Back Calculation

BACKCALCULATION OF PAVEMENT MODULI
USING GENETIC ALGORITHMS
A Dissertation
Presented to
The Graduate Faculty of The University of Akron
In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy
Wael Alkasawneh
May, 2007
BACKCALCULATION OF PAVEMENT MODULI

USING GENETIC ALGORITHMS
Wael Alkasawneh
Dissertation
Approved:
Accepted:
Advisor
Dr. Ernian Pan
Department Chair
Dr. Wieslaw Binienda
Committee Member
Dr. Wieslaw Binienda
Dean of the College

Dr. George Haritos
Committee Member
Dr. Graham Kelly
Dean of the Graduate School

Dr. George Newkome
Committee Member
Dr. Kevin Kreider
Date
Committee Member
Dr. Abdullah Malkawi
Committee Member
Dr. Ping Yi
Committee Member
Dr. Daren Zywicki
ii
ABSTRACT
The static elastic moduli of pavement layers can be considered to be among the
most controversial physical properties in pavement engineering. In addition, pavement
analysis using the static elastic moduli of the constituting layers is widely known and
accepted by engineers and practitioners due to its simplicity. Nondestructive tests are
commonly performed on existing pavements to measure the surface deflections, which in
turn are used to backcalculate the elastic moduli of the pavement layers. However, the
accuracy of the backcalculated moduli is dependent on the backcalculation procedure and
the associated seed moduli. None of the existing classical backcalculation methods can
find the actual pavement moduli due to the theoretical limitations of the existing
methods. These limitations include the convergence to local optima due to the use of seed
moduli, which in turn lead to erroneous pavement moduli. The genetic algorithms can be
used to optimize the search domain of the backcalculated moduli to avoid the premature
convergence to local optima. The use of genetic algorithms in pavement engineering is
new and no guidelines or thorough investigations have been carried out to address all the
aspects and challenges associated with the backcalculation procedure using the genetic
algorithms. This study can be considered as the first comprehensive work that deals with
all aspects of both pavement and genetic algorithms and how to merge them. In addition,
this work can be considered as the first state of the art work on the backcalculation of
pavement moduli using genetic algorithms. nuiuiiuiiiuiuiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiii.
iii
In this study, the use of genetic algorithms has been studied thoroughly to address
all the important parameters and operators that affect the backcalculation process. In
addition, recommendations and findings regarding all the details needed to carry out the
backcalculation process were identified and discussed thoroughly. New novel methods to
study the interaction between the genetic operators and parameters and their effect on the
backcalculation process were developed.
Recommendations regarding the genetic operators as well as the genetic
parameters were presented throughout the work. In addition, the AASHTO recommended
ranges of pavement moduli were modified based on the study results to suit the GAs
backcalculation process.
On the other hand, a new novel method was developed to automate the
backcalculation process. The automation of the backcalculation process was aimed at
reducing the number of inputs needed to carry out the backcalculation process and to
make it more appealing to be used in practice. A new Dynamic Parameterless Genetic
Algorithm (DPGA) was developed as part of this work. The new DPGA can be extended
to many other applications of genetic algorithms including robotics and optimizations.
A new program (BackGenetic3d) was developed based on the novel
MultiSmart3D program developed by the Computation and Simulation Group at the
University of Akron. The new program is the first in the world that can backcalculate the
pavement moduli of pavement systems with any arbitrary number of layers, loading
conditions, and loading configurations. Existing classical programs use backcalculation
procedures that lead to local optima and limited to a maximum of 5 pavement layers and
one loading circle with uniform pressure. mmmmmmmmmmmmmmmmmmmmmmm.
iv
DEDICATION
To my wife Nermeen
To my parents
To my brother Zaids memory
ACKNOWLEDGEMENTS
I would like to extend my deepest thanks and gratitude to my advisor, Dr. Ernian
Pan, for his continuous and generous support and inestimable suggestions and guidance
during this work. I would also like to express my gratitude to my committee members,
Dr. Wieslaw K. Binienda, Dr. Graham Kelly, Dr. Kevin Kreider, Dr. Abdullah Malkawi,
Dr. Ping Yi, and Dr. Daren Zywicki, for their time and support. During the course of this
work, all of the members gave me many very helpful suggestions and recommendations.
Also I would like to thank Mr. Roger Green from Ohio Department of
Transportation for his precious suggestions and continuous support. I also wish to thank
my colleagues Feng Han and Ronghua Zhu, for their help and support.
I am very grateful to my wife Nermeen, for her love and patience during the PhD
period. Last but not least, I would like to thank my parents, my brothers and sisters, and
my nephews and nieces, for their constant encouragement, support and never-ending
love.
vi
TABLE OF CONTENTS
Page
LIST OF TABLES........................................................................................... xiii
LIST OF FIGURES. ......................................................................................... xv
CHAPTER
I. INTRODUCTION. .......................................................................................1
1.1 Statement of the Problem...............................................................................1
1.2 Objectives .....................................................................................................4
1.3 Scope of the Work ........................................................................................4
1.4 Outlines of the Dissertation ..........................................................................6
II. PAVEMENT ANALYSIS . ........................................................................8
2.1 Introduction...................................................................................................8
2.2 Elasticity Theory.........................................................................................10
2.2.1 Hookes Law.......................................................................................10
2.2.2 Boussinesqs Equation ........................................................................10
2.2.3 Charts Method.....................................................................................14
2.3 Method of Equivalent Thickness (MET) ....................................................15
2.4 Finite Element Method ...............................................................................17
2.5 Discrete Element Method ...........................................................................18
2.6 Visco-Elasticity Method .............................................................................19
vii
2.6.1 Maxwell Model ..................................................................................20

2.6.2 Kelvin-Voigt Model ...........................................................................21
2.6.3 Burgers Model ...................................................................................22
2.6.4 Creep Testing ....................................................................................22
2.7 Layered Systems .........................................................................................23
2.7.1 Elastic Multilayer Theory ..................................................................23
2.7.2 Factors Affecting Elasticity ...............................................................25
2.7.2.1 Anisotropy .............................................................................25
2.7.2.2 Shear Sensitivity ...................................................................26
2.8 In-situ Deflection Measurements ................................................................27
2.8.1 The Benkelman Beam ........................................................................28
2.8.2 The Dynaflect System ........................................................................28
2.8.3 The Falling Weight Deflectometer (FWD) ........................................29
2.9 Elastic Multilayer Programs .......................................................................33
2.9.1 Forward Calculation Programs ..........................................................34
2.10 Backcalculation of Elastic Moduli ............................................................35
2.10.1 Classical Backcalculation Methods .................................................36
2.10.1.1 Simplified Methods .............................................................37
2.10.1.2 Gradient Relaxation Methods .............................................38
2.10.1.3 Direct Interpolation Methods ..............................................41
2.10.2 Elastic Multilayer Backcalculation Programs ..................................41
2.10.3 Limitations of The Backcalculation Programs ................................44
2.10.4 Backcalculated Resilient Modulus versus Lab Testing ...................47
viii
2.11 Conclusions ...............................................................................................49

III. EFFECT OF THE ROOT-MEAN-SQUARE ERROR ON
PAVEMENT RESPONSE .....................................................................51
3.1 Introduction..................................................................................................51
3.2 Error Tolerance in The Backcalculation Procedure.....................................52
3.3 Backcalculation Study .................................................................................53
3.3.1 Pavement Fatigue Prediction ..............................................................65
3.3.2 Rutting Damage ..................................................................................67
3.3.3 Fatigue and Rutting Prediction ...........................................................68
3.4 Conclusions .................................................................................................69
IV. INTRODUCTION TO GENETIC ALGORITHMS. ..............................71
4.1 Introduction .................................................................................................71
4.2 Biological Background ................................................................................72
4.2.1 DNA....................................................................................................73
4.3 Evolution Theory and Computer Science ...................................................73
4.4 Genetic Algorithms .....................................................................................74
4.4.1 GA Background ..................................................................................76
4.4.2 GA Fundamental Theory (The Schema Theorem) .............................77
4.4.3 Schemata Evaluation...........................................................................83
4.4.4 Components of Genetic Algorithms ...................................................85
4.4.4.1 Population Size .....................................................................86
4.4.4.2 Selection Methods .................................................................87
4.4.5 Genetic Operators and Parameters......................................................89
4.4.5.1 Elitism ...................................................................................89
ix
4.4.5.2 Crossover ..............................................................................90

4.4.5.3 Mutation ................................................................................91
4.4.5.4 Niching .................................................................................92
4.5 GA Difficulties ............................................................................................96
4.6 Conclusions .................................................................................................97
V. BACKCALCULATION OF LAYER MODULI USING THE
BACKGENETIC3D PROGRAM .............................................................98
5.1 Introduction .................................................................................................98
5.2 BackGenetic3D Program Components ........................................................99
5.3 Backcalculation Mechanism .....................................................................101
5.4 Population Size .........................................................................................102
5.5 Crossover Probability ................................................................................104
5.6 Mutation ....................................................................................................105
5.7 Creep Mutation .........................................................................................106
5.8 Selection Procedure ..................................................................................106
5.9 Niching .....................................................................................................107
5.10 Elitism ...................................................................................................108
5.11 Population Seeding .................................................................................108
5.12 Random Number Generator ....................................................................109
5.13 BackGenetic3D Program Verification ....................................................109
5.14 Conclusions .............................................................................................113
VI. OPTIMIZATION OF GENETIC ALGORITHMS OPERATORS AND

PARAMETERS FOR THE BACKCALCULATION OF THE PAVEMENT
ELASTIC MODULI . ............................................................................114
6.1 Introduction ...............................................................................................114
6.2 Selection of GA Operators.........................................................................117
6.3 Backcalculation of Pavement Moduli .......................................................120
6.4 Genetic Operators Interaction and Performance .......................................121
6.5 Sensitivity Analysis ..................................................................................125
6.5.1 Behavior of Small Populations ........................................................128
6.5.2 Behavior of Large Populations ........................................................143
6.5.3 Behavior of Large Populations Using Different
Ranges of Moduli .............................................................................156
6.5.4 Best Fitness and Average Fitness Comparison ................................170
6.6 Conclusions ...............................................................................................178
VII. FACTORS AFFECTING MODULI BACKCALCULATION . .........180
7.1 Introduction .........................................................................................180
7.2 Chromosome Length............................................................................181
7.2.1 Chromosome Length and GA Performance......................................182
7.2.1.1 Genetic Algorithm Performance .........................................183
7.3 Importance of The Fitness Function .........................................................192
7.3.1 Search and Selection Mechanisms....................................................193
7.3.2 Least Square Error ............................................................................198
7.3.3 Mathematical Background ................................................................199
7.3.4 Selection of The Fitness Function.....................................................202
7.4 The Effect of Increasing the Number of Layers .......................................205
xi
7.5 The Effect of the Number of Sensors ........................................................209

7.6 Conclusions ...............................................................................................211
VIII. PARAMETERLESS GENETIC ALGORITHM FOR THE
BACKCALCULATION OF ELASTIC MODULI . ..........................213
8.1 Introduction ...............................................................................................213
8.2 Genetic Algorithms Difficulties.................................................................214
8.3 Diversity in Genetic Algorithms ...............................................................216
8.4 Interaction Between Genetic Operators ....................................................218
8.5 Automation of GA ...................................................................................220
8.5.1 Adaptive Population Sizing .............................................................224
8.5.1.1 Population Sizing using Schema Variances ........................225
8.5.1.2 Population Sizing using Life Time .....................................226
8.5.1.3 Population Sizing using Competing Subpopulation ...........226
8.5.1.4 Population Sizing Using Parameterless GA (PGA) ............227
8.6 The Dynamic Parameterless Genetic Algorithm .......................................229
8.7 Conclusions................................................................................................235
IX. SUMMARIES AND CONCLUSIONS.................................................237
9.1 Summaries..................................................................................................237
9.2 Conclusions ...............................................................................................238
9.3 Future Studies ...........................................................................................240
REFERENCES .............................................................................................242
xii
LIST OF TABLES
Table
Page
2.1
Boussinesqs Equations for a Point Load . ............................................12
2.2
Typical Poissons ratios of pavement materials ....................................42
2.3
Typical values of modulus of elasticity for pavement materials . .........42
3.1
Parameters of the Original (Exact) Layers ............................................55
3.2
Parameters of the Backcalculated Layers . ............................................56
3.3
Relative Errors in Moduli and RMSEs using the Exact (Table 3.1) and
Backcalculated (Table 3.2) Moduli .......................................................56
3.4
Coordinates and Locations of the Response Points . .............................56
3.5
Exact Responses from the Original (Exact) Pavement Profile . ............57
3.6
Response (Top Row) and Relative Error (Bottom Row in Parenthesis)

Using Case 1 Backcalculated Layer Model . .........................................58
3.7
Response (Top Row) and Relative Error (Bottom Row in Parenthesis) Using
Case 2 Backcalculated Layer Model .....................................................59
3.8
3.9
3.10
Response (Top Row) and Relative Error (Bottom Row in Parenthesis)

Using Case 5 Backcalculated Layer Model . .........................................62
3.11
Comparison of Fatigue Life Using Backcalculated and Exact Pavement

Moduli ...................................................................................................68
xiii
3.12
Comparison of Rutting Failure Using Backcalculated and Exact

Pavement Moduli . ................................................................................69
5.1
Four layer pavement system. ...............................................................110
5.2
Range of GA inputs used in the four-layer pavement .........................111
5.3
Ranges of the moduli . .........................................................................111
5.4
Results of the backcalculation programs . ..........................................112
6.1
Summary of Pavement Backcalculation Studies . ..............................123
6.2
Range of GA inputs used by Reddy et al. (2004) . ..............................126
6.3
Values of parameters used in the sensitivity analysis ........................127
6.4
Moduli ranges of backcalculated moduli ............................................127
6.5
Moduli ranges of backcalculated moduli ...........................................156
6.6
Typical values of modulus of elasticity for pavement materials . ......158
7.1
Genetic parameters used in the backcalculation process (3 layers) ....185
7.2
Moduli ranges for the backcalculation process (3 layers) . .................185
7.3
Data of the 23-layer pavement example . ............................................186
7.4
Genetic parameters used in the backcalculation process (23 layers) ..186
7.5
Moduli ranges for the backcalculation process (23 layers) . ...............187
7.6
Data of the 4-layer pavement example . ..............................................206
7.7
Genetic parameters used in the backcalculation process (4 layers) ....206
7.8
Moduli ranges for the backcalculation process (4 layers) . .................206
xiv
LIST OF FIGURES
Figure
Page
2.1
Notation for Boussinesqs equation (Ullidtz, 1987). .............................12
2.2
Influence lines for a circular loaded area (Holtz and Kovacs, 1981).........15
2.3
Typical transformations in the MET method, after Ullidtz (1987)........16
2.4
Typical multilayer system, after Ullidtz (1987).....................................16
2.5
Burmister two-layer stress curves (Yoder, 1959). .................................24
2.6
The Benkelman beam, after Bendana et al. (1994)................................28
2.7
The Dynaflect system, after Bendana et al. (1994)................................29
2.8
FWD typical sensors configuration. ......................................................30
2.9
Actual stress distribution under pavement systems, after Ullidtz (1987). .32
4.1
Graphical presentation of a schema example ........................................79
4.2
Schema changes due to crossover .........................................................81
4.3
Example of a two-point crossover . .......................................................90
4.4
Example where the two-point crossover fails to work ..........................90
4.5
Triangular-sharing function. After Goldberg and Richardson (1987) . .94
4.6
Power-sharing functions. After Goldberg and Richardson (1987) . ......94
4.7
Simple genetic algorithm performance with (a) and without (b)

sharing. After Goldberg and Richardson (1987) ...................................95
5.1
Flow chart of the genetic algorithm implemented in the

BackGenetic3D program .....................................................................100
xv
5.2
Chromosomal presentation of the trial solutions encoded as binary bits.102
5.3
Population sizing relations as a function of number of layers and

binary string length of genes ...............................................................104
5.4
Best fitness and deflection RMS relations with the number of generations
for the 4-layer pavement system . ........................................................112
6.1
Search Space of a 2-Layer Pavement Section......................................117
6.2
The 3-layer pavement system used in the sensitivity analysis, after

Reddy et al. (2004) ..............................................................................126
6.3
Generation average fitness (a) and RMSE of the backcalculated moduli (b)
under variable mutation rates and constant crossover rate (Pc=0.0) for a small
population . ..........................................................................................132
6.4
under variable mutation rates and constant crossover rates (Pc=0.25) for a
small population ..................................................................................133
6.5
6.6
6.7
small population ...................................................................................136
6.8
under constant mutation rates (Pm=0.001) and variable crossover rates for a
6.9
6.10
xvi
6.11
6.12
small population for a small population ..............................................141
6.13
6.14
large population . .................................................................................145
6.15
6.16
6.17
under variable mutation rates and constant crossover rates (Pc =0.75) for a
6.18
6.19
large population . ..................................................................................150
6.20 Generation average fitness (a) and RMSE of the backcalculated moduli (b)
6.21
under constant mutation rates (Pm=0. 1) and variable crossover rates for a
xvii
6.22
6.23
6.24
6.25
large population (Different ranges of moduli) . ...................................159
6.26
large population. (Different ranges of moduli) . ..................................160
6.27
6.28
large population. (Different ranges of moduli) . ...................................162
6.29
6.30
6.31 Generation average fitness (a) and RMSE of the backcalculated moduli (b)
6.32
under constant mutation rates (Pm=0. 1) and variable crossover rates for a
xviii
6.33
6.34
6.35
6.36
Generation best fitness of the backcalculated moduli under a constant

mutation rate (Pm=0.001) and variable crossover rates for a large
population. . ..........................................................................................173
6.37

population . ..........................................................................................173
6.38

population . ...........................................................................................174
6.39

population . ...........................................................................................174
6.40

population . ...........................................................................................175
6.41

population . ...........................................................................................175
6.42
Generation best fitness of the backcalculated moduli under variable

mutation rates and a constant crossover rate (Pc=0.0) for a large
population . ...........................................................................................176
6.43

population . ...........................................................................................176
xix
6.44

population . ...........................................................................................177
6.45

population . ...........................................................................................177
6.46

population . ...........................................................................................178
7.1
The 3-layer pavement system used in the chromosome length analysis, after
Reddy et al. (2004) ..............................................................................185
7.2
for the 3-layer example . .....................................................................190
7.3
for the 23-layer example . ...................................................................191
7.4
Generation average fitness and RMSE of the backcalculated moduli for the
large population in the 3-layer example using (a) Pc=0.25 and Pm=0.001,
and (b) Pc=0.25 and Pm=0.7 . ...............................................................196
7.5
Generation average fitness and RMSE of the backcalculated moduli for the
small population in the 3-layer example using (a) Pc=0.25 and Pm=0.001,
and (b) Pc=0.25 and Pm=0.7 . ...............................................................197
7.6
Generations RMSE of the backcalculated moduli for the 3-layer

example . ..............................................................................................205
7.7
for the 3-layer and 4-layer pavement systems . ..................................208
7.8
The RMSE of the backcalculated moduli for the 3-layer pavement system
using different number of sensors (RMSE method) . .........................210
7.9
The RMSE of the backcalculated moduli for the 3-layer pavement system
using different number of sensors (AREA method) . .........................211
8.1
Mutation probability as a function of generation number . ................223
8.2
for the 23-layer example using the PGA method ................................233
xx
8.3
for the 23-layer example using the DPGA method .............................234
xxi
CHAPTER I
INTRODUCTION
1.1 Statement of the Problem

The backcalculation of pavement moduli from surface deflection measurements
using nondestructive tests has been used for more than four decades to assess and manage
existing pavements and to design overlays. Unfortunately, the backcalculated pavement
moduli lack the accuracy in spite of the existence of many backcalculation programs
employing different backcalculation procedures and algorithms.
The value and accuracy of the backcalculated moduli are highly dependent on the
backcalculation procedure (Chou and Lytton, 1991; Harichandran et al., 1993). All of the
classical backcalculation procedures require seed moduli to initiate the backcalculation
process. Different seed moduli often lead to different backcalculated moduli which in
turn lead to different pavement designs and evaluations adding more challenges to
engineers.
The main problems that any classical backcalculation procedure faces are
convergence, accuracy, and the number of layers in the backcalculation program. The
selection of the seed moduli controls the convergence of the backcalculation procedure to
pavement moduli that minimizes the mean square error (objective function) between the
measured
deflection
basin
and
the backcalculated deflection basin using the

1
backcalculated moduli. It is known that more than one solution can satisfy the objective
function criterion in backcalculating the pavement moduli due to the multimodal nature
of the backcalculation search space where many local optima exist. In turn, the arrival at
local optima will lead to inaccurate pavement moduli that can be as twice as the
accurate value. On the other hand, the maximum number of layers than can be used in
any backcalculation program does not exceed 5 layers with recommendations to use 3
layers to reduce the error associated with the backcalculation process. In some cases,
increasing the number of the layers in the backcalculation process is desirable to obtain
more representative variation of the moduli with depth.
Genetic algorithms (GAs) can be used to backcalculate the pavement moduli by
searching the entire search space of the objective function using guided random search
techniques. The GAs is based on the Darwinian theory and formulated on the mechanics
of genetics and natural selection (Holland, 1975; Goldberg, 1989). In nature, species
adapt themselves, under the effect of the surrounding environment and conditions, for
survival. In GAs, randomly generated solutions are used to generate a pool of feasible
solutions that evolve from one generation to another using algorithms that mimic the
natural selection in nature such as mutation and crossover (reproduction).
Very few researches have been performed to study the backcalculation of the
pavement moduli using GAs (Fwa et al., 1997; Kameyama et al., 1997; Reddy et al.,
2004; Tsai et al., 2004). These researches were geared toward studying the feasibility of
using the GAs method as a backcalculation procedure and toward finding appropriate
values for the GAs parameters. The researches were successful at pointing out that the
2
GAs method is a robust and accurate method to backcalculate the pavement moduli using
measured deflections. However, the researches failed to select the appropriate GAs
parameters due to the complexity of the search space which is dependent on the forward
calculation procedure, the number of pavement layers, the ranges of the backcalculated
moduli, and the used objective function. In addition, the interaction between the genetic
parameters themselves and the effect of the genetic parameters and operators on the
performance of the backcalculation process add more complexity to the backcalculation
process. Therefore, the GAs method is not used widely since it is not designed, yet, to be
used by practitioners rather than researchers and more research is still ongoing to study
the many factors that affect the backcalculation process.
The lack of enough research and guidelines on how to use the GAs in the
backcalculation of the pavement moduli stimulates the need to develop a robust
backcalculation program that is user friendly and at the same time accurate. This can be
done by identifying the main factors that affect the GAs performance and by studying
thoroughly the performance of the backcalculation procedure. In addition, a successful
procedure that can be used in practice should be easy to use and powerful enough to
eliminate all the drawbacks associated with other classical backcalculation procedures
and programs. The new backcalculation program (BackGenetic3D) is based on the
MultiSmart3D forward calculation program developed by the Computer Modeling and
Simulation Group at the University of Akron (Pan and Alkasawneh, 2006).
1.2 Objectives
The objectives of this research are to:
1) Study the uniqueness and effect of the root mean square error as an objective function
to backcalculate the static pavement moduli.
2) Study the complexity of the search space of the backcalculated moduli.
3) Study the factors that affect the genetic algorithms performance when used as a
backcalculation procedure.
4) Study the interaction between the genetic algorithms parameters and operators.
5) Select a new objective function to enhance the backcalculation performance.
6) Develop a new backcalculation program (BackGeneitc3D) that is user friendly
implementing more accurate genetic operators.
7) Generalize the backcalculation procedure to include any arbitrary number of pavement
layers, loading conditions, loading configuration, and number of sensors.
8) Automate the backcalculation process by eliminating the dependency of the program
on the users inputs (Parameterless Genetic Algorithms-PGA).
8) Develop a new PGA that can be used in the backcalculation of the pavement moduli
and any other genetic algorithms application.
9) Validate the performance and accuracy of the new backcalculation program using
several pavement cases.
1.3 Scope of the Work
The work involved in this study can be grouped into two parts. The first part
includes studying the performance of the genetic algorithms thoroughly by identifying all
4
the factors that can affect the backcalculation procedure. The second part includes the
development of a robust backcalculation program that is both accurate and user friendly.
In addition, the second part includes the automation of the backcalculation procedure by
reducing the number of users inputs and hence eliminating the need for a user who is
well experienced with genetic algorithms.
A literature review on the backcalculation methods of pavement layer moduli is
performed. The limitations and drawbacks of each backcalculation method are identified
and discussed. All of the genetic parameters and operators were studied thoroughly using
different examples. In addition, 3-, 4-, and 23-layer pavement systems were analyzed and
backcalculated to study the mechanism of the backcalculation procedure and the
performance of the genetic algorithms.
A new program was developed using the GAs concepts where new genetic
operators and parameters were introduced for the first time to backcalculate the pavement
moduli. The new program (BackGenetic3D) can be used to backcalculate the static
pavement moduli for any number of layer, loading conditions, and loading configuration.
Current classical backcalculation programs are limited to one circular load, 5-layer
system, and uniform loading only. The BackGenetic3D program assumes that the
pavement layers can be analyzed assuming homogeneous, linear, and elastic materials.
A new powerful Dynamic Parameterless Genetic Algorithm (DPGA) was
developed based on the PGA developed by Harik and Lobo (1999). The new DPGA
eliminates the need for the GAs parameters by automating them. The new method is very
powerful and robust and has been studied using a 23-layer pavement system.
5
1.4 Outlines of the Dissertation

Chapter II presents the literature review on forward and backcalculation analysis
methods of pavement systems assuming homogeneous, linear, and elastic materials. In
addition, the limitations of the backcalculation methods are discussed.
Chapter III studies the uniqueness of the backcalculated pavement moduli using
the root mean square error as the objective function. The effects of the root mean square
error and the backcalculated moduli on the rutting and fatigue in flexible pavements are
investigated as well.
Chapter IV reviews the fundamentals of genetic algorithms and the main used
parameters and operators. In addition, the interaction between the operators and
parameters are discussed.
Chapter V presents the main features and the techniques implemented in the
BackGenetic3D program. The program was verified using results from other classical
backcalculation programs.
Chapter VI introduces a new method to select the genetic parameters. Large and
small population sizes were used to backcalculate the pavement moduli of a 3-layer
pavement system. The performance of the backcalculation process was discussed
thoroughly. In addition, guidelines were provided on the selection of the moduli ranges
for the backcalculation process.
Chapter VII investigates the effect the chromosome length, the fitness function,
the number of layers, and the number of surface deflection points (sensors) on the
performance of the backcalculation process. Examples using 3- and 23-layer pavement
6
systems were analyzed. Guidelines and recommendations were provided as well.

Chapter VIII introduces new innovative methods to automate the backcalculation
process. The Parameterless Genetic Algorithm (PGA) and the Dynamic Parameterless
Genetic Algorithm (DPGA) were introduced and compared using a 23-layer pavement
system. Guidelines and recommendations on the use of the PGA and the DPGA are
provided as well.
Chapter IX presents summaries and conclusions of the research work.
CHAPTER II
PAVEMENT ANALYSIS
2.1 Introduction
Roads are very important in everyday transportation activities. Durability of
pavement and driving convenience are very essential in pavement design and
rehabilitation. The cost of pavement material and design is largely affected by the
material type which can range widely based on the roadway design parameters. A
maintenance-free pavement is still difficult and more research is needed to find an
economical maintenance-free pavement.
Performance and durability of existing and new pavements are very important to
increase the design life of pavements and to reduce the roadway hazards. In the US, a big
portion of the transportation agencies budget is spent every year on pavement
maintenance and improvement. In Ohio, for example, the length of public roads increased
from approximately 117,000 miles to 123,000 miles between 2000 and 2003. However,
the design procedures and practice depended largely on procedures and practices that are
more than 50 years old. It can be argued that the need for more practical and up-to-date
procedures is inevitable these days than ever due to the complexity of roadways design
and the advancement in both technology and behavioral mechanisms of pavements.
New design approaches and procedures have been developed recently by The
National Cooperative Highway Research Program (NCHRP) and the American
Association of State Highway and Transportation Officials (AASHTO) Joint Task Force
on Pavements. The new effort has resulted in replacing the old empirical-based pavement
design procedures with mechanistic-empirical (M-E) based procedures in the new
Mechanistic-Empirical Pavement Design Guide (MEPDG). The M-E approach is a very
powerful approach since it combines the actual observed behavior of pavements and the
analytical modeling techniques that are widely accepted. The M-E procedures are more
reasonable and realistic than the empirical-based procedures since they allow for more
consideration of the effects of the actual site traffic distribution, climate, material types,
structure, and other design features.
Pavement behavior is a key point in the new MEPDG guide. Pavement behavior
and empirical procedures have been developed based on laboratory and full scale field
tests that provide a valuable data base for analytical procedures. Transportation agencies
under the new MEPDG are given more freedom to tailor the design procedures to suit
their roads and pavement conditions.
The M-E procedures can be considered as new design procedures and can be used
to verify the design parameters with the actual parameters including traffic loading,
material properties, and climatic conditions.
Understanding the pavement structural responses along the pavement section and
the load capacity of the pavement requires suitable theories that represent the actual
mechanical behavior. The following discusses some of the available theories that have
9
been used to simulate pavement structural responses.

2.2 Elasticity Theory
Elasticity theory was the first theory to be used for pavement analysis. Applying
this theory to study the structural response of the pavement assumes that the pavement
material is elastic. Elasticity theory works well as long as the stress-strain ratio is
constant. This indicates that the elasticity theory is well suited for pavement sections that
do not undergo stresses greater than the failure stresses. Equations derived from elasticity
theory use the same basic theory but different assumptions for material properties and
geometry (Bendana et. al, 1994). Pavement analysis using the elasticity theory is
normally performed using Hookes law and Boussinesq theory.
2.2.1 Hookes Law
Hookes law is based on the assumption that the stress-strain ratio is constant in
the uniaxial load case of the material. Hookes law was derived assuming that the
material is perfectly elastic and homogeneous.
2.2.2 Boussinesqs Equation
Boussinesq developed equations to compute stresses within a homogeneous,
isotropic, linearly elastic half space under a point load acting perpendicular to the surface.
The half-space assumption indicates an infinitely large area and an infinite depth. The
value of the stress is given by (Holtz and Kovacs, 1981):
10
P (3 / 2 )
Eq. (2.1)
r2 5 / 2
2
z (1 + 2 )
z
where P=point load, z=depth from ground surface to the stress point, r=horizontal
P (3 z 3 )
z =
=
2 (r 2 + z 2 )5 / 2
distance from the point load to the stress point.

Boussinesq developed other equations to compute the state of shear stresses,
normal strains, and displacements under a point load in the elastic material as shown in
Table 2.1 and Figure 2.1. It can be seen from the table that the normal strains,
displacements, tangentional stresses, and radial stresses depend on the Poissons ratio
and/or the modulus of elasticity E while the vertical stress and shear stresses are
independent of the Poissons ratio and the modulus of elasticity.
CL
Figure 2.1, Notation for Boussinesqs equation (Ullidtz, 1987)

11
Table 2.1, Boussinesqs Equations for a Point Load

Normal Stresses
z =
3P
cos 3
2
(2R )
r =
P
1 2
3 cos sin 2
2
1 + cos
(2R )
P
1
(1 2 ) cos +
2
1 + cos
(2R )
1
P
v = ( z + r + t ) =
(1 + )cos
3
(3R 2 )
t =
Shear Stresses
3P
rz =
cos 2 sin
2
(2R )
rt = tz = 0
Normal Strains
z =
r =
(1 + )P (3 cos3 2 cos )
(2ER 2 )
(1 + )P
(2ER )
2
(1 + )P
3 cos3 + (3 2 )cos
1 2
1 + cos
1 2
(2ER )
1 + cos
(1 + )P (1 2 ) cos
v = z + r + t =
(ER 2 )
Displacements
t =
dz =
cos +
(1 + )P (2(1 ) + cos 2 )
(2ER)
(1 + )P cos sin (1 2 )sin
dr =
(2ER)
1 + cos
dt = 0
12
Boussinesqs equation can be extended to other loading conditions such as line

load over a finite area (Newmarks equation; Newmark, 1935). Westergaard (1938)
derived an equation for stresses under a point load in elastic homogeneous half-spaces
with a Poissons ratio equal to zero. His equation is given in Eq. (2.2) where the terms are
similar to the terms in Boussinesqs equation. Both Boussinesqs and Westergaards
equations provide almost the same stress value for r/z1.5, while for r/z<1.5
Boussinesqs equation provides larger values than Westergaards equation (Holtz and
Kovacs, 1981).
Boussinesq (1876, 1885) developed a more realistic theory for stresses in granular
material assuming a variable shear modulus that changes as the confining stresses
change. This assumption resembles actual stresses in materials where the shear modulus
increases as the stresses increase.
Elasticity theory and Boussinesqs equations were studied by other researchers to
check the validity of the theory and the accuracy of the assumptions. Frolich (1934)
showed that the radial stress in Boussinesqs theory is the major principal stress which is
inversely proportional to the square of the distance from the point load. Frolich (1934)
generalized the theory by introducing a concentration factor (n) as shown in the following
equations.
nP
cos n
(2R 2 )
nP
r =
cos n 2 sin 2
2
(2R )
t = 0
nP
R =
cos n 2
2
(2R )
z =
13
Eq. (2.2)
Eq. (2.3)
Eq. (2.4)
Eq. (2.5)
P=
Eq. (2.6)
If the modulus changes with the hydrostatic stress or the major principal stress is
raised to a power factor, two cases arise: The first case is resulted from the assumption
that the modulus changes with the hydrostatic stress which results in a positive power
factor and hence a dilation of the material. The second case is resulted from the
assumption that the modulus changes with the major principal stress which produces a
negative power factor and hence a stress dispersion rather than stress concentration
as proposed by Boussinesqs theory (Ullidtz, 1998).
2.2.3 Charts Method
Boussinesq solution was extended by different researchers to account for more
general loading conditions. Solutions for a circular loaded area can be obtained by
integrating the Boussinesq solution for a point load. In the influence lines method,
solutions for stresses and deflections can be obtained by referring to charts that had been
developed by simplifying the existing conditions under a circular loaded area and by
using the elastic theory as seen in Figure 2.2. These assumptions include neglecting the
Poissons ratio (Foster and Ahlin, 1954), or including the Poissons ratio (Ahlvin and
Ulery, 1962). The Poissons ratio has relatively little effect on stresses and strains in the
half-space elastic theory and therefore can be neglected for simplicity (Huang, 1993).
14
Figure 2.2, Influence lines for a circular loaded area (Holtz and Kovacs, 1981).
2.3 Method of Equivalent Thickness (MET)
This method was first developed by Odemark (Ullidtz, 1987). The MET method
transforms a system with different moduli to a system with one modulus in order to apply
the classical elasticity theories to the system. This method involves two steps in which
the interface plays a key rule for the transformation. In Figure 2.3, stresses and strains
above the first layer can be found by using
the same modulus of elasticity and

15
Poissons ratio for the second layer as in the first layer and assuming a half-space case.
Stresses and strains in the second layer or at the interface can be found by transforming
the first layer into a layer with the same modulus of elasticity and Poissons ratio as in the
second layer but with a new thickness based on the original stiffness of the first layer.
h1, E1,
E1 ,
h 1 , E1 ,
E2,
he, E2,
E2 ,
Figure 2.3, Typical transformations in the MET method, after Ullidtz (1987).
Theoretically, the MET method can be applied to any system with any number of
layers as shown in Figure 2.4.
h1, E1 ,1
h2, E2 ,2
h3, E3 ,
3
:
hn, En
,n
:
Figure 2.4, Typical multilayer system, after Ullidtz (1987).
16
The equivalent thickness of the transformed layer based on the original stiffness
of the layer can be found using the following equation (Ullidtz, 1987):
he , n = f
n 1
i =1
E
hi i
En
1
3
Eq. (2.7)
where he , n is the equivalent thickness for n layers, hi , Ei and En are shown in Figures 2.3
and 2.4, f is the correction factor discussed below.
The MET method is primarily based on the assumption that stress distribution
below the transformed layers will be the same since an equivalent stiffness is used for
each layer. Stresses and strains are assumed to be linear in each layer. However, the MET
method is an approximate method and stresses and strains estimated based on this method
should be corrected to improve the agreement with layered elastic theory. The correction
factor should be estimated based on the number of layers in the system, Poissons ratio,
modulus of elasticity, and layer thickness. In general, the correction factor for the first
interface in a two-layer system is recommended to be 0.9 while it is 1.0 in a multilayer
system. The correction factor for other interfaces in a multilayered system is
recommended to be 0.8 (Ullidtz, 1987). The MET method can overestimate or
underestimate stresses and strains in the layer and at the interface and hence can produce
misleading results unless a good correction factor is available.
2.4 Finite Element Method
Finite Element Method (FEM) has been used to study the responses of flexible
pavements using any material constitutive law. In the FEM method, the geometry under
17
study is discretized into small elements connected by nodes to resemble the actual
geometry or domain. Constitutive laws then applied to govern the behavior of the
material, and stresses and strains can be estimated accordingly. The advantage of the
FEM analysis comes from the ability to simulate any loading condition, static and
dynamic, and any geometrical variation, and local discontinuities such as cracks or joints.
The FEM deals with materials as continuum and therefore it simplifies the actual
behavior of granular material.
Most practical FEM modeling involves only two-dimensional analysis of
pavement sections due to cost, time, and modeling limitations associated with the three
dimensional analysis. Three-dimensional analysis involves the discretization of the
domain using sophisticated meshing techniques that adversely affect the time and cost
needed to study the pavement response. Two-dimensional analysis assumes axissymmetric equilibrium conditions that limit the simulation of the full geometry when
having local discontinuities.
2.5 Discrete Element Method
Advanced numerical simulations assume continuum medium when dealing with
pavement modeling, which indicates a compatibility assumption unless special
considerations are allowed to account for discontinuities such as using elements in the
finite element method.
Granular materials like asphalt can be described using the granular material
physical behavior. Physical behavior of granular material includes normal and shear
forces, and translational and rotational displacements between grains. Strains in granular
18
materials are negligible under normal stresses.

