Sunteți pe pagina 1din 221

STRUCTURAL DESIGN CODE CALIBRATION USING RELIABILITY BASED

COST OPTIMIZATION

by
Engin Akta

B.S. in Civil Engineering, Middle East Technical University, Turkey, 1993
M.S in Civil and Environmental Engineering, University of Pittsburgh, 1996

Submitted to the Graduate Faculty
of the School of Engineering
in partial fulfillment of
the requirements for the degree of
Doctor
of
Philosophy

University of Pittsburgh
2001
The author grants permission
to reproduce single copies.
___________________
Signed

iii
ACKNOWLEDGEMENTS

I would like to give my sincere thanks to my advisor, Dr. Fred Moses, for his
continuous contributions to my research. Without his guidance, support and
encouragement, throughout the course of this research and all my M.S. and Ph. D. studies
here at the University of Pittsburgh none of this would have been possible. I am honored
and grateful to have had the opportunity to work under his supervision.
I would like to thank the members of my Dissertation committee, Dr. Christopher
J. Earls, Dr. Jeen-Shang Lin, Dr. Dipo Onipede, Jr. and Dr. Luis E. Vallejo for their time,
constructive input and contributions.
I am very thankful to Dr. Michel Ghosn of CUNY, for his contributions in the
course of this research.
I would like to thank to my wife Seda Akta, for her continuous support, patience
and love and I would like to apologize to my daughter dil Akta for not being next to her
for about a year during the last phase of this study. I would like to thank my parents
Mefharet and Abdurrahman Akta, my mother-in-law kran Sekin, my sister Mfide
Budak, my brother Ersin Akta and my brother-in-law Serdar Akar for their endless love
and support. I also would like to thank my grandmother mmehan encan who taught
me to go after my dreams and to my uncle Fahri encan for being my mentor.
I would like to thank all my officemates for their friendship, which I will cherish
for the rest of my life.

iv
I am also indebted to zmir Institute of Technology in Turkey for providing
scholarship support to my graduate study at the University of Pittsburgh.

v



ABSTRACT



Signature

Prof. Dr. Fred Moses

STRUCTURAL DESIGN CODE CALIBRATION USING
RELIABILITY BASED COST OPTIMIZATION

Engin Akta, Ph.D.
University of Pittsburgh

Structural design codes provide minimum requirements to assure at least a certain
level of safety. Codes may need some revisions due to availability of new data on random
variables such as material or loading and improvements in analytical tools used in the
design process, access to new or better construction materials, or changing demand on the
vi
use of the structure. In current Load and Resistance Factor Design (LRFD) practice, the
code calibration is carried out by using predefined target (reliability) levels. Defining the
target safety level is not clear since codes involve multiple load combinations including
combination of time dependent loads, and a dilemma arises as to which target safety
index should be used in calibrating load factors for these combinations.
This study investigates use of reliability based cost optimization in calibration of
design codes. Load factors that make the total expected cost a minimum are the optimum
load factors and the corresponding safety index is the optimum safety index for each load
case. No predefined target safety levels are used herein; however, the contribution of the
current code in the calibration process is satisfied by deducing the failure cost that a
current level of safety implies in the current code. The most basic gravity load
combination is chosen to deduce the failure cost. First Order Reliability Method (FORM)
is used for reliability analysis, and the combination of time dependent loads is derived
with the Ferry-Borges Method. Since application of Ferry-Borges method requires
independency between load events, the part coming from the time independent modeling
uncertainties are separated herein from the time dependent part and calculations are
carried out by combining the FerryBorges Method with a Nested Reliability Analysis.
The approach is illustrated with application to a bridge design specification including
dead, live, wind and earthquake load combinations. Optimum total cost including failure
cost and safety indices are compared to existing code format. A recommended load factor
table is presented as a product of the calibration process.
vii
DESCRIPTORS

Code Calibration Code Optimization
Conditional Probability Cost Optimization
Ferry-Borges Load Model First Order Reliability Method
Load and Resistance Factor Design Nested Reliability Analysis
Rackwitz-Fiessler Algorithm Reliability Based Cost Optimization
Structural Reliability Time Dependent Reliability Analysis


viii
TABLE OF CONTENTS

ACKNOWLEDGEMENTS...........................................................................................iii
ABSTRACT.................................................................................................................... v
LIST OF FIGURES...................................................................................................... xiv
LIST OF TABLES ....................................................................................................... xvi
NOMENCLATURE....................................................................................................xxii
1.0 INTRODUCTION................................................................................................. 1
1.1 Background ................................................................................................. 1
1.2 Problem Statement ...................................................................................... 6
1.3 Objectives.................................................................................................... 7
1.4 Literature Review...................................................................................... 10
1.5 Methodology ............................................................................................. 15
1.6 Recommended Load Factors Table........................................................... 19
2.0 STRUCTURAL RELIABILITY THEORY AND PRACTICE.......................... 22
2.1 Structural Reliability ................................................................................. 22
2.2 Structural Codification.............................................................................. 23
2.3 Code Calibration ....................................................................................... 26

ix
2.4 Simulation Methods .................................................................................. 33
2.5 Load Combination..................................................................................... 34
2.5.1 Turkstras Rule.............................................................................. 35
2.5.2 Ferry-Borges Process .................................................................... 37
2.5.3 Wens Load Coincidence Method................................................. 41
2.6 Modifications to Ferry-Borges Model....................................................... 42
2.6.1 Mixed Type Distributions ............................................................. 42
2.6.2 Separation of Time Variant and Invariant Random
Variables........................................................................................ 44
2.6.3 Evaluation of P
f
by Numerical Integration or Nested
Reliability Analysis....................................................................... 46
2.6.4 Logarithmic Approximation to Yearly Safety Indices.................. 52
2.7 Sensitivity Analysis................................................................................... 54
2.8 System Reliability..................................................................................... 55
3.0 STRUCTURAL OPTIMIZATION..................................................................... 58
3.1 General Optimization Formulation ........................................................... 58
3.2 Solving Nonlinear Programming Problem................................................ 59
3.3 Sequential Quadratic Programming .......................................................... 60
3.4 Optimization Using MATLAB................................................................. 61

x
4.0 RELIABILITY BASED STRUCTURAL OPTIMIZATION AND
APPLICATION IN CODE CALIBRATION ..................................................... 63
4.1 Formulation of the Expected Total Cost ................................................... 65
4.1.1 Initial Cost ..................................................................................... 66
4.1.2 Failure Cost ................................................................................... 68
4.1.3 Normalization of the Terms in Total Cost Function ..................... 70
4.2 Optimization Problem............................................................................... 71
4.3 Solution Procedure .................................................................................... 72
4.4 Numerical Example................................................................................... 75
4.4.1 Separation of Modeling Uncertainty for Live Load...................... 77
4.4.2 Separation of Modeling Uncertainty for Wind Load .................... 79
4.4.3 Statistical Data after Separation of Random Variables................. 80
4.4.4 Load Combination Cases .............................................................. 80
4.4.5 Calibration Process........................................................................ 81
4.4.6 Recommended Load Factors......................................................... 96
4.4.7 Discussion of Results .................................................................... 97
5.0 APPLICATION OF CODE CALIBRATION BY USING
RELIABILITY BASED COST OPTIMIZATION TO AASHTO
BRIDGE SPECIFICATIONS............................................................................. 99

xi
5.1 AASHTO Bridge Specifications ............................................................... 99
5.2 Design Format and Design Checks for AASHTO LRFD....................... 102
5.3 Load and Resistance Data ....................................................................... 105
5.3.1 Resistance Data ........................................................................... 106
5.3.2 Dead Load Data........................................................................... 107
5.3.3 Live Load Model ......................................................................... 108
5.3.4 Wind Load Model ....................................................................... 113
5.3.5 Earthquake Load Model .............................................................. 117
5.4 Cost Data................................................................................................. 123
5.5 Probabilistic Total Cost Model ............................................................... 124
5.5.1 Strength III: Combination of Dead and Wind Load.................... 129
5.5.2 Strength V: Combination of Dead, Live and Wind Load ........... 137
5.5.3 Extreme Event I: Combination of Dead, Live and
Earthquake Load ......................................................................... 146
5.6 Optimum Load Factors Table and Discussion of Results....................... 160
6.0 SENSITIVITY ANALYSIS.............................................................................. 162
6.1 Effect of Initial Cost Slope Change, C
I
on K
G
...................................... 162
6.2 Effect of Gravitational Load Cost Factor, K
G
on Deduced
Failure Cost Factor, g.............................................................................. 163

xii
6.3 Effect of Design Life, T on Deduced Failure Cost Factor, g .................. 164
6.4 Effect of Real Interest Rate, j on Deduced Failure Cost Factor,
g. .............................................................................................................. 165
6.5 Effect of Gravitational Load Cost Factor, K
G
on Optimum
Load Factors............................................................................................ 166
6.6 Effect of Design Life, T on Optimum Load Factors ............................... 167
6.7 Effect of Real Interest Rate, j on Optimized Load Factors and
Safety Index............................................................................................. 168
6.8 Effect of Failure Cost Factor, g on Optimized Load Factors and
Safety Index............................................................................................. 169
6.9 Effect of Separation of Time Variant and Invariant Parts on
Probability of Failure Evaluation............................................................ 170
6.10 Effect of Separation of Time Variant and Invariant Parts on
Optimum Load Factors............................................................................ 172
6.11 Effect of COVs on Optimum Load Factors ........................................... 172
7.0 CONCLUSIONS AND FUTURE RESEARCH
RECOMMENDATIONS .................................................................................. 176
7.1 Conclusions ............................................................................................. 176
7.2 Future Research....................................................................................... 178

xiii
APPENDIX................................................................................................................. 180
BIBLIOGRAPHY....................................................................................................... 191

xiv
LIST OF FIGURES

FigureNo Page No
1 Algorithm for current calibration practice of a structural design code
(Melchers (1999)).................................................................................................. 28
2 The load and resistance factors for a predefined target safety level
(Ellingwood et al. (1980)) ..................................................................................... 29
3 The safety index vs. cost .................................................................................... 33
4 Ferry-Borges Model ............................................................................................. 40
5 Mixed Type Distribution...................................................................................... 42
6 Approximation to safety index, ......................................................................... 53
7 Parallel and Series Systems.................................................................................. 57
8 Component Safety index, vs. sum of total cost factors, TCF......................... 87
9 Component Safety Index vs. ICF, FCF and TCF......................................... 88
10 Kinzua Bridge, McKean County, PA (Built in 1900) ........................................ 100
11 HL-93 truck and lane loading (AASHTO LRFD(1994))................................... 112
12 Modeling uncertainties, X
W
for wind load.......................................................... 115
13 Extremal Type II fit to single earthquake event ................................................. 122

xv
14 Effect of variable separation on probability of failure calculations .................... 171

xvi

LIST OF TABLES

Table No: Page No
1 Load factor table for ASCE 7-95 (1996)................................................................. 8
2 Current Code Load Factors for AASHTO............................................................ 19
3 Recommended Load Factors................................................................................. 20
4 The statistical data for sample analysis ................................................................. 48
5 Comparison of number of discrete points used (range={-4 to
+4})................................................................................................................... 49
6 Comparison of number of discrete points used (range={-3 to
+3})................................................................................................................... 49
7 Cpu-time for NRA and Numerical Integration ..................................................... 52
8 Logarithmic Approximations to safety index, beta along the lifetime.................. 54
9 Lifetime Statistical Data (T=75 years) .................................................................. 75
10 Statistical data after time variant and invariant distinction................................... 80
11 Data for calculation of gravitational load cost factor, K
G
.................................... 83
12 K
G
values ............................................................................................................... 83

xvii
13 Deducing an average g value using Equation (4-29.a).......................................... 85
14 Deducing an average g value using Equation (4-29.b) ......................................... 85
15 Optimum component and
W
(g=58) for Wind Load Alone Case...................... 91
16 Optimum system and
W
(g=3528) for Wind Alone Case.................................. 91
17 Current code safety levels and costs for D+W case.............................................. 93
18 Optimum component ,
D
and
W
(g=58) for Dead +Wind Load Case ............... 93
19 Optimum system ,
D
and
W
(g=3528) for Dead +Wind Load Case.................. 94
20 Optimum component and
W
(
D
=1.25, g=58) for Dead +Wind Load
Case ....................................................................................................................... 95
21 Optimum system and
W
(
D
=1.25, g=3528) for Dead +Wind Load
Case ....................................................................................................................... 96
22 Recommended Load Factors................................................................................. 97
23 Resistance Statistical Parameters (Nowak 1993) ................................................ 107
24 Dead Load Statistic Data (Nowak (1993)).......................................................... 108
25 The statistical data at time T for one truck loading............................................. 110
26 Live Load Model Statistical Data........................................................................ 112
27 Wind speed statistical data (Ellingwood et al.(1980)) ........................................ 116
28 Wind load statistical data .................................................................................... 117

xviii
29 Distribution for a single earthquake event .......................................................... 121
30 Earthquake Load Statistical Data ........................................................................ 123
31 Calculated KG values.......................................................................................... 127
32 Deduced failure cost factor, g values (component safety) .................................. 128
33 Deduced failure cost factor, g values (system safety) ......................................... 128
34 Statistical data for D+W case.............................................................................. 132
35 Current Code safety levels and total cost factors for D+W................................. 132
36 Optimized component safety index,
comp
and load factors,
D
and
W

for D+W for g=30. .............................................................................................. 133
37 Optimized component safety index,
comp
and load factors,
D
=1.25
and
W
for D+W for g=30.................................................................................... 134
38 Optimized system safety index,
sys
and load factors,
D
and
W
for
D+W for g=1727. ................................................................................................ 135
39 Optimized system safety index,
sys
and wind load factor,
W
for D+W
for
D
=1.25 and g=1727. ..................................................................................... 136
40 Statistical data for D+W+L case ......................................................................... 138
41 The safety index and cost values for D+W+L case using current code
load factors .......................................................................................................... 141

xix
42 Optimized component safety index,
comp
and load factors,
W
, and
L

for D+W+L (
D
=1.25 g=30)................................................................................ 142
43 Optimized component safety index,
comp
and wind load factor,
W
for
D+W+L (
D
=1.25,
L
=1.35 g=30). ...................................................................... 143
44 Optimized system safety index,
sys
and load factors,
W
, and
L
for
D+W+L (
D
=1.25 g=1727). ................................................................................ 144
45 Optimized system safety index, sys and wind load factor,
W
for
D+W+L (
D
=1.25,
L
=1.35 g=1727). .................................................................. 145
46 Statistical data for D+L+E case........................................................................... 147
47 Current code safety index and cost factors for D+E+L case g=1727 ................. 150
48 Optimized system safety index,
sys
and load factors,
E
and
L
for
D+E+L (
D
=1.25, g=30). ..................................................................................... 151
49 Optimized system safety index,
sys
and load factors,
E
and
L
for
D+E+L (
D
=1.25, g=1727).................................................................................. 153
50 Current code safety index and cost factors for D+E+L case g=1727
(with updated range)............................................................................................ 155
51 Optimized component safety index,
sys
and load factor,
E
for D+E+L
(
D
=1.25,
L
=0.50, g=30 with updated range). ................................................... 156

xx
52 Optimized system safety index,
sys
and load factor,
E
for D+E+L
(
D
=1.25,
L
=0.50, g=1727 with updated range). ............................................... 158
53 Optimum Load Factors........................................................................................ 161
54 K
G
values for changing C
I
................................................................................ 163
55 Deduced g values for different K
G
values ........................................................... 164
56 Failure cost factor, g for different design life, T values...................................... 164
57 The effect of real interest rate, j on failure cost factor, g (D+W
combination)........................................................................................................ 165
58 Effect of K
G
on optimum load factors and optimum ranges for D+W
case...................................................................................................................... 166
59 Effect of T on optimum load factors and optimum ranges (D+W
combination)........................................................................................................ 167
60 Optimized safety index, and load factors,
D
and
W
for different j
values (D+W combination) ................................................................................. 168
61 Optimized safety index and load factors for different g values
(D+W+L combination)........................................................................................ 170
62 Optimized safety index and load factors for different g values (D+E+L
combination)........................................................................................................ 170
63 Statistical data for separation example................................................................ 171

xxi
64 Comparison of separation of time variant and invariant random
variables .............................................................................................................. 172
65 Resistance, R COV effect on optimum (D+W combination).............................. 173
66 Dead Load, D COV effect on optimum (D+W combination)............................. 174
67 Wind Speed, V COV effect on optimum (D+W combination) ........................... 175
68 Wind Load Modeling, X
W
COV effect on optimum (D+W
combination)........................................................................................................ 175

xxii
NOMENCLATURE

Symbol Explanation
A acceleration coefficient
C
f
cost of failure
C
I
initial cost
COV coefficient of variation
C
p
pressure coefficient
C
sm
elastic response coefficient
C
T
total Cost
D dead load random variable
D
n
nominal dead load
E earthquake random variable
E
n
nominal earthquake load
E
z
exposure coefficient
f number of failures in MC Simulation
FCF failure cost factor
g deduced failure cost ratio (C
F
/C
0
)

xxiii
g limit state function
G gust factor
G limit state function
H Hessian matrix
h limit state function
ICF initial cost factor
I
D
dynamic effect of live load on bridge
IM dynamic amplification factor
j discount rate
K marginal cost slope
L live load random variable
L
n
nominal live load
P
f
probability of failure
P
fi
probability of failure in i
th
year
p
i
probability of load i exceeding the limit state
P
s
probability of survival
Q
E
time dependent earthquake load effect
Q
L
time dependent live load effect

xxiv
Q
W
time dependent wind load effect
R resistance random variable
R
m
response modification factor
R
n
nominal resistance
S Load effect
S site coefficient
TCF total cost factor
T
m
m
th
vibration mode period
V wind velocity
V
n
nominal wind velocity
W weight of the structure
W wind load random variable
W
n
nominal wind load
X
E
earthquake load modeling uncertainty
X
E
earthquake load modeling uncertainty
X
i
random variables
x
i
*
design values for random variables
X
L
live load modeling uncertainty

xxv
X
W
wind load modeling uncertainty
Z(t) total load effect at time t
CI change in initial cost

L
change in live load factor
nominal load to nominal dead load ratio
safety index
resistance factor
load factor

i0
reference load factors
marginal cost slope ratio

I
rate of occurrence for load i

j
Lagrange multipliers
mean of a random variable
rate of occurrence of events per unit time
standard deviation of a random variable



1
1.0 INTRODUCTION

1.1 Background
Engineered structures have to be designed and used according to specified
requirements. The collection of these requirements is referred herein as the codes or
specifications. The first written document on building regulations is the Code of
Hammurabi and it is dated back to about 2000 B.C. (Nowak and Lind (1995)
*
, Ersoy
(1991)). Although it is far from defining any technical requirements, this Code strictly
emphasizes safety by providing severe penalties such as the death penalty for any
individual responsible for a structural failure causing a fatality. This clearly indicates that
safety was an important issue even in ancient civilized societies.
Structures have been built since civilization was first established. There are many
structures, which were built centuries ago and still stand. Experience has been a key
parameter in the course of structural design development. Every failure has been a lesson
for the constructor. Deterministic experience was the most acceptable safety decision
criteria. The use of probabilistic methods in structural mechanics was only introduced
into design practice in the late 1960's. Since then it is also important to introduce past
experience by scientific approaches. The developments in structural analysis techniques,
statistical data collections, reliability theory, faster computing abilities etc. now make it
possible to generate more robust structural design codes.

*
Parenthetical references refer to the bibliography.


2
Pre-1970 codes were based on experience rather than on probabilistic background;
both the lack of data and the theoretical basis for reliability analysis are the main reasons
for not earlier achieving more sophisticated design codes. Applicability of reliability
theory to structural design began with researchers including Freudenthal (Freudenthal
(1956), Freudenthal et al. (1966)), Cornell (1969), Turkstra (1970), Lind (1970), Moses
(Moses and Kinser (1967), Moses and Stevenson (1970)) etc. The evolution of the design
codes comes to its current stage by use of the Reliability Theory. The loads, resistances,
mathematical models, and analysis methods are all treated as random variables. In order
to express these variables in the design process it is important to consider this
randomness. In reliability analysis, the statistical data used for describing loads and
resistances must also contain the variability due to modeling and analysis in addition to
the randomness of the physical event. The effects of the physical events are modeled
mathematically, but a mathematical model may not replicate the real effects perfectly.
Modeling is an approximation to the real behavior of a phenomenon, such as earthquake
or wind effects on structures; using such approximate models introduces uncertainty into
the system. Also, analysis techniques applied to predict behavior of a physical event or
structures do not replicate the exact behavior with the exception of few cases. Analysis
may produce conservative results relative to what is observed, or the results of the
analysis may also be unconservative compared to observations. In any case, analysis
techniques introduce uncertainties into the design picture.
It is not enough to use the probabilistic approach alone to calibrate a new code;
deciding on the safety levels that a new code should adopt is a challenging step in the


3
calibration process. It is common practice to reach a decision based on experience and
intuition. Although, experience and intuition are always very important, rationalization of
the decision process where possible is worth pursuing. For example, there is an
inconsistency in current codes; different load combinations have different level of safety.
Increasing safety level would introduce higher costs; therefore, lower safety levels are
adapted to keep cost feasible, such as in case of load combinations involving wind or
earthquake loads. It is one of the objectives of this research to overcome the above-
mentioned inconsistency.
Design codes are categorized into four different levels (Melchers (1996)). A Level I
code is in a deterministic format and load factors (in US) and partial coefficients (in
Europe) are used to design the structures. Current codes such as AISC-LRFD (1986),
API-LRFD (1989), AASHTO-LRFD(1994) etc. are in the Level I format and designers
do not involve any probabilistic calculations in use of design codes in this format. Partial
factors as in Europe or load and resistance factors (LFRD) as in US are used in design
checks. The safety is defined with a safety index, , which is simply for a normal
distribution, a distance of mean reserve capacity to zero (failure) in terms of the number
of standard deviations. A typical LRFD check is as in Equation (2-5). Level II format
deals with the probabilistic nature of random variables. Distribution types with mean and
standard deviation are directly used to represent the random variables and reliability is
estimated by an approximate solution such as First Order Reliability Method (FORM)
which approximates the probability of failure using first order terms of Taylor Series
expansion of the limit state function (See Appendix). A conversion to normal distribution


4
is performed at the most probable failure points during the evaluation of the probability
of failure. A Level III format involves a full probabilistic approach; probability of failure
is calculated using a convolution integral. Level IV codes involve all available tools
mentioned above as well as economical data; a minimum total cost or maximum utility is
used as the code objective. It is proposed herein to use Level IV format to calibrate a
Level I (LRFD) code. Total cost is minimized to find optimum safety level and load
factors.
Throughout the text, structural design codes will be referred to as design tools to
proportion the structural members. These codes prescribe the minimum strength
requirements to assure a reasonable safety level as defined by the code committee. In
code calibration practice it is common to use the current codes safety levels as
predefined safety levels for calibration process. Safety index, is calculated along the
design space defined by design points, which are represented by different values of
nominal environmental load to nominal dead load ratio. For example using different
values of L
n
/D
n
ratios can represent bridges with different span lengths. L
n
/D
n
ratio is
smaller for larger span bridges where dead load governs, and it is higher for shorter span
bridges. The safety indices calculated for discrete design points using the present code are
averaged and become the target safety level,
t
for calibration of a new code. Herein an
alternate approach to replace the current practice is studied i.e., using reliability based
cost optimization to calibrate a new code. The sum of expected total costs (initial plus
expected failure cost) is minimized. Since it is hard to quantify the failure cost, the
simplest load combination case is used as a reference and the failure cost is deduced from


5
that reference combination case. The reason for selecting the basic load combination case
is that it would be the most agreed load combination. Using the total cost model
presented herein, the failure cost that basic combination case imposes can than be the
failure cost for all load combination cases covered in the calibration. Doing so means that
even if the safety indices for different combinations are not uniform, the costs of
consequences for these combinations are identical. This helps to implement the trade-off
between the safety and cost explicitly and rationally.
This dissertation contains the following chapters:
Reliability Theory and it is applications are described in second chapter. All the
necessary tools to use in the approach used herein are explained.
In Chapter 3, the optimization techniques are investigated. The software
MATLAB that is also used for the optimization is also introduced in this chapter.
The Reliability Based Cost Optimization with its development in last three
decades is presented in the Chapter 4. The proposed model for code calibration is also
given in detail in this chapter.
In Chapter 5, application of the method to the AASHTO Bridge Specification is
described. The statistical data on random variables are provided in this chapter.
In Chapter 6, the sensitivity analyses are described for the proposed method is
explained. The outputs are investigated for changes in applied data, and results are
discussed.


6
In last chapter discussions on the method and recommendations for future
research are presented.
In the Appendix, the basic reliability problem and the Rackwitz-Fiessler
Algorithm is presented.

1.2 Problem Statement
Structural design specifications impose minimum safety requirements; therefore the
design code sets the standard for the engineering practice. Since the early days of the
application of reliability theory to codified design, the question of how safe is safe
enough is asked both by researchers and engineers. The question is still unanswered and
both researchers and practicing engineers are still working on reaching a sound solution,
but the answer does not seem so straightforward. There is always a risk of failure with
any structure; in other words the probability of failure of a structure is always greater
than zero. So, absolute reliability can never be attained. Since the probability of failure
can be expressed quantitatively, this allows bringing the economical considerations into
the picture. Increasing the safety level is costly, so there must be a balance between the
safety and the cost.
Forssell (1924) stated that a design should be made to maximize the utility to the
owner, including the expected losses (loss of structure, tangible losses, etc.). The
expected cost of the losses can be expressed by multiplying the probability of failure, P
f

with the cost of the losses. Because the loss has not occurred yet and in order to quantify


7
the probable failure cost of losses, the expected value has to be calculated. Forsells
statement has played a very important role in optimization of the structural design codes.
Consideration of the expected losses in the optimization process would allow reaching a
balance between the safety and economy. The expected losses include property damage,
function losses, and personal injuries and life losses. Intangible losses such as personal
injuries and life losses are very hard to determine. Assigning monetary values for a
human life is not easy, and creates many debates. Since intangible losses also have to be
included in order to reach a reasonable solution the structural optimization is rarely
achieved explicitly in current applications.
In order to reach more sound design codes it is important to explicitly express the
code optimization through the total cost optimization. In this study the ways to find a
reasonable approach to that problem is investigated. In some examples of current design
practice the structures are optimized on a project-by-project basis. Generally this
optimization cannot violate the minimum requirements dictated by the code. Therefore if
a code is not optimized, such structural optimization does not give the real optimum
solution.

1.3 Objectives
In this research it is aimed to establish a procedure to calibrate structural design
codes. A balance between safety and economy is sought using Reliability Based Cost
Optimization.


8

Table 1 Load factor table for ASCE 7-95 (1996)
Dead Live Snow Wind EQ
1 Dead 1.60 - - - -
2 Dead+Live 1.20 1.60 - - -
3 a. Dead+Snow+Live 1.20 0.50 1.60 - -
b. Dead+Snow+Wind 1.20 - 1.60 0.80 -
4 Dead+Live+Wind 1.20 0.50 - 1.30 -
5 a. Dead+EQ+Live 1.20 0.50 - - 1.50
b. Dead+EQ+Snow 1.20 - 0.20 - 1.50
6 a. Dead-Wind 0.90 - - 1.30 -
b. Dead-EQ 0.90 - - - 1.50
Load Combination


A sample of a load factor table can be seen in Table 1. The load combinations and
corresponding load factors are given in a code along with the code specified nominal
design loads. In the calibration the load factors such as in Table 1, are calibrated with the
updated statistical data to achieve a reliability criteria.
Current applications of the code calibration provided pre-assigned target safety
indices. These safety levels are mostly deduced from the previous versions of the codes
by selecting some sample designs, designing those samples according to current code,
and finding the average safety index for those sample designs. The target safety levels are
different for loads such as gravity, wind, earthquake, etc. For example in ANSI A.58
(Ellingwood et al. (1980)) the target safety index for combinations of dead load and snow
load is 3.0. On the other hand, the target safety index drops to 2.5 and 1.75, when wind
load and earthquake loads are involved, respectively. It is important to decide on the
target safety levels; a consistency should be sought in the decision process. The common


9
trend in calibration process is to reach uniform target safety levels for the combinations
(Ditlevsen (1997)).
It has also to bear in mind that different loads may affect the structure differently;
for instance, the effect of lateral load such as wind and earthquake might have a higher
relative impact on the cost than the gravitational loads. Also, the loads have different
statistical data (such as COV), which affects cost of failure differently because of the
influence on probability of failure. It might be highly costly to strengthen a structure to a
high reliability target for earthquake then to gravitational loads. In this study the target
safety levels are not predefined, instead the target safety levels are attained implicitly,
using the balance between cost and safety.
In code calibrations, for the load combinations involving time dependent load
effects, probabilistic load combination techniques such as Turkstras Rule, Ferry-Borges
Method and Wens Load Coincidence Method are used; Ferry-Borges Method is selected
to perform load combination analysis for this research. This method relies on the
independence among the repetitions of load events. This is not precise because of
modeling uncertainties involved. A better representation of time dependent load effect is
used herein by separating modeling uncertainties that are constant throughout lifetime
and time dependent events such as environmental events (wind speed, earthquake
acceleration, etc.) or vehicular weight load.
Available data on statistics of resistance and load events along with cost data such
as cost slopes for different type of loads were gathered to use in the proposed calibration
procedure.


