayes, 2006) 50 7,092 132
PFPSO (Section 4.3) 1 92 1022

PFPSO was able to capture 1022 non-dominated solutions in 92 seconds for the
720-activity problem with a single processor on a desktop computer. Whereas, over
a cluster of 50 processors, the global parallel GA provided 303 Pareto solutions in
55,296 seconds, and the coarse-grained GA obtained 132 non-dominated solutions
in 7,091 seconds for 720-activity problem. Results of PFPSO for the 720-activity
problem is summarized in Table 4.16. However, due to length considerations only
the first nine solutions out of a total of 1022 showing mode selection of only 17
activities out of 720 are illustrated as follows.

Table 4.16 – Non-dominated solutions of PFPSO for 720-activity problem.

Pareto Front Mode Selection

Duration Cost Activities
(day) ($) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 …
4160 5,358,800 5 1 1 1 2 3 1 5 5 3 2 4 1 3 1 3 3
4161 5,353,800 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
4162 5,348,800 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
4163 5,343,650 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
4164 5,338,800 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
4165 5,333,800 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
4166 5,328,750 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
4167 5,323,800 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
4168 5,318,800 5 1 1 1 1 3 1 5 5 3 2 4 1 3 1 3 3
⁝

110
Obtained by PFPSO, 304, 536, and 1022 non-dominated solutions constituting the
Pareto fronts of 180-activity, 360-activity, and 720-activity problems are illustrated
in Figure 4.6.

180-activity problem 360-activity problem 720-activity problem

5,700
5,200
4,700
Total Cost ($1,000)

4,200
3,700
3,200
2,700
2,200
1,700
1,200
700
800 1,800 2,800 3,800 4,800 5,800 6,800

Duration (days)

Figure 4.6 – Pareto fronts located by PFPSO for 180, 360, and 720-activity problems.

The mixed integer model mentioned in Section 3.4.2 was also used to determine the
optimal costs for both 360-activity and 720-activity problems; nevertheless, optimal
costs of these problems were not obtained within 72 hours. As is clear, not only the
computation time requirement of the proposed Pareto front particle swarm
optimizer was significantly less than earlier approaches but was also able to produce
a large number of good feasible Pareto front solutions for the large-scale problems.
Due to the effectiveness and efficiency of the proposed particle swarm optimization
method, it is expected to contribute to optimal planning of real-life-scale
construction projects. To the best of author’s knowledge, the proposed PFPSO is
one of the first methods capable of capturing high quality Pareto solutions for the
large-scale DTCTP within seconds.

111
A more comprehensive study on the performance of PFPSO is given in Section
5.2.5 using new sets of RanGen2 instances discussed in Section 5.2.1. Based on
performance indices demonstrated in Section 5.2.2, effectiveness and efficiency of
this approach is measured and compared with a new heuristic algorithm which is
presented in Section 5.1.4.

112
 CHAPTER 5

COST-SLOPE HEURISTIC METHOD FOR DTCTP

Respecting the state of the existing research on TCTP in the construction industry,
it can be observed that the meta-heuristic approaches are the most prominent
methods used in a multitude of studies. Nonetheless, it is broadly acknowledged
that evolutionary algorithms are very sensitive to configuration of their parameters.
In real-life situations, the experimental process for configuration of parameters for
best values may become a tedious and arduous task. Yet, the parameters need to be
retuned for each new problem at hand which might reduce the practicability of the
meta-heuristic approaches. On the other hand, experimentation of DPSO (Section
3.4) and PFPSO (Section 4.4) revealed that the exceptional performances of the
noted approaches were largely resulting from their heuristic modules. It was also
discovered that there exist only a handful of heuristic exemplars in the existing
literature which is limited to the methods proposed by Fondahl (1961), Siemens
(1971), Goyal (1975), Moselhi (1993), and Bettemir and Birgonul (2017).
However, none of the earlier heuristics mentioned have the capacity to tackle real-
life-scale Pareto front discrete TCT problems since most of them are designed for
rather simple and continuous deadline problems. The discrete Pareto front TCTP is
considered as the most salient type this trade-off problem due to its practical
relevance and also because of its potential for articulation of managers’ propensities
which provides them with tools for selection of the best solution with respect to
their preferences. These are some of the very reasons that inspired development of
a new heuristic approach, hereafter called Cost-Slope Heuristic, for solution of
TCTP in this thesis which is discussed in Section 5.1.

113
Moreover, apart from the fact that the existing approaches have seldom been
applied for solution of large-scale DTCTPs, it is interpreted that the dearth of real-
life-scale problems could possibly be another major reason for the lack of studies
on realistic problems. Despite the fact that some studies have included problems
including up to 720, 2000, and 6300 activities, all of the employed large-scale
problems are generated using small-scale base networks and are generated by
copying the core problem in serial several times; hence, these problems are believed
to have limitations in reflecting the complexity of the real-life construction projects.
To the respect of this, for better evaluation of the capabilities of the proposed
optimization models, new sets of multi-mode large-scale DTCT problems have
been generated in this thesis by means of random network generator, viz., RanGen2.
The systematically generated large-scale instances comprise complex networks and
realistic sets of time-cost alternatives which are elucidated in Section 5.2.1.

On the other hand, it is also observed that the majority of the earlier research on
DTCTP not only employ instances with small problem networks, but also the
detailed performance evaluation on accuracy and efficiency of most of the
presented approaches appear to be lacking, especially for the Pareto front problem.
Unlike a large body of the existing literature, performance measurements are
carried out by taking a more holistic approach that involves a set of performance
comparison indices. In order to compare methods rigorously and to measure
performances on a quantitative bases, efficiency, accuracy, diversity, and
cardinality of the obtained solutions are evaluated using the indices explained in
Section 5.2.2.

Meanwhile, it is not possible to accurately assess quality of the solutions obtained

from heuristic or meta-heuristic algorithms short of exact procedures since they are
the only approaches guaranteeing optimality of the solutions. Owing to this very
reason, they play a crucial role in performance evaluation of non-exact optimization
algorithms. Therefore, a Mixed-Integer Linear programming technique with unique

114
features is also proposed for obtaining the optimal results of the practiced instances
which is described in Section 5.2.3.

The computational experiments include comparative studies on the performance of

the Cost-Slope Heuristic (Section 5.1), DPSO (Section 3.3) and PFPSO (Section
4.3). Benchmark instances attained from the literature as well the new sets of
RanGen2 instances presented in Section 5.2.1 are used to conduct the comparative
studies which prove remarkable efficiency and exceptional efficacy of the proposed
Cost-Slope Heuristic for real-life practices.

5.1. Cost-Slope Heuristic

A heuristic method is defined as a logical sequence of steps – relying on rules of

thumb, execution of which yields an optimum or near-optimum solution
(Demeulemeester and Herroelen, 2002). The literature on heuristic algorithms for
the TCTP is virtually restricted to Fondahl (1961), Siemens (1971), Goyal (1975),
and Moselhi (1993). These methods assume linear time-cost relationships and are
incapable of handling discrete TCT problems. The most recent heuristic approach
is proposed by Bettemir and Birgonul (2017) which is one of the very few methods
designed for unraveling the discrete TCT problems. Their method which is called
Network Analysis Algorithm (NAA) can tackle deadline and cost minimization
DTCT problems. To the best of author’s knowledge, there exists no alternative
multi-objective heuristic method designed for Pareto oriented optimization of
DTCTP other than the SAM-based approaches proposed by the author previously.
Two concepts are introduced herein for reducing the size of the network and for
faster CPM calculations which are described in detail in Section 5.1.1.1 and Section
5.1.1.2, respectively. The details of the methodology implemented to develop the
proposed Cost-Slope Heuristic, abridged as CS-Heuristic, for deadline and Pareto
front TCTP are explained in Section 5.1.3 and Section 5.1.4, respectively.

115
5.1.1. Network Reduction Techniques

Just as an increase in the scale of the problem causes an exponential growth in

computational burden, a decrease in scale results in an exponential decline in
computational burden. The simplification of the problem which is known as the
network reduction technique, despite maintaining the abovementioned utility, has
not received sufficient attention. Based on the methods proposed by Rothfarb,
Frank, Rosebaum, Steiglitz, and Kleitman (1970) and Frank, Frisch, Van Slyke, and
Chou (1971) two clever network reduction techniques are proposed in this chapter
which will be explained in the ensuing sections. Unlike preceding work on this area
which considered AoA networks, this thesis study proposes simplification methods
for AoN notation systems. According to the proposed method, serial/parallel
reducible problems can be simplified through replacing certain activities by
merging them into an equivalent activity. In order to elucidate these concepts, a
numerical example is presented based on the sample case problem introduced in
Figure 3.1 of Section 3.3.1. This instance is exercised by assuming a completion
deadline of 45 days, with delay penalty and indirect cost amounts of $2,000 and
$1,000/day, respectively.

5.1.1.1. Serial Merging Technique

In a network if there exist two activities j 1, j 2  Acts such that j 1 ’s unique

successor is j 2 and j 2 ’s unique predecessor is j 1 , this network is said to be serial
reducible through which activities j 1 and j 2 can be merged into an equivalent

activity j 1  . The replaced activity j 1  will adopt the predecessors and successors
from activities j 1 and j 2 , respectively. Each time-cost alternative of j 1 ,
k j1  [1, m( j 1)] , is then combined with each time-cost component of j2 ,

k j 2  [1, m( j 2 )] , forming the maximum m( j 1 )  m( j 1)  m( j 2 ) number of candidate

time-cost alternatives for the equivalent activity j 1  . Each of the time-cost

116
components of activity j 1  , k j1  [1, m( j 1 )] , is achieved by summation of the

times, d k j   d k j1  d k j 2 , and costs, dc k j   dc k j1  dc k j 2 , of the corresponding

1 1

time-cost alternatives. In case of ties two similar but not identical approaches are
implemented for the CS-Heuristic described in Section 5.1.3 and Section 5.1.4 and
for the Mixed-Integer Linear Programming presented in Section 5.2.3. For CS-
Heuristic, in case of a tie, all but one of the duplicate alternatives with
d k j   d (k 1) j  and dc k j   dc (k 1) j  are eliminated. A dominance rule is also
1 1 1 1

implemented to eliminate the dominated components as k j1  (k  1) j1 , which

indicates if d k j   d (k 1) j  while dc k j   dc (k 1) j  and one of these inequalities

1 1 1 1

holds strictly, (k  1) th alternative of the equivalent activity j 1  will be discarded.

Whereas, for the Mixed-Integer Linear Programming in case of a tie, all but one of
the duplicate alternatives with d k j   d (k 1) j  and dc k j   dc (k 1) j  are eliminated.
1 1 1 1

A dominance rule is also implemented to eliminate the dominated components as

k j1  (k  1) j1 , which indicates if d k j   d (k 1) j  while dc k j   dc (k 1) j  , (k  1) th
1 1 1 1

alternative of the equivalent activity j 1  will be eliminated. The flowchart of the

proposed serial merging technique is illustrated graphically in Figure 5.6 of Section
5.1.3.

The procedure is best elucidated by a case example which is presented in Section

3.3.1. The original network shown in Figure 3.1 is serial reducible since Activity-1
and Activity-3 satisfy the above conditions. Activity-1 has only Activity-3 as its
succeeding activity, and Activity-1 is the unique predecessor of Activity-3. It must
be also pointed out that in this example the numerator of those activities with ID
numbers greater than three are decreased by one. As shown in Figure 5.1, the
equivalent activity, Activity-1, fetches the predecessors of the original Activity-1
(i.e., Start) and successors of the original Activity-3 (i.e., Activity-6 which is
represented as Activity-5 in the reduced network).

117
 1

1. (28, $73,600)
2. (30, $69,000)
3. (32, $66,600)
4. (34, $62,000)
5. (40, $54,600)
6. (42, $50,000) 3 5
7. (54, $47,000)
Start 1. (5, $52,000) 1. (3, $36,000) Finish
2. (8, $34,000) 2. (5, $24,000)
2

1. (7, $38,000)
2. (10, $26,000)
3. (12, $20,000) 4 ID

1. (10, $40,000) 1. (Dur., Cost)

2. (15, $24,000) m. (Dur., Cost)
3. (16, $22,000)

Figure 5.1 – Serial merge applied to the Case Example.

In the absence of the mode elimination rule mentioned earlier, the equivalent
Activity-1 would have a maximum of 3  3  9 alternatives. However, combining
the second mode of the original Activity-1 with the third mode of the original
Activity-3 yields the same duration of 42 days achieved by the combination of the
third mode of the original Activity-1 with the first mode of the original Activity-3.
Since the cost of the first combination is added as $50,000 and the cost of the second
combination is summed as $66,000, the second combination is discarded. In
addition, summation of the third mode of the original Activity-1 with the second
mode of the original Activity-3 yields a component with a duration of 46 days and
direct cost of $59,000 which is dominated by the sixth mode of the equivalent
Activity-1 with less duration and direct cost amounts. The solution space of the
original network comprises 324 realizations; however, the solution space of the
reduced network includes 252 different realizations. As is clear, the smaller solution
space of the reduced network contributes to significantly faster computations.

118
5.1.1.2. Parallel Merging Technique

In a network if there exist two activities j 1, j 2  Acts such that both j 1 and j 2
share a unique predecessor while sharing exactly the same set of successors, this
network is said to be parallel reducible through which activities j 1 and j 2 can be

merged into an equivalent activity j 1  . The replaced activity j 1  will adopt the
single predecessor and the set of successors of the merged activities j 1 and j 2 .
Each time-cost alternative of j 1 , k j1  [1, m( j 1)] , is then combined with each time-

cost component of j2 , k j 2  [1, m( j 2 )] , forming the maximum

m( j 1 )  m( j 1)  m( j 2 )  1 number of candidate time-cost alternatives for the

equivalent activity j 1  .Each of the time-cost components of activity j1  ,

k j1  [1, m( j 1 )] , is achieved by using an iterative scheme which initially starts by

combining the all-normal modes of the j 1 and j 2 activities, i.e., k j1  m( j 1) and

k j 2  m( j 2 ) . In each iteration of the pairwise combination, the larger duration is set

as the duration of the new mode, d k j   max{d k j1 , d k j 2 } , while the cost of the new
1

mode is calculated by adding the direct costs of the time-cost alternatives,

dc k j   dc k j1  dc k j 2 . Thereafter, pairwise combination is carried out by
1

incrementing the index of the driving time-cost component, k j  {k j1, k j 2} , which

had the greater amount of duration in the previous iteration, {k j | d k j   d k j } .

1

Thereby, the same procedure is repeated until the index of both the activities are
promoted to their maximum amounts, viz., until all-crashed time-cost modes of the
merged activities are experimented for combination, i.e., k j1  1 and k j 2  1 .

Unlike serial merge, no elimination or dominance rule is incorporated since no ties

occur in the above described procedure. The flowchart of the proposed parallel
merging technique is illustrated graphically in Figure 5.5 of Section 5.1.3.

119
The procedure is best illustrated using a case example which is presented in Section
3.3.1. The original network shown in Figure 3.1 is parallel reducible since Activity-
4 and Activity-5 satisfy the above conditions. Both the Activity-4 and Activity-5
has only Activity-2 as their preceding activity, and Activity-4 and Activity-5 only
have Activity-6 in their set of successors which is common to both the activities 4
and 5. It must be also pointed out that in this example the numerator of those
activities with ID numbers greater than five are decreased by one. As shown in
Figure 5.2, the equivalent activity, Activity-4, fetches the unique predecessor of the
original Activity-4/Activity-5 (i.e., Activity-2) as well as the successors of the
original Activity-4/Activity-5 (i.e., Activity-6 which is represented as Activity-5 in
the reduced network).

1 3

1. (12, $21,600) 1. (16, $52,000)

2. (14, $17,000) 2. (20, $45,000)
3. (26, $14,000) 3. (28, $33,000) 5

Start 1. (3, $36,000) Finish

2. (5, $24,000)
2 4

1. (7, $38,000) 1. (10, $74,000)

2. (10, $26,000) 2. (15, $58,000)
3. (12, $20,000) 3. (16, $56,000) ID

1. (Dur., Cost)
m. (Dur., Cost)

Figure 5.2 – Parallel merge applied to the Case Example.

The equivalent Activity-4 would have a maximum of 2  3  1  4 alternatives.

However, both the durations for both of the time-cost alternatives of the original
Activity-4 (i.e., 5 and 8) are less than the duration of the all-crashed mode of the
original Activity-5 (i.e., 10). Resultantly, the equivalent Activity-4 includes three
time-cost modes with durations taken from the original Activity-5, while the costs
are calculated by the summation of costs of the original Activity-5 with the least-
cost mode of the original Activity-4 which is the second mode in this case. As

120
mentioned earlier, the solution space of the original network comprises 324
realizations; however, the solution space of the reduced network includes 162
different realizations. As is clear, the considerably smaller solution space of the
reduced network contributes to significantly faster computations.

5.1.2. Partial-CPM Calculator

The computational bottleneck in solution of real-life-scale TCT problems is the

fitness evaluation of the generated solutions due to the iterative and repetitive
computation of the network using the critical path method. Computational
complexities arise from the necessity to calculate the longest path of the network
since any variation in selection of the alternatives modifies the project schedule.
These incessant modifications require rescheduling the project for potential changes
in its total cost and total duration amounts. It is obvious that any rescheduling
process demands reanalyzing the network using the critical path method. Repetitive
classical CPM calculations even with the modern computers is, therefore, not a
convenient method for solution of DTCTP. In order to reduce the computational
burden of the repeated CPM calculations, new approaches are implemented in the
proposed CS-Heuristic. In the conventional CPM calculation module of the
proposed method, where applicable, instead of applying forward and backward pass
calculations, only the forward pass is applied and the total duration of the project is
determined using the early finish date of the last activity on the project network. By
means of this procedure, computation time is reduced for conditions where it is not
necessary to carry out the complete CPM calculations which include computation
of the late dates and floats. Original to this thesis, the proposed CS-Heuristic is also
complemented with a unique CPM-esque approach to accelerate the solution
process. A significant contribution of this thesis study is development of this unique
network analyzer which is called the Partial-CPM calculator. Networks of the
problems are calculated using this robust and efficient method which improves the
overall computation time substantially.

121
Primarily, the distinction between a modified activity and an updated activity must
be made as the first term refers to an activity with a modified setting of its time-cost
alternatives; whereas, the second term refers to an activity with updated early/late
start/finish dates. Rather than analyzing the whole network for changes, the
proposed Partial-CPM technique examines only a small portion of the network as
an alternative execution mode gets selected for an activity. In forward pass, this
technique only updates early dates of the successors to the modified activity which
have early-start dates smaller than early-finish date of the modified activity. In a
similar fashion, the same adjustments are also applied to successors of the updated
activities. In backward pass, Partial-CPM only updates late dates of the
predecessors of the modified activity which have late-finish dates greater than late-
start date of the modified activity. The same adjustments are also applied to
predecessors of the updated activities. Following the forward pass and the backward
pass calculations, total floats of all the updated activities are recalculated. To the
best of author’s knowledge, no such study exists in the literature. The main
advantage of using the proposed Partial-CPM to DTCTP is that the CPM
calculation of the entire network needs only be completed once, thereby the
influence of modifications is analyzed by only experimenting a small fraction of
the network.

Partial-CPM calculation procedure is described using a case example which is

presented in Section 3.3.1. Figure 5.3 demonstrates the PERT chart for the all-
crashed schedule. By uncrashing Activity-2 to its next option, its duration increases
by three days (effective DDiff amount) to 10 days.

122
 13 16 28

1 12 12 13 0 28

1 0 12 8 5 12 29 3 31

Start 4 6 Finish

1 7 7 24 16 28 29 0 31

12 11 18 8 10 17 ES Dur. EF

5 ID

19 11 28 LS TF LF

Figure 5.3 – PERT chart of all-crashed schedule for the Case Example.

Resultantly, its early-finish date increases from 7 to 10. According to Partial-CPM,

in the first forward pass iteration, early dates of successors to Activity-2, i.e.,
Activity-4 and Activity-5 need to be updated if their early-start dates are smaller
than 10. The current early-start dates of both Activity-4 and Activity-5 are shown
to be 8, which satisfies the updating condition in forward pass of the Partial-CPM.
Thus, this date is updated to 10  1  11 and the early-finish date for each of these
activities are updated by adding up their duration amounts to their early-start dates
(Figure 5.4). Hence, early-finish dates of Activity-4 and Activity-5 are updated as
15 and 20, respectively. In the next iteration of forward pass, the successors to the
updated activities, i.e., Activity-6 is evaluated. Since its early-start date of 29 is
already greater than early-finish dates of both Activtiy-4 and Activity-5, and there
exist no other successors to the updated activities, forward pass is terminated. In the
first backward pass cycle, current late-finish date of the predecessors of the updated
activities, i.e., Acitvity-2, is checked against late-start dates of both the updated
activities Activity-4 and Activity-5. Backward pass is then terminated since late-
finish date of 18 for Activity-2 is not greater than late-start dates of 24 and 19 for
Activity-4 and Activtiy-5, respectively. Following the forward pass and the

123
backward pass calculations, total floats of all the updated activities are recalculated.
It is shown that for this case example, Partial-CPM is capable of rescheduling the
project by analyzing and updating only the dates/floats which are underscored in
Figure 5.4. It is shown that the proposed technique, rather than rescheduling the
entire network, is able to obtain the same rescheduled network by updating only 8
components (early dates, late dates, and floats) out of a total of 30.

13 16 28

1 12 12 13 0 28

1 0 12 11 5 15 29 3 31

Start 4 6 Finish

1 10 10 24 13 28 29 0 31

9 8 18 11 10 20 ES Dur. EF

5 ID

19 8 28 LS TF LF

Figure 5.4 – PERT chart of the updated schedule for the Case Example.

5.1.3. CS-Heuristic for Deadline DTCTP

A new Cost-Slope Heuristic is presented herein for cost minimization and deadline
discrete TCT problem. This method incorporates the project deadline, since, in
practice there is a completion deadline stipulated in the contract for the majority of
the projects. In realistic projects, delay penalties which are usually in the form of
liquidated damages are applied in case the project duration exceeds the
predetermined deadline and the incentives are the bonus payments made for each
day saved from the specified deadline. Generally, there is a negative correlation

124
between the deadline and the complexity of the problem. That is, projects with
larger predetermined thresholds are simpler than the ones with shorter deadlines.
Regarding the above fact, the completion deadline of the practiced problems is
calculated in the proposed heuristic algorithm as follows. Firstly, the all-normal
schedule with the largest critical path length, CPM max , is calculated. Secondly, the
all-crashed schedule with the shortest critical path, CPM min , is computed. Finally,
the completion deadline is set to be equal to the average of the earliest allowable
completion time of the project and the latest possible finishing time.