In 1978, Cundall (1978) proposed the Distinct Element Method which uses
gravity, external forces, mass of particles, center of gravity of particles, and moment of
inertia of particles with Newtons laws to describe the movement and interaction between
grains. Discrete element simulation is carried out using steps in which forces and
displacements are calculated for the medium and used as inputs for the next step.
Ullitdz (1995,1998), using circular disks to simulate grains, showed that the
Distinct Element Method produces much larger tensile strains close to the axes of the
load than continuum mechanics. This shows that continuum mechanics has some
limitations when predicting stresses and strains in particulate media. In addition, it was
shown that plastic strains are calculated simultaneously with the resilient or elastic
strains, as well as strains at failure. However, this advanced simulation technique is far
from being available to engineers and is still a research tool.
2.6 Visco-Elasticity Method
Materials deformation can be elastic, plastic, viscous, and viscoelastic. Based on
the physical behavior, these deformations can be categorized as energy-storage processes
and dissipative viscous processes (time-dependent). Viscoelastic behavior is time
dependent while elastic deformations can be time-independent. Plastic materials are
somewhat viscous but for simplicity they can be assumed to be time-independent.
Viscoelastic models are composed of different parts including (Ancey, 2005):
Spring: according to Hookes law, the strain ( ) is proportional to the applied stress ( )
according to:
19
= E
Eq. (2.8)
with E the elastic modulus. In this law, deformations are time independent and elastic
elements represent the possibility of storing energy.
Dashpot: the response of the dashpot, the plunger of which is pushed at velocity is
described by
=
Eq. (2.9)
where ( ) is the viscosity.
The dashpot element represents a dissipative process that occurs as a result of the
relative motion between particles. This motion induces friction when there is contact
between elements or viscous dampening if there is an interstitial fluid.
Viscoelaticity is described by one or more of the above two elements. If the
spring and dashpot are mounted in series, the resulting model is called the Maxwell
model and it is best suited for viscoelastic fluids. If the spring and dashpot are mounted in
parallel, the resulting model is called the Kelvin-Voight model and it is best suited for
viscoelastic solids. These models are elementary models and can be combined to get
more representative models like the Burgers model. These methods can describe some
aspects of the physical behavior, but they still have their own limitations.
2.6.1 Maxwell Model
In this model the dashpot and spring are connected in series, therefore the total
deformation is the sum of both deformations:
d 1 d
=
+
dt E dt
Eq. (2.10)
20
which has the following solution:
(t ) = Ke
Et
E (t '
t )
+ Ee
(t ')dt '
Eq. (2.11)
where K is an integration constant, the lower boundary in the integral is arbitrary. K is

equal to zero when the stress is finite at time t=( ). The following two cases arise from
this equation (Ancey, 2005):
1 for steady state, this equation simplifies to the Newtonian equation = .
2 for sudden changes in the stress, the time derivative dominates.
Equation 2.11 can be written as (t ) =
where t r =
tr
t'
r
tr
'(t ')dt '= (t t ') (t ')dt '
is a relaxation time. The term within the brackets is called the relaxation
modulus and the integral takes the form of a convolution product of (t ) =
t
tr
e
and .
tr
If the above equation is used in the creep testing it will be reduced to:
1 t
+
E
Eq. (2.12)
In the Maxwell model, stresses at time t depend on strains at time t and on the
strain rate at past time t 'but to within a weighting factor that decays exponentially.
2.6.2 Kelvin-Voigt Model
The deformation is described using:
=
where t r =
1 e
t
tr
is the relaxation time.

21
Eq. (2.13)
2.6.3 Burgers Model

This model is a combination of the Maxwell model and the Kelvin-Voight
models. Deformations are described using
t
=
where t r =
2
E2
t
1
1
+
1 e tr +
E1 E2
1
Eq. (2.14)
is the relaxation time.
2.6.4 Creep Testing

Creep testing is used to describe the response behavior of solids. In this test, a
constant stress is suddenly applied to the material and the strain variation over a range of
time is then monitored (Ancey, 2005). Results from the creep test can be used to describe
the three distinct responses of the material during the testing. These responses are: (1) the
immediate elastic response, (2) delayed elastic response (glassy behavior), where the
deformation rate slows with time but becomes steady after long time, (3) the steady state
viscous response (when the shear rate of the material is constant so the material is in
steady flow).
Visco-elastic models are used to describe material responses during creep testing.
Maxwell model is used to describe the immediate elastic response and the steady state
response while Burgers model is used to describe the three responses including the glassy
response.
22
2.7 Layered Systems

It is known that the modulus of elasticity of soil and pavement materials is not
constant but changes as a function of different factors such as the stress level, moisture
content, and temperature (Ullidtz, 1987). On the other hand, the assumption that
pavement and subgrade materials are linear does not resemble the actual conditions.
Depending on the stress level and strains, the physical response of soil and pavement
materials can be categorized as elastic, plastic, viscous, and viscoplastic. Therefore,
modeling the pavement and subgrade using only elastic theory might result in inaccurate
predictions. This problem can be tackled by assuming nonlinear elastic behavior and by
using different techniques to handle the nonlinearity.
2.7.1 Elastic Multilayer Theory
This method was first proposed by Burmister (1943, 1945) as an effort to tackle
the limitations of Boussinesqs method. Burmister (1945) simplified the conditions
between two layers assuming that all layers are isotropic, elastic, and homogeneous. The
top layer, in a two layer system, was assumed to be infinite in extent in the horizontal
direction but of finite thickness in the vertical direction. The bottom layer, in a two-layer
system, was assumed to be of infinite extent in both horizontal and vertical directions. In
addition, Burmister (1945) assumed that the shear and normal stresses outside the limits
of the surface loading are equal to zero. Continuity conditions along the interface
between layers were considered using two cases. In the first case, full continuity of stress
and displacement across the interface was considered assuming full contact between the
two layers and a fully activated shear resistance between the layers. In the second case,
23
continuity of normal stresses and normal displacements was only considered by assuming
a frictionless interface between the layers.
Vertical stresses based on the two-layer theory are shown in Figure 2.5. As it can
be seen from the figure, Burmisters method gave more accurate results than
Boussinesqs method since infinite half-space condition is not applicable in pavement
and subbase layers because the top layer is always of a finite depth. On the other hand,
Burmisters method emphasizes the importance of modeling the interface between the
two layers. Boussinesqs method overestimated the vertical stresses at the interface by
more than 20% compared to stresses from Burmisters method.
Figure 2.5, Burmister two-layer stress curves (Yoder, 1959).
Actual pavement behavior and conditions are different than those used in the
multilayer theory. Pavement materials occupy a finite width and do not have a constant
thickness. Spatial variations of material
geometry and local discontinuities are not

24
considered when using this method. In addition, idealizing the asphalt and soil material as
homogeneous, linear elastic, and isotropic is far from real material behavior. Asphalt has
a viscoelastic behavior while granular materials have a non-linear behavior. On the other
hand, stresses within the pavement section are not considered in the multilayer theory.
Horizontal stresses due to wheel loading and dynamic loads due to road geometry should
be considered to correctly simulate the non-linear and viscoelastic responses, respectively
(COST 333,1999). Anisotropy of pavement materials due to construction methods and
due to the nature of the materials is inevitable and should be considered to account for
stress changes in all directions within the pavement section.
Burmister extended his theory to three-layer systems by deriving the settlement
equation at the surface of the ground only. Based on Burmisters theory, Acum and Fox
(1951) derived a closed form solution for interface stresses under the center of a circular
loaded area. In 1962, Schiffman (1962) developed a solution for multilayer elastic
systems, which was used by many researchers to develop computer programs for
pavement analysis.
2.7.2 Factors Affecting Elasticity

Some of the factors that affect the multilayer elasticity are described below.
2.7.2.1 Anisotropy
Anisotropy has been studied by Van Cauwelaet (1980) assuming a constant ratio
between vertical and horizontal moduli. Van Cauwelaet (1980) derived the following
25
equations to incorporate the effect of anisotropy due to a point load. The equations show
the solutions below the centerline of a uniformly distributed load.
z =
szP
2 (1 s )
z =
szP
2EV (1 s )
dz =
sP
2EV (1 s )
( sz
2 2
+ r2
s2 ( + )
( sz
2 2
+r
+
s z +r
2 2
) (z
3
1
2
+ r2
1 +
) (z
3
+r
Eq. (2.15)
Eq. (2.16)
1 +
Eq. (2.17)
z2 + r2
2
s=
2 2
Eq. (2.18)
where EV is the modulus in compression (vertical), Eh is the modulus in bending

(horizontal), is the Poissons ratio, and = EV / Eh . For loose to medium dense
granular soil with an internal friction angle equal to 30, was found to be 2.25. The
definitions of s, z, and r are shown in Figure 2.1.
2.7.2.2 Shear Sensitivity
Shear sensitive material is a material that has an E/G ratio larger than 2(1+ )
where E is the Youngs modulus, G is the shear modulus, and ( ) is the Poissons ratio.
Misra and Sen (1975) proposed the following equations for stresses and displacements in
shear sensitive materials with a typical value of 6 (Misra and Sen, 1975) for E/G. The
solutions for a point load are:
z =
zP
2 ( )
(z
2 2
26
+ r2
) (z
3
2 2
+ r2
Eq. (2.19)
z =
(1 + ) zP
2E ( )
dz =
(1 + ) P
2E ( )
2A
(z
2 2
+r
z +r
2 2
2B
) (z
3
2 2
+r
Eq. (2.20)
Eq. (2.21)
z + r2
2 2
and for a uniform distributed circular load are
z =
o
z
z
+

2z2 + a2
2z2 + a2
Eq. (2.22)
z =
(1 + ) o
B
( )E
dz =
(1 + ) o
A 2 z 2 + a 2 z B 2 z 2 + a 2 z
( )E
((
z
2z2 + a2
1 A
) (
z
2z2 + a2
Eq. (2.23)
))
Eq. (2.24)
where 2 and 2 are the roots of the following equations.

x 2 + ( K 2) x + 1 = 0 where
E
1
(1 + )G
1 K
K =
1
K 2
A= 2
1
K 2
B= 2
1
K=
2.8 In-situ Deflection Measurements

Nondestructive testing is important to measure or describe the material
characteristics with minimal damage. In pavement, nondestructive testing plays a major
role since cracks and mechanical damage are minimized when pavement cores are
reduced. In addition, taking cores and
samples from roads that cover many miles

27
can be costly and hence the nondestructive testing provides means by which time and
cost can be reduced.
Many systems are available to measure deflections of pavements to back
calculate the elastic modulus (E), the foundation support (k or CBR), and the degree of
load transfer. Most popular pavement deflection systems include the Benkelman Beam,
the Dynaflect, and the Falling Weight Delectometer.
2.8.1 The Benkelman Beam
This is a classical system to measure pavement deformations due to a static load.
Figure 2.6 shows the Benkelman Beam (Bendana et al., 1994).
Figure 2.6, The Benkelman beam, after Bendana et al. (1994).

2.8.2 The Dynaflect System
In this system, sinusoidal vibrations are transmitted to the pavement through steel
wheels, and then vertical velocities of the pavement surface at various positions are
measured using geophones. Vertical velocities are converted to deflections and the
28
measured deflections (deflection basin) can be estimated. Figure 2.7 shows the Dynaflect
system (Bendana et. al., 1994).
Figure 2.7, The Dynaflect system, after Bendana et al. (1994).

2.8.3 The Falling Weight Deflectometer (FWD)
This system was first introduced 30 years ago in France to test flexible pavements
(Ullidtz, 1987). In this system, impact loads (impulse forces) are applied to a loading
plate then the vertical velocities of the pavement surface are measured at different
locations using geophones. In the LTPP deflection-testing program, geophones are placed
at 0, 203, 305, 457, 610, 914, and 1524 mm from the center of the loading plate (Von
Quintus and Simpson, 2002). FWD systems, normally, use seven to nine geophones
within a distance less than 2 meters with the first sensor placed below the center of the
loading plate.
Impulse forces can be varied by adjusting the height of the dropped weights. The
pulse load is transferred to the plate through a set of springs and is applied within a short
loading time (28 milliseconds). Other values that are measured besides the vertical
velocities include the transit time and the duration. Vertical velocities are converted
29
to deflections and the deflection basin can be estimated. Heavy FWD (HFWD) can be
used to simulate one wheel of a fully loaded Boeing 747 by producing a maximum
instantaneous dynamic force of up to 250 kN in a short loading time between 20 and 25
milliseconds. Figure 2.8 shows a typical configuration for the FWD test.
0.305 m.
0.203 m.
0.102 m.
0.152 m.
0.152 m.
0.305 m.
0.610 m.
Loading Plate Contact Area

Sensor
Figure 2.8, FWD typical sensors configuration.
The FWD is believed to generate the most realistic loading function that can be
generated by a nondestructive test to simulate actual moving wheel loads (Hoffman,
1983). In addition, Lytton et al. (1986) rated the FWD as the best overall pavementtesting device while Bentsen et al. (1989) commented on the versatility and the time
effectiveness of the FWD device.
30
Pavement responses are measured using a deflection basin which can be created
by measuring the responses using uniformly or non-uniformly distributed sensors. More
representative basins can be created by placing more sensors around the point where the
impulse force is applied. Deflection basin magnitude and shape are dependent on the
layer thickness, Poissons ratio, layer moduli, moduli ratio, and the depth to the stiff layer
or the bedrock (Bendana et. al., 1994). Geophones close to the point where the impulse
force is applied (inner geophones) are used to measure the composite response of the
whole pavement layers. Geophones far from the point where the impulse force is applied
(outer geophones) are used to measure the response of the deeper layers. It is a common
practice to normalize response data (deflections) to make test results adaptation to other
loading conditions easier.
The backcalculated modulus of elasticity, from deflections, is very sensitive to the
measured deflections. Calibration of small deflection errors on the order of 1m-2m
will affect the modulus of elasticity significantly (Bendana et. al., 1994).
Stress distribution under pavements is not uniform and controlled by many factors
such as the subgrade material, compaction, pavement material, moisture content, and
temperature. The actual stress distribution below the pavement can be determined by
measuring pavement deflections at different distances from the center of an applied load.
Actual stress distributions under pavements are given in Figure 2.9 (Ullidtz, 1987) where
the surface pressure distribution is shown as p while the diameter of the loaded are is 2a.
31
Figure 2.9, Actual stress distribution under pavement systems, after Ullidtz (1987).
In plate load testing, if the distance from the center of the circular plate is larger
than twice the radius of the plate, then, practically, the distributed load can be treated as a
point load and hence Boussinesqs theory can be used to estimate the modulus of
elasticity based on surface deflections as described in the following equation
E=
P (1 2 )
rd (r ,0)
Eq. (2.25)
where d (r ,0) is the surface deflection at distance r from the center of the load.
If the modulus of elasticity of the subgrade is not identical at different distances
from the applied load, then the assumption that the subgrade is a linear elastic half-space
is not valid and either the subgrade is non-linear elastic or it consists of different layers
32
(Ullitdz, 1987). Another reason for non-linearity is the presence of a rigid layer at some
depth. Elasticity theory suggests that the deflection at a depth equal to the distance from
the load is almost equal to the deflection at the surface (Ullitdz, 1998) according to Eq.
(2.25).
However, Bodare and Orrje (1985) indicated that the difference between the
theoretical dynamic modulus from the FWD and the theoretical static modulus for a
homogeneous half-space is minimal when the time to peak load is relatively long. The
work of Bush (1980), Roesset and Shao (1985), and Mamlouk (1985) indicated that the
difference between the measured dynamic deflections and the predicted static deflections
is minimal when a rigid elastic layer is placed at a finite depth below the pavement
system.
2.9 Elastic Multilayer Programs

Elastic multilayer programs based on elasticity theory can be divided into forward
calculation programs and back-calculation programs. In the forward calculation
programs, pavement response is estimated based on known pavement parameters
including the modulus of elasticity. In the back-calculation programs, inverse techniques
are applied to core forward calculation procedures to estimate the modulus of elasticity
based on measured deformations.
33
2.9.1 Forward Calculation Programs

Different programs have been developed for multilayer elastic analysis including
the following:
1- MMOPP: The Mathematical Model of Pavement Performance (MMOPP) was
developed by Ullidtz (1987). This program is primarily a research tool and has not yet
been used in practice. The program can be used to predict longitudinal roughness, rutting
and cracking of a pavement consisting of an asphalt surface course, a granular base layer
and subgrade (Danish Road Institute, 2001). Different parameters can be used to study
the pavement response including the variation of pavement layer thickness, elastic
stiffness, plastic parameters and dynamic load variations along the length of the road.
2-CHEVRON: This program was written based on the work of Michelow (1963). The
program capabilities include computing vertical, tangential, radial, and shear stresses, and
vertical and radial strains. Programs limitations included the incapability of computing
displacements and principal stresses. The program was limited to 5 layers in its early
stages then was expanded to fifteen (15) layers in 1967 (Ullitdz, 1987). The CHEVRON
model was used by the Asphalt Institute in 1991 to develop the Mechanistic-Empirical
Thickness Design Manual (Asphalt Institute, 1991).
3-ELSYMS: This program was built based on CHEVRON with more capabilities. These
capabilities included the ability to compute displacements and principal stresses and the
ability to handle multi-wheel loads.
4-BISTRO: This program was written based on the work of Schiffman (1962). The
program capabilities included computing stresses, strains, and displacements, assuming

34
rough interfaces between layers. The program can account for multi-wheel loads.
5-BISAR: This program was built based on BISTRO with more capabilities. The
program can consider the tangentional surface loadings and include layer interface. This
program was developed by Shell International Petroleum Company Limited and used in
their Pavement Design Manual (Shell, 1978).
6-MultiSmart3D: This program was developed by the Computer Modeling and
Simulation Group at the University of Akron (Pan and Alkasawneh, 2006). This program
is based on the innovative computational and mathematical techniques for multilayered
elastic systems (e.g., Pan, 1989a,b, 1990, 1997). The MultiSmart3D program uses the
propagator matrix method to solve only two 22 systems of linearly algebraic equations
no matter how many layers are in the pavement system. In addition, adaptive Gauss
quadrature is used along with an acceleration approach for fast and accurate calculation
of integrations. The program is capable of calculating the pavement responses at any
arbitrary number of points at any location in the pavement system due to any arbitrary
loading configuration (uniform or nonuniform loading) and shape (regular or irregular
shapes). Moreover, the program can be used for any arbitrary number of layers with any
thickness.
2.10 Back Calculation of Elastic Moduli

Elastic modulus is an important property of pavement materials. Different
methods have been proposed by researchers to estimate the elastic modulus based on
laboratory bending tests and empirical equations (Bonnaure et al., 1977), wave
35
propagation methods (Avramesco, 1968; Heukelom and Foster, 1960; Jones, 1963, 1965;
Jones and Thrower, 1965; Szendrei and Freeme, 1970; Cho and Lin, 2001), and the
Falling Weight Deflectometer (FWD).
Elastic modulus cannot be determined accurately based on empirical equations
since actual field conditions, loading conditions, and traffic conditions vary along the
same section. In addition, some of the input parameters in the empirical equations cannot
be determined for an existing pavement while it can be determined during the mix design
stage of the pavement. For practical purposes, elastic moduli need to be determined on
site to measure the interaction and effect of different factors.
Wave propagation methods are non-destructive insitu testing methods used to
determine the elastic modulus and the thickness of the pavement. These methods are
based on measuring the phase difference between the vibration source that transmits
sinusoidal stress waves of a frequency set into the test object, and the vibration of the
receivers, which are placed at different distances from the source (Martincek, 1994). The
vibration sources, usually, have small excitation energy which makes the test nondestructive as the resulting dynamic stresses are low and cannot affect the state of the
pavement.
2.10.1 Classical Backcalculation Methods

Methods for the backcalculation of the pavement moduli using data from
nondestructive testing can be grouped into three general methods (Anderson, 1988): (1)
simplified methods, (2) gradient relaxation methods, and (3) direct interpolation methods.
36
These methods rely on some assumptions and simplifications that should be made to
facilitate the backcalculation process. For example, the thickness of each layer is assumed
to be constant (uniform) and normally is obtained based on the construction records or
pavement coring. The main steps of the backcalculation procedure consists of:
1-Define the input parameters of the pavement system including: thickness of each layer,
Poissons ratio, etc.
2-Assume moduli seed values for the pavement system. Seed moduli values can be
assumed based on experience or based on typical moduli values. Moduli values can be
different based on the forward method implemented in the backcalculation program.
3-Calculate the pavement deflections, using the forward program, at the FWD geophone
locations (along the surface).
4-Compare the calculated deflections with the measured deflections. If the difference
between the calculated and measured deflections is acceptable, then the assumed layer
moduli are the actual moduli. Otherwise, the assumed layer moduli are not the actual
moduli and the assumed moduli should be refined.
5-Repeate steps 2 through 4 if necessary.
2.10.1.1 Simplified Methods

The simplified methods are backcalculation methods that depend on the
equivalent thickness and the deflection geometry methods as the forward calculation
methods. Ullidtz (1973) developed a backcalculation computer program named ELMOD
based on the equivalent thickness method while Van der Loo (1982) developed another
37
program using the method of surface curvature to backcalculate the layer moduli. These
programs have the same drawbacks that are associated with their forward calculation
methods. Therefore, both programs are not widely used due to reliability issues.
2.10.1.2 Gradient Relaxation Methods

The gradient relaxation methods are the most common method for the
backcalculation of pavement moduli. The method was first proposed by Michelow
(1963). Other programs developed later including BISDEF, ELSDEF, and CHEVDEF
(Bush, 1980; Bush and Alexander, 1985), which use BISAR, ELSYM5, and CHEVRON,
respectively, as their forward calculation programs.
The main idea is to form gradient matrices then to solve these matrices for
solutions that give the minimal difference between calculated and measured deflections.
The relation between the predicted deflection j at sensor j and the layer moduli (E) for a
pavement system of n layers is given by (Anderson, 1988):
j = f ( E1 , E2 , E3 ,..., En )
Eq. (2.26)
The deviation in deflection at sensor j between the measured ( D j ) and the predicted
deflection is:
j = D j j = f ( E1 , E2 , E3 ,..., En )
Eq. (2.27)
The sum of the squares of all sensors deviations is:

m
j =1
j2 =
m
j =1
[D j f ( E1 , E2 , E3 ,..., En )]
where m is the number of sensors (deflection points).

38
Eq. (2.28)
The above equation, then, is derived with respect to the unknown moduli (partial
derivatives) to form a solution matrix of size n equation. Then, the partial derivatives are
approximated numerically by forming gradient equations. The gradient equations are
formed by calculating the deflections of an initial set of moduli ( E o ). Another set of
moduli values ( E1 ) is assumed where the moduli values are all the same as in set ( E o )
but with one of the moduli changed. Then, the deflections of the second set are
calculated. The deflections at sensor j as a function of the gradient equation can be
estimated using:
Eq. (2.29)
j = A ji + S ji log10 ( E i )
o
where Ei is the unknown modulus of layer i, S ji =
i j
log10 Ei log10 Ei
, Aji = i S ji Ei ,
1
Ei is the assumed modulus for layer i in the initial set of moduli, Ei is the assumed
o
modulus for layer i in the second set of moduli, i is the calculated deflection at sensor j
o
for Ei , i is the calculated deflection at sensor j for Ei .

The deflection at sensor j as a function of all unknown moduli values Ei :
o
j = j +
n
i =1
S ji log10 Ei log10 Ei
Eq. (2.30)
Then, j can be estimated using one of the unknown moduli (e.g., modulus of layer
number 3):
o
j = A j 3 + S j 3 log10 E3
Based on Eqs. (2.29) and Eq. (2.30), Eq. (2.31) can be written as:
39
Eq. (2.31)
j = Aj 3 + S j 3 log10 E3 +
n
i =1
Eq. (2.32)
Then, Eq. (1.29) becomes:

n
j =1
j =
2
m
j =1
D j A j 3 S j 3 log10 E3
i =1
Eq. (2.33)
The squared errors in deviation in the above equation can be minimized by taking
the partial derivative with respect to each of the unknown moduli, then setting the partial
derivatives equal to zero.
Then, the following matrix should be solved:
[B]{E}={C}
Eq. (2.34)
where:
Ck =
m
j =1
S jk D j A j 3 S j 3 log10 E3
Bki =
m
j =1
3
i =1
S ji log10 Ei
S jk S ji
Eq. (2.35)
Eq. (2.36)
where i, k are the layer numbers, and Ek is the unknown layer moduli associated with
equations (2.34) through (2.36).
The above procedure is dependent on the forward calculation method. Increasing
the number of layers will increase the time spent searching for the solution.
40
2.10.1.3 Direct Interpolation Methods

This method was developed by Uzan et al. (1988) based on the computer program
MODULUS (Uzan, 1986). In this method, the user selects the structural response model
and then the moduli are backcalculated using direct interpolation from a set of solutions.
The method requires the existence of a table of pre-calculated responses for a wide range
of moduli combinations. Therefore, the accuracy of the program can be limited since the
accuracy of the interpolation will be dependent on the number of data points and the
number of cases considered to create the table.
2.10.2 Elastic Multilayer Backcalculation Programs
Most backcalculation programs apply iterative and inverse techniques on the
multilayer elastic theory to estimate the modulus of elasticity. As a minimum, back
calculation methods using multilayer elastic theory require the thickness of the layer and
the Poissons ratio to be known. Typical values of the Poissons ratio and the modulus of
elasticity for different pavement materials are given in Table 2.2 (AASHTO, 1993) and
Table 2.3, respectively. Guidelines for backcalculating the elastic properties of pavement
from the FWD test are described in ASTM D5858.
41
Table 2.2, Typical Poissons ratios of pavement materials,

modified after (AASHTO, 1993).
Material
Portland Cement
Concrete
Hot Mix Asphalt/
Asphalt Treated
Base
Cement Stabilized
Base
Range
Remarks
0.10-0.20
0.15-0.45
0.15-0.30
Granular
Base/Subbase
0.30-0.40
Subgrade Soils
0.30-0.50
Typical
Value
0.15
Depends on temperature. For temperatures

less than 30C use 0.15; for temperatures
higher than 50C use 0.45.
When sound free of cracks use 0.15; with
cracks use 0.30
For crushed material use low values; for
unprocessed rounded gravel/sands use
high values
For cohesionless soils use values near
0.30; for very plastic/cohesive clays use
0.50.
0.35
0.20
0.35
0.40
Table 2.3, Typical values of modulus of elasticity for pavement materials, after
(AASHTO, 1993).
Range
Typical Value
Material
(MPa)
(MPa)
Hot-Mix Asphalt
1,500-3,500
3,000
Portland Cement Concrete
20,000-55,000
30,000
Asphalt-Treated Base
500-3,000
1,000
Cement-Treated Base
3,500-7,000
5,000
Lean Concrete
7,000-20,000
10,000
Granular Base
100-350
200
Granular Subgrade Soil
50-150
100
Fine-Grained Subgrade Soil
20-50
30
Methods for back calculating the elastic modulus of a layered pavement system
can be divided based on the inverse approach, as the following (von Quintus and
Simpson, 2002):
1-The equivalent thickness method (e.g. ELMOD and BOUSDEF programs).
2-The optimization method (e.g. MODULUS and WESDEF programs).
42
3-The iterative method (e.g. MODCOMP and EVERCALC programs).

The following is a brief description of some of the available back calculation
programs:
1- MODCOMP 3: This program is used to back calculate the modulus of elasticity of
pavement layers based on the work of Irwin (Irwin, 1983, 1992). The modulus of
elasticity is back calculated using iterative methods where the modulus is adjusted until it
matches the measured deflections from the geophones (sensors). The program can handle
data from up to ten deflection sensors at up to six load levels. The maximum number of
layers with unknown moduli can be up to twelve (12) layers with a recommendation of
using five or six unknown moduli only.
2-WES5: This program is a linear elastic program that was developed by the Corps of
Engineers. This program uses algorithms that make the calculations faster than many
other programs and has no copy right restrictions.
3-MODULUS: This program was developed in 1986 (Uzan, 1986) based on the Corps of
Engineers WES5 program. The program can handle up to five layers and can be used to
estimate the optimal number of sensors to be used in the backcalculation process. In
addition, the program can, automatically, estimate the depth of the stiff layer (rock) and
the weighted factors for each layer. The program uses Hooke-Jeeves pattern search
algorithm which provides a definite convergence for the backcalculated values.
4-BISDEF: This program is based on the calculation process in the BISAR program
(Bush, 1985). The program uses the iterative gradient relaxation method to backcalculate
the moduli. Sensitivity analysis showed that the program is sensitive to the initial moduli
43
values (seed moduli) and in many cases did not converge (Anderson, 1988).
5-ELSEDEF: This program is based on the calculation process in ELSYM5 (Lytton et.
al., 1990).
6-CHEVDEF: This program is based on the calculation process in CHEVRON (Warren
and Diekmann, 1963).

7-COMDEF: This program was developed in 1989 (Anderson, 1989) to back calculate
the modulus of elasticity using a stored database of deflection basins generated with
CHEVRON. In this method, a set of pavement system parameters is selected, including
the layer thickness and the modulus of elasticity for each layer, and the deflection basin is
calculated for that set and stored in a database. Then the measured data from the in-situ
deflection testing are compared with those in the stored database to estimate the required
values. However, the program did not work well with cases that involved thick layers or
very shallow deflection basins (small deflections) making the program inappropriate to
some of the airfield pavements due to the limitations of the database.
8-ILLI-BACK: This program was developed to back calculate the slab modulus and the
subgrade reaction modulus for a two-layer rigid pavement system (Ioannides et. al.,
1989). The program uses the area of the deflection basin and the maximum deflection as
inputs to Westergaards deflection equation assuming a unique relationship between the
area and the radius of relative stiffness of the pavement system.
2.10.3 Limitations of The Backcalculation Programs
Back calculation programs suffer from different limitations based on the
assumptions and simplifications used to develop the program which reflects, in most
44
cases, the limitations of the core multilayer elastic programs. One way to overcome these
limitations was proposed previously (Von Quintus and Killingsworth, 1997) using
correction factors based on comparisons between laboratory-measured and backcalculated elasticity moduli. However, these correction factors were dependent on the
backcalculation program and hence should be used carefully.
Some of the limitations observed in the backcalculation process carried out in
different studies (Von Quintus and Simpson, 2002; George, 2003) are listed below.
1-Many programs are limited by the number and the thickness of layers used in the
analyzed pavement system.
2-Many programs assume a linear elastic behavior with few considering the non-linear
behavior. Unfortunately, programs that account for viscoelastic (time-dependent) or
elastoplastic (inelastic) behavior of materials have not been successful in estimating
consistent and reliable results.
3-Actual modulus of elasticity of the layer is not calculated because many factors
affecting the modulus cannot be included in the current used methods. These factors
include stress sensitivity, discontinuities and anomalies such as variations in layer
thickness, localized segregation, cracks, and the combination of similar materials into a
single layer. Therefore, the back calculated modulus is known as the effective or
equivalent modulus.
4-Most programs use static back calculations whereas the applied load in the FWD is a
dynamic load. Standardized back calculation procedures are available only for the static
case (ASTM D5858) while the dynamic case is more complex.
45
5-A large number of iterations is required, if the model is based on the linear elastic
assumption, at varying load levels (different drop heights) to identify the stress sensitivity
of unbound pavement materials and soils. In other words, forward analysis (deflection
calculation) of pavement systems using linear response models is relatively fast while
back calculation using linear models to estimate the nonlinear properties require
relatively large number of iterations.
6-Most programs are developed based on quasi-nonlinear response models. This indicates
that the modulus of elasticity for a layer is the same in the horizontal as well as in the
vertical directions even if the state of the stress within the layer is changing. Variation of
the modulus of elasticity within the pavement layer has not been considered. In addition,
the modulus can vary by changing the load level.
7-Solution convergence is not guaranteed for many programs. The starting modulus (the
seed) should be adjusted based on the users experience and judgment to arrive at a
reasonable modulus that fit the elastic modulus range for the layer and the desired root
mean square error (RMSE). This makes the backcalculation process highly dependent on
the user rather than completely automated.
8-Accuracy of the back calculated properties using elastic models vary with depth.
Accuracy of the back-calculated results may vary between 1 percent to 20 percent with
the high value for the surface layer and the low value for the subgrade layer.
9-Most programs are unable to model plastic flow or lateral movement of the underlying
materials close to the surface. This limitation can result in the back calculation of
unrealistically high elastic moduli.
46
10- Modulus of elasticity is stress dependent in coarse and fine-grained unbound

materials. Hot mix asphalt (HMA) layers are assumed to be linear elastic and therefore
they are stress independent. However, backcalculation programs show stress sensitivity
(stress dependent elasticity modulus) that is dependent on the drop height in the FWD
test. This contradiction in behavior is attributed to limitations in the backcalculation
process which ignores the interface condition near the surface and the inability to
simulate any damage in the bound surface layers rather than the true stress sensitivity of
the layer.
11-Subdividing the subgrade into more than two layers can significantly improve the
matching between the measured and calculated deflection basins resulting in low root
mean square errors. However, most programs have limitations regarding the maximum
number of layers.
12-All back calculation programs are sensitive to the rigid layer location below the
subgrade/subbase interface. However, the SHRP recommended (SHRP-P-655, 1993)
placing the rigid layer at a depth of 50 feet below the surface if no information is
available.
13- The backcalculated moduli is not unique. Different backcalculated moduli can be
obtained for the same root mean square error.
2.10.4 Backcalculated Resilient Modulus versus Lab Testing
The difference between backcalculated subgrade resilient modulus from the FWD
and laboratory resilient modulus has been studied by many researcher (Daleiden et al.,
1994; Akram et al., 1994; Nazarian et al., 1995; Von Quintus and Killingworth, 1997;
47
Seeds et al., 2000; George, 2003).