10
It is common practice to perform calibration considering only reliability of a
member rather than a system. It would also be important to investigate the system
reliability because the structures are a system of the elements and the reliability of a
structure is the reliability of a system. In this research reliability analysis are mostly
based on element safety, and the effect of system safety has not been studied extensively.
In order to generalize the results of the calibration process, sensitivity analysis are
run and the effect of the variations in the assumed or deduced parameters are
investigated.
As a product of the calibration process, load factors that give the optimal code are
illustrated.

1.4 Literature Review
Structural design codes are widely used. Although every designer should look for
an optimized design, he or she is constrained with the minimum proportioning
requirements imposed by the current code. So a real optimum design highly depends on
an optimum code.
Reliability theory and the optimization techniques have to be used together in
order to optimize a code. Loads and resistance show variability in nature. Therefore
assuming the loads and resistance deterministic would not be a realistic approach.
Reliability theory helps to capture the probabilistic nature of the loads and resistance.


11
Freudenthal's (1947) work can be assumed to be the first step to introduce the
reliability approach to structural design. His later work (Freudenthal (1956), Freudenthal
et al. (1966)) helped researchers to realize the structural reliability concept can be an
efficient way to deal with structural design and analysis. Cornell's (1969) work on the
probabilistic code approach for ACI gave a momentum to the applications of structural
reliability theory to structural design. But lack of data made it difficult to apply. Turkstra
(1970) defined the modern approach to the codification and used a Bayesian approach for
the decision analysis. Bayesian approach is very suitable for structural design because it
uses the available experience and data to improve the estimate of structural reliability in
the decision process. Bayesian approach can simply be identified as systematically
combining judgment, experience and indirect data with the observed data (Ang and Tang
(1975)). Bayesian Approach is one of the vital issues that make a continuous
improvement in structural design codes.
The idea of using optimization along with the reliability analysis has been an
important step to reach balanced designs. Forsell (1924) long before the structural
reliability was in the picture formulated the design process as a minimization of total cost
that covers construction, maintenance, and expected failure cost. The work of Moses and
his colleagues (Moses and Kinser (1967), Moses (1969), Moses and Stevenson (1970))
showed how to integrate the optimization with reliability analysis. Rosenbluth and
Mendoza (1971) formulated the optimum of structural design using minimization of the
expected total cost. Ravindra and Lind (1972) introduced the optimization of code
concept as, minimization of the sum of expected total costs of all structures designed


12
using a structural code. Researchers started to accept the notion of probabilistic based
design and in 1970s the researchers were focused on creating the basis for a probabilistic
and more rationale structural design codes.
American Concrete Institute (1977) is the first organization that adopted the
probability based code concept in USA, but it is the ANSI A.58 Building Code
Requirements for Minimum Design Loads in Buildings and Other Structures (1980) that
set an example for calibrating the design codes. Ellingwood et al. (1980) predefined the
target safety levels using the current code then, and determined the load factors required
for that target safety level, and then obtained the resistance factors. The LRFD studies
involving code calibration followed these same footsteps that Ellingwood and his
colleagues developed.
The number of probability based codes following LRFD philosophy start to
increase. In 1986 American Institute for Steel Construction migrated to a LFRD based
design code. The American Petroleum Institute, API code in 1989 then followed AISC.
Next, American Association of State Highways Officials, AASHTO (1994) migrated to a
load and resistance factor based specification for design of highway bridges.
Developments in the First Order Second Moment (FOSM) reliability methods, and
collection of statistical data helped to provide more rational codes.
Although the theoretical basis of code calibration matured during the last three
decades, the selection of target safety index is still a subject of many debates. The
decision process to decide about the target safety index is controversial and still more
political than it is scientific. Another problem arises since the load combinations do not


13
show a uniform safety level, which brings the question; are the structures less safe for
loads such as wind or earthquake, because it is the current practice to have lower target
safety indices for the combination of loads such as wind and earthquake. The reason
having lower target safety levels is due to the fact that there is a trade off between safety
and cost. This is an implication of putting cost and reliability issues together in the
decision process. But unfortunately, this is only done qualitatively. Intuition is the main
factor in decision process for current practice.
Ellingwood (1994) in his paper summarized the past accomplishments and the
future challenges on codification. He pointed out that target reliability selection is
difficult and has always been the soft side of the probability-based codes. According to
Ellingwood (1994) decisions based on the intuition and judgment fails when the rare
events like earthquake and winds are the basis of the design. He points, Better estimates
of limit state probability are also required in support of life cycle cost analyses.
Ditlevsen (1997) also stressed that the decision on the target reliability levels has
to be based on decision theoretical principles such as expected total cost optimization. He
points that the formal codes safety levels can be used to back calculate the cost of
failure. Ditlevsen (1997) states that Since the costs of consequences of structural
damage or collapse embrace the intangible socio-economic costs such as injury or loss of
human lives and possibly irremediable damages on the environment, there seem to be no
other way than to establish some fix-points in the most frequent existing practice and
calibrate the reliability level to these fix-points. He also proposes to migrate from
partial safety factor code to a full probabilistic code to obtain a code with uniform safety.


14
Rackwitz (2000) gave the basis for the code making process. In his model he also
used the total cost minimization as the decision tool to determine the target safety for the
calibration process where the maintenance and reconstruction costs are also considered.
The renewal processes are used to represent the failure in time, which assumes
independent and identical distribution of load processes, and each reconstruction assumes
a new structure. In load combination analysis, time variant loads are represented by a
Poisson process and outcrossing rates are used to define the failure probability for these
loads.
Researchers including Kanda and Ellingwood (1991), Kanda (1996), Wen and
Kang (1997), Ang and De Leon (1997), Rackwitz (2000), proposed to establish the cost
of failure through the current practice by assigning direct monetary values for the failure.
This approach brings the problem of the assigning values for the intangible cost which is
a political issue and also, it is difficult to assign a value for a life loss or injuries. The
appraisal for human life is mostly based on the work by Viscusi (1993). In his work
Viscusi estimates the value of human life using the labor market data. The outcome of
that approach would be so sensitive to the type of structure, location, type of failure and
to injury and fatality costs, etc. Therefore reaching to a generalized solution would be
difficult and open to many debates. It is believed, it would be more rational to use the
current code as a fix point and back calculate the failure cost using the risk that society is
currently willing to accept.
It is proposed in this study to deduce an implied cost of failure through the most
basic and most agreed load combination from the previous code; such as dead plus live


15
load combination in the AASHTO Bridge Specifications (Aktas et al. (2000, 2001a,
2001b)). There is a consensus on the safety level of this load combination case, so the
risk amount a society is willing to pay to avoid failure can be deduced. As mentioned
earlier considering the trade-off between safety and cost, it is not realistic to assume a
uniform safety level for all the combination cases. When it is not costly to increase
safety, the safety level may be increased as in present criteria wherein a connection has a
higher safety index than other elements (see AISC). Assuming the failure represents the
complete failure, the cost of failure should be the same for all the combinations. The
expected losses do not change from one to another load combination (such as
gravitational loads, earthquake, wind, etc.), i.e. consequences of system failure should be
same. Basically the proposed method is based on that assumption and used herein for
calibrating a reliability-based consistent structural design code.

1.5 Methodology
An optimized code means that the recommended load and resistance factors give
the optimum solution for the total of the expected designs described in a design space.
The total costs for every design has been summed and the optimization is processed for
the sum of the total costs of the designs. Each individual design may not in themselves be
the optimum design. Therefore, the total cost over the applicability of the design code in
the whole design space is optimized instead of individual designs.


16
The total cost of the design consists of the initial cost, C
I
that is a linear function
of the load and resistance factors, and expected failure cost, C
F
that is a product of the
cost of failure, C
f
and probability of failure, P
f
. The expected total cost function is given
in Equation (1-1).
f f I T
P C C C + = (1-1)
Cost of failure is the cost incurred due to a failure, such as property damage, loss
of function, and fatality. The failure can occur anytime during the lifetime of a structure;
hence the probability of failure is calculated annually. In order to make a comparison all
costs have to be on the same basis, such as present worth. Therefore, assuming a constant
cost of failure throughout the design life of the structure, the failure costs are calculated
annually and discounted to present worth value at the time of construction. In order to
simplify the problem a constant nominal real interest rate that is the rate adjusted for
inflation is introduced for the design life. The total cost now has a linear part and a non-
linear part of the probability of failure calculation through FORM iteratively. Non-linear
programming is used to find the optimum load factors.
The calculation of the annual probability of failure involves finding the probability
of failure of the structure in n years, where n is from 1 to lifetime, and the differences of
the two consecutive years gives the probability of failure of a structure specifically in that
year.
Loads including live, wind, earthquake, etc. show variation in both in time and
space. The possibility of all the load effects to affect the structure at their expected


17
maximum values simultaneously is very rare. Therefore, the combined stochastic
variation of loads has to be expressed reasonably. Load combination methods such as
Turkstra's rule and the Ferry- Borges Method can be used to analyze and assess the
combined action of the loads.
It is also important to emphasize that the time dependent loading variables also
contain a part due to uncertainties in modeling, which do not change in time. Therefore
stochastic load terms showing variation throughout the design life are separated into two
parts; time dependent and independent as in wind load case where the wind speed is
converted to pressure on the structure; the wind speed varies with time but the modeling
variable does not. Because the load combination processes depend on the assumption of
independency between load pulses, in order to be consistent with that assumption the
loads needs to be separated into time dependent and independent parts. The Turkstras
rule which is a deterministic approximation of the load combination process based on the
assumption that, when one the load is at its maximum, others are at their arbitrary values,
but it does not allow considering the modeling uncertainties. Therefore the Ferry-Borges
Method is modified and combined with Nested Reliability Analysis to account for the
modeling uncertainties. The conditional probability concept is used to calculate the
probability of failure for the combination of time dependent loads with the modeling
uncertainties.
In this study, the existing code is used to deduce the cost of failure. This allows
validating the current practice into a new code. Consistency along the different load


18
combinations would also be attained, because the cost of failure used would be same for
all the load combinations.
AASHTO bridge design specifications are used to illustrate the proposed
approach. The AASHTO LRFD specification consists of load combinations including
dead, live, wind and earthquake. The first load combination involves gravitational loads;
dead and live load. Live load is the effect of heavy vehicles crossing the span. The work
on the calibration of the AASHTO in 1993 was mostly based on this combination case.
The lack of the available data made it not possible to do an extensive calibration for other
load combinations. The calibrations of the AASHTO specification for extreme loads (e.g.
earthquake, wind, vessel collision, etc.) are now studied under the NCHRP-12-48 project.
The target safety index used for the dead and live load combination case was reported to
be 3.5 (Nowak (1993)). This value is used herein to back calculate the cost of failure by
assuming that the optimum solution to that specific load combination is at the safety level
of 3.5. This assumption is based on the Linds postulate that the current code is already
close to the optimum (Lind (1977)). The deduced cost of failure will hold for all the load
combination cases. The optimized load factors are found for the other load combination
cases and the load factors are developed herein.
One final step is to check for the sensitivity of the optimized load factors to the
random variable data and other assumed input data such as relative cost value and the
discount rate. The analyses are rerun for the changing values and the behavior of the
optimized load factors is investigated.



19
1.6 Recommended Load Factors Table
In this study an example for the proposed calibration approach is presented. The
AASHTO Specification has current code load factors given in Table 2. As a result of the
calibration load factors tabulated in Table 3 are recommended.
The dead load factor for all the combinations studied are kept at the current code
specified value of 1.25. In the load combination case of dead and wind load effects, the
wind load factor has been increased from 1.40 to 1.60, because the safety level obtained
using the current code values gives a high expected failure probability and expected total
cost. The recommended load factors are concluded using the system safety (system safety
index is equal to component safety index plus one) and assuming wind load marginal cost
slope is three times the marginal gravitational load cost slope. The current code gives an
expected total cost of 4.4, where the recommended load factors reduces that to 4.2, which
corresponds to a 5% reduction in the sum of expected total cost for the design space.

Table 2 Current Code Load Factors for AASHTO
Load Combination
D

L

W

D+L 1.25 1.75 --- ---


D+W 1.25 --- 1.40 ---
D+L+W 1.25 1.35 0.40 ---
D+L+E 1.25 0.50 --- 1.00
Load Factors




20
Table 3 Recommended Load Factors
Load Combination
D

L

W

D+L 1.25 1.75 --- ---


D+W 1.25 --- 1.60 ---
D+L+W 1.25 1.35 1.10 ---
D+L+E 1.25 0.50 --- 1.45
Load Factors


In dead, live and wind load combination case the load factors required to keep a
balance between cost and safety are higher than the current code specified values (see
Table 2). The current code load factor value of 0.40 for wind load is due to the
assumption that there would be no truck present on the bridge when the wind speed is
higher than 55 mph. Since Ferry-Borges Method does not allow the change in the limit
state function in time, this constraint is introduced by having lower values for W
n
/D
n

range. Doing so decreases the domination of the wind load in the reliability analysis,
therefore the probability of wind speed star values to go beyond the 55 mph would be
relatively low. Considering system safety (component safety index plus one) it is
recommended to increase the wind load factor to 1.10 in order to have minimum
expected total cost. The sum of expected total costs using current code factors for
combination of dead, wind and live loads is about 29.2. The recommended load factors
for this case lower the sum of expected costs to less than one-half of the current code cost
to 13.0.
In the last combination case, which is the combination of the dead, live and
earthquake load, the safety index calculated is the system safety index, because the


21
earthquake model used in this study already accounts for the system effects. The load
factors for dead and live loads are kept at current code values of 1.25 and 0.50,
respectively. Since the earthquakes marginal cost is the dominating factor the
optimization is run for the earthquake load factor. The load factor for the earthquake load
is found to be 1.44 and 1.45 is recommended as the calibration result. The recommended
earthquake load factor of 1.45 lowers the sum of expected total cost 30%, from the
current codes cost of 27.0 to 20.0.



22
2.0 STRUCTURAL RELIABILITY THEORY AND PRACTICE

2.1 Structural Reliability
Structural reliability is based on the rational quantification of a structures chance
to satisfy the performance requirements. In the design process of structures, which is
mainly proportioning of the elements, the strength of material used and the loads they
will be subjected are not known exactly. These input show variation in space and time
and can be best represented by probabilistic and statistical approaches.
Freudenthal (1947) introduced the reliability concept to structural engineering in
the United States. The structural reliability theory had shown a slow improvement until
mid 1960s, but since then structural reliability and its applications draw considerable
amount of attention from researchers. Especially following the work for the American
National Standard on Minimum Design Loads (Ellingwood et al. 1980), it has been more
accepted by the engineering practice.
Structural reliability is basically the probability of demand not to be greater than
capacity. Demand is the load effects on the structure and capacity is the resistance or
strength of the structure. Structural reliability can be represented by either the probability
of survival, P
s
or the probability of failure, P
f
. The relation between these two
probabilities is
s f
P P =1 (2-1)


23
In structural reliability since P
s
is so close to 1 it is preferable to use the probability of
failure to not lose the significance of the calculated probability values.
As mentioned earlier both demand (load effects) and capacity (strength) show
variation in time (wind, snow, earthquake loads, deterioration of structures over time
reduces the resistance), in location (earthquake prone regions are susceptible to higher
earthquake loads), use (live loads in office and residential buildings, or traffic loads on
bridges). Unfortunately to quantify the variability is not easy and most important this
variance is not unique. The load effects and the strength are both random in nature. For
example, it is a slim chance to test two elements produced, and obtain the same resistance
from these two. Therefore, the best solution to represent variations is using the statistical
theory.
Historically, the main decision tool for designers was performance experience.
One of the main reasons that structural reliability theory and applications show a slow
improvement until 1960s is the lack of sufficient data to represent the design variables
rationally. As the data from the observations of environmental effects and from the
laboratory tests of construction materials of common practice become more abundant, it
made it possible to represent variables reasonably.

2.2 Structural Codification
Although the approach behind the current structural codes is probabilistic, in
order to ease the application and obtain uniformity among the designers, codes are


24
written in deterministic formats. The procedure is called Load and Resistance Factor
Design, LRFD in US, and Partial Safety Factor Design in Europe. Following the code
specifications allows designers to attain a sufficient level of safety. In a sense, structural
design codes may serve as a legal shield to a designer.
Research on structural design codes showed a considerable increase in 1970s
after establishing First Order Second Moment Theory, FOSM, which is also called First
Order Reliability Method, FORM and the gathering of statistical data on random
variables. FOSM is based on representation of the limit state function, G (Equation (A-
6)) in a linear form by using the first term of the Taylor Series Expansion (Equation (A-
8)). The safety is represented by an index, that is simply the number of standard
deviations from origin to mean of a standard normal distributed variable. So the FOSM or
FORM name comes from using the linear (first order) approximation of limit state
function and using the first and second derivatives of a normal distribution, mean and
standard deviation, to define the probability of failure (or safety). When the second
term of the Taylor Series expansion in Equation (A-8) is used to evaluate the limit state
function, the solution becomes more complex. When the second order terms are used, the
method is called Second Order Reliability Method (SORM). In most reported cases, the
accuracy obtained using FORM is sufficient for the Code calibration issues. Therefore
FORM is used in this study instead of SORM. This conclusion is supported by
comparisons with Monte Carlo Simulation.
The improvements in codes are still in progress and studies in US and Europe are
still looking for improved structural design codes. In order to represent the loads and


25
resistances more realistically more data is required. As the time goes by, researchers will
access more accumulated data to represent the loads better. The resistance does reflect the
reality rather well since the low coefficient of variation, COV values, especially for
factory produced members, show that controlled production helps to reduce the
deviations from the nominal properties such as strength, geometry, etc. Most likely, since
the computing power is easy to access with improving technology, even designers may be
asked by design codes to use probabilistic approaches to verify the safety of the
individual designs. At present, some national codes such as the Dutch code allow using
probabilistic computations to check the designs (Hoogenboom and Kasbergen (1998)).
Codes are used through the safety-checking format, which are the mathematical
representation of a limit state. Limit state might be an ultimate limit state (ULS) or
serviceability limit state (SLS). Ultimate limit state represents the physical limit of an
element before failure. When an element is beyond that limit it is nonfunctional or lost. In
case of serviceability it is the level of functioning that is in question; floor excessive
vibration or deflection is an example of a serviceability limit state.
The current safety format used in US is called LRFD. American Concrete
Institute (ACI), American Petroleum Institute (API), American Institute of Steel
Construction (AISC), and American Association of State Highway and Transportation
Officials (AASHTO) now provide the LRFD format for their design specification. The
safety checking equation for LRFD proposed by Galambos and Ravindra (1978) was
im
n
i
i n
S R

=

1
(2-2)


26
where is the resistance factor, R
n
is the nominal resistance,
i
are load factors and S
im

are the mean load effects. This design check equation is applied for different types of
load cases, for different members such as beams, columns, etc. and connections. The
design check equation allow a designer to determine minimum required design resistance
among the different load combinations, such as dead and live load effects, dead and wind
load effects, etc.

2.3 Code Calibration
In a code calibration process, the steps of Ellingwood et al. (1980) are mostly
followed. These steps are composed of selecting the design space; use of existing code to
design sample cases; defining the limit states; statistical data evaluation; use of reliability
methods (FOSM theory or simulation techniques); selection of target safety levels by
using current applications (if the current code gives satisfactory designs the safety index
obtained from the sample designs are averaged and used as the target safety level for
these designs) and finally deciding about the load and resistance factors by minimizing
the deviations from the target safety level for the range of designs (Melchers (1999)). The
current code calibration process can be summarized as the algorithm shown in Figure 1,
and a calibration example of the dead, wind and live load combination using current
calibration practice is given in Figure 2.
It is important to define what type of material and structures are covered by the
code. The codes cover a wide range of design space but in order to make a code more


27
efficient limiting it to a reasonable range would be appropriate. If the deviations from the
target reliability index are high even after minimization it might be an indication that the
range shall be narrower.
Design experience is a vital input of code calibration. It is important to use
experience to improve the codes, because it is the application that can check the
satisfaction of a design code. Code calibration can be viewed as an ongoing process, the
new challenges, disasters, and observations can help to improve the current code. The
information gathered need to be reflected during calibration of a new code. Bayesian
Theory, basically updating the present data by available new information is the keystone
of the calibration.


28

Figure 1 Algorithm for current calibration practice of a structural design code
(Melchers (1999))



29



Figure 2 The load and resistance factors for a predefined target safety level
(Ellingwood et al. (1980))


30
A design code checks the safety through the limit state equations. Failure modes,
such as flexure in a beam or plate buckling, etc. have to be defined and necessary
analytical and empirical equations are used to rationalize the safety check. A LRFD code
contains the limit state check equations, nominal loads and resistance values and load and
resistance factors.
In current calibration practice, if present applications give successful results then
the safety index obtained with the present code can be prescribed as the target safety
levels for the new code. In this study instead of obtaining just safety levels, the failure
cost is also deduced from the current code. This value is used to express total cost, which
is then minimized along the design space defined by the possible ranges of the random
variables. It is a common practice to represent the nominal loads as a ratio with respect to
the dead load to define the design space. For instance in ANSI A58 (Ellingwood et al.
(1980)) the loads such as live, wind and snow loads are represented by a factor of dead
load ranging between 0.2 and 5. Assume that the design requirement is in the form of
n W n L n D n
W L D R + + (2-3)
where R
n
, D
n
, L
n
and W
n
are the nominal values for the resistance, dead, live, and wind
loads respectively and s are the corresponding load factors. Both sides of the Equation
(2-3) are divided by nominal dead load D
n
and that gives
n
n
W
n
n
L
n
n
D
n
n
D
W
D
L
D
D
D
R

(2-4)
So the design equation is normalized by the nominal dead load. Therefore, instead of
dealing with actual nominal values the ratios of resistance and loads to dead load are


31
used. It is useful to use the normalized form (Equation (2-4)) to generalize the calibration
process rather than working with a single case. So Equation (2-4) can be rewritten as
n W n L n D n
W L D R + + (2-5)
where the nominal value for the dead load, D
n
is equal to unity and the nominal values for
the other loads and the resistance is a factor of 1.0. The same procedure for limit state
function; for the design criteria given in Equation (2-3), limit state function, g is
' ' ' ' ' L W D R g = (2-6)
where prime represent the non-normalized random variables. Dividing and multiplying
random variables with their nominal values and normalizing with dead load nominal
value, D
n
gives
|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|
=
n
n
n n
n
n n
n
n n
n
n n
D
L
L
L
D
W
W
W
D
D
D
D
D
R
R
R
D
g
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
(2-7)
n n n n
L L W W D D R R g = (2-8)
where R, D, W and L are the normalized random variables. Since the original random
variables are normalized by their nominal values the mean for these random variables are
equal to the bias of these random variables. Also, nominal values R
n
, W
n
and Ln are
normalized with respect to D
n
values so they are equal to a factor of D
n
.
The normalization eases the consideration of a range for the possible nominal
values. For instance in case of highway bridge specifications the ranges of nominal loads
such as L, W etc. to the nominal dead loads represents the different spans for the highway


32
bridges; a low L
n
/D
n
or W
n
/D
n
ratio represent that the bridge has a long span and dead
load dominates the design equation (2-8).
The approach followed herein requires the optimization of the total cost, which is
given in Equation (2-9)
{ }

+ =
space Design Space Design
f F I T
P C C C (2-9)
where, C
T
, C
I
and C
F
are total, initial and failure costs and P
f
is the failure probability for
a single realization of the design space. One has to realize when optimization of a code is
concerned, the optimization should cover the full design space; therefore, selecting load
and resistance factor for an optimum code may not give the optimum design for every
individual design problem covered in a code. Involvement of probability of failure makes
the minimization a Reliability Based Cost Optimization; because the failure associated
cost is assessed as the multiplication of the cost of consequences, C
F
due to a failure
times the probability of a failure to occur. The optimum solution is as shown in Figure 3.
The optimum solution is at the point where the slopes of the initial cost and cost of failure
are equal and opposite, and the corresponding safety index is called the optimum safety
level.






33

C
I
C
T
C
F
Cost

opt.
C
T(min)


Figure 3 The safety index vs. cost


2.4 Simulation Methods
Simulation is a tool for especially complex reliability problems; however, it is
mostly preferred to confirm the results obtained out of FORM or SORM solutions. But
when the limit state function is highly non-linear it may be necessary to use a simulation
technique to compute failure probabilities. In this study simulation is used only for
confirming the outcome of the conditional probability calculations performed for the
Ferry-Borges Method.


34
The simulation methods are based on realizing random numbers for random
variables by using the random number generators and testing those with limit state
function, G such as Equation (A-6). When G is calculated repeatedly from enough points
the probability distribution of the G can be obtained. It is important to keep generated
numbers independent. The probability of failure is defined as the P(G<0), therefore the
number of failure cases is f out of n independent trials. Then the probability of failure can
be represented by
n
f
P
n
f

= lim (2-10)
As the number of trials increase the results get better. This method is named as
Direct Monte Carlo Method, which is also called crude Monte Carlo Method. The
number of necessary trials increases as the number of random variables increases. In
order to reduce the number of trials required, methods such as Importance Sampling,
Adaptive Sampling, Antithetic Variates Methods etc. are reported (Melchers (1999)).

2.5 Load Combination
Structures are subject to different types of loads during their lifetime. In design
phase, the structure should be designed for a prediction of the future loads. Loads such as
wind, earthquake, live load etc. show a wide variation along the time in addition to their
randomness in nature. In a combination case when only one time variant load is involved
it is necessary to consider the lifetime maximum load effect for that load. But when more
than one time variant load is involved in a load combination than it is not accurate to


35
consider all these loads at their lifetime maximum. Since the simultaneous occurrence of
the time variant loads at their maximum is rare, the total maximum effect of the
combined loads is not the sum of the maximum load effects of these time variant loads.
For example, there is a very low chance to have a strong earthquake along with extreme
high wind speeds. In early design codes load combinations were taken care of by
applying a reduction factor to the sum of individual design load effects (e.g. in ACI
codes 1970 edition a factor of 70% for combination of live load and wind (or
earthquake) loads), which was based on experience and intuition Wen(1977). Three
methods are currently used in load combination problems, Turkstras Rule, Ferry-Borges
Model and Wens Load Coincidence Method.
2.5.1 Turkstras Rule
Turkstras Rule is a deterministic approach to solve a Load Combination problem,
and has been used extensively in code developments because of its simplicity in
application. Since the probability of occurrence of time variant loads at their maximum
simultaneously is so small Turkstra (1972) proposed to approximate the combined effect
by assuming one of the loads is at its maximum and the rest are at their arbitrary point in
time values. A set of sub-combinations is created and the maximum of this set is the
maximum effect of the combined loads. Assuming the total effect Z(t) at time t is
( ) ( ) ( ) ( ) t X t X t X t Z
n
L + + =
2 1
(2-11)
then the maximum value of Z can be found by


36

+ +
+ +
+ +
=
max 2 1
max 2 1
2 max 1
max max
n
n
n
X X X
X X X
X X X
Z
L
M
L
L
(2-12)
where X
i
represents the instantaneous, and X
maxi
represent the maximum values of load
variable i. Consider the combination of the dead load, live load and wind load; dead load
does not show variation throughout the design life of a structure so assuming time
independent dead load is reasonable. On the other hand live load and wind load shows
variations during the life of a structure. So the combined effect of the above-mentioned
loads would be the maximum of dead load effect combined with the maximum of
combination of live load at its maximum with point in time value of wind load or wind
load at its maximum with point in time value of live load. The combination of these loads
can be represented as follows;

+
+
+ =
+ + =
max
max
max max
) ( ) ( ) (
W L
W L
D Z
t W t L D t Z
apt
apt
(2-13)
where D, L and W represent dead, live and wind loads, respectively and subscript apt is
the arbitrary point in time value for the corresponding load.
Turkstras rule has been widely used in code calibration processes; in most of
todays codes this rule has been the load combination technique (ANSI-A58, AISC,
AASHTO, etc.). Turkstras rule has an important drawback that avoids using this rule in
this research. Turkstras rule does not allow introducing the modeling uncertainties into
the picture because it absolutely depends on the independence of the loads repetitions, but


37
the modeling makes the load effects correlated. Although Turkstras rule is easy to
apply, it is difficult to estimate the accuracy of the method. It may be conservative or
unconservative depending on the number of the repetitions of each individual load case
and relative uncertainty of the magnitude of each load.
2.5.2 Ferry-Borges Process
Ferry-Borges Method models the load process as rectangular pulses that change
after prescribed, equal and deterministic intervals (Turkstra and Madsen (1980)). It is
assumed that load intensities are independent and constant during the interval. The
lifetime of the structure is an integer multiple of the load intervals, and also the individual
loads intervals are integer factors of each other as shown in Figure 4.
In Figure 4 the t
1
, t
2
, t
3
are the single event duration times and T is the lifetime of
the structure. The combination of X
1
(t), X
2
(t) and X
3
(t) is given as in Equation (2-11). The
maximum effect of Z(t) can be written as;
( ) ( ) ( ) ( ) { }
)
`

)
`

+ + = t X t X t X t Z
t t T T
3 2 1
3 2
max max max max (2-14)
The cumulative distribution CDF of the maximum effect of the combination of
the loads can be calculated using a convolution integral (Melchers (1999)). For example
the maximum CDF for ( ) ( ) { }
)
`

+ t X t X
t
3 2
3
max can be calculated by
( ) ( ) ( ) [ ] d x F x f x F
t
t
X
x
X
3
2
3 2
max
=


(2-15)


38
In the above equation the probability of the maximum of X
2
(t) + X
3
(t) in t
2
to be
less than x is calculated. ( ) [ ]
3
2
3
t
t
X
x F gives the probability of maximum in t
2
/t
3

repetitions of the load event is less than (x-), and ) (
2
x f
X
is the probability density
function of X
2
, and the multiplication is integrated from to x.
The calculation of Equation (2-15) becomes time consuming as the number of
variables increases. Instead of calculating the convolution integral by numerical
integration, FORM theory can be used to find the result. The calculation of probability of
failure, P
f
, algorithm developed by Rackwitz and Fiessler (1978) based on FORM will be
used. The effects of the load repetitions are taken into account during the conversion of
non-normal variates to normal variates. For instance the CDF of X
3
in time span t
2
is
( ) ( ) [ ]
3
2
3 2
*
3 ) at t (
*
3
t
t
X
x F x CDF = (2-16)
where
*
i
x represent the design point or most probable point which can be defined as the
point at the limit state function with the highest probability density (Melchers (1999)). At
time t
2
the combined effect of X
2
and X
3
has a CDF of
( ) ( ) ( ) [ ]
3
2
3 2 2
*
3 2
*
) at t (
*
3
*
2
t
t
X X
x F x F x x CDF + = + (2-17)
and at t
1
the CDF of the combined effect of X
2
and X
3
is
( ) ( ) ( ) [ ]
2
1
3
2
3 2 1
*
3
*
2 ) at t (
*
3
*
2
t
t
t
t
X X
x F x F x x CDF
)
`

+ = + (2-18)
the combined effect of X
1
, X
2
, and X
3
at t
1
is then


39
( ) ( ) ( ) ( ) [ ]

)
`

+ + = + +
2
1
3
2
3 2 1 1
*
3
*
2
*
1 ) at t (
*
3
*
2
*
1
t
t
t
t
X X X
x F x F x F x x x CDF (2-19)
finally the combined effect of X
1
, X
2
, and X
3
in lifetime T has a CDF of
( ) ( ) ( ) ( ) [ ]
n
t
t
t
t
X X X
x F x F x F x x x CDF

)
`

+ + = + +
2
1
3
2
3 2 1
*
3
*
2
*
1 ) T at (
*
3
*
2
*
1
(2-20)
The equivalent normal mean and standard deviation is found according to the CDF
values calculated by Equations 2-16 to 2-20.
The Ferry-Borges Method is still a simplification because the time durations for
every load pulse is assumed to be constant and without fluctuations during the duration.
But it is more accurate compared to Turkstras Rule, because in Turkstras Rule time
duration and rate of occurrence are not considered at all. Ferry-Borges Method is suitable
to handle the separated time dependent and independent parts of the load effects.
Therefore it is preferred to use this method in implementation of the proposed cost model
in this study.