The first step of the CS-Heuristic involves serial/parallel merging of the network
activities. Based on the techniques presented in Section 5.1.1.1 and Section 5.1.1.2
CS-Heuristic reduces the network of the problem to a simpler equivalent network.
Thereafter it calculates the completion deadline based on the above described
procedure. In the next step, CS-Heuristic calculates and records the cost-slopes (
CS ), and crash amounts (duration difference DDiff , and cost difference CDiff )
for every time-cost option of all the activities. The cost-slopes are calculated
according to Eq. (3.1) given in Section 3.2. All these options are then indexed as
follows. Time-cost alternatives are sorted in descending order by CDiff values then
by CS values in descending order; thereby they are treated with an index named
UFF  index . Similar to the previous stage, time-cost alternatives are sorted in
descending order by CS values then by DDiff values in ascending order, thereby
an index called UCS  index is assigned to each alternative. According to
UFF  index , a component providing a greater cost saving is uncrashed first; in
case of a tie, the component with a greater cost-slope rate is selected; if the tie is
still not broken, the first component is selected. According to UCS  index , a
component with a greater cost-slope rate is uncrashed first; in case of a tie, the
component providing a smaller duration reduction is selected; if the tie is still not
broken, the first component is selected.

125
Unlike earlier heuristic approaches, in lieu of starting from the all-normal schedule,
CPM max , CS-Heuristic initiates from the all-crashed schedule, CPM min , by
selecting the shortest/costliest time-cost alternative for the activities. Contrary to
previous heuristics, in every cycle of CS-Heuristic, each new schedule is built upon
an already non-dominated solution. Duration of this schedule is determined using
CPM technique and the total cost is calculated by explicitly including the incentive
and disincentive payments in the project cost formulation. Since CS-Heuristic starts
from the all-crashed schedule, it is designed to optimize the solutions by performing
an uncrashing scheme. The uncrashing scheme incorporates two different phases,
namely, uncrash free-float ( UFF ) and uncrash cost-slope ( UCS ). Through UFF
phase those activities with total floats greater than or equal to their effective
(immediate available) crash amounts, DDiff jk  TF j , are determined. The

determined activities are then uncrashed to their next available option, by

uncrashing one activity at a time, with respect to their UFF  index values. After
performing each UFF , CS-Heuristic uses Partial-CPM technique (Section 5.1.2) to
update the modified activities. The UFF process is continued until either all the
activities get fully uncrashed or their total floats become less than their effective
crash amounts. It is after this step that the heuristic stores the current schedule as a
solution. While each solution, Sol , obtained from the uncrashing phases are
recorded in an archive called Solutions , an external repository, Result , is
implemented to record the least-cost solution. Accordingly, the least-cost Sol of
Solutions is stored in Result , subsequently, the schedules stored in Solutions are
wiped out at the end of each iteration.

After the first schedule is generated and stored in Result , CS-Heuristic performs
UCS process to generate new schedules unless the status of the unique solution
stored in Result is “Closed”. Status of each solution which gets selected as a base
for generation of new schedules, is switched to “Closed” at the end of each UCS
phase, making it unavailable for re-selection. New schedules are generated based
on the solutions that iteratively replace the single schedule recorded in Result , as

126
long as they are not “Closed”. Through UCS phase those activities that are
uncrashable, i.e., not fully uncrashed, are determined. The determined activities are
then uncrashed to their next available option, by uncrashing one activity at a time,
with respect to their UCS  index values. After performing each UCS , CS-
Heuristic implements the short procedure described in Section 5.1.2 to calculate the
completion time of the obtained schedule. In case there exists no Sol stored in
Solutions with the same duration, CPM calculation is carried out which is then
followed by the UFF process as mentioned earlier. It is after this step that the
heuristic stores the current schedule as a solution. However, through UCS phase,
if the same duration already exists in Solutions within the same cycle, that schedule
is discarded without being stored as a Sol in the archive. The UCS process is
continued until all the determined uncrashable activities are uncrashed to their
longer durations. After completing each UCS phase, the status of solution,
Sol  Result , which was selected at the start of the cycle is switched to “Closed”.
UCS and UFF phases are iteratively applied to the identified uncrashable
activities,  Sol  (Solutions or Result ) : status  Closed , until obtaining a solution,
all of the activities of which are uncrashed to their longest durations. CS-Heuristic
terminates by returning the ultimate Sol stored in Result with “Closed” status. The
flowcharts of parallel merge, serial merge, and uncrash free-float ( UFF ) modules
of the proposed CS-Heuristic are presented in Figure 5.5, Figure 5.6, and Figure
5.7, respectively. The pseudo-code of the proposed CS-Heuristic for cost
minimization and deadline DTCTP is illustrated in Figure 5.8.

127
 Parallel Merge

No Merging Active Yes
and j ≤ S

Act. j has Two Yes
Succs.: Sc1j, Sc2j
No
Succs. and
Preds. of  Sc1j and
No Sc2j  are Equal
Increment j
Yes

Merge Sc1j and Sc2j

Return

Figure 5.5 – Flowchart of the proposed parallel merging technique.

Serial Merge

No Merging Active Yes
and j ≤ S

Act. j has Yes
One Succ.: Sc1j
No

Sc1j has One Pred.
No
Increment j
Yes

Merge Sc1j and Sc2j

Return

Figure 5.6 – Flowchart of the proposed serial merging technique.

128
 Uncrash Free-Float

No Yes
j≤S

Yes
k≤m
No
Increment j
DDiff ≤ Total-Float
No

Increment k Yes

Perform Partial- Uncrash based on

Calculate Cost
CPM UFF Index

Return

Figure 5.7 – Flowchart of the uncrash free-float module of the proposed CS-Heuristic.

Begin;
For ∀ File in Directory
For ∀ j ∈ [1, S]
For ∀ k ∈ [1, m]
Retrieve Values;
End;
End;
For ∀ j ∈ Acts
Calculate Predecessors;
End;
Sort Acts;
If Perform Parallel Merge = True
For ∀ j ∈ Acts
If j has 2 Successors
If Successors of Successors to j ∧ Predecessors of Successors to j are same
Merge Successors to j;
End;
End;
End;
End;
If Perform Serial Merge = True
For ∀ j ∈ Acts

Figure 5.8 – Pseudo-code of the proposed CS-Heuristic for deadline DTCTP.

129
 If j has 1 Successor
If Successor to j has 1 Predecessor
Merge j and its Successor;
End;
End;
End;
End;
Calculate Project Deadline;
For ∀ j ∈ Acts
For ∀ k ∈ Modes
Calculate CS, DDiff, CDiff;
Sort and Index by CDiff then by CS for Uncrash Free-Float;
Sort and Index by CS then by DDiff for Uncrash Cost-Slope;
End;
Select 1st Mode;
End;
Calculate CPM;
Calculate Dur/Cost;
While ∃ j ∈ Acts: DDiff ≤ Total-Float
Add to Uncrashables;
Perform Uncrash Free-Float;
Perform Partial-CPM;
Calculate Cost;
Break;
Store Result;
While Result ≠ Closed
For ∀ Act ∈ Result: Fully-Uncrashed ≠ True
Add to Uncrashables;
End;
Perform Uncrash Cost-Slope;
Perform Partial-CPM;
Calculate Dur;
If ∃ Sol ∈ Solutions: Sol.Dur = Dur
;
End;
Else
Calculate CPM;
Calculate Cost;
While ∃ j ∈ Acts: DDiff ≤ Total-Float
Add to Uncrashables;
Perform Uncrash Free-Float;
Perform Partial-CPM;
Calculate Cost;
Break;
Store Sol in Solutions;
End;
Close Result;

Figure 5.8 – Pseudo-code of the proposed CS-Heuristic for deadline DTCTP (Continued).

130
 If ∃ Sol ∈ Solutions: Sol.Cost ≤ Result.Cost
Result = Sol;
End;
Break;
Return Result;
End;
End;

Figure 5.8 – Pseudo-code of the proposed CS-Heuristic for deadline DTCTP (Continued).

The proposed CS-Heuristic is described by a case example which is presented in

Section 3.3.1. The original network shown in Figure 3.1 is used to elucidate UFF
and UCS schemes of the CS-Heuristic. For the all-crashed schedule, the project
duration is 31 days and the total cost is $270,600 as shown in Table 5.1.

Table 5.1 – Candidate solutions found by CS-Heuristic for deadline DTCTP.

# of Continue Crash Dur. Direct Indirect Total Cost

Activity Uncrash
Schedule from Level (day) Cost ($) Cost ($) ($)
1 0 Crashed 31 239,600 31,000 270,600
1.i 1 4 0 M1 to M2 31 221,600 31,000 252,600
1.ii 1.i 5 0 M1 to M2 31 205,600 31,000 236,600
1.iii 1.ii 2 0 M1 to M2 31 193,600 31,000 224,600
1.iv 1.iii 2 0 M2 to M3 31 187,600 31,000 218,600
1.v 1.iv 5 0 M2 to M3 31 185,600 31,000 216,600
2 1.v 6 1 M1 to M2 33 173,600 33,000 206,600
- 1.v 1 1 M1 to M2 33 - - -
3 1.v 3 1 M1 to M2 35 178,600 35,000 213,600
4 2 1 2 M1 to M2 35 169,000 35,000 204,000
5 3 3 2 M1 to M2 37 166,600 37,000 203,600
6 4 3 3 M1 to M2 39 162,000 39,000 201,000
7 4 1 3 M2 to M3 47 166,000 47,000 213,000
8 5 1 4 M1 to M2 39 162,000 39,000 201,000
9 5 3 4 M2 to M3 45 154,600 45,000 199,600
10 6 3 5 M2 to M3 47 150,000 47,000 201,000*
11 6 1 5 M2 to M3 51 159,000 51,000 210,000
12 9 1 6 M1 to M2 47 150,000 47,000 201,000*
*Exceeds deadline by 2 days

131
In the Schedule-1 which consists of crashed modes, activities 2, 4, and 5 are non-
critical activities. According to UFF , among the non-critical activities, Activity-4
with a CDiff of $18,000 provides the greatest cost saving as shown in Table 5.2,
and is uncrashed first, to its second mode (M2).

Table 5.2 – Cost-slopes, DDiff s and CDiff s of crash modes.

Activity Crash Mode Cost-slope ($/Day) DDiff (day) CDiff ($)

1 M1 to M2 2,300 2 4,600
1 M2 to M3 250 12 3,000
2 M1 to M2 4,000 3 12,000
2 M2 to M3 3,000 2 6,000
3 M1 to M2 1,750 4 7,000
3 M2 to M3 1,500 8 12,000
4 M1 to M2 6,000 3 18,000
5 M1 to M2 3,200 5 16,000
5 M2 to M3 2,000 1 2,000
6 M1 to M2 6,000 2 12,000

The same procedure is applied to Activity-5 which is uncrashed to its second mode
(M2), similarly, Activity-2 is uncrashed to its second mode (M2), as well. In
schedules Schedule-1.iv and Schedule-1.v, Activity-2 and Activity-5 are uncrashed
to their third option (M3), respectively. The resulting Schedule-1.v has a duration
of 31 days and a total cost of $216,600 (Table 5.1) which is recorded in Result as
the first UFF cycle terminates. Following UFF phase, five uncrashing options
could be identified for UCS including Activity-1: M1 to M2, Activity-1: M2 to
M3, Activity-3: M1 to M2, Activity-3: M2 to M3, and Activity-6: M1 to M2.
However, M2 to M3 crash options of Activity-1 and Activity-3 are excluded from
the uncrashables, since, only the effective (immediate available) crash options are
considered throughout UCS phase. Therefore, succeeding UFF phase, Activity-6
with largest cost-slope of $6,000/day is uncrashed through the UCS process and
stored as Schedule-2 in Solutions . The next least-cost-slope activity is Activity-1,
however, uncrashing it to its second mode (M2) results in a solution with a duration

132
of 33 days which already exists as Schedule-2 in Solutions ; hence, this schedule is
discarded. The final uncrashable activity in this cycle of UCS is Activity-3, which
is uncrashed to its second mode (M2). For none of the schedules of the first UCS
cycle, UFF process was applied since there were no uncrashable non-critical
activities in the network. As mentioned earlier, status of each solution which gets
selected as a base for generation of new schedules, is switched to “Closed” at the
end of each UCS phase. Accordingly, at the end of the first UCS cycle the status
of Schedule-1.v is changed to “Closed” and the solution in Result is replaced by
Schedule-2 having the least total cost found so far. Thereafter, the components of
the Solutions are erased. These iterative cycles are repeated for the remainder of
the time-cost options. Since Schedule-10 and Schedule-12 exceed the deadline of
45 days by two days, the overall cost is increased from $197,000 to $201,000 for
these solutions. Finally, the CS-Heuristic terminates by returning the last Sol
recorded in Result – Schedule-9, highlighted in boldface in Table 5.1 – with a
duration of 45 days and a total cost of $199,600.

5.1.4. CS-Heuristic for Pareto front DTCTP

Similar to the heuristic approach presented in Section 5.1.3, a new Cost-Slope

Heuristic is presented herein for Pareto front discrete TCT problem. This method
also incorporates the project deadline, since, in practice there is a completion
deadline stipulated in the contract for the majority of the projects. In realistic
projects, delay penalties which are usually in the form of liquidated damages are
applied in case the project duration exceeds the predetermined deadline and the
incentives are the bonus payments made for each day saved from the specified
deadline. Generally, there is a negative correlation between the deadline and the
complexity of the problem. That is, projects with larger predetermined thresholds
are simpler than the ones with shorter deadlines. Regarding the above fact, the
completion deadline of the practiced problems is calculated in the proposed
heuristic algorithm as follows. Firstly, the all-normal schedule with the largest

133
critical path length, CPM max , is calculated. Secondly, the all-crashed schedule with
the shortest critical path, CPM min , is computed. Finally, the completion deadline is
set to be equal to the average of the earliest allowable completion time of the project
and the latest possible finishing time.

The first step of the CS-Heuristic involves serial/parallel merging of the network
activities. Based on the techniques presented in Section 5.1.1.1 and Section 5.1.1.2
CS-Heuristic reduces the network of the problem to a simpler equivalent network.
Thereafter it calculates the completion deadline based on the above described
procedure. In the next step, CS-Heuristic calculates and records the cost-slopes (
CS ), and crash amounts (duration difference DDiff , and cost difference CDiff )
for every time-cost option of all the activities. The cost-slopes are calculated
according to Eq. (3.1) given in Section 3.2. All these options are then indexed as
follows. Time-cost alternatives are sorted in descending order by CDiff values then
by CS values in descending order; thereby they are treated with an index named
UFF  index . Similar to the previous stage, time-cost alternatives are sorted in
descending order by CS values then by DDiff values in ascending order, thereby
an index called UCS  index is assigned to each alternative. According to
UFF  index , a component providing a greater cost saving is uncrashed first; in
case of a tie, the component with a greater cost-slope rate is selected; if the tie is
still not broken, the first component is selected. According to UCS  index , a
component with a greater cost-slope rate is uncrashed first; in case of a tie, the
component providing a smaller duration reduction is selected; if the tie is still not
broken, the first component is selected.

Unlike earlier heuristic approaches, in lieu of starting from the all-normal schedule,
CPM max , CS-Heuristic initiates from the all-crashed schedule, CPM min , by
selecting the shortest/costliest time-cost alternative for the activities. Contrary to
previous heuristics, in every cycle of CS-Heuristic, each new schedule is built upon

134
an already non-dominated solution. Duration of this schedule is determined using
CPM technique and the total cost is calculated by explicitly including the incentive
and disincentive payments in the project cost formulation. Since CS-Heuristic starts
from the all-crashed schedule, it is designed to optimize the solutions by performing
an uncrashing scheme. The uncrashing scheme incorporates two different phases,
namely, uncrash free-float ( UFF ) and uncrash cost-slope ( UCS ). Through UFF
phase those activities with total floats greater than or equal to their effective
(immediate available) crash amounts, DDiff jk  TF j , are determined. The

determined activities are then uncrashed to their next available option, by

uncrashing one activity at a time, with respect to their UFF  index values. After
performing each UFF , CS-Heuristic uses Partial-CPM technique (Section 5.1.2) to
update the modified activities. The UFF process is continued until either all the
activities get fully uncrashed or their total floats become less than their effective
crash amounts. It is after this step that the heuristic stores the current schedule as a
solution. While each solution, Sol , obtained from the uncrashing phases are
recorded in an archive called Solutions , an external repository, Pareto , is
implemented to record the non-dominated solutions. Accordingly, the non-
dominated Sol s of Solutions are copied to Pareto and the solutions stored in
Pareto are sorted according to their durations in ascending order; subsequently,
the schedules stored in Solutions are wiped out at the end of each iteration.

After the first set of non-dominated schedules are generated and stored in Pareto ,
CS-Heuristic performs UCS process on the first non-dominated solution – with the
least duration – archived in Pareto to generate new schedules unless the status of
the selected solution is “Closed”. Status of each solution which gets selected as a
base for generation of new schedules, is switched to “Closed” at the end of each
UCS phase, making it unavailable for re-selection. New schedules are generated
based on the non-dominated solutions which are already stored in Pareto , as long
as they are not “Closed”. Through UCS phase those activities that are uncrashable,
i.e., not fully uncrashed, are determined. The determined activities are then

135
uncrashed to their next available option, by uncrashing one activity at a time, with
respect to their UCS  index values. After performing each UCS , CS-Heuristic
implements the short procedure described in Section 5.1.2 to calculate the
completion time of the obtained schedule. In case there exists no Sol stored in
Solutions with the same duration, CPM calculation is carried out which is then
followed by the UFF process as mentioned earlier. It is after this step that the
heuristic stores the current schedule as a solution in Solutions . However, through
UCS phase, if the same duration already exists in Solutions within the same cycle,
that schedule is discarded without being stored as a Sol in the archive. The UCS
process is continued until all the determined uncrashable activities are uncrashed to
their longer durations. After completing each UCS phase, the status of solution,
Sol  Pareto , which was selected at the start of the cycle is switched to “Closed”
and the dominated solutions are removed from Solutions . UCS and UFF phases
are iteratively applied to any of the uncrashable activities of the non-dominated
solutions,  Sol  (Solutions or Pareto) : status  Closed , until obtaining a
solution, all of the activities of which are uncrashed to their longest durations. CS-
Heuristic terminates by returning the ultimate non-dominated Sol s stored in
Pareto with “Closed” statuses. The flowcharts of parallel merge, serial merge, and
uncrash free-float ( UFF ) modules of the proposed CS-Heuristic are presented in
Figure 5.5, Figure 5.6, and Figure 5.7 of Section 5.1.3, respectively. The proposed
CS-Heuristic approach for Pareto front DTCTP is graphically explained as a
flowchart in Figure 5.9 and the pseudo-code of this method is demonstrated in
Figure 5.10.

136
 Start

Retrieve 
Parallel Merge Serial Merge Calculate Deadline
Project Info.

No Yes Calculate CS, DDiff,  Index by CDiff, then 

j≤S CDiff by CS for UFF

Yes
k≤m
No Index by CS, then 
Increment j Increment k
by DDiff for UCS

Fully Crash all  Uncrash Free-

Calculate CPM Calculate Dur/Cost
Activities Float

No
Yes Uncrash based on  Perform Partial-
Store Sol in Pareto Sol in not Closed
UCS Index CPM

Same Dur in  No
Calculate Dur Calculate CPM Calculate Cost
Solutions

Yes

Uncrash Free- Store Sol in  Remove dominated 

Close Sol
Float Solutions Sol’s from Solutions

No

Copy Solutions to  Yes max Sol index 

Increment Sol index
Pareto reached

Remove dominated  Output 
Sol’s from Pareto Pareto

End

Figure 5.9 – Flowchart of the proposed CS-Heuristic for Pareto front DTCTP.

137
Begin;
For ∀ File in Directory
For ∀ j ∈ [1, S]
For ∀ k ∈ [1, m]
Retrieve Values;
End;
End;
For ∀ j ∈ Acts
Calculate Predecessors;
End;
Sort Acts;
If Perform Parallel Merge = True
For ∀ j ∈ Acts
If j has 2 Successors
If Successors of Successors to j ∧ Predecessors of Successors to j are same
Merge Successors to j;
End;
End;
End;
End;
If Perform Serial Merge = True
For ∀ j ∈ Acts
If j has 1 Successor
If Successor to j has 1 Predecessor
Merge j and its Successor;
End;
End;
End;
End;
Calculate Project Deadline;
For ∀ j ∈ Acts
For ∀ k ∈ Modes
Calculate CS, DDiff, CDiff;
Sort and Index by CDiff then by CS for Uncrash Free-Float;
Sort and Index by CS then by DDiff for Uncrash Cost-Slope;
End;
Select 1st Mode;
End;
Calculate CPM;
Calculate Dur/Cost;
While ∃ j ∈ Acts: DDiff ≤ Total-Float
Add to Uncrashables;
Perform Uncrash Free-Float;
Perform Partial-CPM;
Calculate Cost;
Break;
Store Sol in Pareto;

Figure 5.10 – Pseudo-code of the proposed CS-Heuristic for Pareto front DTCTP.

138
 While ∃ Sol ∈ Pareto: Sol ≠ Closed
For ∀ Act ∈ Sol: Fully-Uncrashed ≠ True
Add to Uncrashables;
End;
Perform Uncrash Cost-Slope;
Perform Partial-CPM;
Calculate Dur;
If ∃ Sol ∈ Solutions: Sol.Dur = Dur
;
End;
Else
Calculate CPM;
Calculate Cost;
While ∃ j ∈ Acts: DDiff ≤ Total-Float
Add to Uncrashables;
Perform Uncrash Free-Float;
Perform Partial-CPM;
Calculate Cost;
Break;
Store Sol in Solutions;
End;
Close Sol;
For ∀ Sol ∈ Solutions: Non-dominated ≠ True
Remove Sol;
End;
Copy Solutions to Pareto;
Break;
For ∀ Sol ∈ Pareto: Non-dominated ≠ True
Remove Sol;
End;
Return Pareto;
End;
End;

Figure 5.10 – Pseudo-code of the proposed CS-Heuristic for Pareto front DTCTP (Continued).

The proposed CS-Heuristic is explained using the same case problem which is
presented in Section 3.3.1. The original network shown in Figure 3.1 is used to
elucidate UFF and UCS processes of the CS-Heuristic. For the all-crashed
schedule, the project duration is 31 days and the total cost is $270,600 as shown in
Table 5.3.