It was found by Daleiden et al. (1994), Akram et al.(1994), Seeds et al. (2000)
and Nazarian et al.(1995) that there is no unique relation between the backcalculated
resilient modulus and the laboratory measured resilient modulus. The ASSHTO Guide
(1993) suggested that the backcalculated modulus is three times the laboratory modulus,
while Von Quintus and Killingworth (1997) suggested using factors, to match the
backcalculated and the laboratory moduli, developed using a multi layer elastic program.
However, the suggested correction factors are highly dependent on the backcalculation
program and should be used carefully.
Besides the spatial variations and nonlinearity of materials that contribute to this
contradiction, the FWD test measures deflections due to stresses in 3D directions while
deflections in conventional multilayer back calculation programs are assumed to be in 2D
instead. Conventional laboratory testing procedures estimate the modulus based on
measurements in the vertical direction only. Programs that consider the 3D effect and the
variation of the resilient modulus in both the horizontal directions (anisotropy) should be
developed to overcome such problems.
Variation of the modulus of elasticity in subgrade material can be observed even
if the subgrade contains one type of material. This variation can be observed when the
moisture content of the material varies with depth within the subgrade (George, 2003). In
the backcalculation as well as the forward calculation procedures, the assumption that the
subgrade layer has a constant modulus with depth, is based on the need for simplification
since most programs have limitations on the number of layers.
48
Another reason for anisotropy in subgrade material is compaction. It is a common

practice to prepare the subgrade material by compacting the subgrade lifts during
construction, which can be from 6 to 12 inches based on the used equipment and
material. Compaction energy applied to each layer will change the stress distribution and
hence will change the modulus of elasticity in both directions resulting in an anisotropic
state.
2.11 Conclusions
Elasticity theory has been used for many years by engineers to analyze pavement
responses due to traffic loading due to its simplicity and cost effectiveness. In addition,
field and lab test data showed that the elasticity theory can provide accurate results and
can be practical more than other methods since less controlling parameters are needed.
The elasticity theory was further implemented in computer programs to facilitate the
analysis of pavement systems. These computer programs are limited to a certain number
of pavement layers due to the limitations associated with the mathematical formulation of
the closed form solutions. The maximum number of layers that can be used in pavement
programs is 20 layers. This limited number of layers limits the modeling of pavement
systems where temperature variation with depth is observed. Therefore, the need for a
more robust and flexible forward calculation (calculates response form current known
moduli) program should be a priority.
However, the backcalculation of the pavement layer moduli, using the available
pavement programs and backcalculation techniques, showed many drawbacks. These
drawbacks include the difficulty of finding the optimal solution and the possibility of
49
having more than one set of moduli for the same set of pavement deflections. This
difficulty stems from the fact that seed moduli are used in the backcalculation process
leading to a local solution rather than a global solution for the problem. In addition, most
backcalculation techniques are not computationally effective or robust leading to
divergence (no solution is obtained) in some cases. Therefore, the development of a
robust backcalculation program (calculates moduli from known deflections) that can be
used for a wide spectrum of pavement loading and modeling conditions is vital.
50
CHAPTER III
EFFECT OF THE ROOT-MEAN-SQUARE ERROR
ON PAVEMENT RESPONSE
3.1 Introduction
The elastic modulus is an important property of pavement materials. Different
methods have been proposed by researchers to estimate the elastic modulus based on
laboratory bending tests and empirical equations (Bonnaure et al., 1977), wave
propagation methods (Cho and Lin, 2001), and the Falling Weight Deflectometer (FWD).
The elastic modulus in pavements cannot be determined accurately based on
empirical equations since actual field conditions, loading conditions, and traffic
conditions vary. In addition, some of the input parameters in the empirical equations
cannot be determined for an existing pavement. Therefore, for practical purposes, elastic
moduli need to be determined using in-situ methods to include the effect of different field
factors.
The FWD test is currently the most widely used nondestructive method in
pavement engineering. The test involves applying impact loads (impulse forces) to a
loading plate and measuring the vertical displacement of the pavement surface at
different locations using velocity sensors. The FWD system normally uses seven to nine
51
geophones within a distance less than 2 meters with the first sensor below the center of
the loading plate. The measured deflections from the FWD test along the pavement
surface are then utilized to backcalculate the modulus of elasticity in each layer. This
method, however, suffers from different limitations since backcalculating the modulus of
elasticity does not always ensure an accurate estimate of the modulus where a seed
modulus is required for each layer in the backcalculation procedure. Therefore, the
backcalculation of the elastic moduli does not provide a unique solution and in many
cases is user-dependent. It is common to assess the accuracy of the backcalculated
elasticity moduli by assessing the accuracy between the measured deflections from the
FWD and the calculated deflections using the backcalculated set of elastic moduli.
In this study we investigate the current practice of rejecting and accepting the
backcalculated elasticity moduli based on a tolerance value between the measured and
calculated deflections.
3.2 Error Tolerance in The Backcalculation Procedure
The backcalculation of elasticity moduli is commonly carried out by assuming a
set of pavement-layer moduli (seed moduli) that can produce a deflection basin similar to
the measured one from the FWD test. In order to minimize the error between the
measured and calculated deflections, the relative root-mean-square error (RMSE) is used
to control the convergence of the backcalculated deflections and to assess the acceptance
and rejection of the final set of pavement moduli. The RMSE is computed by:
52
RMSE =
1
n
n
i =1
d i Di
Di
100%
Eq. (3.1)
where, RMSE is the relative root-mean-square error, n is the total number of the
deflection measurement points, di is the backcalculated deflection at point i, and Di is the
measured deflection at point i. When the RMSE value decreases, the accuracy of the
backcalculated elasticity moduli is assumed to increase as the error between the measured
and calculated deflections decreases.
In the Long Term Pavement performance (LTPP) test sections, an RMSE of 3%
was used as an acceptable error (Von Quintus and Simpson, 2002). In addition, Von
Quintus and Simpson (2002) showed that the selection of a 2% RMSE does not
necessarily result in convergence in the backcalculated elasticity moduli in all cases. In
general, their results indicated (Von Quintus and Simpson, 2002) that RMSE values less
than 3% have little effect on the average backcalculated elastic moduli. In practice,
RMSE values larger than 1% can be achieved quickly (Harichandran et al., 1994).
Therefore, the most commonly used value for the RMSE is between 1% and 3%.
However, it is believed that achieving lower RMSE will always enhance the
backcalculated elastic moduli and therefore more accurate results can be obtained.
3.3 Backcalculation Study
The majority of the available research investigates the difference between the
RMSE values from different backcalculation programs (Fwa et al. 1997), the effect of
other factors on the quality of the deflection data (Mehta and Roque 2003), the effect of
the seed generation on the RMSE values
(Fwa and Rani, 2005), and the effect of

53
other factors on the FWD data. To the best of the authors knowledge, however, the effect
of the backcalculated elastic moduli and the associated RMSE on the strain and stress
responses in flexible pavements has not been discussed so far.
In this chapter, to study the effect of the RMSE on the strains and stresses in
flexible pavements, a three-layer pavement section was selected. The flexible pavement
section and the backcalculated elastic moduli were reported by Anderson (1988) and
were shown in Tables 3.1 and 3.2. In this study, the responses and the associated error at
a total of 11 points (as shown in Tables 3.3 and 3.4) were calculated for each case:
responses along the ground surface at common locations of the velocity sensors used in
the FWD test along the pavement profile (points 1 through 7), at the middle of the AC
layer (point 8), at the bottom of the AC layer (point 9), at the middle of the base layer
(point 10), and at the top of the subgrade (point 11). For all cases, responses were
calculated for a circular load with a radius of 150 mm and a pressure of 690 kPa. The
Poissons ratio for all layers was equal to 0.35.
The pavement response was calculated using the MultiSmart3D program. The
MultiSmart3D program is a fast and accurate software tool developed by the Computer
Modeling and Simulation Group at the University of Akron, and it is based on the
innovative computational and mathematical techniques for multilayered elastic systems
(Pan 1989a,b; Pan 1990; Pan 1997). The program is capable of analyzing any pavement
system regardless of the number of layers, the thickness of each layer, the number of
response points, and the shape of the applied pressure at the surface of the pavement.
54
The RMSE for each case was calculated using Eq. 3.1 and shown in Table 3.3. In
addition, Table 3.3 shows the relative errors using the exact and backcalculated elastic
moduli. The relative error is defined as:
RE =
Exact Calculated
x100%
Exact
Eq. (3.2)
This definition of the relative error is also used to compare the calculated strains and
stresses based on the exact and backcalculated models. In other words, Calculated is the
response using the backcalculated elastic moduli, and Exact is the response using the
exact elastic moduli.
Pavement responses from the exact and backcalculated elastic moduli (Cases 1
through 5) are shown in Tables 3.5 through 3.10. In addition, Tables 3.5 through 3.10
show the relative errors associated with each response.
Case
Case 1
Case 2
Case 3
Case 4
Case 5
Table 3.1, Parameters of the Original (Exact) Layers.

AC
Base
Exact Modulus (MPa)
Thickness Thickness
(mm)
(mm)
AC
Base
Subgrade
381.0
381.0
381.0
381.0
381.0
152.4
152.4
152.4
152.4
152.4
3447.3785
55
13789.5140
68.9476
Table 3.2, Parameters of the Backcalculated Layers.

Backcalculated Modulus (MPa)
Case
AC
Base
Subgrade
Case 1
Case 2
Case 3
Case 4
Case 5
3506.1839
3359.2980
3656.5723
3410.8018
3657.1101
13507.0427
15717.2743
12088.8670
13371.5676
12017.7062
69.0096
68.6304
69.4302
68.7063
69.6715
Table 3.3, Relative Errors in Moduli and RMSEs using the Exact (Table 3.1) and
Backcalculated (Table 3.2) Moduli.
AC
Base
Subgrade
RMSE
Case
(%)
(%)
(%)
(%)
1
2
3
4
5
1.71
2.55
6.07
1.06
6.08
2.05
13.98
12.33
3.03
12.85
0.09
0.46
0.70
0.35
1.05
0.24
0.22
0.65
0.69
0.85
Table 3.4, Coordinates and Locations of the Response Points.

x
y
Z
Point
(mm)
(mm)
(mm)
Location
1
2
3
4
5
6
7
0.0
304.8
609.6
914.4
1219.2
1524.0
1828.8
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
190.5
0.0
0.0
380.9
10
0.0
0.0
457.2
11
0.0
0.0
533.5
56
AC Surface
Middle of AC
Layer
Bottom of AC
Layer
Middle of Base
Layer
Top of Subgrade
Layer
Table 3.5, Exact Responses from the Original (Exact) Pavement Profile.
uz
x
y
z
x
y
z
Point
(mm)
( m/m) ( m/m) ( m/m) (kPa)
(kPa)
(kPa)
1
2
3
4
5
6
7
8
9
10
11
-57.77
-7.00
-8.44
-3.32
-0.04
1.74
2.70
22.57
1.60
10.32
21.43
-57.7
-27.08
-18.22
-14.08
-10.94
-8.56
-6.76
22.57
1.59
10.32
21.43
-62.49 -677.95 -677.95 -690.00

18.35 -64.75 -116.02 0.00
14.35 -58.21 -83.17
0.00
9.37
-32.39 -59.87
0.00
5.91
-15.20 -43.02
0.00
3.67
-4.93 -31.24
0.00
2.18
1.32
-22.83
0.00
-88.36 -71.10 -71.10 -354.37
-21.33 -49.96 -49.97 -108.51
-12.91 197.33 197.33 -39.95
-85.56 -1.45
-1.45
-6.91
57
0.21
0.17
0.16
0.15
0.13
0.12
0.11
0.19
0.18
0.18
0.18
Table 3.6, Response (Top Row) and Relative Error (Bottom Row in Parenthesis) Using
Case 1 Backcalculated Layer Model.
uz
x
y
z
x
y
z
Point
(mm)
( m/m) ( m/m) ( m/m) (kPa)
(kPa)
(kPa)
1
2
3
4
5
6
7
8
9
10
11
-57.10
(1.173)
-7.04
(0.592)
-8.34
(1.233)
-3.27
(1.408)
-0.04
(7.515)
1.72
(1.039)
2.67
(1.054)
22.16
(1.823)
1.91
(19.686)
10.51
(1.910)
21.56
(0.611)
-57.10
(1.173)
-26.87
(0.788)
-18.07
(0.777)
-13.96
(0.854)
-10.84
(0.876)
-8.49
(0.875)
-6.70
(0.865)
22.16
(1.823)
1.91
(19.686)
10.51
(1.910)
21.56
(0.611)
-61.13
(2.179)
18.26
(0.505)
14.22
(0.922)
9.28
(0.960)
5.86
(0.901)
3.64
(0.833)
2.17
(0.739)
-86.76
(1.808)
-21.17
(0.742)
-13.14
(1.777)
-85.49
(0.089)
-679.52
(0.232)
-65.72
(1.501)
-58.59
(0.651)
-32.59
(0.610)
-15.32
(0.742)
-4.99
(1.047)
1.31
(0.722)
-71.03
(0.099)
-47.60
(4.722)
197.19
(0.074)
-1.42
(1.720)
58
-679.52
(0.232)
-117.21
(1.021)
-83.88
(0.851)
-60.34
(0.794)
-43.37
(0.806)
-31.50
(0.829)
-23.03
(0.858)
-71.03
(0.099)
-47.60
(4.722)
197.19
(0.074)
-1.42
(1.720)
-690.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
-353.91
(0.129)
-107.55
(0.879)
-39.50
(1.118)
-6.90
(0.248)
0.21
(0.465)
0.17
(0.223)
0.16
(0.197)
0.15
(0.186)
0.13
(0.171)
0.12
(0.158)
0.11
(0.146)
0.19
(0.317)
0.18
(0.241)
0.18
(0.255)
0.18
(0.264)
uz
x
y
z
x
y
z
Point
(mm)
( m/m) ( m/m) ( m/m) (kPa)
(kPa)
(kPa)
-58.31
(0.930)
-6.75
2
(3.632)
-8.70
3
(3.091)
-3.52
4
(6.211)
-0.17
5
(300.543)
1.66
6
(4.465)
2.66
7
(1.604)
23.21
8
(2.829)
0.30
9
(81.365)
9.17
10
(11.155)
20.15
11
(5.994)
1
-58.31
(0.930)
-27.02
(0.227)
-18.23
(0.088)
-14.17
(0.660)
-11.05
(1.009)
-8.67
(1.265)
-6.86
(1.484)
23.21
(2.829)
0.30
(81.365)
9.17
(11.155)
20.15
(5.994)
-65.18
(4.306)
18.18
(0.926)
14.50
(1.039)
9.53
(1.719)
6.04
(2.154)
3.77
(2.728)
2.26
(3.539)
-91.06
(3.060)
-21.15
(0.831)
-11.52
(10.792)
-83.18
(2.781)
-672.90 -672.90
(0.745) (0.745)
-62.04 -112.48
(4.186) (3.051)
-57.74
-81.45
(0.802) (2.061)
-32.48
-58.97
(0.265) (1.500)
-15.45
-42.52
(1.594) (1.180)
-5.25
-30.96
(6.419) (0.894)
-22.69
0.98
(0.620)
(25.332)
-71.82
-71.82
(1.021) (1.021)
-58.94
-58.94
(17.970) (17.970)
199.20
199.20
(0.945) (0.945)
-1.52
-1.52
(4.868) (4.868)
59
-690.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
-356.17
(0.509)
-112.31
(3.509)
-41.63
(4.217)
-6.77
(2.037)
0.21
(0.035)
0.17
(0.407)
0.16
(0.307)
0.15
(0.175)
0.13
(0.052)
0.12
(0.066)
0.11
(0.180)
0.19
(0.302)
0.18
(0.468)
0.18
(0.405)
0.18
(0.350)
uz
x
y
z
x
y
z
Point
( m/m) ( m/m) ( m/m) (kPa)
(kPa)
(kPa) (mm)
1
2
3
4
5
6
7
8
9
10
11
-55.76
-55.76
(3.484) (3.484)
-7.24
-26.58
(3.401) (1.870)
-8.02
-17.84
(5.037) (2.058)
-3.07
-13.71
(7.473) (2.602)
-10.62
-0.05
(219.214) (2.883)
1.75
-8.30
(0.508) (3.054)
-6.54
2.66
(1.470) (3.182)
21.21
21.21
(6.017) (6.017)
3.29
3.29
(106.255) (106.255)
11.58
11.58
(12.266) (12.266)
22.58
22.58
(5.388) (5.388)
-57.53
(7.948)
18.21
(0.787)
13.92
(3.001)
9.04
(3.532)
5.91/5.69
(3.709)
3.53
(3.963)
2.09
(4.322)
-82.83
(6.256)
-21.23
(0.479)
-14.42
(11.683)
-86.86
(1.516)
-685.22
(1.072)
-68.93
(6.460)
-59.42
(2.085)
-32.78
(1.229)
-15.28
(0.522)
-4.81
(2.409)
1.55
(17.314)
-70.20
(1.260)
-37.34
(25.253)
195.01
(1.178)
-1.34
(7.540)
60
-685.22
(1.072)
-121.30
(4.548)
-86.03
(3.444)
-61.61
(2.914)
-44.19
(2.703)
-32.04
(2.539)
-23.38
(2.398)
-70.20
(1.260)
-37.34
(25.253)
195.01
(1.178)
-1.34
(7.540)
-690.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
-352.01
(0.665)
-103.77
(4.370)
-37.85
(5.251)
-6.97
(0.797)
0.21
(1.228)
0.17
(0.380)
0.16
(0.391)
0.15
(0.457)
0.13
(0.513)
0.12
(0.568)
0.11
(0.621)
0.19
(0.669)
0.18
(0.394)
0.18
(0.473)
0.18
(0.533)
uz
x
y
z
x
y
z
Point
(mm)
( m/m) ( m/m) ( m/m) (kPa) (kPa) (kPa)
1
2
3
4
5
6
7
8
9
10
11
-58.46
(1.195)
-7.08
(1.125)
-8.48
(0.480)
-3.30
(0.460)
0.00
(100.104)
1.79
(3.007)
2.75
(1.993)
22.84
(1.163)
1.79
(12.346)
10.60
(2.766)
21.88
(2.097)
-58.46
(1.195)
-27.42
(1.238)
-18.42
(1.114)
-14.22
(0.985)
-11.03
(0.881)
-8.63
(0.788)
-6.80
(0.700)
22.84
(1.163)
1.79
(12.346)
10.60
(2.766)
21.88
(2.097)
-63.09
(0.951)
18.58
(1.215)
14.48
(0.913)
9.43
(0.709)
5.94
(0.495)
3.68
(0.221)
2.18
(0.161)
-89.29
(1.051)
-21.67
(1.604)
-13.27
(2.757)
-86.87
(1.524)
61
-678.32
(0.055)
-67.82
(0.116)
-58.02
(0.316)
-32.18
(0.662)
-15.01
(1.273)
-4.77
(3.326)
1.45
(9.928)
-70.86
(0.327)
-48.76
(2.392)
196.66
(0.342)
-1.45
(0.222)
-678.32
(0.055)
-116.20
(0.155)
-83.13
(0.046)
-59.75
(0.196)
-42.89
(0.324)
-31.10
(0.450)
-22.70
(0.577)
-70.86
(0.327)
-48.76
(2.392)
196.66
(0.342)
-1.45
(0.222)
-690.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
-354.14
(0.064)
-108.05
(0.417)
-39.78
(0.415)
-6.98
(0.968)
0.22
(0.860)
0.17
(0.790)
0.16
(0.730)
0.15
(0.671)
0.13
(0.614)
0.12
(0.557)
0.11
(0.504)
0.20
(0.841)
0.19
(0.826)
0.18
(0.816)
0.18
(0.805)
Case 5 Backcalculated Layer Model
uz
x
y
z
x
y
z
Point
( m/m) ( m/m) ( m/m) (kPa) (kPa) (kPa) (mm)
1
2
3
4
5
6
7
8
9
10
11
-55.77
(3.472)
-7.23
(3.321)
-7.99
(5.284)
-3.05
(8.057)
0.07
(256.180)
1.76
(1.183)
2.67
(1.168)
21.22
(5.981)
3.35
(109.897)
11.63
(12.772)
22.64
(5.667)
-55.77
(3.472)
-26.58
(1.857)
-17.83
(2.091)
-13.70
(2.679)
-10.61
(2.996)
-8.29
(3.202)
-6.53
(3.363)
21.22
(5.981)
3.35
(109.897)
11.63
(12.772)
22.64
(5.667)
-57.50 -685.30 -685.30

(7.988) (1.085) (1.085)
18.21 -68.92 -121.33
(0.793) (6.448) (4.570)
13.91 -59.33 -85.99
(3.102) (1.936) (3.394)
9.02 -32.70 -61.55
(3.705) (0.946) (2.807)
5.68 -15.20 -44.12
(3.964) (0.007) (2.545)
3.51
-4.75 -31.97
(4.321) (3.761) (2.331)
2.08
1.60 -23.32
(4.825) (21.260) (2.138)
-82.82 -70.10 -70.10
(6.270) (1.401) (1.401)
-21.27 -36.94 -36.94
(0.299) (26.065) (26.065)
-14.49 194.73 194.73
(12.191) (1.320) (1.320)
-86.94 -1.34
-1.34
(1.614) (7.575) (7.575)
-690.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
0.00
(0.000)
-351.94
(0.685)
-103.63
(4.495)
-37.81
(5.352)
-7.00
(1.180)
0.21
(1.367)
0.17
(0.553)
0.16
(0.588)
0.14
(0.677)
0.13
(0.755)
0.12
(0.832)
0.11
(0.906)
0.19
(0.820)
0.18
(0.554)
0.18
(0.636)
0.18
(0.700)
In Tables 3.5 through 3.10, values without parentheses are the pavement
responses using the backcalculated moduli in Table 3.2 (Cases 1 to 5) while values in
parentheses are the relative errors compared to responses using the original (exact)
pavement profile. Cells that are highlighted with gray show the response with a relative
error higher than 2% whilst values in bold show the response with a relative error higher
than 10%.
62
The results in Tables 3.3, and 3.5 through 3.9 show that even if the RMSE (of
displacements) value is kept less than 1%, the resulting response of strains and stresses
can largely differ than the exact response. Cases 1 and 2 (Tables 3.6 and 3.7,
respectively) have relatively the same RMSE in deflection but have different relative
errors in strains and stresses. In addition, the magnitude and location of the relative errors
vary randomly between Cases 1 and 2. Case 1 showed a relative error of 19.686% in the
horizontal strain (x, or y) at the bottom of the AC layer (point 9) while Case 2 showed a
higher relative error of 81.365% in the horizontal strain at the same point. On the other
hand, at the ground surface (point 5), Case 2 showed a relative error of 300.543% in the
horizontal strain which is approximately 40 times the relative error of the horizontal
strain at the same point in Case 5 (7.515%). The relative error in Case 1 was higher than
2% in 9.1% of the response points and higher than 10% in 3% of the response points; in
Case 2 the relative error was higher than 2% in 48.9% of the response points and higher
than 10% in 13.6% of the response points, all indicating a variable variation in the error.
This high variation in the strain and stress errors can be explained by high relative error
of the backcalculated elastic moduli in Case 2 as compared to that in Case 1 (Table 3.3).
Similar random variation in the relative errors can be also observed at a higher
RMSE for Cases 3 and 4 as shown in Tables 3.8 and 3.9 respectively where the RMSE
was almost the same (less than 1%). However, the relative error in Case 3 was higher
than 2% in 65.2% of the response points and higher than 10% in 13.6% of the response
points while in Case 4 the relative error was higher than 2% in 19.7% of the response
points and higher than 10% in 4.5% of the response points, indicating a higher variation
63
than Case 1. Similarly, the high variation in the strain and stress errors was attributed to
the high variation in the relative error of the backcalculated elastic moduli rather than the
RMSE values.
Comparing Cases 3 and 5 (Tables 3.8 and 3.10) one can observe different RMSE
values but similar relative errors in the backcalculated elastic moduli. Furthermore, in
both cases the magnitude and location of the relative errors in the strains and stresses
were nearly identical. This finding shows that controlling the displacement RMSE does
not necessarily reduce the relative errors associated with the pavement response (strains
and stresses) at either the surface or along the pavement profile. This would require a
better control of the backcalculated moduli. However, controlling the error associated
with the backcalculated elastic modulus is not an easy task since the exact modulus is not
known. Even with known elastic moduli from lab testing, the variation between the
backcalculated and exact moduli can be high. The difference between backcalculated
subgrade elastic moduli from the FWD and laboratory elastic moduli has been studied by
many researchers (Daleiden et al. 1994; Von Quintus and Killingworth 1997). It was
found that there is no unique relation between the backcalculated and laboratory
measured resilient moduli. The ASSHTO Guide (1993) suggested that the backcalculated
modulus is three times the laboratory modulus, whilst Von Quintus and Killingworth
(1997) suggested that one could use some correction factors calculated from any
multilayered elastic program to match the backcalculated and laboratory moduli.
However, the suggested factors are highly dependent on the backcalculation program and
should be used with caution. On the other hand, Stolle (2002) showed that the moduli of
64
the base and the subgrade layers have the largest contribution to the measured FWD
deflection.
It should be noted that in all cases the effect of the backcalculated moduli on the
vertical stress (z) was relatively small as compared to their effect on the horizontal
strains, vertical strains, and horizontal stresses.
3.3.1 Pavement Fatigue Prediction
The damage of flexible pavements can be assessed by predicting the number of
loads needed to initiate cracks (fatigue cracking). The Shell Model (Bonnaure et al. 1980)
and the Asphalt Institute Model (Shook et al. 1982) are frequently used for fatigue
cracking in flexible pavements.
The Shell Model is based on two different loading modes, as given by Eqs. 3.3
and 3.4, below:
Shell Constant Strain Model:
N = 13909 A f K
Es
1.8
Eq. (3.3)
and Shell Constant Stress Model:

N = A f K
Es
1.4
Eq. (3.4)
where N and N are the number of load repetitions to fatigue cracking using the constant
strain and constant stress analysis, respectively, Af and K are material constants, t is the
tensile strain at the critical location and Es is the stiffness of the material (i.e. elastic
modulus). The constant strain model is applicable to thin AC layers usually less than 51
65
mm, whilst the constant stress model is applicable to thick AC layers usually more than
203 mm. The Shell Model was calibrated and generalized for any thickness as given
below (MEPDG, 2004):
N f = A f KF
Es
1.4
Eq. (3.5)
where Nf is the number of load repetitions to fatigue cracking, and F is a constant that
depends on the layer thickness and the material stiffness.
The Asphalt Institute Model is given below:
N f = 0.00432C
1
t
3.291
1
Es
0.854
Eq. (3.6)
where, similarly, Nf is the number of load repetitions to fatigue cracking, C is a material

constant, t is the tensile strain at the critical location, and again Es is the material
stiffness. The Asphalt Institute Model can be used for any thickness.
It can be seen from the above equations that the critical tensile strain and the
stiffness of the AC layer are the main factors affecting the number of load repetitions
needed to initiate fatigue failure. The effect of the backcalculated set of elastic moduli on
the fatigue in flexible pavements can be studied by finding the ratio between the
estimated number of repeated loads (Nf) using the backcalculated set of elastic moduli
and that using the exact set of elastic moduli. In other words, the ratio is equal to Nfb
(backcalculated set of elastic moduli) over Nfe (exact set of elastic moduli).
66
3.3.2 Rutting Damage

Rutting in flexible pavement is considered as a functional deterioration. Rutting is
mainly predicted by calculating the vertical strains at the top of the subgrade and then
estimating the allowable load repetitions until a certain rutting threshold is met. For
example, Shook et al. (1982) assumed a rutting depth of 10 mm in their method, while
Potter and Donald (1985) assumed 20-30 mm rutting depth.
Recently, the results from the test sections at MnROAD were used to develop a
method to predict the number of allowable load repetitions until rutting failure using a
rutting depth of 13 mm as shown in the following relation (Skok et al., 2003):
N r = (5.5).10
15
3.929
Eq. (3.7)
where Nr is the number of allowable load repetitions until rutting failure, and v is the
maximum compressive strain at the top of the subgrade layer.
It can be seen, from the above equation, that the vertical strain at the top of the
subgrade layer is very important to predict the lifetime of the pavement due to rutting.
Similar to the fatigue case, the effect of the backcalculated elastic moduli on the rutting
can be studied by finding the ratio between the estimated number of repeated loads (Nr)
using the backcalculated set of elastic moduli and that using the exact set of elastic
moduli. In other words, the ratio is equal to Nrb (backcalculated set of elastic moduli)
over Nre (exact set of elastic moduli).
67
3.3.3 Fatigue and Rutting Prediction

The fatigue and rutting of the pavement are studied for the five cases as
summarized in Tables 3.1 and 3.2, and the results are listed in Table 3.11. It is observed
from Table 3.11 that, for all cases except Case 2, using either the Shell Model or the
Asphalt Institute Model, the fatigue life will be largely underestimated based on the
backcalculated moduli as compared to those based on the exact moduli. Furthermore,
comparing Case 2 to Case 1 it can be seen that even for the same displacement RMSE the
predicted number of repeated loads for fatigue can be largely overestimated rather than
underestimated. The fatigue prediction results in Table 3.11 also show that even for a
small relative error (less than 2.1% in Case 1) in the backcalculated set of elastic moduli
the fatigue life of the pavement can be underestimated by 60% using the Shell model and
by 45% using the Asphalt Institute Model, indicating a very high sensitivity of the fatigue
life to the relative error in elastic moduli rather than the RMSE.
Table 3.11, Comparison of Fatigue Life Using Backcalculated and Exact Pavement
Moduli.
Nfb/Nfe
Nfb/Nfe
Case
Shell
Asphalt Institute RMSE (%)
Case 1
0.407
0.554
0.24
Case 2 4449.469
251.951
0.22
Case 3
0.027
0.092
0.65
Case 4
0.559
0.682
0.69
Case 5
0.025
0.087
0.85
The effect of the RMSE value and the relative error in pavement moduli on
rutting can be observed from Table 3.12. It can be seen that even for a very small RMSE
(Cases 1 and 2 where the RMSE is 0.22% and 0.24%, respectively), the rutting failure
68
prediction based on the backcalculated moduli was overestimated by 11.7% in Case 2

while it was overestimated by 0.3% in Case 1, indicating a high sensitivity of the rutting
life on the relative error in moduli. In addition, the results show that as the RMSE (error)
increases, the underestimation of the rutting increases, as can be observed by comparing
the rutting results in Table 3.12 for Cases 2 to 5.
Table 3.12, Comparison of Rutting Failure Using Backcalculated and Exact Pavement
Moduli.
Case
Nrb/Nre
RMSE (%)
Case 1
1.003
0.24
Case 2
1.117
0.22
Case 3
0.943
0.65
Case 4
0.942
0.69
Case 5
0.939
0.85
3.4 Conclusions
This study shows that the use of the RMSE is not enough to secure an accurate
backcalculation of the pavement elastic moduli. Large discrepancies can exist in the
predicted pavement strains and stresses using the backcalculated and exact elastic moduli.
As a result, even RMSE values less than 1% can significantly affect the fatigue and
rutting predictions in flexible pavements.
The effect of the RMSE is suitable for controlling the fitness of the
backcalculated deflection basin to that measured in the field while the use of the relative
error in the elastic moduli is more appropriate. However, the availability of laboratory
measured elastic moduli does not guarantee more appropriate backcalculated elastic
moduli. This variation adds more uncertainty when dealing with data from the FWD test,
69
and should be the future endeavor in pavement engineering.
70
CHAPTER IV
INTRODUCTION TO GENETIC ALGORITHMS
4.1 Introduction
The theory that proposes that organisms characteristics change over time
(generations) is called the evolution theory. According to the theory, new species can be
created from the parent species due to changes in the inheritable characteristics
transferred by genes and chromosomes.
The theory promotes the idea of natural selection as proposed by Darwins book
(1859) The Origin of Species. In his book Darwin proposed that organisms with better
characteristics survive and reproduce over generations. This means that certain
characteristics can become dominant over time as a result of changes in the surrounding
environmental conditions (Haldane, 1953; Lande and Arnold, 1983; Futuyma, 2005).
Mendel and later George Darwins half-cousin Francis Galton contributed to the
evolution theory by studying the mechanism of heredity in plant through experimentation
and statistical means. Mendel original work was aimed at studying plant hybridization
using peas. Galton, on the other hand, extended the work of Darwin to humans by
71
creating a large database of human characteristics (including height, fingerprints,

weightetc.) to study the heredity of human traits from one generation to another. Both
works in addition to the work by Darwin are considered the founding stones of the
evolution theory and genetics.
Regardless of the ongoing debate concerning the theory itself, some concepts of
the theory are interesting since they provide some tools and ideas to other sciences.
Nature has always inspired scientists to be creative and to explore more innovative ideas
and solutions to many scientific challenges. The evolution theory is no exception.
Evolution can happen over billion of years while computer simulations that resemble the
evolution theory can happen over few seconds to several days taking advantage of the
ongoing improvements computers.
4.2 Biological Background
All cellular forms of life are not created from nothing. These cellular forms are
created based on certain rules that guarantee the same general characteristics of the base
cells (parents). However, the new organisms (children) will always inherit either the
same or modified copies of the parent original characteristics. Therefore, the presence of
organisms with identical characteristics (even in clones) is impossible in nature. This

variation is necessary in nature since species with the same characteristics transferring
from one generation to the next will not survive diseases and environmental changes
when they happen.
The transfer of the parents information and the creation of new children with
different characteristics are the backbones of the genetic algorithms and programming.
72
The following sections explain the biological terms frequently used in genetic
programming.
4.2.1 DNA
All genetic instructions that are needed for the formation and development of new
organisms are stored in a nucleic acid known as the Deoxyribonucleic acid (DNA). The
DNA is made of sugar and phosphate groups which form a skeleton that carries
molecules, known as bases, connected to each other in certain sequences that are unique
in each organism. The unique sequences represent the encoded genetic information that
transfer from the parents to the children.
The DNA can be described as a polymer chain that consists of small repeating
units which form long strands of bases that can be up to hundred of millions of base pairs.
There are four main bases (nucleotides), mainly: (1) the Adenine (A), (2) the Thymine
(T), (3) the Cytosine(C), (4) and the Guanine (G). These bases connect the DNA strands
together. The bond between two strands consists of two different bases. On the other
hand, these bases can be found in groups called codons which consist of three nucleotides
and form the genetic code of the organism. Since there are 4 main bases and each codon
is made of three bases, the possible number of codons is 64 (43=64).
4.3 Evolution Theory and Computer Science
Scientists were fascinated by the evolution theory not only because of its ability to
explain some biological aspects of nature but because it can be used as a powerful tool in
mathematics. The pioneer work of computer scientists in the 1960s and 1970s was
73
geared toward innovative ideas for mathematical optimization similar to the work
pioneered by Darwin, Mendel, and Galton. The main ideas involved creating a large
number of candidate solutions (population) then applying different operators to simulate
natural selection and hence to arrive at the most optimal solution for the problem in hand.
This pioneer strategy was described by many computer scientists such as Rechenberg
(1965) and Fogel et al. (1966). In modern sciences, the field of evolutionary computation
is a branch of artificial intelligence. Evolution in computer science is mainly carried out
using operators that simulate reproduction, mutation, recombination (crossover) and
natural selection.
In addition to the use of the evolution theory as an optimization tool, scientists
(Smith, 1980; Cramer, 1985; Koza, 1992) used the evolution theory as a machinelearning tool where several computer programs can be optimized using genetic
programming. In genetic programming, programs can be written in any language and
then natural selection and other evolution operators are applied to select the fittest
program.
4.4 Genetic Algorithms
The work on evolution theory in the 1950s and 1960s was geared toward
solving specific problems rather than creating generalized solution procedures. In the late
1960s John Holland was fascinated by the idea of evolution and the possibility of
transforming the knowledge into computer applications. Holland (1975) was able to
establish the basic theoretical foundation of the genetic algorithms by the use of the
schema theory (plural schemas or schemata). In his original work, Holland assumed that
74
a specie population can be represented by chromosomes which consist of strings of ones

and zeros. The chromosome contains several genes (strings) that are instances of alleles
(0 or 1). Evolution theory was then applied using natural selection and operators such as
crossover, mutation, and inversion.
The distinction between Hollands genetic algorithms and the previous work in
evolution theory was remarkable since Holland was able to simulate the natural selection
and the evolution derived operators such as the crossover, mutation, and inversion on the
parent chromosomes and the offspring chromosomes as well. The previous work by
Rechenberg (1965) and Fogel et al. (1966) used mutation to create new offspring. In
addition, the population size was assumed to consist of two to few parents only.
Evolutionary algorithms (including genetic algorithms) are superior compared to
the random walk methods, when dealing with complex search spaces, even though they
are based on random numbers. A specified domain is searched using evolutionary
algorithms where good search points are captured and used as guidance to search for
better points within the search space. However, evolution algorithms cannot be used for
all problems since the availability of customized algorithms (dealing with specific
problems) can be superior. The evolutionary algorithms are best suited to deal with
problems where a closed form solution method is unknown or when methods cannot be
used to search for an optimal solution due to the large space of the search domain.
Details of the fundamental theory of genetic algorithm and the most common
operators in genetic algorithms are presented in the following sections.
75
4.4.1 GA Background
Genetic algorithms are robust and randomized search algorithms based on the
evolution theory (natural selection) and natural genetics (Goldberg, 1989). Many
calculus-based solutions are available for many problems ranging between simple to
complex. However, calculus-based search methods can be inefficient when used to search
for the optimal solution in a complex space since a large number of trials should be
investigated in order to arrive at the most optimal solution, which makes the use of the
selected method unrealistic. Other methods have been proposed for optimization using
different search techniques.
The key points that make the GAs more superior than many other methods for
optimizing functions of large search space and complex behavior are summarized below
(Goldberg, 1989):
1-Parameters in the GAs are encoded using binary/real encoding methods and therefore
the parameters themselves are not used. If the problem has more than one parameter, each
parameter is encoded as a string of l-bit strings then a series of strings attached to each
other is formed to create a chromosome. The entire collection of chromosomes of an
organism forms the genome. Information encoded in the genome is called genotype.
2-A population of points is used to search for the optimal solution and not a single point.
3-No need for function derivatives, function continuity, changes to the optimized
function, or any other auxiliary knowledge of the optimized function to perform the
search. GAs use objective functions to measure the suitability of the solution (fitness).
The GAs work with strings and their associated objective function values.
76
4-GAs use probabilistic transition rules, not deterministic rules.