40



t
3
t
3
t
3
t
3
t
3
t
3
t
3
t
3
t
3
t
3
t
3
t
3
t
2
t
2
t
2
t
2
t
2
t
2
t
1
t
1
t
1
X
3
(t)
X
2
(t)
X
1
(t)
t
t
t
t
2
/ t
3
=2
t
1
/ t
2
=2 T / t
1
=n

Figure 4 Ferry-Borges Model




41
2.5.3 Wens Load Coincidence Method
This model is proposed by Wen (1977), to calculate the probability of failure
when combination of two or more time dependent loads are concerned. The probability of
failure can be calculated by using (Wen (1981))
( )

(
(

|
|
.
|

\
|
+ +

= = =
T p p T P
n
i
n
i
n
j
ij ij i i
1 1 1
exp exp 1 L (2-21)
where P is the probability of failure for a duration of time T, n is the number of loads
involved in combination,
i
is the rate of occurrence for load i, p
i
is the probability of
load i exceeding the limit state when considered alone and
ij
and p
ij
are the rate of
occurrence and probability of failure respectively, given that loads i and j act
simultaneously. This method is a more sophisticated approach to load combination than
the other two methods, but more mathematical work is involved with this method. The
previous two methods are preferred to Wens Load Coincidence Method because of
simplicity in application. This is important since the calculations for cost analysis should
be done for every year throughout the design life. Also, Wens Load Coincidence Method
assumes independence between the events, but modeling uncertainties make events
dependent. In order to consider correlation, this method has to be derived using non-
Poisson events. Since the Ferry-Borges Method can be modified to consider correlation
between events in this research, the Ferry-Borges Method is preferred for reliability
analysis.



42
2.6 Modifications to Ferry-Borges Model
2.6.1 Mixed Type Distributions
The Ferry-Borges Model is used to represent the time dependent load effects such
as wind, earthquake etc. and the probability of not having an event is generally not zero.
The mixed type distributions are used to represent the event occurrence and not-event
occurrence together (see Figure 5).
T
X
f
X
f
X*
offstate
p
on state
q=1-p

Figure 5 Mixed Type Distribution

A mixed type distribution has a non-zero probability of not occurring, such as
wind load or earthquake loads. The probability of X<x is the multiplication of probability
of having an event and probability that event being less than x. The Ferry-Borges Method
is also capable of dealing with the mixed type distributions, but it is reported slow
convergence or even non-convergence (Rackwitz & Fiessler (1978)). If the load events
follow Poissonian model, then only on state case is considered and the possible number


43
of events can be calculated using the arrival rate. For X the probability of exceeding a
value x in time t may be written as
( )
t
X
e x F t T x X P
t

= = < > 1 1 ) ; ( (2-22)
where is the mean rate of occurrence of events per unit time. For extreme values of x,
t
e

is close to 1 and t consists of m repetitions, and Equation (2-24) can be approximated
by
( ) ( )
m
X
m t
x F e t T x X P = = < >

1 ) 1 ( 1 1 ) ; (

(2-23)
so, the probability of X being less than x is
( ) ( ) ( )
m
X X
x F x F t T x X P
t
= < < ) ; ( (2-24)
For example the duration of a windstorm is assumed to be 4 hour and according to Belk
and Bennett (1991), storms occur at an average rate of 0.0176 each hour (152 storms
every year). So, in order to find the maximum CDF in a year, the cumulative distribution
for a storm event is raised to a power of 152.
( ) ( ) ( )
152
1 1
) 1 , ( x F x F year T x X P
event year
X X
= < < (2-25)
In combination of wind load with live load, then the number of live load events in
a 4 hr. period has to be considered. The live load event distribution is a mixed type
distribution as given in Figure 5. Consider a bridge example, the number of trucks in a
day is estimated to be 1000 (Moses (2000)); in 4 hr. that would be 167 (1000/6) truck
loadings. Since for the live load the number of events is known and these events are
Poissonian and assumed independent from each other, a 4 hour distribution can be
calculated by raising the single event distribution to the number of events in 4 hours.


44
Doing so helps to eliminate the consideration of the mixed type directly in the Ferry-
Borges Method and adverse effect of mixed type distribution on the convergence rate of
FORM calculations while using Rackwitz-Fiessler Algorithm.
2.6.2 Separation of Time Variant and Invariant Random Variables
The correlation between the individual load effects due to modeling uncertainties
may produce conservative results. For example, consider wind load effects. The velocity
of the maximum wind used to model the load effects is independent from storm to storm,
but the modeling uncertainties that are not time variant are the same for every storm and
this makes the load effects correlated. The mathematical model tries to duplicate
realizations of the real life, but the model most of the time is not perfect, or is an
empirical one, so the real life situation may be different than the mathematical model
realization. For instance, in predicting earthquake or wind effects on a structure the
ground motion and wind speed data are used, respectively. This prediction depends on
the mathematical model used to convert, for example, the wind speed to pressure on a
structure. It is hard to claim that this conversion fully replicates the real life. So, predicted
and observed show a variation, and this variation is best represented by modeling
uncertainties introduced to the reliability analysis. The modeling uncertainties do not
show variance in time. Since the Ferry-Borges Model depends on the independency of
the single load realizations the time dependent r.v.s are separated into time dependent
and independent parts and only the latter are used in the evaluation of the convolution
integral in Equation (2-15).


45
The Ferry-Borges Model can be used without modification in a case when there is
only one time dependent load in a load combination case. When the number of time
dependent loads is two or more, conditional probability is used to evaluate the probability
of failure. Assume a load combination having three random variables Y
1
and Y
2
, where Y
1

and Y
2
are time dependent r.v.s. Since both Y
1
and Y
2
consist of time independent
modeling uncertainties these r.vs are separated as follows;
2 2 2
1 1 1
Q X Y
Q X Y
=
=
(2-26)
where X={X
1
,X
2
} is the time independent and Q={Q
1
,Q
2
} is the time dependent parts of
Y. The conditional probability of failure ( ) X Q P
f
can be written for a limit state function
G(X,Q) as;
( ) ( )


=
0 ) , ( x q G
f
dq x q f P
X Q
X Q (2-27)
where ( ) x q f
X Q
is the joint distribution of Q, given X. Equation (2-27) can also be
written in the form of a convolution integral given in Equation (2-15). The Ferry-Borges
Method is capable of solving Equation (2-27) for a realization of X, but the unconditional
probability of failure is required for reliability analysis. The unconditional probability of
failure can be evaluated by using expectation. The probability of failure is
( ) [ ] ( ) ( )dx x f P P E P
f f f X
Y
X Q X Q

= = (2-28)
where f
X
(x) is the joint distribution function of X={X
1
, X
2
}. Numerical Integration or
Nested Reliability Analysis can be used to evaluate the integration in Equation (2-28).


46
2.6.3 Evaluation of P
f
by Numerical Integration or Nested Reliability Analysis
Evaluation of expected probability of failure given by Equation (2-28) can be
performed by using Numerical Integration or Nested Reliability Analysis. Both of these
approaches is explained in this section and a comparison of the results (safety index, )
and computing time required for them are provided for selecting the most feasible
method for analysis.
Numerical Integrations requires performing reliability analysis for realizations of
X. It is important to decide on how many discrete realizations are required to evaluate the
Equation (2-28) accurately. It is desirable to use the minimum number of discrete points
to evaluate numerical integration. In cost analysis the reliability calculations are
performed for each year in a design life to find the annual probability and annual cost of
failure. So, for a design life of 75 years the numerical integration is evaluated for 75
times, and also to bear in mind, when a code is minimized the 75 numerical integration
needs to be carried for many design points (such as different nominal live load to nominal
dead load or nominal wind load to nominal dead load ratios).
Another issue to be considered in numerical integration is the boundaries for the
integration. Assuming the modeling uncertainties follow a Normal Distribution first
including a range {mean4*standard deviation to mean+4*standard deviation}, then
{mean3*standard deviation to mean+3*standard deviation} are investigated. First
boundary range covers about the 99.994 % of probable values, and second case covers
about 99.97 %.


47
In order to decide on the number of integration points a sample reliability
analyses is run and sensitivity to number of integration points for different integration
boundaries is investigated. Let us consider the combination of dead load, live load and
wind load case. The limit state function is in the form
W L D R G = (2-29)
where R, D, L and W are resistance, dead load, live load and wind load, respectively.
Resistance, R and Dead Load, D are time independent, on the other hand Live Load, L
and Wind Load, W are time-dependent. However, both L and W contain modeling
uncertainties, which are time independent. In order to increase accuracy in reliability
analysis the modeling uncertainties need to be separated. So, after separation Equation (2-
34) can be written as;
W W L L
Q X Q X D R G = (2-30)
where Q={Q
L
, Q
W
} is the time dependent part and X={X
L
, X
W
} is the time independent
modeling uncertainties. The following statistical data is assumed for analysis








48
Table 4 The statistical data for sample analysis
R.V. Distr. Type Bias COV
R Logn. 1.12 0.10
D Normal 1.03 0.08
Q
L
Logn. 0.50 0.16
X
L
Normal 1.00 0.12
Q
W
Extr. I 0.25 0.40
X
W
Normal 1.00 0.25


where bias is the mean to nominal ratio, and COV is the standard deviation to mean ratio,
Q
L
and Q
W
are 4 hr. data and the nominal values for these are expected maximum 75 and
50 year events, respectively. It is assumed that a windstorm lasts 4 hr. with 152
windstorms in a year. Therefore for a year t(Q
W
) / t(Q
L
) = 1, and n=(1 yr.) / t(Q
W
) =152,
and for 75 years n=75*152=11400.
Reliability analysis results using Ferry-Borges Method along with the numerical
integration for different number of discrete points for years {1 2 5 10 20 50 75} for a
range of {mean4*standard deviation to mean+4*standard deviation} are given in Table
5 and for same years beta values are given in Table 6 considering the range {mean
3*standard deviation to mean+3*standard deviation}. In both Tables N is the number of
discrete points used in integration.
The results for the case of N=11 points with range {mean4*standard deviation to
mean+4*standard deviation} tabulated in Table 5 are seem to be accurate enough to
perform reliability analysis.



49

Table 5 Comparison of number of discrete points used (range={-4 to +4})
Time(yrs.) N=41 N=31 N=21 N=13 N=11 N=9
1 3.28 3.28 3.28 3.28 3.27 3.31
2 3.10 3.10 3.10 3.10 3.09 3.13
5 2.86 2.86 2.86 2.86 2.85 2.90
10 2.68 2.68 2.68 2.68 2.67 2.72
20 2.50 2.50 2.50 2.51 2.49 2.55
50 2.27 2.27 2.27 2.27 2.25 2.32
75 2.17 2.17 2.17 2.17 2.15 2.22




Table 6 Comparison of number of discrete points used (range={-3 to +3})
Time (yrs.) N=21 N=13 N=11 N=9 N=7 N=5
1 3.28 3.28 3.28 3.28 3.26 3.35
2 3.10 3.10 3.11 3.11 3.08 3.17
5 2.87 2.87 2.87 2.87 2.84 2.94
10 2.69 2.69 2.69 2.70 2.66 2.76
20 2.51 2.51 2.51 2.52 2.48 2.58
50 2.27 2.27 2.27 2.29 2.23 2.35
75 2.17 2.17 2.17 2.19 2.13 2.25




50
An observation made during the analysis for deciding what should be the
minimum discrete numbers used for accurate results also helped to reduce the
computation time in the reliability analysis. A logarithmic decay in beta values is
observed during the lifetime, so instead of repeating the analysis 75 times a logarithmic
approximation is made to yearly safety index values as discussed in Section 1.6.3.
Another way of solving Equation (2-28) is using Nested Reliability Analysis. In
order to solve the conditional probability problem let us define a limit state function, h
( ) X + = u h (2-31)
where u is a standard normal variate and (X)is the safety index given the X. The
solution for limit state, h is same as Equation (2-28), Der Kiureghian (1990) shows that
these two solutions are identical as
( )
( ) ( ) ( )
( ) [ ]
( ) ( ) ( )dx x f P dx x f du u f dudx x f u f
f
P
u
u
u
f
X
X
X
X
X Q
X
X Q
X Q

=
(
(

+
1
0
(2-32)
In nested reliability analysis, first the inner reliability problem defined ( ) X Q
f
P is
solved and then outer reliability problem defined by limit state function, h is solved. The
inner reliability problem is time dependent; therefore Ferry-Borges Method is used to
evaluate it. For outer reliability problem a regular FORM analysis can be used. The
evaluation process involves calculating inner time-dependent reliability problem for
every design points (star values). The inner reliability problem is used as a function in the
outer reliability problem limit state, h defined in Equation (2-31). Therefore in order to
evaluate the derivatives of limit state function, h at design points, probability of failure


51
for inner reliability problem has to be calculated at design points for X. Convergence is
reached quickly and therefore this method requires much less computing time compared
to Numerical integration. Next, a comparison for two approaches is provided to indicate
the advantages of Nested Reliability Analysis (NRA).
Let us assume the data provided in
Table 4 for analysis of dead, wind and live load combination. A personal
computer with a P-III 733 cpu is used to run the analysis. In Table 7 both safety index for
first year and lifetime are provided along with the cpu-time used to calcuate. The load
factors used are the code defined 1.25, 0.40 and 1.35 for dead, wind and live load
respectively. The ratios between the cpu-time for Numerical integration and NRA are
between 4.6 to 8.1 to for
1
and 5.6 to 8.9 for
75
. It is important to realize that
optimization analysis require repeated calculation for design sample spaces for different
combinations of load factors. So the differences in cpu-time make a quite impact on total
time of the calculations. Table 7 shows the cpu-time advantage of Nested Reliability
Analysis to Numerical Integration clearly. In this research Reliability Analysis preferred
to Numerical Integration to evaluate expected probability of failure and corresponding
safety indices.






52


Table 7 Cpu-time for NRA and Numerical Integration
Wn/Dn Ln/Dn

1
cputime
75
cputime
1
cputime
75
cputime
0.50 0.50 3.05 1.90 2.18 1.70 3.06 9.97 2.20 11.79
0.50 1.00 3.99 2.03 3.22 1.70 3.99 10.83 3.23 13.49
0.50 2.00 4.66 2.23 4.05 1.83 4.69 12.79 4.07 16.30
1.00 0.50 2.06 2.26 0.91 1.69 2.09 10.38 0.93 13.09
1.00 1.00 3.00 1.62 1.92 2.61 3.02 11.37 1.94 14.56
1.00 2.00 4.13 1.74 3.22 2.46 4.14 12.70 3.24 17.50
2.00 0.50 1.02 1.44 -0.29 2.62 1.04 10.36 -0.26 15.25
2.00 1.00 1.79 1.48 0.49 2.33 1.82 11.96 0.52 16.58
NRA Numerical Integration


2.6.4 Logarithmic Approximation to Yearly Safety Indices
The observation made during the analysis in previous section helps to reduce the
computation time considerably. Since the cost analysis require annual probability of
failure, probability of failure in every year has to be calculated, such as probability of
failure in 5 years, 10 years or 75 years. In order to make the analysis as effective as
possible, the computation time has to be reasonable, and any accurate approximations are
used when possible. Considering the data in
Table 4, the yearly safety indices are approximated by two points along the design
life. The results to these approximations are given in Figure 6 and Table 8. All four
approximations gives good close results to Numerical Integration results by using two
points. The results using 1 and 75 years seem to be closest to numerical integration
results. Therefore in reliability analysis instead of finding probability of failure for every


53
year using numerical integration, only probability of failure in 1 year and 75 years is
calculated by using numerical integration and the rest will be approximated by a
logarithmic fit. The fit will be in the form of
( ) ( ) 1 ) log( * + = t a t (2-33)
where a can be calculated determined by
( ) ( )
( ) 75 log
1 75
= a (2-34)


Figure 6 Approximation to safety index,




54


Table 8 Logarithmic Approximations to safety index, beta along the lifetime
Time(yrs.) FORM+N.I. 1&2 1&5 1&10 1&75
1 3.275 3.275 3.275 3.275 3.275
2 3.099 3.099 3.097 3.097 3.097
5 2.863 2.866 2.863 2.861 2.862
10 2.683 2.670 2.685 2.683 2.683
20 2.504 2.513 2.507 2.505 2.505
50 2.269 2.280 2.272 2.269 2.270
75 2.166 2.177 2.169 2.165 2.166
Log Approximation using years 1&i
(i=2,5,10,75)


The logarithmic approximation to safety indices by using 1 and 75 years safety
indices values cut the computation time considerably, and helps to increase optimization
efficiency.

2.7 Sensitivity Analysis
In code calibration the load factors obtained should not to be so sensitive to the
assumed statistical data that it may be difficult to verify. Therefore the results of the
calibration should be checked for possible changes.


55
The easiest way of running sensitivity analysis is perturbing the variables about
the values used and obtaining the behavioral change of the results, which are the
optimized load factors and safety index.
In this study the behavior changes to different variables is studied. The sensitivity
to COV R, D, L and W are investigated. Also, the effect of changes in failure cost ratio
(failure cost to current code cost ratio) which is deduced from existing code using the
simplest load combination, and discount rate j used in the analysis are studied for
sensitivity.

2.8 System Reliability
It is common practice to use element reliability in code calibration. Actual
structures are the combination of elements; therefore failure of an element does not
necessarily mean structure will fail. The systems can be classified as series systems and
parallel systems. A real structure is a combination of series and parallel systems.
In a series system, failure of an element is also the failure of the system. On the
other hand, failure of one element may not cause a parallel system to fail. All the
members in the parallel system must fail for a complete system failure. The load
redistribution takes place after an element fails, and remaining elements carry the load
(See Figure 7).
It is reasonable to comment that in real structures actual safety levels may be
higher than calculated. The P
f
for series system can be written as


56
fi
n
i
fsys
P P
1 =
= U (2-35)
and for a parallel system,
fi
n
i
fsys
P P
1 =
= I (2-36)
Unfortunately when a code is considered it is not easy to formulate the system
reliability of structures because of the wide variety available in design possibilities. A
rational way of introducing system safety into code calibration was also considered in
this study. For example, the redundancy of a bridge structure after failure of an element
plays a key role in safety of a bridge. The simplest way to deal with that in bridge system
is to assume a relationship for the system safety index such as (Nowak (1993))
1 + =
element system
(2-37)
It is assumed that for all the load combination cases considered in this study
Equation (2-37) holds. Note that in the load combination case including earthquake load,
the load effect model used for earthquake analysis already accounts for system safety.








57



S
R
1
R
2 R
3
Parallel System
R
1
R
2
R
3
S
Series System

Figure 7 Parallel and Series Systems



58
3.0 STRUCTURAL OPTIMIZATION

In this chapter structural optimization is discussed. It is not the intention to give a
full description on optimization methods, but rather to give a brief explanation of the
optimization method followed in this study.
The optimization process can be divided into two sections mainly; Linear and
Non-linear programming. The iterative nature of the FORM analysis makes the
optimization problem in this study a non-linear one. In the next sections general
optimization problem formulation and the solution techniques to the non-linear problem
is discussed.

3.1 General Optimization Formulation
An optimization problem consists of design variables, objective function, and
constraints to variables. Design variables (the load factors in this study) are the main
unknowns so that changes affect the outcome of a design such as cost or weight. An
optimal solution is the set of the variables that gives the optimum (e.g. minimum cost,
minimum weight or maximum utility) outcome. The general optimization problem can be
formulated as


59
( )
u l
e i
e i
x x x
n n i x c
n i x c st
x C

+ =
= =
, , 1 0
, 2 , 1 0 ) ( .
) ( min
L
L
(3-1)
where C(x) is the objective function that has to be optimized. The c
i
(x) is the constraints
that the optimization problem is subjected. The first set of constraints (i=1,,n
e
) are the
equality constraints, where c
i
gives a scalar quantity about x, and the second set of
constraints (i=n
e
+1,,n) are the inequality constraints. Some of the constraints may be
active and some inactive for the solution. The upper and lower bounds (x
l
, x
u
) for the
design variables helps to reduce the design space where the optimal solution is sought.
The problem is called Linear Programming when the objective function and the
constraints are both linear. The case when both the objective and constraints are non-
linear, which is called Nonlinear Programming is the problem that is harder to solve. This
is solved by an iterative procedure, and in every step simpler sub-problems are solved.

3.2 Solving Nonlinear Programming Problem
In solving the NP problem the Kuhn-Tucker equations are considered. If the
problem is a convex problem the Kuhn-Tucker optimality conditions are necessary and
sufficient for an optimal solution. For the Equation (3-1), define the Lagrangian function
for the problem,
( ) ( ) ( ) ( )
e
e
e
n j j
n
n j
j j
n
j
j
S x c x c x C s x

+ = =
+ + + =

1 1
) , , ( (3-2)


60
where
j
are the Lagrange multipliers, and S are the slack variables. There is a set of
Lagrange Multipliers that makes the Lagrangian stationary with respect to x
j
,
j
.
n j for
x
c
x
c
x
C
x
j
i
n
n i
i
j
i
n
i
i
j j e
e
L 1 0
1
*
1
*
= =



+ = =
(3-3)
( ) ( )
e i i
n i for x c x c
i
L 1 ; 0 , 0
* *
= = = (3-4)
( ) ( ) n n i for x c x c
e i i
j
i i
, , 1 ; 0 , 0 , 0
* * *
L + = = (3-5)
The solution to Kuhn-Tucker equations is used in most non-linear programming
algorithms. Next a solution technique called Sequential Quadratic Programming (SQP) is
discussed.

3.3 Sequential Quadratic Programming
Sequential Quadratic Programming is a very powerful solution technique, and it is
high performance solution that has been tested by Schittowski (1985) for various types of
problem. The method shows a better performance compared to other techniques.
The method based on the reducing the problem into sub-problems that can be
solved using Quadratic Programming Solution. It starts with updating the Hessian Matrix
of the Lagrangian function (see Equation 3.6), and then solving quadratic programming
problem and than line search for the optimum solution.
The detailed information on the solution technique can be found in many
textbooks, see e.g. Arora (1989), Kirsch (1993).


61
(
(
(
(
(
(
(

n n n
n
x x x x
x x x x
H
2
2
2
1
2
1 1
2
L
M O M
L
(3-6)
In analysis the software MATLAB is used. This software allows creating
subroutines that can use built in functions to run analysis. In the optimization part a built
in function based on the Sequential Quadratic Programming available with the
optimization toolbox of the MATLAB (1999) is used.

3.4 Optimization Using MATLAB
MATLAB consists of many built in functions that allow user to apply directly
without writing lines of codes. The built in function used in this research for the
optimization is fmincon, which is capable of solving constrained nonlinear
multivariable function. This function is very easy to use and allows the objective function
to be in an iterative function, such as a FORM analysis.
The fmincom function can be used to formulate the general problem given in
Equation (3-1). Let us redefine Equation (3-1).
ub x lb
beq x Aeq
b x A
x ceq
x c t s x C

.
.
0 ) (
0 ) ( . ) ( min
(3-7)


62
where C(x) is a function that returns a scalar, c(x) and ceq(x) is the functions that returns
vectors and all three can be nonlinear in nature, A, Aeq are matrices and b, beq, lb, ub are
vectors. The problem can be formulized in MATLAB by using the following syntax.
[ ] ( ) nonlcon ub lb beq Aeq b A x fun output exitflag fval x , , , , , , , 0 , fmincon , , , =
where x is the vector of the design variables that gives the optimum solution, fval the
value of the minimum C(x), exitflag is the indicator of the convergence (if greater than 0
the function is converged to a solution, if equals to 1 than the maximum iteration number
is reached and if less than zero, function did not converge to a solution), output (contains
the information about the number of iterations and number of function evaluations are
involved, etc.), the fun is the string that is the name of the user defined function, that
is used as the objective function, and x0 is the starting vector for the design variables.
In this study the objective function that is the sum of total cost functions contains
implicit functions such as probability of failure. The probability of failure is calculated
throughout the FORM analysis, which is an iterative function and it has to be calculated
every year separately as time becomes an important issue. Because the function has to be
evaluated may times, even for one set of design variables (such as load factors in this
research) the function has to be evaluated for different parameters in the design space
(different W
n
/D
n
, L
n
/D
n
values) and summed over that space. It is this sum that has to be
minimized as discussed in next chapter.


63
4.0 RELIABILITY BASED STRUCTURAL OPTIMIZATION AND
APPLICATION IN CODE CALIBRATION

How safe is safe enough? This question has to be answered in order to spend a
reasonable amount in construction. The cost and safety of a structure is highly correlated;
there is a trade-off between these two. It is not feasible to build structures that are much
safer than they should be. Although people always want to feel safe with their structures
but increasing safety comes with high price tag. The simplest way of finding the
minimum cost of a structure is applying deterministic optimization. But the resistance
and loads are random in nature; therefore the deterministic approaches are not sufficient
to find an optimum solution. The safety also has to be an active part in optimization
process, and a balance between cost and safety has to be attained. The Reliability Based
Structural Optimization (RBSO) is the right tool to use both cost and safety together to
find the actual optimum, because it is a fact that if a structure fails than consequences has
to be also paid. Therefore when the cost is considered possible failure scenarios also need
to be thought and the expected cost (failure cost times the probability of failure) due to
these failures need to be added on top of initial cost. Therefore the expected total cost has
to be minimized rather than just the construction cost. But unfortunately it is not that easy
to consider the expected failure cost, because that type of cost brings a lot of discussion
and argument with it.
Reliability Based Cost Optimization, RBSO can also be used in calibration of
structural design codes, because in a calibration process a balance between safety and


64
cost should be sought implicitly or explicitly. Since RBSO seeks that balance rationally,
it can be a valuable asset of the calibration process. In current design code calibration
practice the safety levels are determined by target safety indices. For instance, in
calibration of AASHTO Specifications by Nowak a target safety index of 3.5 is used for
combination of dead and live (truck) load. A Safety index of 3.5 was the current codes
average safety level, and aimed as the calibrations target. This target was selected using
the current codes. Design samples are the discrete points along the probable design
configurations. For instance in bridge engineering practice, bridges have different spans.
In order to represent the range a code covers, representative discrete samples are used in
the calibration. Different nominal live (truck) load to dead load ratio is used to
distinguish between short and long span bridges. The lower values of nominal live to
dead load ratio are for the long span bridges where the self-weight of the bridge
dominates, and higher values for shorter span bridges. Any economic aspects are taken
into consideration through an intuitive approach. For example, in ANSI A58 where a
target safety index of 3.0 is used for D and D+L+S, and a target safety index of 2.5 for
D+L+W and 1.75 for D+L+E cases (D L, S, W, and E stands for dead, live, snow, wind
and earthquake loads, respectively). Lower safety index values for last two load
combinations is due to fact that having higher values would significantly increase the cost
of the structure, therefore lower target safety index values are adapted for these cases. On
the other hand, connections where increasing the target is not reflected in the cost of
structure significantly, a higher value (about 4.0) is adapted. Although in calibration of


65
ANSI A.58 economy is not considered explicitly but during the decision of target levels
the economy is addressed intuitively.
The soundest way of dealing with a calibration processes should be rationalizing the
economy and safety aspect together. Rosenblueth (1976) stated that optimum reliability
will be the basis for future codes, and it is clear that trend is directed to that direction. In
this research an approach that allows using RBSO in calibration process consistently is
studied. The formulation of the cost and safety is given in proceeding sections.