139
 Table 5.3 – Candidate solutions found by CS-Heuristic for deadline DTCTP.

# of Continue Crash Dur. Direct Indirect Total Cost

Activity Uncrash
Schedule from Level (day) Cost ($) Cost ($) ($)
1 0 Crashed 31 239,600 31,000 270,600
1.i 1 4 0 M1 to M2 31 221,600 31,000 252,600
1.ii 1.i 5 0 M1 to M2 31 205,600 31,000 236,600
1.iii 1.ii 2 0 M1 to M2 31 193,600 31,000 224,600
1.iv 1.iii 2 0 M2 to M3 31 187,600 31,000 218,600
1.v 1.iv 5 0 M2 to M3 31 185,600 31,000 216,600
2 1.v 6 1 M1 to M2 33 173,600 33,000 206,600
- 1.v 1 1 M1 to M2 33 - - -
3 1.v 3 1 M1 to M2 35 178,600 35,000 213,600
4 2 1 2 M1 to M2 35 169,000 35,000 204,000
5 3 3 2 M1 to M2 37 166,600 37,000 203,600
6 4 3 3 M1 to M2 39 162,000 39,000 201,000
7 4 1 3 M2 to M3 47 166,000 47,000 213,000
8 5 1 4 M1 to M2 39 162,000 39,000 201,000
9 5 3 4 M2 to M3 45 154,600 45,000 199,600
10 6 3 5 M2 to M3 47 150,000 47,000 201,000*
11 6 1 5 M2 to M3 51 159,000 51,000 210,000
12 9 1 6 M1 to M2 47 150,000 47,000 201,000*
*Exceeds deadline by 2 days

In the Schedule-1 which consists of crashed modes, activities 1, 3, and 6 are on the
critical path. According to UFF , among the non-critical activities, Activity-4 with
a CDiff of $18,000 provides the greatest cost saving as shown in Table 5.2 given
in Section 5.1.3. Thus, Activity-4 is uncrashed first, to its second mode (M2). The
same procedure is applied to Activity-5 which is uncrashed to its second mode
(M2), similarly, Activity-2 is also uncrashed to its second mode (M2). In schedules
Schedule-1.iv and Schedule-1.v, Activity-2 and Activity-5 are uncrashed to their
third option (M3), respectively. The resulting Schedule-1.v has a duration of 31
days and a total cost of $216,600 (Table 5.3) which is recorded in Pareto as the
first UFF cycle terminates. Following UFF phase, five uncrashing options could
be identified for UCS including Activity-1: M1 to M2, Activity-1: M2 to M3,
Activity-3: M1 to M2, Activity-3: M2 to M3, and Activity-6: M1 to M2. However,
M2 to M3 crash options of Activity-1 and Activity-3 are excluded from the

140
uncrashables, since, only the effective (immediate available) crash options are
considered throughout UCS process. Therefore, succeeding UFF phase, Activity-
6 with largest cost-slope of $6,000/day is uncrashed to its second mode (M2) in the
course of UCS phase and stored as Schedule-2 in Solutions . The next least-cost-
slope activity is Activity-1, however, uncrashing it to its second mode (M2) results
in a solution with a duration of 33 days which already exists as Schehule-2 in
Solutions ; hence, this schedule is discarded. The final uncrashable activity in this
cycle of UCS is Activity-3, which is uncrashed to its second mode (M2). For none
of the schedules of the first UCS cycle, UFF process was applied since there were
no uncrashable non-critical activities in the network. At the end of the first UCS
cycle the non-dominated Sol in Solutions , i.e., Schedule-2, is copied to Pareto .
As mentioned earlier, status of each solutions which gets selected as a base for
generation of new schedules, is switched to “Closed” at the end of each UCS phase.
Accordingly, the status of Schedule-1.v is then changed to “Closed” and the
components of the Solutions are erased. In the second cycle of UCS , the least-cost
schedule is copied to now blank Solutions , unless its status is “Closed”. In this
case, Schedule-2 is copied to the repository to carry out UCS . Based on this
schedule, Schedule-4 is generated by uncrashing Activity-1 to its second mode
(M2). Next, Activity-3 is uncrashed to its second mode (M2). Since Activity-3 is
already uncrashed to its second mode (M2) in the preceding cycle of UCS , it is
now possible for this activity to be uncrashed to its third mode (M3) in the third
cycle of UCS , which results in Schedule-6 as shown in Table 5.3. The final
uncrashable activity in this cycle of UCS is Activity-1. The resulting schedule has
a duration of 47 and total cost of $213,000, which is already dominated by
Schedule-6 stored in Solutions . In fourth cycle of UCS , Schedule-5 is used as the
base schedule. Uncrashing Activity-1 with the largest CS results in exactly the
same solution already recorded in Pareto as Schedule-6. The final uncrashable
activity of cycle five is Activity-3 which is uncrashed to its third option (M3). In
the sixth cycle of UCS , Schedule-10 is obtained, completion time of which exceed
the deadline by two days. The next schedule achieved in this cycle include

141
Schedule-11 as shown in Table 5.3. Through the last cycle of UCS , Schedule-12 is
obtained which also exceeds the deadline of 45 days by two days, with an overall
cost of $201,000. Finally, the CS-Heuristic terminates by returning the last non-
dominated Sol s recorded in Pareto which includes Schedule-1.v, Schedule-2,
Schedule-4, Schedule-5, Schedule-6, and Schedule-9. The highlighted schedules
shown in boldface in Table 5.3 constitute the final set of non-dominated solutions
found by the CS-Heuristic for the case problem.

5.2. Computational Experiments of CS-Heuristic

5.2.1. Generation of New Sets of Instances

It is mentioned earlier that no standard test-bed is available in the literature that

could be considered to be as complex as the real-life projects. Therefore, several
random DTCTP instances are generated systematically in this thesis to measure the
performances and evaluate the capabilities of the proposed optimization algorithms.
There exist numerous random network generators proposed within the literature
including ProGen/Max (Schwindt, 1995), RanGen (Demeulemeester, Dodin, and
Herroelen, 1993; Demeulemeester, Vanhoucke, and Herroelen, 2003), DAGEN
(Agrawal, Elmaghraby, and Herroelen, 1996), and RanGen2 (Vanhoucke, Coelho,
Debels, Maenhout, and Tavares, 2008). Except for DAGEN, all the other network
generators generate AoN type of networks. Since this thesis uses AoN notation
system, DAGEN generator is dismissed. On the other hand, it is shown in the study
by Demeulemeester et al. (2003) that RanGen is more advantageous compared with
other earlier instance generators including ProGen/Max and DAGEN and
topological sorting of the networks generated by means of RanGen are more
complicated. In a more recent study, Vanhoucke et al. (2008) show their RanGen2
generator outperforms ProGen (Kolisch, Sprecher, and Drexl, 1995), RanGen, and
RiskNet (Tavares, 1999) generators with respect to the total amount of networks
generated. RanGen2 is capable of generating networks with complexities on par

142
with the original RanGen. In addition, RanGen2 is shown to best the original
RanGen since it starts from a larger pool of possible networks and is able to generate
different networks, whereas, RanGen is able to generate exactly one network per
run. Accordingly, in this thesis, RanGen2 random network generator (Figure 5.11)
is implemented to generate strongly random activity networks.

Figure 5.11 – RanGen2 Interface.

RanGen2 is actually a modified version of RanGen which incorporates alternative

topological indicators. RanGen2 aims at generating different topological structures
using predefined values for I 1 , I 2 , I 3 , I 4 , I 5 , and I 6 parameters. These
parameters are defined as follows (Vanhoucke et al., 2008):

- I 1 : Network size indicator: Number of activities;

- I 2 : Serial/parallel indicator: Closeness of a network to a serial ( I 2  1 , a chain
of activities) or parallel ( I 2  0 , no precedence relations) graph;

143
- I 3 : Activity distribution indicator: Distribution of the activities over the
network ( I 3  0 , uniform distribution of activities);
- I 4 : Short precedence relations indicator: Presence of short precedence
relations;
- I 5 : Long precedence relations indicator: Presence of long precedence relations;
- I 6 : Topological float: Topological float of activities in the network ( I 6  0 , all
activities have topological float of zero).

New sets of DTCT problems are created as txt files by combining complex
networks – generated by means of RanGen2 – with realistic sets of time-cost
alternatives – developed by means of a code written in C# based on Eq. (5.1) given
below. Characteristics of the generated random instances can be summarized as
follows:

- I 1 : Number of activities include five levels of 50, 100, 200, 500, and 990;
- I 2 : Serial/Parallel factor is varied from 0.2 (almost parallel) to 0.8 (almost
serial) in steps of 0.2, i.e., 0.2, 0.4, 0.6, and 0.8;
- I 3 to I 6 : These parameters are drawn randomly from U(0,1) in order to make
a balanced representation of different network properties;
- m : Number of time-cost alternatives include three fixed levels of 3, 6, and 9.

For each parameter level combination, i.e., I 1 , I 2 , and m , ten replications of

networks are generated leading to a total of 5  4  3 10  600 instances. In doing
so, a wide range of complexity with respect to the topological structure is captured.
The reason behind generation of instances with 990 activities is simply because
RanGen2 is unable to generate samples for greater I 1 values. Each of the project
sizes can only have a fixed number of alternatives as 3, 6, or 9 which are generated
by means of non-increasing convex time-cost functions using the procedure
described in Akkan et al. (2005). Fixed number of time-cost alternatives are

144
assigned for the activities since according to Vanhoucke and Debels (2007), such
problems are more difficult to solve than the instances that involve a random
number selected from a predefined range. As stated by Vanhoucke and Debels
(2007), variation in the number of modes through the activities reduces the
complexity of the problems.

For each activity, the time-cost alternatives are generated as follows: First the
number of modes ( m ) is decided for all the activities, i.e., in an instance all the
activities include either 3, 6, or 9 alternatives. Then, the durations of these modes
are randomly sampled from DU(1,54) (i.e., discrete uniform distribution with
parameters 1 and 54) as follows: The range 1-54 is divided into m number of
intervals. Duration for the normal mode is selected randomly from the last interval,
and the duration of each of the succeeding alternatives are assigned randomly from
the corresponding preceding intervals; viz., duration for the crashed mode is
selected from the first interval using a random scheme. After determining all the
duration amounts, the costs of the modes are generated sequentially, starting with
the normal mode (least-cost alternative) by multiplying its duration by a value (
dc Int ) randomly sampled from DU(500, 2000) . Direct costs of the subsequent
alternatives are generated using Eq. (5.1).

dc jk  dc j(k 1)  CS jk  dc Int  (d j(k 1)  d jk )

(5.1)
 j  {1,..., S} ,  k  {2,..., m( j)}

where dc jk is the direct cost of the k th alternative of the j th activity; dc j(k 1) is

the direct cost of the (k  1) th alternative of th j th activity; CS jk is the cost-slope

percentage of k th option of j th activity which is randomly sampled from the

uniform distribution U(0.1, 0.5) as follows: The range 10%-50% is divided into
(m  1) number of intervals, CS jk for the k th alternative is selected randomly from

the first interval and CS jk of each of the succeeding alternatives are selected

145
randomly from the corresponding subsequent intervals which secures an
incremental cost-slope pattern; dc Int is a value randomly selected from
DU(500, 2000) ; d j(k 1) and d jk are durations of (k  1) th and k th options,

respectively. According to Eq. (5.1), mutually incomparable modes are generated

for the activities. Besides, the implemented incremental cost-slopes scheme is
capable of reflecting the realistic decline in productivity rate, since, activities are
usually crashed by overstaffing or by working overtime. As a result of this, the
realistic crash options lead to a convex solution space. The generated instances, akin
to real-life projects, have a convex time-cost relationship for the projects as a whole
and for each of its activities.

Completion deadlines are also included in the new sets of instances, since, in
practice there is a completion deadline stipulated in the contract for the majority of
the projects. In realistic projects, delay penalties which are usually in the form of
liquidated damages are applied in case the project duration exceeds the
predetermined deadline. Generally, there is a negative correlation between the
deadline and the complexity of the problem. That is, projects with larger
predetermined thresholds are simpler than the ones with shorter deadlines.
Regarding the above fact, the completion deadlines of the generated instances are
calculated as follows. Firstly, the all-normal schedule with the largest critical path
length, CPM max , is calculated. Secondly, the all-crashed schedule with the shortest
critical path, CPM min , is computed. Finally, the completion deadline is set to be
equal to the average of the earliest allowable completion time of the project and the
latest possible finishing time. The indirect cost rate is set as $1000/day for all the
instances and a twofold of the indirect cost is used as the amount of the delay
penalty, i.e., $2,000/day. The order of numbering for each setting of the generated
instances is illustrated in Table 5.4. As presented in Section 5.2.3 and Section
5.2.5.4, behavior of the proposed exact and heuristic/meta-heuristic methods are
dissimilar for each corresponding setting of the parameters. According to the
findings of Section 5.2.3, the nature of mixed-integer linear programming technique

146
tends to be more suitable for pseudo-serial networks with greater I 2 values that
include smaller number of time-cost modes. Contrary to mixed-integer linear
programming technique, results of Section 5.2.5.4 reveal that the nature of the
proposed heuristic and meta-heuristic approaches are more suitable for solution of
pseudo-parallel networks with smaller I 2 values that comprise larger number of
time-cost alternatives. Thus, problems 21 to 30 are considered to be the most
complicated, and the problems 91 to 100 are regarded as the simplest problems for
the proposed exact algorithm. The contrary, problems 91 to 100 are experienced to
be the most complex, while, problems 21 to 30 to be the least complicated instances
for the proposed heuristic/meta-heuristic solution procedures.

Table 5.4 – Complexity of the generated instances.

I2
# of Modes 0.2 0.4 0.6 0.8
3 1 to 10 31 to 40 61 to 70 91 to 100
6 11 to 20 41 to 50 71 to 80 101 to 110
9 21 to 30 51 to 60 81 to 90 111 to 120

5.2.2. Performance Indices

Obviously, the results obtained from multi-objective optimizations comprise a set

of solutions rather than a single optimal solution. Of the multi-objective
optimization problems, Pareto front DTCTP yields a sequence of solutions for the
problems with conflicting objectives of time and cost. The obtained sets of results
for Pareto oriented DTCTP are mutually incomparable and non-dominated with
respect to multiple objectives of the project. Consequently, the definition of quality
is complicated within this context and for Pareto front DTCTP it cannot be
evaluated using the concept of optimality which is exercised in single objective
optimizations. Owing to this fact, numerous performance indices have been
proposed for performance evaluation of multi-objective methods which mainly

147
engage three aspects of the generated solutions as follows (Zitzler, Deb, and Thiele,
2000):
- The number of the located Pareto optimal solutions – measured using
cardinality indices;
- The accuracy of the captured Pareto fronts – measured using accuracy indices;
- The diversity (distribution and spread) of the achieved Pareto optimal solutions
– measured using diversity indices.

There is also a second classification criterion which groups performance metrics

into unary and binary indexes. The first group defines a value by considering a
single set of solutions while the second group assigns relative values for two
comparable sets. Overall Non-dominated Vector Generation (ONVG) and Error
Ratio (ER) are among indices mainly used for measuring cardinality of the Pareto
fronts (Knowles and Corne, 2002; Okabe, Jin, and Sendhoff, 2003). Generational
Distance (GD) and Inverted Generational Distance (IGD) are among the indexes
mainly utilized in evaluation of the accuracy of the solutions (Zitzler et al., 2000,
Okabe et al. 2003, Zhang and Li 2010, Riquelme, Von Lucken, and Baran, 2015).
Diversity of the Pareto fronts are measured using indices that include Range
Variance (RV),  , Uniform Distribution (UD), and Chi-Square-Like Deviation
(Zitzler et al., 2000; Okabe et al., 2003; Zhang and Li, 2010). Hypervolume (HV),
also known as S -metric is able to capture all three aspects of cardinality, accuracy,
and diversity of the Pareto solutions (Zitzler and Thiele, 1999; Riquelme et al.,
2015). According to Riquelme et al. (2015), HV, GD, Epsilon Indicator (  ), and
IGD are the most practiced indexes in the domain of multi-objective optimization.
Knowles and Corne (2002) after studying various indexes recommend utilization
of R-metrics (Hansen and Jaskiewicz, 1998) and HV in performance evaluations.
By comparing a multitude of performance indices, Okabe et al. (2003) argue that a
single unary index cannot adequately reflect all the three aspect of the Pareto fronts
and that binary indices are more suitable for comparing different sets of solutions.
It is also discussed that the abovementioned indices might occasionally deliver

148
misleading information since they are not free of cons/caveats (Knowles and Corne,
2002). Meanwhile, it is observed that the convergence speed is not emphasized
sufficiently in the literature, however, it is one of the most important aspects of an
optimization method. For any practical application of a solution method, the
effectiveness needs to be accompanied with the efficiency.

Respecting the findings on the metrics mentioned above, a more holistic approach
is taken toward performance evaluation of the approaches introduced in this thesis
by employing a collection of unary and binary performance indices along with the
CPU times. The unary ONVG metric is used to measure cardinality of the obtained
frontiers. A binary metric called ND pct is also introduced in this thesis to measure

the cardinality of the Pareto fronts. ND pct measures the percentage of the non-

dominated solutions achieved by different optimization methods after discarding

the dominated solutions from the unified front, UF . In the cases when true Pareto
fronts are available, two sorts (unary and binary types) of Average Percent
Deviation (APD) values are calculated to assess accuracy of the solutions. The
binary index, APD bin , is calculated for the cost figures based only on the
completion duration amounts located mutually by all of the different approaches;
and the unary APDs are measured for each approach independently. A normalized,
HV-based approach called Hyperarea Ratio (HR) (Veldhuizen, 1999) is adopted to
measure convergence and diversity of the frontiers. Unary HR is measured by
HV PF
calculating the ratio as ; where HV PF is the Hypervolume of the
HVTP
approximated Pareto front and HVTP is the Hypervolume of the true Pareto front.
In the cases when the true Pareto fronts are unavailable, the binary HR is calculated
HV PF
as ; where HVUF is the Hypervolume of the unified front consisting of the
HVUF
individual solutions in all the approximate frontiers, excluding the dominated ones.
Hypervolume indicates the partition of the solution space bounded by the non-
dominated frontier and a reference point ( R ) (Zitzler and Thiele, 1998). For each

149
Sol  Pareto , an area is constructed using a reference point and Sol as the diagonal
corner of the area. The reference point for the first area, i.e., R , is defined by
increasing each of the maximum objective function values by 0.5%. The maximum
objective function values are derived with regard to both PF and TP in case the
optimal front is at hand, or it is set by considering UF if the true Pareto is
unavailable. Intersection points of the preceding Sol s and R are appointed as the
reference points for the subsequent areas. The union of the constructed areas yields
the Hypervolume for that frontier. Consequently, the value of HR belongs to U(0,1)
; where values close to one suggest the obtained frontier is close to the best-known
Pareto front (either TP or UF ). Higher values for ONVGs, ND pct , and HR are

desirable; whereas, lower values for both the unary and binary APDs and CPU time
are preferred. All the performance evaluation procedures have been implemented
in C# programming language using Microsoft Visual Studio 2013.

The procedure exercised for Hypervolume calculation of the solution sets is

elucidated by means of a hypothetical case example. A shown in Table 5.5, a true
Pareto, TP , is assumed to comprise three optimal solutions, on the other hand, two
approximated Pareto fronts, PF a and PFb , are assumed to include three and five
solutions, respectively.

Table 5.5 – Hypothetical Pareto fronts.

TP PF a PFb
Dur. (day) Cost ($) Dur. (day) Cost ($) Dur. (day) Cost ($)
44 648,000 44 648,000 44 648,000
53 621,000 53 642,000 53 621,000
64 620,000 62 641,000 64 620,000
64 637,000
69 620,000

As shown in Figure 5.12, the coordinates of the reference point R are calculated
with regard to the maximum duration and cost amounts among all the three sets of

150
TP , PF a , and PFb as follows: R Dur.  1.05  69  72.45 and
R Cost  1.005  648, 000  651, 240 . As a side note, R Dur. in this case example is
calculated by augmenting the maximum duration by 5% to obtain an improved
visualization of the Hypervolumes; though, both R Cost and R Dur. are set by
increasing the worst objective functions by 0.5% in the following sections of this
thesis.

Figure 5.12 – Hypothetical Hypervolume comparison for PF a .

As demonstrated in Figure 5.12, the area highlighted in red resembles HV PFa and

the area highlighted in green shows HVTP . HR for this set of solutions is calculated
163810
as HR a   0.48 . Similarly, Hypervolume of PFb is shown in red in
334610
Figure 5.13 against the same Hypervolume of TP shown in green. In Figure 5.12
and Figure 5.13 Hypervolumes are extruded to enhance the graphical representation
of the areas.

151
 Figure 5.13 – Hypothetical Hypervolume comparison for PFb .

334610
HR for PFb is calculated as HR b   1 . This is a clear indication of why
334610
a multitude of performance metrics must be implemented. Although the ONVG of
PF a is greater than that of PFb , it has been shown that the second set of solutions
with HR  1 have an exact compliance with the true Pareto front. PFb with an HR
value of one is shown to outclass PF a with a lower HR value of 0.48, despite
capturing fewer number of non-dominated solutions. Therefore, this thesis study
uses the aforementioned performance indexes in a complementary fashion to come
up with a more precise verdict on the performances of the proposed approaches. To
the best of author’s knowledge, Zhang and Li (2010), Kalhor et al. (2011), Fallah-
Mehdipour et al. (2012), and Mungle et al. (2013) are among the few studies within
the construction management literature reporting on different performance indices
of their proposed optimization approaches.

152
5.2.3. Mixed-Integer Linear Programming Technique

Since all the extensions of DTCTP are Non-deterministic polynomial-time hard

(NP-hard) problems in the strong sense (De et al., 1997), only a few studies as
discussed in Section 2.3 present exact methods for the DTCTP, especially for the
more complex Pareto front problem. The nature of Mixed-Integer Linear
Programming model, abridged as MILP (also known as MIP), is well-suited for the
solution of DTCTP. The incremental cost-slopes scheme of realistic projects which
captures the decline in productivity rate of the crashed activities lead to a convex
solution space. The convexity of the solution space guarantees converging to global
optimal solutions by means of a MILP algorithm. Any type of time-cost trade-off
problem can be modeled using MILP and solved by means of a compatible
commercial optimization software. MILP models for DTCTP problems generally
include four to five parameters per each activity (Bettemir, 2009); thus, the large-
scale problems involve a considerably high number of parameters. Nonetheless,
MILP is guaranteed to locate the optima for DTCTP unless there are physical
limitations in the memory of the computing devices. Inasmuch as the exact methods
are the only approaches guaranteeing optimality of the solutions, they play a crucial
role in experimentation of heuristic and meta-heuristic algorithms. In order for
adequate evaluation of the effectiveness of the proposed approaches by means of
the performance metrics introduced in Section 5.2.2, a MILP model is developed
using Gurobi solver version 6.0.5.