Most of the global search methods fail to find the optimal solution especially in
multimodal (many-peaked) search spaces. This can be attributed to the fact that in GA
numbers are encoded as strings and therefore a population of strings is searched
simultaneously rather than point by point (Goldberg, 1989). This simultaneous search
technique can be described as climbing more than one peak at the same time looking for
the optimal solution and hence the power of search and the probability of finding the
optimal solution are high. Each population of strings is used to create a new population of
strings using genetic algorithms and the fitness of the population is evaluated by the
objective function, then another population is created and so on.
The main operators that are used in a simple GA are the crossover and mutation
operators. It should be noted that, even though genetic algorithms are used to aid in
optimizing problems, some argue that they do not optimize (De Jong, 1993) but rather
search the landscape problem for the best solution.
4.4.2 GA Fundamental Theory (The Schema Theorem)
The schema theory (plural schemata or schemas) was first introduced as the
fundamental theory of genetic algorithms by Holland (1975). The schema theorem (or the
Fundamental Theorem of Genetic Algorithms) was introduced to describe in a
mathematical sense why genetic algorithms work. The Schema Theorem divides the
search space into subspaces then quantifies the subspaces and explains the movement
mechanism of individuals between subspaces. It should be noted that the standard
schemata theorem assumes bit flip mutation, single-cross crossover, and proportional
77
selection.
Genetic algorithms depend mainly on a bit string representation of solutions and
on schema. The schema is a similarity template that shows the similarity between
different strings in the same population at certain string positions (Holland, 1975;
Goldberg, 1989). Since each individual in the population is encoded using binary coding
(base 2), similarities can occur between different strings that either created randomly for
the population or those that resulted from applying the natural selection operators
(crossover
and
mutation).
For
(124+123+022+021+020=18)
example
can
be
string
similar
to
described
other
strings
by
11000
such
as
{11001,11011,11111,111000etc} at the first two positions (1 and 1). Therefore, one

can assume that the similar strings can follow the format of 11*** where the * indicates a
do not care symbol. This symbol is not coded in programming the GA but it is used to
illustrate the possibility of having two different values at the string position where * is
encountered, namely {0,1}. Using this notation, one can create strings (schemata) using
{0,1,*}. In general for a string of length (l), there are 2l similarity templates (schemata)
because each position of the template can take either its actual value or a do not care
value (*). In addition, a population of size n contains between 2l and n.2l schemata
depending upon the population diversity (Goldberg, 1989). The purpose of locating
similarities between strings is to guide the search besides information from the objective
function (fitness). A graphical presentation of schema is shown in Figure 4.1. As can be
seen since there are two do not care symbols (*), the number of possible new strings is
22=4.
78
String
1 1 1 1 0 0
1 * 1 * 0 0
1 0 1 1 0 0
Schema
1 0 1 0 0 0
1 1 1 0 0 0
Figure 4.1, Graphical presentation of a schema example.
Not all schemata are same. Differences between schemata can be explained based
on two properties; namely: (1) the order of a schema, and (2) the length of a schema. The
order of a schema is equal to the number of fixed positions in the string (the number of 0
and 1), while the length of a schema is equal to the total number of positions minus the
number of the do not care positions (*). The number of strings in a schema can be
calculated using the following equation (White, 1993):
ns = 2 l
os
Eq. (4.1)
where ns is the number of strings in a schema, l is the total length of the string, and os is
the number of the fixed positions (the number of 0 and 1)-order of schema. The defining
length of a schema is the longest distance between two fixed positions.
Equation 4.1 indicates that the order of schema controls the dimension of the
strings in a schema. In addition, the order of schema (os) plays important role in
calculating the probability of survival of the schema for mutation. On the other hand, the
length of a schema plays an important role in calculating the probability of survival of the
schema for crossover.
79
Therefore, schemata are used to analyze the combined effect of the selection,
crossover, and mutation on building blocks in the population. Genetic algorithms
operators can disrupt the schemata depending on the length of the schema and the size of
the population. In GAs, schemata of strings with high fitness values will be selected more
(in a fitness-based selection) and therefore the disruption of the schemata can occur due
to the mutation operator while destroying the schemata is possible due to the crossover
operator. In general, the number of schemata processed in each generation with the effect
of natural selection operators can be n3.
The effect of the GAs operators on the schemata can be studying the behavior of a
schema H in population x1,x2,n(t) where t is a time step at which there exist m(H,t)
examples of a particular schema (Goldberg, 1989). In addition, assume that each
f ( xi )
.
f (x j )
chromosome Ci will be selected (fitness-based selection) with a probability

j
Therefore, for a population of size n there exists:
f (H )
f (x j )
m( H , t + 1) = m( H , t ).n.
Eq. (4.2)
where f(H) is the average fitness of all strings in x1,x2,n(t) matching H. Assuming the
population average to be f =
f ( x j ) / n , Eq. (4.2) can be written as:

j
m( H , t + 1) = m( H , t )
f (H )
f
Eq. (4.3)
Assume that for a schema H, it will be above the average fitness an amount cf where c is
a constant. Therefore, Eq. (4.3) can be written as:
80
m( H , t + 1) = m( H , t ).
f + cf
= (1 + c)m( H , t ) = (1 + c) t m( H ,0)
f
Eq. (4.4)
It can be seen from Eq. (4.4), even though mutation and crossover not included, that the
number of trials to above-average schemata increases exponentially due to reproduction.
Crossover can either disrupt or destroy the schema. Schema disruption can occur
when the crossover of the first and second mates produces a chromosome (string) with
the do not care (*) at positions that can alter slightly the string rather changing the
string entirely while large change in the string value can destroy the chromosome.
Schema disruption and destroying are shown in Figure 4.2 when mating strings A and H1,
and A and H2, respectively. It can be seen, a schema can survive when the crossover
position is located outside the defining length (l).
A =100 011
H1 = * * * 1 0 *
H2 = * 0 * * * *
Figure 4.2, Schema changes due to crossover.
Based on the above discussion, the survival probability under simple crossover ( p s ) can
be given by (Goldberg, 1989):
ps = 1
(H )
l 1
Eq. (4.5)
and the probability that a schema will be destroyed ( p d ) can be given by:
pd =
(H )
Eq. (4.6)
l 1
81
where (H ) =os (order of schema), and l-1 is the possible positions of the cut point
(crossover point) within a schema.
If crossover is performed by random choice using a crossover probability ( pc ) Eq. (4.5)
becomes:
ps 1 pc .
(H )
l 1
Eq. (4.7)
Using Eqs. (4.3) and (4.7), the combined effect of the crossover and mutation, assuming
independence of the crossover and mutation operations, can be expressed by:
m( H , t + 1) m( H , t )
f (H )
(H )
1 pc .
l 1
f
Eq. (4.8)
The above equation shows that the rate of growth or decay of schema (H) depends
on a multiplication factor. Moreover, the equation indicates that schemata with aboveaverage performance (fitness) and short defining length (l) will be sampled at
exponentially increasing rates.
On the other hand, mutation can disrupt the schema since positions will be altered
randomly. The mutation can affect only the fixed positions of a schema. Therefore, the
disruption survival probability of a schema with a defining length o(H ) and probability
of mutation p m is:
pt = (1 p m ) o ( H )
Eq. (4.9)
It can be noticed that for very small values of p m ( p m <<1), Eq. (4.9) can be
approximated using:
pt 1 p m .o( H )
82
Eq. (4.10)
Therefore, under the effect of reproduction, crossover, and mutation, the expected
number of copies that a schema can receive in the next generation is given by:
m( H , t + 1) m( H , t )
f (H )
(H )
1 pc .
o( H ) p m
l 1
f
Eq. (4.11)
The above expression can lead to the Schema Theorem or the Fundamental
Theorem of Genetic Algorithms. The theorem states that Short, low-order, aboveaverage schemata receive exponentially increasing trials in subsequent generations
(Goldberg, 1989). The theorem indicates that short, low-order schemata analyze the
search space in genetic algorithms. On the other hand, Holland (1975) showed that, based
on the Schema Theorem, standard genetic algorithm minimizes the allocation of trials to
the wrong schemata, which in turn minimize the expected loss of computation time.
The GAs schemata can be viewed as partitions of the genome space or as groups
of search points that have some syntactic features.
4.4.3 Schemata Evaluation
Another advantage of GAs is the fact that the number of the evaluated schemata is
larger than the number of individuals in the population (called implicit parallelism),
which makes the GAs very efficient.
The probability that a fixed schema of order (H ) =i matches at least one string
in the population is given by:
P(H matches at least one string) = 1- P( H matches none of the n strings)

= 1- P( H matches not one string)n
= 1- (1- P( H matches one string))n
83
1
= 1- 1
k
Eq. (4.12)
and the expected number of schemata of order i in n strings is:

1
i k 1 1
k
l
Eq. (4.13)
Therefore, the total number of expected schemata is:
S ( k , l , n) =
1
i ki 1 1
k
l
i =0
Eq. (4.14)
for an alphabet of size k, where l is the length of a string, and n is the string population
size (Goldberg, 1989). When n , Eq. (4.14) converges to (k+1)l, which is the total
number of schemata of length l over an alphabet of size k.
The use of binary encoding in GAs facilitate the use of many more schemata
compared to using higher level representation which in turn covers a wider search space
(Holland, 1975). The strings of a space of (n) points can be encoded as strings of length
( l = log k n ) using a k-ary alphabet. As mentioned before, each string has 2l schemata
matches resulting in a total expected number of schemata equal to (k+1)l when n .
Both the l and the (k+1)l are maximum when k=2 (binary representation). The previous
conclusion agrees with Goldbergs (Goldberg, 1989) principle of minimal alphabets
where he suggests that the smallest alphabet should be used to permit a natural expression
of the problem.
84
4.4.4 Components of Genetic Algorithms

Genetic algorithms are efficient and highly effective search and optimization
tools. Genetic algorithms differ from other search and optimization algorithms and tools
in the way they handle the solutions. In GAs tentative solutions (chromosomes) are
encoded before being manipulated by genetic operators.
Then, chromosomes are
manipulated using genetic operators such as crossover and mutation. In GA function

optimization, an objective function should be determined in order to evaluate the
goodness of the individuals (chromosomes) of the population. This can be done using the
fitness function. The fitness function in optimization problems normally is taken as the
function itself.
A standard genetic algorithm works as the following (Mitchell, 1999):
1-Generate a random population of size n of l-bit chromosomes. The chromosomes
represent the trial solutions to the problem.
2-Use the fitness function [f(x)] to calculate the fitness of each chromosome (x) in the
population (n).
3-Create n offspring individuals by repeating the following steps:
i-Select two parents (chromosomes) from the current population based on the
fitness values. The two parents will be biased in favor of fitness. The parent
chromosomes may be selected more than once.
ii-Crossover the selected two parents to form two offspring. The crossover takes
place at a randomly chosen point (using the crossover probability pc). When no
crossover is desired (crossover probability is equal to zero), the two offspring will
85
be the exact copies of their respective parents.

iii-Mutate the two resulting offspring chromosomes at each locus with probability
pm then place the resulting chromosomes in the new population.

4-Use the new population to replace the current population.
5-Go to step 2. Repeat if necessary.
Each run of the genetic algorithm consists of many iterations where each iteration
resembles a generation of chromosomes. The total number of evaluations (number of
generations multiplied by the population size) is problem dependent. Genetic algorithms
are randomized methods and therefore each run can result in different results and hence
results are reported as average fitness of population and as average solution value(s) for a
certain population size.
4.4.4.1 Population Size
A population in genetic algorithm contains a certain number of chromosomes
(candidate solutions). As the number of chromosomes increases, the diversity of the
population increases which in turn increases the survival of the population. The survival
is attributed to the fact that a large population will have individuals with diverse
characteristics that enable them to evolve and survive against any environmental changes
since no dominant gene will transform from one generation to another.
However, large populations can slow down the search for the fittest (with highest
possible fitness value) and can be computationally intensive. On the other hand, small
populations tend to have dominant genes (genes with strong characteristics) causing early
unfavorable convergence of the solution. This can be attributed to the fact that dominant
86
solutions will restrict the search space, which will lead to local optima. Therefore,
solutions in small populations will depend mainly on the mutation to produce new
chromosomes that do not become dominant quickly, which in turn slows the search for
the global optima.
4.4.4.2 Selection Methods
The selection of the chromosomes that will undergo crossover to produce new
chromosomes (candidate solutions) can be carried out using many methods among which
the roulette wheel and the tournament selection methods are the most common. In
his original work Holland (Holland, 1975) used the fitness-proportionate selection where
the fitness of the individual was divided by the average fitness of the population creating
a relative fitness for each individual.
Roulette Wheel Selection
The roulette wheel selection is used to implement the fitness-proportionate

selection. Each individual of the population is assigned a slice of the wheel that is
proportional to the individuals fitness. Then, the wheel is spun n times (n is the
population size) and the slice (and its associated individual) under the wheel marker is
selected to be a parent for the next generation. However, this method can result in
unlikely distribution of offspring such that all offspring are allocated to one individual
due to the stochastic nature of the method.
Assume that the fitness of individual i is fi, then the sum of all fitness values in the
population (n = population size) is given by:
87
F=
n
i =1
Eq. (4.15)
fi
Assume a running sum Tj :
Tj =
j
i =1
fi
j=1,,n
Eq. (4.16)
where Tj+1 Tj j, and Tn=F.
Then generate a random number R [0.0,F], and locate the element Tj for which:
Tj-1<R<Tj
Eq. (4.17)
The above procedure requires that the objective fitness value of the function to be
positive definite value. The roulette wheel sector in this method is given a size of
s=
1 fi
2 F
Eq. (4.18)
while spinning the wheel is carried out using the random number generator.
In the above procedure, the probability of stopping the wheel in sector j is
proportional to the angular extent of the sector, and hence the normalized fitness of
individual j.
Tournament Selection
The tournament selection method (Goldberg and Deb, 1991) is more common and
efficient than the roulette wheel selection method. The tournament selection method is
computationally more effective since there is no need to compute the individual fitness of
the chromosome and the total fitness of the population in order to implement the method.
88
In this method, two individuals are chosen randomly from the population. A random
number between 0 and 1 is selected (c) and compared to a constant parameter (j>0.5 in
order to provide selection pressure). If the value of the random number is less than the
value of the constant parameter (c<j), the individual with the higher fitness value is
selected to be a parent. At the end of the process, both parents are returned back to the
population with the possibility of being selected again in another iteration.
Goldberg and Deb (1991) showed that the binary tournament selection converges
in O(log n) generations (n is the population size), while unscaled fitness-proportionate
schemata converge in O(n log n) time.
4.4.5 Genetic Operators and Parameters

The main genetic operators and parameters are discussed in the following
sections.
4.4.5.1 Elitism
In standard GAs, there is a possibility that new generations may replace good
individuals with lower fit characteristics (smaller fitness) increasing the time needed to
search for good individuals. This problem can be solved by using the elitist strategy
where the best r% of the population always survives to the next generation. This is done
in practice by saving in temporary storage the fittest individual of the parent population,
and artificially reinserting it at the end of the generational iteration.
89
4.4.5.2 Crossover
The crossover operator is intended to disrupt or destroy the schemata based on the
crossover strategy. As can be expected, more than one crossover strategy exists since the
crossover process can be carried out by mating the chromosomes at random positions.
The simplest crossover strategy is the single-point crossover, which is based on
one-point crossover operator that cuts the chromosome into two parts at random string
positions. However, the single-point crossover has limited ability since in some cases it
can change entirely the beneficial effect of the schemata.
The second crossover strategy is the two-point crossover, which is based on two
cut points. The resulting three string parts are then combined with three other parts from
another chromosome to form a new chromosome as shown in Figure 4.3. As can be seen,
the single-point crossover cannot be used to combine the two schemata in Figure 4.3.
However, the two-point crossover fails to work in schemata where a multiple crossover is
needed to combine the chromosomes, as shown in Figure 4.4.
1***0
*01**
1 001 0
0 010 1
1 010 0
0 001 1
Figure 4.3, Example of a two-point crossover.
1*0*0
*0*0*
10010
00101
10000
00111
Figure 4.4, Example where the two-point crossover fails to work.
90
An alternative to the single-point and the two-point crossover is the uniform

crossover (shown in Figure 4.4). In the uniform crossover the location of the encoded
feature on a chromosome is irrelevant compared to the single-point and two-point
crossovers where the disruption possibility of the schema increases as the number of the
cut positions increases. In addition, the problem of local minima in the fitness function
can be avoided using the uniform crossover since good features that are encoded some
distance apart on a chromosome are preserved (Syswerda, 1989).
Crossover is normally implemented using a crossover probability or rate. The
crossover probability in a single-crossover represents the probability that two parents will
crossover in a single point (regardless of the crossover strategy). The crossover
probability in a multiple-crossover represents the number of points at which a crossover
takes place.
In practice, a probabilistic test is carried out to compare between the crossover
rate (0.0
pc
1.0) and a random number generated in each crossover process (0.0
1.0). If the generated random number is less than the crossover rate (r
pc), the crossover
process is carried out; otherwise (r > pc) the two offspring become exact copies of the
two parents. For example, a value of pc =x % means that x % of the population members
will participate in crossovers to create new members while the remaining percent (1-x %)
will be directly accepted as new members.
4.4.5.3 Mutation
Mutation of chromosomes can be carried out by flipping the bit at a random
position. For example if the position contains 0, the mutation will result in a 1 at that
91
position. In early evolution strategies, random mutation was the only source of variation
while in genetic algorithms it plays a secondary (minor) role compared to the role of the
crossover operator. However, in evolutionary strategy methods, mutation alone (without
crossover) was used to solve complex behaviors (Ray, 1991).
In practice, a probabilistic test is carried out to compare between the mutation rate
(0.0
pm 1.0) and a random number generated in each crossover process (0.0
If the generated random number is less than the mutation rate (r
1.0).
pm), the mutation is
carried out; otherwise (r > pm) the mutation is not carried out.
4.4.5.4 Niching
Niching in biology describes the way a population responds to the surrounding
resources and enemies. When there is a wealth of resources available for the population it
is natural for the population to grow. On the other hand, when enemies surround the
population or resources are not adequately available, the population tends to decrease.
Therefore, niching explains the organisms role or job in its habitat or environment.
In genetic algorithms, niching solves the problem of focusing the search to one
subdomain in the search space which in turn leads to early unfavorable convergence of
the search. Early convergence in genetic algorithms can be caused by genetic drift due to
stochastic errors caused by small population sizes. Therefore, it is important to give focus
to other subdomains in the search space, to avoid the genetic drift, which can be done
using sharing.
Modern niching schemes in genetic algorithms are based mainly on either De
Jongs (1975) crowding scheme or Goldberg and Richardsons (1987) sharing scheme. In
92
the crowding scheme, similar individuals (bit-by-bit comparison) are replaced to allow
new members to enter the population. Sharing scheme uses a sharing function to
determine the neighbouhood and degree of sharing for each string in the population. The
derated fitness of each individual before the selection takes place as suggested by
Goldberg and Richardon (1987) is given by:
f s ( xi ) =
f ( xi )
n
j =1
Eq. (4.19)
s(d ( xi , x j )
where f ( xi ) is the fitness function, and s (d ( xi , x j ) is the sharing as shown in Figure 4.5.
The sharing is equal to the relative bit difference between the two chromosomes;
i.e. the number of different bits (the Hamming distance between the strings) divided by
the string length. Goldberg and Richardson suggested different power law sharing
functions as shown in Figure 4.6. The most common power law sharing function is (1- ).
Figures 4.5 and 4.6 indicates that strings close to an individual (a Hamming
distance of zero) require a high degree of sharing (close to one) while those far from the
individual require a very small degree of sharing (close to zero). Therefore, this scheme
(Eq. 4.19) limits the uncontrolled growth of particular species (a class of organism with
common characteristics) within a population since individuals that are in the same
neighborhood contribute to one anothers share count, thus derating one anothers fitness
value. The niching effect is shown in Figure 4.7 for a function with equal peaks.
93
Share
s(d(xi,xj))
1.0
0.0
0.0
share
Distance
dij=||xi-xj||
Figure 4.5, Triangular-sharing function. After Goldberg and Richardson (1987).

1
1-4
Share
s()
1-
1-
1/4
0
0
Relative Difference
=dij/share
Figure 4.6, Power-sharing functions. After Goldberg and Richardson (1987).
94
1.1
(a)
1.0
0.9
0.8
f(x)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0
0.2
0.4
0.6
0.8
1.0
x
1.1
(b)
1.0
0.9
0.8
f(x)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Figure 4.7, Simple genetic algorithm performance with (a) and without (b) sharing. After
Goldberg and Richardson (1987).
95
4.5 GA Difficulties
The optimization of a function to determine the global optimum can be difficult
when the function has several local optima and when the function is non-differentiable.
Such functions are best optimized using genetic algorithms since no differentiability is
required and functions with more than one optimum can be optimized. However, the
efficiency of optimizing a function using genetic algorithms is dependent on the features
of the fitness landscape and the GA parameters and operators (Holland, 1975).
Goldberg (1993) studied the difficulties that can reduce the efficiency of genetic
algorithms. These difficulties include: isolation of desired solution, deception,
multimodality, problem size, and the search bias.
The isolation of the desired solution can occur when the optimal solution is
globally isolated and hence the surrounding points in the search landscape do not provide
any information about the right search direction, and therefore, the sampling of the search
space can be very difficult (Goldberg, 1989).
The optimized function can be deceptive when the resulting low order schemata,
using the GA, leads the search process towards a false optimum rather than a global
optimum. Deceptive functions cause undesirable convergence to local optima.
The existence of several optima in the search landscape (modality) can increase
the difficulty of the GA search process for a global optimal solution. Increasing the
modality of the search landscape can increase the probability of premature convergence.
The selection of the population size is very important since small populations can
have low diversity and hence the premature convergence to local optima is likely, while
96
large populations are time consuming.

Genetic algorithms search bias is dependent on the interaction between the GA
parameters and the performance of the GA operators which are very complex and
problem dependent.
4.6 Conclusions
Genetic algorithms are robust algorithms that can be used to optimize the search
domains of functions. Genetic algorithms mimic the natural selection in nature by
creating a pool of parents with characteristics that can be inherited by their children. Each
individual is a string of binary code that can mate with another chromosome and mutate
with time to create individuals with characteristics better than the parents.
The optimized unknown in the genetic algorithms problems can have a range
rather than a seed value making the search for an optimal solution more powerful since a
global solution can be obtained and divergence or local optima are prohibited when
compared to other methods. In addition, genetic algorithms do not require transforming
the optimized function from one form to another (such as derivative) to obtain the
solution. These features of the genetic algorithms make them appealing to be used in the
backcalculation of pavement layer moduli using the Falling Weight Deflectometer
(FWD) test data.
97
CHAPTER V
BACKCALCULATION OF LAYER MODULI USING THE BACKGENETIC3D
PROGRAM
5.1 Introduction
BackGenetic3D is a Fortran program developed by the author to backcalculate
the layers moduli of any pavement system with no restrictions regarding the number of
layers, thickness of layers, location of the response points, number of loading circles, the
shape of the loaded area (circle, triangle ,..etc.), and the configuration of the applied
loading (uniform, nonuniform). The code consists of approximately 4100 source code
lines. The program is based on the MultiSmart3D forward multilayer elastic program
developed by the Computer Modeling and Simulation Group at the University of Akron
(Pan and Alkasawneh, 2006). The MultiSmart3D program is based on the innovative
computational and mathematical techniques for multilayered elastic systems (e.g., Pan,
1989a,b, 1990, 1997).
BackGenetic3D is a very advanced and fast program that is capable of
backcalculating the elastic moduli accurately. The program searches the domain of the
elastic moduli for possible solutions (moduli) and determines the optimal solution using a
guided stochastic search technique. The search technique is based on the genetic
98
algorithm technique proposed by Holland (1975) with many improvements to handle the
complexity associated with the backcalculation of layers moduli. The implemented
search technique is designed to avoid the local optima that can be encountered when
using other methods such as the Simplex method.
The program is user-friendly and can be used with the typical inputs that are used
in any elastic backcalculation program. In the BackGenetic3D program there is no need
for seed moduli and a range of the seed moduli is required while in other programs, that
do not use the GA, a seed modulus is needed for each layer. Using seed moduli can lead
to local optima or can lead to divergence rather than convergence.
The genetic operators needed to perform the backcalculation should be provided
by experienced users or can be estimated using special techniques as those suggested in
this work. More details regarding the capability and important guidelines on how to use
the BackGenetic3D program in backcalculating the elastic moduli are discussed
throughout this thesis.
5.2 BackGenetic3D Program Components
The BackGenetic3D program consists of many components that work together to
backcalculate the elastic moduli. The backcalculation procedure starts with the selection
of the parameters of the unknown layers including the Poissons ratio, layer thickness,
and typical ranges of the elastic moduli based on the description of the layer (material
type). Figure 5.1, shows the main components of the program and the sequence of the
components. As can be seen, the program follows the natural selection by creating a pool
of parents then mating the parents to create new offspring (chromosomes). The fittest
99
Start
Input Pavement Data
Input Genetic Parameters
Generate the Initial Pool of

Parent Solutions
Use the Fitness Function to

Evaluate the Parent Pool
Generate Chromosomes using

Crossover and Mutation
Create Next Parent Pool by

Selecting Chromosomes
based on Fitness and the
Selection Method
No
Has stopping
Criteria Been
Met?
Yes
Stop
Figure 5.1, Flow chart of the genetic algorithm implemented in the BackGenetic3D
program.
100
offspring are then selected as the best candidates that can survive the surrounding
environment represented by the constraints (moduli ranges) and the fitness function. The
fitness function facilitates the selection by providing a feedback about the quality of each
offspring and if it will be beneficial to keep it for the next generation to form the new
parent pool. The stopping criterion in BackGenetic3D is either the generation number or
the fitness value.
5.3 Backcalculation Mechanism
The program starts by randomly generating the initial pool of parents using a
random number generator, which is desirable since no bias will exist among the selected
individuals and the generation will cover a wide range of the search domain. This
approach is different than the available backcalculation programs that do not use the GA
where seed moduli should be provided in order to start the backcalculation process.
Selecting the seed moduli can create a bias toward one or all of the selected moduli,
which can be far from the optimum solution. The program uses Mersenne Twister
random generator (Matsumoto and Nishimura, 1998) to create the parents pool.
The individuals form the genotype. The genotype is a chromosomal presentation
of the solutions where each genotype consists of genes. Each chromosome contains the
trial solutions encoded in a binary string format. For example, if the pavement system
contains three layers, there will be three unknowns and hence the chromosome will
consist of three genes (moduli). The program deals with chromosomes when executing
the genetic algorithm rather than genes. Figure 5.2, represents a chromosome that
contains three genes which is used to solve a three-layer pavement system.
101
Chromosome
A1 A2 A3
Gene A1
0 1 0 0 1 0 Gene A3
1 1 1 0 0 0
Gene A2
1 1 0 1 0 1
Figure 5.2, Chromosomal presentation of the trial solutions encoded as binary bits.
5.4 Population Size
The size of the parent pool represents the size of the generation in the genetic
algorithm and is determined by the user. The size of the generation is used as an input
parameter by the random generator to create the initial parent pool. The generation can be
of any size depending on the users experience and on the complexity of the problem.
However, increasing the population size will increase the computational time compared
to smaller population sizes. It is recommended to use large population sizes as increasing
the size will increase the diversity and the distribution of the parent pool. Diversity of the
population is highly recommended for pavement moduli backcalculation since a wider
range of the search domain will be covered and hence the probability of getting stuck
with local optima will be minimized and the need for high selection pressures through the
genetic operators will be reduced.
Goldberg and Deb (1991) and Goldberg et al. (1992) suggested using the
following equation to determine the size of the population (n) assuming that the function
can be viewed as the sum of m independent subfunctions:
102
n = O(m k )
Eq. (5.1)
where m = l / k ,l is the total number of bits in a chromosome, is the alphabets

cardinality, and k is the order of the schemata which can be approximated as the string
length of each gene (modulus). Using binary coding, the above equation can be reduced
to:
l
n = O( 2 k )
k
Eq. (5.2)
The effect of the number of layers on the population size assuming k=2,4,8,10 is
shown in Figure 5.3. As can be seen, the population size increases as the number of
layers increases and as the binary string length of genes increases indicating that the
population size is problem dependent. However, the figure indicates the worst-case
scenario and does not necessarily indicate the population size required to backcalculate
the moduli. Experience with pavement moduli backcalculation, in this work, showed that
the required population size is less than those shown in Figure 5.3 by many folds.
103
14000
k=2
k=4
k=8
k=10
Population Size (n)
12000
10000
8000
6000
4000
2000
0
0
10
12
14
Number of Layers
Figure 5.3, Population sizing relations as a function of number of layers and binary string
length of genes.
5.5 Crossover Probability

Crossover in genetic algorithms is a very important operator since it is responsible
for combining the good characteristics of the existing individuals (chromosomes) and
hence works as features enhancer. Different crossover techniques can be used in genetic
algorithms including the single-point crossover, the two-point crossover, and the uniform
crossover. The single-point and the uniform crossovers are implemented in
BackGenetic3D. In the single-point crossover a point is selected at random along the
chromosome length, then two children (offspring) are produced by selecting and
combining one part from each parent. If the single point-crossover is used, it is
recommended to use high probability values (more than 0.5) to ensure more diverse
104
population at the initial stages of the backcalculation and hence avoid premature
convergence. In general, values less than 0.6 are rarely used (Eiben et al., 1999).
The uniform crossover allows a variable number of crossover points and therefore
avoids any bias toward any of the string bits. Syswerda (1989) showed that the uniform
crossover is better than the single-point and the two-point crossover. In addition, he
argued that the schema in the uniform crossover get destroyed at a faster rate than in the
single-point and two-point crossovers and get created at a far faster rate.
5.6 Mutation
Mutation was normally used as the main operator in evolution strategies where
the diversity of the population is controlled by one factor. In genetic algorithms, mutation
allows the production of new chromosomes with new characteristics that are different
from those in the parents. Mutation ensures a controlled diversity in the population to
avoid premature convergence. On the other hand, mutation is one of the means by which
genetic algorithms exploit the search domain for the best solutions.
BackGenetic3D uses jump mutation where the probability of mutating each
chromosome is equal to the mutation probability. In mutation, a binary string in the

chromosome is flipped to its opposite value. For example, if the mutated binary string is
1 it will be flipped to 0 while if the string is 0 it will be flipped to 1. By flipping binary
strings, the value of the gene changes dramatically. When the mutation probability is
equal to one (Pm=1.0), all chromosomes will be mutated resulting in a complete random
(not guided) generation of chromosomes. It is recommended to use low mutation
probabilities since diversity is desirable without the destruction of the schema. Mutation
105
can be viewed as an occasional alteration of the chromosomes. It is common to use a

mutation probability equal to either the inverse of the population size or the inverse of the
chromosome length.
5.7 Creep Mutation

The BackGenetic3D program is the first genetic algorithm program to use creep
mutation in the backcalculation of pavement moduli. In creep mutation the values of the
chromosomes are altered slightly by shifting the values backwards or forwards to the
neighboring value. The rate of change imposed on the chromosomes using the creep
mutation is less than that using the uniform mutation. It should be noted that the creep
mutation acts on the decoded gene (phenotype) while the mutation works on the encoded
chromosome (genotype). A common value for creep mutation is equal to the ratio
between the length of the gene binary string to the population size.
5.8 Selection Procedure

The selection procedure is very important in genetic algorithms since it
determines the mating chromosomes for the crossover process in order to create new
chromosomes (solutions). The BackGenetic3D program uses the tournament selection
method. In tournament selection, two chromosomes are chosen at random and then
compared based on their fitness values. Using other selection methods such as the
roulette wheel selection, the linear ranking selection, and steady state selection were
excluded. Goldberg and Deb (1991) showed that the tournament selection is highly
106
efficient and requires less population sizes to converge compared to many other selection
methods suggested in literature. In the tournament selection a significant selection
pressure is applied while the pitfalls of fitness ordering or ranking are avoided
(Falkenauer, 1998).
5.9 Niching
Niching is used for the first time in the backcalculation of pavement moduli in
BackGenetic3D. Niching is a user defined option and can be either used or turned off.
Niching helps in preserving temporally useful diversity. Niching is helpful when the size
of the population used is tends to cause a premature convergence. In such situations,
niching is highly recommended
The implemented niching in the BackGenetic3D program is based on Goldberg
and Richardsons (1987) sharing scheme which is more appropriate for multimodal
functions. Sharing scheme uses a sharing function to determine the neighbouhood and
the degree of sharing of each string in the population. The derated fitness of each
individual before selection takes place as suggested by Goldberg and Richardon (1987) is
given by:
f s ( xi ) =
f ( xi )
n
j =1
Eq. (5.3)
s(d ( xi , x j )
where f ( xi ) is the fitness function, and s (d ( xi , x j ) is the sharing. The sharing is equal to
the relative bit difference between the two chromosomes; i.e. the number of different bits
(the Hamming distance between the strings) divided by the string length.
107
Goldberg and Richardson suggested different power law sharing functions. The
most common power law sharing function is (1- ). The value of
is user defined in
BackGenetic3D. However, a value of =1.0 is commonly used and is the default in

BackGenetic3D.
5.10 Elitism
Elitism strategy is used for the first time in the backcalculation of pavement
moduli in BackGenetic3D. Elitism is a user defined option and can be either used or
turned off. In the elitism strategy, the chromosome with the best fitness is always
retained. Therefore, elitism can only be used after calculating the fitness of the
chromosomes. Elitism is beneficial in reducing the required number of generations since
convergence to the best solution is achieved quickly. In addition, it is recommended to
use elitism especially when the selected genetic operators tend to disrupt the schema of
the new chromosomes such as when the mutation probability is high.
5.11 Population Seeding

The initial population (parent pool) is generated randomly at the start of each run.
However, BackGenetic3D has the capability of seeding the population using information
from a previous run. In this technique, the last chromosomes are retained externally and
used as the initial parent pool for the next generation. Population seeding is a user defined
option and can be either used or turned off. The default option is no population seeding.
108
5.12 Random Number Generator

The genetic algorithms are based on generating stochastic random numbers to
spatially distribute the parent pool. In addition, random numbers are used during many
steps in the genetic algorithm computation process. Therefore, it is important to ensure
that the used random generator is highly stochastic and the probability of generating the
same number or gene is extremely low.
The BackGenetic3D program uses Mersenne Twister (MT) random generator as
proposed by Matsumoto and Nishimura (1998). The MT random generator employs a
powerful and fast algorithm that is capable of generating very high quality
pseudonumbers. The algorithm was designed to solve many of the disadvantages
associated with other random number generators. The MT algorithm is very powerful
since it has a proved and extremly large period of (219937-1) and the probability of
duplicating the value is negligeble. On the other hand, the algorithm is suitable to be used
in computer implementations when speed is of a concern since the generation of random
numbers is fast. The seed value of each random generation is automated and obtained
using the clock of the computer (time, day, and year) to gurantee complete randomness
during the run.
5.13 BackGenetic3D Program Verification
The validity of the BackGenetic3D program has been assessed by comparing its
performance with the performance of other available programs and backcalculation
procedures. A robust backcalculation program should include a robust backcalculation
procedure to minimize or eliminate (if possible) the problems of local optima and seed
109
moduli effect on the backcalculation results. In reality, all of the popular commercially
available backcalculation programs share the similar backcalculation procedure.
However, the main key in these programs is the estimation of the seed moduli in order to
minimize the effect of the seed moduli on the backcalculated moduli realizing the
drawback of the seed moduli and the premature convergence to local optima.
A four-layer composite pavement system was used to validate the backcalculation
procedure as shown in Table 5.1. The pavement system was analyzed by Harichandran et
al. (1993). Harichandran et al. (1993) backcalculated the pavement moduli using three
backcalculation programs (MICHBACK, MODULUS, EVERCALC). MICHBACK was
developed by Harichandran et al. (1993) while MODULUS and EVERLCALC were
developed by Uzan et al. (1989) and Sivaneswaran et al. (1991), respectively. The
backcalculation was performed using a deflection basin generated using a load of 40.03
kN applied to a circular plate of radius 150.1 mm. The deflection basin consisted of seven
surface deflections at radial distances of 0,203.2, 304.8, 457.2, 609.6, 914.1, and 1524
mm from the center of the loaded plate.
Table 5.1, Four layer pavement system.
Thickness
Actual Modulus
Pavement Layer
Poissons Ratio
(mm)
(MPa)
AC
152.4
0.35
3,447.38
Slab
254.0
0.25
31,026.40
Base
203.2
0.4
172.37
Subgrade
0.45
51.71
In addition to the above programs, BackGenetic3D was used to backcalculate the
pavement moduli using the GA parameters shown in Table 5.2 and the pavement moduli
110
ranges shown in Table 5.3.
Table 5.2, Range of GA inputs used in the four-layer pavement.

Parameter
Value
Population size
512
Number of generations (G)
150
Probability of crossover (Pc)
0.5
Probability of mutation (Pm)
0.0
Table 5.3, Ranges of the moduli.