4.1 Formulation of the Expected Total Cost
The total cost contains initial construction cost, maintenance cost and expected
failure cost. Maintenance cost is not included herein in order to keep problem
manageable. Total cost, C
T
can be formulated as;

=
+ =
T
i
ij
fi
f I T
e
P
C C C
1
(4-1)
where C
I
is the initial cost of the structure, C
f
is the failure cost of the structure and
assumed constant for every year throughout the design life, T, P
fi
is the probability of
failure that a structure fails in i
th
year, e
ij
is the discount factor to bring the failure cost
from i
th
year to present worth and j is the discount rate (real rate=nominal inflation).
The discount rates in developed countries are between 2% to 4%, and FEMA 227 (1992)
assumes a discount rate about 3% for public works. In this study a rate equal to 3%
throughout the design life is selected to represent the trend. It is important to add the cost


66
in the same basis to make the cost comparable. Therefore, the future worth are discounted
to present worth for making a decision.
In a code format the design is controlled by the limit state functions imposed by
the code as described in Chapter 2. For example in case of LRFD the limit state for
combined gravity and wind effects is:
R
n

D
D
n
+
L
L
n
+
W
W
n

where is the resistance factor, R
n
is the nominal resistance,
D
,
L
,
W
are the load factors
for dead, live and wind load and D
n
, L
n
and W
n
are the nominal dead, live and wind load
effects respectively. As discussed in Chapter 2 a code needs to cover a range of possible
of designs rather than a single design. Instead of using the real load values the loads are
normalized with respect to dead load and than reasonable ranges are used to replicate the
design space. Since the size, geometry, location and function of the structure vary the
initial cost and failure cost, these costs are normalized by a base cost, C
0
.
4.1.1 Initial Cost
The initial cost is mainly the construction cost of a structure; that is the cost
associated with the structural configuration and size. The initial cost, C
I
can be
represented by the following formula (Kanda and Ellingwood (1991)),
|
|
.
|

\
|
|
|
.
|

\
|
+ = 1 1
0
0
D
D
I
Q
Q
K C C (4-2)


67
C
o
is the base cost which is the construction cost of the structure associated with the
nominal reference load effects, Q
D0
and Q
D
is the calculated total design load effect. K is
the normalized cost factor. Nominal load effects, Q
D0
and Q
D
can be calculated by using
nominal loads, base load factors and applied load factors as follows;
{ }{ } { }{ }
n D n D
X Q X Q = =
0 0
(4-3)
where ,
0
are the load factor row vectors for reference and the applied case and X
n
is
the nominal load column vector. Equation (4-3) can be rewritten as
kn k n n D kn k n n D
X X X Q X X X Q
0 2 02 1 01 0 2 2 1 1
L L + + = + + =
(4-4)
where
1
,
2
are load factors such as
D
,
L
, and
W
,etc. ,
01
,
02
, are reference load
factors such as
D0
,
L0
,
W0
, etc., X
in
are the nominal load values such as D
n
, L
n
, W
n
, etc.
The cost effect of the type of the load to the structure may be different. The lateral loads
such as wind and earthquake create a higher impact on the cost than the gravitational
loads; marginal cost slope for environmental loads are higher than the gravitational loads
generally. Thus not all X
i
s contribute the same percentage to the construction cost; let us
assume that loads X
i
( m i L 1 = ) has a cost factor of K
i
, then Equation (4-2) can be
arranged as follows
( ) ( )
|
|
.
|

\
|
(

|
|
.
|

\
|
+
+
+ =
m mn n
m m mn m n
I
X X
X X
K C C
0 01 1
0 01 1 1 1
0
1


L
L
(4-5)
and normalizing the nominal loads with X
1
gives;
( ) ( )
|
|
.
|

\
|
(

|
|
.
|

\
|
+
+
+ =
m mn n
m m mn m n
I
K C C
0 01 1
0 01 1 1 1
0
1


L
L
(4-6)


68
where is the ratio of the ith load to first load. The cost factor, K and can be estimated
by studying the effect of the changing load factors in current code. There are studies
reported on the effect of load factors to marginal cost such as Lind (1976), Ferrito (1984),
Moses(1989), etc. The current code load factors can be used as the reference load factors
in calculation of initial cost. In this study the current load factors are used as reference
case, therefore the base cost, C
0
is the current code implied cost of a design.
4.1.2 Failure Cost
The expected failure cost, C
f
is indeed hard to estimate, especially when a code is
considered, because the failure consequences are not unique from one structure to
another. Therefore determining the cost of failure, C
f
is one of the key issues of the cost
minimization process discussed herein. This term has to include the economic cost of
failure, including intangible costs and consequential costs. The importance of a structure
and the costs of keeping that structure in service are also contained within the C
f
value.
The studies like Kanda and Ellingwood (1991), Kanda (1996), Ang and De Leon (1997),
Lee et al. (1998), Ellingwood (1999) have tried to quantify the failure cost by estimating
probable consequences. Especially for intangible costs most of these studies were based
on the data provided by Viscusi (1993). It is clearly not easy to quantify C
f
especially
when human loss is considered and the issue of quantifying C
f
becomes more of a
political issue rather than an engineering one. According to Linds postulate the current
code is already close to optimum Lind (1977), therefore the safety level in the current
code can be assumed at the optimum level. Ditlevsen (1997) suggests using some fixed


69
points from the current code to determine the cost of failure. Rackwitz (2000) also
emphasize on using Linds postulate in determining failure cost. Ditlevsen (1997) also
points out that given an optimal decision by the authorities, the intangible costs can be
backcalculated for a code format. But a procedure for obtaining failure costs from the
current code is not outlined.
In this research, instead of estimating the values of C
f
by using real cost values,
previous codes and designs and performance histories are used to deduce an implied
failure cost value and this value is used as a fixed point assuming the code is optimum. If
the current code is not an optimum that would also be reflected into our calibration
process but at least having the same failure cost for all load combinations maintains a
consistency. Further studying the sensitivity of the calibration to the deduced value of
failure cost would help to reach a better understanding of the calibration values. For
example, one load combination case such as the application of the dead load alone using
the AISC building design code, or the combination of dead and live loads acting together
using the AASHTO bridge specifications, can be used to deduce the C
f
value, and that C
f
value is consistently applied to all

the combinations.
Therefore, a current code with its predefined target safety level for a particular
combination is assumed to present a balanced cost with a sufficient level of safety. The
sum of total costs for design space is assumed to be the minimum by using the current
code values for load factors, which are also assumed to be the optimized load factors for
current code. The above assumption requires the derivative of sum of total costs with
respect to load factors should equal to zero. Then the deduction of C
f
is a matter of


70
solving the derivative equation for C
f
. The implied failure cost can then be deduced using
the total cost function and assuming it is minimized at the current codes load and
resistance factors. Deducing the C
f
value from the current code allows code writers to
reach a consistent code for all other load combinations and avoids the issue of assigning
monetary values to the cost of failure. The latter can be a politically intractable problem,
which would create grounds for many discussions and would not necessarily lead to an
acceptable solution.
Herein, the failure cost value deduced by backcalculation is introduced in the
optimization process of the rest of the load combination cases to be included in the design
code. Thus, the proposed approach maintains consistency in the design code by
considering the respective marginal initial cost slopes (K term in Equation. (4-2)). It is
important to note the safety index will not be uniform for load combination cases. In the
approach followed herein, it is not sought to reach uniform safety indices, but rather the
cost of the consequences of a failure would be the same for all load combinations.
4.1.3 Normalization of the Terms in Total Cost Function
As mentioned above a design code must cover a wide range of design
applications, therefore, the terms used in Equation (4-2) need to be generalized. The total
cost function (Equation (4-1)) is normalized by the base cost C
0
, and a total cost factor,
TCF is expressed as;


71
( ) ( )
1 & 1
1
1 1
1 0 01 1
0 01 1 1 1
0
= =
+
|
|
.
|

\
|
|
|
.
|

\
|
+
+
+ = =

=



n
i
ij
fai
m m
m m m m T
e
P
g K
C
C
TCF
L
L
(4-7)
0
C
C
g
F
= (4-8)
where the failure cost ratio, g is the ratio of the failure cost, C
F
, to the reference initial
cost, C
0
. Equation (4-3) is used herein to deduce the implied failure cost ratio, g for the
reference case of the design code. This value of g is then used consistently to optimize
load factors for all other load combination cases.

4.2 Optimization Problem
The normalized form of the total cost function is valid for a single design (a single
combination of Ln/Dn and Wn/Dn ratios for instance). Ravindra and Lind (1971) defined
the code optimization as the minimization of the sum of expected total costs of all
structures deigned using a structural code, so in order to optimize the code, the whole
design space needs to be considered not just a single design. Either optimizing the
individual designs and then averaging the load factors obtained, or optimizing the sum of
the total cost factors, TCF given in Equation (4-3) throughout the design space and
obtaining the load factors can be chosen as the solution procedure. Since the first
approach needs more optimization calculations then the latter one and also optimization
of the sum is a society approved and good goal to achieve, in this study optimization of


72
the sum of the total cost factors is preferred. The objective function can be formulated as
follows;
[ ]
[ ] ( )
[ ] [ ] [ ]
upper lower
TCF

space design
min to
Find
(4-9)
The load factors, (such as
D
,
L
,
W
etc.) making the Equation (4-9) minimum
will be the optimum load factors that give the balanced solution. Bounds for the load
factors are used in order to avoid unreasonable values of for Equation (4-9). Since the
load factors would probably not differ too much from current code values, it would be
easy to set reasonable upper and lower boundaries for them.
The total cost factor, TCF contains the expected failure cost, for which the
probability of failure, P
f
needs to be calculated for every year during the lifetime.
Reliability analyses are performed by using FORM; the Rackwitz-Fiessler algorithm (See
Appendix), which uses Ferry-Borges Method is preferred in this study.

4.3 Solution Procedure
In solving the optimization problem Equation (4-9), the following steps are
followed;
A. Find normalized cost of failure for basic load combination


73
Pick the single most basic load combination case and assume current safety
level () for this load combination case and load factors are optimum.
Define the design space for this basic load combination (such as in dead and
live load case a L
n
/D
n
ratio ranging between 0.5 to 2.0)
Obtain the expression for TCF in the design space using load factors
prescribed by the current code for the basic load combination (Sum Equation (4-
3) along the design space).
Taking numerical derivative of TCF (sum of Equation (4-3) along the design
space) with respect to load factors and equate to zero about the safety level. The
resulting expression is in terms of g, so g can be determined from that
expression obtained from that expression as seen below.
[ ] ( )
[ ]
[ ] ( )
[ ]
0
space design 1
space design
=

T
i
ij
f
e P
g
TCF


[ ] ( )
[ ]
[ ] ( )
[ ]

=
space design 1
space design
T
i
ij
f
I
e P
C
g






74
B. Optimize TCF to find [] for each load combination. Also note the corresponding

average
.
Define the design space for each load types in load combinations (W
n
/D
n
, E
n
/D
n

ratio ranges)
Use the deduced g and statistical data available along marginal cost slopes, K
and s and set the expression for TCF by summing Equation (4-3) along the
design space.
Find the annual probability failure for each single design. When only one time
dependent load is involved in load combination than a regular FORM algorithm
can be used to find the annual failure probabilities. Otherwise the Ferry-Borges
Method explained in Chapter 2 is capable of dealing with time dependent load
variables by using the independent rectangular pulses.
In order to have more accurate results the time dependent loads needs to be
decomposed to their time invariant and variant parts. The load effects have a
time dependent part (such as wind speed, wave height, etc.) and time
independent parts due to modeling which are result of uncertainties of the
converting for instance, wind speed to bending moment on an element. So the
time variant and invariant parts have to be distinguished in order not to violate
the basic principle of the Ferry-Borges Method, that is, the independence of the
single load event recurrences.


75
The Ferry-Borges Method (Rackwitz and Fiessler Algorithm) need to be used
together with either Nested Reliability Analysis or Numerical Integration to
calculate the expectation of the probability failure given by Equation (2-33).
Using the optimization procedure find the load factors that gives a minimum
TCF for each load combination cases.
For every load combination case use optimized load factors and calculate safety
index, for each single design and then average that to find the
average
for that
load combination case.

4.4 Numerical Example
In order to present the implementation of the proposed method and clarify the
approach the following example is studied. A fictitious code that consists of three load
combinations such as dead plus live load (D+L), wind load alone (W) and dead plus wind
load (D+W) is analyzed. Lifetime statistical data used are as in Table 9.
Table 9 Lifetime Statistical Data (T=75 years)
Random Variable Symbol Distr. Type Bias COV
Resistance R Lognormal 1.12 0.10
Dead Load D Normal 1.03 0.08
Live Load L Lognormal 1.00 0.18
Wind Load W Extreme Type I 0.87 0.33




76
The current load factors and design check equations used are as follows for the
above mentioned three load combination cases.
n n n
n n
n n n
W D R
W R
L D R
40 . 1 25 . 1
40 . 1
75 . 1 25 . 1
+

+
(4-10)
It is required to calibrate the respective load factors to reach an optimum code.
The random variables L and W show variation during the lifetime. The reliability based
cost optimization as studied in this study is used to optimize the above-mentioned code.
In order to use reliability-based cost optimization failure probability needs to be
calculated for every year during the lifetime of the structure. The FORM analyses allow
using order statistics to find the annual probabilities. Also as seen in the application of the
proposed approach to calibration of AASHTO LRFD Specifications (1994), when more
than one time dependent load is involved in a load combination case, a special treatment
such as Turkstras rule, or Ferry-Borges Method is required for the combination process.
The time dependent load effects like live load or wind load effect actually is a synthesis
of time dependent and independent random variables. Environmental events are time
dependent, but when converting the environmental events to load effects uncertainties
due to modeling are introduced. Both in using in order statistics and Ferry-Borges
Method an important assumption is made that the events are independent of each other.
Unfortunately, since the random variables consist of the time invariant random variables,
assumption of independency among events is violated. Therefore, a separation is required
to use the independency of events assumption. Next, the separation of the live wind load
is explained.


77
4.4.1 Separation of Modeling Uncertainty for Live Load
The lifetime live load is given as in Table 9. The modeling uncertainty is same
throughout the lifetime. Therefore, if the modeling is separated from the time variant
data, the time variant data would be available to use in order statistics, which can be
stated as
( )
k
k year in
Q CDF Q CDF
1
) ( = (4-11)
where Q is the time variant random variable and CDF is the cumulative distribution
function for the distribution of the random variable Q. Equation (4-11) is very helpful for
the whole calibration process. In FORM analysis, Equation (4-11) ensures to use the
exact cumulative distribution for finding annual probabilities, which are crucial in
calculating annual cost of failure during calculation of total failure cost. Let us say the
75-year live load effect, L
75
can be written as
75 75 L ML
Q X L = (4-12)
where X
ML
is the modeling uncertainty and Q
L75
is the 75 year live load effect without
modeling. Since L
75
contains both time variant and invariant terms load events are
correlated, and
( ) ( ) [ ]
75
1 75
L CDF L CDF (4-13)
but following holds,
( ) ( ) [ ] ( ) ( ) [ ]75
1
75 1
75
1 75 L L L L
Q CDF Q CDF or Q CDF Q CDF = = (4-14)


78
so Equations (4-11 to 14) clearly state that separation of the time variant and invariant
terms in time dependent random variables is important not only for Ferry-Borges Model
but for the annual failure probability calculations. Having the lifetime statistical data and
using the second part of the Equation (4-14) makes it very easy to use for any year n
during the lifetime as
( ) ( ) [ ]75
75
n
L Ln
Q CDF Q CDF = (4-15)
Now let us reconsider Equation (4-12) where 75 years live load effect is
expressed as the multiplication of modeling random variable X
ML
and time variant Q
L75
,
which is the result of independent 75-year events. It is assumed that the modeling r.v.,
X
ML
has a bias of 1.0 and a COV of 12%. The statistical data for L
75
are a bias of 1.0 and a
COV of 18% (Table 9). Rewriting Equation (4-12) gives
ML
L
X
L
Q f
75
75
= =
(4-16)
Let us use the Taylor series expansion to evaluate bias and COV of Q
L75
(Ang and Tang
(1975)).
( ) ( )
14 . 0
134 . 0 0 . 1 * 12 . 0
0 . 1
0 . 1
0 . 1 * 18 . 0
0 . 1
1
75
75 75
2 2
2
2
2
2
2
2
2
2
75
2

= |
.
|

\
|
|
.
|

\
|
=
|
|
.
|

\
|

+
|
|
.
|

\
|

L
ML L
Q
X
ML
L Q
X
f
L
f


(4-17)
( )
( ) ( )
0 . 1 99 . 0
0 . 1 * 12 . 0
0 . 1
0 . 1
2 0 . 1 * 18 . 0 0
2
1
0 . 1
2
1
,
75
75 75 75
2
3
2
2
2
2
2
2
75
2

)
`

|
|
.
|

\
|

+
|
|
.
|

\
|

+
L
ML ML L
Q
X
ML
L L X Q
say so
X
f
L
f
f


(4-18)


79
therefore the bias is 1.0 and a COV of 0.13 and assumed that the distribution of the Q
L75

is lognormal.
4.4.2 Separation of Modeling Uncertainty for Wind Load
The separation of the modeling uncertainty is also carried out for the wind load.
Wind load can be expressed as the wind load statistic data; the procedure in ANSI-A58 is
followed in this study. The Equation (4-19) represents the wind load effect.
2
V GC cE W
p Z
= (4-19)
where c is a constant, E
z
is exposure coefficient, G is gust factor, C
p
is the pressure
coefficient and V is the wind speed at a height of 33 ft. above ground (ANSI A-58).
Equation (4-19) basically converts the wind speed to wind load effects on structure. The
coefficients multiplying velocity square all assumed to follow a normal distribution and
have bias of 1. The COV of the C
p
, G, E
z
are 0.12, 0.11, and 0.16 (Ellingwood et al.
(1980)). These are the terms that contribute to modeling uncertainties. There is also a
factor of 0.85 to account for the possibility of having the wind in a direction other than
the unfavorable direction. The Equation (4-19) is rewritten as,
75 75 W MW
Q X W = (4-20)
where X
MW
is the modeling uncertainties and Q
W75
is a function of the wind speed
squared. X
MW
, is the combination of C
p
, G and E
z
and follows a normal distribution with a
bias of 1.0 and a COV of 23 %. In ANSI A58 both the annual and 75-year wind load
effect are assumed to follow Gumbel distribution. It is assumed herein Q
W75
also follows
a Gumbel distribution and has 0.87 and 24% as its bias and COV values. In Belk and


80
Bennett (1991), it is stated that the mean rate of occurrence is 0.0176 per hr and the mean
individual duration for a storm is 3.76 hr. In this study the mean individual duration is
assumed as 4 hr.
4.4.3 Statistical Data after Separation of Random Variables
After separating the time dependent random variables into time variant and
invariant random variables the statistical data are as in Table 10. In code calibration
process the data presented in Table 10 are used for all load combination cases.

Table 10 Statistical data after time variant and invariant distinction
R.V. Distr. Type Bias COV
R Lognormal 1.12 0.10
D Normal 1.03 0.08
Q
L75
Lognormal 1.00 0.13
X
ML
Normal 1.00 0.12
Q
W75
Extreme Type I 0.87 0.24
X
MW
Normal 1.00 0.23


4.4.4 Load Combination Cases
The load combination cases are also modified for the random variables given in
Table 10, and Equation (4-10) takes the form as
Wn MWn W n D n
Wn MWn W n
Ln MLn L n D n
Q X D R
Q X R
Q X D R
3 3
2
1 1

+
(4-21)


81
The calibration process has to be run for the combination cases in Equation (4-21)
and the optimum load factors calculated by minimizing the sum of the total cost factor
over the set of possible designs. The set of the possible designs is called the design space
and should cover the designs that the code aims to address. The designs points are
represented by the nominal load values to nominal dead load value such as L
n
/D
n
, W
n
/D
n
,
etc.
4.4.5 Calibration Process
Calibration of the given code in this example is carried out first assuming that the
load factors,
D0
=1.25 and
L0
=1.75 given in Equation (4-1) are the optimum load factors
and the sum of the total cost over the design space represented by 4 equally likely design
cases namely L
n
/D
n
={0.5, 1.0, 1.50, 2.0} is a minimum. So the g given in the next
equation defines the ratio of the failure cost to base cost, C
0
that the society is accepting
with code.

+
|
|
.
|

\
|

+
+
+ =
ij
fi
L D
L D
G
e
P
g K TCF 1 1
0 0
(4-22)
where K
G
is the normalized cost factor for gravitational load, and is the nominal live
load to nominal dead load ratio in the range of 0.5 to 2.0, g is the failure cost ratio, i is the
year, j is the real interest rate assumed as 3%, and P
f
is the failure probability that is
calculated using the limit state function
L ML
Q X D R G = (4-23)


82
is used along with the FORM analysis for every year during the lifetime. A study
conducted on bridges by Moses (1989) showed that 25% change in
L
produce a change
of 2% in initial cost of bridges. This is used as a representative cost slope input for the
example.
Initial cost factor, ICF of a design is;
|
|
.
|

\
|

+
+
+ = 1 1
0 0 L D
L D
G
K ICF


(4-24)
The ratio of change in initial cost, C
I
to base cost C
0
due to change in live load
can be stated as;
0 0 0
0
] 1
1
[
C
C
K
I
L D
L
L
L D
G

=
+
|
|
.
|

\
|
+ +


(4-25)
Rearranging Equation (4-25) gives
1
0 0
0
0
1
1

(
(
(
(
(

+
|
|
.
|

\
|
+ +


L D
L
L
L D
I
G
C
C
K (4-26)
Introducing the data in Table 11 into Equation (4-26) the gravitational cost factor,
K
G
is obtained for different values of as in Table 12. Averaged along the range of ,
the factor K
G
is calculated as 0.14, and this value is used for the rest of calibration process
in the total cost factor function.



83

Table 11 Data for calculation of gravitational load cost factor, K
G

D0

L0

0 L
L



0
C
C
I



1.25 1.75 0.25 0.02 0.5,1.0,1.5,2.0


Table 12 K
G
values
K
G
0.50 0.194
1.00 0.137
1.50 0.118
2.00 0.108
Average K
G
= 0.14


Next step in the calibration process is deducing the failure cost, g and the total
cost factor, TCF function is going to be used in deducing g is as;

+ |
.
|

\
|

+
+
+ =
ij
fi
L D
e
P
g TCF 1
75 . 1 25 . 1
14 . 0 1


(4-27)
The sum of TCF along the design space given as Ln/Dn=={0.5,1.0,1.5,2.0} is a
minimum at the current code values of the load factors for the load combination case 1,
and an average safety level, of 3.5 is attained. Differentiating TCF with respect to load
factors and equating to zero gives,


84
0
75 . 1 25 . 1
1
14 . 0 =

+ |
.
|

\
|
+
=

ij
D
fi
D
e
P
g
TCF

(4-28.a)
0
75 . 1 25 . 1
14 . 0 =

+ |
.
|

\
|
+
=

ij
L
fi
L
e
P
g
TCF

(4-28.b)
Rearranging Equations (4-28.a) and (4-28.b),
1
75 . 1 25 . 1
1
14 . 0

(
(

|
.
|

\
|
+
=

D
ij
fi
e P
g

(4-29.a)
1
75 . 1 25 . 1
14 . 0

(
(

|
.
|

\
|
+
=

L
ij
fi
e P
g

(4-29.b)
Equation (4-29.a) or (4-29.b) can be used to deduce the failure cost factor. Consider
Equation (4-29.b), it can be rewritten as
1
1
75 . 1 25 . 1
1
14 . 0
75 . 1 25 . 1
14 . 0

(
(

|
.
|

\
|
+
=
(
(

|
.
|

\
|
+
=

n
ij
fi
L
n
n
ij
fi
R
e P
R
R
e P
g

(4-29.c)
As seen in Equation (4-29.c) the Equations (4-29.a and b) give the same result for g, as
tabulated in Table 13 and Table 14. The g changes with . Since the g needs to represent
possible values an average value of the deduced values of g is used for the calibration
process. Using a g value in calibration of the other load combinations does not utilize a
uniform safety level for these load combinations, but a consistent and uniform
consequence of failure is maintained among them.



85
Table 13 Deducing an average g value using Equation (4-29.a)

0.5 0.47 -591.199 38.95


1.0 0.33 -1365.237 63.71
1.5 0.26 -1807.072 65.29
2.0 0.21 -2141.212 63.11
Average g= 58
|
.
|

\
|
+
=
75 . 1 25 . 1
1
A
1

(
(

=

D
ij
fi
e P
B

B A g * * 14 . 0 =



Table 14 Deducing an average g value using Equation (4-29.b)

0.5 0.24 -1182.398 38.95


1.0 0.33 -1365.237 63.71
1.5 0.39 -1204.715 65.29
2.0 0.42 -1070.606 63.11
Average g= 58
|
.
|

\
|
+
=

75 . 1 25 . 1
A
1

(
(

=

L
ij
fi
e P
B

B A g * * 14 . 0 =




The 58 value found means that the expected failure to cost is 58 times the initial
cost which corresponds to current code design. For instance for a design of $100,000, the
failure cost is expected to be around $5,800,000. It is verified that the total cost factor,
TCF is a minimum for dead plus live load case at the current safety level of =3.5 by


86
using the deduced g and calculated K
G
(See Figure 8). Using the data given in Table 10
and the limit state function given in Equation (4-23), the following sum of the total cost
factors given in equation below is optimized.
( )
( )

+
|
|
.
|

\
|

+
+
+ =
i
L D fi
L D
L D
L D
L D L D
e
P
TCF
that and and find
03 . 0
0 0
,
58 1 14 . 0 1 ) , ( min
,





(4-30)
The optimum safety index value is 3.50 as can be seen in Figure 8. The safety
index, vs. sum of initial cost factors, ICF, sum of factor of cost of failure, FCF, and
sum of total cost factor, TCF are depicted in Figure 9. The minimum TCF is obtained
when slope of ICF is equal to negative slope of FCF. The optimum load factors are at
1.25 and 1.75 for dead and live load, respectively. These are the values in current code,
so this verifies the procedure to deduce g.
The same procedure to deduce g value is applied for system safety as well. The
system safety index is assumed as component safety index plus one. So, for a system
safety optimum index value is 4.50, using Equation (4-29a or b) with the system safety
index (component safety index plus one) the g value is deduced as 3528. The expected
total cost optimization is performed for both component and system safety.




87

Figure 8 Component Safety index, vs. sum of total cost factors, TCF


88



Figure 9 Component Safety Index vs. ICF, FCF and TCF



89
The calibration procedure is applied to the second load combination given in
Equation (4-21). There is only one load involved in this case, wind load. The current code
wind load factor gives a component safety index of 1.86. The initial cost factor for
current code load factor is 1.0 because the current code factor is assumed to be the base
design, and the total cost function is normalized with base cost, C
0
. The expected failure
and total costs are calculated as 0.75, and 1.75, respectively for the current code
considering component safety (g=58). Expected failure cost factor and total cost factor
increases to 2.54 and 3.54, respectively if system safety (g=3528) is considered.
The optimization analysis for finding minimum expected total cost and
corresponding safety index and load factor is run for different wind cost slope to
gravitational cost slope ratios namely,
W
=K
W
/K
G
={0.5 1.0 2.0 3.0} for both component
safety (deduced g=58), and system safety (deduced g=3528).
The TCF function used in this load combination is as Equation (4-31). Since there
is one design point in this case the summation is not used.
ij
fi
WO
W
W
e P g K TCF

+
|
|
.
|

\
|
+ = 1 1

(4-31)
The optimization problem is
( )
( )

=
+
|
|
.
|

\
|
+ =
75
1
03 . 0
0
1 1 ) ( min
i
i
W fi
W
W
W W
W W
e
P
g K TCF
that and find


(4-32)
The optimum safety indices and load factors are tabulated in Table 15 and Table
16 for component and system safety, respectively. Optimum component and system


90
values decrease with the increasing K
W
/K
G
values. The cost of increasing wind load factor
would be higher than increasing the load factor for gravitational loads. For component
safety (g=58) if marginal cost factor for wind load is assumed to be equal to gravitational
marginal cost factor of 0.14, the optimum level for safety is at 3.10 and a load factor of
2.15 for wind load is required for this safety index. If the K
W
/K
G
is assumed to be 2 than
the optimum wind load factor drops to 2.00. Since the wind load factor has relatively
higher effect on the initial cost a K
W
/K
G
value of 3 is assumed. Then, the load factor for
wind would be 1.92 with a component safety index of 2.77. For system safety the
optimum load factors increase compared to component safety, this increase is small if the
K
W
/K
G
is small, but as the K
W
/K
G
increase the increase is about 0.1. Assuming K
W
/K
G
is
equal to 3.0 the optimum wind load factor is found to be 2.03 with a corresponding
system safety of 3.93. The optimized wind load factor decreases the expected total cost
considerably, from the current code; optimization considering component safety
decreases the TCF from 1.75 to 1.22 a 30% decrease, and for system safety this decrease
is more significant, from 3.75 to 1.24, the TCF is lowered to nearly one third of the
current codes.