Primarily, the proposed MILP employs the serial/parallel merging technique

presented in Section 5.1.1.1 and Section 5.1.1.2 to reduce the network of the
problem to a simpler equivalent network. As mentioned in Section 5.1.1.1, MILP
takes a serial merging approach similar to that of CS-Heuristic described in Section
5.1.3 and Section 5.1.4; however, for the MILP in case of a tie, all but one of the
duplicate alternatives with d k j   d (k 1) j  and dc k j   dc (k 1) j  are eliminated. A
1 1 1 1

dominance rule is also implemented to eliminate the dominated components as

153
k j1  (k  1) j1 , which indicates if d k j   d (k 1) j  while dc k j   dc (k 1) j  , (k  1) th
1 1 1 1

alternative of the equivalent activity j 1  will be eliminated. This procedure is

explained using the case example presented in Section 3.3.1. It is shown in Figure
5.1 that the equivalent Activity-1 would have included seven time-cost alternatives
in case CS-Heuristic’s merging technique was applied. However, for MILP, the
dominated time-cost alternative with a duration of 46 days and direct cost of
$59,000 – resulted from adding the third mode of the original Activity-1 to the
second mode of the original Activity-3 – is retained and added to the available
options (Figure 5.14). The reason behind not discarding the dominated component
is simply to prevent generation of incomplete true Pareto fronts. The solution space
of the original network comprises 324 realizations; however, the solution space of
the reduced network includes 288 different realizations. As is clear, the size of the
solutions space shrinks only marginally which might provide a rather small benefit
and might not justify the extra efforts involved in merging the activities serially.
Owing to this very reason, only the parallel merging technique is exercised for
optimal solution of the practiced instances in Section 5.2.5.

1. (28, $73,600)
2. (30, $69,000)
3. (32, $66,600)
4. (34, $62,000)
5. (40, $54,600)
6. (42, $50,000)
3 5
7. (46, $59,000)
Start 8. (54, $47,000) 1. (5, $52,000) 1. (3, $36,000) Finish
2. (8, $34,000) 2. (5, $24,000)
2

1. (7, $38,000)
2. (10, $26,000)
3. (12, $20,000) 4 ID

1. (10, $40,000) 1. (Dur., Cost)

2. (15, $24,000) m. (Dur., Cost)
3. (16, $22,000)

Figure 5.14 – Serial merge applied to the Case Example.

154
After reducing the problem network, MILP calculates the completion deadline
based on the procedure described in Section 5.1.3. The proposed MILP model is
based on the modified formulation that explicitly includes bonus ( db ), delay
penalty ( dp ), and daily indirect cost ( ic ) rates. The modified formulation for MILP
(Bilir, 2015) is presented in Eqs. (1.1)-(1.6) in Section 1.1.2. Gurobi version 6.0.5
is adopted among different commercial optimization software for solution of the
MILP model for DTCTP. Gurobi is a high-speed powerful solver with an inclusive
library for different programming environments such as C# and C++. Gurobi
provides a flexible licensing along with a major collection of mathematical
programming solvers, such as MILP. On the other hand, Gurobi is shown to be the
fastest solver (Mittelmann, 2013) among different optimization software (e.g.,
CPLEX, MATLAB) to capture optimality, feasibility, and infeasibility of a set of
benchmark problems (MIPLIB2010) which can be found in the “MILP
Benchmark” section of the site maintained by Mittelmann.

Typically, Gurobi is designed to read problems in LP-file format. However, an

online converter (written as a block of code) is implemented to enable the MILP
method to read from the generic txt file format which is universally used in all of
the approaches proposed in this thesis. Presolve process, which is a preprocessing
technique frequently applied within the MILP-models, is experienced to contribute
to a larger computational time. Therefore, it is excluded from the optimization
model. Two MILP models are designed and developed for deadline and Pareto front
DTCTP separately. For Pareto front problem, upper and lower boundaries are set
for the duration of the problem in accord with all-normal, CPM max , and all-crashed,
CPM min , schedules, respectively. Moreover, a horizon-varying approach is also
implemented for unraveling Pareto front problem which involves iterative solution
of deadline DTCTP by varying D dl from CPM min to CPM max , in steps of one.
Pareto oriented MILP is also complemented with an upper-bound calculation
mechanism for the cost of the feasible durations. The value of the upper-bound is
designed to update to the cost figure located in the preceding iteration of the

155
horizon-varying approach. This is facilitated by means of the Cutoff parameter of
Gurobi which basically indicates that the solver should only consider solutions with
total costs less than the specified amount. Inclusion of the upper-bound rules out
the need for elimination of the dominated solutions from the final Pareto, since, it
only locates the non-dominated solutions in the course of the optimization.

Simulation routines of the proposed MILP method are coded in C# and compiled
within Visual Studio 2013 on a 64-bit platform. All of the tests are carried out on a
desktop computer with a P9X79 Chipset motherboard, 16 GB 667 MHz DDR3
random-access memory (RAM), Intel Core i7-3.40 GHz CPU, and 64-bit Windows
10 operating system. MILP is executed solely (no other software is ran
simultaneously) on a single processor and overclocking is not performed. All the
runs are truncated after a CPU time of one hour which is enforced using TimeLimit
parameter of Gurobi. If it exceeds the runtime limit, the algorithm will terminate by
reporting a non-optimal status.

The new sets of instances presented in Section 5.2.1 are fitted into the proposed
MILP model in order to locate the optimal solutions for deadline and Pareto front
DTCTPs. The percentage of the instances solved within the enforced computation
time limit of one hour are demonstrated for deadline and Pareto front DTCTPs in
Table 5.6 and Table 5.8, respectively. These tables summarize percentages for
every combination of different network sizes, network complexities, and mode
numbers. The average CPU times required by MILP method for tackling every ten
replicate problems of each parameter level combination are illustrated in Table 5.7
and Table 5.9 for deadline and Pareto front DTCTPs, respectively. Unavailable
values are tabulated as ‘na’ in Table 5.7 and Table 5.9.

156
 Table 5.6 – Percentage of RanGen2 instances optimally solved for deadline DTCTP.

% of Solved Instances
# of Acts. 50 100 200 500 990
I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
# of Modes
3 100 100 100 100 100 100 100 100 100 50 100 100 0 100 100 100 0 0 0 100
6 100 100 100 100 100 100 100 100 70 0 100 100 0 20 20 100 0 0 0 100
9 100 100 100 100 90 100 100 100 20 20 100 100 0 30 20 100 0 0 0 100

157
Table 5.7 – Average CPU time of MILP for RanGen2 instances solved for deadline DTCTP.

Avg. CPU time (s)

# of Acts. 50 100 200 500 990
I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
# of Modes
3 0.14 0.07 0.06 0.03 4.81 0.73 0.17 0.07 461.92 1333.29 0.61 0.12 na 46.33 13.14 0.44 na na na 1.50
6 1.13 0.28 0.24 0.08 243.77 8.29 1.46 0.15 718.51 na 4.01 0.25 na 615.82 1407.16 3.08 na na na 12.24
9 1.27 0.29 0.43 0.11 177.43 20.89 3.39 0.21 171.36 1440.83 10.19 0.40 na 1149.87 598.14 5.22 na na na 27.10
 Table 5.8 – Percentage of RanGen2 instances optimally solved for Pareto front DTCTP.

% of Solved Instances
# of Acts. 50 100 200 500 990
I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
# of Modes
3 100 100 100 100 0 90 100 100 0 0 100 100 0 0 0 100 0 0 0 100
6 100 100 100 100 0 0 70 100 0 0 30 100 0 0 0 80 0 0 0 0
9 50 100 100 100 0 0 60 100 0 0 0 100 0 0 0 0 0 0 0 0

158
Table 5.9 – Average CPU time of MILP for RanGen2 instances solved for Pareto front DTCTP.

Avg. CPU time (s)

# of Acts. 50 100 200 500 990
I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
# of Modes
3 50.27 18.43 14.28 11.26 na 1516.12 91.81 40.55 na na 355.33 118.01 na na na 664.24 na na na 2377.88
6 1210.03 158.37 79.80 26.04 na na 949.52 89.37 na na 2175.79 246.95 na na na 2376.24 na na na na
9 2468.87 293.68 188.24 37.96 na na 1453.82 125.76 na na na 359.68 na na na na na na na na
The contents of Table 5.6 and Table 5.7 reveal MILPs successful convergence to
optimum solutions for deadline problems that include up to 100 activities by solving
an overall 99.16% of 50-activity and 100-activity instances within the imposed
runtime limit. It is displayed that on average, MILP requires 0.35 and 38.45 seconds
to locate the optimum solutions for 50-activity and 100-activity problems,
respectively. The performance of MILP is consistent for the larger network of size
200 and 500. MILP is shown to be able to solve all of the 200-activity and 500-
activity problems with I 2 values of 0.6 and 0.8 which contain virtually serial
networks. Overall, 43.33% of 200-activity and 25% of 500-activity problems with
smaller I 2 values are solved to optimality which indicates the networks with almost
parallel graphs require greater processing times. MILP is experimented to locate
the global optimum solutions in 376.50 and 426.58 seconds on average, for 200-
activity and 500-activity instances, respectively. MILP is also shown to be able to
unravel 25% of the instances with 990 activities, comprising 3 to 9 time-cost
alternatives. CPU time limit of one hour is experienced to be insufficient for the
more complex (with regard to the nature of MILP) 990-activity instances with I 2
values of 0.2 to 0.6; however, MILP is displayed to be capable of solving 3 sets of
990-activity problems, i.e., 30 instances, with I 2 values of 0.8 in an average
processing time of 13.61 seconds.

The contents of Table 5.8 and Table 5.9 shed some light on the performance of
MILP over capturing the true Pareto front of the complex RanGen2 instances. It is
revealed that MILP is virtually able to locate the true Pareto fronts for all the 50-
activity instances in an average processing time of 379.77 seconds. Solving 50% of
the most complex (with regard to the nature of MILP) set of 50-activity problems
with I 2 values of 0.2 and 9 time-cost modes, MILP obtains exact solutions for an
overall 95.83% of the 50-activity instances. MILP is shown to be able to solve all
the pseudo-serial 100-activity and 200-activity problems with I 2 values of 0.8.
MILP is displayed to be able to capture optimal frontiers for 53.33% of the 100-

159
activity problems with smaller I 2 values of 0.4 and 0.6. With an overall average
computation time of 609.56 seconds, MILP is experimented to require more than
one hour to solve 100-activity problems with I 2 values of 0.2. 43.33% of the 200-
activity problems with I 2 value of 0.6 are solved for Pareto front problem within
the enforced time limit. With an overall average processing time of 651.15 seconds,
runtime limit of one hour is experienced to be inadequate for capturing the true
Pareto fronts for pseudo-parallel 200-activity problems with smaller I 2 values of
0.2 and 0.4. The successful performance of MILP is consistent for the larger sets of
instances with 500 and 990 activities that include 3 modes with network of I 2 = 0.8.
MILP requires an overall average computation time of 1520.24 seconds to locate
the efficient frontiers for 15% of the 500-activity problems. An overall average of
2377.38 seconds is recognized to be sufficient for solution of 8.3% of the largest
set of instances with 990 activities.

In the light of the above interpretations, an overall picture of MILP’s performance

can be obtained as the nature of this method tends to be more suitable for pseudo-
serial networks with greater I 2 values that include smaller number of time-cost
alternatives. Thus, problems 21 to 30 are experimented to be the most complicated,
and the problems 91 to 100 to be the simplest problems for the proposed exact
algorithm.

In the absence of the upper-bound (Cutoff parameter) and parallel merging (Section
5.1.1.2) techniques, the proposed MILP is able to obtain optimal Pareto fronts for
207 number of instances in an overall average processing time of 656.06 seconds;
however, implementation these techniques increases the number of solved
problems by more than 19% while reducing the overall average CPU time by more
than 18%. To the best of author’s knowledge, this is the first contribution where
global optimal costs and true Pareto fronts are captured for real-life-scale instances
that are based upon the complex RanGen2 networks. With regard to the summary
of the past research given in Table 2.1, it can also be observed that no previous

160
exact method, other than Bilir’s (2015) approach, has successfully been applied to
large-scale DTCT problems. Nevertheless, MILP is proved to be capable of locating
the optima for instances with up to 990 activities.

5.2.4. Cost Minimization and Deadline DTCTPs

Computational experiments are carried out to evaluate the performance of the

proposed CS-Heuristic method for cost minimization and deadline DTCTPs using
a set of benchmark instances acquired from the literature, based on the performance
metrics presented in Section 5.2.2. The proposed optimization algorithm is coded
in C# and compiled within Visual Studio 2013 on a 64-bit platform. All of the tests
are carried out on a desktop computer with a P9X79 Chipset motherboard, 16 GB
667 MHz DDR3 random-access memory (RAM), Intel Core i7-3.40 GHz CPU, and
64-bit Windows 10 operating system. CS-Heuristic is executed solely (no other
software is ran simultaneously) on a single processor and overclocking is not
performed.

5.2.4.1. Small-Scale Benchmark Problems

The performance of the proposed CS-Heuristic for deadline DTCTP is first tested
using the small-scale benchmark instances which are commonly used in the
literature (Elbeltagi et al., 2005; Zheng et al., 2005; El-Rayes and Kandil, 2005;
Kandil and El-Rayes, 2006; Elbeltagi et al., 2007; Ng and Zhang, 2008; Xiong and
Kuang, 2008; Afshar et al., 2009; Fallah-Mehdipour et al., 2012; Sonmez and
Bettemir, 2012; Zhang and Ng, 2012; Monghasemi at al., 2015) for performance
evaluations. Readers are referred to Section 3.4.2 for details on the practiced small-
scale instances including 18a, 18b, and 18c problems. Performance of CS-Heuristic
is compared with DPSO algorithm (Section 3.3) which is shown in Section 3.4.2 to
be able to outperform any of the previous optimization methods. Snapshots of the

161
performance of the results for problems 18a, 18b, and 18c are given in Table 5.10,
Table 5.11, and Table 5.12, respectively.

Table 5.10 – Performance of CS-Heuristic for problem 18a.

Algorithm CPU Time (s) APD (%)

DPSO (Section 3.3) 0.4 0.00
CS-Heuristic (Section 5.1.3) ~0 0.00

Table 5.11 – Performance of CS-Heuristic for problem 18b.

Algorithm CPU Time (s) APD (%)

DPSO (Section 3.3) 0.4 0.00
CS-Heuristic (Section 5.1.3) ~0 0.00

Table 5.12 – Performance of CS-Heuristic for problem 18c.

Algorithm CPU Time (s) APD (%)

DPSO (Section 3.3) 0.4 0.00
CS-Heuristic (Section 5.1.3) ~0 0.00

According to the contents of the above tables, it is observed that CS-Heuristic is

also able to locate the global optimum solutions for problems 18a, 18b, and 18c;
however, by running on the same desktop computer, the processing time of this
method is revealed to be less than that of the DPSO algorithm for all the small-scale
instances.

5.2.4.2. Medium-Scale Benchmark Problems

The performance of the proposed CS-Heuristic for deadline DTCTP is also tested
using the medium-scale problem of Sonmez and Bettemir (2012). Readers are
referred to Section 3.4.3 for the explanations on the exercised medium-scale
problems including 63a and 63b problems. Results of CS-Heuristic are compared
with the DPSO algorithm (Section 3.3), detailed results of which was demonstrated
in Section 3.4.3. DPSO was conceded to be more effective and efficient than both

162
the GA and HA methods of Sonmez and Bettemir (2012). Results of CS-Heuristic
for the medium-scale instances are tabulated in Table 5.13. Unavailable values are
tabulated as ‘na’ in Table 5.13.

Table 5.13 – Performance of CS-Heuristic for problems 63a and 63b.

63a 63b
Algorithm CPU APD CPU APD
Time (s) (%) Time (s) (%)
NAA (Bettemir and Birgonul, 2017) na 0.04 na 0.07
Parallel GA (Agdas et al., 2018) 111 0.26 88.8 1.24
DPSO (Section 3.3) 1.3 0.02 1.3 0.05
CS-Heuristic (Section 5.1.3) 0.01 0.08 0.01 0.07

The comparison of CS-Heuristic with the state-of-the-art methods proves that the
proposed CS-Heuristic is among the top performing algorithms for the medium-
scale deadline DTCTP. CS-Heuristic is displayed to outperform DPSO – using the
same PC – and Parallel GA of Agdas et al. (2018) with regard to the required
processing time. Bettemir and Birgonul (2017) do not report the computation time
requirement of their Network Analysis Algorithm (NAA); though, its convergence
capabilities are observed to be virtually equal to CS-Heuristic. Besides, accuracy of
the solutions obtained by CS-Heuristic while being on par with DPSO, is shown to
be significantly better than Parallel GA of Agdas et al. (2018).

5.2.4.3. Large-Scale Benchmark Problems

The large-scale problems used for performance measurement of the proposed CS-
Heuristic include problems comprising 630, 1800, 3150, and 6300 activities. The
630-activity, 3150-activity, and 6300-activity problems are generated by copying
the 63-activity problem of Sonmez and Bettemir (2012) ten, 50, and 100 times in
serial, respectively. Readers are referred to Section 3.4.4 for the description of the
63-activity-based large-scale problems. The 1800-activity problem, on the other
hand, is generated by copying the 18-activity problem of Hegazy (1999) 100 times

163
in serial. Readers are referred to Section 4.4.2 for details on the base 18-activity
problem.

The performance of CS-Heuristic is tested against DPSO (Section 3.3) and Parallel
GA of Agdas et al. (2018). In Section 3.4.4, it is illustrated that DPSO is
significantly faster and more effective that GA and HA approaches of Sonmez and
Bettemir (2012). Performance of the results for 630-activity problems are
summarized in Table 5.14.

Table 5.14 – Performance of CS-Heuristic for problems 630a and 630b.

630a 630b
Algorithm CPU Time APD CPU Time APD
(s) (%) (s) (%)
Parallel GA (Agdas et al., 2018) 364.2 2.76 396.6 2.29
DPSO (Section 3.3) 14.6 0.33 14.6 0.34
CS-Heuristic (Section 5.1.3) 5.23 0.05 5.15 0.10

Agdas et al. (2018) experimented 630-activity problems by applying parallel

computing on a high-performance computing facility that included eight CPU
cores. Though, both DPSO and CS-Heuristic running on a single CPU are shown
to remarkably outperform Parallel GA of Agdas et al. (2018) with substantially
smaller deviation and runtime amounts. CS-Heuristic is also observed to best DPSO
with respect to both accuracy and the convergence speed.

The performance of CS-Heuristic is also investigated using the 1800-activity

problems which are used by Agdas et al. (2018). This problem is examined under
two different conditions. In problem 1800a, the indirect cost rate is $200/day, the
delay penalty is set as $20,000/day, and the incentive payment is assumed as
$1,000/day. Whereas, in problem 1800b, the indirect cost rate is assumed to be
$1,500/day. CS-Heuristic is experimented against the Parallel GA of Agdas et al.
(2018), results of which are presented in Table 5.15.

164
 Table 5.15 – Performance of CS-Heuristic for problems 1800a and 1800b.

1800a 1800b
Algorithm CPU Time APD CPU Time APD
(s) (%) (s) (%)
Parallel GA (Agdas et al., 2018) 20,952 7.05 21,024 14.72
CS-Heuristic (Section 5.1.3) 40.91 0.00 40.66 0.00

CS-Heuristic is experienced to be able to locate the global optimum solutions for

both the 1800a and 1800b problems within a processing time of less than 41
seconds. Performance of CS-Heuristic is incomparable with Parallel GA (Agdas et
al., 2018) for this large-scale problem, since, Parallel GA can find solutions with
considerably larger deviations in substantially longer computation times of more
than five hours.

The CS-Heuristic is further experimented using the 3150-activity problems of

Agdas et al. (2018). This problem is examined under two different conditions.
Indirect cost rate for problems 3150a and 3150b are set as $2,300/day and
$3,500/day, respectively. Performance of CS-Heuristic is compared with Parallel
GA of Agdas et al. (2018) and the results are given in Table 5.16.

Table 5.16 – Performance of CS-Heuristic for problems 3150a and 3150b.

3150a 3150b
Algorithm CPU Time APD CPU Time APD
(s) (%) (s) (%)
Parallel GA (Agdas et al., 2018) 32,940 6.5 33,876 4.73
CS-Heuristic (Section 5.1.3) 548.36 0.04 529.43 0.10

It is discovered that the proposed CS-Heuristic can converge to global optimum

solutions with only fractional deviations for 3150-activity problems. For this
problem, as well, CS-Heuristic is shown to greatly outperform Parallel GA of
Agdas et al. (2018) on the grounds of efficiency and effectiveness. Computational
time of Parallel GA running on eight cores of a high-performance computing
facility is measured to be nine hours more than that of CS-Heuristic.

165
Finally, performance of the proposed CS-Heuristic evaluated using the 6300-
activity problems of Agdas et al. (2018). This problem is also implemented under
two different settings. Indirect cost rate for problems 6300a and 6300b are assigned
as $2,300/day and $3,500/day, respectively. As shown in Table 5.17, performance
of CS-Heuristic is measured by directing comparisons with Parallel GA of Agdas
et al. (2018).

Table 5.17 – Performance of CS-Heuristic for problems 6300a and 6300b.

6300a 6300b
Algorithm CPU Time APD CPU Time APD
(s) (%) (s) (%)
Parallel GA (Agdas et al., 2018) 59,112 7.66 60,336 6.96
CS-Heuristic (Section 5.1.3) 4484.56 0.04 4292.81 0.10

The successful performance of CS-Heuristic is also consistent for the 6300-activity

problems. It is proved that the proposed CS-Heuristic is capable of converging to
global optimal solutions with only fractional deviations for problems 6300a and
6300b. Performance of CS-Heuristic is unmatched by Parallel GA (Agdas et al.,
2018) for the largest and the most complex problem implemented. There is
remarkable gap between the performance of the proposed CS-Heuristic and the
Parallel GA for solution of this instance. Processing time requirement of Parallel
GA running on eight cores of a high-performance computing facility is revealed to
be more than 15 hours over the CPU time requirement of the proposed CS-
Heuristic. The average percent deviation of Parallel GA is measured to be almost
191 and 70 times the amount of APDs for CS-Heuristic over problems 6300a and
6300b, respectively.