Layer
Moduli (MPa)
AC
1,000-6,000
Slab
10,000-60,000
Base
100-400
Subgrade
10-150
The GA fitness function was assumed to be:

Minimize:
RMSE =
1
m
m
i =1
d i Di
Di
Eq. (5.4)
where RMSE is the root mean square error, m is the number of measuring sensors, d i is
the backcalculated deflection at point i, and Di is the measured deflection at point i. In
addition, the relative error (RE) of each backcalculated modulus was estimated using:
RE =
Ec Em
100%
Em
Eq. (5.5)
where Ec and Em are the backcalculated and the actual moduli, respectively.
The relations between the number of generations and the best fitness and the
associated RMSE of the generation are shown in Figure 5.4 while the relative error of
111
1.000
0.025
0.995
0.020
Best fitness
Deflection RMS
0.990
0.015
0.985
0.010
0.980
0.005
0.975
Deflection RMSE
Fitness
each modulus is shown in Table 5.4.
0.000
0
20
40
60
80
100
120
140
160
Generation
Figure 5.4, Best fitness and deflection RMS relations with the number of generations for
the 4-layer pavement system.
Table 5.4, Results of the backcalculation programs.
Program
AC
(MPa)
Slab
(MPa)
Base
(MPa)
Subgrade
(MPa)
3443.03
31131.9
158.59
51.75
(0.13)*
(0.34)*
(7.99)*
(0.08)*
3639.74
30827.1
67.57
52.40
MODULUS
(5.58)*
(0.64)*
(60.80)*
(1.33)*
10908.93
15838.0
90.95
51.83
EVERCALC
(216.44)*
(48.95)*
(47.24)*
(0.23)*
3445.656310 31036.018184 172.800925 51.606580
BackGenetic3D
(0.05)*
(0.031)*
(0.25)*
(0.2)*
* Relative error (%) of the backcalculated moduli.
MICHBACK
112
Deflection
RMSE
(%)
0.007
0.068
1.526
0.0014
It can be seen from Figure 5.4 and Table 5.4 that the BackGenetic3D was able to
optimize the entire search domain to locate the optimal solutions. In addition, the results
show that even that the other programs were able to achieve the commonly recommended
RMSE of less than 3%, the discrepencies between the actual and the backcalculated
moduli can be extremly high. The backcalculation results of EVERCALC showed the
highest discrepancy with an error of 216.44% in the top layer. In addition, the results
indicate that even if the RMSE is less than 1% the discrepancy can be high due to the
local optima effect as shown by MODULUS results where the RMSE is 0.068 while the
modulus relative error of the base layer is 60.80%.
5.14 Conclusions
A new pavement backcalculation program is developed based on the genetic
algorithms method. The new program implements many new operators and parameters
that have been used for the first time to backcalculate the pavement moduli. In addition, a
powerful and robust random number generator (Mersenne Twister Random Number
Generator) was implemented as well.
The verification problem showed that the BackGenetic3D program is superior
than many other programs that backcalculate the pavement moduli using classical
backcalculation methods. The BackGenetic3D program is designed to optimize the entire
search space looking for the global optima rather than the local optima.
113
CHAPTER VI
OPTIMIZATION OF GENETIC ALGORITHMS OPERATORS AND PARAMETERS
FOR THE BACKCALCULATION OF THE PAVEMENT ELASTIC MODULI
6.1 Introduction
Carrying out a successful genetic algorithm run to find the optimal parameters of
a function requires the use of the genetic algorithm operators properly and the selection
of the right population size. Genetic algorithm operators and strategies include crossover,
mutation, niching, elitism, and many other operators. The selection of the genetic
algorithm parameters has not been generalized in the form of rules or guidelines. This can
be attributed to the nature of each problem where the operators and strategies work as
tools that help in optimizing the search space.
The genetic algorithm operators and strategies can differ from one problem to
another. The interaction between the parameters can be one of the reasons why the same
operators cannot be generalized to all problems. In biology, two genes can have two
different effects at different loci in the genome; however, the two genes can still be
strongly linked. Interaction between genes can lead to the manifestation of certain
characteristics of the organism (epistatic gene) or the suppression of another genes
characters (hypostatic gene). In genetic algorithms, the same can be observed when the
114
optimized function contains many parameters interacting with each other (especially
nonlinear functions).
On the other hand, parameters of the optimized function are encoded as strings of
l-bit which combined consist one chromosome. Therefore, assuming a problem with three
parameters where the string length of each parameter is 18 bits, the total length of the
chromosome will be 54 bits (18 bits x 3 parameters). It is known that the size of the
search space for a bit string encoding of length l is 2l (Goldberg, 1989), which forms a
hypercube (l-dimensional hypercube) where the genetic algorithm samples the corners of
the hypercube (Whitley, 1994).
The search space size can be very large based on the exponential relation 2l. For
example, Winston (1992) pointed out that the effective size of the search space of
possible board configurations in chess is in the order of 2400. This large number is
extremely impractical and hence random search techniques should be used. Therefore,
the biased random search in genetic algorithms is suitable for such cases. The selection of
the genetic operators and strategies will play a large role in directing the search in very
large or small search spaces depending on the size of the space.
On the other hand, the mechanism of each genetic algorithm operator affects the
value of the parameter itself. For example, the values of the crossover operator when
single-crossover or multiple-crossover methods are used will be different for the same
problem. Therefore, when operators are used, the value of the operator should be taken
into account together with the method (strategy) used to implement that operator.
115
The interaction of different techniques with the GA operators can affect the value
of the operators. For example, the implementation of elitism and niching (sharing)
(Goldberg, 1989; Goldberg and Richardson, 1987) can largely affect the values of the GA
operators.
The complexity of the search space is another important factor that affects the
value of the operator. For example, searching the space of multimodal functions depends
on the complexity of the function. Multimodal functions are functions with more than one
optima with the possibility of having a single or more than one global optima with same
value. Thus, in constrained search spaces, the degree of modality (number of optima)
depends on the window (limits) of the search space, as shown in Figure 6.1. For example,
if the search space in a region is to be investigated, the value of the crossover operator
will be different than that applied to search another region within the same search space.
On the other hand, single or double crossover operations can be destructive on
individuals (solutions) near different optima.
116
Figure 6.1, Search Space of a 2-Layer Pavement Section.

6.2 Selection of GA Operators
Designing a GA to optimize a problem requires the selection of the crossover and
the mutation operators, the size of the population, and the number of generations. These
requirements impose a dilemma to the users since no standard method or guidelines are
available. However, several attempts have been carried out since Holland (Holland, 1975)
placed the foundations of the GAs.
The first systematic work to study the interaction between the GA operators was
carried out by De Jong (1975). De Jong used 5 test functions as shown below:
F1( X ) =
3
i =1
xi
, 5.12 xi 5.12
F 2( X ) = 100( x1 x 2 ) 2 + (1 x1 ) 2 , 2.048 xi 2.048
117
Eq. (6.1)
Eq. (6.2)
F 3( X ) =
5
i =1
F 4( X ) =
30
i =1
integer ( xi ) , 5.12 xi 5.12
Eq. (6.3)
, 1.28 xi 1.28
Eq. (6.4)
, 65.536 xi 65.536
Eq. (6.5)
ixi + Gauss (0,1)

1
F 5( X ) = 0.002 +
25
j =1
j+
2
i =1
( xi aij )
In his work, De Jong (1975) used roulette wheel selection, single-point crossover,
and simple mutation. It was concluded that, though, small populations can lead to quick
results, large populations are more recommended since they can lead to more accurate
results. The GA parameters were recommended based on the performance of all of the
five test functions (Eq. 6.1 through Eq. 6.5). The recommended values for crossover
probability (per pair of parents) and mutation probability (per bit) were 0.6 and 0.001,
respectively, while the recommended population size ranged between 50 and 100. In
addition, De Jong recommended the use of low mutation rates all the time. The
recommended mutation probability was equal to the inverse of the population size
(pm=1/n).
Grefenstette (1986) extended the work of De Jong (1975) by optimizing the GA
parameters of the five functions (Eq. 6.1 through Eq. 6.5) using the GA as an
optimization tool. Grefenstette (1986) used elitism to enhance the search ability. The
recommended values for crossover probability (per pair of parents) and mutation
probability (per bit) were 0.95 and 0.01,
respectively,
118
while
the
recommended
population size was 30. Grefenstette (1986) recommended the use of higher crossover
and mutation probabilities, which contradict the recommended values by De Jong (1975).
Grefenstette (1986) concluded that higher crossover probabilities (close to pc=1.0) can be
used when more accurate selection procedures are used to reduce the stochastic errors of
sampling.
Schaffer et al. (1989), studied many test functions including some of De Jongs
(De Jong, 1975) test functions to provide recommendations on the GA parameters. The
recommended values for crossover probability and mutation probability were 0.75-0.95
and 0.005-0.01, respectively, while the recommended population size ranged between 20
and 30. However, Schaffer et al. (1989) concluded that the selected GA parameters were
independent of the tested functions.
Other studies included recommendations on the selection of some of the GA
parameters. Mhlenbein (1992) and Bck (1992), working separately, recommended the
use of a mutation probability equal to the inverse of the string (chromosome) length
(pm=1/l). This recommended value is best suited for (1+1) GAs where each single parent
chromosome produces a single child by means of mutation. Deb (2005) suggested using
crossover values between 0.7 and 1.0.
As can be seen from the above review, different studies suggested different
parameter sets. None of the studies was able to show that the suggested parameters are
the optimal parameters that can be used for any problem. However, many GA users use
these parameters as a rule of thumb for optimization. It should be emphasized that
manual tuning of parameters is required especially for functions that have not been
119
studied before.
6.3 Backcalculation of Pavement Moduli
The backcalculation of the pavement moduli involves the use of forward
multilayer elastic programs. Trial moduli are used to calculate the deflections at the
Falling Weight Delectometer (FWD) sensors and then the fitness function is used in the
genetic algorithm to quantify the quality of the backcalculated moduli. The search space
in the pavement backcalculation problem is multimodal and hence the required genetic
parameters should be dealt with care. The use of genetic parameters from studies other
than those dealing with the same problem can be misleading since the complexity of the
problem and the size of the search space are unique compared to the studied functions in
many investigations.
Optimizing the pavement backcalculation problem is challenging since the size of
the search space can be very large compared to search spaces in standard genetic
algorithm optimized functions such as those studied by De Jong (1975), Grefenstette
(1986), and Schaffer et al. (1989) where the search space was limited to one region. In
the pavement backcalculation problem, each modulus has its own search region making
the interaction between the studied regions an extra challenge to the genetic algorithm
method. Pavement backcalculation problems include, normally, at least two unknown
elastic parameters (elastic moduli of layers) while other parameters are known or
assumed (layer thicknesses, Poissons ratios).
On the other hand, the selection of the search region is not the same for all moduli
(all elastic layers in the pavement system). The search space (region) for each modulus
120
should take into account the physical characteristics and the type of the layer in the
system. For example, the search space for the concrete-pavement layer will be different
than that of the base, subbase, and the subgrade layers due to differences in the physical
properties that change the elasticity characteristics of each layer. However, due to the
dependency of the elastic parameters on temperature/moisture and mechanical distresses
in the pavement (micro/macro cracks), the search space should be expanded to account
for any possible moduli that solve the problem. This in turn increases the effort required
to solve the problem since more modal points will exist in the search space. Therefore,
one can expect different genetic parameters for different search spaces. Figure 6.1 shows
the search space for a 2-layer pavement system over the selected search range. As can be
seen from the figure, the difficulty of locating the best solution (optimal solution) is
largely affected by the search window within the search space. As the search window
increases the difficulty increases.
6.4 Genetic Operators Interaction and Performance

The backcalculation of the pavement layers moduli using genetic algorithms is
very robust and effective if the right genetic operators are used. The selection of high or
low operators can highly influence the convergence of the objective function and hence
all genetic operators should be chosen carefully. For example, by definition, the selection
of high mutation rates (probabilities) can increase the randomness of the backcalculated
solutions (diversity of the chromosomes/individuals) due to the high changing rate of the
binary code that represents the parameters. On the other hand, the selection of very low
121
mutation rates can reduce the change in the chromosomes characteristics and hence will
reduce the convergence rate to the best solution (best fit of the objective function).
However, the mutation rate is not the only operator that affects the performance
and the behavior of the backcalculation process using genetic algorithms. The effect of
each parameter and the interaction between the parameters cannot be studied explicitly
using mathematical formulations due to the complexity of the problem and the complex
search space of the backcalculated parameters. Therefore, the interaction between the
genetic operators can be done more practically using sensitivity analysis.
The objectives of the scarce available pavement backcalculation researches using
genetic algorithms were aimed at demonstrating the applicability of the method (Fwa et
al., 1997;Kameyama et al., 1997; Reddy et al., 2004; Tsai et al., 2004). In these
researches, the selection of the genetic parameters was based on trial and error (in Reddy
et al., 2004, 1200 trials were used) to optimize the selected operators. The selected
operators were chosen based on the value of the fitness function and the previous
knowledge of the pavement layer moduli. The fittest solution was determined by
comparing the backcalculated moduli with the known moduli.
However, such approach suffers from several major drawbacks. The main
objective of any backcalculation procedure is to find the fittest parameters (optimum
layer moduli) where the parameters are unknown beforehand. The second drawback is
the lack of any guidelines on how to make the selection of the genetic operators more
practical rather than carrying out several economically inefficient and time consuming
runs since, in practice, the only reference for solution quality control is the displacement
122
objective function.
Another drawback is the generalization of one-problem results to other problems.
For example, the results of De Jongs (1975) test functions cannot be used to optimize all
other functions and problems since different search spaces will result in different
complexities. In addition, moduli backcalculation using genetic algorithms showed
different optimum genetic operators, as shown in Table 6.1. It can be seen that the
population size range is 60-500, the generation range is 40-593, the mutation probability
range is 0.02-0.15, and the crossover probability range is 0.85 to 0.9. The inconsistency
in the genetic parameters is evident regardless of the number of backcalculated moduli
(number of layers). The computational effort in Table 6.1 was computed, as suggested by
Deb (1995), for the optimum genetic operators. The computational effort is a measure of
the required effort and time by each set of genetic parameters to converge to the best
solution.
Table 6.1, Summary of Pavement Backcalculation Studies.
No. of
Layers
Mutation
Probability
(Pm)
Crossover
Probabilit
y
(Pc)
Study
Population
Size
No. of
Generations
(G)
Fwa et al. (1997)
60
40-150
0.15
0.85
Fwa et al. (1997)

60
120
Kameyama et al.
50
42-593
(1997)
Reddy et al. (2004)
100
60
Reddy et al. (2004)
60
60
Tsai et al. (2004)
500
50
*Computational Effort: CE=GPmPc
0.15
0.85
5.115.3
15.3
Variable
Variable
NA
3
3,4
3
0.02
0.10
NA
0.90
0.74
NA
1.08
4.44
NA
123
CE*
Reddy et al. (2004) optimized the genetic operators using a 3-layer pavement
system (100,60,0.02,0.90). In their work, additional analysis using the 3-layer system was
carried out to estimate the optimum genetic operators (60,60,0.10,0.74) that can be used
for 3- and 4-layer systems. The resulting two different sets of genetic operators show that
the selected genetic operators are not unique even for the same problem and search space
and they may not be optimal. The computational effort (CE) of the two sets can be 4.11
times the other (4.44/1.08). On the other hand, for the same number of layers, results
from Fwa et al. (1997) and Reddy et al. (2004) indicated that the computational effort
can be 3.44 times a possible available solution (15.3/4.44).
The above conclusions and comparisons assume that the search will converge to
the best solution. Fitting the displacement functions was the main objective during the
genetic algorithms iterations while the optimum solution was selected by comparing the
backcalculated moduli and the original known moduli of the pavement. The fitness
functions used in all studies are shown below:
Fwa et al. (1997):
Minimize RMSE =
Kameyama et al. (1997): Minimize RMSE k =
Reddy et al. (2004):
Minimize
RMSE =
1
m
m
i =1
Minimize
i =1
m
RMSE =
(d i Di )2
m
i =1
124
wi (d ik Dik ) 2
i =1
Tsai et al. (2004):
d i Di
Di
(d i Di )2
Eq. (6.6)
Eq. (6.7)
Eq. (6.8)
Eq. (6.9)
the backcalculated deflection at point i, Di is the measured deflection at point i, w is the
weight factor at point i, and k is the chromosome number.
As can be seen from the above equations, the fitness function in all studies was in
general the same.
6.5 Sensitivity Analysis
Sensitivity analysis was carried out to study the effect of the genetic operators on
the search space and to study the interaction between the genetic operators to aid in the
selection of the parameters even if the modulus of the elasticity is unknown beforehand,
which is the goal of the backcalculation analysis and the common case in practice.
The sensitivity analysis was carried out using the problem studied by Reddy et al.
(2004), shown in Figure 6.2. Reddy et al. (2004) investigated the backcalculation of the
elastic moduli of a 3-layer pavement system using different trials to obtain the optimum
genetic operators. A total of 1200 parameter sets were created in their study using the
ranges shown in Table 6.2. The optimum parameters shown in Table 6.2 are the
parameters of the best solution. The best solution is defined as the solution with the
highest fitness and the lowest average moduli error. The fitness function used by Reddy
et al. (2004) was minimized using:
Maximize
f =
1
1 + RMSE
Eq. (6.10)
where f is the fitness of the solution based on the root mean square error as estimated in
Eq. (6.8). The RMSE was calculated for six points along the surface of the pavement
125
located at 0,0.3,0.6,0.9,1.2, and 1.5 m from the center of the loading plate.
The effect of each genetic operator was studied by varying one of the operators
while the others are constant. Two populations (6 and 100) were used during the analysis
to study the effect of the population size on the operators as well. The number of
generations was selected to be 200 for the small population size and 100 for the large
population size, resulting in a total number of evaluations equal to 1,200 and 10,000,
respectively. The range of the parameters used in the sensitivity analysis is shown in
Table 6.3. The length of each chromosome was equal to 45.
Figure 6.2, The 3-layer pavement system used in the sensitivity analysis, after Reddy et
al. (2004).
Table 6.2, Range of GA inputs used by Reddy et al. (2004).
Parameter
Range
Optimum
Population size
20-160
100
20-160
60
0.6-0.95
0.9
0.001-0.2
0.02
126
Table 6.3, Values of parameters used in the sensitivity analysis.

Parameter
Values
Population size
6,100
200*,100**
0.0-1.0
0.001,0.01,0.1,0.5,0.7,1.0
* Small population.
** Large population.
As can be seen the crossover probability values ranged between no crossover (Pc=
0.0) to 100% crossover (Pc=1.0) and the mutation values probability values ranged
between very slow mutation (Pm= 0.001) to 100% mutation (Pc=1.0).
The goodness of the selected parameters using the fitness function given by Eq.
6.10 was further assessed using the following equation:
RMSE E =
m
i =1
(Ec i E m i )2
Eq. (6.11)
where RMSE E is the root mean square error of the backcalculated moduli of elasticity,
and Eci and Emi are the backcalculated and the pavement system moduli (actual moduli),
respectively, for layer i.
Moduli ranges of the studied pavement system used in the backcalculation
procedure are shown in Table 6.4.
Table 6.4, Moduli ranges of backcalculated moduli.
Moduli Range
Layer
(MPa)
Top Layer
1200-2800
Base Layer
240-560
Subgrade Layer
40-100
127
6.5.1 Behavior of Small Populations

The crossover probability was varied between 0.0 and 1.0 for the small population
size (population size=6) as shown in Figures 6.3 through 6.13. Each figure shows the
relations between the generation average fitness and the RMSEE of the backcalculated
moduli under a certain set of mutation and crossover rates. The generation average fitness
explains the average convergence behavior of the backcalculation procedure.
Figure 6.3a shows the effect of the mutation rate on the generation average fitness
(will refer to as the fitness) when the crossover rate is equal to zero, which implies no
crossover during the evaluations (runs). It can be seen that as the mutation rate increases,
in this example, the fitness decreases due to the increase of the randomized flipping of the
digits of the chromosomes, which in turn randomizes the backcalculated solutions
(moduli) and hence it increases the diversity of the chromosomes. When the mutation rate
is equal to one a complete chaos is produced since all chromosomes will be subjected
to mutation and hence the new generation will be completely different than the old
generation and none of the fittest chromosomes (solutions) will be preserved for the next
generation. In addition, decreasing the mutation rate increased the fitness since variation
was controlled to a minimum level and hence premature (early false) convergence to the
best solution (since no crossover) was somehow minimized.
According to De Jong (1975) the optimal mutation probability is equal to the
inverse of the population size, which is equal to 0.167 or 1/6. Mhlenbein (1992) and
Bck (1992) recommended an optimal mutation probability equals to the inverse of the
chromosome length, which is 0.02 or 1/45. It can be observed that for the studied
128
problem (3-layer pavement system) the random effect of the mutation rate on the fitness
was less observed when the mutation rate was less than 0.1. Mutation rates less than 0.01
showed the highest fitness values compared to all other mutation rates. The best mutation
rate of the studied set of parameters in Figure 6.3a was 0.001, which is 1/20 times the
recommended value by Mhlenbein (1992) and Bck (1992), and 1/167 the
recommended value by De Jong (1975).
However, Figure 6.3b shows a different effect of the fitness function on the root
mean square error of the backcalculated moduli (RMSEE). The results suggest that the
possibility of having RMSEE values close to zero (perfect match between the
backcalculated and the model elastic moduli) increases as the mutation rate increases
which contradicts the fitness values behavior in Figure 6.3a where the fitness decreased
as the mutation rate increased. In addition, a mutation rate of 0.01 showed the highest
RMSEE error while a mutation rate of 1.0 showed the highest possibility of zero RMSEE
error. This behavior suggests that high fitness values (Eq. 6.10) can be associated with
incorrect backcalculated moduli. This shows that different sets of layers moduli can be
far from the optimal layers moduli but still at the same time satisfy the fitness function.
On the other hand, Figures 6.4 through 6.7 show the effect of the variation of the
mutation rate while the crossover rate is constant (Pc =0.25,0.5,0.75, and 1.0). In general,
increasing the mutation rate beyond 0.01 increased the randomness of the fitness values
while mutation probabilities less than 0.01 exhibited higher fitness values (Figures 6.4a
through 6.7a). In addition, increasing the crossover rate reduced the RMSEE error
regardless of the mutation rate (Figures 6.4b through 6.7b). It is evident from the figures
129
that the mutation rate can enhance the values of the fitness and the RMSEE error when the
crossover is more than 0.5.
Figure 6.8 presents the relation between a constant very low mutation rate
(Pm=0.001) and variable crossover rates ranging between zero and one. A uniform
crossover probability was used as suggested by Spears and De Jong (1991) in which an
exchange happens at each bit position with a probability of Pc. It can be seen that
increasing the crossover probability for a mutation probability of 0.001 increased the
fitness values for all of the Pc values. The fittest solutions (using Eq. 6.10) had the highest
RMSEE errors (Figure 6.8b), with values up to 60% of the backcalculated moduli. The
lowest RMSEE errors were obtained when the mutation probability was 0.001 and the
crossover probability was higher than 0.75 (0.75 and 1.0). This indicates that, in small
populations, high crossover probabilities are more appropriate when the mutation
probability is very small.
Increasing the mutation probability more than 0.001 resulted in more variation of
the chromosome bits and hence increased the diversity of the new generations as shown
in Figures 6.9a through 6.13a where the fitness variation did not follow a certain trend.
High mutation rates reduced the transfer of the elite chromosomes (chromosomes with
high fitness) from one generation to another. In addition, the variation of the RMSEE
errors increased as the crossover probability increased while the mutation probability was
constant (Figures 6.9b through 6.13b). In general, the RMSEE errors decreased regardless
of the crossover probability when the mutation probability was higher than 0.1. On the
other hand, the fitness decreased regardless of the crossover probability when the
130
mutation probability was higher than 0.1. These behaviors suggest again that the use of
the fitness function alone in practice can be deceptive and hence convergence to the
optimal solution can be difficult.
It can be seen that for small populations, there is a need to increase the mutation
probability since more diversity can be induced when the mutation probability increases.
However, when the mutation rate is equal to 1.0, a complete random generation of
chromosomes will be produced.
When the population size is small the possibility that similar genes dominate the
gene pool increases due to the lack of diversity among the genes and hence the possibility
of the premature convergence increases. In general, when the crossover probability is
zero and the mutation probability is very small, the GA search becomes a hill-climbing
search. On the other hand, when the crossover probability is zero and the mutation
probability is very high, the GA search becomes a random search.
131
Generation Average Fitness
1.00000000
0.99999995
0.99999990
Pc=0,Pm=0.001
Pc=0,Pm=0.01
Pc=0,Pm=0.1
Pc=0,Pm=0.5
Pc=0,Pm=0.7
Pc=0,Pm=1.0
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
Modulus of Elasticity-RMSE (%)
100
Pc=0,Pm=0.001
Pc=0,Pm=0.01
Pc=0,Pm=0.1
Pc=0,Pm=0.5
Pc=0,Pm=0.7
Pc=0,Pm=1.0
80
60
(b)
40
20
0
0
50
100
150
200
Generation
Figure 6.3, Generation average fitness (a) and RMSE of the backcalculated moduli (b)
under variable mutation rates and constant crossover rate (Pc=0.0) for a small population.
132
1.00000000
0.99999995
0.99999990
Pc=0.25,Pm=0.001
Pc=0.25,Pm=0.01
Pc=0.25,Pm=0.1
Pc=0.25,Pm=0.5
Pc=0.25,Pm=0.7
Pc=0.25,Pm=1.0
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0.25,Pm=0.001
Pc=0.25,Pm=0.01
Pc=0.25,Pm=0.1
Pc=0.25,Pm=0.5
Pc=0.25,Pm=0.7
Pc=0.25,Pm=1.0
80
60
(b)
40
20
0
0
50
100
150
200
Generation
under variable mutation rates and constant crossover rates (Pc=0.25) for a small
population.
133
1.00000000
0.99999995
0.99999990
Pc=0.5,Pm=0.001
Pc=0.5,Pm=0.01
Pc=0.5,Pm=0.1
Pc=0.5,Pm=0.5
Pc=0.5,Pm=0.7
Pc=0.5,Pm=1.0
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0.5,Pm=0.001
Pc=0.5,Pm=0.01
Pc=0.5,Pm=0.1
Pc=0.5,Pm=0.5
Pc=0.5,Pm=0.7
Pc=0.5,Pm=1.0
80
(b)
60
40
20
0
0
50
100
150
200
Generation
population.
134
1.00000000
0.99999995
0.99999990
Pc=0.75,Pm=0.001
Pc=0.75,Pm=0.01
Pc=0.75,Pm=0.1
Pc=0.75,Pm=0.5
Pc=0.75,Pm=0.7
Pc=0.75,Pm=1.0
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0.75,Pm=0.001
Pc=0.75,Pm=0.01
Pc=0.75,Pm=0.1
Pc=0.75,Pm=0.5
Pc=0.75,Pm=0.7
Pc=0.75,Pm=1.0
80
60
(b)
40
20
0
0
50
100
150
200
Generation
population.
135
1.00000000
0.99999995
0.99999990
Pc=1.0,Pm=0.001
Pc=1.0,Pm=0.01
Pc=1.0,Pm=0.1
Pc=1.0,Pm=0.5
Pc=1.0,Pm=0.7
Pc=1.0,Pm=1.0
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=1.0,Pm=0.001
Pc=1.0,Pm=0.01
Pc=1.0,Pm=0.1
Pc=1.0,Pm=0.5
Pc=1.0,Pm=0.7
Pc=1.0,Pm=1.0
80
60
(b)
40
20
0
0
50
100
150
200
Generation
population.
136
1.00000000
0.99999995
0.99999990
Pc=0,Pm=0.001
Pc=0.25,Pm=0.001
Pc=0.5,Pm=0.001
Pc=0.75,Pm=0.001
Pc=1.0,Pm=0.001
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0,Pm=0.001
Pc=0.25,Pm=0.001
Pc=0.5,Pm=0.001
Pc=0.75,Pm=0.001
Pc=1.0,Pm=0.001
80
60
40
20
(b)
0
50
100
150
200
Generation
under constant mutation rates (Pm=0.001) and variable crossover rates for a small
population.
137
1.00000000
0.99999995
0.99999990
Pc=0,Pm=0.01
Pc=0.25,Pm=0.01
Pc=0.5,Pm=0.01
Pc=0.75,Pm=0.01
Pc=1.0,Pm=0.01
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0,Pm=0.01
Pc=0.25,Pm=0.01
Pc=0.5,Pm=0.01
Pc=0.75,Pm=0.01
Pc=1.0,Pm=0.01
80
60
40
20
(b)
0
50
100
150
200
Generation
population.
138
1.00000000
0.99999995
0.99999990
Pc=0,Pm=0.1
Pc=0.25,Pm=0.1
Pc=0.5,Pm=0.1
Pc=0.75,Pm=0.1
Pc=1.0,Pm=0.1
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0,Pm=0.1
Pc=0.25,Pm=0.1
Pc=0.5,Pm=0.1
Pc=0.75,Pm=0.1
Pc=1.0,Pm=0.1
80
60
40
20
(b)
0
0
50
100
150
200
Generation
population.
139
1.00000000
0.99999995
0.99999990
Pc=0,Pm=0.5
Pc=0.25,Pm=0.5
Pc=0.5,Pm=0.5
Pc=0.75,Pm=0.5
Pc=1.0,Pm=0.5
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0,Pm=0.5
Pc=0.25,Pm=0.5
Pc=0.5,Pm=0.5
Pc=0.75,Pm=0.5
Pc=1.0,Pm=0.5
80
60
(b)
40
20
0
0
50
100
150
200
Generation
population.
140
1.00000000
0.99999995
0.99999990
Pc=0,Pm=0.7
Pc=0.25,Pm=0.7
Pc=0.5,Pm=0.7
Pc=0.75,Pm=0.7
Pc=1.0,Pm=0.7
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
80
Pc=0,Pm=0.7
Pc=0.25,Pm=0.7
Pc=0.5,Pm=0.7
Pc=0.75,Pm=0.7
Pc=1.0,Pm=0.7
60
(b)
40
20
0
0
50
100
150
200
Generation
population for a small population.
141
1.00000000
0.99999995
0.99999990
Pc=0,Pm=1.0
Pc=0.25,Pm=1.0
Pc=0.5,Pm=1.0
Pc=0.75,Pm=1.0
Pc=1.0,Pm=1.0
0.99999985
(a)
0.99999980
0
50
100
150
200
Generation
100
Pc=0,Pm=1.0
Pc=0.25,Pm=1.0
Pc=0.5,Pm=1.0
Pc=0.75,Pm=1.0
Pc=1.0,Pm=1.0
80
60
(b)
40
20
0
0
50
100
150
200
Generation
population.
142
6.5.2 Behavior of Large Populations

The effect of the population size was further studied using a large population size
of 100 and a total number of generations of 100 (total evaluations=10,000). The
crossover probability was varied between 0.0 and 1.0 for the large population size as
shown in Figures 6.14 through 6.24. Each figure shows the relation between the
generation average fitness and the RMSEE of the backcalculated moduli under a certain
set of mutation and crossover rates.
It can be seen that increasing the population size (approximately 16.67 times the
small population size) reduced the effect of the mutation on the fitness. In Figure 6.14a,
the fitness changed slightly compared to the change in fitness in Figure 6.3a (small
population). In addition, in this example, the effect of large mutation probabilities
increased as the number of generations increased, indicating a destructive effect of high
mutation probabilities on the schema of the chromosomes. The same destructive effect of
high mutation probabilities can be seen in small populations as well (Figure 6.3a).
However, the rate of schema destruction in small populations is higher than that in large
populations. On the other hand, the small change in fitness (in Figures 6.14 through 6.24)
in large populations lead to a relatively (about 10% change in RMSEE) large change in the
backcalculated moduli indicating high sensitivity to the change in the fitness.
Furthermore, the RMSEE error sensitivity to the change in fitness is higher in small
populations compared to that in large populations.
Figures 6.14 through 6.18 show that increasing the crossover in large populations
can increase the fitness values and, at the same time, can reduce the associated RMSEE
143
errors. However, the effect of increasing the crossover was pronounced when the
crossover probability was less than 0.5 while the effect became negligible with
probabilities higher than 0.5 and less than 1.0. When the crossover probability is 1.0, all
individuals in the population will be subjected to crossover. The change in the GA
performance due to the change in the crossover probability is attributed to the change in
the selection pressure. It should be noted that, in tournament selection two individuals are
selected and compared based on the fitness values and the fittest is selected for mating. In
addition, when the mutation probability is equal to 1.0, a complete random generation of
chromosomes will be produced.
Figures 6.15 and 6.16 shows that the best performance of the GA can be achieved
(for a population size=100) when the crossover probability was between 0.25 and 0.5
while the mutation probability was less than 0.01.
144
1.00000000
Pc=0,Pm=0.001
Pc=0,Pm=0.01
Pc=0,Pm=0.1
Pc=0,Pm=0.5
Pc=0,Pm=0.7
Pc=0,Pm=1.0
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
Pc=0,Pm=0.001
Pc=0,Pm=0.01
Pc=0,Pm=0.1
Pc=0,Pm=0.5
Pc=0,Pm=0.7
Pc=0,Pm=1.0
50
(b)
40
30
20
0
20
40
60
80
100
Generation
under variable mutation rates and constant crossover rates (Pc=0.0) for a large population.
145
1.00000000
Pc=0.25,Pm=0.001
Pc=0.25,Pm=0.01
Pc=0.25,Pm=0.1
Pc=0.25,Pm=0.5
Pc=0.25,Pm=0.7
Pc=0.25,Pm=1.0
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0.25,Pm=0.001
Pc=0.25,Pm=0.01
Pc=0.25,Pm=0.1
Pc=0.25,Pm=0.5
Pc=0.25,Pm=0.7
Pc=0.25,Pm=1.0
50
40
30
20
0
20
40
60
80
100
Generation
under variable mutation rates and constant crossover rates (Pc=0.25) for a large
population.
146
1.00000000
Pc=0.5,Pm=0.001
Pc=0.5,Pm=0.01
Pc=0.5,Pm=0.1
Pc=0.5,Pm=0.5
Pc=0.5,Pm=0.7
Pc=0.5,Pm=1.0
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0.5,Pm=0.001
Pc=0.5,Pm=0.01
Pc=0.5,Pm=0.1
Pc=0.5,Pm=0.5
Pc=0.5,Pm=0.7
Pc=0.5,Pm=1.0
50
40
30
20
0
20
40
60
80
100
Generation
147
1.00000000
Pc=0.75,Pm=0.001
Pc=0.75,Pm=0.01
Pc=0.75,Pm=0.1
Pc=0.75,Pm=0.5
Pc=0.75,Pm=0.7
Pc=0.75,Pm=1.0
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0.75,Pm=0.001
Pc=0.75,Pm=0.01
Pc=0.75,Pm=0.1
Pc=0.75,Pm=0.5
Pc=0.75,Pm=0.7
Pc=0.75,Pm=1.0
50
40
30
20
0
20
40
60
80
100
Generation
under variable mutation rates and constant crossover rates (Pc =0.75) for a large
population.
148
1.00000000
Pc=1.0,Pm=0.001
Pc=1.0,Pm=0.01
Pc=1.0,Pm=0.1
Pc=1.0,Pm=0.5
Pc=1.0,Pm=0.7
Pc=1.0,Pm=1.0
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=1.0,Pm=0.001
Pc=1.0,Pm=0.01
Pc=1.0,Pm=0.1
Pc=1.0,Pm=0.5
Pc=1.0,Pm=0.7
Pc=1.0,Pm=1.0
50
40
30
20
0
20
40
60
80
100
Generation
population.
149
1.00000000
Pc=0,Pm=0.001
Pc=0.25,Pm=0.001
Pc=0.5,Pm=0.001
Pc=0.75,Pm=0.001
Pc=1.0,Pm=0.001
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0,Pm=0.001
Pc=0.25,Pm=0.001
Pc=0.5,Pm=0.001
Pc=0.75,Pm=0.001
Pc=1.0,Pm=0.001
50
40
30
20
0
20
40
60
80
100
Generation
under constant mutation rates (Pm=0.001) and variable crossover rates for a large
population.
150
1.00000000
Pc=0,Pm=0.01
Pc=0.25,Pm=0.01
Pc=0.5,Pm=0.01
Pc=0.75,Pm=0.01
Pc=1.0,Pm=0.01
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0,Pm=0.01
Pc=0.25,Pm=0.01
Pc=0.5,Pm=0.01
Pc=0.75,Pm=0.01
Pc=1.0,Pm=0.01
50
40
30
20
0
20
40
60
80
100
Generation
population.
151
1.00000000
Pc=0,Pm=0.1
Pc=0.25,Pm=0.1
Pc=0.5,Pm=0.1
Pc=0.75,Pm=0.1
Pc=1.0,Pm=0.1
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0,Pm=0.1
Pc=0.25,Pm=0.1
Pc=0.5,Pm=0.1
Pc=0.75,Pm=0.1
Pc=1.0,Pm=0.1
50
40
30
20
0
20
40
60
80
100
Generation
under constant mutation rates (Pm=0. 1) and variable crossover rates for a large
population.
152
1.00000000
Pc=0,Pm=0.5
Pc=0.25,Pm=0.5
Pc=0.5,Pm=0.5
Pc=0.75,Pm=0.5
Pc=1.0,Pm=0.5
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0,Pm=0.5
Pc=0.25,Pm=0.5
Pc=0.5,Pm=0.5
Pc=0.75,Pm=0.5
Pc=1.0,Pm=0.5
50
40
30
20
0
20
40
60
80
100
Generation
population.
153
1.00000000
Pc=0,Pm=0.7
Pc=0.25,Pm=0.7
Pc=0.5,Pm=0.7
Pc=0.75,Pm=0.7
Pc=1.0,Pm=0.7
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
(b)
Pc=0,Pm=0.7
Pc=0.25,Pm=0.7
Pc=0.5,Pm=0.7
Pc=0.75,Pm=0.7
Pc=1.0,Pm=0.7
50
40
30
20
0
20
40
60
80
100
Generation
population.
154
1.00000000
Pc=0,Pm=1.0
Pc=0.25,Pm=1.0
Pc=0.5,Pm=1.0
Pc=0.75,Pm=1.0
Pc=1.0,Pm=1.0
0.99999995
(a)
0.99999990
0.99999985
0.99999980
0
20
40
60
80
100
Generation
60
Pc=0,Pm=1.0
Pc=0.25,Pm=1.0
Pc=0.5,Pm=1.0
Pc=0.75,Pm=1.0
Pc=1.0,Pm=1.0
50
(b)
40
30
20
0
20
40
60
80
100
Generation
population.
155
6.5.3 Behavior of Large Populations Using Different Ranges of Moduli

The effect of the search domain on the performance of the backcalculation
process was further studied using different ranges of pavement moduli, as shown in Table
6.5. The large population size was compared according to the change in moduli ranges to
further understand the complexity of the search domain. Figures 6.25 through 6.35 show
the results of the new ranges of moduli.
Table 6.5, Moduli ranges of backcalculated moduli.
New Moduli
Old Moduli
Layer
Range (MPa)
Range (MPa)
Top Layer
800-3200
1200-2800
Base Layer
160-640
240-560
Subgrade Layer
28-112
40-100
It can be seen that changing the moduli ranges by increasing the search domain of
each moduli increased the fitness and reduced the associated RMSEE error. In general the
RMSEE errors reduced by approximately 85% by changing the ranges of the moduli while
the increase in the fitness values was relatively small.
Figures 6.14 and 6.25 indicate different performances of the GA. In Figure 6.14b
at low mutation probabilities (less 0.1), and while there is no crossover (Pc=0), the effect
of the mutation was almost negligible on the fitness values and moderate on the
associated RMSEE error while in Figure 6.25b the effect was moderate and high,
respectively.
Figures 6.26 through 6.29 indicate that the fitness values increase and the
associated RMSEE errors decrease by increasing the crossover probability. On the other
156
hand, Figures 6.26 through 6.29 and Figures 6.15 through 6.18 show different responses
to the mutation and crossover probabilities. In Figures 6.26 through 6.29 increasing the
mutation increased the fitness and reduced he associated RMSEE error while in Figures
6.15 through 6.18, the increase in the mutation probability had the opposite effect by
reducing the fitness and increasing the associated RMSEE errors. The same observations
can be drawn by comparing Figures 6.30 through 6.35 and Figures 6.19 through 6.24.
The above example shows that due to the complexity of the search domain and
the interaction between the GA parameters and operators, recommending one set than can
be used for any problem is not feasible. In addition, the selection of the moduli range is
as important as selecting the genetic operators and parameters.
Increasing the range of the unknown pavement moduli is recommended since the
algorithm will be able to search for the optimal solution within a wide range of solution
(as in the current ranges of moduli) rather than getting stuck with a subdomain (as in the
previous ranges of moduli) leading to incorrect conclusions and backcalculated moduli.
However, increasing the moduli ranges will increase the effort of the GA and can lead to
different results than desired. The physical characteristics of the layers should be used as
guidelines when selecting the appropriate ranges of moduli. Furthermore, Table 6.5 can
be used as a guideline for the selection of the appropriate ranges of moduli based on the
description of the pavement material. However, based on the authors experience and to
give more flexibility to the GA search, Table 6.6 included some recommended ranges to
be used in the backcalculation of the pavement moduli using GA.
157
Table 6.6, Typical values of modulus of elasticity for pavement materials.