91

Table 15 Optimum component and
W
(g=58) for Wind Load Alone Case
K
W
/K
G

comp

W
ICF FCF TCF
0.5 3.30 2.31 1.0453 0.0113 1.0565
1.0 3.10 2.15 1.0754 0.0222 1.0976
2.0 2.90 2.00 1.1208 0.0438 1.1646
3.0 2.77 1.92 1.1553 0.0651 1.2204


Table 16 Optimum system and
W
(g=3528) for Wind Alone Case
K
W
/K
G

sys

W
ICF FCF TCF
0.5 4.33 2.33 1.0467 0.0090 1.0557
1.0 4.18 2.21 1.0815 0.0177 1.0991
2.0 4.03 2.10 1.1394 0.0346 1.1740
3.0 3.93 2.03 1.1887 0.0513 1.2400


The last combination case given in Equation (4-21) is the combination of dead
and wind load, and TCF of this combination case can be expressed as;
( ) ( )

|
|
.
|

\
|
+
|
|
.
|

\
|
+
+
+ =
=

W W
i
ij
fi
W W D
W W W W D D
G
e P g K TCF



75
1 0 0
0 0
1 (4-33)
where
W
is the K
W
/K
G
ratio. Equation (4-33) is calculated for the possible design values
given as
W
=W
n
/D
n
={0.10 0.25 0.50 1.00 2.00} using current load factors for both
component and system safety, and the results are given in Table 17. The summation
given in Equation (4-33) is minimized for
W
=K
W
/K
G
={0.5,1.0,2.0,3.0} first for
component safety with deduced g=58, and then for system safety with deduced g=3528.
Again the system safety is assumed to be component safety index plus 1. The optimum


92
safety index and load factor values are given in Table 18 and Table 19. For both case the
wind load factor decreases and dead load factor increases as the K
W
/K
G
ratio increases,
but the average safety index do not show the same trend starting with the K
W
/K
G
=2.0.
For high W
n
/D
n
values since increasing the wind load factor is expensive the dead load
factor is increased instead. A considerable increase in the dead load factor is required in
order to lower the expected failure cost to attain a balance between initial cost and
expected failure cost for the cases with high Wn/Dn values. This increase in dead load
also increases the safety levels to relatively higher levels for the cases where wind load
does not dominate (lower Wn/Dn values). Therefore the average safety index values start
to increase with increasing K
W
/K
G
. If component safety is considered for K
W
/K
G
=2 the
optimum dead load factor is 1.40 and wind load factor is 1.86, which correspond to a
TCF of 5.22. For same K
W
/K
G
considering system safety optimum dead load factor is
1.38 and wind load factor is 1.95, these optimum load factors give the TCF as 5.23. If
K
W
/K
G
=3.0, then dead load factor, wind load factor and TCF are 1.46, 1.76 and 5.23,
respectively for component safety and 1.44, 1.88 and 5.24 respectively for system safety.





93
Table 17 Current code safety levels and costs for D+W case

D

W
W
n
/D
n

comp

sys
ICF FCF TCF FCF TCF
1.25 1.40 0.10 2.66 3.66 1.0000 0.1468 1.1468 0.2609 1.2609
0.25 2.74 3.74 1.0000 0.0790 1.0790 0.1242 1.1242
0.50 2.56 3.56 1.0000 0.1251 1.1251 0.2301 1.2301
1.00 2.31 3.31 1.0000 0.2512 1.2512 0.5795 1.5795
2.00 2.11 3.11 1.0000 0.4112 1.4112 1.1231 2.1231
2.48 3.48 = 5.0000 1.0134 6.0134 2.3177 7.3177 Average =
Component System



Table 18 Optimum component ,
D
and
W
(g=58) for Dead +Wind Load Case
K
W
/K
G

D

W
W
n
/D
n

comp
ICF FCF TCF
0.5 1.29 2.19 0.10 3.33 1.0085 0.0150 1.0235
0.25 3.72 1.0136 0.0024 1.0160
0.50 3.62 1.0207 0.0034 1.0242
1.00 3.44 1.0311 0.0067 1.0379
2.00 3.31 1.0437 0.0108 1.0545
3.49 = 5.1177 0.0383 5.1560
1.0 1.34 2.01 0.10 3.48 1.0148 0.0086 1.0233
0.25 3.73 1.0208 0.0023 1.0232
0.50 3.53 1.0280 0.0048 1.0329
1.00 3.29 1.0367 0.0117 1.0485
2.00 3.12 1.0451 0.0212 1.0663
3.43 = 5.1455 0.0487 5.1942
2.0 1.40 1.86 0.10 3.76 1.0225 0.0028 1.0253
0.25 3.84 1.0277 0.0015 1.0291
0.50 3.52 1.0326 0.0050 1.0377
1.00 3.19 1.0374 0.0164 1.0538
2.00 2.97 1.0411 0.0348 1.0759
3.46 = 5.1613 0.0605 5.2219
3.0 1.49 1.76 0.10 4.17 1.0295 0.0005 1.0300
0.25 4.07 1.0313 0.0006 1.0318
0.50 3.60 1.0327 0.0037 1.0364
1.00 3.17 1.0339 0.0175 1.0514
2.00 2.88 1.0348 0.0455 1.0803
3.58 = 5.1622 0.0677 5.2299
Average =
Average =
Average =
Average =





94

Table 19 Optimum system ,
D
and
W
(g=3528) for Dead +Wind Load Case
K
W
/K
G

D

W
W
n
/D
n

sys
ICF FCF TCF
0.5 1.29 2.19 0.10 4.36 1.0090 0.0125 1.0215
0.25 4.74 1.0141 0.0013 1.0155
0.50 4.64 1.0213 0.0022 1.0235
1.00 4.45 1.0317 0.0052 1.0369
2.00 4.32 1.0443 0.0095 1.0538
4.50 = 5.1204 0.0307 5.1512
1.0 1.33 2.06 0.10 4.47 1.0145 0.0076 1.0221
0.25 4.74 1.0213 0.0014 1.0226
0.50 4.57 1.0293 0.0031 1.0324
1.00 4.34 1.0390 0.0089 1.0479
2.00 4.18 1.0483 0.0182 1.0665
4.46 = 5.1523 0.0391 5.1914
2.0 1.38 1.95 0.10 4.67 1.0217 0.0028 1.0245
0.25 4.82 1.0289 0.0009 1.0298
0.50 4.56 1.0358 0.0033 1.0391
1.00 4.27 1.0425 0.0122 1.0547
2.00 4.07 1.0477 0.0291 1.0768
4.48 = 5.1766 0.0483 5.2250
3.0 1.44 1.88 0.10 4.95 1.0278 0.0007 1.0285
0.25 4.98 1.0333 0.0004 1.0337
0.50 4.61 1.0378 0.0025 1.0404
1.00 4.25 1.0416 0.0131 1.0547
2.00 4.01 1.0443 0.0371 1.0814
4.56 = 5.1849 0.0538 5.2387
Average =
Average =
Average =
Average =


In order to make a recommendation for load factors, dead load factor is fixed at
current code specified 1.25, and calibration process is repeated for component safety
(g=58) and system safety (g=3528). Optimum wind load factors with average safety
levels are presented in Table 20 and Table 21. Since the wind load marginal cost slope is
relatively higher than the gravitational load cost slope, the K
W
/K
G
is assumed to be 3.0.
For component safety the optimum wind load factor is found to be 2.05 with a TCF of
5.29. When the system safety is considered the optimum wind load factor is 2.12 with a
TCF of 5.30. Since the failure of a component does not necessarily mean a complete


95
failure, system safety is considered to reach a conclusion about the wind load factor. The
optimum wind load factor of 2.12 reduces the TCF 28% compared to current code wind
load factor.

Table 20 Optimum component and
W
(
D
=1.25, g=58) for Dead +Wind Load Case
K
W
/K
G

D

W
W
n
/D
n

comp
ICF FCF TCF
0.5 1.25 2.30 0.10 3.17 1.0048 0.0269 1.0317
0.25 3.67 1.0110 0.0029 1.0139
0.50 3.65 1.0196 0.0030 1.0227
1.00 3.52 1.0322 0.0050 1.0372
2.00 3.42 1.0474 0.0073 1.0547
3.49 = 5.1151 0.0451 5.1601
1.0 1.25 2.18 0.10 3.11 1.0079 0.0339 1.0417
0.25 3.56 1.0171 0.0044 1.0215
0.50 3.53 1.0280 0.0049 1.0329
1.00 3.39 1.0412 0.0082 1.0494
2.00 3.28 1.0539 0.0121 1.0661
3.37 = 5.1481 0.0634 5.2116
2.0 1.25 2.09 0.10 3.06 1.0127 0.0401 1.0528
0.25 3.48 1.0249 0.0061 1.0310
0.50 3.43 1.0366 0.0069 1.0435
1.00 3.28 1.0479 0.0119 1.0598
2.00 3.17 1.0566 0.0178 1.0744
3.28 = 5.1787 0.0829 5.2615
3.0 1.25 2.05 0.10 3.03 1.0164 0.0433 1.0598
0.25 3.44 1.0298 0.0071 1.0369
0.50 3.39 1.0409 0.0082 1.0491
1.00 3.24 1.0503 0.0141 1.0644
2.00 3.12 1.0568 0.0212 1.0780
3.24 = 5.1943 0.0939 5.2882
Average =
Average =
Average =
Average =








96

Table 21 Optimum system and
W
(
D
=1.25, g=3528) for Dead +Wind Load Case
K
W
/K
G

D

W
W
n
/D
n

sys
ICF FCF TCF
0.5 1.25 2.33 0.10 4.19 1.0049 0.0273 1.0322
0.25 4.71 1.0114 0.0016 1.0131
0.50 4.69 1.0204 0.0017 1.0221
1.00 4.56 1.0334 0.0032 1.0366
2.00 4.46 1.0492 0.0051 1.0543
4.52 = 5.1194 0.0389 5.1583
1.0 1.25 2.23 0.10 4.13 1.0084 0.0350 1.0434
0.25 4.61 1.0182 0.0026 1.0208
0.50 4.58 1.0298 0.0029 1.0327
1.00 4.44 1.0439 0.0054 1.0493
2.00 4.34 1.0574 0.0088 1.0662
4.42 = 5.1577 0.0547 5.2123
2.0 1.25 2.16 0.10 4.09 1.0138 0.0421 1.0559
0.25 4.54 1.0271 0.0037 1.0308
0.50 4.50 1.0400 0.0042 1.0441
1.00 4.36 1.0523 0.0080 1.0603
2.00 4.25 1.0618 0.0132 1.0751
4.35 = 5.1951 0.0712 5.2662
3.0 1.25 2.12 0.10 4.07 1.0182 0.0457 1.0639
0.25 4.51 1.0330 0.0043 1.0373
0.50 4.47 1.0453 0.0049 1.0502
1.00 4.32 1.0557 0.0096 1.0653
2.00 4.21 1.0629 0.0159 1.0788
4.31 = 5.2151 0.0805 5.2956
Average =
Average =
Average =
Average =


4.4.6 Recommended Load Factors
The product of a calibration process is the recommended load factors. As a result
for this example presented herein the load factors given in Table 22 are recommended.
The load factors tabulated in Table 22 is based on the system safety; if the component
safety were used to reach a decision the recommended wind load factors would be
slightly lower.



97

Table 22 Recommended Load Factors
Combination Case
D

L

W
D+L 1.25 1.75 ----
W ---- ---- 2.05
D+W 1.25 ---- 2.10


4.4.7 Discussion of Results
The optimization is performed for both component and system safety. The
deduced g values are 58 and 3528 for component and system safety, respectively. In
deduction of g using D+L case and optimization of the expected total costs for W and
D+W load combination cases the system safety is assumed to be component safety index
plus one. If the component safety is considered, the optimum wind load factors are
slightly lower than the when the system safety is considered. It is preferred to use system
safety herein to conclude about the recommended load factors, since the system safety
represents the reliability of a whole system, nor just a single element.
The optimization clearly indicates the trade-off between safety and economy.
Higher g values, which mean the consequences of failure are more costly, cause a higher
optimum safety index. On the other hand when the consequences have a lower cost the
safety index decreases. The difference between deduced g values for component safety
and system safety is due to the consequences the component and system safety represent.
The marginal cost slopes, K
W
and K
G
also plays a role in finding the optimum. This effect
can be best seen in the second load combination, which is the wind load alone case. An


98
increase of K
W
for example, increases the slope of ICF. The optimum point where slopes
of ICF and FCF equal and opposite, has a lower value, and lower
W
and value are
obtained. If deduced failure cost factor, g is increased for both component and system
safety FCF curve become steeper and the optimum point where the ICF and TCF slopes
are equal and opposite, has a higher value, which results in a higher safety index and high
optimum
W
.






99
5.0 APPLICATION OF CODE CALIBRATION BY USING RELIABILITY
BASED COST OPTIMIZATION TO AASHTO BRIDGE SPECIFICATIONS

5.1 AASHTO Bridge Specifications
Bridges are very important to the lifeline of cities and countries. It appears today
that there has always been a specification or standard for bridge engineering but this is
not the reality. The examples of early wooden bridges and railroad bridges were mostly
built based on experience without available standards. But during the 1870s about 40
bridges failed every year in the United States (Gies (1963)). This relatively high number
brought attention to the fact that there has to be a standard in building bridges. Although
there had been many successful applications such as Kinzua Railroad Bridge seen in
Figure 10, having a rate of 1 failure out of every 4 bridges built was high, and
unacceptable. A failure rate that high made bridges unsafe places in the eyes of the
public. Even though the failure rate was high the engineering practice could not easily
respond to come up with an effective plan. Finally, the American Railroad Engineering
Association (AREA) was the first to accept a standard railroad loading in 1903 (Barker
and Puckett (1997)). Then in 1914, the American Association of State Highway Officials
(AASHO) was formed.
The first modern standards for highway bridges in the U.S. were published in
1931 by AASHO as Standard Specifications for Highway Bridges and Incidental
Structures, and updated periodically since then. There are 16 standard specification


100
editions with the last one published in 1996. In 1963 the institutions name changed to
American Association of Highway and Transportation of Officials (AASHTO), which
make the organization involved with all types of transportation modes.

Figure 10 Kinzua Bridge, McKean County, PA (Built in 1900)

The design philosophy followed in the Standard Specifications is the Allowable
Stress Design (ASD). The ASD is based on the factor of safety concept, and the
resistance such as yield stress or moment capacity was divided by a factor of safety and it
is used to express the allowable limit. The nominal or working load applied should be
less or equal to allowable value. ASD design philosophy deals with the material and the
factor of safety is mostly found intuitively, rather than probabilistically. Recent editions
of the AASHTO Code did provide a load factor design alternative. It is similar to the
LRFD format but factors were fixed arbitrarily without probabilistic analysis.


101
It is a fact that both the materials and loads have many uncertainties in nature. The
idea of dealing with the uncertainties in structural design process with probabilistic
approaches began in late 1960s and 1970. Therefore reaching a sound design
specification become attractive with available tools developed by the researchers around
the world. It is the trend in todays engineering world to migrate from the ASD to a more
sophisticated design philosophy. The ANSI A-58 had been the pioneering document in
this trend and it is followed by many successive applications such as American Petroleum
Institution (API) and American Institute of Steel Construction (AISC). AASHTO
published the First Edition of LRFD Bridge Design Specifications in 1994. The structural
materials and loads are investigated in probabilistic perspective mostly following the
steps outlined in ANSI-A58 (Ellingwood et al. (1980)).
In his work Nowak (1993) has shown that for dead and live truck load combination
the bridge safety index, for LRFD is more uniform than the ASD along the design
space defined by a range of span lengths of bridges. There is a current study ongoing on
Design of Highway Bridges for Extreme Events (NCHRP 12-48) to calibrate the load
combinations other than the basic combination case (Strength I; Dead and Live Load
combination). In calibration study used to develop the 1994 AASHTO LRFD
Specifications, calibration was preceded using reliability-based approach particularly the
Strength I limit state. The load factors for the rest of the limit state cases were selected
from AASHTO Standard Specifications, common practice and the results obtained in
other codes (Ghosn et al. 2001).



102
5.2 Design Format and Design Checks for AASHTO LRFD
In AASHTO LRFD Specifications, the design checks are performed for different
limit states. The limit states considered are strength, service, extreme events and fatigue.
There are five load combinations for strength, three load combinations for service, two
load combinations for extreme events and fatigue.
The loads are categorized as permanent and transient loads. The permanent loads
are weight of structural and non-structural components, wearing surfaces, utilities, earth
fill, and earth surcharge, etc. The transient loads are wind load, ice load, vehicular live
load, earthquake, vessel collision force, etc.
All the limit states have to satisfy the LRFD design check as in Equation (5-1).
i i n
Q R

(5-1)
where is the resistance factor, R
n
is the nominal resistance factor, is the load
modification factor,
i
is the load factors given in Table 1, Q
i
is the force effects from the
loads such as dead, live truck load, wind load, etc. The load modification factor, has
three components, ductility factor,
D
, redundancy factor,
R
and operational importance
factor,
I
. The load modification factor is the multiplication of these three components.
The load modification factor is then
95 . 0 =
I R D
(5-2)


103
It is desirable to have ductile structures. Ductility prevents structures to reach a sudden
failure. In AASHTO for nonductile structures a
D
value of 1.05 or higher is
recommended and a value not less than 0.95 is used for the ductile structures.
The redundancy factor,
R
is used to account for the available alternative load paths
in case one main carrying path has been damaged. In the cases of redundant structures the
factor is 0.95 and for non-redundant structures it is 1.05.
In the non-strength limit states both ductility factor and redundancy factor are taken
as 1.0.
The operational importance factor,
I
is used to impose the importance of the use of
a bridge. In extreme events and strength limit states, for an operationally important bridge
use 1.05 or larger value and for a non-important bridge a value not less than 0.95 is used.
In all other limit states a value of 1.0 is used.
The limit states mentioned in the AASHTO LRFD can be summarized as follows:
Strength I: This is the basic loading case and the dead load and the regular vehicular live
load combination is considered.
Strength II: This limit state deals with the owner specified special vehicle loads.
Strength III: A wind speed velocity exceeding 55 mph is considered in this limit state,
which prevents a significant live load on the bridge. So wind load along with the dead
load is considered omitting live load.


104
Strength IV: In this limit state long span bridges are targeted, where the dead load to live
load ratio is very high.
Strength V: Normal vehicular live load with a 55 mph wind limit is studied in this limit
state.
Extreme Event I: This limit state deals with the load combination with earthquake.
Extreme Event II: The ice load, vessel and vehicle collision, certain hydraulic events are
studied along with a reduced live load to account for the low probability of simultaneous
occurrence of maximum live load with an extreme event.
Service I: The normal operational loads are considered at their nominal values with a
wind speed of 55 mph, and the deflections, crack widths in concrete are studied as well.
Service II: This limit state is used to control the steel yielding and slip of slip-critical
connections under vehicular live load.
Service III: Crack control for prestressed concrete structures.
Fatigue: Check for fatigue and fracture under repetitive vehicular live load.
In this study the following limit states is studied and targeted in a calibration
example; Strength I, III, V and Extreme Event I. Therefore the load combinations and
current load factors are (AASHTO LRFD (1994))
LL EQ DC
WL WS LL DC
WS DC
LL DC
50 . 0 00 . 1 25 . 1 : I Event Extreme
4 . 0 4 . 0 35 . 1 25 . 1 : V Strength
40 . 1 25 . 1 : III Strength
75 . 1 25 . 1 : I Strength
+ +
+ + +
+
+
(5-3)


105
Dead Load, DC, vehicular live load, LL, wind load on structure, WS, wind load
acting on live load, WL, and earthquake load, EQ are involved in Equation (5-3). The
current load factors for Strength I are obtained by using the classical code calibration
procedure, the rest of the load cases load factors are selected from existing deterministic
criteria. First, some selective examples were designed according to the existing code and
the average safety index along these examples is used as the target safety index for the
first edition of AASHTO LRFD (1994). A target safety index equal to 3.5 used in the
calibration of Strength I by Nowak (1993).
It is also important to consider the design life of the structure; the design life
considered for a bridge is 75 years. The design life information will come into picture
when the cost model is used for the bridges and the overall lifetime risk is computed.
In this study a Level IV calibration, which involves optimization of the summation
of the expected total cost is used. The total cost covers both initial construction cost and
expected failure cost. The load factors that give the minimized sum of the total costs are
the optimized load factors and averaged safety index, along the design space will be the
optimum safety level.

5.3 Load and Resistance Data
The statistical data on the resistance and load is mainly from the Nowaks work
and also some information is obtained from the other structural codes such as ANSI A58,


106
ASCE 7-95, and National Earthquake Hazard Reduction Program, NEHRP and the active
research NCHRP Project 12-48 on AASHTO LRFD.
5.3.1 Resistance Data
The resistance of an element or a connection is considered as the product of the
material factor, M, fabrication factor, F and analysis or professional factor, P as
( ) PMF R R
n
= (5-4)
The factors F and M are combined and the bias (mean over nominal), B
R
and
coefficient of variation (standard deviation over mean), V
R
are calculated as in Table 23
using Equation (5-5).
2 2
P FM R
FM P R
V V V
B B B
+ =
=
(5-5)
The distribution type used in Nowaks (1993) study is lognormal. In this study,
the resistance is also represented by lognormal probability distribution and a bias of 1.12
and a COV of 0.10 are used as the parameters for a representative example.








107

Table 23 Resistance Statistical Parameters (Nowak 1993)
B
FM
V
FM
B
P
V
P
B
R
V
R
Moment (compact) 1.095 0.075 1.020 0.060 1.120 0.100
Moment (non-compact) 1.085 0.075 1.030 0.060 1.120 0.100
Shear 1.120 0.080 1.020 0.070 1.140 0.105
Moment 1.070 0.080 1.050 0.060 1.120 0.100
Shear 1.120 0.080 1.020 0.070 1.140 0.105
Moment 1.120 0.120 1.020 0.060 1.140 0.130
Shear w/steel 1.130 0.120 1.075 0.100 1.200 0.155
Shear no steel 1.165 0.135 1.200 0.100 1.400 0.170
Moment 1.040 0.045 1.010 0.060 1.050 0.075
Shear w/steel 1.070 0.100 1.075 0.100 1.150 0.140
Type of structure
FM P R
Non-Composite Steel girders
Composite Steel girders
Reinforced Concrete
Prestressed concrete



5.3.2 Dead Load Data
Dead load is the self-weight of the materials used in the bridge construction, such
as weight of factory made materials, cast-in-place concrete, wearing surface, etc. The
values used by Nowak (1993) are tabulated in Table 24.





108

Table 24 Dead Load Statistic Data (Nowak (1993))
Component B
D
V
D
Factory-made members 1.03 0.08
Cast-in-place members 1.05 0.1
Asphalt 3.5 inch (mean thickness) 0.25
Miscellaneous 1.03-1.05 0.08-0.10

Nowak (1993) assumed a normal distribution for the total of dead and live load
but did not specify a distribution type for dead load separately. In this study, a normal
distribution is assumed for dead load. A representative value of dead load with a bias of
1.03 and 0.08 is used in the analysis in this study.
5.3.3 Live Load Model
The effect of truck live load on bridges is a complex issue and the span length of
the bridge; transverse and longitudinal position and axle configuration of a truck are
among the variables that affect the live truck load effect (Nowak (1993)). In calibration
of AASHTO specification according to LRFD philosophy the live load effects received a
considerable amount of attention. Nowak (1993) recommended the HL-93 (see Figure
11), which contains a uniform load of 640 lb/ft in addition to the standard HS-20 design
truck already used in AASHTO Standard Specifications. The safety index values
obtained by using HL-93 were more uniform for different spans compared to ones with
HS-20.


109
The live load effect can be studied in three parts; static effect due to the weight of
the truck(s), live load analysis factor (also contains site-to-site variability) and dynamic
effect due to truck movement on the bridge. In his work Nowak (1993) has used a COV
around 19% to 20%, which contains the static, dynamic and modeling effects. Among the
variables contributing to live load only the static part which is due to truck weight is time
dependent, and in order to be able to use in the cost analysis that part has to be separated
from the other load random variables. The live load can be expressed as;
L L D
Q X I L = (5-6)
where I
D
is the dynamic r.v., X
L
is the modeling uncertainties and the Q
L
is the truck
weight r.v..
Nowak (1993) provided tables for 75-year effects on bridges for different span
lengths in single or multi span bridges by extrapolating the data from Ontario Ministry of
Transportation. The COV values for these values are not provided. The number of heavy
trucks for a day used in Nowaks analysis is 1000 events/day and 1/15 (67 events/day) of
these events are assumed as multiple presence side-by-side truck loading events. Moses
(2000) estimated that the truck weights used in calibration of AASHTO follows a Normal
Distribution with a mean value of 68 kips and a standard deviation of 18 kips. This data is
used in research conducted in City University of New York (CUNY) on Calibration of
AASHTO Specifications for Extreme Events (NCHRP 12-48). The expected maximum
weight corresponding to the standard HL93 load gives the bias for truck loading effect.
The COV of the truck weight for different return periods is calculated by using the


110
assumption of independence of the single events as follows; if a single truck load has a
cdf and pdf values F(x) and f(x), respectively than in T years the cdf and pdf would be

( ) ( )
( ) ( ) ( ) x f x F T x f
x F x F
T
T
T
T
1 * 365 * 1000
* 365 * 1000
* * 365 * 1000

=
=
(5-7)
the cdf in T years us calculated by raising power of the 1 events cdf to 1000*365*T, and
the pdf is found by differentiating cdf with respect to x. The mean and standard deviation
at time T can be evaluated by
( ) ( )
( ) ( )
2
1 * 365 * 1000
2
1 * 365 * 1000
) ( * 365 * 1000
) ( * 365 * 1000
T
T
T
T
T
dx x f x F T x
dx x f x F T x

=
=

(5-8)
and the statistical data for single truck loading are obtained are tabulated in Table 25.

Table 25 The statistical data at time T for one truck loading
T COV
1 event 68.0 18.0 0.265
1 yr 151.9 4.6 0.031
5 yrs 157.7 4.4 0.028
10 yrs 160.0 4.3 0.027
20 yrs 162.4 4.2 0.026
50 yrs 165.4 4.1 0.024
75 yrs 166.7 4.0 0.024


The 75-year truck weight has a mean of 167 kips and a COV of 2.4%. These
values are very close with the values obtained in analysis at CUNY (mean is 165 kips and
a COV of less than 3%) using a similar procedure. Herein a 3% value is used for COV as
well in order to decide about the modeling uncertainties for live load model. As the return


111
period increases the distribution for the truck weight approaches an extreme value
distribution, and the variation from the mean value for that return period decreases.
However, Ghosn et al. (2001) has shown that in order to replicate the tables given
by Nowak (1993) for one lane and two lane loading to HL-93 truck loading ratio, a bias
of 0.79 (mean truck weight for a single event to HL-93 loading) and a COV of 10% for
single lane occurrence and a bias of 1.58 (mean truck weight for side-by-side occurrences
event to HL-93 loading) with a COV of 7% are needed. Herein these values are assumed
for the one lane and two-lane loading.
Hwang and Nowak (1991), and Nowak (1993) modeled the dynamic effect as a
function of three parameters; road surface roughness, bridge dynamics and vehicle
systems. Numerical simulations were used to investigate the dynamic effects and the
following statistical data is found for calibration purposes; the dynamic amplification
factor has a mean value of 13% more than the static effect with a COV of 10% for single
event and a mean value of 9% more than the static effect with a COV of 6% for side-by-
side two-lane occurrence.
The data gathered above is used to conclude the modeling uncertainties used in
reliability analysis (Ghosn et al. (2001)). Nowak (1993) has used a 19% to 20% COV for
side-by-side occurrence and with the 6% COV for dynamic amplification and 3% COV
for truck weight gives a range of COV for modeling uncertainties including site-to-site
variation is 17.8% (= 2 ^ 03 . 0 2 ^ 06 . 0 2 ^ 19 . 0 ) to 18.8% (= 2 ^ 03 . 0 2 ^ 06 . 0 2 ^ 20 . 0 ), and


112
as the average an 18% value is assumed for modeling uncertainties following a lognormal
distribution.
The statistical data for live load model given in Equation (5-6) is summarized in
Table 26.