Comparative studies reveal superiority of the proposed CS-Heuristic over earlier

approaches as well as the DPSO method. It is obvious that not only the computation
time requirement of the innovative optimization model presented in this section is
remarkably less than the earlier approaches, but also it is able to locate high quality
solutions for all the practiced cost minimization/deadline problems. Owing to its

166
unprecedented efficacy and exceptional accuracy, Cost-Slope Heuristic is expected
to contribute to optimal planning of real-life-scale construction projects. To the best
of author’s knowledge, the proposed CS-Heuristic optimization model is the first
method that outperforms state-of-the art meta-heuristics and is capable of
unraveling large-scale problems comprising thousands of activities within
practically reasonable timeframes with only fractional deviations.

5.2.5. Pareto front DTCTP

Computational experiments are carried out to measure the performance of the

proposed CS-Heuristic method for Pareto front DTCTPs using a set of benchmark
and case problems acquired from the literature as well as the RanGen2 instances
introduced in Section 5.2.1, based on the performance indices presented in Section
5.2.2. The proposed optimization algorithm is coded in C# and compiled within
Visual Studio 2013 on a 64-bit platform. All of the tests are carried out on a desktop
computer with a P9X79 Chipset motherboard, 16 GB 667 MHz DDR3 random-
access memory (RAM), Intel Core i7-3.40 GHz CPU, and 64-bit Windows 10
operating system. CS-Heuristic is executed solely (no other software is ran
simultaneously) on a single processor and overclocking is not performed.

5.2.5.1. Small-Scale Benchmark Problems

The performance of the proposed CS-Heuristic for Pareto front DTCTP is first
tested using the small-scale benchmark instances that include 18 activities and up
to five time-cost modes. This problem is widely used in the literature (Elbeltagi et
al., 2005; Zheng et al., 2005; El-Rayes and Kandil, 2005; Kandil and El-Rayes,
2006; Elbeltagi et al., 2007; Ng and Zhang, 2008; Xiong and Kuang, 2008; Afshar
et al., 2009; Fallah-Mehdipour et al., 2012; Sonmez and Bettemir, 2012; Zhang and
Ng, 2012; Monghasemi at al., 2015) for performance evaluations. Readers are
referred to Section 4.4.2 for details on the practiced small-scale instances including

167
18d, 18e, 18f, and 18g problems. Performance of CS-Heuristic is compared with
PFPSO model (Section 4.3) which is shown in Section 4.4.2 to be able to surpass
any of the earlier optimization methods in the literature. Results of the performance
assessment over problems 18d, 18e, 18f, and 18g are displayed in Table 5.18.

Table 5.18 – Performance comparison of CS-Heuristic for small-scale problems.

ND pct APD APD bin CPU

Method Problem ONVG HR
(%) (%) (%) Time (s)
PFPSO 18d 39 100 0 0 1 2
(Section 4.3) 18e 44 100 0 0 1 2
18f 18 100 0 0 1 2
18g 4 100 0 0 1 2
CS-Heuristic 18d 39 87.18 0.01 0.01 1 0.03
(Section 5.1.4) 18e 43 84.09 0.01 0.01 0.99 ~0
with PM*
18g 15 83.33 0 0 1 ~0
18d 4 100 0 0 1 ~0
CS-Heuristic 18d 39 87.18 0.01 0.01 1 0.01
(Section 5.1.4) 18e 42 75 0.02 0.02 0.99 0.01
with SM**
18f 10 55.56 0 0 0.99 ~0
18g 4 100 0 0 1 ~0
*Parallel Merge
**Serial Merge

Performance of the proposed CS-Heuristic is experimented by using either of the

merging techniques of serial and parallel, independently. Compared with PFPSO,
both variants of CS-Heuristic are found to be able to locate the same number of
non-dominated solutions for problems 18d and 18g. For the rest of the small-scale
problems, the number of the obtained Pareto solutions by PFPSO are slightly more
than CS-Heuristic with parallel merge which is followed by CS-Heuristic with
serial merge. For problems 18d, 18e, and 18f, PFPSO is observed to perform
marginally better than the CS-Heuristics with regard to the performance indexes,
excluding the computation time. Results of all the methods are identical for problem
18g with respect to any of the metrics. Nonetheless, executed on the same desktop
computer, CS-Heuristics are experienced to require considerably less processing

168
time for all the 18d, 18e, 18f, and 18g problems which might justify the slight
differences in the remainder of the performance indices. CS-Heuristic with serial
merge, on the other hand, is shown to perform very close to CS-Heuristic with
parallel merge; though, according to ND pct values, parallel merge variant is

capable of capturing solutions of higher quality for problems 18e and 18f.

5.2.5.2. Medium-Scale Benchmark Problem

The performance of the proposed CS-Heuristic for Pareto front DTCTP is also
evaluated using the medium-scale problem of Kandil and El-Rayes (2006). Readers
are referred to Section 4.4.3 for the explanations on the exercised medium-scale
180-activity problem. Results of both the variants of CS-Heuristic are compared
with PFPSO method (Section 4.3) detailed results of which was illustrated in
Section 4.4.3. PFPSO was validated to be more effective and efficient than both the
GP-GA and CG-GA approaches of Kandil and El-Rayes (2006). Results of CS-
Heuristics for the medium-scale instance are tabulated in Table 5.19.

Table 5.19 – Comparison of the results for 180-activity problem.

ND pct CPU
APD APD bin
Algorithm ONVG HR Time
(%) (%) (%) (s)
PFPSO (Section 4.3) 304 34.86 0.09 0.09 0.99 21
CS-Heuristic (Section 5.1.4) with PM* 586 69.95 0.27 0.21 0.99 1.37
CS-Heuristic (Section 5.1.4) with SM** 586 69.95 0.27 0.21 0.99 1.25
*Parallel Merge
**Serial Merge

Both variants of the proposed CS-Heuristic approach are able to capture 586 non-
dominated solutions in slightly over one second with average deviations of less than
0.3%. CS-Heuristics are able to position a rather larger set of non-dominated
solutions along the efficient frontier compared with PFPSO. All the approached are
observed to achieve solutions with comparable diversities along the frontiers since

169
they share the same values for HR index. Performed on the same PC, the processing
time requirements of CS-Heuristics are validated to be remarkably less than 21
seconds of PFPSO approach. Although both the unary and binary average deviation
amount of PFPSO are smaller than CS-Heuristic, it is indicated by ND pct metric

that the number of non-dominated solutions in the final unified front is significantly
higher for CS-Heuristic. Faster convergence speed, larger ONVG, and larger
ND pct values of CS-Heuristic are believed to make up for its slightly larger

deviation amounts. Computational time of serial variant of CS-Heuristic is

experimented to be marginally shorter than the parallel variant for this instance.

5.2.5.3. Large-Scale Benchmark Problems

Experiments on large-scale benchmark instances comprise the 360-activity and

720-activity problems created by Kandil and El-Rayes (2006). Readers are referred
to Section 4.4.4 for the description of the 18-activity-based large-scale problems.
Efficacy and efficiency of the two variants of CS-Heuristic are validated in
comparison with PFPSO method (Section 4.3). PFPSO was displayed in Section
4.4.4 to substantially outperform both the GP-GA and CG-GA approaches of
Kandil and El-Rayes (2006). Performance of the results for 360-activity problem is
summarized in Table 5.20.

Table 5.20 – Comparison of the results for 360-activity problem.

ND pct CPU
Algorithm ONVG HR Time
(%) (s)
PFPSO (Section 4.3) 536 32.48 0.99 43
CS-Heuristic (Section 5.1.4) with Parallel Merge 1176 71.71 0.99 11.26
CS-Heuristic (Section 5.1.4) with Serial Merge 1176 71.71 0.99 10.04

Unary and binary APDs are not reported for this problem since the true Pareto fronts
were not available. Both the CS-Heuristics with parallel and serial merge
techniques are able to obtain 1176 non-dominated solutions in slightly over 10 and

170
11 seconds respectively. CS-Heuristics are able to capture more than two times the
amount of non-dominated solutions achieved by PFPSO. All the optimization
methods are revealed to be able to locate non-dominated solutions with
commensurate diversities along the fronts since they share the same values for HR
metric. Performed on the same desktop computer, the processing time requirements
of CS-Heuristics are experienced to be almost one fourth of the 43 seconds required
by PFPSO method. With respect to the values calculated for ND pct index, the

number of non-dominated solutions in the final unified front for CS-Heuristic is

more than two times the amount of solutions positioned by PFPSO. CS-Heuristic
with serial merge, on the other hand, is observed to require slightly less processing
time for solution of this problem in comparison with the parallel variant.

The CS-Heuristic is further experimented using the large-scale 720-activity

problems of Kandil and El-Rayes (2006). Performance of CS-Heuristics are
compared with PFPSO (Section 4.3) and the results are given in Table 5.21.
PFPSO’s exceptional convergence capabilities was proved in Section 4.4.4 which
outperformed GP-GA and CG-GA approaches of Kandil and El-Rayes (2006) by a
large margin.

Table 5.21 – Comparison of the results for 720-activity problem.

ND pct CPU
Algorithm ONVG HR Time
(%) (s)
PFPSO (Section 4.3) 1022 27.74 0.99 92
CS-Heuristic (Section 5.1.4) with Parallel Merge 2356 75.36 0.99 108.29
CS-Heuristic (Section 5.1.4) with Serial Merge 2356 75.36 0.99 95.47

Unary and binary APDs are not reported since the true Pareto fronts were not known
for this problem. The successful performances of CS-Heuristics are also consistent
for the large-scale 720-activity problems. It is displayed that the proposed CS-
Heuristics are capable of capturing 2356 non-dominated solutions along the
efficient frontier within less than two minutes. Both the variants of CS-Heuristic

171
are able to achieve more than two times the amount of solutions located by PFPSO.
While computational time requirement of CS-Heuristic with serial merge is
virtually the same as PFPSO, the parallel variant is shown to demand moderately
larger processing time on the same desktop computer. Indicated by equal HR
values, all the optimization methods are capable of locating non-dominated
solutions with comparable distribution and spread characteristics along the frontier.
In accord with the values calculated for ND pct index, the number of non-dominated

solutions in the final unified front for CS-Heuristic is slightly shy of three time the
amount of solutions obtained by PFPSO.

Comparative studies reveal that the performance of CS-Heuristic is unmatched by

any of the previous approaches including the PFPSO algorithm. Not only the
computation time requirement of the innovative multi-objective optimization model
presented in this section is substantially less than the earlier approaches, but it is
also able to produce a large number of high quality non-dominated solutions for all
the practiced Pareto front problems. Owing to its unprecedented efficacy and
exceptional accuracy, Cost-Slope Heuristic is expected to contribute to optimal
planning of realistic construction projects. To the best of author’s knowledge, the
proposed CS-Heuristic optimization model is the first method capable of tackling
large-scale problems consisting of hundreds of activities within reasonably short
timespans and practically viable deviations.

5.2.5.4. New Sets of Instances

In this section, the new sets of realistic instances are used to compare the
performance of PFPSO (Section 4.3) with CS-Heuristic (Section 5.1.4) on the basis
of the performance comparison metrics presented in Section 5.2.2. Initially, all the
instances are fitted into CS-Heuristic with parallel merge; afterwards, processing
time required for solution of each of the 600 instances are extracted to be used as a
runtime limit for PFPSO method. Hence, PFPSO is slightly modified to

172
accommodate new termination criteria. It is designed to terminate either if the
enforced runtime limit or the maximum number of iterations is reached. The values
for various performance indices for PFPSO and the two variants of CS-Heuristic,
i.e., serial and parallel, over the new complex instances are reported in the following
tables as follows. Table 5.22 and Table 5.23 compare cardinality of the solutions
using ONVG and ND pct values, respectively. Table 5.24 and Table 5.25 shed light

on the accuracy of the results based on APD and APD bin values, respectively. Table
5.26 illustrates the convergence and diversity of the Pareto fronts using HR metric.
Finally, Table 5.27 displays the convergence speed of each of the approaches by
running on the same desktop computer. These tables summarize values for every
combination of different network sizes, network complexities, and mode numbers.
The average values for every ten replicate problems per each parameter setting are
elucidated for the proposed optimization approaches. Unavailable values are
tabulated as ‘na’ in Table 5.24 and Table 5.25.

173
 Table 5.22 – Comparison of ONVG values.

ONVG

# of Acts. 50 100 200 500 990

I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

# of Modes

PFPSO 3 45 23 11 4 80 43 18 7 145 91 32 7 393 173 70 15 699 400 131 16

6 96 60 28 8 181 119 42 15 303 241 93 15 902 377 191 3 1590 1022 248 1

174
9 136 87 45 17 254 183 77 19 431 368 139 24 1273 550 308 35 2226 1540 201 1

CS-H-PM 3 80 64 40 11 190 159 80 32 313 341 127 30 1128 271 175 63 1496 779 346 126

6 164 130 77 25 324 285 91 34 555 555 98 28 1695 200 147 82 2166 587 207 127

9 200 149 93 28 385 330 98 27 618 564 59 26 1900 154 85 70 2330 519 175 144

CS-H-SM 3 80 63 43 14 190 158 89 38 313 348 138 46 1127 316 210 96 1497 842 328 137

6 164 137 85 30 324 292 113 47 555 562 158 50 1689 306 165 78 2150 752 192 73

9 200 155 96 37 385 345 135 45 618 607 123 45 1902 222 113 51 2409 628 122 71
 Table 5.23 – Comparison of ND pct values.

ND pct (%)

# of Acts. 50 100 200 500 990

I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

# of Modes

PFPSO 3 0 0.48 2.42 7.73 0 0 0.41 1.86 0 0 0.13 1.38 0 0.10 0 0.37 0 0.86 3.46 0.09

6 0.12 0.07 1.21 2.63 0 0 0 1.95 0 0 3.20 1.62 0 12.86 29.60 0 0 33.52 41.57 0

175
9 0.05 0.39 1.85 3.36 0 0 3.25 4.95 0 0.64 19.40 7.70 0 48.81 62.40 16.47 0.48 54.98 33.79 0

CS-H- 3 100 94.45 75.60 67.33 100 82.60 48.89 44.08 100 53.02 38.41 39.42 97.69 23.37 21.33 22.64 93.93 37.55 17.75 21.75
PM
6 99.88 81.47 50.64 34.29 100 64.44 31.89 33.63 100 42.78 17.09 17.48 72.23 12.28 6.73 26.31 77.88 10.75 17.60 45.76

9 100 67.82 50.89 36.35 100 59.96 21.05 20.47 100 39.18 10.63 13.50 64.31 11.58 4.54 32.57 92.29 7.63 42.58 66.26

CS-H- 3 100 89.15 76.57 84.14 100 92.54 71.94 79.15 100 72.34 79.73 89.28 98.90 83.75 77.37 83.49 96.90 59.71 75.57 78.30
SM
6 99.88 92.83 67.69 65.66 100 72.73 72.08 76.44 100 52.23 80.49 88.89 77.71 78.33 52.64 75.65 72.72 47.19 37.48 53.79

9 100 77.61 74.74 73.62 100 47.41 66.54 87.10 100 59.14 73.09 84.48 71.65 36.66 30.92 53 92.66 29.71 18.19 34.25
 Table 5.24 – Comparison of APD values.

APD (%)

# of Acts. 50 100 200 500 990

I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

# of Modes

PFPSO 3 1.79 2.08 0.51 0.10 na 1.50 0.89 0.20 na na 1.33 0.36 na na na 0.51 na na na 0.48

6 1.33 2.34 0.90 0.38 na na 1.01 0.47 na na 1.76 0.82 na na na na na na na na

176
9 1.25 2.33 0.94 0.49 na na 0.94 0.44 na na na 0.82 na na na na na na na na

CS-H-PM 3 0.17 0.04 0.05 0.03 na 0.13 0.11 0.06 na na 0.20 0.04 na na na 0.15 na na na 0.13

6 0.20 0.12 0.15 0.10 na na 0.28 0.12 na na 0.81 0.26 na na na 0.20 na na na na

9 0.26 0.24 0.23 0.10 na na 0.26 0.12 na na na 0.27 na na na na na na na na

CS-H-SM 3 0.17 0.04 0.06 0.01 na 0.12 0.07 0.04 na na 0.16 0.05 na na na 0.15 na na na 0.15

6 0.20 0.11 0.12 0.05 na na 0.25 0.08 na na 0.76 0.23 na na na 0.25 na na na na

9 0.26 0.23 0.17 0.10 na na 0.25 0.12 na na na 0.30 na na na na na na na na

 Table 5.25 – Comparison of APD bin values.

APD bin (%)

# of Acts. 50 100 200 500 990

I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

# of Modes

PFPSO 3 1.79 1.73 0.32 0.01 na 1.24 0.59 0.07 na na 0.66 0.07 na na na 0.09 na na na 0.03

6 1.33 2.18 0.48 0.09 na na 0.51 0.08 na na 0.83 0.04 na na na na na na na na

177
9 1.25 2.12 0.54 0.14 na na 0.38 0.11 na na na 0.05 na na na na na na na na

CS-H-PM 3 0.18 0.02 0.01 0.00 na 0.10 0.06 0.00 na na 0.10 0.00 na na na 0.02 na na na 0.01

6 0.21 0.11 0.07 0.02 na na 0.13 0.02 na na 0.53 0.02 na na na na na na na na

9 0.27 0.22 0.12 0.05 na na 0.14 0.06 na na na 0.03 na na na na na na na na

CS-H-SM 3 0.18 0.02 0.00 0.00 na 0.10 0.03 0.00 na na 0.09 0.00 na na na 0.02 na na na 0.01

6 0.21 0.10 0.06 0.01 na na 0.12 0.02 na na 0.47 0.02 na na na na na na na na

9 0.27 0.20 0.11 0.05 na na 0.11 0.04 na na na 0.03 na na na na na na na na

 Table 5.26 – Comparison of HR values.

HR

# of Acts. 50 100 200 500 990

I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

# of Modes

PFPSO 3 0.79 0.59 0.57 0.55 0.77 0.67 0.57 0.56 0.68 0.66 0.46 0.47 0.81 0.49 0.44 0.49 0.74 0.66 0.73 0.38

6 0.86 0.67 0.64 0.57 0.84 0.75 0.59 0.58 0.71 0.79 0.55 0.36 0.88 0.72 0.91 0.00 0.80 0.93 0.74 0.00

178
9 0.88 0.71 0.67 0.63 0.86 0.80 0.68 0.54 0.74 0.85 0.74 0.43 0.91 0.92 0.96 0.49 0.84 0.96 0.48 0.00

CS-H-PM 3 0.98 0.99 0.96 0.91 1.00 0.97 0.93 0.87 1.00 0.99 0.86 0.79 1.00 0.96 0.91 0.62 1.00 0.98 0.88 0.63

6 0.98 0.98 0.90 0.86 1.00 0.99 0.81 0.72 1.00 0.99 0.71 0.49 1.00 0.79 0.72 0.83 1.00 0.74 0.66 0.96

9 0.99 0.96 0.84 0.73 1.00 0.99 0.74 0.57 1.00 0.98 0.63 0.41 1.00 0.58 0.49 0.88 1.00 0.62 0.72 0.99

CS-H-SM 3 0.98 0.98 0.95 0.95 1.00 0.97 0.95 0.93 1.00 1.00 0.89 0.88 1.00 1.00 0.99 0.76 1.00 0.98 0.98 0.69

6 0.98 0.98 0.93 0.88 1.00 1.00 0.88 0.82 1.00 0.99 0.90 0.71 1.00 0.97 0.87 0.88 1.00 0.88 0.72 0.99

9 0.99 0.97 0.86 0.83 1.00 0.99 0.86 0.72 1.00 0.99 0.94 0.58 1.00 0.82 0.65 0.96 1.00 0.77 0.72 0.98
 Table 5.27 – Comparison of CPU times.

CPU Time (s)

# of Acts. 50 100 200 500 990

I2 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8

# of Modes

PFPSO 3 0.06 0.04 0.02 0.01 0.45 0.35 0.16 0.07 2.18 2.99 0.97 0.22 59.68 11.65 8.07 2.87 222.09 131.69 55.08 18.41

6 0.15 0.09 0.06 0.02 0.89 0.70 0.22 0.11 4.22 5.14 0.86 0.36 110.12 10.40 7.32 4.35 402.96 110.07 35.74 21.96

179
9 0.21 0.14 0.10 0.06 1.19 0.94 0.36 0.24 5.32 5.61 1.14 0.94 134.76 11.64 8.33 7.59 482.16 111.85 45.65 39.67

CS-H- 3 0.06 0.04 0.02 0.01 0.45 0.35 0.16 0.07 2.18 2.99 0.97 0.22 59.68 11.65 8.07 2.87 222.09 131.69 55.08 18.41
PM
6 0.15 0.09 0.06 0.02 0.89 0.70 0.22 0.11 4.22 5.14 0.86 0.36 110.12 10.40 7.32 4.35 402.96 110.07 35.74 21.96

9 0.21 0.14 0.10 0.06 1.19 0.94 0.36 0.24 5.32 5.61 1.14 0.94 134.76 11.64 8.33 7.59 482.16 111.85 45.65 39.67

CS-H- 3 0.06 0.03 0.03 0.32 0.46 0.34 0.16 1.30 2.17 2.93 1.20 10.09 58.79 11.85 8.54 107.61 218.20 131.32 48.02 838.27
SM
6 0.14 0.11 0.25 8.40 0.88 0.72 1.20 31.70 4.23 5.21 15.23 211.04 108.31 26.32 50.27 1847.83 392.90 141.49 146.22 8924.87

9 0.20 0.25 0.97 18.61 1.18 1.06 5.81 77.00 5.32 7.24 58.34 427.82 133.36 76.16 182.71 3361.96 494.23 181.61 520.39 12973.30
The interpretation of the results tabulated in Table 5.22 to Table 5.27 can be
summarized as follows. It is revealed that for the complete set of 50-activity
problems, while being executed within the same timeframes, parallel variant of CS-
Heuristic is able to capture 88.8% more non-dominated solutions than PFPSO. Per
120 problems of this size, parallel variant of CS-Heuristic with an average CPU
time of 0.08 seconds, operates at least two seconds faster than the serial variant, on
average; though, serial variant is observed to be able to achieve 4% more non-
dominated solutions. On average, PFPSO, parallel, and serial variants are
discovered to account for 1.7%, 71.5%, and 83.5% of the non-dominated solutions
over the final unified Pareto front, respectively. Per 120 number of 50-activity
problems, parallel and serial CS-Heuristics with comparable average deviations of
0.12% and 0.14%, are shown to be more accurate than PFPSO with an average APD
of 1.2%. This remark is also true for the binary APD, since, PFPSO happens to have
an average deviation of 1% from the optimal costs, for the mutually located duration
amounts. This index for both CS-Heuristics is evaluated to be 0.1%. Solutions
obtained by parallel and serial CS-Heuristics tend to be more accurate, well-
distributed and widely spread with average HR values of 0.92 and 0.94; whereas
this value for PFPSO is measured as 0.67.