Range*
Recommended
Material
(MPa)
Range (MPa)
Hot-Mix Asphalt
1,500-3,500
1,000-4,000
Portland Cement Concrete
20,000-55,000
18,000-60,000
Asphalt-Treated Base
500-3,000
300-3,5000
Cement-Treated Base
3,500-7,000
2,500-8,000
Lean Concrete
7,000-20,000
6,000-25,000
Granular Base
100-350
80-450
Granular Subgrade Soil
50-150
30-250
Fine-Grained Subgrade Soil
20-50
10-100
* After (AASHTO, 1993).
158
1.00000000
(a)
Pc=0,Pm=0.001
Pc=0,Pm=0.01
Pc=0,Pm=0.1
Pc=0,Pm=0.5
Pc=0,Pm=0.7
Pc=0,Pm=1.0
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0,Pm=0.001
Pc=0,Pm=0.01
Pc=0,Pm=0.1
Pc=0,Pm=0.5
Pc=0,Pm=0.7
Pc=0,Pm=1.0
20
15
10
0
0
20
40
60
80
100
Generation
under variable mutation rates and constant crossover rates (Pc=0.0) for a large population
(Different ranges of moduli).
159
1.00000000
(a)
Pc=0.25,Pm=0.001
Pc=0.25,Pm=0.01
Pc=0.25,Pm=0.1
Pc=0.25,Pm=0.5
Pc=0.25,Pm=0.7
Pc=0.25,Pm=1.0
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0.25,Pm=0.001
Pc=0.25,Pm=0.01
Pc=0.25,Pm=0.1
Pc=0.25,Pm=0.5
Pc=0.25,Pm=0.7
Pc=0.25,Pm=1.0
20
15
10
0
0
20
40
60
80
100
Generation
under variable mutation rates and constant crossover rates (Pc=0.25) for a large
population. (Different ranges of moduli).
160
1.00000000
(a)
Pc=0.5,Pm=0.001
Pc=0.5,Pm=0.01
Pc=0.5,Pm=0.1
Pc=0.5,Pm=0.5
Pc=0.5,Pm=0.7
Pc=0.5,Pm=1.0
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0.5,Pm=0.001
Pc=0.5,Pm=0.01
Pc=0.5,Pm=0.1
Pc=0.5,Pm=0.5
Pc=0.5,Pm=0.7
Pc=0.5,Pm=1.0
20
15
10
0
0
20
40
60
80
100
Generation
(Different ranges of moduli).
161
1.00000000
Pc=0.75,Pm=0.001
Pc=0.75,Pm=0.01
Pc=0.75,Pm=0.1
Pc=0.75,Pm=0.5
Pc=0.75,Pm=0.7
Pc=0.75,Pm=1.0
(a)
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0.75,Pm=0.001
Pc=0.75,Pm=0.01
Pc=0.75,Pm=0.1
Pc=0.75,Pm=0.5
Pc=0.75,Pm=0.7
Pc=0.75,Pm=1.0
20
15
10
0
0
20
40
60
80
100
Generation
162
1.00000000
(a)
Pc=1.0,Pm=0.001
Pc=1.0,Pm=0.01
Pc=1.0,Pm=0.1
Pc=1.0,Pm=0.5
Pc=1.0,Pm=0.7
Pc=1.0,Pm=1.0
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=1.0,Pm=0.001
Pc=1.0,Pm=0.01
Pc=1.0,Pm=0.1
Pc=1.0,Pm=0.5
Pc=1.0,Pm=0.7
Pc=1.0,Pm=1.0
20
15
10
0
0
20
40
60
80
100
Generation
163
1.00000000
(a)
Pc=0,Pm=0.001
Pc=0.25,Pm=0.001
Pc=0.5,Pm=0.001
Pc=0.75,Pm=0.001
Pc=1.0,Pm=0.001
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0,Pm=0.001
Pc=0.25,Pm=0.001
Pc=0.5,Pm=0.001
Pc=0.75,Pm=0.001
Pc=1.0,Pm=0.001
20
15
10
0
0
20
40
60
80
100
Generation
164
1.00000000
(a)
Pc=0,Pm=0.01
Pc=0.25,Pm=0.01
Pc=0.5,Pm=0.01
Pc=0.75,Pm=0.01
Pc=1.0,Pm=0.01
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0,Pm=0.01
Pc=0.25,Pm=0.01
Pc=0.5,Pm=0.01
Pc=0.75,Pm=0.01
Pc=1.0,Pm=0.01
20
15
10
0
0
20
40
60
80
100
Generation
165
1.00000000
(a)
Pc=0,Pm=0.1
Pc=0.25,Pm=0.1
Pc=0.5,Pm=0.1
Pc=0.75,Pm=0.1
Pc=1.0,Pm=0.1
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0,Pm=0.1
Pc=0.25,Pm=0.1
Pc=0.5,Pm=0.1
Pc=0.75,Pm=0.1
Pc=1.0,Pm=0.1
20
15
10
0
0
20
40
60
80
100
Generation
under constant mutation rates (Pm=0. 1) and variable crossover rates for a large
166
1.00000000
(a)
Pc=0,Pm=0.5
Pc=0.25,Pm=0.5
Pc=0.5,Pm=0.5
Pc=0.75,Pm=0.5
Pc=1.0,Pm=0.5
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0,Pm=0.5
Pc=0.25,Pm=0.5
Pc=0.5,Pm=0.5
Pc=0.75,Pm=0.5
Pc=1.0,Pm=0.5
20
15
10
0
0
20
40
60
80
100
Generation
167
1.00000000
Pc=0,Pm=0.7
Pc=0.25,Pm=0.7
Pc=0.5,Pm=0.7
Pc=0.75,Pm=0.7
Pc=1.0,Pm=0.7
(a)
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
(b)
Pc=0,Pm=0.7
Pc=0.25,Pm=0.7
Pc=0.5,Pm=0.7
Pc=0.75,Pm=0.7
Pc=1.0,Pm=0.7
20
15
10
0
0
20
40
60
80
100
Generation
168
1.00000000
(a)
Pc=0,Pm=1.0
Pc=0.25,Pm=1.0
Pc=0.5,Pm=1.0
Pc=0.75,Pm=1.0
Pc=1.0,Pm=1.0
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
25
Pc=0,Pm=1.0
Pc=0.25,Pm=1.0
Pc=0.5,Pm=1.0
Pc=0.75,Pm=1.0
Pc=1.0,Pm=1.0
20
(b)
15
10
0
0
20
40
60
80
100
Generation
169
6.5.4 Best Fitness and Average Fitness Comparison

The use of the best fitness instead of the average fitness in the sensitivity analysis
to optimize the selection of the genetic parameters can be argued to be less favorable.
Understanding the meaning of both the average fitness and the best fitness can help in
understanding the advantages and the disadvantages of using both in the sensitivity
analysis. The best fitness value represents the fitness of the fittest chromosome (moduli)
within the population while the average fitness value represents the average of all fitness
values (high and low) within the population. The selection process in the genetic
algorithms favors the chromosomes with high fitness values, therefore, the fittest
chromosome will be transferred (sometimes with the help of elitism) from one generation
to the other. However, the crossover and the mutation processes will work together to
produce chromosomes with higher fitness values from the selected chromosomes. This
indicates that the fitness of the individuals within the population will increase with time
and one of the best solutions will dominate the population after several generations an
can lead to local convergence if the population size is small. The convergence to a
solution does not mean always, as was shown previously, that the fittest chromosome is
the correct solution even if the fitness value is high.
Therefore, if one uses the best fitness as the only measure to select the genetic
algorithms parameters, the best fitness can be misleading since it will show the
performance of the best chromosome only while the mechanism of the search process and
the interaction between the genetic parameters can not be studied. In addition, since the
genetic algorithms method works as a guided stochastic search technique, it is more
170
appropriate to study the behavior of averages rather than single values. The average
fitness value shows the performance of the entire population and how the selection,
crossover, and mutation processes interact with each other within the current search
space. When the best fitness used in the sensitivity analysis, the selection of the number
of generations and the population size can be done as soon as the best fitness dominates
for several generations assuming that the solution converged to a local subregion within
the search space. Therefore, the selection of the associated population size and the
number of generations can be done more efficiently if the entire population (average
fitness) is considered where the convergence to a local subregion can be observed
closely.
On the other hand, the average fitness of the population is a good indicator of the
convergence of the population to local optima. It can be said that when the average
fitness of the population is close to the best fitness, the resulting solution is the optimal
solution. In other words, when the average fitness of the population does not change with
time (from generation to generation), the resulting solution is the optimal solution using
the current GA parameters and operators. Therefore, the average fitness is more
appropriate to study the convergence of the search and hence the selection of the required
population size can be done efficiently.
The diversity of the individuals in the population can be studied easily using the
average fitness rather than the best fitness. The variation of the average fitness with
generations indicates that no chromosome is dominant and hence the population is
diverse. The GA parameters, GA operators, and the population size control the diversity
171
of the population. When the average fitness of several populations is constant (or almost
constant), the population becomes dominated by one or few chromosomes and hence the
effect of increasing the number of generations is negligible since the GA parameters and
operators will not be able to produce new chromosomes. The best fitness gives limited
information regarding the diversity of the population and hence it cannot be used to study
the performance of the GA parameters and operators. Therefore, the average fitness is
more appropriate to study the interaction between the GA parameters and their effect on
the GA search process.
Figures 6.36 through 6.46 show the best fitness value in each generation during
the genetic algorithms search process, for the pavement section and GA parameters
described in Section 6.5.2. The figures show that the use of the best fitness in the
sensitivity analysis to select the GA parameters is not appropriate compared to the
average fitness values as shown in Figures 6.14 through 6.24. Figures 6.36 through 6.46
show that the change in the best fitness value can be slow when both the crossover and
the mutation probabilities are low while the change is high when either the crossover
probability or the mutation probability is high. In addition, the best fitness (for sensitivity
analysis) relations in Figures 6.36 through 6.46 may suggest the same performance of the
genetic parameters on the search space (eg. Figure 6.41) where the best fitness values
from different genetic parameter sets have close values. However, the use of the average
fitness values in Figures 6.14 through 6.24 show different trends and better conclusions
can be obtained.
172
Generation Best Fitness
0.999999998
0.999999996
0.999999994
0.999999992
Pc=0,Pm=0.001
Pc=0.25,Pm=0.001
Pc=0.5,Pm=0.001
Pc=0.75,Pm=0.001
Pc=1.0,Pm=0.001
0.999999990
0.999999988
0
20
40
60
80
100
Generation
Figure 6.36, Generation best fitness of the backcalculated moduli under a constant
mutation rate (Pm=0.001) and variable crossover rates for a large population.
0.999999998
0.999999996
0.999999994
Pc=0,Pm=0.01
Pc=0.25,Pm=0.01
Pc=0.5,Pm=0.01
Pc=0.75,Pm=0.01
Pc=1.0,Pm=0.01
0.999999992
0.999999990
0.999999988
0.999999986
0
20
40
60
80
100
Generation
173
0.999999998
0.999999996
0.999999994
Pc=0,Pm=0.1
Pc=0.25,Pm=0.1
Pc=0.5,Pm=0.1
Pc=0.75,Pm=0.1
Pc=1.0,Pm=0.1
0.999999992
0.999999990
0.999999988
0.999999986
0.999999984
0.999999982
0
20
40
60
80
100
Generation
0.999999998
0.999999996
Pc=0,Pm=0.5
Pc=0.25,Pm=0.5
Pc=0.5,Pm=0.5
Pc=0.75,Pm=0.5
Pc=1.0,Pm=0.5
0.999999994
0.999999992
0.999999990
0.999999988
0.999999986
0.999999984
0
20
40
60
80
100
Generation
174
1.000000000
Pc=0,Pm=0.7
Pc=0.25,Pm=0.7
Pc=0.5,Pm=0.7
Pc=0.75,Pm=0.7
Pc=1.0,Pm=0.7
0.999999999
0.999999998
0.999999997
0.999999996
0.999999995
0.999999994
0.999999993
0.999999992
0
20
40
60
80
100
Generation
0.999999998
0.999999996
0.999999994
Pc=0,Pm=1.0
Pc=0.25,Pm=1.0
Pc=0.5,Pm=1.0
Pc=0.75,Pm=1.0
Pc=1.0,Pm=1.0
0.999999992
0.999999990
0.999999988
0.999999986
0.999999984
0
20
40
60
80
100
Generation
175
0.999999998
0.999999996
0.999999994
0.999999992
0.999999990
Pc=0,Pm=0.001
Pc=0,Pm=0.01
Pc=0,Pm=0.1
Pc=0,Pm=0.5
Pc=0,Pm=0.7
Pc=0,Pm=1.0
0.999999988
0.999999986
0.999999984
0
20
40
60
80
100
Generation
Figure 6.42, Generation best fitness of the backcalculated moduli under variable mutation
rates and a constant crossover rate (Pc=0.0) for a large population.
0.999999996
0.999999994
Pc=0.25,Pm=0.001
Pc=0.25,Pm=0.01
Pc=0.25,Pm=0.1
Pc=0.25,Pm=0.5
Pc=0.25,Pm=0.7
Pc=0.25,Pm=1.0
0.999999992
0.999999990
0.999999988
0.999999986
0.999999984
0.999999982
0
20
40
60
80
100
Generation
176
0.999999998
0.999999996
0.999999994
Pc=0.5,Pm=0.001
Pc=0.5,Pm=0.01
Pc=0.5,Pm=0.1
Pc=0.5,Pm=0.5
Pc=0.5,Pm=0.7
Pc=0.5,Pm=1.0
0.999999992
0.999999990
0.999999988
0
20
40
60
80
100
Generation
0.999999998
0.999999996
0.999999994
Pc=0.75,Pm=0.001
Pc=0.75,Pm=0.01
Pc=0.75,Pm=0.1
Pc=0.75,Pm=0.5
Pc=0.75,Pm=0.7
Pc=0.75,Pm=1.0
0.999999992
0.999999990
0.999999988
0
20
40
60
80
100
Generation
177
0.999999998
0.999999996
0.999999994
Pc=1.0,Pm=0.001
Pc=1.0,Pm=0.01
Pc=1.0,Pm=0.1
Pc=1.0,Pm=0.5
Pc=1.0,Pm=0.7
Pc=1.0,Pm=1.0
0.999999992
0.999999990
0
20
40
60
80
100
Generation
6.6 Conclusions
The backcalculation of the moduli of a pavement system is a very complex
process since no closed solution is available or possible for the backcalculation process.
In addition, multimodal search spaces can be very complex and may show premature
convergence. Therefore, the selection of the genetic operators is very important since
poor operators can lead to a premature convergence to a local optimum rather than a
global optimum as a result of schema deceptiveness.
It was shown that achieving the same fitness value using different sets of the
genetic operators and parameters is possible. However, defining a set of parameters to be
used with all GA optimizations showed to be difficult and far from being practical.
178
The proposed sensitivity analysis showed a more practical and easy tool to
investigate the effects of the genetic parameters and operators on the performance of the
genetic algorithm and search space. In addition the sensitivity analysis is more
appropriate than creating hundreds of trial and error runs based on sets of the genetic
operators and parameters. For example, in the trial and error method reported by Reddy et
al. (2004) approximately 1200 runs were used to determine the optimal genetic operators
and parameters. Using the proposed sensitivity analysis in this chapter, the GA operators
and parameters can be determined using 30 runs only.
In addition, the selection of the pavement moduli ranges is as important as
selecting the GA operators and parameters. It is recommended that typical values of
material moduli be used in the GA backcalculation procedure to avoid local optimal
solutions.
179
CHAPTER VII
FACTORS AFFECTING MODULI BACKCALCULATION
7.1 Introduction
Sensitivity analysis was performed to study the effect of the genetic parameters
(population size, total number of generations, mutation rate, and crossover rate).
However, other genetic parameters can affect the quality and performance of the
optimization process. In the previous pavement studies by Fwa et al. (1997), Kameyama
et al. (1997), Reddy et al. (2004), and Tsai et al. (2004), the trials were performed
ignoring the effect of the chromosome length, the effect of increasing the number of
layers, and the effect of the fitness function. The performance of the genetic algorithms
does not depend on the performance of the population size and the mutation and
crossover probabilities only. The genetic and evolutionary algorithms work as the natural
selection in nature where each operator works on selecting the fittest individuals that can
survive the existing conditions and environment. Therefore, relying mainly on one
operator or on a partial set of operators can make the selection of the fittest individuals (if
the survival of the fittest is retained) a lengthy process. However, in genetic algorithms
the development of new methods is still ongoing to better simulate the natural selection
and to produce more effective algorithms and operators. Thus, one should consider the
180
interaction between the most effective factors that help in solving the problem and
reducing the possibility of early false (premature) convergence.
7.2 Chromosome Length

Chromosomes are formed by encoding the randomly generated solutions (trial
pavement moduli) into binary bits (0 and 1). The chromosome length is equal to the total
number of digits (length of binary bits) of each encoded solution multiplied by the
number of solutions. Therefore, mutation and crossover rates can have different effects
on the chromosome length depending on the length of the chromosome. For example, a
011011 chromosome (length=6) will have a different response to mutation and crossover
compared to a 0111011101 chromosome (length=10) when encoding two solutions where
3 binary bits were used for each solution in the first chromosome while 5 binary bits were
used for each solution in the second chromosome.
The mutation will be carried out by randomly flipping the binary bits (1 becomes
0 and 0 becomes 1), therefore, the probability of flipping the same position in the
chromosome is higher when the chromosome length is small (for example 6 versus 10)
while the probability of flipping the same position in long chromosomes is lower. This
variation in probability will affect the diversity of the chromosomes in the new
generation and hence can lead to a premature convergence (false convergence).
On the other hand, the length of the chromosome will affect the number of
possible solutions associated with that chromosome according to the schema theorem.
According to the theorem, using binary codes, there exist 2l possible solutions associated
181
with each gene of length l. Therefore, a gene 5 string-long will have 25=32 possibilities
while a gene 20 string long will have 220=1,048,576 possibilities (trial solutions).
Increasing the chromosome length is expected to increase the candidate solutions that the
genetic algorithm can explore randomly. Using a complete exhaustive search will be
impossible for large chromosome lengths due to time and computational constrains.
The effect of the chromosome length was ignored in all of the pavement
backcalculation studies assuming negligible effect. The chromosome length in Fwa et al.
(1997), Kameyama et al. (1997), Reddy et al. (2004), and Tsai et al. (2004) was not
reported indicating that the studies neglected the effect of the chromosome length
assuming negligible or no effect on the performance of the backcalculation process. The
BackGenetic3D program can be used for unlimited number of layers and hence the
number of backcalculated parameters will be unlimited based on the user preference and
the problem setup. The effect of the chromosome length should be investigated since
short chromosomes can affect the performance of the genetic algorithm operators and
hence the backcalculated moduli.
7.2.1 Chromosome Length and GA Performance
The effect of the chromosome length is studied using two different problems
where the length of the chromosome was varied while the performance of the genetic
operators and the backcalculated moduli was observed.
The first example involved the study of the performance of a small number of
layers. The example of Reddy et al. (2004) was used to investigate the effect of the
chromosome length on a 3 layer pavement system. In the second example, the effect of
182
the chromosome length on the performance of the backcalculation process was evaluated
using 23 layers. Both examples are discussed in details in the following sections.
The best solution is defined as the solution with the highest fitness and the lowest
average moduli error. The used fitness function is:
RMSE =
m
i =1
Maximize f =
(d i Di )2
1
1 + RMSE
Eq. (7.1)
Eq. (7.2)
the backcalculated deflection at point i, Di is the measured deflection at point i, and f is
the fitness of the solution based on the root mean square error as estimated in Eq. (7.1).
The goodness of the selected parameters using the fitness function in Eq. (7.2)
was further assessed using the following equation:
RMSE E =
3
i =1
(Ec i E m i )2
Eq. (7.3)
where RMSE E is the root mean square error of the backcalculated elasticity moduli, and
Eci and Emi are the backcalculated and the sensitivity pavement system moduli (actual),
7.2.1.1 Genetic Algorithm Performance
The 3-layer pavement system shown in Figure 7.1 was analyzed using the genetic
operators shown in Table 7.1 and the moduli ranges shown in Table 7.2. These values
were selected arbitrarily. However, the selected genetic operators showed relatively good
183
performance for large populations based on the sensitivity analysis in a previous chapter.
The selected parameters may not be the optimal parameters and used here to facilitate
studying the GA performance only. The RMSEE was calculated for six points along the
surface of the pavement located at 0,0.3,0.6,0.9,1.2, and 1.5 m from the center of the
loading plate.
The data of the 23-layer pavement system is given in Table 7.3 while the genetic
operators and the moduli ranges of the backcalculated layers are shown in Tables 7.4 and
6.5, respectively. The 23-layer pavement system represents the modulus variation in
pavement due to temperature variation with depth. The arbitrarily selected genetic
operators showed relatively good performance for large populations based on the
sensitivity analysis in a previous chapter. The selected parameters may not be the optimal
parameters and used here to facilitate studying the GA performance only. The radius of
the loaded plate was 0.1105 m while the applied load was 690kPa. The RMSEE was
calculated for seven points along the surface of the pavement located at
0,0.203,0.305,0.457,0.610, 0.914, and 1.524 m from the center of the loading plate.
184
Figure 7.1, The 3-layer pavement system used in the chromosome length analysis, after
Reddy et al. (2004).
Table 7.1, Genetic parameters used in the backcalculation process (3 layers).

Parameter
Value
Population size
100
100
0.75
0.001
Table 7.2, Moduli ranges for the backcalculation process (3 layers).

Moduli Range
Layer
(MPa)
Top Layer
1200-2800
Base Layer
240-560
Subgrade Layer
40-100
185
Table 7.3, Data of the 23-layer pavement example.

Thickness
Modulus
Description
Layer Number
(m)
(MPa)
1
0.0075
5058
2
0.0075
4846
3
0.0075
4642
4
0.0075
4446
5
0.0075
4258
6
0.0075
4078
7
0.0075
3906
8
0.0075
3742
9
0.0075
3586
10
0.0075
3438
Asphalt Concrete
11
0.0075
3298
12
0.0075
3166
13
0.0075
3042
14
0.0075
2926
15
0.0075
2818
16
0.0075
2718
17
0.0075
2626
18
0.0075
2542
19
0.0075
2466
20
0.0075
2398
Base Layer
21
0.25
700
Subbase Layer
22
0.25
300
Subgrade Layer
23
100
Poissons
ratio
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
0.30
Parameter
23 Layers
Population size
100
10
0.75
0.001
186
Layer
Moduli range
Description
Number
(MPa)
1
2000-8000
2
1900-7800
3
1900-7400
4
1800-7100
5
1700-6800
6
1600-6500
7
1600-6300
8
1500-5600
9
1400-5700
10
1400-5500
Asphalt Concrete
11
1300-5300
12
1300-5000
13
1200-5000
14
1200-5000
15
1100-4500
16
1100-4300
17
1100-4200
18
1000-4100
19
1000-4000
20
1000-4000
Base Layer
21
280-1120
Subbase Layer
22
120-480
Subgrade Layer
23
40-160
The 3-layer pavement system was studied using five different chromosome
lengths. The assumed chromosome lengths were 18,30,45,60, and 90 binary bits
corresponding to parameter lengths of 6,10,15,20, and 30 binary bits. Therefore, a wide
range of chromosome lengths was covered in the analysis to extensively study the effect
of the chromosome length on the backcalculation process.
Figure 7.2 shows the effect of the chromosome length and the relation between
the number of generations, the generation average fitness, and the moduli root mean
187
square error (RMSEE) for the 3-layer example. Figure 7.2a indicates that there is a
negligible effect of the chromosome length on the average fitness of the population when
the number of generations is larger than 50 (corresponding to more than 5000
evaluations). However, the figure indicates that the required number of generations to
achieve high fitness values decreases as the length of the chromosome increases. This
performance indicates that the genetic operators perform better when the chromosome
length is larger than 30 binary bits (a trial modulus of 10-binary bits long) and the
number of generations is less than 50. In addition, for large number of generations (larger
than 50) the chromosome length effect on the fitness value is negligible.
On the other hand, Figure 7.2b shows the effect of the chromosome length on the
RMSEE of the moduli. As shown in the figure, the best performance (lowest RMSEE) was
achieved when the chromosome length was more than 30 binary bits and the number of
generations is larger than 50.
The 23-layer pavement system was studied using five different chromosome
lengths. The assumed chromosome lengths were 46, 230,345,690, and 1150 binary bits
corresponding to encoded solution lengths of 2,10,15,30, and 50 binary bits, respectively.
Figure 7.3 shows the effect of the chromosome length and the relation between
the number of generation, generation average fitness, and the root mean square error
(RMSEE) for the 23-layer example. The figure shows that the use of the fitness function
as the only measure of the quality and accuracy of the backcalculated moduli can be
misleading. In Figure 7.3a, the fitness of the backcalculated displacements was the
highest when the chromosome length was 46 while the corresponding RMSEE value was
188
the highest (233%) in Figure 7.3b indicating a poor performance. On the other hand, for
chromosome lengths larger than 46, the fitness values increased as the chromosome
lengths increased (Figure 7.3a) while the RMSEE values decreased as the chromosome
lengths increased (Figure 7.3b). The effect of the chromosome length was negligible
when the length was larger than 345 binary bits corresponding to an encoded solution
length of 15 binary bits.
In addition, Figure 7.3a shows that when the chromosome length is 46, the
algorithm converges to a local optimum very quickly. This can be attributed to the fact
that with a gene length of 2, the corresponding possibilities will be 22=4, making the
search space relatively small and inappropriate for a problem of 23 unknown moduli.
Moreover, the crossover and mutation processes were not able to provide enough
diversity to avoid the premature convergence. On the other hand, Figure 7.3 shows that it
is possible to achieve high fitness values while the backcalculated moduli are far from the
actual ones.
It can be concluded that the chromosome length affects the performance of the
backcalculation of the moduli when the number of layers is either small (for example 3
layers) or large (for example 23 layers). Based on the above findings, it is recommended
that each trial solution be encoded using 15 binary bits or more to increase the accuracy
of the backcalculated moduli by reducing the RMSEE values. Therefore, the chromosome
length when backcalculating the moduli should be larger than 15n where n is the total
number of the unknown moduli or layers.
189
1.00000000
(a)
Chrom=18
Chrom=30
Chrom=45
Chrom=60
Chrom=90
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
20
40
60
80
100
Generation
14
(b)
Chrom=18
Chrom=30
Chrom=45
Chrom=60
Chrom=90
12
10
8
6
4
2
0
0
20
40
60
80
100
Generation
Figure 7.2, Generation average fitness (a) and RMSE of the backcalculated moduli (b) for
the 3-layer example.
190
0.99999894
0.99999892
0.99999890
Chrom=46
Chrom=230
Chrom=345
Chrom=690
Chrom=1150
0.99999888
0.99999886
(a)
0.99999884
0.99999882
0.99999880
0.99999878
0.99999876
0
10
Generation
250
200
Chrom=46
Chrom=230
Chrom=345
Chrom=690
Chrom=1150
150
(b)
100
50
0
0
10
Generation
the 23-layer example.
191
7.3 Importance of The Fitness Function

The genetic operators in genetic algorithms mimic the natural selection in nature
by selecting the fittest individuals in the population and by transferring the selected
individuals from one generation to the next. The genetic operators work as hill climbers
in the local optima domain within the global search domain.
The fitness function in genetic algorithms directs the search process toward the
location of the global optimum in the search domain. Therefore, if the used fitness
function is not appropriate for the optimization of the problem, the direction of the search
will deviate from the right direction and hence the global optima will not be found due to
a premature convergence toward a local optimum. The fitness function should be
assigned properly to ensure effective optimization of the search space.
In multimodal optimization problems, the selection of the fitness function is vital
since the complexity of the search domain can cause a premature convergence toward
one of the local optima. When the premature convergence occurs, the genetic operators
work on the local search domain (subdomain of the global search domain) and hence the
best optimum in that domain is found. Due to the stochastic nature of the genetic
algorithms, the search for the global optimum can be a matter of luck if the selection
function does not have the ability to properly direct the hard work of the genetic
operators to find the optimum.
In addition, the premature convergence is characterized by a remarkable similarity
of the individuals indicating low diversity of the population (Minqiang and Jisong, 2001).
When the similarity of the individuals in the population increases, the ability of the
192
genetic operators to create better individuals decreases, reducing the order of the created
schema. High mutation probabilities have been used to improve the genetic algorithm
performance when dealing with complex multimodal optimizations but with little
improvement (Minqiang and Jisong, 2001). It was shown in a previous chapter that the
use of high mutation rates in the backcalculation of the moduli can be inappropriate
sometimes appropriate, since the destruction of the schemata in the backcalculation
process increases by increasing the mutation probability.
7.3.1 Search and Selection Mechanisms
In addition to the directional effect of the fitness functions, they are used to
measure the quality of the selected individuals (chromosomes) and hence they provide
feedback to the genetic algorithm. The feedback is then used by the selection process to
create the next generation of chromosomes. Tournament selection is the most common
method of selection in genetic algorithms. In the tournament selection a number of
individuals is selected then the individual with the best fitness is selected based on its
fitness value provided by the fitness function. Tournament selection (implemented in
BackGenetic3D) is performed on all individuals that are involved in the reproduction
process including the crossover and mutation.

The genetic operators create new individuals (children) with better characteristics
than the previous ones (parents) with some inherited characteristics or totally new
characteristics based on the operators values and the interaction between the operators.
However, the effect of the new individuals (chromosomes) can be minimal on the fitness
values as measured by the fitness function due to the nature of the fitness landscape. In
193
smooth fitness landscapes, small and large changes to the chromosomes cause in turn
small and large changes, respectively, to the associated fitness values (strong causality)
indicating that there is a correlation between the change in the chromosome and the
change in the fitness. Strong causality exists mainly in smooth landscapes. Weak
causality indicates cases where changes in the chromosome cause changes in the fitness
with little or no correlation between both changes (Kauffman, 1993). In some search
domains the landscape is rough (not smooth) and hence the possibility of converging to a
local subdomain with a local optimum is more likely. The selection of the fitness function
largely affects the values associated with the chromosomes and hence affects the rate of
change of the fitness. Therefore, it is more appropriate to select a fitness function
characterized by strong casualty since small change rates are more desirable (Rosca and
Ballard, 1995).
In order to understand the characteristics of the search landscape of the fitness and
the search landscape of the associated RMSEE errors, Figures 7.4 and 7.5 are drawn for
large and small population sizes based on the results of the 3-layer example (original
ranges of moduli). The figures show that the fitness landscape is smooth regardless of the
population size (Figures 7.4a and 7.5a). On the other hand, Figure 7.4b show smooth
search landscapes for the RMSEE regardless of the mutation probability. However, Figure
7.5a shows that the search landscape for the RMSEE can be smooth or rough at different
generation numbers indicating different sensitivities to the change in the fitness values.
Since the only information that we have in the real moduli backcalculation problems are
the displacements data and the feedback from the fitness function, the need for smooth
194
landscapes for both the fitness and the associated RMSEE errors is vital. Therefore, large
populations are favorable when backcalculating the pavement moduli since the quality of
the feedback by the fitness function can somehow be used to measure the quality of the
backcalculated moduli.
195
12
(a) Population Size=100

Pc=0.25,Pm=0.001

RMSE
10
0.99999998
8
0.99999996
6
0.99999994
4
0.99999992
0.99999990
1.00000000
0
0
20
40
60
80
100
Generation
1.00000000
10
RMSE
0.99999998
0.99999996
0.99999994
0.99999992
0.99999990
0
0
20
40
60
80
100
Generation
Figure 7.4, Generation average fitness and RMSE of the backcalculated moduli for the
large population in the 3-layer example using (a) Pc=0.25 and Pm=0.001, and (b) Pc=0.25
and Pm=0.7.
196
(b) Population Size=100

Pc=0.25,Pm=0.7
60
50
0.99999998
40
0.99999996
30
0.99999994
20
0.99999992
(a) Population Size=6
Pc=0.25,Pm=0.001
0.99999990
0
20
10

RMSE
1.00000000
0
40
60
80
100
Generation
40
0.99999998
30
0.99999996
20
0.99999994
0.99999992
10
0.99999990
0
0.99999988

RMSE
(b) Population Size=6

Pc=0.25,Pm=0.7
0.99999986
0
20
40
1.00000000
60
80
100
Generation
Figure 7.5, Generation average fitness and RMSE of the backcalculated moduli for the
small population in the 3-layer example using (a) Pc=0.25 and Pm=0.001, and (b) Pc=0.25
and Pm=0.7.
197
The selection of the genetic operators and the fitness function is a problem that
has been introduced as the exploration-exploitation trade-off problem (Epstein, 1967;
Holland, 1992). In genetic algorithms, the fitness function and the genetic operators
perform the exploration and exploitation, respectively. Genetic algorithms perform best
when the search space exploration initially is performed by weakly causal changes, while
the exploitation is later performed by strongly causal changes (Rosca and Ballard, 1995).
However, weak causality is more common in genetic algorithm problems.
7.3.2 Least Square Error
The least square error has been used as the fitness function in all of the moduli
backcalculation studies using genetic algorithms (Fwa et al., 1997 ; Kameyama et al.,
1997 ; Reddy et al., 2004 ; Tsai et al., 2004). The fitness functions used in all studies are
shown below:
Fwa et al. (1997):
Minimize RMSE =
Kameyama et al. (1997): Minimize RMSE k =
Reddy et al. (2004):
1
m
Minimize
Minimize
i =1
RMSE =
RMSE =
d i Di
Di
wi (d ik Dik ) 2
i =1
i =1
Tsai et al. (2004):
(d i Di )2
m
i =1
(d i Di )2
Eq. (7.4)
Eq. (7.5)
Eq. (7.6)
Eq. (7.7)
the backcalculated deflection at point i, Di is the measured deflection at point i, wi is the
198
weight factor at point i, and k is the chromosome number.

As can be seen, all of the above functions used the least square errors as a
measure of the quality of the backcalculated displacements and provided feedback to the
genetic algorithm to direct the search.
The principal idea behind the selection of the least square error is the
minimization of the associated errors with each chromosome. When the least square error
is minimized, the error between the measured deflectionts from the FWD test is minimal,
and hence it is commonly assumed that the difference between the actual moduli and the
backcalculated moduli is minimal as well. However, this assumption was shown to be not
true since the displacement basin measured in the FWD test can be produced using more
than one combination of layers moduli and hence the solution is not unique. In other
words, the same fitness value of the chromosome (backcalculated moduli) can be
obtained using different moduli. Therefore, it is important to understand the mechanism
of the fitness function to determine the effectiveness of the feedback process.
7.3.3 Mathematical Background
In the least square error method a straight line is used to fit between a set of
paired observations (Chapra and Canale, 1988) such as (xi,yi). If a straight line is assumed
to connect between the paired observations, the straight line can be written as:
y = a0 + a1 x + e
Eq. (7.8)
where a0 and a1 are the intercept of the straight line with the y axis and the slope of the
straight line ,respectively, while e is the residual between the model and the observations
which can be written as:
199
e = y a0 a1 x
Eq. (7.9)
Therefore, to fit a line through the paired observations, the error associated with
the fitting should be minimal to obtain the best fit. The best fit can be achieved by
minimizing the residual errors, for all data points (n), which can be written as:
n
i =1 i
e =
i =1
( yi a0 a1 x)
Eq. (7.10)
However, the above equation can be satisfied with any line (except a perfectly
vertical line) that passes a point in the middle of the line given by Eq. 7.8. Therefore, the
above criterion is not satisfactory since no unique best fit is available and hence should
not be used.
The above criterion can be modified to:
n
i =1
ei =
n
i =1
y i a0 a1 x
Eq. (7.11)
where the absolute values of the residuals are minimized to obtain the best fit straight
line. Again, this criterion will not yield a unique solution since two lines can have the
same absolute residual differences but with different slopes and still satisfy the above
criterion.
The problem of the criterion presented in Eq. 7.11 can be solved by minimizing
the sum of the squares of the residuals:
n
2
i =1 i
e =
n
i =1
( y i a0 a1 x) 2
Eq. (7.12)
which yields a unique line for a given set of observation points.