Table 26 Live Load Model Statistical Data
Variable BIAS COV Distr. Type
X
L
1.00 18% Lognormal
I
D1
(one truck single event) 1.13 10% Normal
I
D2
(side-by-side single event) 1.09 6% Normal
Q
L1
(one truck single event) 0.79 10% Normal
Q
L2
(side-by-side single event) 1.58 7% Normal


640 lb/ft
8 kips
32 kips 32 kips
14' 14-30'

Figure 11 HL-93 truck and lane loading (AASHTO LRFD(1994))



113
5.3.4 Wind Load Model
Nowak (1993) suggested that although the wind load model based on the report
by Ellingwood et al. (1980) is gathered for building structures for most cases the same
model can be used for bridges as well. In order to obtain the wind load statistic
parameters, the procedure in ANSI-A58 is followed herein. Equation (5-9) represents the
wind load effect.
( )
2
V X GC cE W
v p Z
= (5-9)
where c is an analysis coefficient, E
z
is exposure coefficient, G is gust factor, C
p
is the
pressure coefficient, V is the wind speed at a height of 33 ft. above ground (ANSI-A58),
and X
V
is the statistical uncertainties in V and all the terms are random variables.
Equation (5-9) basically converts the wind speed to wind load effects on structure, and
the random variables multiplying velocity square are all assumed to follow a normal
distribution and have bias of 1 and the COV of the c, C
p
, G, E
z
, X
V
are 0.05, 0.12,
0.11,0.16, and 0.075 (Ellingwood et al. (1980)) and all are time independent. The wind
load effect given in Equation (5-9) is rewritten as
( )
2 2 2
V X V X GC cE W
W v p z
= = (5-10)
where X
W
is the modeling uncertainty, so the wind load model is now represented in two
parts, time dependent and independent r.v.s, and a bias of 1.04 and a COV of 0.27 is
calculated for X
W
by following the procedure described below. Since time dependent part
has been separated, the independence property can be used in Ferry-Borges Method, and
the probability of failure can be evaluated at any time during the lifetime of the structure.


114
In order to use the chosen method of solution efficiently for Ferry-Borges Method
(using either Nested Reliability Analysis or Numerical Integration) the time independent
random variables are combined together. The multiplication of several normally
distributed random variables approach to a lognormal distribution. A FORM
approximation is used to approximate the multiplication of the random variables; a limit
state function as g=X
W
- cE
z
GC
p
X
2
v
is used to evaluate (1-P
f
) for different values of X
W
,

and gives the CDF of cE
z
GC
p
X
2
v
. Calculated safety index, (corresponds to standard
variate, S) vs. X
W
are plotted (plot is the CDF of X
W
) in a lognormal probability paper as
given in Figure 12. The plot is close to a linear line; therefore, the assumption of having a
lognormal distribution for X
W
is accurate.
The median of the X
W
is 1.00 and the X
W
value with a 0.84 probability of being
less than is 1.30, and the parameters (mean of ln(X
W
)) and (standard deviation of
ln(X
W
)) can be calculated by using
2624 . 0
00 . 1
30 . 1
ln ln
0 1 ln ln
84 . 0
=
|
.
|

\
|
=
|
|
.
|

\
|
=
= = =
m
m
W
W
W
x
x
x

(5-11)
and the mean and standard deviation for X
W
can be calculated as
( ) ( ) ( ) ( ) 28 . 0 1 2 ^ 2624 . 0 exp 2 ^ 04 . 1 1 exp
04 . 1 2624 . 0
2
1
0 exp
2
1
exp
2 2
2 2
= = =
=
|
.
|

\
|
+ =
|
.
|

\
|
+ =


(5-12)
so, the modeling uncertainties has a mean of 1.04 and a standard deviation of 0.28.



115


Figure 12 Modeling uncertainties, X
W
for wind load

Ellingwood et al. (1980) used seven sample city annual wind speed statistics
given in Table 27. The design wind speed is determined as a 3 second gust speed at 33 ft
above the ground with a exceedance probability of 0.02 which corresponds to a 50 year
mean recurrence interval (ASCE 7-95); for given sites in Table 27 a design speed of 90
mph is used. The mean values in Table 27 are divided by the nominal wind speed and the
wind speed is presented with the random variable V/V
n
that has a bias of 1.0. In order to
show the application of the calibration process the average value for the mean annual


116
wind speed to nominal wind speed V/V
n
and COV are used. The assumption of the wind
speeds following an Extreme Type I distribution is also used herein.
In the study done by Belk and Bennett (1991), it is stated that the mean rate of
occurrence is 0.0176 per hr that corresponds to 152 events in a year. The occurrence
duration for the windstorms ranges 3.76 hrs to 5.40 hrs, where the former duration is for a
single storm and the latter is for a cluster of storms. In this study the mean individual
windstorm duration is assumed as 4 hrs. So in load combination analysis a constant 4 hr.
rectangular pulse will represent a single wind load event, and the expected number of
windstorms in a year is 152.

Table 27 Wind speed statistical data (Ellingwood et al.(1980))
Site
V

V
/ V
n
COV
Baltimore, MD 55.9 0.62 0.12
Detroit, MI 48.9 0.54 0.14
St. Louis, MO 47.4 0.53 0.16
Austin, TX 45.1 0.50 0.12
Tucson, AZ 51.4 0.57 0.17
Rochester, NY 53.5 0.59 0.10
Sacremento, CA 46.0 0.51 0.22
Average 49.7 0.55 0.15


In load combination analysis when the wind load effect is combined with another
time dependent load the Ferry-Borges Method is used and this method requires the single
event statistics, which lasts 4 hours. The 4 hour distribution is not approximated
explicitly but using the advantage of the wind events independency, the distribution can


117
be evaluated inside the Ferry-Borges Method. At a design point v*, first, annual pdf and
cdf are calculated for 1 year using the distribution for 1 year (f
1year
(v*) and F
1year
(v*))and
then the single event can be back calculated from these values as
( ) ( ) ( )
( ) ( ) ( ) ( ) * *
152
1
*
* *
1
1
152
1
1 1
152
1
1 1
v f v F v f
v F v F
year year event
year event
|
.
|

\
|

=
=
(5-13)
The statistical data used in Reliability Analysis can be tabulated as in Table 28, V
is the random variable for V/V
n
and it is the average for seven sites.

Table 28 Wind load statistical data
Random Variable Distr. Type Bias COV
X
W
Lognormal 1.04 0.28
V Extr. Type I 0.55 0.15


5.3.5 Earthquake Load Model
In AASHTO (1994) the earthquake loads are defined as the horizontal force
effects determined by the elastic response coefficient, C
sm
and it is adjusted by a response
modification factor, R. The elastic seismic response coefficient is given by
A
T
AS
C
m
sm
5 . 2
2 . 1
3 / 2
= (5-13)


118
where T
m
is the m
th
vibration mode period in seconds, A is the acceleration coefficient and
obtained by using the maps provided by NEHRP, S is the site coefficient and takes into
account the effect of the soil type on C
sm
. The load effect is then
m
sm
R
W C
E = (5-14)
where W is the total or effective weight of the structure. The response modification
factor, R
m
accounts for the system behavior under a dynamic load. The advantageous
effects of ductility and redundancy are accounted through, R
m
; therefore, in reliability
analysis the safety index obtained from load combinations involving earthquake
corresponds to the system safety rather than the element safety.
The acceleration coefficient, A is the main parameter that defines the probability
distribution of the earthquake force effect. The design acceleration used herein is the one
having a 10% probability of being exceeded in 50 years, which means an event having a
475 years mean recurrence rate. This recurrence rate corresponds to essential bridges. For
crucial bridges which are those which must stay open to traffic after an earthquake, a
design earthquake with a longer return period is required; for such crucial bridges a 2500-
year return period event is used. The acceleration coefficients can be determined from the
maps provided by NEHRP(1988). The data for these maps contains acceleration
coefficients for different return period events with a magnitude over 4 in Richter scale,
which believed may cause structural damage (Ellingwood and Rosowsky (1996)). In
order to clarify the load model a location is chosen; let us choose San Francisco, where
the Peak ground Acceleration (PGA) for 10% probability of exceedance in 50 year is


119
53% g. The maximum effective peak ground acceleration for 50 year period follows a
Type II probability distribution function (Ellingwood and Rosowsky (1996)).

|
|
.
|

\
|

=
x
e X F ) (
max
(5-15)
where is the shape parameter and is the characteristic extreme value. The 0.53%g
value can be used to obtain for an assumed in Equation (5-15). In San Francisco the
probability of an PGA of 0.53g is 10% so the F
max
(0.53g)=0.90. The is between 2.3 to
3.3, where latter is the value used for the Western U.S., where the earthquakes are mostly
associated with faults (Ellingwood and Rosowsky (1996)). For San Francisco, value of
3.3 would be appropriate; therefore, for X=0.53g the mean, value is
268 . 0 )) 9 . 0 ln( ( * 53 . 0 )) ( ln (
3 . 3
1 1
max
= = =

X F x (5-16)
Then, the maximum distribution for the 50 years can be written as
(
(

|
.
|

\
|

=
3 . 3
268 . 0
max
) (
x
e x F (5-17)
The earthquakes are assumed to follow a Poisson model, which would relate time
to probability distribution. According to Melchers (1999) the probability of maximum
distribution of a Poissonian event can be written in terms of an individual event as
follows;
( ) ( ) [ ] x F T
X
e x F

=
1
max
) (

(5-18)


120
where is the earthquake occurrence rate and F
x
(x) is the individual earthquake intensity
distribution, which can be written in terms of maximum event distribution in T years.
Rearranging Equation (5-18) allows obtaining single earthquake intensity distribution
( ) ( ) [ ] x F
T
x F
X max
ln
1
1

+ = (5-19)



The rate of occurrence is assumed to be 0.4/year (1 in 2.5 years) for San
Francisco. So, the using Equation (5-19), the Fx(x) is as in Table 29. The data in Table 29
can be fitted by an Extreme Type II distribution (see Figure 13) with characteristic value
of 0.12 and a slope of 3.92. So the statistical parameters can be calculated as in Equation
(5-20).
43 . 0
15 . 0
0647 . 0
28 . 0
53 . 0
15 . 0
53 . 0 nominal 0647 . 0 15 . 0 mean
92 . 3 12 . 0
1
1
2
1
1
1
) exp( ) ( .
2 2 2
= = = =
= = =
= =
(

|
.
|

\
|

|
.
|

\
|
=
|
.
|

\
|
=
|
|
.
|

\
|
=

COV BIAS
k and
k k
k
mean
x
x F distr assumed
n
n n n
n
k
n

(5-20)





121
Table 29 Distribution for a single earthquake event
Acc./g F
x
(x)
0.15 0.660599
0.20 0.868654
0.25 0.937105
0.30 0.965540
0.35 0.979280
0.40 0.986664
0.45 0.990959
0.50 0.993614
0.55 0.995338
0.60 0.996501
0.65 0.997313
0.70 0.997896
0.75 0.998325
0.80 0.998646
0.85 0.998892
0.90 0.999082
0.95 0.999232
1.00 0.999352






122
Figure 13 Extremal Type II fit to single earthquake event

Ghosn et al. (2001) estimates that the response modification factor, R
m
has a COV
of 34%, and a according to Priestly and Park (1987) the actual R
m
is about 1.5 higher than
the code defined ductility levels, therefore a bias of 1.5 is used for R
m
and a normal
distribution is assumed. Since A (acceleration) will dominate the equation, therefore A
will be used as the time variant part and following Ellingwood et al. (1980) a 20% COV
is assumed for modeling uncertainty. The earthquake load model is
m
E E
R
Q X
E
1
= (5-21)
and the statistical data are tabulated in Table 30.



0.1
1
0 2 4 6 8
Standard Extremal Variate S
G
P
A
*

g
=0.12
0.2


123

Table 30 Earthquake Load Statistical Data
R.V. DISTR. BIAS COV
Q
E(1 event)
Extreme Type I 0.28 0.43
X
E
Normal Distr. 1.00 0.20
R Normal Distr. 1.50 0.34



5.4 Cost Data
The total cost model includes the marginal cost slope and the cost of failure.
Although the first one can be obtained by means of cost analysis on bridges, the cost of
failure is hard to estimate. Cost of failure estimation is rather a political issue than a
technical one, because inclusion of intangible costs such as loss of life, injury, etc. brings
a lot of debate.
Study conducted on bridges by Moses (1989) showed that on the average, a 25%
change in the live load factor produces a change of 2% in initial cost of bridges. This data
is used to calculate the cost slope for the live load effect and assumed to represent
gravitational cost slope. For lateral load effects there is no data to conclude about the
marginal cost slope, therefore in cost analysis the marginal cost slope for lateral loads are
represented as a factor of gravitational marginal cost slope, and analysis are repeated for
different ratios namely {0.5 1.0 2.0 3.0}.


124
As mentioned above, the failure cost is the most ambiguous part of the data.
Since, still there is no consensus on the estimations used by Kanda (1996), Ang and Leon
(1997), Lee et al. (1998) and Ellingwood (1999). Instead of using the estimates an
approach that deduces the cost of failure from the agreed safety level already in the code.
It is believed that a code body can agree on the load factors and safety levels obtained by
using the approach presented herein. The selection is aided with the sensitivity analysis
presented in next chapter, for changing parameters such as real nominal interest rate, j,
gravitational load cost factor, K
G
, the ratio of wind load cost ratio to gravitational load
ratio, design life, T etc.

5.5 Probabilistic Total Cost Model
The approach used in this research is actually two stages; first, to assume the most
basic load combination cases safety level is a satisfactory and therefore, deduce the cost
of failure from that basic case. Second, introduce that deduced value to other load
combination case and obtain consistent load factors and safety levels for these cases.
Among the load cases given in Equation (5-3), the case involving dead and live
load is the one that would be mostly agreed on; therefore, it is used to deduce the
information for failure cost. This allows using the current code as a reference and passing
the already available successful practice to the new calibration.
The live load data are two sets, one for single event and another for the side-by-
side occurrence. In order to decide which one of these sets to use in the calibration


125
process; the dead and live load case is run with the code prescribed load factors of 1.25
and 1.75 for dead and live load respectively, and the resistance factor, is assumed as
1.00. For illustration purposes, single lane loading is selected.
In his study Nowak (1993) used a target component reliability index of 3.5.
Therefore, it is also assumed herein that for this load combination case the total cost is
minimized at the safety index of 3.5. According to AASHTO LRFD Specification design
resistance can be calculated by
( )( ) IM HL m D R
L n D n
+ + = 1 93 (5-22)
where m is the multiple presence factor, and single lane load it is equal to 1.2, HL93 is
the static effect (specified as L
n
) obtained using standard design truck and lane loading,
and IM is the dynamic amplification factor and it is assumed as 33%. The IM value is
applied only to static load effect for design truck, not for lane loading. Therefore a
reduction factor is used to simulate that, and this reduction factor is calculated by
averaging the safety index value along the equally likely design samples given as
L
n
/D
n
={0.5 1.0 1.5 2.0}. The factor is found be 0.92, therefore design resistance can be
rewritten as
( ) ( )( ) ( ) 93 47 . 1 1 93 92 . 0 HL D IM HL m D R
L n D L n D n
+ = + + = (5-23)
then the cost factor is;
( ) ( )( )
( )

+
|
|
.
|

\
|
+
+
+ =
ij
fi
L L D
L L L D D
G
e
P
g K TCF
0 0
0 0
47 . 1
47 . 1
1


(5-24)


126
where K
G
is the normalized cost factor for gravitational load (dead, live loads), and
L
is
the nominal live load to nominal dead load ratio taken herein in the range of 0.5 to 2.0 to
cover the possible ranges of span lengths. As the span length decreases the nominal live
load to dead load ratio increases. The probability of failure P
fi
is calculated using the limit
state function
G=R-D-L (5-25)
through FORM analysis. The information by Moses (1989) mentioned in the previous
section is used in Equation (5-25) to determine the normalized cost factor for the
gravitational load, K
G
;
( ) ( )( )
( )
( )( )
( )
0 0 0 0 0
0 0 0
47 . 1
47 . 1
47 . 1
47 . 1
C
C
K K
I
L L D
L L
G
L L D
L L L D D
G

=
|
|
.
|

\
|
+

=
|
|
.
|

\
|
+
+




(5-26)
( )
( )( )
0
0 0
47 . 1
47 . 1
C
C
K
I
L L
L L D
G

|
|
.
|

\
|

+
=


(5-27)
where
DO
and
LO
are the current code defined values as 1.25 and 1.75 for dead and live
load respectively.
L
is a change in the live load factor and C
I
is the change in the cost
due to that, and
L
={0.5 1.0 1.5 2.0}. The value of K
G
is calculated for given values of

L
; a change in live load factor as shown by
L
=0.25
L0
, has a corresponding initial cost
change represented by C
I
=0.02. Averaged along the range of
L
, the factor K
G
is
calculated as 0.12 (see Table 31). Kanda and Ellingwood (1991) give the marginal cost
factor as approximately 0.2, assuming a central safety factor of 3.0. It is believed that this
0.2 value does not distinguish between the gravitational and lateral loads, therefore it is
assumed that it is on the average, 0.12 would be accurate enough.


127

Table 31 Calculated K
G
values
K
G
0.50 0.16
1.00 0.12
1.50 0.11
2.00 0.10
Average= 0.12


An important assumption is that the total cost is a minimum for this load
combination using Linds Postulate. Thus, differentiating TCF in Equation (5-24) with
respect to the dead and live load factors about the current target safety index (3.5) should
be equal to zero due to the assumption that the current load factors and the target safety
level are optimal. Therefore, equating the derivative of TCF to zero gives the optimum
failure cost ratio, g as (see Equation 5-24).
( ) ( )
G
L L D
ij
fi
L
L G
L L D
ij
fi
D
K
e
P
K
e
P
g
0 0
1
0 0
1
47 . 1 47 . 1

+
(

|
|
.
|

\
|

=
+
(

|
|
.
|

\
|

=


(5-28)
The failure cost ratio, g values deduced using Equation (5-28) along with the code
specified
D0
=1.25 and
L0
=1.75 are averaged for all the values and found equal to 30
(see Table 32). The same procedure is applied also considering the system safety
assuming the system safety index is equal to component safety index plus 1.0. In the
evaluation of annual probability of failure in Equation (5-28) the same approximated
system safety index values are use. The deduced value increases to 1727 (see Table 33)


128
when system safety is considered. Both these deduced values for g are used in calibration
analysis. When g=30 is used the component safety is evaluated to obtain annual
probability of failure. When g=1727 is used the system safety index (component safety
index plus one) is used to evaluate the annual probability of failure. Load combination
with earthquake is an exception in that the safety index evaluated in earthquake analysis
corresponds to system safety. Therefore g=1727 needs to be used in optimization of
earthquake case without increasing the safety index by one.

Table 32 Deduced failure cost factor, g values (component safety)
g
0.50 30.81 3.56
1.00 32.09 3.52
1.50 29.50 3.47
2.00 27.67 3.44
Average= 30 3.50


Table 33Deduced failure cost factor, g values (system safety)
g
0.50 1871 4.56
1.00 1881 4.52
1.50 1653 4.47
2.00 1505 4.44
Average= 1727 4.50


The determined K
G
value as 0.12 and the deduced g values as 30 and 1727 for
component and system safety, respectively are used for the rest of the load combinations


129
in Equation (5-3). Doing so helps to maintain consistency among the load combination
cases because the failure consequences for all the load combination cases are assumed to
be the same. This does not give a uniform safety index for all the combination cases but it
is consistent because for every load combination a failure would result in the same cost
consequences. This optimization is carried out in the next sections.
5.5.1 Strength III: Combination of Dead and Wind Load
The strength III limit state is for combination of the dead and wind load. All the
elements and connections have to satisfy the design check
n W n D n
W D R + (5-29)
where D
n
is equal to 1.0 and W
n
is a ratio of D
n
. The limit state function for this load
combination case is
n
n
W
n
n
D
W
V X D
D
R
R G
2
) 1 ( = (5-30)
All the random variables are normalized to their nominal values and therefore nominal
values of the random variables are 1.0. Current load factors are 1.25 and 1.40 for dead
load and wind load, respectively. Normalized initial cost can be written as follows for this
combination case;
( ) ( )
|
|
.
|

\
|
+
+
+ =
W
W W W D
G
I
K
C
C


40 . 1 25 . 1
40 . 1 25 . 1
1
0
(5-31)
where
W
=K
W
/K
G
and K
W
is the cost factor for wind load, and
W
=W
n
/D
n
. Introducing
the deduced g value of 30 into the TCF equation into Equation (5-32)


130
( ) ( )

+
|
|
.
|

\
|
+
+
+ =
ij
fi
W W
W W W D
G
e
P
g K TCF


140 25 . 1
40 . 1 25 . 1
1 (5-32)
The optimization process is repeated for different values of marginal wind cost slope to
marginal gravitational load cost factor,
W
along the design space namely
W

=W
n
/D
n
={0.50 1.00 2.00 5.00}. The limit state function in Equation (5-30) along with the
statistical data given in Table 34 is used in FORM analysis. The safety index and total
cost factor values are provided for current code defined load factor values in Table 35 for
information. Optimized load factors and safety indices for
W
are tabulated in Table 36
for g=30 and in Table 38 for g=1727. As the marginal cost slopes of wind load become
higher the dead load factor increases because it becomes cheaper to use high dead load
factor to increase the safety. The dead load factor 1.25 is same in most AASHTO load
combinations, therefore calibration analysis are also run by keeping the dead load factor
at the code specified 1.25 value for both component and system safety (g=30 and
g=1727).
In Table 36, as the wind cost factor become expensive the dead load factor
increases and especially for a
W
=3.0 optimum dead load factor is 2.36, which is high.
The reason for having dead load factor so high is due to lowering the expected failure
cost for high W
n
/D
n
values. In Table 37, dead load factor is fixed at 1.25; and the wind
load factor drops from 1.88 to 1.49 as the marginal cost slope increase from 0.5 to 3.0.
Assuming
W
=3.0 would lead to an optimized wind load factor of 1.50.
If system safety is introduced into the calibration analysis a value of 1727 is used
for the deduced g value. The optimum load factors obtained considering system safety are


131
tabulated in Table 38. The scattering in safety index values is the main reason for an
increase in optimum load factors. Since high values of W
n
/D
n
governs the optimization, in
order to increase the safety indices (to lower expected failure cost) the wind load factor
needs to be increased. However, when it is more costly to increase the wind load factor
(such as for
W
=3.0) the dead load factor is increased instead. The amount of increase
required for the dead load factor is more than if a higher wind load factor is used. In
order to reach a recommended value, the dead load factor is fixed at the current code
specified value of 1.25 and optimization is repeated considering system safety. Results
are presented in Table 39. For the marginal wind load to gravitational load cost ratio,
W

equal to 2.0, the optimum wind load factor is 1.67 and an average system safety index of
4.05 is attained. For
W
equal to 3.0 the optimum wind load is 1.60 and corresponding
system safety index is 3.96. Assuming a
W
of 3.0 an optimum wind load factor of 1.60 is
appropriate to conclude as a final value. Therefore, the load factor for combination of
dead plus wind load is 1.25 and 1.60 for dead and wind load, respectively with an
average system safety of about 3.95 and component safety of 2.95. The sum of the total
cost factor decreases from 4.36 to about 4.24, a 3% decrease by changing wind load
factor from the current 1.40 to 1.60.






132



Table 34 Statistical data for D+W case
R.V. Distr. BIAS COV
R Logn. 1.12 0.10
D Norm 1.03 0.08
X
W
Norm 1.04 0.28
V Extr. I 0.55 0.15



Table 35 Current Code safety levels and total cost factors for D+W
Component System
g=30 g=1727

W

comp

sys

D

W
TCF TCF
0.5 2.95 3.95 1.25 1.40 1.0197 1.0241
1.0 2.73 3.73 1.25 1.40 1.0388 1.0578
2.0 2.57 3.57 1.25 1.40 1.0630 1.1088
5.0 2.45 3.45 1.25 1.40 1.0889 1.1710
Average = 2.68 3.68 TCF= 4.2104 4.3618








133

Table 36 Optimized component safety index,
comp
and load factors,
D
and
W
for D+W
for g=30.

W
=K
W
/K
G
W
n
/D
n

comp

D

W
TCF
0.50 0.50 3.04 0.83 2.30 1.00286
1.00 3.34 0.83 2.30 1.00661
2.00 3.45 0.83 2.30 1.01776
5.00 3.51 0.83 2.30 1.02949

avg
= 3.34
TCF=
4.05673
1.00 0.50 3.22 1.17 1.78 1.01464
1.00 3.13 1.17 1.78 1.02423
2.00 3.05 1.17 1.78 1.03403
5.00 2.99 1.17 1.78 1.04339

avg
= 3.10
TCF=
4.11629
2.00 0.50 3.72 1.59 1.43 1.02392
1.00 3.21 1.59 1.43 1.02589
2.00 2.86 1.59 1.43 1.03891
5.00 2.60 1.59 1.43 1.06550

avg
= 3.10
TCF=
4.15422
3.00 0.50 4.61 2.36 1.13 1.04375
1.00 3.70 2.36 1.13 1.01521
2.00 2.98 2.36 1.13 1.00250
5.00 2.37 2.36 1.13 1.06879

avg
= 3.41
TCF=
4.13025






134

Table 37 Optimized component safety index,
comp
and load factors,
D
=1.25 and
W
for
D+W for g=30.

W
=K
W
/K
G
W
n
/D
n

comp

D

W
TCF
0.50 0.50 3.50 1.25 1.88 1.01025
1.00 3.34 1.25 1.88 1.01604
2.00 3.22 1.25 1.88 1.02199
5.00 3.14 1.25 1.88 1.02779

avg
= 3.30
TCF=
4.07607
1.00 0.50 3.34 1.25 1.73 1.01520
1.00 3.16 1.25 1.73 1.02433
2.00 3.03 1.25 1.73 1.03405
5.00 2.94 1.25 1.73 1.04364

avg
= 3.12
TCF=
4.11722
2.00 0.50 3.17 1.25 1.58 1.02027
1.00 2.98 1.25 1.58 1.03387
2.00 2.83 1.25 1.58 1.04909
5.00 2.73 1.25 1.58 1.06448

avg
= 2.93
TCF=
4.16772
3.00 0.50 3.07 1.25 1.49 1.02184
1.00 2.86 1.25 1.49 1.03828
2.00 2.71 1.25 1.49 1.05755
5.00 2.60 1.25 1.49 1.07744

avg
= 2.81
TCF=
4.19511



135

Table 38 Optimized system safety index,
sys
and load factors,
D
and
W
for D+W for
g=1727.

W
=K
W
/K
G
W
n
/D
n

sys

D

W
TCF
0.50 0.50 4.12 0.88 2.27 1.00245
1.00 4.36 0.88 2.27 1.00700
2.00 4.45 0.88 2.27 1.01757
5.00 4.49 0.88 2.27 1.02853

avg
= 4.35
TCF=
4.05555
1.00 0.50 4.26 1.15 1.85 1.01415
1.00 4.19 1.15 1.85 1.02439
2.00 4.13 1.15 1.85 1.03497
5.00 4.08 1.15 1.85 1.04521

avg
= 4.16
TCF=
4.11872
2.00 0.50 4.66 1.49 1.57 1.02598
1.00 4.25 1.49 1.57 1.03247
2.00 3.98 1.49 1.57 1.04768
5.00 3.79 1.49 1.57 1.07393

avg
= 4.17
TCF=
4.18006
3.00 0.50 5.41 2.10 1.33 1.04640
1.00 4.64 2.10 1.33 1.03064
2.00 4.08 2.10 1.33 1.02735
5.00 3.62 2.10 1.33 1.08746

avg
= 4.44
TCF=
4.19184




136

Table 39 Optimized system safety index,
sys
and wind load factor,
W
for D+W for

D
=1.25 and g=1727.