Another interesting remark to note is that, for the 50-activity problems there is an
increasing pattern in ONVG, APD, and APD bin values of all the solution
procedures with increasing mode numbers and decreasing I 2 parameters. This
observation is also valid for the CPU times of the approaches, with the exception
of serial variant of CS-Heuristic requiring more processing times for problems with
larger I 2 values. For CS-Heuristics and PFPSO, HR values are determined to be
larger for problems with smaller I 2 s. Furthermore, HR values for PFPSO and CS-
Heuristics are discovered to vary conversely for problems with I 2 values of 0.6 and
0.8. For these problems, HR of CS-Heuristics are observed to be inversely
proportional to the number of modes, while, HR for PFPSO is directly related to
the mode numbers.

180
Considering the whole set of 100-activity problems, parallel variant of CS-Heuristic
is able to locate 96% more non-dominated solutions than PFPSO within the same
execution time. Per 120 number of 100-activity problems, parallel variant of CS-
Heuristic with an average CPU time of 0.47 seconds, operates at least nine seconds
faster than the serial variant, on average; nevertheless, serial variant is observed to
be able to obtain 6% more non-dominated solutions. On average, PFPSO, parallel,
and serial variants are found out to account for 1%, 58.9%, and 80.5% of the non-
dominated solutions located over the final unified Pareto front, respectively. Per
120 problems of this size, parallel and serial CS-Heuristics with comparable
average deviations of 0.15% and 0.13%, are discovered to contribute to higher
levels of exactness compared to PFPSO with an average APD of 0.77%. Regarding
the average binary APD values, the above statement is also valid for the average
deviations from the exact costs of same-duration solutions located by different
methods. PFPSO happens to have an average APD bin of 0.42%, whereas, this
metric for parallel and serial CS-Heuristics are measured as 0.07% and 0.06%,
respectively. Solutions obtained by CS-Heuristics – specifically the serial variant –
are observed to be more accurate, well-distributed and widely spread, with higher
average HR values. HR for PFPSO is evaluated to be 0.68, while, this value for
parallel and serial CS-Heuristics are determined to be 0.88 and 0.92, respectively.

Moreover, an alternate conclusion can also be drawn for the 100-activity problems
that there exists an increasing pattern in ONVG, APD, and APD bin values of all the
practiced methods with increasing mode numbers and decreasing I 2 rate. This
observation is also valid for the computational times of the models, with the
exception of serial variant of CS-Heuristic demanding significantly more CPU
times for instances with larger I 2 values. For PFPSO, HR values are discovered to
be inversely proportional to I 2 rate and directly proportional to the number of
modes. For CS-Heuristics, however, HR values are determined to be inversely
proportional to both I 2 parameter and the number of alternatives, with the

181
exception of problems with I 2 values of 0.2, all of which have HR values equal to
1.

It is identified that for the entire set of 200-activity problems, parallel variant of CS-
Heuristic is capable of capturing 75.4% more non-dominated solutions than PFPSO
by running within the same CPU time. For 120 problems with 200 activities, serial
variant of CS-Heuristic is found out to require significantly longer runtimes.
Compared to overall average of 2.5 seconds for parallel variant, serial variant is
experimented to require a further one minute to locate the Pareto fronts;
nonetheless, serial variant is revealed to be able to locate 7% more non-dominated
solutions. On average, PFPSO, parallel, and serial variants are discovered to
account for 2.8%, 47.6%, and 81.6% of the non-dominated solutions over the final
unified Pareto front, respectively. Parallel and serial CS-Heuristics with literally the
same average deviations of 0.31% and 0.3%, are conceded to be more accurate than
PFPSO with an average APD of 1%. The same scheme is also confirmed for the
binary APD, since, PFPSO happens to have an average APD bin of 0.33%. This
index for parallel and serial CS-Heuristics are assessed to be 0.13% and 0.12%,
respectively. Solutions obtained by CS-Heuristics – especially the serial variant –
are observed to be more accurate, well-distributed and widely spread, with greater
average HR values of 0.82 and 0.9; however, HR for PFPSO is assessed to be 0.62.

Furthermore, another noteworthy conclusion can also be pointed out for the 200-
activity problems that there exists an increasing pattern in ONVG, APD, and
APD bin values of all the exercised methods with increasing mode numbers and
decreasing I 2 parameter. This remark is also true for the processing times of the
methods, with the exception of serial variant of CS-Heuristic which operates within
considerably larger CPU times for instances with larger I 2 values. The average HR
values for all the approaches over the problems with I 2 values of 0.8 are observed
to be inversely proportional to the number of modes. This is also true for parallel

182
variant of CS-Heuristic over the problems with I 2 values of 0.6. For PFPSO, HR
values are discovered to be inversely proportional to I 2 rate and directly
proportional to the number of modes over the problems with I 2 s of 0.2 to 0.6. For
CS-Heuristics, except for problems with I 2 of 0.6, HR values are determined to be
either one or very close to one with no obvious trend in between.

For the complete set of 500-activity problems, parallel variant of CS-Heuristic is

able to locate 39% more non-dominated solutions than PFPSO within the same time
interval. Per 120 number of 500-activity problems, parallel variant of CS-Heuristic
with an average CPU time of 31.39 seconds, runs within only 6.3% of the average
computation time of serial variant; though, serial variant is observed to be able to
capture 5% more non-dominated solutions. On average, PFPSO, parallel, and serial
variants are discovered to provide 14.2%, 32.9%, and 68.3% of the non-dominated
solutions positioned along the final unified Pareto front, respectively. Per 120
problems with 500 activities, parallel and serial CS-Heuristics with commensurate
average deviations of 0.17% and 0.2%, are discovered to contribute to solutions of
higher quality compared to PFPSO with an average APD of 0.51%. Unlike 50, 100,
and 200 activity problems, average APD value for parallel variant is less than the
serial CS-Heuristic for 500-activity instances. This trend is also confirmed for the
binary APD, since, PFPSO happens to have an average deviation of 0.09% from
the optimal costs, for the mutually located duration amounts. This index for both
CS-Heuristics is calculated to be 0.02%. Solutions obtained by CS-Heuristics –
particularly the serial variant – are observed to be more accurate, well-distributed
and widely spread, with greater average HR values of 0.81 and 0.9; however, HR
for PFPSO is measured to be 0.66.

Additionally, another remark can also be addressed for the 500-activity problems
that there exists an increasing scheme for ONVG values of all the practiced methods
with decreasing I 2 rate. For PFPSO, ONVG is observed to increase with the growth
of mode numbers. For CS-Heuristics, however, the number of located non-

183
dominated solutions are discovered to decrease with the growth of the number of
modes, with the exception of problems with I 2 s of 0.2. Since no 500-activity
problems with I 2 values of 0.2 to 0.6 are solved to optimality, no clear pattern is
captured for APDs and APD bin s of these problems. However, it is observed that
deviation amounts of CS-Heuristics are directly proportionate to the number of
modes. Computational times of the models are addressed to have an increasing
pattern with increasing mode numbers and decreasing I 2 parameter, with the
exception of serial variant of CS-Heuristic for problems with I 2 rates of 0.6 and
0.8. Serial variant is experienced to require substantially greater runtimes for
instances with larger I 2 values. While no clear scheme is detected for HR values
of PFPSO, CS-Heuristics are observed to be inversely proportional to I 2 rate and
the number of modes, except for the problems with I 2 values of 0.8 for which HRs
increase with the growth of mode numbers. Besides, for both the CS-Heuristics,
problems with I 2 values of 0.2 are shown to have HR values of one.

Lastly, it is acknowledged that for the whole set of 990-activity problems, parallel
variant of CS-Heuristic is capable of obtaining 11% more non-dominated solutions
than PFPSO by running within the same CPU time. For 120 problems including
990 activities, parallel variant of CS-Heuristic with an average processing time of
139.77 seconds, operates within only 6.7% of the average computation time of the
serial variant. Despite its significantly larger computational burden, serial variant
captures only 2% more non-dominated solutions than the parallel variant. On
average, PFPSO, parallel, and serial variants are determined to deliver 14%, 44.3%,
and 58% of the non-dominated solutions located along the final unified Pareto front,
respectively. For 990-activity problems, parallel and serial CS-Heuristics with close
average deviations of 0.13% and 0.15%, are noted to be more accurate than PFPSO
with an average APD of 0.48%. Unlike 50, 100, and 200 activity problems, average
APD value for parallel variant is smaller than the serial CS-Heuristic for 990-
activity instances. This remark is also true for the binary APD, since, PFPSO

184
happens to have an average deviation of 0.03% from the exact costs, for the jointly
located same-duration solutions. This index for both CS-Heuristics is measured to
be 0.01%, on average. Solutions obtained by CS-Heuristics are validated to be more
accurate, well-distributed and widely spread, with larger average HR values of 0.85
and 0.89. HR metric for PFPSO, on the other hand, is measured to be 0.6.

In addition, it can also be interpreted from the values of the performance metrics
for the 500-activity problems that there exists an increasing scheme for ONVG
values of all the practiced approaches with decreasing I 2 rate. For PFPSO, ONVG
is discovered to increase with the growth of mode numbers. This is specifically true
for problems with I 2 values of 0.2 and 0.6. For CS-Heuristics, however, the number
of obtained non-dominated solutions are revealed to decrease with the growth of
the number of modes, with the exception of problems with I 2 s equal to 0.2.
Inasmuch as no 990-activity problems with I 2 values of 0.2 to 0.6 are solved to
optimality, no distinct pattern is captured for APDs and APD bin s of these instances.
Computational times of the models are determined to have an increasing pattern
with increasing mode numbers. For problems with I 2 values of 0.2, processing
time of all the models are directly proportionate to the number of modes. However,
for the remaining I 2 values, there exists an inverse relation between the average
CPU time and the number of modes, with the exception of serial variant of CS-
Heuristic. Serial variant is experimented to contribute to remarkably larger
computation times. While no clear pattern is identified for HR values, this index is
noticed to be significantly lower for problems with I 2 values of 0.8 for PFPSO
method. In addition, for problems with I 2 values of 0.2, CS-Heuristics are shown
to be able to capture frontiers with HR values equal to one.

In the light of the above interpretations, an overall picture of PFPSO’s and CS-
Heuristic’s performances can be obtained as follows. For the entire collection of the
RanGen2 instances, both variants of CS-Heuristic are shown to be able to achieve

185
higher number of Pareto solutions, serial variant even more so than the others. It is
also discovered that the number of non-dominated solutions located along the final
unified frontier are significantly larger for CS-Heuristics. In fact, PFPSO only
accounted for less than 7% of these solutions, while serial variant positioned 23%
more than the parallel CS-Heuristic. Despite high levels of accuracy for all the
experimented methods, CS-Heuristics are confirmed to be more successful in
converging to the true Pareto fronts with fractional deviations. Average deviations
for parallel and serial CS-Heuristic are shown to be literally the same; whereas,
PFPSO’s average deviation value is calculated to be five to six times the amount
for CS-Heuristics. Well-distributed and widely spread fronts are only obtained by
means of CS-Heuristic since the area of the solution space covered by PFPSO,
poorly represents the Hypervolume of the true/best Pareto front.

Although the values for ONVG, ND pct , and HR performance metrics are

sporadically higher for the serial variant of CS-Heuristic, it is experienced to

contribute to remarkably higher computational costs. In fact, serial variant is
conceded to operate within an average of 15 times the CPU time of parallel CS-
Heuristic. The reason behind this major increase in CPU time is due to the
enormously large number of serially merged activities for the problems with
pseudo-serial networks. It is observed that for higher I 2 values, viz., 0.6 and 0.8,
the number of serial activities satisfying the conditions for the serial merge can
reach up to almost 50% of the original network size. For instance, some 990-activity
problems are observed to be reduced by merging 487 number of serial activities. As
is clear, this technique is not suitable for networks with serial graphs and that the
marginal improvements in the obtained solutions might not justify the additional
processes and the extra efforts involved in merging the activities serially.

Inclusion of parallel merge, on the other hand, is experimented to contribute to

lower computational times, although by meager amounts, with the exception of
problems with 500 and 990 activities and larger I 2 parameters. Parallel merge is

186
discovered to slightly increase the computation time for 500-activity problems, and
even more so for 990-activity instances, with I 2 values of 0.8. The reason behind
this slight growth in computation time is due to surging number of parallel reducible
activities. As mentioned in Section 5.1.1.1, the first step for parallel merging is
defined as to search for activities including only two successors. For the problems
generated by means of RanGen2, the pseudo-parallel networks with smaller I 2
values comprise a multitude of successors (generally more than two); however, as
the value of I 2 increases, the number of successors for the activities of pseudo-
serial networks reduce to either one or two. That is the reason why a large
percentage of the activities on the pseudo-serial networks satisfy the conditions for
parallel merging. Meanwhile, it is also observed that both the serial and parallel
merging techniques practically increase ONVG, ND pct , and HR values for

problems with 500 and 990 activities. Furthermore, for the new sets of RanGen2
instances, implementation of Partial-CPM technique (Section 5.1.2) is experienced
to reduce the overall average computation time of CS-Heuristic by 5.2%.

With respect to the number of the obtained non-dominated solutions and the
diversity of the captured Pareto fronts, the nature of PFPSO and CS-Heuristic
approaches, in contrast to MILP, appear to be more suitable for pseudo-parallel
networks with smaller I 2 values that include greater number of time-cost
alternatives. However, for this setting of problems the deviation and the
computation time amounts are observed to increase, although by meager amounts.
Comparative studies reveal that the performance of CS-Heuristic is unmatched by
any of the previous approaches including the PFPSO algorithm. Owing to the
unprecedented accuracy and diversity of the obtained Pareto fronts, as well as its
exceptional efficiency, Cost-Slope Heuristic is expected to contribute to optimal
planning of realistic large-scale construction projects. To the best of author’s
knowledge, the proposed CS-Heuristic optimization model is the first method that
outperforms highly capable meta-heuristics and is able to tackle large-scale
problems consisting of hundreds of activities that are based upon the complex

187
RanGen2 networks, within reasonably short extents of time and practically viable
deviations.

5.2.5.5. Case-Problems

Robustness of the proposed PFPSO and CS-Heuristics for Pareto front DTCTP are
also validated using two real construction projects. The proposed solution
procedures are tested on the same desktop computer. The proposed approaches are
first experimented using a warehouse construction project which is obtained from
Chen and Weng (2009). This project consists of 37 activities with up to two time-
cost alternatives. The indirect cost rate is $600/day, the delay penalty is set as
$500/day, and the completion deadline is assumed as 240 days. The exact results of
MILP (Section 5.2.3) and the approximations of CS-Heuristics (Section 5.1.4) and
PFPSO (Section 4.3) are presented in Table 5.28.

Table 5.28 – Comparison of the results for the first case problem.

ND pct CPU
APD APD bin
Algorithm ONVG HR Time
(%) (%) (%) (s)
MILP (Section 5.2.3) 1 0.06
PFPSO (Section 4.3) 1 100 0 0 1 0.02
CS-Heuristic (Section 5.1.4) with PM* 1 100 0 0 1 0.02
CS-Heuristic (Section 5.1.4) with SM** 1 100 0 0 1 0.02
*Parallel Merge
**Serial Merge

As shown in Table 5.28, all the proposed approaches are capable of capturing the
true Pareto front of this real project which includes a single solution with a duration
of 174 days and an overall cost of $253,400. The solution times of parallel and serial
CS-Heuristics and PFPSO are discovered to be the same.
The second real construction project used for experimentation of the proposed
approaches include a process plant project which is acquired from Abbasi-Iranagh
(2015). This project consists of 519 activities with up to four time-cost alternatives.

188
The indirect cost rate is $10,000/day, the delay penalty is assumed as $20,000/day,
and the completion deadline is defined to be 540 days. Performance of results for
MILP (Section 5.2.3), CS-Heuristics (Section 5.1.4), and PFPSO (Section 4.3) are
summarized in Table 5.29.

Table 5.29 – Comparison of the results for the second case problem.

ND pct CPU
APD APD bin
Algorithm ONVG HR Time
(%) (%) (%) (s)
MILP (Section 5.2.3) 1 1.10
PFPSO (Section 4.3) 2 0 1.09 1.09 0.34 0.07
CS-Heuristic (Section 5.1.4) with PM* 1 100 0.43 0.43 0.72 0.07
CS-Heuristic (Section 5.1.4) with SM** 1 0 0.49 0.49 0.69 0.13
*Parallel Merge
**Serial Merge

As shown in Table 5.29, MILP is able to capture the true Pareto front for the second
real project in 1.1 seconds. The optimal Pareto front for this project comprise a
unique non-dominated solution with a duration of 626 days and an overall cost of
$12,620,417. Being executed within the same computational time, PFPSO is
discovered to be able to locate two non-dominated solutions compared to a single
non-dominated solution of all the other approaches. PFPSO, despite contributing to
a greater ONVG value, is observed to provide solutions with slightly higher
deviations. The area of the solution space covered by PFPSO is also observed to
poorly represent the Hypervolume of the true Pareto front. Parallel CS-Heuristic is
shown to be able to provide the best approximate Pareto front for this project with
the greatest ND pct , smaller unary and binary deviation, and larger HR by running

within a shorter timeframe.

Based on the extremely successful performance of the proposed algorithms for the
real projects, implications can be drawn out as the realistic projects tend to be
relatively simpler than the instances generated in Section 5.2.1. The deviation and
the runtime amounts are conceded to be small even for a large-scale 519-activity

189
construction project. According to the results obtained for the two real construction
projects, practicability and real-life applicability of the proposed approaches are
firmly validated. The proposed optimization models enjoy the fastness and are able
to capture optimal/near optimal sets of non-dominated solutions for the multi-
objective DTCTP. The author proposes to employ the proposed CS-Heuristic and
PFPSO methods in place of the MILP when the optimality is not regarded as the
most crucial concern.

190
 CHAPTER 6

INTEGRATION OF THE PROPOSED METHODS INTO

MICROSOFT PROJECT

Despite the fact that any scientific decision support tool would have a pivotal role
in the decision-making process, none of the commercial scheduling software
packages (e.g., Microsoft Project, Primavera) include tools or modules for time-
cost trade-off analyses of the scheduling problems. To the respect of this, this
chapter is dedicated to the development of a tool which bridges the gap between the
scheduling software and the DTCTP optimization techniques. Therefore, the
proposed discrete time-cost trade-off problem optimization algorithms are
integrated into the Microsoft Project 2013 which is a widely used commercial
planning software in the construction industry. Integration is facilitated by means
of an add-in which is capable of solving two variants of DTCTP, namely, cost
minimization/deadline and Pareto front problems. The integrated modules include
the proposed CS-Heuristics (Section 5.1.3 and Section 5.1.4), DPSO (Section 3.3),
and PFPSO (Section 4.3) approaches which are integrated by means of an add-in
implemented in C# programming language using Microsoft Visual Studio 2013. By
means of the built add-in, users will readily be able to visualize the optimized
schedules for their projects.

After installation of the created add-in, a ribbon named “TCTP” will appear among
the existing tabs of the Microsoft Project. Selecting this new tab will direct users to
a menu which includes two groups of “Heuristic” and “Meta-Heuristic”. Either of
the mentioned groups comprise two buttons of “Optimize” and “Pareto front”.

191
As the names imply, CS-Heuristic methods are classified under “Heuristic”, while
DPSO and PFPSO are placed under “Meta-Heuristic” group. The “Optimize”
button under “Heuristic” group invokes CS-Heuristic for deadline DTCTP (Section
5.1.3) while “Pareto front” button recalls the CS-Heuristic for Pareto front DTCTP
(Section 5.1.4). Likewise, the “Optimize” button under “Meta-Heuristic” group
invokes DPSO (Section 3.3) while “Pareto front” button activates the PFPSO
(Section 4.3) module. As displayed in Figure 6.1, a “Quick-Solve” option is also
included under the “Meta-Heuristic” group on the main panel. When checked, this
option avails PFPSO module to run up to five times faster by adjusting the
parameters of this module. The user interface (UI) of the developed add-in is
illustrated in Figure 6.1.

Figure 6.1 – User Interface of the Microsoft Project Add-in.

18a problem which is introduced in Section 3.4.2 is implemented to exemplify the

application of the designed TCTP add-in. Precedence relationships are defined for

192
the project activities through the standard procedure. As shown in Figure 6.2, time-
cost alternatives of the activities are entered into the corresponding fields as
follows. For each activity, “Duration 1” and “Cost 1” are filled in with the time and
cost of the normal mode, respectively. That is, modes with the longest durations
and the least direct costs are entered first. The remaining time-cost alternatives are
entered into the succeeding cells (i.e., “Duration 2”/”Cost 2” to “Duration
10”/”Cost 10”) in ascending order with regard to their direct cost figures. By
default, Microsoft Project holds “0 days” and “$0.00” for any unfilled entries of
time and cost, respectively. Hence, for the activities with fewer options than the
maximum supported number of modes, i.e., ten, the extra fields are left unmodified.

Figure 6.2 – Defining time-cost alternatives for Microsoft Project Add-in.

As shown in Figure 6.3, by clicking on any of the buttons, “Project Details” window
pops-up which prompts users to enter the rate of the indirect cost ($/day), project
deadline (days), and the amount of delay penalty ($/day).

193
 Figure 6.3 – Project Details window of Microsoft Project Add-in.

In case the cost minimization/deadline problem is practiced by means of either CS-

Heuristic or DPSO modules, TCTP add-in, after acquiring the project information
proceeds to a new window (Figure 6.4) which displays the unique optimal solution
for the single-objective time-cost trade-off problem.

Figure 6.4 – Optimal solution window of Microsoft Project Add-in.

By pressing the “View Schedule” button on “Optimum Solution” window, users

are directed to the optimal solution which demonstrates the optimal selection of the
time-cost options that yields the optimum schedule. Information on the generated
schedule, including early dates, late dates, and floats, can be obtained from
Microsoft Project which also provides the visual view of the activities in the form
of a Gantt chart (Figure 6.5).

194
 Figure 6.5 – Optimal schedule generated by Microsoft Project Add-in.

In case the Pareto front problem is practiced for the active project by means of either
CS-Heuristic or PFPSO modules, TCTP add-in, after acquiring the project
information proceeds to a new window which is illustrated in Figure 6.6. The
“Pareto Front Solutions” window provides the sequence of the non-dominated
solutions which are listed in ascending order with regard to duration amounts.