The individual residual values computed by Eq. 7.12 are the squares of the
distance between the data points and the straight line. In addition, the quality of the best
200
fit can be evaluated by measuring the standard error of the estimate which can be
computed by:
n
sy / x =
i =1
( y i a0 a1 x) 2
Eq. (7.13)
n2
which measures the error of the predicted value of y using the value of x by quantifying
the spread of the data around the regression line. Another measure that quantifies the
spread of the data around the mean ( y ) is the standard deviation ( S t ) given by:
( y i y )2
St =
n 1
The mean can be calculated by y =
i =i
yi
Eq. (7.14)
. However, it is still possible to obtain a
standard deviation ( S t ) that is close to 1 while the relation between the y and x is not
even linear (Chapra and Canale, 1988).
The above discussion shows that the least squares method has some drawbacks
that can prohibit the finding of a unique solution. Even if a unique solution is available by
the least squares method, the solution will result in the best match between the
calculated deflection basins from the backcalculated moduli and the measured deflection
basin which does not ensure the finding of the right moduli of the pavement. It was
shown that different elastic moduli can result in the same displacements residual and
hence the same least square errors.
201
7.3.4 Selection of The Fitness Function

It was shown that the fitness function has a significant effect on the genetic
algorithm performance. Therefore, the fitness function should be selected properly to
enhance the genetic algorithm performance by directing the search process to the
subdomain of the best solution (optimum solution).
The following functions were identified as possible fitness functions that can be
used to explore the search domain:
1
1 + RMSE
Eq. (7.15)
f =
1
100(1 RMSE ) + 1
Eq. (7.16)
Maximize
f =
1
1000(1 RMSE ) + 1
Eq. (7.17)
Case 4:
Maximize
f =
1
0.1(1 RMSE ) + 1
Eq. (7.18)
Case 5:
Maximize
f =
1
0.01(1 RMSE ) + 1
Eq. (7.19)
Case 6:
Maximize
f =
1
0.001(1 RMSE ) + 1
Eq. (7.20)
Case 7:
Minimize
f =
Case 8:
Minimize
f = r2 = 1
Case 1:
Maximize
f =
Case 2:
Maximize
Case 3:
RMSE
N
Eq. (7.21)
RMSE
m
i =1
202
(D
D)
Eq. (7.22)
Case 9:
f =
Minimize
i =1
Case 10:
f =
Minimize
i =1
Case 11:
f =
Minimize
i =1
Case 12: Minimize AREA =
1
D0 r1 +
2 D0
N 1
i =1
d i Di
Eq. (7.23)
(d max Di )2
Eq. (7.24)
(d i Dmax )2
Eq. (7.25)
Di (ri +1 ri ) + DN (rN rN 1 )
Eq. (7.26)
where N is the number of sensors (measured deflection points), d i is the backcalculated

deflection at point i, Di and D are the measured deflections at sensor i and the average
deflection of all sensors, respectively, d max and Dmax are the maximum backcalculated and
measured deflections of all sensors, respectively, and ri is the distance between the center
of the loading plate and sensor i. The RMSE in the above equations is given by Eq. (7.6).
Case 1 was proposed by Reddy et al. (2004) to minimize the RMSE error. Cases 2
through 6 have the same formula but with different coefficients (100,1000,0.1,0.01, and
0.001). Those cases (Cases 2 through 6) represent nonlinear fitness amplifications that
have been used to increase the sensitivity of the selection process to any variation in the
gradient of the fitness surface in cases where the surface is flat (Daridi et al., 2004).
Cases 7 and 8 use the correlation coefficient as the objective function. Cases 9 through 11
are different forms of the least square error.
Case 12 uses the AREA parameter as the objective function since it combines the
effect of several measured deflections. The AREA parameter was first proposed by
Hoffman
and
Thompson
(1981)
for
interpreting flexible pavement deflection

203
basins. The AREA parameter has a length dimension since it is normalized using one of
the measured deflections to remove the effect of the load.
Figure 7.6 shows the effect of the fitness function on the average backcalculated
moduli. As the figure shows, the fitness functions are not equal in their effect even
though the least square error or a modified form of it was used in all cases except Case
12. The REMSEE error of one fitness function (eg. Case 6) can be more than twice that
from another fitness function (eg. Case 12). This indicates that the exploration
mechanism in some cases performed better in guiding the genetic algorithm toward the
best average solution while other functions guided the genetic algorithm toward other
domains that have local optima.
In addition, the convergence of the fitness functions varied from one function to
another. In Case 12 the fitness function converged quickly to the optimal solution while
the other functions showed some stochastic behavior in the REMSEE error. On the other
hand, all cases showed an increase in the REMSEE error with the increase of the
generation while Case 12 showed steady REMSEE error values.
The function represented by Case 12 has shown the best performance among the
studied functions. Case 11 showed lower REMSEE error values at the beginning of the
optimization but then showed high REMSEE error values later on. Moreover, Case 12
showed strong causality indicating that a small change rate exists in the search space of
the moduli, which is desirable (Rosca and Ballard, 1995). Since the use of the GA does
not have any clue about the actual moduli, the fitness function represented by Case 12
is recommended over the other tested functions.
204
140
Case 1
Case 2
Case 3
Case 4
Case 5
Case 6
Case 7
Case 8
Case 9
Case 10
Case 11
Case 12
120
100
80
60
40
20
0
0
20
40
60
80
100
120
140
Generation
Figure 7.6, Generations RMSE of the backcalculated moduli for the 3-layer example.
7.4 The Effect of Increasing the Number of Layers

Increasing the number of layers in the pavement system will increase the number
of the backcalculated moduli and hence the number of unknowns. In addition, increasing
the number of unknowns increases the complexity of the search domain and can affect
the performance of the genetic parameters and operators. Therefore, it is vital to
understand the effect of increasing the number of unknown moduli on the performance of
the backcalculation process to better select more appropriate genetic parameters and
hence to design more effective backcalculation algorithms.
205
In order to study the effect of the number of layers, the 3-layer pavement system
shown in Figure 7.1 was modified to a 4-layer pavement system by adding an additional
subbase layer below the base layer and above the subgrade layer. The data of the new 4layer pavement system is shown in Table 7.6 while the used genetic parameters and
moduli ranges are shown in Tables 7.7 and 7.8, respectively. The effect of adding the
layer was studied using a large population size (population size of 100).
Table 7.6, Data of the 4-layer pavement example.
Layer
Thickness Modulus
Poissons
Description
Number
(m)
(MPa)
Ratio
Top Layer
1
0.175
2000
0.30
Base Layer
2
0.400
400
0.30
Subbase Layer
3
0.200
200
0.30
Subgrade Layer
4
70
0.30
Parameter
Value
Population size
100
100
0.75
0.001
Moduli Range
Layer
(MPa)
Top Layer
1200 - 2800
Base Layer
240 - 560
Subbase Layer
120 - 280
Subgrade Layer
40 - 100
206
Figure 7.7 shows that using the same genetic operators and parameters that were
used in the 3-layer pavement system to backcalculate the unknown pavement moduli in
the 4-layer pavement system can lead to a better performance. The figure indicates that
the complexity of the search space in the 4-layer system can be explored and exploited
more effectively even though the used genetic operators and parameters showed a
relatively poor performance when they were used to explore and exploit the 3-layer
system which is relatively less complex. In addition, no rule of thumb can be drawn from
increasing or reducing the number of the unknown moduli on the performance of the
genetic backcalculation process. This indicates that no unique set of genetic operators can
be used blindly with any pavement system due to the complexity of the search space, the
interaction between the genetic operators, and the effect of the genetic operators on the
search space.
207
1.00000000
(a)
0.99999995
0.99999990
0.99999985
3 Layers; Pc=0.75,Pm=0.001
4 Layers; Pc=0.75,Pm=0.001
0.99999980
0
20
40
60
80
100
Generation
60
(b)
3 Layers; Pc=0.75,Pm=0.001
4 Layers; Pc=0.75,Pm=0.001
50
40
30
20
10
0
0
20
40
60
80
100
Generation
the 3-layer and 4-layer pavement systems.
208
7.5 The Effect of the Number of Sensors

The effect of the number of sensors on the backcalculated moduli was studied
using the 3-layer pavement system and assuming two different objective functions. The
first objective function is based on the RMSE error (best-fit case) as given by Eq. (7.15)
while the second objective function is based on the AREA method (AREA case) as given
by Eq. (7.25).
The relations between the number of generation and the root mean square error of
the backcalculated moduli are shown in Figures 7.8 and 7.9 for the best-fit case and the
AREA case, respectively. Both methods indicated that increasing the number of sensors
will increase the accuracy of the backcalculated moduli. However, the effect of the
number of sensors on the accuracy of the backcalculated moduli is clearly evident in the
AREA case where the accuracy (lowest RMSE) was the highest when all of the sensors
(6 sensors) were included in the backcalculation process and was the lowest (high
RMSE) when only one sensor was included.
Figure 7.8 shows the highest accuracy when all of the sensors (6 sensors) were
included in the backcalculation process. However, there is no distinct trend between the
increase in accuracy and the number of sensors when the number of sensors is less than 6.
This indicates that the best-fit case employs a poor objective function that affects the
exploitation and exploration for the optimal solution.
On the other hand, Figure 7.9 shows the sensitivity of the backcalculation process
to the number of sensors. The figure indicates that increasing the number of sensors can
greatly affect the performance of the backcalculation process and the interaction between
209
the genetic operators and parameters to explore and exploit the search space. In other
words, the complexity of the search space is affected by the number of sensors which in
turn affects the accuracy of the backcalculated moduli. In Figure 7.9, doubling the
number of sensors (3 sensors to 6 sensors) reduced the error by 50%. However, the figure
indicates that there is a threshold for the number of sensors after which the reduction in
error can be negligible (5 sensors and 6 sensors).
It can be said that increasing the number of sensors increases the accuracy of the
backcalculated moduli. In addition, the selection of the objective function in the
backcalculation process affects the optimal number of sensors as well as the accuracy of
the backcalculated moduli.
100
1 Sensor
2 Sensors
3 Sensors
4 Sensors
5 Sensors
6 Sensors
80
60
40
20
0
0
50
100
150
200
250
300
Generation
Figure 7.8, The RMSE of the backcalculated moduli for the 3-layer pavement system
using different number of sensors (RMSE method).
210
80
3 Sensors
4 Sensors
5 Sensors
6 Sensors
60
40
20
0
0
20
40
60
80
100
120
140
Generation
Figure 7.9, The RMSE of the backcalculated moduli for the 3-layer pavement system
using different number of sensors (AREA method).
7.6 Conclusions
The backcalculation of the pavement moduli is sensitive to many factors other
than the genetic algorithms parameters and operators. It was found that the
backcalculation process is sensitive to the chromosome length, the number of layers, the
number of sensors, and the objective (fitness) function.
It is evident that increasing the chromosome length would increase the accuracy
of the backcalculated moduli since a larger search space would be explored. Increasing
the number of layers can either increase or decrease the complexity of the genetic
operators work to find the optimal solution. The AREA method showed the best
211
performance when used as the fitness function compared to many other objective
functions.
The effect of increasing the number of sensors depends on the used fitness
function. However, there is a threshold beyond which the effect of the number of sensors
will be negligible.
212
CHAPTER VIII
PARAMETERLESS GENETIC ALGORITHM FOR THE BACKCALCULATION OF
ELASTIC MODULI
8.1 Introduction
Genetic algorithms are very powerful algorithms for the optimization of complex
functions and applications. The mechanism of genetic algorithms is similar to that from
natural selection theory where a population first is randomly created (candidate
solutions), and then mating between the individuals (chromosomes) starts based on the
characteristics (fitness value) of the individuals. The mating process produces new
individuals that can survive the surrounding environment and conditions and should
ensure that the best characteristics that can handle the environment will transfer (will be
inherited) to the next generation (offspring).
The reproduction of chromosomes in genetic algorithms is controlled by the
genetic operators such as the crossover and the mutation. These operators work together
to exploit the search space by providing diversity among the individuals of the generation
to allow a wider spatial investigation of the subdomain (within the search space) that was
pointed to by the fitness function. However, the interaction between the genetic operators
is very complex since each operator has its own effect on the search space and the
213
schema of the newly generated chromosomes. For example, using high mutation
probabilities can disturb the schema which in turn will lead to a complete random
generation of the chromosomes which in turn will have an adverse effect on the genetic
algorithm ability to locate the best solution since less characteristics will be transferred
from one generation to the next.
The interaction between the genetic operators will control the population size
needed to solve the problem in hand. This can be understood by looking at the genetic
operators as tools for finding the best solution within a subdomain and hence for
convergence. These tools can be either highly effective or poorly designed where the
time needed to reach the peak or the valley is highly impractical. Therefore, designing the
best tools will be a challenge since each problem has its unique characteristics and
should be considered carefully.
The selection of the right genetic operators is not a trivial task. The population
size, the search domain, the physical behavior of the optimized problem, and the desired
time to convergence all work as constraints to the selection of the genetic operators to be
used in practical applications.
8.2 Genetic Algorithms Difficulties
The complexity of the problem, the genetic operators, and the population size are
the most common problems when using the genetic algorithms since the interaction
between these factors is complex. The content and the context of the genetic algorithm
problem and operators are additional unseen factors that affect the problem. The content
of a problem is represented by the physical behavior of the problem while the context is
214
represented by the way that the genetic operators interact with each other (Gustafson,
2004). Therefore, one cannot generalize the selected genetic operators that work with one
problem to other problems since each problem has its unique content and context. The
content and context of a problem can highly affect the relation between the fitness values
and the chromosomes where it can become either sensitive or not sensitive.
The population size of any problem influences both the context and the content of
the problem. The genetic operators (context) work on the chromosomes of the population
and therefore, increasing the population size will tend to create new individuals that
behave differently than those generated in a small population. This can be attributed to
the higher diversity among the individuals in large populations compared to small ones.
In turn, the fitness values associated with large populations comprise a different search
landscape than the ones encountered in small populations. The ability of the genetic
algorithms to reach optimal solutions is limited. This limitation arises from the behavior
of the genetic operators, the fitness function, and the population size where they can lead
to a premature convergence.
The increase of the individuals with the same or close fitness values will reduce
the ability of the selection procedure. In the selection procedure individuals are compared
based on their fitness, therefore, increasing the individuals with the same fitness will
make the selection process more random since no preference is given to any of the
selected chromosomes. Hence, increasing the fitness with the reduction of the diversity in
the population reduces the effectiveness of the genetic algorithms since the algorithm
become more random rather than guided.
215
Deception is another factor that affect the performance of genetic algorithms

(Goldberg, 1987; Goldberg, 2002). Deception occurs when the fitness of two
chromosomes (solutions) is the same but their values are different, therefore, applying
genetic operators such as crossover produces poor fitness values.
8.3 Diversity in Genetic Algorithms
A population of chromosomes is said to be diverse if the chromosomes consist of
genes (solutions) from as many subdomains as possible. Diversity is believed to be a vital
requirement to avoid premature convergence in genetic algorithms (Ryan, 1994; McKay,
2000). Increasing the diversity can undermine the search process and hence no optimal
solution can be found. On the other hand, decreasing the diversity can either slow down
the convergence to optimal solutions or result in premature convergence.
However, preserving diversity is not trivial since each run can start with a diverse
population over the search domain, but during the run, the population tends to
propagate toward a subdomain and hence will tend to loose diversity. The initial parent
pool is diverse since all individuals are generated randomly. The algorithm starts to
converge when the fitness of the individuals cannot be improved indicating the arrival at
the optimal solution for the given conditions.
As the genetic algorithm loses diversity the search tends to work more on
localized subdomains with close neighborhoods only and loses the ability to work on
larger subdomains. Therefore, the search turns from a guided random genetic search to a
stochastic hill-climbing search. Hill-climbing is a local search method that moves in a
direction related to the local gradient (Goldberg, 1989). In genetic algorithms, hill216
climbing can be obtained if the crossover probability is equal to zero while the mutation
probability is very small. This indicates that when diversity is very minimal, the only
source of diversity would be mutation since crossover will not be able to provide diverse
individuals.
In addition, diversity is affected by the selection pressure (Whitley, 1989).
Increasing the selection pressure increases the convergence ability of the search while
decreasing the selection pressure reduces the ability of the search to converge. When the
selection pressure increases, the search focuses more on the individuals with the highest
fitness and therefore the diversity decreases. In addition, tuning the genetic parameters
for a certain problem is in fact an auxiliary way of tuning the selection pressure. On the
other hand, increasing the population increases the schemata involved in the search and
hence reduces the selection pressure and increases the diversity. The selection procedure
(tournament selection, roulette wheel,etc.) affects the selection pressure. For example,
in the roulette wheel selection method individuals with high fitness can dominate quickly
leading to a premature convergence.
Controlling the diversity of the population can be achieved by controlling the
genetic operators. Controlling diversity showed problematic issues such as knowing the
level of diversity needed to solve the problem before the start of the run. The interaction
between the exploration and the exploitation processes in genetic algorithms impose a
random effect on the generated individuals that make the prediction of the needed level of
diversity difficult.
217
8.4 Interaction Between Genetic Operators

Crossover and mutation are the main genetic operators that are commonly used in
genetic algorithms. These operators are applied after calculating the fitness of each
chromosome in the population. Then, the operators work to increase the number of the
fittest individuals in the population. The schema theorem cannot be used to explain the
mechanisms of the search strategy of genetic algorithm (Mhlenbein, 1991).
The crossover is one of the important genetic operators. It works on chromosomes
by joining parts of chromosomes together to form new chromosomes. The crossover
probability is high in many of the applications in genetic algorithms. On the other hand,
mutation works on chromosomes by flipping the binary strings at random locations to its
opposite binary value.
The work of De Jong (1975) was the first published attempt to tune the genetic
parameters to arrive at a set that can be used for other problems. In his work different
functions were tested to cover a wide range of possible complexity in function
characteristics. The genetic algorithm components and operators included the single-point
crossover, mutation, and the roulette wheel selection method. Many trial runs for
different sets of genetic operators/parameters were carried out to better understand the
behavior of the genetic algorithms including the crossover probability, the mutation
probability, and the population size. The main findings in De Jongs
(1975) work
included:
1- Mutation is needed to replace the lost chromosomes during the generation of new
chromosomes. The mutation probability should be selected carefully and should
218
be low since high mutation rates can result in a random search strategy rather than
a genetic algorithm search.
2- The following parameters can be used to optimize a wide range of problems:
crossover probability=0.6, mutation probability=0.001, and population size=50 to
100.
Although De Jongs work break the ground for future investigations and research, the
recommended genetic parameters cannot be used for all problems as done by many
researchers. For example, the use of roulette wheel will impose a different selection
pressure than the selection pressure by the tournament selection method. In addition, the
search landscape is problem dependent.
Schaffer et al. (1989) expanded the work of De Jong (1975) by adding more functions
to those studied by De Hong. They investigated the performance of the functions using
different set of genetic operators including crossover probabilities ranging between 0.05
and 0.95, mutation probabilities ranging between 0.001 and 0.1. In addition, Schaffer et
al. (1989) used six population sizes including 10,20,30,50,100, and 200. The main
findings in their work included:
1-High mutation probabilities are recommended for small populations while small
mutation probabilities are recommended for large populations.
2-Mutation can be used alone to solve genetic algorithm problems.
3-The following parameters can be used to optimize similar functions: crossover
probability=0.75 to 0.95, mutation probability=0.005 to 0.01, and population size=20 to
30.
219
Schaffer et al. (1989) found, empirically, that an optimal parameter setting for the
studied functions can be achieved using:
ln N + 0.93 ln m + 0.45 ln n = 0.56
Eq. (8.1)
where N is the size of the population, m is the mutation probability, and n is the length of
the chromosome. This relation was proved to be incorrect (Mhlenbein, 1992). However,
the above study was still limited to a certain number of functions and cannot be extended
to all problems. In addition, Schaffer et al. (1989) recognized that the used functions are
insensitive to crossover and sensitive to mutation and hence more complex functions
should be studied.
The effect of the mutation probability on the performance of the genetic
algorithms was studied by Mhlenbein (1992) and Bck (1993). Both studies indicated
that the optimal mutation probability for unimodal functions is 1/l where l is the length of
the chromosome. In addition, Bck (1993) indicated that for multimodal functions,
mutation probabilities equal to 1/l were appropriate.
Ochoa et al. (1999) showed that without recombination and using adaptive
mutation, the mutation probability decreases with time and hence high mutation rates
should be used at the beginning of the run. In addition, they showed that with crossover,
the optimal solution can be achieved with constant low mutation probabilities.
8.5 Automation of GA
The main problem with genetic algorithms is the selection of the required genetic
parameters (genetic operators, population size, and number of generations). The selection
of genetic parameters depends on the user
experience, published literature, and the

220
trial and error technique.
These make the genetic algorithms less appealing to
practitioners since no general guidelines are available to guide the users. Therefore,
tuning the genetic algorithms parameters is the main challenge that yet has to be solved to
guarantee the acceptance of the method by a wide spectrum of practitioners.
The value of the genetic operators can be changed using different methods that
can be grouped into three categories (Eiben et al., 1999):
1-Deterministic Parameter Control Methods: In these methods the operators are changed
based on deterministic rules without taking into consideration the search progress. The
controlling parameter for the modification of the genetic operator is the generation
number.
2-Adaptive Parameter Control Methods: In these methods, the direction and/or magnitude
of the change of the genetic operator is determined based on the search progress.
3-Self-adaptive Parameter Control Methods: In these methods, the genetic operators are
encoded into the chromosomes and hence undergo changes as the search progress.
Some of the relationships that have been suggested in literature using deterministic
parameter control are presented below.
Fogarty and Terence (1989) proposed the following relation to change the
mutation probability (Pm) based on the generation number (t):
Pm =
1
0.11375
+
240
2t
Eq. (8.2)
Hesser and Manner (1991) proposed the following relation to change the mutation
probability (Pm) based on the generation number (t):
221
Pm =
e 2
Eq. (8.3)
where is the population size, t is the generation number, L is the chromosome length in
bits, and ( , ,and ) are constants.
Bck and Schutz (1996) proposed the following relation for mutation probability
(Pm) based on the generation number (t):
l2
Pm = 2 +
t
T 1
Eq. (8.4)
where T is the maximum allowed number of generation. This relation constrains the
change of the mutation between 0.5 at t=0 to 1/L at t=T.
Figure 8.1 shows the relation between the generation number and the mutation
probability as suggested by Fogarty and Terence (1989) and Bck and Schutz (1996). As
can be seen, Fogarty and Terence (1989) method resulted in a mutation probabilities of
0.061042 at generation t=1 and a mutation probability of approximately 0.004167 for all
generations at and beyond t=18. Therefore Fogarty and Terence (1989) suggested relation
gives limited number of small mutation probabilities throughout the generations and
hence will have effect at the beginning of the GA run while it will be constant after
generation 18. On the other hand, Bck and Schutz (1996) suggested relation showed
higher mutation probabilities than those by Fogarty and Terence (1989). In addition,
Bck and Schutz (1996) mutation probabilities were higher for short chromosomes (l=8)
compared to long chromosomes (l=16 or 32). Therefore, it is expected to see higher
diversity in the population when the chromosome length is relatively short. Both
222
mutation relations showed that the mutation probability relation should be selected
carefully since each problem may respond differently to the rate of change in the
mutation probability.
0.6
Fogarty and Terence (1989)
Back and Schutz (1996); Chrom Length=8
Mutation Probability - Pm
0.5
0.4
0.3
0.2
0.1
0.0
0
20
40
60
80
100
Generation - t
Figure 8.1, Mutation probability as a function of generation number.
The resulted genetic parameters from the manual tuning of parameters using the
trial and error technique showed some drawbacks (Eiben et al., 1999). The genetic
parameters are independent and hence it is impossible to try all of the different parameter
combinations systematically. In addition, the trial and error process is time consuming
since a very large number of combinations should be used. On the other hand, the
selected parameter combinations will aid in finding the optimal parameter set among the
223
searched possibilities and the optimal parameter set may not be found.
The genetic parameters can be constant during certain stages of the search while
they can be adaptive during other stages (Syswerda, 1991; Bck, 1992; Eiben et al.,
1999). For example, large mutation probabilities are needed during the initial stages of
the search to aid in the exploration of the search space while small mutation probabilities
are needed during the final stages to aid in finding the optimal solution.
Adaptive parameter control was first introduced by Rechenberg (1973). In his
work he used a variable mutation probability that was dependent on the rate of successful
mutations in a population. Julstrom (1995) changed the value of the crossover and
mutation probabilities based on their contribution to the fitness of the new generation.
Srinivas and Patniak (1994) proposed assigning mutation and crossover
probabilities for each chromosome based on its fitness and the fitness of the population.
Another way of self-adapting was first introduced by Schwefel (1997). He proposed a
new method where the encoded chromosome carried also its mutation probability which
was subjected to crossover and mutation with the chromosome. Bck et al. (2000)
extended Schwefel (1997) work by encoding the mutation and crossover probabilities
with the chromosome. The mutation and crossover probabilities were subjected to
evolution during the progress of the run.
8.5.1 Adaptive Population Sizing
The population size is a very important parameter in genetic algorithms. If the
population size is underestimated, the algorithm will never converge or will converge to a
local optimum since the representation of the search space in the genetic pool is less than
224
what is required, and therefore, regardless of the genetic operators or the selection
method, the newly generated schemata will not provide the required diversity in the
population. On the other hand, if the population size is overestimated, the algorithm will
spend significant time searching for the optimal solution. It is typical to expect larger
population sizes in problems of high complexity such as multimodal problems. In other
words, increasing the possibility of convergence due to the large number of local optima
increases the need for larger population size. Therefore, the selection of the appropriate
size is as important as selecting the genetic operators or the selection strategy.
However, selecting the right population size for practical problems is very
difficult since the search landscape is not known in advance. A common way of handling
this problem is using different population sizes to figure out the effect of the population
size on the performance of the algorithm. This trial and error method adds more
complexity to the problem of defining the appropriate parameters and operators for the
genetic algorithm since different parameter/operator sets can produce different effects.
Therefore, changing the size of the population can add an additional variable that need to
be optimized.
Different adaptive population sizing methods have been proposed in literature.
The following discusses some of the proposed methods.
8.5.1.1 Population Sizing using Schema Variances
This method was first proposed by Smith (1993) and Smith and Smuda (1995)
based on the work of Goldberg et al. (1993). In this method the size of the population
was adjusted as the search for the optimal solution progressed. They proposed an
225
algorithm that sizes the population based on the expected selection loss and a specified
target accuracy should be provided by the user. The expected selection loss between two
competing schemata is the probability that the lower fit schema gets selected weighted by
the fitness difference between them.
8.5.1.2 Population Sizing using Life Time
Arabas et al. (1994) created a new variable called the remaining life time (RLT),
which was assigned to each chromosome. The RLT of each chromosome was updated at
the beginning of each generation using a bi-linear formula depending on whether the
fitness of the chromosome was less than the average fitness of the population or not.
Chromosomes with RLT values close to zero are eliminated from the generation. A
reproduction factor was used in which part of the population was allowed to reproduce.
Different reproduction factors showed different effects on the performance of the genetic
algorithm. The method requires an initial population size to be provided by the user.
8.5.1.3 Population Sizing using Competing Subpopulation
Schlierkamp and Muhlenbein (1996) proposed running more than one population
at the same time. After several generations, the run with the best fitness is increased
while the other generations are reduced. Hinterding et al. (1996) proposed running three
populations simultaneously with initial size ratios of 1:2:4. The fitness values of the three
populations are compared after several runs and the size of the best run is doubled, the
size of the worst run is halved, and the size of the run in between is maintained.
226
8.5.1.4 Population Sizing Using Parameterless GA (PGA)

This method was proposed by Harik and Lobo (1999) and Lobo (2000) assuming
that the solution quality increases as the population size increases. They further assumed
that for two identical genetic algorithms with two different population sizes (P1, P2)
where P1< P2, if P1 spends more evaluations (generation multiplied by population size)
than P2 and the average fitness of P1 is less than P2, then there is a strong evidence that
that the analysis of P1 is carried out using an undersized population. The algorithm steps
are as the following:
1-Run two population sizes simultaneously with Pi=(1/2) Pi+1, where i is the population
number.
2- Run population Pi four times population Pi+1 (number of generations).
3-Inspect the average fitness of the population after several evaluations. If the average
fitness of Pi< Pi+1 then eliminate Pi and double Pi+1.
4-Repeat until the desired fitness is achieved.
According to the schema theorem, the survival probability of a schema can be
increased either by increasing the selection rate or lowering the crossover probability
(Harik and Lobo, 1999; Lobo, 2000). Both the crossover and the selection rate control the
growth of the building blocks from generation to another. In addition, the selection rate
controls the amount of bias towards better individuals while the crossover probability
controls the amount of recombination or mixing. However, if the selection rate is high
the algorithm will pay more attention to the individuals with the highest fitness values
while it will pay less attention if the selection rate is too low. An appropriate selection
227
rate should ensure that the growth ratio of building blocks is greater than 1.
The net growth ratio on schema H at generation t is (Harik and Lobo, 1999; Lobo,
2000):
( H , t )[1 ( H , t )]
Eq. (8.5)
where ( H , t ) is the effect of the selection operator, ( H , t ) is the disruption factor on the
schema.
The above equation is a simplified form of the schema theorem, which can be written,
using the selection rate s and the probability of crossover Pc, as:
s (1 Pc )
Eq. (8.6)
Harik and Lobo (1999) and Lobo (2000) suggested using s=4 and Pc=0.5 in the PGA
which give a net growth of 2. Mutation probability was ignored for simplicity in the
original work by Harik and Lobo (1999) and Lobo (2000). However, Lobo (2000)
recommended using a mutation probability equal to the inverse of the chromosome length
(1/l).
In addition, the PGA takes advantage of the fact that increasing the population
size makes the sampling of the building blocks better since the error in decision-making
reduces. Population sizing equations proposed in literature (Goldberg et al., 1992; Harik
et al., 1997) are difficult to apply in practice. Population sizing equations assume proper
mixing of the building blocks. In addition, applying the population sizing equations
requires knowing the maximum deception in the problem and the selective advantage of a
building block over its most tough competitor (Lobo, 2000).
228
The selection of an adaptive function is very difficult since many factors are
involved during the search process and the interaction between these factors is highly
complex. In addition, the adaptive parameters are controlled mainly by the generation
number while the actual progress in solving the problem and the performance of the
genetic algorithms are ignored (Eiben et al., 1999).
8.6 The Dynamic Parameterless Genetic Algorithm
The optimization of pavement layer moduli by backcalculating the moduli using
the FWD data is a complex task since the resulting function is multimodal. Multimodal
functions are difficult to optimize due to the high probability of converging to a local
optima during the search process. Due to the complexity of the problem, tuning the
genetic parameters and operators can be a complex task, as well, which may limit the use
of the method by practitioners. The genetic algorithm method is robust and very reliable
in finding the global optima if the optimal genetic operators and parameters are used
making it more attractive than the available backcalculation procedures. Therefore, the
Parameterless Genetic Algorithm method is appealing since none of the parameters needs
human tuning.
A new PGA method is proposed. The new PGA method is intended to increase
the efficiency of the PGA and to reduce the time associated with the method. The new
PGA method is called the Dynamic Parameterless Genetic Algorithm (DPGA). The
DPGA is based on Harik and Lobos (1999) and Lobos (2000) parameterless GA, and
includes the following steps:
1-Run population Pi for a number of generations until the average fitness of two
229
consecutive generations is less than a specified tolerance value or negligible.