W
=K
W
/K
G
W
n
/D
n

sys

D

W
TCF
0.50 0.50 4.53 1.25 1.92 1.00971
1.00 4.37 1.25 1.92 1.01540
2.00 4.26 1.25 1.92 1.02148
5.00 4.18 1.25 1.92 1.02762

avg
= 4.34
TCF=
4.07422
1.00 0.50 4.41 1.25 1.73 1.01523
1.00 4.24 1.25 1.73 1.02462
2.00 4.11 1.25 1.73 1.03499
5.00 4.02 1.25 1.73 1.04567

avg
= 4.19
TCF=
4.12051
2.00 0.50 4.28 1.25 1.67 1.02245
1.00 4.09 1.25 1.67 1.03739
2.00 3.96 1.25 1.67 1.05470
5.00 3.86 1.25 1.67 1.07304

avg
= 4.05
TCF=
4.18757
3.00 0.50 4.20 1.25 1.60 1.02689
1.00 4.01 1.25 1.60 1.04605
2.00 3.87 1.25 1.60 1.06909
5.00 3.77 1.25 1.60 1.09450

avg
= 3.96
TCF=
4.23653







137
5.5.2 Strength V: Combination of Dead, Live and Wind Load
The strength V limit state is for combination of dead load, wind load and
vehicular live load. All the elements and connections have to satisfy the design check
n W n L n D n
W L D R + + (5-33)
Current load factors are 1.25, 1.35 and 0.4 for dead load, live load and wind load,
respectively. Normalized initial cost can be written as follows for this combination case;
( ) ( )( ) ( )
( )
|
|
.
|

\
|
+ +
+ +
+ =
W L
W W W L L D
G
I
K
C
C


4 . 0 47 . 1 35 . 1 25 . 1
40 . 0 35 . 1 47 . 1 25 . 1
1
0
(5-34)
where
W
and
W
are the wind load marginal cost slope to gravitational load cost slope
ratio (K
W
/K
G
) and nominal wind load to nominal dead load ratio (W
n
/D
n
), respectively
and
L
is the nominal live load to nominal live load ratio (L
n
/D
n
=(HL93 effect)/D
n
).
Introducing the deduced g value of 30 into the TCF equation as
( ) ( )( ) ( )
( )

+
|
|
.
|

\
|
+ +
+ +
+ =
ij
fi
W L
W W W L L D
G
e
P
g K TCF


40 . 0 47 . 1 35 . 1 25 . 1
40 . 0 35 . 1 47 . 1 25 . 1
1 (5-35)
the optimization process is repeated for different values of
W
=K
W
/K
G
along the design
space namely
L
=L
n
/D
n
={0.5 1.0 1.5 2.0} and
L
=W
n
/D
n
={0.5 1.0 1.5} (the range for
this case is different than the previous combination case because in AASHTO it is
assumed that the over the wind speeds 55 mph there would be no truck loading on bridge)
and the limit state function in Equation (5-36) along with the statistical data given in
Table 40 is used in FORM analysis. It is assumed that the maximum load effect would
occur when wind is on.


138
|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|
=
n
n
W
n
n
L L D
n
n
D
W
V X
D
L
Q X I D
D
R
R G
2
(5-36)
where all the random variables given have a nominal value of 1.0. Since there are two
time dependent loads Ferry-Borges process (Equation (2-14)) described in Chapter 2
should be used for this limit state. The Ferry-Borges Method along with the Nested
Reliability Analysis is used to find the probability of failure (Equation (2-28)). Using the
Equation (2-28) the limit state function for D+W+L combination case can be written as
in Equation (5-37).

Table 40 Statistical data for D+W+L case
R.V. Distr. BIAS COV
R Logn. 1.12 0.10
D Normal 1.03 0.08
X
L
Logn. 1.00 0.18
I
D
Normal 1.13 0.10
Q
L
Normal 0.79 0.10
X
W
Norm 1.04 0.28
V Extr. I 0.55 0.15


( )
W L D
X X I u h , , + = (5-37)
where u is a standard normal variate with zero mean and unit standard deviation, and
is a function of the dynamic amplification and modeling factor for live load and modeling
factor wind load and a limit state function of
2 * * *
V x Q i x D R g
W L L
D
= (5-38)


139
where star values are design values (most probable values) for the solution of limit state
function, h. Since the solution is iterative for every star value Equation (5-38) is solved to
give using the Ferry-Borges Method (Rackwitz-Fiessler algorithm).
First, the expected total cost is calculated along with safety indices for current
code values as given in Table 41; there are 12 sample designs used to represent the design
space. Since the current code load factor values used as the reference load factors in cost
model and expected total cost function is normalized with base cost, C
0
, the initial cost
factors for each individual design sample are equal to 1.0. Therefore, the sum of initial
cost factors is equal to 12.0 for design space and the sum of expected failure cost factor,
FCF is 3.95, which makes the sum of expected total cost factor, TCF 15.95. Most of the
contribution to FCF is from the samples where W
n
/D
n
is higher than L
n
/D
n
values. This is
very clear especially when W
n
/D
n
is 1.5 and L
n
/D
n
is 0.50, this sample design has the
lowest safety index and nearly 50% of FCF is due to this sample alone. It is easy to
estimate that during the optimization process load factors need to be increased in order to
lower FCF and consequently TCF.
Deduced cost of failure factor, g=30 for component safety is used to optimize the
TCF for the code. In Table 42, the results for optimum safety index, and load factors
are presented considering component safety. Optimization is run for marginal wind load
cost slope raging from one half to 3 times the marginal gravitational cost slope. The
reason having
W
=0.50 and
W
=1.0 is to clearly show the trend in load factors. Real
interest of the calibration is focused on the values of
W
=2.0 and
W
=3.0.


140
Optimum load factors for
W
=2.0 are found to be 0.90 and 1.56 for wind and live
load, respectively, but when
W
=3.0 the live load factor jumps to 2.20 where wind load
factor is 0.43. When increasing wind load factor becomes expensive the live load factor
increases. Since the governing cases are the ones where nominal wind load is higher than
live load, this increase need to be relatively high. The reason for that is the effect of live
load factor on safety index for that sort of design samples is relatively low.
When cost optimization is repeated considering system safety the same trend as in
component safety is also seen. Even in
W
=3.0 case, the wind load factor falls below the
current specified wind load factor of 0.40 but the live load takes a value of 3.0 which is
very high. In order to avoid this kind of jump the number of the sample designs in the
design space has to be increased but this would affect the efficiency of the method
explained herein, As the number of sample designs increases the time required for
analysis also increases. The next best thing to do is to keep the live load factor at a
constant value during the optimization process. The live load factor is fixed at current
code specified load factor of 1.35 and the optimization analyses are repeated both for
component and system safety with results tabulated in Table 43 and Table 45. When
component safety is considered, optimum wind load factors are found to be 1.08 and 0.99
for
W
of 2.0 and 3.0, respectively. It is more realistic to consider system safety in order
to conclude on the calibrated load factors.
W
is assumed to be 3.0, so the wind load
factor can be chosen as 1.10. Therefore the optimum load factors are 1.25, 1.10 and 1.35
for dead, wind and live loads, respectively. The optimum load factors reduces the


141
expected total cost factor 55%, from current codes 29.2 to optimum expected total cost
of 13.0.



Table 41 The safety index and cost values for D+W+L case using current code load
factors
W
n
/D
n
Ln/Dn
comp.

sys.

D

W

L
ICF FCF TCF FCF TCF
0.50 0.50 2.51 3.51 1.25 0.40 1.35 1.0000 0.0964 1.0964 0.1803 1.1803
1.00 2.86 3.86 1.25 0.40 1.35 1.0000 0.0449 1.0449 0.0640 1.0640
1.50 2.79 3.79 1.25 0.40 1.35 1.0000 0.0607 1.0607 0.0939 1.0939
2.00 2.69 3.69 1.25 0.40 1.35 1.0000 0.0831 1.0831 0.1413 1.1413
1.00 0.50 1.68 2.68 1.25 0.40 1.35 1.0000 0.6678 1.6678 2.5419 3.5419
1.00 2.31 3.31 1.25 0.40 1.35 1.0000 0.1565 1.1565 0.3470 1.3470
1.50 2.62 3.62 1.25 0.40 1.35 1.0000 0.0792 1.0792 0.1366 1.1366
2.00 2.74 3.74 1.25 0.40 1.35 1.0000 0.0648 1.0648 0.1029 1.1029
1.50 0.50 1.15 2.15 1.25 0.40 1.35 1.0000 1.8702 2.8702 11.0876 12.0876
1.00 1.80 2.80 1.25 0.40 1.35 1.0000 0.5133 1.5133 1.7660 2.7660
1.50 2.21 3.21 1.25 0.40 1.35 1.0000 0.2005 1.2005 0.4843 1.4843
2.00 2.46 3.46 1.25 0.40 1.35 1.0000 0.1144 1.1144 0.2238 1.2238

avg
= 2.32 3.32
=
12.0000 3.9518 15.9518 17.1696 29.1696

min
= 1.15 2.15

max
= 2.86 3.86
System Safety Component Safety
g=30 g=1727








142

Table 42 Optimized component safety index,
comp
and load factors,
W,
and
L
for
D+W+L (
D
=1.25 g=30).

W
W
n
/D
n Ln/Dn
D

W

L ICF FCF TCF
0.50 0.50 0.50 3.77 1.25 1.65 1.23 1.01100 0.00100 1.01200
1.00 3.64 1.25 1.65 1.23 1.00460 0.00290 1.00750
1.50 3.22 1.25 1.65 1.23 1.00110 0.01450 1.01570
2.00 2.94 1.25 1.65 1.23 0.99890 0.03720 1.03620
1.00 0.50 3.46 1.25 1.65 1.23 1.02440 0.00340 1.02780
1.00 3.68 1.25 1.65 1.23 1.01470 0.00140 1.01620
1.50 3.75 1.25 1.65 1.23 1.00920 0.00140 1.01070
2.00 3.59 1.25 1.65 1.23 1.00570 0.00350 1.00920
1.50 0.50 3.29 1.25 1.65 1.23 1.03590 0.00630 1.04220
1.00 3.49 1.25 1.65 1.23 1.02380 0.00300 1.02680
1.50 3.64 1.25 1.65 1.23 1.01660 0.00170 1.01840
2.00 3.72 1.25 1.65 1.23 1.01190 0.00130 1.01320

avg
= 3.52 = 12.15780 0.07760 12.23590
1.00 0.50 0.50 3.55 1.25 1.23 1.36 1.02080 0.00250 1.02330
1.00 3.62 1.25 1.23 1.36 1.01510 0.00310 1.01820
1.50 3.31 1.25 1.23 1.36 1.01190 0.01030 1.02220
2.00 3.11 1.25 1.23 1.36 1.00990 0.02120 1.03110
1.00 0.50 3.09 1.25 1.23 1.36 1.03810 0.01290 1.05090
1.00 3.45 1.25 1.23 1.36 1.02800 0.00350 1.03150
1.50 3.62 1.25 1.23 1.36 1.02220 0.00250 1.02460
2.00 3.56 1.25 1.23 1.36 1.01840 0.00390 1.02230
1.50 0.50 2.83 1.25 1.23 1.36 1.05290 0.02950 1.08240
1.00 3.16 1.25 1.23 1.36 1.03950 0.00990 1.04940
1.50 3.40 1.25 1.23 1.36 1.03160 0.00430 1.03580
2.00 3.55 1.25 1.23 1.36 1.02640 0.00270 1.02910

avg
= 3.35 = 12.31480 0.10630 12.42080
2.00 0.50 0.50 3.51 1.25 0.9 1.56 1.03220 0.00290 1.03510
1.00 3.82 1.25 0.9 1.56 1.02820 0.00140 1.02960
1.50 3.61 1.25 0.9 1.56 1.02600 0.00340 1.02940
2.00 3.46 1.25 0.9 1.56 1.02470 0.00580 1.03050
1.00 0.50 2.84 1.25 0.9 1.56 1.05260 0.02880 1.08130
1.00 3.41 1.25 0.9 1.56 1.04330 0.00410 1.04740
1.50 3.73 1.25 0.9 1.56 1.03800 0.00160 1.03950
2.00 3.78 1.25 0.9 1.56 1.03450 0.00170 1.03620
1.50 0.50 2.48 1.25 0.9 1.56 1.07010 0.08690 1.15700
1.00 2.99 1.25 0.9 1.56 1.05670 0.01780 1.07450
1.50 3.36 1.25 0.9 1.56 1.04890 0.00500 1.05390
2.00 3.62 1.25 0.9 1.56 1.04370 0.00190 1.04560

avg
= 3.38 = 12.49890 0.16130 12.66000
3.00 0.50 0.50 3.96 1.25 0.43 2.2 1.03290 0.00050 1.03340
1.00 4.83 1.25 0.43 2.2 1.04530 0.00000 1.04530
1.50 4.72 1.25 0.43 2.2 1.05210 0.00000 1.05220
2.00 4.67 1.25 0.43 2.2 1.05650 0.00000 1.05650
1.00 0.50 2.84 1.25 0.43 2.2 1.03240 0.02870 1.06110
1.00 3.88 1.25 0.43 2.2 1.04430 0.00070 1.04490
1.50 4.51 1.25 0.43 2.2 1.05100 0.00000 1.05110
2.00 4.95 1.25 0.43 2.2 1.05540 0.00000 1.05540
1.50 0.50 2.19 1.25 0.43 2.2 1.03200 0.18840 1.22050
1.00 3.19 1.25 0.43 2.2 1.04330 0.00900 1.05230
1.50 3.83 1.25 0.43 2.2 1.05000 0.00080 1.05080
2.00 4.30 1.25 0.43 2.2 1.05440 0.00010 1.05450

avg
= 3.99 = 12.54960 0.22820 12.77800





143
Table 43 Optimized component safety index,
comp
and wind load factor,
W
for D+W+L
(
D
=1.25,
L
=1.35 g=30).

W
W
n
/D
n Ln/Dn
D

W

L TCF FCF TCF
0.50 0.50 0.50 3.74 1.25 1.44 1.35 1.01270 0.00120 1.01390
1.00 3.76 1.25 1.44 1.35 1.00900 0.00180 1.01090
1.50 3.40 1.25 1.44 1.35 1.00700 0.00750 1.01450
2.00 3.17 1.25 1.44 1.35 1.00570 0.01710 1.02280
1.00 0.50 3.32 1.25 1.44 1.35 1.02350 0.00560 1.02920
1.00 3.65 1.25 1.44 1.35 1.01710 0.00170 1.01880
1.50 3.81 1.25 1.44 1.35 1.01340 0.00120 1.01460
2.00 3.71 1.25 1.44 1.35 1.01110 0.00220 1.01320
1.50 0.50 3.10 1.25 1.44 1.35 1.03280 0.01220 1.04500
1.00 3.39 1.25 1.44 1.35 1.02430 0.00440 1.02880
1.50 3.60 1.25 1.44 1.35 1.01930 0.00200 1.02130
2.00 3.74 1.25 1.44 1.35 1.01600 0.00120 1.01720

avg
= 3.53 = 12.19190 0.05810 12.25020
1.00 0.50 0.50 3.55 1.25 1.24 1.35 1.02070 0.00250 1.02330
1.00 3.60 1.25 1.24 1.35 1.01480 0.00330 1.01810
1.50 3.29 1.25 1.24 1.35 1.01140 0.01110 1.02250
2.00 3.08 1.25 1.24 1.35 1.00940 0.02300 1.03240
1.00 0.50 3.09 1.25 1.24 1.35 1.03830 0.01250 1.05090
1.00 3.45 1.25 1.24 1.35 1.02790 0.00350 1.03140
1.50 3.61 1.25 1.24 1.35 1.02190 0.00260 1.02450
2.00 3.54 1.25 1.24 1.35 1.01800 0.00410 1.02220
1.50 0.50 2.85 1.25 1.24 1.35 1.05350 0.02830 1.08180
1.00 3.17 1.25 1.24 1.35 1.03960 0.00980 1.04940
1.50 3.40 1.25 1.24 1.35 1.03150 0.00430 1.03580
2.00 3.55 1.25 1.24 1.35 1.02610 0.00280 1.02900

avg
= 3.35 = 12.31310 0.10780 12.42130
2.00 0.50 0.50 3.37 1.25 1.08 1.35 1.03360 0.00500 1.03850
1.00 3.47 1.25 1.08 1.35 1.02390 0.00550 1.02940
1.50 3.20 1.25 1.08 1.35 1.01850 0.01540 1.03390
2.00 3.01 1.25 1.08 1.35 1.01510 0.02950 1.04460
1.00 0.50 2.88 1.25 1.08 1.35 1.06200 0.02540 1.08740
1.00 3.27 1.25 1.08 1.35 1.04510 0.00690 1.05200
1.50 3.43 1.25 1.08 1.35 1.03540 0.00500 1.04040
2.00 3.40 1.25 1.08 1.35 1.02920 0.00710 1.03620
1.50 0.50 2.60 1.25 1.08 1.35 1.08650 0.05970 1.14620
1.00 2.96 1.25 1.08 1.35 1.06410 0.01970 1.08380
1.50 3.21 1.25 1.08 1.35 1.05090 0.00850 1.05940
2.00 3.36 1.25 1.08 1.35 1.04230 0.00570 1.04800

avg
= 3.18 = 12.50660 0.19340 12.69980
3.00 0.50 0.50 3.27 1.25 0.99 1.35 1.04340 0.00750 1.05090
1.00 3.39 1.25 0.99 1.35 1.03090 0.00740 1.03830
1.50 3.15 1.25 0.99 1.35 1.02400 0.01860 1.04260
2.00 2.97 1.25 0.99 1.35 1.01960 0.03400 1.05360
1.00 0.50 2.75 1.25 0.99 1.35 1.08030 0.03880 1.11910
1.00 3.16 1.25 0.99 1.35 1.05840 0.01030 1.06870
1.50 3.33 1.25 0.99 1.35 1.04590 0.00740 1.05320
2.00 3.31 1.25 0.99 1.35 1.03780 0.00960 1.04740
1.50 0.50 2.45 1.25 0.99 1.35 1.11190 0.09340 1.20540
1.00 2.83 1.25 0.99 1.35 1.08300 0.02990 1.11290
1.50 3.10 1.25 0.99 1.35 1.06590 0.01270 1.07860
2.00 3.25 1.25 0.99 1.35 1.05470 0.00860 1.06330

avg
= 3.08 = 12.65580 0.27820 12.93400




144
Table 44 Optimized system safety index,
sys
and load factors,
W,
and
L
for D+W+L
(
D
=1.25 g=1727).
W
W
n
/D
n
Ln/Dn D W L
ICF FCF TCF
0.50 0.50 0.50 4.60 1.25 1.26 1.37 1.01150 0.00140 1.01280
1.00 4.67 1.25 1.26 1.37 1.00880 0.00170 1.01040
1.50 4.36 1.25 1.26 1.37 1.00730 0.00780 1.01500
2.00 4.15 1.25 1.26 1.37 1.00630 0.01970 1.02610
1.00 0.50 4.14 1.25 1.26 1.37 1.02040 0.01110 1.03150
1.00 4.51 1.25 1.26 1.37 1.01540 0.00210 1.01750
1.50 4.68 1.25 1.26 1.37 1.01260 0.00120 1.01380
2.00 4.62 1.25 1.26 1.37 1.01070 0.00220 1.01290
1.50 0.50 3.89 1.25 1.26 1.37 1.02800 0.03220 1.06030
1.00 4.22 1.25 1.26 1.37 1.02130 0.00790 1.02930
1.50 4.46 1.25 1.26 1.37 1.01740 0.00270 1.02010
2.00 4.62 1.25 1.26 1.37 1.01480 0.00130 1.01610
avg
= 4.41 = 12.17450 0.09130 12.26580
1.00 0.50 0.50 4.61 1.25 1.29 1.36 1.02240 0.00130 1.02360
1.00 4.67 1.25 1.29 1.36 1.01620 0.00170 1.01790
1.50 4.35 1.25 1.29 1.36 1.01290 0.00810 1.02100
2.00 4.14 1.25 1.29 1.36 1.01070 0.02110 1.03180
1.00 0.50 4.16 1.25 1.29 1.36 1.04090 0.01000 1.05090
1.00 4.52 1.25 1.29 1.36 1.03010 0.00200 1.03200
1.50 4.69 1.25 1.29 1.36 1.02390 0.00120 1.02510
2.00 4.62 1.25 1.29 1.36 1.01990 0.00220 1.02210
1.50 0.50 3.92 1.25 1.29 1.36 1.05680 0.02820 1.08510
1.00 4.24 1.25 1.29 1.36 1.04240 0.00720 1.04960
1.50 4.47 1.25 1.29 1.36 1.03400 0.00250 1.03650
2.00 4.62 1.25 1.29 1.36 1.02840 0.00120 1.02960
avg
= 4.42 = 12.33860 0.08670 12.42520
2.00 0.50 0.50 4.58 1.25 1.05 1.51 1.03740 0.00150 1.03890
1.00 4.81 1.25 1.05 1.51 1.03060 0.00080 1.03150
1.50 4.56 1.25 1.05 1.51 1.02690 0.00300 1.02990
2.00 4.40 1.25 1.05 1.51 1.02450 0.00650 1.03100
1.00 0.50 3.99 1.25 1.05 1.51 1.06390 0.02140 1.08530
1.00 4.49 1.25 1.05 1.51 1.05030 0.00230 1.05260
1.50 4.76 1.25 1.05 1.51 1.04250 0.00080 1.04330
2.00 4.77 1.25 1.05 1.51 1.03740 0.00100 1.03850
1.50 0.50 3.67 1.25 1.05 1.51 1.08670 0.07880 1.16540
1.00 4.11 1.25 1.05 1.51 1.06790 0.01260 1.08040
1.50 4.43 1.25 1.05 1.51 1.05680 0.00300 1.05970
2.00 4.66 1.25 1.05 1.51 1.04950 0.00100 1.05050
avg
= 4.44 = 12.57440 0.13270 12.70700
3.00 0.50 0.50 5.69 1.25 0.30 3.00 1.05220 0.00000 1.05220
1.00 6.93 1.25 0.30 3.00 1.07950 0.00000 1.07950
1.50 8.09 1.25 0.30 3.00 1.09450 0.00000 1.09450
2.00 8.17 1.25 0.30 3.00 1.10410 0.00000 1.10410
1.00 0.50 4.41 1.25 0.30 3.00 1.04140 0.00340 1.04480
1.00 5.62 1.25 0.30 3.00 1.07010 0.00000 1.07010
1.50 6.36 1.25 0.30 3.00 1.08650 0.00000 1.08650
2.00 6.90 1.25 0.30 3.00 1.09720 0.00000 1.09720
1.50 0.50 3.64 1.25 0.30 3.00 1.03220 0.08620 1.11830
1.00 4.85 1.25 0.30 3.00 1.06180 0.00040 1.06220
1.50 5.59 1.25 0.30 3.00 1.07920 0.00000 1.07920
2.00 6.12 1.25 0.30 3.00 1.09070 0.00000 1.09070
avg
= 6.03 = 12.88940 0.09000 12.97930




145

Table 45 Optimized system safety index,
sys
and wind load factor,
W
for D+W+L
(
D
=1.25,
L
=1.35 g=1727).

W
W
n
/D
n Ln/Dn
D

W

L ICF FCF TCF
0.50 0.50 0.50 4.75 1.25 1.45 1.35 1.01290 0.00060 1.01360
1.00 4.77 1.25 1.45 1.35 1.00920 0.00100 1.01020
1.50 4.41 1.25 1.45 1.35 1.00710 0.00620 1.01340
2.00 4.18 1.25 1.45 1.35 1.00580 0.01790 1.02370
1.00 0.50 4.34 1.25 1.45 1.35 1.02390 0.00450 1.02840
1.00 4.66 1.25 1.45 1.35 1.01740 0.00100 1.01840
1.50 4.82 1.25 1.45 1.35 1.01370 0.00060 1.01430
2.00 4.72 1.25 1.45 1.35 1.01120 0.00130 1.01260
1.50 0.50 4.12 1.25 1.45 1.35 1.03330 0.01190 1.04530
1.00 4.41 1.25 1.45 1.35 1.02470 0.00340 1.02810
1.50 4.62 1.25 1.45 1.35 1.01960 0.00120 1.02090
2.00 4.76 1.25 1.45 1.35 1.01630 0.00060 1.01690

avg
= 4.55 = 12.19510 0.05020 12.24580
1.00 0.50 0.50 4.61 1.25 1.31 1.35 1.02230 0.00130 1.02360
1.00 4.65 1.25 1.31 1.35 1.01590 0.00180 1.01770
1.50 4.33 1.25 1.31 1.35 1.01230 0.00900 1.02140
2.00 4.11 1.25 1.31 1.35 1.01010 0.02380 1.03380
1.00 0.50 4.17 1.25 1.31 1.35 1.04130 0.00960 1.05080
1.00 4.52 1.25 1.31 1.35 1.03000 0.00200 1.03200
1.50 4.67 1.25 1.31 1.35 1.02360 0.00130 1.02480
2.00 4.60 1.25 1.31 1.35 1.01940 0.00240 1.02180
1.50 0.50 3.94 1.25 1.31 1.35 1.05750 0.02650 1.08400
1.00 4.24 1.25 1.31 1.35 1.04270 0.00700 1.04970
1.50 4.47 1.25 1.31 1.35 1.03390 0.00250 1.03640
2.00 4.61 1.25 1.31 1.35 1.02810 0.00130 1.02940

avg
= 4.41
= 12.33710 0.08850 12.42540
2.00 0.50 0.50 4.47 1.25 1.17 1.35 1.03810 0.00250 1.04060
1.00 4.54 1.25 1.17 1.35 1.02710 0.00310 1.03020
1.50 4.25 1.25 1.17 1.35 1.02100 0.01280 1.03380
2.00 4.05 1.25 1.17 1.35 1.01720 0.03090 1.04810
1.00 0.50 4.01 1.25 1.17 1.35 1.07040 0.01970 1.09020
1.00 4.38 1.25 1.17 1.35 1.05120 0.00390 1.05510
1.50 4.53 1.25 1.17 1.35 1.04020 0.00250 1.04270
2.00 4.48 1.25 1.17 1.35 1.03310 0.00420 1.03730
1.50 0.50 3.75 1.25 1.17 1.35 1.09820 0.05740 1.15560
1.00 4.08 1.25 1.17 1.35 1.07280 0.01430 1.08710
1.50 4.32 1.25 1.17 1.35 1.05780 0.00500 1.06280
2.00 4.47 1.25 1.17 1.35 1.04800 0.00290 1.05090

avg
= 4.28 = 12.57510 0.15920 12.73440
3.00 0.50 0.50 4.40 1.25 1.10 1.35 1.05180 0.00370 1.05550
1.00 4.48 1.25 1.10 1.35 1.03690 0.00410 1.04100
1.50 4.21 1.25 1.10 1.35 1.02860 0.01540 1.04400
2.00 4.02 1.25 1.10 1.35 1.02340 0.03560 1.05900
1.00 0.50 3.91 1.25 1.10 1.35 1.09580 0.02960 1.12540
1.00 4.29 1.25 1.10 1.35 1.06960 0.00570 1.07540
1.50 4.46 1.25 1.10 1.35 1.05470 0.00360 1.05830
2.00 4.42 1.25 1.10 1.35 1.04510 0.00560 1.05070
1.50 0.50 3.64 1.25 1.10 1.35 1.13360 0.08880 1.22230
1.00 3.99 1.25 1.10 1.35 1.09900 0.02140 1.12040
1.50 4.24 1.25 1.10 1.35 1.07870 0.00730 1.08600
2.00 4.39 1.25 1.10 1.35 1.06530 0.00430 1.06960

avg
= 4.20 = 12.78250 0.22510 13.00760




146

5.5.3 Extreme Event I: Combination of Dead, Live and Earthquake Load
The Extreme Event I limit state is for combination of dead load, earthquake load
and vehicular live load. All the elements and connection have to satisfy the design check
( )
n E n L n D n
E L D R + + 47 . 1 (5-39)
Current load factors are 1.25, 0.50 and 1.0 for dead load, live load and earthquake
load, respectively. Normalized initial cost can be written as follows for this combination
case;
( ) ( )( ) ( )
( )
|
|
.
|

\
|
+ +
+ +
+ =
E L
E E E L L D
G
I
K
C
C


00 . 1 47 . 1 50 . 0 25 . 1
00 . 1 50 . 0 47 . 1 25 . 1
1
0
(5-40)
where
E
and
E
are the earthquake load marginal cost slope to gravitational load cost
slope ratio (K
E
/K
G
) and nominal earthquake load to nominal dead load ratio (E
n
/D
n
),
respectively and
L
is the nominal live load to nominal live load ratio (L
n
/D
n
). This
combination case is special, because earthquake analysis already takes into account
system effects by using the response modification factor, R
m
, and the safety index
obtained corresponds to system safety.
In order to see the effect of g on the analysis both deduced g values of 30 and
1727 are used. Actually, the 1727 value is the appropriate value to use for analysis
because of it is for system safety. The TCF equation is



147
( ) ( ) ( )

+
|
|
.
|

\
|
+ +
+ +
+ =
ij
fi
E L
E E E L L
G
e
P
g K TCF


00 . 1 50 . 0 25 . 1
00 . 1 50 . 0 25 . 1
1
0
(5-41)
the optimization process is repeated for different values of
E
=K
E
/K
G
along the design
space namely
L
=L
n
/D
n
={ 0.5 1.0 1.5 2.0} and
E
=E
n
/D
n
={0.5 1.0 2.0 5.0}and the limit
state function in Equation (5-42) along with the statistical data given in Table 46 is used
in FORM analysis.

Table 46 Statistical data for D+L+E case
R.V. Distr. BIAS COV
R Logn. 1.12 0.10
D Norm 1.03 0.08
X
L
Logn. 1.00 0.18
I
D
Norm 1.13 0.10
Q
L
Norm 0.79 0.10
X
E
Norm 1.00 0.20
R
m
Norm 1.50 0.34
Q
E(1 event)
Extr. II 0.28 0.43


n
n
m
E E
n
n
L D L
n
n
D
L
R
Q X
D
L
Q I X D
D
R
R G = (5-42)
All the random variables in above limit state function have a nominal value of 1.0
since they are normalized with their nominal values. As in the previous combination case
since there are two time dependent loads a load combination process described in Chapter
2 should be used for this limit state. Nested Reliability is used to find the probability of
failure (Equation (2-28)).