195
 Figure 6.6 – Pareto front solutions window of Microsoft Project Add-in.

The “Pareto Front Solutions” window prompts users to select the desired non-
dominated solution(s) from the schedules listed (Figure 6.7).

Figure 6.7 – Selection of a non-dominated solution achieved by Microsoft Project Add-in.

By pressing the “View Schedule” button for each of the selected solutions listed on
“Pareto Front Solutions” window, users are directed to each non-dominated
solution individually which reveal the arrangements of time-cost alternatives for the
desired schedules. Details on the generated schedule, including early dates, late

196
dates, and floats, can be attained from Microsoft Project which also presents a visual
view of the project activities in the form of a Gantt chart (Figure 6.8).

Figure 6.8 – Schedule for the selected Pareto solution generated by Microsoft Project Add-in.

TCTP add-in is acknowledged to enhance the practicability of the proposed DTCTP

optimization algorithms. Benefiting from the presented add-in, users of Microsoft
Project will readily be able to visualize the optimized schedules for the practiced
projects; hence, the proposed methods are supposed to be more readily accepted
and used by the parties to construction projects. By means of the developed add-in,
the new optimization models are envisioned to be applicable in real projects and to
suit the actual practices of construction managers. It is expected that these
approaches might prove to be an efficient and effective base for exerting this highly
challenging problem.

197
198
 CHAPTER 7

CONCLUSIONS

The significance of the counteracting aspects of time and cost for construction
projects is highly emphasized since project success is chiefly related to these
factors. A key process for efficacious realization of the anticipated outcomes of
time and cost is acknowledged to be the preparation of exhaustive plans and
impeccable schedules. It is declared that construction projects are prone to major
financial losses short of optimal schedules. Of the schedule optimization
methodologies, exertion of time-cost trade-off problem is reckoned to play a crucial
role in securing the pre-specified objectives of time and cost. It is a problem solving
and decision-making science which provides the management with a quantitative
basis for decisions on selection of the optimal time-cost alternatives. Despite its
widely accepted practical significance, no major real-life applicable approach is
discovered within the construction management literature. The author criticizes a
large body of the existing research due to the insufficient details on test
configurations as well as the inadequate size of the practiced problems. The
literature on large-scale time-cost trade-off problem is discovered to be virtually
void, with only a few studies using large-scale problems that are generated by
cloning simple instances. In spite of practicability of discrete multi-objective
variant of the time-cost trade-off problem, no real-life applicable contribution is
observed within the earlier research. A few methods that are implemented for real-
life-size large-scale problems, are conceded to lead to unrealistically significant
computational efforts. Though, for any solution method to be practically viable,
accuracy needs to be accompanied with the efficiency. Notwithstanding the fact

199
that achieving an adequate schedule boils down to utilization of scientific decision
support tools, no commercial scheduling software provides tools for time-cost
trade-off analyses. Accordingly, in order to bridge the gap between the theoretical
and practical relevance in time-cost trade-off problem, new exact, heuristic, and
meta-heuristic approaches with sound convergence capabilities are proposed, some
of which are integrated to a popular scheduling software package within the context
of this thesis.

Apart from the fact that the existing approaches have seldom been applied for
solution of large-scale DTCTPs, it is interpreted that the dearth of real-life-scale
problems could possibly be another major reason for the lack of studies on realistic
problems. Despite the fact that some studies have included problems including up
to 720, 2000, and 6300 activities, all of the employed large-scale problems are
generated using small-scale base networks and are generated by copying the core
problem in serial several times; hence, these problems are anticipated to have
limitations in reflecting the complexity of the real-life construction projects. In
order to have a better understanding of the behavior and capabilities of the proposed
optimization models, in addition to the existing benchmark and case problem, new
sets of multi-mode large-scale DTCT problems including up to 990 activities have
been generated by means of RanGen2 random network generator. The
systematically generated large-scale instances comprise complex networks and are
treated with realistic sets of time-cost alternatives. In order to compare methods
rigorously and to quantitatively measure performance of different approaches over
the benchmark and the RanGen2 instances, performance metrics are employed. The
incorporated performance comparison indices are designed to measure cardinality,
accuracy, diversity, and efficiency of the optimization models. Accuracy-based
performance indices necessitate acquisition of exact optimal solutions; to this end,
an exact optimization method is developed.

200
The proposed exact method is based on Mixed-Integer Linear Programming which
engages Gurobi solver. Different variants of this method are introduced for two
paradigms of cost minimization/deadline and Pareto front discrete time-cost trade-
off problems. Since, exact procedures are the only methods guaranteeing optimality
of the solutions, the proposed exact methods are mainly used for validation of the
performance of the developed heuristic and meta-heuristic approaches. The
proposed exact methods are equipped with a new merging technique which
exponentially decreases the scale of the practiced problems. The multi-objective
variant of the proposed exact method is also equipped with an efficient upper-bound
calculator designed to reduce the size of the solution space. In the light of the
implemented techniques, computation time of the exact methods are significantly
reduced. The presented single-objective exact method is shown to be able to
successfully converge to optimal solutions for 100% of 50-activity, 100% of 100-
activity, 71.66% of 200-activity, 49.16% of 500-activity, and 25% of 990-activity
deadline DTCTPs within the enforced runtime limit of one hour. Similarly, the
multi-objective exact method is experimented to effectively capture true Pareto
fronts for 95.83% of 50-activity, 51.66% of 100-activity, 35.83% of 200-activity,
15% of 500-activity, and 8.3% of 990-activity Pareto front DTCTPs within the
imposed CPU time limit of one hour. It is interpreted that the natures of these
methods tend to be more suitable for pseudo-serial networks that include smaller
number of time-cost alternatives. Implementation of the upper-bound and the
merging techniques are confirmed to increases the number of solved problems by
more than 19% and to reduce the overall average CPU time by more than 18%. To
the best of author’s knowledge, this is the first contribution where global optimal
costs and true Pareto fronts are captured for real-life-scale instances that are based
upon the complex RanGen2 networks.

It is alleged that the existing literature is not rich with particle swarm optimization
exemplars with the capacity to tackle realistic large-scale DTCT problems. To the
respect of this, different PSO algorithms are proposed for two extensions of cost

201
minimization/deadline and Pareto front DTCT problems. Both the introduced PSOs
are equipped with unique semi-deterministic initialization techniques and use new
principles for presentation and position-updating of the particles. The discrete PSO
which is designed for the single-objective DTCTP, is complemented with the
modified-SAM heuristic; whereas, the multi-objective Pareto front PSO is
enhanced using the simplified Heuristic. The trajectories and velocities of the
presented PSO-based approaches are defined as probabilities.

The solid convergence capabilities of DPSO is first validated for solution of small,
medium, and large-scale benchmark problems as well as new sets of instances
generated by means of ProGen/Max random instance generator. Later, its
performance is measured against the proposed Cost-Slope Heuristic. The
comparison of DPSO with the state-of-the-art methods proved that DPSO is among
the best, if not the best, meta-heuristic approach for single-objective DTCTP with
respect to both solution quality and computation time. DPSO is shown to be able to
produce good feasible solutions in acceptable timeframes even for the complex
ProGen/Max instances. An average deviation of 0.21% from the optima of 500-
activity problems is considered to be practically reasonable which establishes
DPSO as an effective and robust alternative for real-world applications. To the best
of author’s knowledge, the proposed DPSO is one of the first methods capable of
obtaining high quality solutions for the large-scale single-objective DTCTPs within
seconds.

The sound convergence capabilities of PFPSO is first illustrated for solutions of

benchmark problems attained from the literature. Later, its performance is
compared to the proposed Cost-Slope Heuristic using the new sets of RanGen2
instances as well as case problems. The computational tests involving benchmark
problems revealed that the proposed PFPSO can provide significantly larger
number of non-dominated solutions for small, medium, and large-scale problems,
and remarkably outperformed the well-developed methods. The results revealed

202
that the computation time requirement of PFPSO is considerably less than that of
the existing methods. It is demonstrated that PFPSO is capable of locating high
quality non-dominated solutions which are either optimal or very close to the
optimal costs. To the best of author’s knowledge, this is one of the first
contributions where a meta-heuristic algorithm is able to adequately solve real-life-
scale multi-objective DTCTPs within seconds, a performance which is unmatched
by the previous meta-heuristic methods.

It is broadly acknowledged that evolutionary algorithms are very sensitive to

configuration of their parameters. In real-life situations, the experimental process
for configuration of parameters for best values may become a tedious and arduous
task. Yet, the parameters need to be retuned for each new problem at hand which
might reduce the practicability of the meta-heuristic approaches. The proposed
DPSO and PFPSO methods are not exceptions to this. On the other hand,
experimentation of the PSO-based approaches revealed that their exceptional
performances were largely resulting from their heuristic modules. It was also
observed that none of the previous heuristics have the capacity to tackle large-scale
Pareto front DTCT problems. Consequently, regarded as the chief contribution of
this thesis, different variants of a new Cost-Slope Heuristic are designed and
developed within the context of this thesis. The proposed CS-Heuristics include
parallel and serial versions for both cost minimization/deadline and Pareto front
classes of DTCTPs. CS-Heuristic engages unique scientific and programmable
rules comprising an innovative Partial-CPM technique which is designed to
accelerate the solutions process. Furthermore, similar to the proposed exact
methods, parallel and serial merging techniques are implemented to reduce the scale
and computation cost of the practiced problems.

Comparative studies on the single-objective CS-Heuristic involving a set of small,

medium, and large-scale benchmark problems not only confirmed its soundness,
but also revealed its superiority over earlier state-of-the-art approaches. While it is

203
able to locate high quality solutions for all the practiced cost minimization/deadline
problems, computation time requirement of this method is also demonstrated to be
remarkably less than the earlier approaches. In fact, by running on the same desktop
computer, the processing time of this method is experimented to be less than that
of DPSO. It is proved that the proposed CS-Heuristic by outperforming the highly
capable DPSO method, can converge to global optimal solutions with only
fractional deviations. To the best of author’s knowledge, the proposed CS-Heuristic
optimization model is the first method that outdoes state-of-the-art meta-heuristic
approaches and is capable of unraveling large-scale problems comprising thousands
of activities within practically reasonable timeframes with only fractional
deviations.

Comparative studies on the multi-objective CS-Heuristic involved a set of existing

small, medium, and large-scale benchmark instances, case problems acquired from
the literature, and new sets of RanGen2 instances. A collection of unary and binary
performance metrics was measured in the course of performance evaluations
including cardinality, accuracy, diversity, and efficiency indices. Results revealed
an unmatched performance by CS-Heuristic in comparison with the previous
approaches including PFPSO algorithm. It is concluded that not only the
computation time requirement of the innovative multi-objective CS-Heuristic is
substantially less than the earlier approaches, but it is also able to produce a large
number of high quality non-dominated solutions for all the practiced Pareto front
problems. Compared to results of PFPSO over the entire collection of the RanGen2
instances, both parallel and serial variants of CS-Heuristic are shown to be able to
achieve higher number of Pareto solutions, serial variant even more so than the
others. It is also discovered that the number of non-dominated solutions located
along the final unified frontier are significantly larger for CS-Heuristics. In fact,
PFPSO only accounted for less than 7% of these solutions, while serial variant
captured 23% more than the parallel CS-Heuristic. Despite high levels of accuracy
for all the experimented methods, CS-Heuristics are confirmed to be more

204
successful in converging to the true Pareto fronts with fractional deviations.
Average deviations for parallel and serial CS-Heuristic are shown to be literally the
same; whereas, PFPSO’s average deviation value is calculated to be five to six times
the amount for CS-Heuristics. Well-distributed and widely spread fronts are only
obtained by means of CS-Heuristic since the area of the solution space covered by
PFPSO, poorly represents the Hypervolume of the true/best Pareto front. It is also
interpreted that the nature of PFPSO and CS-Heuristic approaches, in contrast to
the proposed exact method, tend to be more suitable for pseudo-parallel networks
that include greater number of time-cost alternatives. To the best of author’s
knowledge, the proposed CS-Heuristic optimization model is the first method that
outperforms the highly capable meta-heuristic approaches and is able to tackle
large-scale problems consisting of hundreds of activities within reasonably short
timespans and practically viable deviations. Owing to its unprecedented efficacy
and exceptional accuracy, Cost-Slope Heuristic is expected to contribute to optimal
planning of realistic construction projects.

Practicability and real-life applicability of the proposed PFPSO and CS-Heuristics

were firmly validated by means of real construction projects acquired from the
literature. Due to the extremely successful performance of the proposed
optimization approaches over the real projects, it is interpreted that the realistic
projects tend to be relatively simpler than the instances generated within the course
of this thesis. The deviation and the runtime amounts are conceded to be small even
for the large-scale 519-activity construction project. In fact, all the methods are
observed to be able to locate high quality solutions which are either optimal or very
close to the optimal frontier within less than a second.

Integration of the proposed optimization algorithms into Microsoft Project is also

presented in this thesis. An add-in which is capable of solving cost
minimization/deadline and Pareto problems is developed which includes both the
PSO-based methods and the CS-Heuristics. It is acknowledged to enhance the

205
practicability of the proposed DTCTP optimization algorithms. Benefiting from the
presented add-in, users of Microsoft Project will readily be able to visualize the
optimized schedules for the practiced projects; hence, the proposed methods are
supposed to be more readily accepted and used by the parties to construction
projects. By means of the developed add-in, the new optimization models are
envisioned to be applicable in real projects and to suit the actual practices of
construction managers. It is expected that these approaches might prove to be a
robust base for exerting this highly challenging problem.

To wrap up, all the different proposed DTCTP optimization models are shown to
be innovational and efficacious which are highly relevant for real-life applications.
Particularly, the CS-Heuristic is regarded as a pioneering method which presents a
new uncrashing concept along with salient techniques for network reduction and
faster analysis of the networks. Experimentations attest to the efficacy and
efficiency of the proposed methods for successful solution of real-life-scale
problems. While the multi-objective variants of all the presented models enable
articulation of decision makers’ preferences, the author proposes to employ the
suggested heuristic or meta-heuristic approaches in place of the described exact
method when the optimality is not regarded as the most crucial concern.

The proposed models still have some limitations which are to be addressed in future
studies. Firstly, all the developed methods are designed to solve problems with
standard networks. They can only tackle networks with logical relationships of type
finish-to-start, considering no lags in between. However, the proposed models
should consider generalized precedence relationships by covering all the other types
of constraints including finish-to-finish, start-to-finish, and start-to-start logical
relationships. Besides, they should also allow for inclusion of positive and negative
lag times between the activities. Secondly, the proposed Microsoft Project add-in
accepts only up to ten time-cost alternatives. Though, it is possible for a project

206
activity to comprise more than ten options; hence, the proposed add-in should
support a larger number of entries for time-cost options.
Despite exceptional performance of the proposed models, there still remains room
for improvements. The proposed models run on a single core of the CPU. Though,
dividing the parallel calculations and processes of the algorithms into several cores
would further enhance the convergence speed of the optimization approaches.
Much research also remains to be done toward extending the developed models to
incorporate other aspects of construction projects such as uncertainties, resource
limitations, safety, productivity, and quality. Last but not least, development of a
hybrid meta-heuristic algorithms by capitalizing on the solid convergence
capabilities of the proposed CS-Heuristic appears to be a promising research area.

207
208
 REFERENCES

Abbasi-Iranagh, M. (2015). Development of High Performance Heuristic and

Meta-Heuristic Methods for Resource Optimization of Large Scale
Construction Projects. Doctoral dissertation, Middle East Technical
University, Ankara, Turkey.

Abdel-Raheem, M., & Khalafallah, A. (2011). Using Electimize to Solve the Time-
Cost-Tradeoff Problem in Construction Engineering. Proc., ASCE Int.
Workshop on Computing in Civil Engineering, ASCE, pp. 250–257, Reston,
VA.

Afshar, A., Ziaraty, A. K., Kaveh, A., & Sharifi, F. (2009). Nondominated
Archiving Multicolony Ant Algorithm in Time-Cost Trade-Off
Optimization. Journal of Construction Engineering and Management-
ASCE, 135(7), 668-674.

Agdas, D., Warne, D. J., Osio-Norgaard, J., & Masters, F. J. (2018). Utility of
Genetic Algorithms for Solving Large-Scale Construction Time-Cost
Trade-Off Problems. Journal of Computing in Civil Engineering, 32(1),
04017072.

Agrawal, M. K., Elmaghraby, S. E., & Herroelen, W. S. (1996). DAGEN: A

Generator of Testsets For Project Activity Nets. European Journal of
Operational Research, 90(2), 376–382.

209
Akkan, C., Drexl, A., & Kimms, A. (2005). Network Decomposition-Based
Benchmark Results for the Discrete Time-Cost Tradeoff Problem.
European Journal of Operational Research, 165, 339-358.

Aminbakhsh, S. (2013). Hybrid Particle Swarm Optimization Algorithm for

Obtaining Pareto Front of Discrete Time-Cost Trade-Off Problem.
Master’s thesis, Middle East Technical University, Ankara, Turkey.

Aminbakhsh, S., & Sonmez, R. (2016). Discrete Particle Swarm Optimization

Method for the Large-Scale Discrete Time-Cost Trade-off Problem. Expert
Systems with Applications, 51, 177–185.

Aminbakhsh, S., & Sonmez, R. (2017). Pareto Front Particle Swarm Optimizer for
Discrete Time-Cost Trade-Off Problem. Journal of Computing in Civil
Engineering, 31(1), 04016040.

Anagnostopoulos, K. P., & Kotsikas, L. (2010). Experimental Evaluation of

Simulated Annealing Algorithms for the Time-Cost Trade-off Problem.
Applied Mathematics and Computation, 217(1), 260-270.

Ashuri, B., & Tavakolan, M. (2012). Fuzzy Enabled Hybrid Genetic Algorithm–
Particle Swarm Optimization Approach to Solve TCRO Problems in
Construction Project Planning. Journal of Construction Engineering and
Management, 138(9), 1065-1074.

Benders, J.F. (1962). Partitioning Procedures for Solving Mixed Variables

Programming Problems, Numerische Mathematik, 4, 238-252.

210
Bettemir, Ö. H. (2009). Optimization of time-cost-resource trade-off problems in
project scheduling using meta-heuristic algorithm. Doctoral dissertation,
Middle East Technical University, Ankara, Turkey.

Bettemir, Ö. H., & Birgönül, M. T. (2017). Network Analysis Algorithm for the
Solution of Discrete Time-Cost Trade-off Problem. KSCE Journal of Civil
Engineering, 21(4), 1047–1058.

Bilir, M. (2015). A Mixed Integer Programming Method for Pareto Front

Optimization of Discrete Time Cost Trade-off Problem. Master’s thesis,
Middle East Technical University, Ankara, Turkey.

Birbil, S., & Fang, S. C. (2003). An Electromagnetism-Like Mechanism for Global

Optimization. Journal of Global Optimization, 25,263–282

Burns, S. A., Liu, L., & Feng, C. W. (1996). The LP/IP Hybrid Method for
Construction Time-Cost Trade-off Analysis. Construction Management
and Economics, 14(3), 265–276.

Butcher, W.S. (1967). Dynamic Programming for Project Cost-Time Curves.

Journal of the Construction Division, Proceedings of the ASCE, 93, 59-73.

Chassiakos, A. P., & Sakellaropoulos, S. P. (2005). Time-Cost Optimization of

Construction Projects with Generalized Activity Constraints. Journal of
Construction Engineering and Management-ASCE, 131(10), 1115-1124.

Chen, P. H., & Weng, H. (2009). A Two-Phase GA Model for Resource-

Constrained Project Scheduling. Automation in Construction, 18(4), 485–
498.

211
Coello, C. A. C., Pulido, G. T., & Lechuga, M. S. (2004). Handling Multiple
Objectives with Particle Swarm Optimization. IEEE Transactions on
Evolutionary Computation, 8(3), 256–279.

Colorni, A., Dorigo, M., & Maniezzo, V. (1992). Distributed Optimization by Ant
Colonies. Toward a Practice of Autonomous Systems, 134-142.

Construction Industry Institute. (1988). Concepts and Methods of Schedule

Compression. Austin, TX: Construction Industry Institute, University of
Texas.

Crowston, W. B., & Thompson, G. L. (1967). Decision CPM: A Method for

Simultaneous Planning, Scheduling and Control of Projects. Operations
Research, 15, 407-426.

De, P., Dunne, E. J., Ghosh, J. B., & Wells, C. E. (1995). The Discrete Time-Cost
Tradeoff Problem Revisited. European Journal of Operational Research,
81(2), 225-238.

De, P., Dunne, E. J., Ghosh, J. B., & Wells, C. E. (1997). Complexity of the Discrete
Time-Cost Trade off Problem for Project Networks. Operations Research,
45(2), 302–306.

Deckro, R. F., Hebert, J. E., Verdini, W. A., Grimsrud, P. H., & Venkateshwar, S.
(1995). Nonlinear Time Cost Tradeoff Models in Project-Management.
Computers & Industrial Engineering, 28(2), 219-229.

Degirmenci, G., & Azizoglu, M. (2013). Branch and Bound Based Solution
Algorithms for the Budget Constrained Discrete Time/Cost Trade-off
Problem. Journal of the Operational Research Society, 64(10), 1474–1484.

212
Demeulemeester, E. L., De Reyck, B., Foubert, B., Herroelen, W. S., & Vanhoucke,
M. (1998). New Computational Results on the Discrete Time/Cost Trade-
off Problem in Project Networks. Journal of the Operational Research
Society, 49(11), 1153-1163.

Demeulemeester, E. L., Dodin, B., & Herroelen, W. S. (1993). A Random Activity

Network Generator. Operations Research, 41(5), 972-980.

Demeulemeester, E. L., & Herroelen, W. S. (2002). Project Scheduling: A Research

Handbook. Dordrecht, Boston: Kluwer Academic Publishers.

Demeulemeester, E. L., Herroelen, W. S., & Elmaghraby, S. E. (1996). Optimal

Procedures for the Discrete Time Cost Trade-off Problem in Project
Networks. European Journal of Operational Research, 88(1), 50-68.

Demeulemeester, E., Vanhoucke, M., & Herroelen, W. (2003). RanGen: A Random

Network Generator for Activity-On-The-Node Networks. Journal of
Scheduling. Vol. 6, pp. 17–38.

Dragović, N., Vulević, T., Todosijević, M., Kostadinov, S., & Zlatić, M. (2017).
Minimization of Direct Costs in the Construction of Torrent Control.
Technical Gazette, 24, 4(2017), 1123-1128.

Eberhart, R., & Kennedy, J. (1995). A New Optimizer Using Particle Swarm
Theory. Proceedings of the Sixth International Symposium on Micro
Machine and Human Science, 39-43.