2-Run another population (Pi+1) where Pi+1, is twice Pi.
3-Insert the fittest individuals from run Pi to the parents pool of population Pi+1. The
remaining individuals of the parents pool of population Pi+1 are generated randomly
4-Inspect the average fitness of the population after several evaluations. If the average
fitness of two consecutive generations is less than a specified tolerance value or
negligible, stop the run. Otherwise, continue the run.
5-Repeat steps 2 through 3 until the change in the average fitness between two
populations (Pi and Pi+1) is less than a tolerance value or negligible.
The DPGA is intended to increase the population size of the same run
dynamically. By doing so, the diversity of the population is increased dynamically
without destructing the schema of the individuals in the population. In the DPGA
algorithm, the mutation rate is implicitly employed during the backcalculation process by
transferring the best individuals from one run to the other (from Pi to Pi+1) while adding
randomly the same number of individuals to the next run (to Pi+1). Therefore, the
mutation rate is random but at the same time is the highest at the beginning of the run (as
desired) and decreases as the run progress. In addition, the DPGA implements the niching
concept implicitly by giving focus to other than one region within the search space when
the highest fitness is focused at one region.
The effect of the PGA and the DPGA algorithms on the performance of the
backcalculation process was evaluated using the 23-layer pavement system. The example
details were discussed in a previous chapter. The best solution is defined as the solution
230
with the highest fitness and the lowest average moduli error. The used fitness function is:
RMSE =
m
i =1
Maximize f =
(d i Di )2
1
1 + RMSE
Eq. (8.7)
Eq. (8.8)
the backcalculated deflection at point i, Di is the measured deflection at point i, and f is
the fitness of the solution based on the root mean square error as estimated by Eq. (8.7).
The goodness of the selected parameters using the fitness function in Eq. (8.8)
was further assessed using the following equation:
RMSE E =
3
i =1
(Ec i E m i )2
Eq. (8.9)
where RMSE E is the root mean square error of the backcalculated elasticity moduli, and
Eci and Emi are the backcalculated and the sensitivity pavement system moduli (actual),

Figures 8.2 and 8.3 show the relation between the average fitness of the
generation and the RMSE of the moduli with the number of generation for the PGA and
the PDGA algorithms, respectively. In the PDGA runs, the case with a population of 64
was used as the starting population and was similar to that in the PGA case. The figures
show that the PDGA is better than the PGA since higher fitness values and lower moduli
RMSE errors were obtained. In addition, Figure 8.3 shows that the PDGA algorithm is
better in exploring and exploiting the search space than the PGA algorithm (Figure 8.2).
The PDGA algorithm, in this example,
showed RMSE errors approximately 40%
231
lower than those obtained using the PGA since the exploitation and the exploration with
mutation and niching were all employed correctly to the search space.
It is evident that increasing the population size suddenly increases the ability of
the genetic algorithms to explore and exploit other search domains while either focusing
more or ignoring the current domain. In the PGA, ignoring the results of one population
size and moving to the next population size can duplicate the effort by the genetic
algorithms in searching the same subdomains which can be time consuming.
232
(a)
Pc=0.50,Pop=64
Pc=0.50,Pop=128
Pc=0.50,Pop=256
Pc=0.50,Pop=512
Pc=0.50,Pop=1014
Pc=0.50,Pop=2028
0.999998840
0.999998830
0.999998820
0.999998810
0.999998800
0.999998790
0.999998780
0
10
Generation
60
(b)
Pc=0.50,Pop=64
Pc=0.50,Pop=128
Pc=0.50,Pop=256
Pc=0.50,Pop=512
Pc=0.50,Pop=1014
Pc=0.50,Pop=2028
50
40
30
20
10
0
0
10
Generation
the 23-layer example using the PGA method.
233
(a)
Pc=0.50,Pop=128
Pc=0.50,Pop=256
Pc=0.50,Pop=512
Pc=0.50,Pop=1014
Pc=0.50,Pop=2028
0.999999330
0.999999320
0.999999310
0.999999300
0.999999290
0.999999280
0.999999270
0
10
Generation
60
(b)
Pc=0.50,Pop=64
Pc=0.50,Pop=128
Pc=0.50,Pop=256
Pc=0.50,Pop=512
Pc=0.50,Pop=1014
Pc=0.50,Pop=2028
50
40
30
20
10
0
0
10
Generation
the 23-layer example using the DPGA method.
234
8.7 Conclusions
The optimization of the genetic parameters is challenging and can affect the
results of the backcalculation process. The results of one optimization problem cannot be
used with other problems due to the uniqueness of each problem and the complexity of
the search space. On the other hand, the available adaptive methods are developed based
on the success of solving certain problems ranging between simple to complex. However,
multimodal problems can be challenging and none of the available adaptive method can
deal with the complexity of such search spaces.
The automation of the genetic parameters can reduce the effort of the user and can
increase the time efficiency in finding accurate results. The automation of the genetic
parameters can be carried out using population sizing utilizing the PGA algorithm or the
new PDGA algorithm. The PDGA algorithm showed better performance dealing with
complex multimodal search spaces such as those in the pavement backcalculation
problem. At the beginning of each PDGA set of generations, half of the population is
inserted from the previous run while the second half is created randomly. The mutation
probability at the beginning of each set of generations is very high and reduces implicitly
as the backcalculation progresses. In addition, niching and is implicitly satisfied by
giving different than one subdomain some focus through the random distribution of half
of the population. Transferring the fittest individuals from one generation to the other
guarantee that the achieved results from the previous run do not get lost.
The backcalculation of the moduli of a 23-layer pavement system showed
promising future of the new PDGA algorithm in saving time and in achieving accurate
235
results with less effort.
236
CHAPTER IX
SUMMARIES AND CONCLUSIONS
9.1 Summaries
The backcalculation of pavement moduli is very challenging since the multimodal
nature of the search space reduces the ability of the classical backcalculation techniques
in finding the global optima. In this study, the backcalculation of the pavement moduli
using genetic algorithms was investigated thoroughly. The effect of the genetic operators
and parameters on the performance of the backcalculation process was studied as well. In
addition, the complexity of the search space was further studied using different
backcalculation cases.
A new backcalculation program (BackGenetic3D) was developed based on the
genetic algorithms theory that is capable of backcalculating the moduli of any pavement
system with any number of pavement layers, loading configuration, and loading shape.
The BackGenetic3D program employed new genetic operators and parameters such as
the jump mutation, creep mutation, crossover, niching, tournament selection, and elitism.
In addition, a superior random number generator (Mersenne Twister random Number
Generator) was employed to ensure randomness in the generated population and
numbers. The program was verified by analyzing a 3-layer pavement system using the
237
BackGenetic3D program and other programs that use classical backcalculation

techniques.
The BackGenetic3D program was used to further backcalculate the
pavement moduli of different cases and to study the many factors that affect the
backcalculation process.
A new method was proposed to reduce the effort needed to optimize genetic
parameters. In addition, the effects of the chromosome length, the fitness function, the
number of layers, the number of sensors, and the range of the search space on the
performance of the backcalculation process were investigated. The results of 3-, 4-, and
23-layer pavement systems were used to draw the conclusions and findings.
The automation of the genetic algorithms was investigated by reviewing the
available adaptive methods in literature. A new population-sizing algorithm (PDGA) was
developed that eliminates the need for both the mutation probability and the crossover
probability. The size of the population is determined by doubling the population size in
steps while keeping the results of the previous population. The new algorithm was tested
using a 23-layer pavement system.
9.2 Conclusions
Based on the research work performed in this study, the following conclusions
can be drawn:
1) The backcalculation of the pavement moduli using classical backcalculation
techniques can lead to misleading results due to the premature convergence, the
local optima, and the need for seed moduli.
2) The use of the deflection RMSE error with classical backcalculation techniques
238
can be misleading since different moduli sets can be obtained for the same RMSE
error.
3) The least square error should not be used as the objective function when
backcalculating the pavement moduli and the AREA method should be used
instead.
4) Increasing the number of sensors can increase the accuracy of the backcalculated
moduli.
5) Each unknown moduli should be presented by at least 15 bit strings.
6) The complexity of the search space depends on the number of sensors, the number
of layers, ranges of moduli, fitness function, and the population size.
7) Small populations should not be used to backcalculate the pavement moduli and
large populations should be used instead.
8) The sensitivity analysis can be used to reduce the number of trials to optimize the
genetic parameters.
9) Niching and elitism should be used to enhance the performance of the genetic
algorithms in multimodal functions.
10) The automation of the genetic algorithms can be done using the newly developed
Dynamic Parameterless Genetic Algorithm (DPGA).
11) The BackGenetic3D program is an accurate and robust backcalculation program.
The new program is capable of backcalculating the moduli of any pavement
system with any arbitrary number of layers, loading configuration, and load
distribution assuming elastic, linear, and homogeneous layers.
239
9.3 Future Studies

Based on the current work, some attractive problems could be further studied in
the future. The backcalculation of the pavement moduli using field measurements from
instruments installed inside the pavement should be considered. Many of the pavement
sections in Ohio and in the US are instrumented with strain gages and hence the variation
of the deflection within the pavement section and spatially can be obtained.
On the other hand, the backcalculation of the pavement using the actual load
distribution from tires on the pavement sections should be considered. The
BackGenetic3D program can be used to backcalculate the pavement moduli using the
actual load configuration and distribution with data collected either along the surface of
the pavement or within the pavement section.
In addition, the current program assumes elastic, linear, and homogeneous
pavement layers. The nonlinearity of the stresses in the pavement layers should be
considered in the backcalculation procedure especially under high surface loads. A
comparison between backcalculated moduli assuming linear and nonlinear stress
distributions should be carried out.
The current practice uses the surface deflections as the main inputs in the
backcalculation process. This approach showed limitations during the backcalculation
process where unique solutions are hard to find especially when the root mean square
error is used as the objective function. Future research should focus more on developing
better objective functions or on identifying other field measurements that can enhance the
backcalculation process and the accuracy of the backcalculated moduli.
240
The location and number of sensors used in the backcalculation process should be
further investigated. The current practice adopt the use of 6 to 9 sensors to measure the
surface deflections placed at distances that are multipliers of loading plate radius. The
effects of the distances between the sensors and the number of sensors should be further
studied taking into account the load value and the pavement thickness.
The developed genetic algorithms in this study can be used with other advanced
numerical simulation tools such as the finite element to backcalculate the pavement
moduli.
241
REFERENCES
AASHTO Guide for Design of Pavement Structures (1993). American Association of

State Highway and Transportation Officials.
Acum, W.A, Fox, L. (1951). Computation of Load Stresses in a Three-Layer Elastic
System, Geotechnique, Vol.2, pp. 293-300.
Akram, T., Scullion, T., and Smith, R.E. (1994). Comparing Laboratory and
Backcalculated Layer Moduli on Instrumented Pavement Sections, Special Technical
Publication (STP) 1198, ASTM, Philadelphia, PA.
Ancey, C. (2005). Notebook Introduction to Fluid Rheology, cole Polytechnique
Fdrale de Lausanne.
Anderson, M. (1988). Backcalculation of Composite Pavement Payer Moduli, PhD
dissertation, The University of Kentucky, Kentucky.
Anderson, M. (1989). A Database Method for Back Calculation of Composite Pavement
Layer Moduli, Nondestructive Testing of Pavements and Back Calculation of Moduli,
ASTM STP 1026, A.J. Busch III and G.Y Baladi, eds., American Society of Testing and
Materials, Philadelphia.
Arabas, J., Michalewicz, Z., and Mulawka, J. (1994). GAVaPS- A Genetic Algorithm
with Varying Population Size, Proceeding of the 1st IEEE Conference on Evolutionary
Computation, IEEE Press, Vol. 1, pp. 73-78.
Asphalt Institute (1991). Thickness Design: Asphalt Pavements for Highways and
Streets, Manual Series No.1 (MS-1).
Bck, T. (1992). The Interaction of Mutation Rate, Selection, and Self-adaptation within
a Genetic Algorithm. In Manner, R. and Manderick, B., eds., Parallel Problem Solving
from Nature, Amsterdam, North-Holland, pp. 85-94.
242
Bck, T. (1993). Optimal Mutation Rates in Genetic Search, in Proceedings of the Fifth
International Conference on Genetic Algorithms, Stephanie Forrest, ed., Morgan
Kaufmann, pp. 2-8.
Bck, T., and Schutz M. (1996). Intelligent Mutation Rate Control in Canonical Genetic
Algorithm, Proceeding of the International Symposium on Methodologies for Intelligent
Systems, pp. 158-167.
Bck, T., Eiben, A.E., and Van der Vaart, N.A. (2000). An Empirical Study on GAs
without Parameters, In Schenauer, M., Deb, K., Rudolph, G., Yao, X., Lutton,
E.,Merelo, J. J., and Schwefel, H-P. eds., Parallel Problem Solving from Nature PPSN V,
Lecture Notes in Computer Science, Vol. 1917, pp. 315-324.
Bendana, L.J., Yang, W.S., and Lu, J. (1994). Interpreting Data from the Falling Weight
Deflectometer, Engineering Research and Development Bureau, New York State
Department of Transportation, Research Report 160.
Bentsen, R.A., Bush, A.J., and Harrison, J.A. (1989). Evaluation of Nondestructive Test
Equipment for Airfield Pavements. Phase 1. Calibration Test Results and Field Data
Collection, US Army Engineer Waterways Experiment Station, Washington, DC.
Bodare, A. and Orrje, O. (1985). Impulse Load on a Circular Surface in an Infinite Elastic
Body-Closed Solution According to the Theory of Elasticity, Report No. 19, Department
of Soil and Rock Mechanics, Royal Institute of Technology, Stockholm, Sweden.
Bonnaure, F., Gest, G., Gravois, A. and Uge, P. (1977). A New Method of Predicting The
Stiffness of Asphalt Paving Mixtures, Proceedings of the Association of Asphalt Paving
Technologists, No. 46, pp. 64100.
Bonnaure, F., Gravois, A., and Udron, J. (1980). A New Method of Pedicting Te Ftigue
Lfe of Btuminous Mxes. Journal of the Association of Asphalt Paving Technologist, 49,
pp. 499-529.
Boussinesq, J. (1876). Essai theorique sur lequilibre delasticite des massifs
pulverulents compare a celui de massifs solides et sur la poussee des terres sans
cohesion. Memoires des Savants Etrangers, Academie de Belgique, 40, Brussels.
Boussinesq, J. (1885). Application des potentials a letude de lequilibre et du movement
des solides elastique, Gauthier-Villard, Paris.
Burmister, D.M. (1943). The Theory of Stresses and Displacements in Layered Systems
and Applications to the Design of Airport Runways, Proceedings, Highway Research
Board, Vol.23, pp. 126-144.
243
Burmister, D.M. (1945).The General Theory of Stresses and Displacements in Layered

Soil Systems, Journal of Applied Physics, Vol. 16, pps.89-94, 126-127, 296-302.
Bush, A.J. (1980). Nondestructive Testing for Light Aircraft Pavements, Phase II,
Development of the Nondestructive Evaluation Methodology, Report No. FAA-RD-80-9II, US Army Engineer Waterways Experiment Station, Washington, DC.
Bush, A. J. (1985). Computer Program BISDEF, United States Army, Corps of
Engineers, Waterways Experiment Station.
Bush, A.J., and Alexander, D.R. (1985). Pavement Evaluation Using Deflection Basin
Measurements and Layered Theory, Transportation Research Record 1022,
Transportation Research Board, Washington, DC, pp. 16-29.
Chapra, S, and Canale, R. (1988). Numerical Methods for Engineers. McGraw Hill Book
Co.
Cho, Y.S., and Lin, F.B. (2001). Spectral Analysis of Surface Wave Response of
Multilayer Thin Cement Mortar Slab Structures with Finite Thickness. NDT and E
International, 34, pp. 115122.
Chou, Y. J., and Lytton, R.L. (1991). Accuracy and Consistency of Backcalculated
Pavement Layer Moduli, Transportation Research Record 1022, TRB, National Research
Council, Washington, D.C., pp. 1-7.
COST 333 (1999). Development of New Bituminous Pavement Design Method, Final
Report of the Action, European Commission, Brussels, Luxembourg.
Cramer, N.L. (1985). A Representation for The Adaptive Generation of Simple
Sequential Programs, Proceedings of an International Conference on Genetic Algorithms
and the Applications, Grefenstette, John J., ed., Carnegie Mellon University.
Daleiden, J.F., Killingsworth, B.M., Simpson, A.L., and Zamora, R.A. (1994). Analysis
of procedures for establishing in situ subgrade moduli, Transportation Research Record
1462, TRB, National Research Council, Washington, D.C., pp. 102-107.
Danish Road Institute (2001). Verification of Flexible Pavement Response From a Field
Test Report, Report 121.
Daridi, F., Kharma, N., and Salik, J. (2004). Parameterless Genetic Algorithms: Review
and Innovation, IEEE Canadian Review, No. 47, pp. 19-23.
Darwin, C. R. (1859). On The Origin of Species by Means of Natural Selection, or The
Preservation of Favored Races in The Struggle for Life. First edition. London: John
Murray.
244
Deb, K. (1995). Optimization for Engineering Design, Prentice-Hall, New Delhi.

Deb, K. (2005). Practical Optimization Using Evolutionary Methods. Kanpur Genetic
Algorithm Lab (KanGAL), GAL Report No. 2005008, Kanpur, India.
De Jong, K. A. (1975). An Analysis of The Behavior of a Class of Genetic Adaptive
Systems. PhDl dissertation, University of Michigan, Ann Arbor. (University Microfilms
No. 76-9381).
De Jong, K. A. (1993). Genetic Algorithms are Not Function Optimizers. In L.D.
Whitley, ed., Foundations of Genetic Algorithms 2. Morgan Kaufmann, pp. 5-17.
Eiben, A., Hinterding, R. and Michalewicz, Z. (1999). Parameter Control in
Evolutionary Algorithms. IEEE Transactions on Evolutionary Computation, Vol. 3 No.
2, 124-141.
Falkenauer, E. (1998). Genetic Algorithms and Grouping Problems, John Wiley and
Sons.
Federal Highway Administration (FHWA) (1998).
Rehabilitation, Course No. FHWA-NHI-131008.
Techniques
for
Pavement
Federal Highway Administration (FHWA) (1990). Spring Load Restrictions, Pavement

Newsletter No. 17, US Department of Transportation.
Fogarty, T., and Terence, C. (1989). Varying the Probability of Mutation in The Genetic
Algorithm, Proceeding of the Third International Conference on Genetic algorithms, pp.
104-109.
Fogel, L.J., Owens, A.J., and Walsh, M.J. (1966). Artificial Intelligence through
Simulated Evolution. Wiley, New York.
Foster, C.R., and Ahlvin, R.G. (1954). Stresses and Deflections Induced by a Uniform
Circular Load, Proceedings, Highway Research Board, Vol.33, pp. 467-470.
Futuyma, Douglas J. (2005). Evolution. Sunderland, Sinauer Associates, Massachusetts.
Fwa, T.F., Tan, C.Y. and Chan, W.T. (1997). Backcalculation Analysis of Pavement
Layer Moduli using Genetic Algorithms, Transportation Research Record 1905, TRB,
National Research Council, Washington, D.C., pp 134142.
Fwa, T.F., and Rani, T.S. (2005). Seed Modulus Generation Algorithm for
Backcalculation of Flexible Pavement Moduli, Transportation Research Record 1905,
TRB, National Research Council, Washington, D.C., 117-127.
245
Goldberg, D. E. (1987). Simple Genetic Algorithms and The Minimal, Deceptive

Problem. In Davis, L., ed., Genetic Algorithms and Simulated Annealing, Los Altos, CA,
Morgan Kaufmann, pp. 7488.
Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization, and Machine
Learning. Reading, M.A, Addison-Wesley.
Goldberg, D.E. (1993). Making Genetic Algorithms Fly. A lesson from The Wright
Brothers. Advanced Technology for Developers, No. 2, pp. 1-8.
Goldberg, D.E. (2002). The Design of Innovation, Lessons from and for Competent
Genetic Algorithms. Kluwer Academic Publishers, Boston, MA.
Goldberg, D. E. and Deb, K. (1991). A Comparative Analysis of Selection Schemes used
in Genetic Algorithms, Foundations of Genetic Algorithms, G. J. E. Rawlins, ed., pp. 6993.
Goldberg, D. E., Deb, K., and Clark, J. H. (1992). Genetic Algorithms, Noise, and the
Sizing of Populations. Complex Systems, Vol. 6, pp. 333-362.
Goldberg, D.E., and Richardson, J. (1987). Genetic Algorithms with Sharing for
Multimodal Function Optimization. Genetic algorithms and Their Applications:
Proceedings of the Second International Conference on Genetic Algorithms, p. 41-49.
George, K.P, (2003). Falling Weight Deflectometer for Estimating Subgrade Moduli,
Mississippi Department of Transportation, Report No. MS-DOT-RD-03-153.
Grefenstette, J.J. (1986). Optimization of Control Parameters for Genetic Algorithms.
IEEE Transactions on Systems, Man, and Cybernetics, Vol. 16, No. 1, p. 122-128.
Gustafson, M. (2004). An Analysis of Diversity in Genetic Programming. PhD
dissertation, The University of Nottingham, United Kingdom,.
Haldane, J.B. (1953). The Measurement of Natural Selection. Proceedings of the 9th
International Congress of Genetics, Vol. 1, pp. 480487.
Harichandran, R., Mahmood, T., Raab, A., and Baladi, G. (1993). Modified Newton
Algorithm for Backcalculation of Pavement Layer Properties, Transportation Research
Record 1384, TRB, National Research Council, Washington, D.C., pp. 15-22.
Harichandran, R. S., Ramon, C. M., and Baladi, G. Y. (1994). MICHBACK User'
s
Manual, Department of Civil and Environmental Engineering, Michigan State University,
East Lansing, Michigan.
246
Harik, G. R., Cant-Paz, E., Goldberg, D. E., and Miller, B. (1997). The Gamblers Ruin
Problem, Genetic Algorithms, and The Sizing of Populations, Bck, T. ed., Proceedings
of 1997 IEEE International Conference on Evolutionary Computation, New York, IEEE
Press, pp. 712.
Harik, G., and Lobo, F. (1999). A Parameter-less Genetic Algorithm, Banzhaf, W.,
Daida, J., Eiben, A. E., Garzon, M. H., Honavar, V., Jakiela, M., & Smith, R. E., eds.,
GECCO-99: Proceedings of the Genetic and Evolutionary Computation Conference, San
Francisco, CA, Morgan Kaufmann, pp. 258267.
Hesser, J. and Manner, R. (1991). Towards an Optimal Mutation Probability in Genetic
Algorithms, Proceeding of the 1st Parallel Problem Solving from Nature, Springer, pp.
23-32.
Heukelom, W., and Foster, C. (1960). Dynamic Testing of Pavements. ASCE SM
Journal, Vol. 86, No. 1,pp. 1-28.
Hinterding, R., Michalewicz, Z., and Peachy, T. (1996). Self-Adaptive Genetic
Algorithm for Numeric Functions, Proceeding of the Fourth International Conference on
Parallel Problem Solving from Nature, in Lecture Notes from Computer Science,
Springer Verlag, pp. 420-429.
Hoffman, M.S., and Thompson, M.R. (1981). Mechanistic Interpretation of
Nondestructive Testing Deflections, Civil Engineering Studies, Transportation
Engineering Series No. 32, Illinois Cooperative Highway and Transportation Research
Program Series No. 190, University of Illinois, Urbana, IL.
Hoffman, M.S. (1983). Loading Mode Effects on Pavement Deflections. Journal of
Transportation Engineering, Vol. 109, No. 5, pp. 61-668.
Holland, J. H. (1975). Adaptation in Natural and Artificial Systems. Ann Arbor,
University of Michigan Press.
Holland, J.H. (1992). Adaptation in Natural and Artificial Systems, An Introductory
Analysis with Applications to Biology, Control and Artificial Intelligence, MIT Press.
Holtz, R., and Kovacs, W. (1981). An Introduction to Geotechnical Engineering, Prentice
Hall, Inc., New Jersy.
Huang, Y. H. (1993). Pavement Analysis and Design, Prentice Hall, New Jersey.
247
Ioannides, A.M., Barenberg, E.J., and Lay, J.A. (1989). Interpretation of Falling Weight
Deflectometer Results Using Principles of Dimensional Analysis, Proceedings of the
Fourth International Conference on Concrete Pavement Design and Rehabilitation,
Purdue University.
Irwin, L.H. (1983). Users Guide to MODCOMP2, Report No. 83-8, Cornell University
Local Roads Program, Cornell University.
Irwin, L.H. (1992). Instructional Guide for Back-Calculation and the Use of MODCOMP
3, Version 3.6, Report 92-5, Cornell Local Roads Program, Cornell University.
Jones, R. (1963). Following Changes in the Properties of Road Bases and Subbases by
the Surface Wave Propagation Method. Civil Engineering and Public Works Review,
Vol. 58, pp.613, 615,617.
Jones, R. (1965). Thickness and Quality of Cemented Surfacing and Bases Measuring
by a Non-Destructive Surface Wave Method. Civil Engineering and Public Works
Review, Vol. 60, pp.523, 525,527,529.
Jones, R., and Thrower, E. (1965). Effect of Interfacial Contact on the Propagation of
Flexural Waves along a Composite Plate, J. Sound Vib., Vol. 2, No. 2, pp. 167-174.
Julstrom, B.A. (1995). What Have You Done for Me Lately? Adapting Operator
Probabilities in a Steady-state Genetic Algorithm, Proceeding of the Sixth International
Conference on Genetic Algorithms, pp. 81-87, Morgan Kufmann.
Kameyama, S., Himeno, K., Kasahara, A., and Maruyama, T. (1998). Backcalculation of
Pavement Layer Moduli using Genetic Algorithms, 8th International Conference on
Asphalt Pavements, University of Washington, Seattle, Washington, pp 13751385.
Kauffman, S. (1993). The Origins of Order: Self-origination and Selection in Evolution,
Oxford University Press, New York, NY.
Koza, J.R. (1992). Genetic Programming: On the Programming of Computers by Means
of Natural Selection, MIT Press.
Lande, R., Arnold, S. (1983). The Measurement of Selection on Correlated Characters.
Evolution, Vol. 37, pp. 12101226.
Lobo, F. G. (2000). The Parameter-less Genetic Algorithm: Rational and Automated
Parameter Selection for Simplified Genetic Algorithm Operation. Doctoral dissertation,
Universidade Nova de Lisboa, Lisboa.
248
Lytton, R. L., Germann, F.P., Chou, Y.J., and Stoffels, S.M. (1990). Determining
Asphaltic Concrete Pavement Structural Properties by Nondestructive Testing, NCHRP
Report 327, Transportation Research Board.
Lytton, R.L., Roberts, F.L., and Stoffels, S. (1986). Determination of Asphaltic Concrete
Pavement Structural Properties by Nondestructive Testing, Report NCHRP 10-27, Phase
I, Texas Transportation Institute, Transportation Research Board, Washington, DC.
Mamlouk, M.S. (1985). Use of Dynamic Analysis in Predicting Field Multilayer
Pavement Moduli, Transportation Research Record 1043, Transportation Research
Board, Washington, DC, pp. 113-120.
Martincek, G. (1994). Dynamics of Pavement Structures, E & FN SPON, London.
Matsumoto, M. and Nishimura, T. (1998). Mersenne Twister: a 623-dimensionally
Equidistributed Uniform Pseudorandom Number Generator. ACM Transactions on
Modeling and Computer Simulation, Vol. 8, No. 1, pp. 3-30.
McKay, R. (2000). Fitness Sharing in Genetic Programming. Proceedings of the Genetic
and Evolutionary Computation Conference, Whitley, D. et al., eds., Las Vegas, NV,
,Morgan Kaufmann, pp. 435-442.
Michelow, J. (1963). Analysis of Stresses and Displacements in N-Layered Elastic
System Under a Load Uniformly Distributed on a Circular Area, Chevron Oil Research,
Richmond, California.
Minqiang, L., and Jisong, K. (2001). A New Non-monotone Fitness Scaling for Genetic
Algorithm., Progress in Natural Science, Vol. 11, No. 8, pp. 622-630.
Mitchell, M. (1999). An Introduction to Genetic Algorithms, MIT Press.
Misra, D. and Sen, B. (1975). Stresses and Displacements in Granular Materials due to
Surface Load. International Journal of Engineering Science, Vol. 13, pp. 743-761.
Mhlenbein, H. (1991). Evolution in Time and Space-the Parallel Genetic Algorithm,
Foundations of Genetic Algorithms, G. Rawlins, ed, Sand Mateo, Morgan Kaufman, pp.
316-338.
Mhlenbein, H. (1992). How Genetic Algorithms Really Work: Mutation and
Hillclimbing, Parallel Problem Solving from Nature, Manner, R. and Manderick, B. eds.,
Amsterdam, North-Holland, pp. 15-26.
National Research Council (1993). SHRPs Layer Moduli Back Calculation Procedure,
Strategic Highway Research Program, Report SHRP-P-655.
249
Nazarian, S., Yuan, D., and Baker, M.R. (1995). Rapid Determination of Pavement
Moduli with Spectral-Analysis-of-Surface-Waves Method, Research Record 1243-1F,
Center for Geotechnical and Highway Materials Research, University of Texas at El
Paso.
Newmark, N.M. (1935). Simplified Computation of Vertical Pressures in Elastic
Foundations, University of Illinois Engineering Experiment Station Circular 24, Urbana,
Illinois.
Ochoa, G., Harvey, I. and Buxton, H. (1999). On Recombination and Optimal Mutation
Rates, Proceedings of Genetic and Evolutionary Computation Conference (GECCO-99).
Pan, E. (1989a). Static Response of a Transversely Isotropic and Layered Half-space to
General Dislocation Sources. Phys. Earth Planet. Inter., Vol. 58, pp. 103-117.
Pan, E. (1989b). Static Response of a Transversely Isotropic and Layered Half-space to
General Surface Loads. Phys. Earth Planet. Inter., Vol. 54, pp. 353-363.
Pan, E. (1990). Thermoelastic Deformation of a Transversely Isotropic and Layered
Half-space by Surface Loads and Internal Sources. Phys. Earth Planet. Inter., Vol. 60, pp.
254-264.
Pan, E. (1997). Static Green'
s Functions in Multilayered Half-spaces. Applied
Mathematical Modelling, Vol. 21, pp. 509-521.
Pan, E. and Alkasawneh, W. (2006). An Exploratory Study on Functionally Graded
Materials with Application to Multilayered Pavement Design, Report No. FHWA/OH6936-01, Ohio Department of Transportation, Columbus, Ohio.
Potter, D.W, and Donald, G.S. (1985). Revision of The NAASRA Interim Guide to
Pavement Thickness Design, Australian Road Research, Technical Note 2, Vol. 15, No.
2.
Ray, T.S. (1991). Is It Alive or Is It GA, Proceedings of the Fourth International
Conference on Genetic Algorithms, Morgan Kaufmann, pp. 527-534.
Rechenberg, I. (1965). Cybernetic Solution Path of an Experimental Problem, Ministry of
Aviation, Royal Aircraft Establishment, United Kingdom.
Rechenberg, I. (1973). Evolutions Strategie: Optimierung Technischer Systeme Nach
Prinzipien der Biologischen Evolution. Frommann-Holzboog, Stuttgart.
250
Reddy, M.A., Reddy, K.S., and Pandey, B.B. (2004). Selection of Genetic Algorithm
Parameters for Backcalculation of Pavement Moduli. The International Journal of
Pavement Engineering, Vol. 5, No. 2, pp. 81-90.
Roesset, J.M., and Shao, K.Y. (1985). Dynamic Interpretation of Dynaflect and Falling
Weight Deflectometer Tests, Transportation Research Record 1022, Transportation
research Board, National Research Council, Washington, DC, pp. 7-16.
Rosca, J. and Ballard, D. (1995). Causality in Genetic Programming, Proceedings of the
Sixth International Conference on Genetic Algorithms, In Eshelman, L., ed., Pittsburgh,
PA, Morgan Kaufmann, pp. 256-263.
Ryan, C. (1994). Pygmies and Civil Servants, Advances in Genetic Programming,
Kinnear, Jr., K., ed., MIT Press, Cambridge, MA, pp. 243-263..
Schaffer, J. D., Caruana, R. A., Eshelman, L. J., and Das, R. (1989). A Study of Control
Parameters Affecting Online Performance of Genetic Algorithms for Function
Optimization, Proceedings of the Third International Conference on Genetic Algorithms
Schaffer, J. D. ed., San Mateo, CA, Morgan Kaufmann, pp. 5160.
Schiffman, R. L. (1962). General Solution of Stresses and Displacements in Layered
Elastic Systems, Proceedings of the International Conference on the Structural Design of
Asphalt Pavement, University of Michigan, Ann Arbor, USA.
Schlierkamp-Voosen, D., and Muhlenbein, H. (1996). Adaptation of Population Sizes by
Competing Subpopulations, Proceeding of International Conference on Evolutionary
Computation (ICEC96), Negoya, Japan, pp. 330-335.
Schwefel, H-P. (1977). Numerische Optimierung von Computer-modellen Mittels der
Evolutionsstrategie. Basel, Birkaeuser.
Seeds, S.B., Ott, W.C., Milchail, M., and Mactutis, J.A. (2000). Evaluation of Laboratory
Determined and Nondestructive Test Based Resilient Modulus Values From WesTrack
Experiment, Nondestructive Testing of Pavements and Backcalculation of Moduli,
ASTM STP 1375, S. Tayabji and O. Lukanen, eds., ASTM.
Shell International Petroleum Company Limited (1978). Shell Pavement Design Manual.
Shook, J. F., Finn, F. N., Witczak, M. W., and Monismith, C. L. (1982). Thickness
Design of Asphalt Pavements The Asphalt Institute Method. Fifth International
Conference on the Structural Design of Asphalt Pavements, Vol. 1, pp. 17-44.
251
Sivaneswaran, N., Kramer, S.L., and Mahoney, J.P. (1991). Advanced Backcalculation
using a Nonlinear Least Squares Optimization Technique, Presented at the 70th Annual
Meeting of The Transportation Research Board, Washington, D.C.
Skok, E.L., Clyne, T.R., Johnson, E., Timm, D.H., and Brown, M.L. (2003). Best
Practices for The Design and Construction of Low Volume Roads Revised, Report No.
MN/RC-2002-17REV, Minnesota Department of Transportation Research Services
Section, Minnesota.
Smith, S.F. (1980). A Learning System Based on Genetic Adaptive Algorithms, PhD
dissertation, University of Pittsburgh.
Smith, R.E. (1993). Adaptively Resizing Populations: An Algorithm and Analysis,
Technical Report No. 93001, The University of Alabama, Tuscaloosa, AL.
Smith, R. E. and Smuda, E. (1995). Adaptively Resizing Populations: Algorithm,
Analysis, and First Results. Complex Systems, Vol. 9, pp. 4772.
Spears, W.M. and De Jong, K.A. (1991). On the Virtues of Parameterized Uniform
Crossover, Proceeding of the Fourth International Conference on Genetic Algorithms,
R.K. Belew and L.B. Booker, eds., Morgan Kaufmann, pp. 230-236.
Srinivas, M., and Patniak, L. M. (1994). Adaptive Probabilities of Crossover and
Mutation in Genetic Algorithms. IEEE Transactions on Systems, Man and Cybernetics,
Vol. 24, No. 4, pp. 17-26.
Stolle, D. (2002). Pavement Displacement Sensitivity to Layer Moduli. Canadian
Geotechnical Journal, Vol. 39, pp. 1395-1398.
Syswerda, G. (1989). Uniform Crossover in Genetic Algorithms, Proc. of the 3rd Int.
Conf. on Genetic Algorithms, Schaffer, J. D., ed., Morgan Kaufmann, pp. 2-9.
Syswerda, G. (1991). Schedule Optimization using Genetic Algorithms, Handbook of
Genetic Algorithms, Lawrence Davis, ed., Van Nostrand Reinhold, New York, pp. 332349.
Szendrei, M., and Freeme, C. (1970). Dynamic Techniques for Testing Pavement
Structures, National Institute of Road Research, Bulletin 9, South Africa.
Tsai, B., Kannekanti, V., and Harvey, J.T. (2004). Application of Genetic Algorithm in
Asphalt Pavement Design, Transportation Research Board, No. 1891, National Research
Council, Washington, D.C., pp. 112-120.
252
Ullidtz, P. (1973). The Use of Dynamic Plate Loading Tests in Design of Overlays, The
Conference on Road Engineering in Asia and Australia, Kuala Lumpur.
Ullidtz, P. (1987). Pavement Analysis, Developments in Civil Engineering, Elsevier,
Amestrdam.
Ullitdz, P. (1998). Stresses and Strains in a Two-Dimensional Particulate Material, 74th
Annual Meeting of the Transportation Research Board, Washington D.C.
Uzan, J. (1986). Computer Program MODULUS, Appendix E, NCHRP Project 10-27,
Texas Transportation Institute, Texas A&M University System, College Station, Texas,
pp. E1-E22.
Uzan, J., Lytton, R.L., and Germann, F.P. (1989). General Procedure for Backcalculating
Layer Moduli, Nondestructive Testing of Pavements and Backcalculation of Moduli,
ASTM STP 1026, A.J. Buch III and G.Y. Baladi, eds., American Society for Testing and
Materials, pp. 217-228.
Uzan, J., Michalek, C., Paredes, M., and Lytton, R. (1988). A Microcomputer Based
Procedure for Backcalculating Layer Moduli from FWD Data, Research Report 113-1,
Texas Transportation Institute.
Van der Loo, J.M. (1982). Simplified Method for Evaluation of Asphalt Pavements,
Proceedings the Fifth International Conference on Structural Design of Asphalt
Pavements, The Netherlands, Vol. 1, pp. 475-481.
Von Quintus, H.L. and Killingsworth, B. (1997). Design Pamphlet for the
Backcalculation of Pavement Layer Moduli in Support of the 1993 AASHTO Guide for
the Design of Pavement Structures, Publication No. FHWA-RD-97-076, Washington,
DC, Federal Highway Administration.
Von Quintus, H.L., and Simpson, A. (2002). Back-Calculation of Layer Parameters for
LTPP Test Sections Volume II: Layered Elastic Analysis for Flexible and Rigid
Pavements, Report No. FHWA-RD-01-113, Federal Highway Administration.
Warren, T., and Diekmann, W.L. (1963). Numerical Computation of Stresses and Strains
in Multilayer Asphalt Pavement System, Internal Report, Chevron Research Company.
Westergaard, H.M. (1938). A Problem of Elasticity Suggested by a Problem in Soil
Mechanics: A Soft Material Reinforced by Numerous Strong Horizontal Sheets,
Contributions to the Mechanics of Solids, Stephen Timoshenko 60th Anniversary
Volume, Macmillian, New York, pp. 268-277.
253
White, A. (1993). Integrating Automata with Genetic Algorithms in Order to Provide

Adaptive Operators, M.S. Thesis, Ottawa-Carleton Institute for Computer Science,
Carleton University, Ottawa, Ontario.
Whitley, D. (1994). A Genetic Algorithm Tutorial. Statistics and Computing Journal,
Vol. 4, No. 2, pp. 65-85.
Winston, P. (1992). Artificial Intelligence, Third Edition. Addison-Wesley.
Yoder, E.J. (1959). Principles of Pavement Design, John Wiley & Sons, Inc., London.
254

Back Calculation

Uploaded by

Copyright:

Available Formats

Back Calculation

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Back Calculation

Uploaded by

Copyright:

Available Formats

BACKCALCULATION OF PAVEMENT MODULI

USING GENETIC ALGORITHMS

BACKCALCULATION OF PAVEMENT MODULI

Dean of the College

Dean of the Graduate School

2.6.1 Maxwell Model ..................................................................................20

2.11 Conclusions ...............................................................................................49

4.4.5.2 Crossover ..............................................................................90

VI. OPTIMIZATION OF GENETIC ALGORITHMS OPERATORS AND

7.5 The Effect of the Number of Sensors ........................................................209

Boussinesqs Equations for a Point Load . ............................................12

Typical Poissons ratios of pavement materials ....................................42

Typical values of modulus of elasticity for pavement materials . .........42

Parameters of the Original (Exact) Layers ............................................55

Parameters of the Backcalculated Layers . ............................................56

Coordinates and Locations of the Response Points . .............................56

Exact Responses from the Original (Exact) Pavement Profile . ............57

Response (Top Row) and Relative Error (Bottom Row in Parenthesis)

Response (Top Row) and Relative Error (Bottom Row in Parenthesis)

Comparison of Fatigue Life Using Backcalculated and Exact Pavement

Comparison of Rutting Failure Using Backcalculated and Exact

Four layer pavement system. ...............................................................110

Range of GA inputs used in the four-layer pavement .........................111

Ranges of the moduli . .........................................................................111

Results of the backcalculation programs . ..........................................112

Summary of Pavement Backcalculation Studies . ..............................123

Range of GA inputs used by Reddy et al. (2004) . ..............................126

Values of parameters used in the sensitivity analysis ........................127

Moduli ranges of backcalculated moduli ............................................127

Moduli ranges of backcalculated moduli ...........................................156

Typical values of modulus of elasticity for pavement materials . ......158

Genetic parameters used in the backcalculation process (3 layers) ....185

Moduli ranges for the backcalculation process (3 layers) . .................185

Data of the 23-layer pavement example . ............................................186

Genetic parameters used in the backcalculation process (23 layers) ..186

Moduli ranges for the backcalculation process (23 layers) . ...............187

Data of the 4-layer pavement example . ..............................................206

Genetic parameters used in the backcalculation process (4 layers) ....206

Moduli ranges for the backcalculation process (4 layers) . .................206

Notation for Boussinesqs equation (Ullidtz, 1987). .............................12

Typical transformations in the MET method, after Ullidtz (1987)........16

Typical multilayer system, after Ullidtz (1987).....................................16

Burmister two-layer stress curves (Yoder, 1959). .................................24

The Benkelman beam, after Bendana et al. (1994)................................28

The Dynaflect system, after Bendana et al. (1994)................................29

FWD typical sensors configuration. ......................................................30

Graphical presentation of a schema example ........................................79

Schema changes due to crossover .........................................................81

Example of a two-point crossover . .......................................................90

Example where the two-point crossover fails to work ..........................90

Triangular-sharing function. After Goldberg and Richardson (1987) . .94

Power-sharing functions. After Goldberg and Richardson (1987) . ......94

Simple genetic algorithm performance with (a) and without (b)

Flow chart of the genetic algorithm implemented in the

Chromosomal presentation of the trial solutions encoded as binary bits.102

Population sizing relations as a function of number of layers and

Search Space of a 2-Layer Pavement Section......................................117

The 3-layer pavement system used in the sensitivity analysis, after

Generation best fitness of the backcalculated moduli under a constant