213
El-Abbasy, M. S., Elazouni, A., & Zayed, T. (2016). MOSCOPEA: Multi-
Objective Construction Scheduling Optimization Using Elitist Non-
Dominated Sorting Genetic Algorithm. Automation in Construction,
71(Part 2).

Elbeltagi, E., Hegazy, T., & Grierson, D. (2005). Comparison Among Five
Evolutionary-Based Optimization Algorithms. Advanced Engineering
Informatics, 19(1), 43-53.

Elbeltagi, E., Hegazy, T., & Grierson, D. (2007). A Modified Shuffled Frog-
Leaping Optimization Algorithm: Applications to Project Management.
Structure and Infrastructure Engineering, 3(1), 53-60.

El-Rayes, K., & Kandil, A. (2005). Time-Cost-Quality Trade-Off Analysis for

Highway Construction. Journal of Construction Engineering and
Management, 131(4), 477–486.

Eshtehardian, E., Afshar, A., & Abbasnia, R. (2008). Time–Cost Optimization:

Using GA and Fuzzy Sets Theory for Uncertainties in Cost. Construction
Management and Economics, 26(7), 679-691.

Eshtehardian, E., Afshar, A., & Abbasnia, R. (2009). Fuzzy-based MOGA

Approach to Stochastic Time-Cost Trade-off Problem. Automation in
Construction, 18(5), 692–701.

Eusuff, M. M., & Lansey, K. E. (2003). Optimizing of Water Distribution Network

Design Using the Shuffled Frog Leaping Algorithm. Journal of Water
Resources Planning and Management, 129(3), 210–225

214
Falk, J. E., & Horowitz, J. L. (1972). Critical Path Problems with Concave Cost-
Time Curves. Management Science Series B-Application, 19(4), 446-455.

Fallah-Mehdipour, E., Bozorg Haddad, O., Rezapour Tabari, M. M., & Mariño, M.
A. (2012). Extraction of Decision Alternatives in Construction Management
Projects: Application and Adaptation of NSGA-II and MOPSO. Expert
Systems with Applications, 39(3), 2794–2803.

Feng, C. W., Liu, L., & Burns., S. A. (1997). Using Genetic Algorithms to Solve
Construction Time-Cost Trade-off Problems. Journal of Computing in Civil
Engineering, 11(3), 184-189.

Foldes, S., & Soumis, F. (1993). PERT and Crashing Revisited - Mathematical
Generalizations. European Journal of Operational Research, 64(2), 286-
294.

Fondahl, J. M. (1961). A Non-Computer Approach to the Critical Path Method for

the Construction Industry. Technical Report, No. 9, Construction Institute,
Department of Civil Engineering, Stanford University, California.

Frank, H., Frisch, I. T., Van Slyke, R. & Chou, W. S. (1971). Optimal Design of
Centralized Computer Networks. Networks, 1, 43-57.

Fulkerson, D. R. (1961). A Network Flow Computation for Project Cost Curves.

Management Science, 7(2), 167-178.

Geem, Z. (2010). Multiobjective Optimization of Time-Cost Trade-off Using

Harmony Search. Journal of Construction Engineering and Management,
136(6), 711–716.

215
Goldberg, D. E., & Segrest, P. (1987). Finite Markov Chain Analysis of Genetic
Algorithms. Proceedings of the Second International Conference on
Genetic Algorithms and their application, Cambridge, Massachusetts,
United States.

Goyal, S. K. (1975). A Note on A Simple CPM Time-Cost Trade-off Algorithm.

Management Science, 21(6), 718–722.

Hansen, M. P., & Jaszkiewicz, A. (1998). Evaluating the Quality of

Approximations to the Non-Dominated Set. Technical report IMM-
REP1998-7, Technical University of Denmark.

Hazır, Ö., Erel E., & Gunalay, Y. (2011). Robust Optimization Models for the
Discrete Time/Cost Trade-off Problem. International Journal of Production
Economics, 130, 87-95.

Hazır, Ö., Haouari, M., & Erel, E. (2010). Discrete Time/Cost Trade-off Problem:
A Decomposition-based Solution Algorithm for the Budget Version.
Computers & Operations Research, 37(4), 649–655.

Hegazy, T. (1999). Optimization of Construction Time-Cost Trade-off Analysis

Using Genetic Algorithms. Canadian Journal of Civil Engineering, 26(6),
685-697.

Holland, J. H. (1975). Adaptation in Natural and Artificial Systems: An

Introductory Analysis with Applications to Biology, Control, and Artificial
Intelligence. Cambridge, MA, USA: University of Michigan Press.

216
Huang, Y., Zou, X., & Zhang, L. (2016). Genetic Algorithm – Based Method for
the Deadline Problem in Repetitive Construction Projects Considering Soft
Logic. Journal of Management in Engineering, 32(4), 1–9.

Kalhor, E., Khanzadi, M., Eshtehardian, E., & Afshar, A. (2011). Stochastic Time-
Cost Optimization Using Non-Dominated Archiving Ant Colony Approach.
Automation in Construction, 20(8), 1193–1203.

Kandil, A. (2005). Multi-Objective Optimization for Large-Scale Highway

Construction Projects. Doctoral dissertation, University of Illinois, Urbana
Champaign, Urbana, IL.

Kandil, A., & El-Rayes, K. (2006). Parallel Genetic Algorithms for Optimizing
Resource Utilization in Large-Scale Construction Projects. Journal of
Construction Engineering and Management, 132(5), 491–498.

Kelley, J. E. (1961). Critical-Path Planning and Scheduling - Mathematical Basis.

Operations Research, 9(3), 296-320.

Kelley, J. E., & Walker, M. R. (1959). Critical-Path Planning and Scheduling.

Proceedings of Eastern Joint Computer Conference, Vol. 16, 160–173.

Kennedy, J., & Eberhart, R. C. (1995). Particle Swarm Optimization. IEEE

International Conference on Neural Networks Proceedings, Vols 1-6, 1942-
1948.

Kennedy, J., & Eberhart, R. C. (1997). A Discrete Binary Version of the Particle
Swarm Algorithm. Smc '97 Conference Proceedings - IEEE International
Conference on Systems, Man, and Cybernetics, Vols 1-5, 4104-4108.

217
Kerzner, H. (2009). Project Management: A Systems Approach to Planning,
Scheduling, and Controlling. John Wiley & Sons.

Kirkpatrick, S., Gellat, C. D., & Vecchi, M. P. (1983). Optimization by Simulated

Annealing. Science, 220, 671 – 680.

Knowles, J., & Corne, D. (2002). On Metrics for Comparing Nondominated Sets.
IEEE Computer Society - Proceedings of the 2002 Congress on
Evolutionary Computation, CEC 2002, Vol. 1, pp. 711–716.

Kolisch, R., & Hartmann, S. (2006). Experimental Investigation of Heuristics for

Resource-Constrained Project Scheduling: An Update. European Journal of
Operational Research, 174(1), 23–37.

Kolisch, R., Sprecher, A., & Drexl, A. (1995). Characterization and Generation of
a General Class of Resource-Constrained Project Scheduling Problems.
Management Science, 41, 1693–1703.

Koo, C., Hong, T., & Kim, S. (2015). An Integrated Multi-Objective Optimization
Model for Solving the Construction Time-Cost Trade-off Problem. Journal
of Civil Engineering and Management, 21(3), 323–333.

Land, A. H., & Doig, A. G. (1960). An Automatic Method of Solving Discrete

Programming Problems. Econometrica: Journal of the Econometric
Society, 497-520.

Li, H., & Love, P. (1997). Using Improved Genetic Algorithms to Facilitate Time-
Cost Optimization. Journal of Construction Engineering and management,
123(3), 233-237.

218
Liberatore, M., Pollack-Johnson, B., & Smith, C. (2001). Project Management in
Construction: Software Use and Research Directions. Journal of
Construction Engineering and Management, 127(2), 101–107.

Liu, L., Burns, S. A., & Feng, C. W. (1995). Construction Time-Cost Trade-off
Analysis Using LP/IP Hybrid Method. Journal of Construction Engineering
and Management, 121(4), 446-464.

Lock, D. (2007). Project Management: 9Th Edition. Surrey, UK: Gower

Publication.

Menesi, W., Golzarpoor, B., & Hegazy, T. (2013). Fast and Near-Optimum
Schedule Optimization for Large-Scale Projects. Journal of Construction
Engineering and Management, 139(9), 1117–1124.

Meyer, W. L., & Shaffer, L. R. (1963). Extensions of the Critical Path Method
through the Application of Integer Programming. Civil Engineering
Construction Research Series No: 2. University of Illinois, Urbana
Champaign, Urbana, IL.

Meyer, W. L., & Shaffer, L. R. (1965). Extending CPM for Multiform Project
Time-Cost Curves. Journal of Construction Division, ASCE, 91(1), 45-68.

Mittelmann, H. D. (2013). Benchmarks for Optimization Software. Retrieved

March 2014, from http://plato.asu.edu/bench.html

Moder, J. J., Phillips, C. R., & Davis, E. W. (1983). Project Management with CPM,
PERT, and Precedence Diagramming (3rd Ed.). New York: Van Nostrand
Reinhold.

219
Monghasemi, S., Nikoo, M. R., Khaksar Fasaee, M. A., & Adamowski, J. (2015).
A Novel Multi Criteria Decision Making Model for Optimizing Time-Cost-
Quality Trade-off Problems in Construction Projects. Expert Systems with
Applications, 42(6), 3089–3104.

Moselhi, O. (1993). Schedule Compression Using the Direct Stiffness Method,

Canadian Journal of Civil Engineering, 20, 65-72.

Moussourakis, J., & Haksever, C. (2004). Flexible Model for Time/Cost Tradeoff
Problem. Journal of Construction Engineering and Management, ASCE,
130(3), 307-314.

Mubarak, S. (2010). Construction Project Scheduling and Control. John Wiley &
Sons.

Mungle, S., Benyoucef, L., Son, Y. J., & Tiwari, M. K. (2013). A Fuzzy Clustering-
based Genetic Algorithm Approach for Time-Cost-Quality Trade-off
Problems: A Case Study of Highway Construction Project. Engineering
Applications of Artificial Intelligence, 26(8), 1953–1966.

Ng, S. T., & Zhang, Y. S. (2008). Optimizing Construction Time and Cost Using
Ant Colony Optimization Approach. Journal of Construction Engineering
and Management, ASCE, 134(9), 721-728.

Okabe, T., Jin, Y., & Sendhoff, B. (2003). A Critical Survey of Performance Indices
for Multi-Objective Optimisation. The 2003 congress on evolutionary
computation, CEC '03, IEEE Press, vol. 2, pp. 878–85.

220
Panagiotakopoulos, D. (1977). A CPM Time-Cost Computational Algorithm for
Arbitrary Activity Cost Functions. INFOR: Information Systems and
Operational Research, 15(2), 183-195.

Riquelme, N., Von Lucken, C., & Baran, B. (2015). Performance Metrics in Multi-
Objective Optimization. 2015 Latin American Computing Conference,
CLEI, IEEE, pp. 1–11.

Robinson, D.R. (1975). A Dynamic Programming Solution to the Cost-Time

Tradeoff for CPM. Management Science, 22,158-166.

Rothfarb, B., Frank, H., Rosebaum, D., Steiglitz, K., & Kleitman, D. (1970).
Optimal Design of Offshore Natural Gas Pipeline Systems. Operations
Research, 18, 992-1020.

Schwindt, C. (1995). ProGen/Max: A New Problem Generator for Different

Resource-Constrained Project Scheduling Problems with Minimal and
Maximal Time Lags. Research Report, WIOR 449, University of Karlsruhe.

Siemens, N. (1971). A Simple CPM Time-Cost Tradeoff Algorithm. Management

Science, 17(6), B354-B363.

Skutella, M. (1998). Approximation Algorithms for the Discrete Time-Cost

Tradeoff Problem. Mathematics of Operations Research, 23(4), 909-929.

Sonmez, R., & Bettemir, O. H. (2012). A Hybrid Genetic Algorithm for the Discrete
Time-Cost Trade-off Problem. Expert Systems with Applications, 39(13),
11428-11434.

221
Su, Z., Qi, J., & Wei, H. (2017). Simplifying the Nonlinear Continuous Time-Cost
Tradeoff Problem. Journal of Systems Science and Complexity, 30(4), 901–
920.

Szmerekovsky, J.G., & Venkateshan, P. (2012). An Integer Programming

Formulation for the Project Scheduling Problem with Irregular Time-Cost
Tradeoffs. Computers & Operations Research, 39, 1402-1410.

Tavares, L. V. (1999). Advanced Models for Project Management. Boston, MA:

Kluwer Academic Publishers.

Vanhoucke, M. (2005). New Computational Results for the Discrete Time/Cost

Trade-off Problem with Time-Switch Constraints. European Journal of
Operational Research, 165(2), 359-374.

Vanhoucke, M. (2015). Generalized Discrete Time-Cost Tradeoff Problems. In:

Schwindt C, Zimmermann J (eds) Handbook on project management and
scheduling, vol 1, pp 639–658, Berlin: Springer.

Vanhoucke, M., Coelho, J., Debels, D., Maenhout, B., & Tavares, L. V. (2008). An
Evaluation of the Adequacy of Project Network Generators with
Systematically Sampled Networks. European Journal of Operational
Research, 187(2), 511–524.

Vanhoucke, M., & Debels, D. (2007). The Discrete Time/Cost Trade-off Problem:
Extensions and Heuristic Procedures. Journal of Scheduling, 10(4-5), 311-
326.

222
Vanhoucke, M., Demeulemeester, E., & Herroelen, W. (2002). Discrete Time/Cost
Trade-offs in Project Scheduling with Time-Switch Constraints. Journal of
the Operational Research Society, 53(7), 741-751.

Veldhuizen, D. A. V. (1999). Multiobjective Evolutionary Algorithms:

Classifications, Analyses, and New Innovations. Doctoral dissertation,
Graduate School of Engineering of the Air Force Institute of Technology,
Air University, OH.

Xiong, Y., & Kuang, Y. P. (2008). Applying an Ant Colony Optimization

Algorithm-based Multiobjective Approach for Time-Cost Trade-off.
Journal of Construction Engineering and Management, ASCE, 134(2), 153-
156.

Yang, I. T. (2007a). Performing Complex Project Crashing Analysis with Aid of

Particle Swarm Optimization Algorithm. International Journal of Project
Management, 25(6), 637-646.

Yang, I. T. (2007b). Using Elitist Particle Swarm Optimization to Facilitate

Bicriterion Time-Cost Trade-off Analysis. Journal of Construction
Engineering and Management, ASCE, 133(7), 498-505.

Yang, H. H., & Chen, Y. L. (2000). Finding the Critical Path in An Activity
Network with Time-Switch Constraints. European Journal of Operational
Research, 120(3), 603-613.

Zadeh, L. A. (1965). Fuzzy Sets. Information and Control, 8(3), 338-353.

223
Zhang, H., & Li, H. (2010). Multi-Objective Particle Swarm Optimization for
Construction Time-Cost Tradeoff Problems. Construction Management and
Economics, 28(1), 75–88.

Zhang, H., & Xing, F. (2010). Fuzzy-Multi-Objective Particle Swarm Optimization

for Time-Cost-Quality Tradeoff in Construction. Automation in
Construction, 19(8), 1067-1075.

Zhang, L., Zou, X., & Qi, J. (2015). A Trade-Off Between Time and Cost in
Scheduling Repetitive Construction Projects. Journal of Industrial and
Management Optimization, 11(4), 1423-1434.

Zhang, Y., & Ng, S. (2012). An Ant Colony System Based Decision Support
System for Construction Time-Cost Optimization. Journal of Civil
Engineering and Management, ASCE, 18(4), 580–589.

Zheng, D. X. M., Ng, S. T., & Kumaraswamy, M. M. (2004). Applying a Genetic

Algorithm-based Multiobjective Approach for Time-Cost Optimization.
Journal of Construction Engineering and Management, ASCE, 130(2), 168-
176.

Zheng, D. X. M., Ng, S. T., & Kumaraswamy, M. M. (2005). Applying Pareto

Ranking and Niche Formation to Genetic Algorithm-based Multiobjective
Time-Cost Optimization. Journal of Construction Engineering and
Management, ASCE, 131(1), 81-91.

Zitzler, E., Deb, K., & Thiele, L. (2000). Comparison of Multiobjective

Evolutionary Algorithms: Empirical Results. Evolutionary Computation,
8(2), 173–195.

224
Zitzler, E., Laumanns, M., & Thiele, L. (2001). SPEA2: Improving the Strength
Pareto Evolutionary Algorithm. TIK-report 103, Lausanne, Switzerland:
Swiss Federal Institute of Technology.

Zitzler, E., & Thiele, L. (1998). Multiobjective Optimization Using Evolutionary

Algorithms – A Comparative Case Study. Proceedings of 5th International
Conference on Parallel Problem Solving from Nature, (PPSN-V), pp. 292-
301.

Zitzler, E., & Thiele, L. (1999). Multiobjective Evolutionary Algorithms: A

Comparative Case Study and the Strength Pareto Approach. IEEE
Transactions on Evolutionary Computation, 3(4), 257–271.

Zou, X., Fang, S. C., Huang, Y. S., & Zhang, L. H. (2016). Mixed-Integer Linear
Programming Approach for Scheduling Repetitive Projects with Time-Cost
Trade-Off Consideration. Journal of Computing in Civil Engineering,
31(3), 6016003.

225
226
 CURRICULUM VITAE

Saman Aminbakhsh
Date of Birth: 25 Jan 1986
Place of Birth: Tabriz, Iran
E-mail: [email protected]
Cell-Phone: +90 536 761 72 61

EDUCATION

2013-2018 Middle East Technical University (METU), Ankara,

Turkey
Doctor of Philosophy (Ph.D.), Civil Engineering
Construction Engineering and Management
CGPA: 4/4

2010-2013 Middle East Technical University (METU), Ankara,

Turkey
Master’s degree (M.Sc.), Civil Engineering
Construction Engineering and Management
CGPA: 4/4

2004-2009 Islamic Azad University of Tabriz (IAUT), Tabriz, Iran

Bachelor’s degree (B.Sc.), Civil Engineering
CGPA: 16.47/20

227
1997-2004 National Organization for Development of Exceptional
Talents (NODET), Tabriz, Iran

EXPERIENCE

2013-2018 Teaching Assistant

Civil Engineering Department, Construction Engineering
and Management Division, Middle East Technical
University (METU), Ankara, Turkey

2014-2016 Research Assistant

Civil Engineering Department, Construction Engineering
and Management Division, Middle East Technical
University (METU), Ankara, Turkey

PUBLICATIONS

A. SCI-E and SSCI Indexed Journal Articles:

A1. Aminbakhsh, S., Gunduz, M., & Sonmez, R. (2013). Safety risk assessment
using analytic hierarchy process (AHP) during planning and budgeting of
construction projects. Journal of Safety Research, 46, 99-105.

A2. Aminbakhsh S. & Sonmez R. (2016). Discrete Particle Swarm Optimization

Method for the Large-Scale Discrete Time-Cost Trade-Off Problem. Expert
Systems with Applications, 51, 177–185.

228
A3. Aminbakhsh S. & Sonmez R. (2017). A Pareto Front Particle Swarm Optimizer
for Discrete Time-Cost Trade-Off Problem. Journal of Computing in Civil
Engineering, ASCE, 31 (1), 04016040-1- 04016040-10.

A4. A Pareto Front Optimization Heuristic for the Large-Scale discrete Time-Cost
Trade-Off Problem. (Under Preparation)

A5. Optimization of Large-Scale Discrete Time-Cost Trade-Off Problem using

Mixed-İnteger Linear Programming. (Under Preparation)

B. International Conference Proceedings:

B1. Aminbakhsh S. & Sonmez R. (2015). Pareto Oriented Optimization of Discrete

Time-Cost Trade-Off Problem Using Partıcle Swarm Optimization. Procs 31th
Annual ARCOM Conference, Lincoln, UK, 33-40.

B2. Aminbakhsh S., Sonmez R. & Atan T. (2017). An Efficient Heuristic Method
for Pareto Oriented Optimization of Time-Cost Trade-Off Problem for Construction
Projects. Procs. IO2017 The XVIII Congress of the Portuguese Association of
Operational Research, Valença, Portugal.

B3. Aminbakhsh S. & Sönmez R. (2018). Integrating a Meta-Heuristic Method into

Microsoft Project for Time-Cost Trade-off Analyses. 5th İnternational Project and
Construction Management Conference (IPCMC), Kyrenia, Cyprus. (Abstract
Accepted)

C. National Conference Proceedings:

C1. Haghgooei A. & Aminbakhsh S. (2012). Türk Müteahhitlerin İran İnşaat

Sektöründe Karşılaşabilecekleri Risklerin İncelenmesi. 2. Proje ve Yapım Yönetimi
Ulusal Kongresi, İzmir, Turkey.

229
C2. Aminbakhsh S, Sönmez R., Iranagh, M.A., & Rezvankhah E. (2014). Kuş
Sürüsü Optimizasyon Algoritması ile Kesikli Zaman-Maliyet Ödünleşim
Probleminin Çözümü. 3. Proje ve Yapım Yönetimi Ulusal Kongresi, Antalya,
Turkey.
C3. Aminbakhsh S., Sönmez R., & Bilir M. (2016). Tamsayılı Doğrusal
Programlama Yöntemiyle Kesikli Zaman-Maliyet Ödünleşim Probleminin Optimal
Pareto Çözümü. 4. Proje ve Yapım Yönetimi Ulusal Kongresi, Eskişehir, Turkey.

D. Thesis:

D1. Aminbakhsh, S. (2013). Hybrid Particle Swarm Optimization Algorithm for

Obtaining Pareto Front of Discrete Time–Cost Trade-off Problem. Master’s thesis,
Middle East Technical University, Ankara, Turkey.

E. Citing Articles (Web of Science, May 2018):

A1: 43
A2: 9
A3: 3
Total Citations: 55

AWARDS AND HONORS

2016 Awarded the Highest Cited Research certificate from

Elsevier for an article published in Journal of Safety
Research.

2001 Selected for National Organization for Development of

Exceptional Talents (NODET) among over 800,000
participants.

230
TECHNICAL SKILLS

Adobe Photoshop
AIMMS-Advanced Optimization Software
Autodesk Sketchbook
C++/C# Programming
Microsoft Office
Microsoft Project
Microsoft Visio
Microsoft Visual Studio
Primavera PERTMaster-Risk analysis software
Primavera Project Management-P3 and P6
SPSS Statistics

LANGUAGES

Azeri Mother Tongue

English Fluent
Persian Native Language
Turkish Fluent

REFERENCES

References available upon request.

231