148
The probability of failure can be evaluated by using the Nested Reliability
Analysis for the limit state function h is given by;
( ) R X X I u h
E L D
, , , + = (5-43)
where u is a standard normal variate with zero mean and unit standard deviation, and
is a function of the dynamic amplification and modeling factor for live load and modeling
factor and response modification factors for earthquake load and a limit state function of
|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|
=
n
n
m
E E
n
n
L L
n
n
D
E
r
Q x
D
L
Q i x D
D
R
R g
D *
*
* *
(5-44)
where star values are design values (most probable values) for the solution of limit state
function, h. Since the solution is iterative for every star value the Equation (5-38) is
solved using Ferry-Borges Method (Rackwitz-Fiessler algorithm) till the FORM solution
for g converges to a value.
The dead load factor is kept at the current code specified value of 1.25 to keep the
optimization problem two-dimensional. In optimization analysis 16 design samples are
used to represent the design space, the sum of total cost factors found to be 56.13 using
current code values where sum of initial cost factors is 16, and sum of failure cost factors
is 40.13. Safety index varied from 0.84 to 3.12 and have an average of 2.51. For small
values of E
n
/D
n
, current code load factors give very low safety index values. Although the
safety index values are already for system safety, cost optimization is run for both g=30
and g=1727.


149
In Table 48, the results for g=30 is presented; the safety index values show a wide
scatter. The same observation holds for the results presented in Table 49 considering
system safety. The results indicate that the range used to consider design space do not
replicate the actual one; when E
n
/D
n
values are small, the dead and live load combination
would govern therefore the smaller values are omitted and a new design sample of E
n
/D
n

={2, 3, 5, 10} is adapted. Also marginal earthquake cost slope to marginal gravitational
cost slope ratios are
E
={1, 2, 3, 5, 10}, because earthquake is likely to have relatively
higher marginal cost slope values. The current code load factors give the safety indices
and cost factors tabulated in Table 50. The sum of total cost factors is about 27 for
g=1727. Considering the experience in the previous load combination case about not
having too many design samples to avoid high scatter, the live load factor is fixed at
current specified value of 0.50 and the optimization process is repeated. When deduced g
for component safety is considered, optimum earthquake load factors are 0.76 and 0.66
for
E
equals to 5 and 10, respectively. Considering the deduced g for system safety
optimized earthquake load factors are 1.66 and 1.44 for
E
of 5.0 and 10.0, respectively.
Since the marginal earthquake load cost factor is mostly very high compared to marginal
gravitational load cost factor, a value of 10 is assumed for
E
. Therefore, the load factors
are selected as 1.25, 1.44 and 0.50 for dead, earthquake and live load factors,
respectively. The sum of total cost factors drops from the current codes 27 to an
optimized value of 20, which is about a 30% decrease in expected total cost.



150


Table 47 Current code safety index and cost factors for D+E+L case g=1727
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
0.50 0.50 3.00 1.25 1.00 0.50 1.0000 0.1154 1.1154
1.00 1.86 1.25 1.00 0.50 1.0000 2.7765 3.7765
1.50 1.23 1.25 1.00 0.50 1.0000 10.1451 11.1451
2.00 0.84 1.25 1.00 0.50 1.0000 19.5382 20.5382
1.00 0.50 3.12 1.25 1.00 0.50 1.0000 0.0543 1.0543
1.00 2.90 1.25 1.00 0.50 1.0000 0.1455 1.1455
1.50 2.10 1.25 1.00 0.50 1.0000 1.5026 2.5026
2.00 1.58 1.25 1.00 0.50 1.0000 5.0224 6.0224
2.00 0.50 3.01 1.25 1.00 0.50 1.0000 0.0776 1.0776
1.00 3.00 1.25 1.00 0.50 1.0000 0.0819 1.0819
1.50 2.97 1.25 1.00 0.50 1.0000 0.0907 1.0907
2.00 2.91 1.25 1.00 0.50 1.0000 0.1612 1.1612
5.00 0.50 2.93 1.25 1.00 0.50 1.0000 0.1022 1.1022
1.00 2.93 1.25 1.00 0.50 1.0000 0.1029 1.1029
1.50 2.92 1.25 1.00 0.50 1.0000 0.1043 1.1043
2.00 2.92 1.25 1.00 0.50 1.0000 0.1063 1.1063

avg
= 2.51
=
16.0000 40.1271 56.1271

min
= 0.84

max
= 3.12












151


Table 48 Optimized system safety index,
sys
and load factors,
E
and
L
for D+E+L
(
D
=1.25, g=30).

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
0.50 0.50 0.50 4.10 1.25 1.00 0.97 1.0195 0.0003 1.0198
1.00 3.40 1.25 1.00 0.97 1.0332 0.0052 1.0384
1.50 3.03 1.25 1.00 0.97 1.0434 0.0191 1.0625
2.00 2.81 1.25 1.00 0.97 1.0513 0.0393 1.0905
1.00 0.50 3.40 1.25 1.00 0.97 1.0158 0.0041 1.0199
1.00 3.58 1.25 1.00 0.97 1.0277 0.0021 1.0297
1.50 3.64 1.25 1.00 0.97 1.0369 0.0020 1.0389
2.00 3.31 1.25 1.00 0.97 1.0444 0.0071 1.0515
2.00 0.50 3.18 1.25 1.00 0.97 1.0114 0.0090 1.0205
1.00 3.31 1.25 1.00 0.97 1.0207 0.0057 1.0265
1.50 3.41 1.25 1.00 0.97 1.0285 0.0039 1.0323
2.00 3.51 1.25 1.00 0.97 1.0350 0.0027 1.0377
5.00 0.50 3.01 1.25 1.00 0.97 1.0063 0.0162 1.0224
1.00 3.07 1.25 1.00 0.97 1.0119 0.0129 1.0248
1.50 3.13 1.25 1.00 0.97 1.0169 0.0105 1.0274
2.00 3.19 1.25 1.00 0.97 1.0214 0.0086 1.0300

avg
= 3.32
=
16.4243 0.1487 16.5728

min
= 2.81

max
= 4.10
1.00 0.50 0.50 3.87 1.25 0.73 1.05 1.0152 0.0009 1.0161
1.00 3.37 1.25 0.73 1.05 1.0326 0.0058 1.0383
1.50 3.10 1.25 0.73 1.05 1.0454 0.0150 1.0605
2.00 2.94 1.25 0.73 1.05 1.0554 0.0259 1.0813
1.00 0.50 3.24 1.25 0.73 1.05 1.0060 0.0074 1.0134
1.00 3.49 1.25 0.73 1.05 1.0216 0.0025 1.0241
1.50 3.53 1.25 0.73 1.05 1.0337 0.0032 1.0369
2.00 3.28 1.25 0.73 1.05 1.0435 0.0079 1.0514
2.00 0.50 2.94 1.25 0.73 1.05 0.9952 0.0203 1.0156
1.00 3.13 1.25 0.73 1.05 1.0079 0.0107 1.0186
1.50 3.28 1.25 0.73 1.05 1.0184 0.0063 1.0247
2.00 3.41 1.25 0.73 1.05 1.0273 0.0039 1.0312
5.00 0.50 2.68 1.25 0.73 1.05 0.9824 0.0453 1.0277
1.00 2.79 1.25 0.73 1.05 0.9903 0.0323 1.0227
1.50 2.89 1.25 0.73 1.05 0.9974 0.0238 1.0212
2.00 2.97 1.25 0.73 1.05 1.0039 0.0180 1.0218

avg
= 3.18 =
16.2762 0.2292 16.5055

min
= 2.68

max
= 3.87




152

contd

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
2.00 0.50 0.50 3.80 1.25 0.54 1.15 1.0008 0.0012 1.0019
1.00 3.46 1.25 0.54 1.15 1.0236 0.0041 1.0278
1.50 3.26 1.25 0.54 1.15 1.0406 0.0084 1.0490
2.00 3.14 1.25 0.54 1.15 1.0537 0.0128 1.0665
1.00 0.50 3.13 1.25 0.54 1.15 0.9794 0.0105 0.9900
1.00 3.46 1.25 0.54 1.15 1.0011 0.0029 1.0040
1.50 3.56 1.25 0.54 1.15 1.0180 0.0028 1.0208
2.00 3.38 1.25 0.54 1.15 1.0316 0.0055 1.0371
2.00 0.50 2.74 1.25 0.54 1.15 0.9545 0.0373 0.9917
1.00 3.01 1.25 0.54 1.15 0.9730 0.0158 0.9888
1.50 3.21 1.25 0.54 1.15 0.9884 0.0079 0.9963
2.00 3.38 1.25 0.54 1.15 1.0014 0.0044 1.0058
5.00 0.50 2.38 1.25 0.54 1.15 0.9249 0.1070 1.0318
1.00 2.55 1.25 0.54 1.15 0.9370 0.0661 1.0031
1.50 2.69 1.25 0.54 1.15 0.9479 0.0433 0.9912
2.00 2.82 1.25 0.54 1.15 0.9578 0.0296 0.9874

avg
= 3.12
=
15.8337 0.3596 16.1932

min
= 2.38

max
= 3.80
3.00 0.50 0.50 3.86 1.25 0.44 1.24 0.9833 0.0009 0.9842
1.00 3.64 1.25 0.44 1.24 1.0122 0.0021 1.0143
1.50 3.49 1.25 0.44 1.24 1.0336 0.0037 1.0373
2.00 3.39 1.25 0.44 1.24 1.0502 0.0052 1.0554
1.00 0.50 3.11 1.25 0.44 1.24 0.9480 0.0115 0.9595
1.00 3.49 1.25 0.44 1.24 0.9764 0.0025 0.9789
1.50 3.70 1.25 0.44 1.24 0.9985 0.0016 1.0001
2.00 3.56 1.25 0.44 1.24 1.0163 0.0027 1.0190
2.00 0.50 2.65 1.25 0.44 1.24 0.9066 0.0493 0.9559
1.00 2.98 1.25 0.44 1.24 0.9317 0.0174 0.9491
1.50 3.22 1.25 0.44 1.24 0.9525 0.0077 0.9603
2.00 3.41 1.25 0.44 1.24 0.9701 0.0040 0.9741
5.00 0.50 2.20 1.25 0.44 1.24 0.8575 0.1754 1.0329
1.00 2.42 1.25 0.44 1.24 0.8744 0.0948 0.9692
1.50 2.61 1.25 0.44 1.24 0.8896 0.0559 0.9455
2.00 2.76 1.25 0.44 1.24 0.9033 0.0352 0.9386

avg
= 3.16 =
15.3042 0.4699 15.7743

min
= 2.20

max
= 3.86




153

Table 49 Optimized system safety index,
sys
and load factors,
E
and
L
for D+E+L
(
D
=1.25, g=1727).

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
0.50 0.50 0.50 4.30 1.25 1.00 2.27 1.0738 0.0026 1.0764
1.00 6.16 1.25 1.00 2.27 1.1258 0.0000 1.1258
1.50 6.03 1.25 1.00 2.27 1.1644 0.0000 1.1644
2.00 5.96 1.25 1.00 2.27 1.1942 0.0000 1.1942
1.00 0.50 3.87 1.25 1.00 2.27 1.0597 0.0376 1.0973
1.00 4.24 1.25 1.00 2.27 1.1047 0.0077 1.1124
1.50 4.49 1.25 1.00 2.27 1.1399 0.0025 1.1424
2.00 6.21 1.25 1.00 2.27 1.1681 0.0000 1.1681
2.00 0.50 3.51 1.25 1.00 2.27 1.0432 0.1540 1.1972
1.00 3.82 1.25 1.00 2.27 1.0784 0.0464 1.1249
1.50 4.04 1.25 1.00 2.27 1.1077 0.0186 1.1263
2.00 4.21 1.25 1.00 2.27 1.1325 0.0088 1.1413
5.00 0.50 3.18 1.25 1.00 2.27 1.0236 0.5075 1.5312
1.00 3.38 1.25 1.00 2.27 1.0447 0.2525 1.2973
1.50 3.54 1.25 1.00 2.27 1.0638 0.1398 1.2035
2.00 3.67 1.25 1.00 2.27 1.0810 0.0837 1.1646

avg
= 4.41
=
17.6055 1.2617 18.8673

min
= 3.18

max
= 6.21
1.00 0.50 0.50 5.70 1.25 1.96 1.17 1.0553 0.0000 1.0553
1.00 4.72 1.25 1.96 1.17 1.0710 0.0010 1.0720
1.50 4.23 1.25 1.96 1.17 1.0827 0.0098 1.0924
2.00 3.93 1.25 1.96 1.17 1.0917 0.0349 1.1266
1.00 0.50 3.93 1.25 1.96 1.17 1.0668 0.0299 1.0967
1.00 4.07 1.25 1.96 1.17 1.0785 0.0159 1.0944
1.50 4.20 1.25 1.96 1.17 1.0876 0.0094 1.0970
2.00 4.72 1.25 1.96 1.17 1.0949 0.0009 1.0958
2.00 0.50 3.78 1.25 1.96 1.17 1.0803 0.0534 1.1337
1.00 3.88 1.25 1.96 1.17 1.0878 0.0364 1.1242
1.50 3.96 1.25 1.96 1.17 1.0940 0.0258 1.1199
2.00 4.04 1.25 1.96 1.17 1.0993 0.0189 1.1182
5.00 0.50 3.68 1.25 1.96 1.17 1.0963 0.0799 1.1762
1.00 3.73 1.25 1.96 1.17 1.0997 0.0670 1.1667
1.50 3.77 1.25 1.96 1.17 1.1028 0.0566 1.1595
2.00 3.81 1.25 1.96 1.17 1.1056 0.0483 1.1539

avg
= 4.13 =
17.3943 0.4881 17.8825

min
= 3.68

max
= 5.70



154

contd

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
2.00 0.50 0.50 5.50 1.25 1.53 1.34 1.0654 0.0000 1.0654
1.00 4.77 1.25 1.53 1.34 1.0856 0.0007 1.0864
1.50 4.41 1.25 1.53 1.34 1.1007 0.0043 1.1050
2.00 4.19 1.25 1.53 1.34 1.1123 0.0115 1.1238
1.00 0.50 3.81 1.25 1.53 1.34 1.0773 0.0476 1.1249
1.00 4.02 1.25 1.53 1.34 1.0927 0.0200 1.1128
1.50 4.18 1.25 1.53 1.34 1.1048 0.0059 1.1107
2.00 4.77 1.25 1.53 1.34 1.1144 0.0007 1.1151
2.00 0.50 3.62 1.25 1.53 1.34 1.0914 0.1036 1.1950
1.00 3.76 1.25 1.53 1.34 1.1016 0.0594 1.1611
1.50 3.88 1.25 1.53 1.34 1.1101 0.0367 1.1468
2.00 3.98 1.25 1.53 1.34 1.1173 0.0239 1.1413
5.00 0.50 3.47 1.25 1.53 1.34 1.1080 0.1822 1.2903
1.00 3.54 1.25 1.53 1.34 1.1130 0.1392 1.2522
1.50 3.61 1.25 1.53 1.34 1.1175 0.1083 1.2258
2.00 3.67 1.25 1.53 1.34 1.1215 0.0856 1.2071

avg
= 4.07
=
17.6336 0.8296 18.4637

min
= 3.47

max
= 5.50
3.00 0.50 0.50 4.16 1.25 1.24 1.71 1.0709 0.0052 1.0761
1.00 5.32 1.25 1.24 1.71 1.1034 0.0000 1.1035
1.50 5.08 1.25 1.24 1.71 1.1276 0.0001 1.1277
2.00 4.94 1.25 1.24 1.71 1.1462 0.0003 1.1465
1.00 0.50 3.80 1.25 1.24 1.71 1.0739 0.0493 1.1233
1.00 4.09 1.25 1.24 1.71 1.1006 0.0147 1.1153
1.50 4.30 1.25 1.24 1.71 1.1215 0.0058 1.1273
2.00 5.34 1.25 1.24 1.71 1.1382 0.0000 1.1382
2.00 0.50 3.53 1.25 1.24 1.71 1.0774 0.1432 1.2207
1.00 3.75 1.25 1.24 1.71 1.0971 0.0618 1.1589
1.50 3.92 1.25 1.24 1.71 1.1134 0.0310 1.1445
2.00 4.06 1.25 1.24 1.71 1.1273 0.0173 1.1446
5.00 0.50 3.31 1.25 1.24 1.71 1.0816 0.3267 1.4083
1.00 3.43 1.25 1.24 1.71 1.0926 0.2103 1.3029
1.50 3.54 1.25 1.24 1.71 1.1025 0.1418 1.2443
2.00 3.63 1.25 1.24 1.71 1.1115 0.0992 1.2106

avg
= 4.14 =
17.6857 1.1067 18.7927

min
= 3.31

max
= 5.34



155




Table 50 Current code safety index and cost factors for D+E+L case g=1727 (with
updated range)

E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
2.00 0.50 3.01 1.25 1.00 0.50 1.0000 0.9059 1.9059
1.00 3.00 1.25 1.00 0.50 1.0000 0.9558 1.9558
1.50 2.97 1.25 1.00 0.50 1.0000 1.0585 2.0585
2.00 2.91 1.25 1.00 0.50 1.0000 1.8816 2.8816
5.00 0.50 2.93 1.25 1.00 0.50 1.0000 1.1924 2.1924
1.00 2.93 1.25 1.00 0.50 1.0000 1.2010 2.2010
1.50 2.92 1.25 1.00 0.50 1.0000 1.2168 2.2168
2.00 2.92 1.25 1.00 0.50 1.0000 1.2409 2.2409
10.00 0.50 2.90 1.25 1.00 0.50 1.0000 1.3227 2.3227
1.00 2.90 1.25 1.00 0.50 1.0000 1.3243 2.3243
1.50 2.90 1.25 1.00 0.50 1.0000 1.3277 2.3277
2.00 2.89 1.25 1.00 0.50 1.0000 1.3331 2.3331

avg
= 2.93 =
12.0000 14.9607 26.9607

min
= 2.89

max
= 3.01





156

Table 51 Optimized component safety index,
sys
and load factor,
E
for D+E+L
(
D
=1.25,

L
=0.50, g=30 with updated range).

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
1.00 2.00 0.50 3.02 1.25 1.01 0.50 1.0005 0.0153 1.0158
1.00 3.01 1.25 1.01 0.50 1.0005 0.0161 1.0166
1.50 2.98 1.25 1.01 0.50 1.0004 0.0179 1.0183
2.00 2.92 1.25 1.01 0.50 1.0004 0.0216 1.0220
5.00 0.50 2.94 1.25 1.01 0.50 1.0007 0.0201 1.0208
1.00 2.94 1.25 1.01 0.50 1.0007 0.0203 1.0209
1.50 2.93 1.25 1.01 0.50 1.0006 0.0205 1.0212
2.00 2.93 1.25 1.01 0.50 1.0006 0.0209 1.0215
10.00 0.50 2.91 1.25 1.01 0.50 1.0008 0.0223 1.0231
1.00 2.91 1.25 1.01 0.50 1.0008 0.0223 1.0231
1.50 2.91 1.25 1.01 0.50 1.0008 0.0224 1.0232
2.00 2.90 1.25 1.01 0.50 1.0007 0.0225 1.0232

avg
= 2.94
=
12.0075 0.2422 12.2497

min
= 2.90

max
= 3.02
2.00 2.00 0.50 2.89 1.25 0.89 0.50 0.9851 0.0237 1.0087
1.00 2.87 1.25 0.89 0.50 0.9864 0.0253 1.0118
1.50 2.83 1.25 0.89 0.50 0.9876 0.0290 1.0166
2.00 2.73 1.25 0.89 0.50 0.9885 0.0638 1.0524
5.00 0.50 2.79 1.25 0.89 0.50 0.9796 0.0320 1.0116
1.00 2.79 1.25 0.89 0.50 0.9807 0.0323 1.0130
1.50 2.78 1.25 0.89 0.50 0.9816 0.0329 1.0145
2.00 2.78 1.25 0.89 0.50 0.9825 0.0337 1.0162
10.00 0.50 2.76 1.25 0.89 0.50 0.9767 0.0359 1.0127
1.00 2.75 1.25 0.89 0.50 0.9775 0.0360 1.0135
1.50 2.75 1.25 0.89 0.50 0.9781 0.0361 1.0142
2.00 2.75 1.25 0.89 0.50 0.9788 0.0363 1.0151

avg
= 2.79
=
11.7831 0.4170 12.2003

min
= 2.73

max
= 2.89
3.00 2.00 0.50 2.83 1.25 0.84 0.50 0.9685 0.0284 0.9968
1.00 2.81 1.25 0.84 0.50 0.9714 0.0306 1.0019
1.50 2.76 1.25 0.84 0.50 0.9738 0.0357 1.0094
2.00 2.38 1.25 0.84 0.50 0.9758 0.1402 1.1160
5.00 0.50 2.73 1.25 0.84 0.50 0.9569 0.0388 0.9957
1.00 2.73 1.25 0.84 0.50 0.9592 0.0392 0.9984
1.50 2.72 1.25 0.84 0.50 0.9612 0.0399 1.0011
2.00 2.71 1.25 0.84 0.50 0.9630 0.0411 1.0041
10.00 0.50 2.69 1.25 0.84 0.50 0.9509 0.0438 0.9947
1.00 2.69 1.25 0.84 0.50 0.9524 0.0439 0.9963
1.50 2.69 1.25 0.84 0.50 0.9538 0.0441 0.9979
2.00 2.69 1.25 0.84 0.50 0.9551 0.0443 0.9995

avg
= 2.70
=
11.5420 0.5700 12.1118

min
= 2.38

max
= 2.83



157

contd

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
5.00 2.00 0.50 2.72 1.25 0.76 0.50 0.9201 0.0401 0.9602
1.00 2.69 1.25 0.76 0.50 0.9275 0.0439 0.9714
1.50 2.63 1.25 0.76 0.50 0.9336 0.0537 0.9873
2.00 2.17 1.25 0.76 0.50 0.9388 0.2511 1.1899
5.00 0.50 2.61 1.25 0.76 0.50 0.8908 0.0562 0.9470
1.00 2.60 1.25 0.76 0.50 0.8966 0.0569 0.9535
1.50 2.59 1.25 0.76 0.50 0.9017 0.0582 0.9600
2.00 2.58 1.25 0.76 0.50 0.9064 0.0603 0.9667
10.00 0.50 2.56 1.25 0.76 0.50 0.8756 0.0642 0.9398
1.00 2.56 1.25 0.76 0.50 0.8794 0.0643 0.9438
1.50 2.56 1.25 0.76 0.50 0.8830 0.0646 0.9476
2.00 2.56 1.25 0.76 0.50 0.8864 0.0651 0.9515

avg
= 2.57
=
10.8399 0.8786 11.7187

min
= 2.17

max
= 2.72
10.00 2.00 0.50 2.56 1.25 0.66 0.50 0.7726 0.0644 0.8370
1.00 2.52 1.25 0.66 0.50 0.7935 0.0728 0.8664
1.50 2.42 1.25 0.66 0.50 0.8110 0.1256 0.9366
2.00 1.89 1.25 0.66 0.50 0.8257 0.5051 1.3307
5.00 0.50 2.43 1.25 0.66 0.50 0.6892 0.0938 0.7831
1.00 2.42 1.25 0.66 0.50 0.7056 0.0954 0.8010
1.50 2.41 1.25 0.66 0.50 0.7203 0.0983 0.8185
2.00 2.40 1.25 0.66 0.50 0.7336 0.1029 0.8365
10.00 0.50 2.37 1.25 0.66 0.50 0.6460 0.1090 0.7550
1.00 2.37 1.25 0.66 0.50 0.6568 0.1093 0.7661
1.50 2.37 1.25 0.66 0.50 0.6670 0.1100 0.7770
2.00 2.37 1.25 0.66 0.50 0.6766 0.1110 0.7877

avg
= 2.38
=
8.6979 1.5976 10.2956

min
= 1.89

max
= 2.56




158

Table 52 Optimized system safety index,
sys
and load factor,
E
for D+E+L (
D
=1.25,

L
=0.50, g=1727 with updated range).

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
1.00 2.00 0.50 3.83 1.25 2.33 0.50 1.0883 0.0439 1.1322
1.00 3.83 1.25 2.33 0.50 1.0802 0.0443 1.1245
1.50 3.83 1.25 2.33 0.50 1.0734 0.0451 1.1185
2.00 3.82 1.25 2.33 0.50 1.0677 0.0462 1.1139
5.00 0.50 3.80 1.25 2.33 0.50 1.1207 0.0502 1.1709
1.00 3.80 1.25 2.33 0.50 1.1143 0.0502 1.1646
1.50 3.80 1.25 2.33 0.50 1.1086 0.0503 1.1590
2.00 3.80 1.25 2.33 0.50 1.1035 0.0504 1.1539
10.00 0.50 3.79 1.25 2.33 0.50 1.1375 0.0526 1.1901
1.00 3.79 1.25 2.33 0.50 1.1333 0.0526 1.1859
1.50 3.79 1.25 2.33 0.50 1.1293 0.0526 1.1819
2.00 3.79 1.25 2.33 0.50 1.1256 0.0526 1.1782

avg
= 3.81 =
13.2824 0.5910 13.8736

min
= 3.79

max
= 3.83
2.00 2.00 0.50 3.70 1.25 2.02 0.50 1.1348 0.0748 1.2096
1.00 3.70 1.25 2.02 0.50 1.1224 0.0758 1.1982
1.50 3.69 1.25 2.02 0.50 1.1121 0.0775 1.1896
2.00 3.68 1.25 2.02 0.50 1.1033 0.0802 1.1835
5.00 0.50 3.66 1.25 2.02 0.50 1.1842 0.0871 1.2713
1.00 3.66 1.25 2.02 0.50 1.1745 0.0872 1.2617
1.50 3.66 1.25 2.02 0.50 1.1658 0.0874 1.2533
2.00 3.66 1.25 2.02 0.50 1.1579 0.0878 1.2458
10.00 0.50 3.65 1.25 2.02 0.50 1.2099 0.0919 1.3018
1.00 3.65 1.25 2.02 0.50 1.2034 0.0919 1.2953
1.50 3.65 1.25 2.02 0.50 1.1974 0.0920 1.2893
2.00 3.65 1.25 2.02 0.50 1.1917 0.0920 1.2837

avg
= 3.67 =
13.9574 1.0256 14.9831

min
= 3.65

max
= 3.70
3.00 2.00 0.50 3.62 1.25 1.85 0.50 1.1695 0.1020 1.2715
1.00 3.62 1.25 1.85 0.50 1.1539 0.1036 1.2574
1.50 3.61 1.25 1.85 0.50 1.1409 0.1064 1.2473
2.00 3.60 1.25 1.85 0.50 1.1299 0.1109 1.2408
5.00 0.50 3.58 1.25 1.85 0.50 1.2316 0.1201 1.3517
1.00 3.58 1.25 1.85 0.50 1.2194 0.1203 1.3398
1.50 3.58 1.25 1.85 0.50 1.2085 0.1208 1.3292
2.00 3.58 1.25 1.85 0.50 1.1985 0.1214 1.3200
10.00 0.50 3.56 1.25 1.85 0.50 1.2638 0.1274 1.3912
1.00 3.56 1.25 1.85 0.50 1.2557 0.1274 1.3831
1.50 3.56 1.25 1.85 0.50 1.2481 0.1275 1.3756
2.00 3.56 1.25 1.85 0.50 1.2410 0.1276 1.3686

avg
= 3.58 =
14.4608 1.4154 15.8762

min
= 3.56

max
= 3.62



159

contd

E
E
n
/D
n
Ln/Dn
D

E

L
ICF FCF TCF
5.00 2.00 0.50 3.52 1.25 1.66 0.50 1.2199 0.1505 1.3703
1.00 3.51 1.25 1.66 0.50 1.1996 0.1534 1.3530
1.50 3.51 1.25 1.66 0.50 1.1828 0.1587 1.3415
2.00 3.49 1.25 1.66 0.50 1.1686 0.1672 1.3358
5.00 0.50 3.47 1.25 1.66 0.50 1.3005 0.1801 1.4806
1.00 3.47 1.25 1.66 0.50 1.2847 0.1805 1.4652
1.50 3.47 1.25 1.66 0.50 1.2705 0.1813 1.4518
2.00 3.47 1.25 1.66 0.50 1.2576 0.1825 1.4402
10.00 0.50 3.45 1.25 1.66 0.50 1.3423 0.1922 1.5345
1.00 3.45 1.25 1.66 0.50 1.3318 0.1922 1.5240
1.50 3.45 1.25 1.66 0.50 1.3220 0.1924 1.5143
2.00 3.45 1.25 1.66 0.50 1.3127 0.1926 1.5053

avg
= 3.48
=
15.1930 2.1236 17.3165

min
= 3.45

max
= 3.52
10.00 2.00 0.50 3.38 1.25 1.44 0.50 1.2899 0.2540 1.5438
1.00 3.37 1.25 1.44 0.50 1.2632 0.2606 1.5238
1.50 3.36 1.25 1.44 0.50 1.2410 0.2730 1.5140
2.00 3.34 1.25 1.44 0.50 1.2222 0.2938 1.5160
5.00 0.50 3.32 1.25 1.44 0.50 1.3961 0.3114 1.7076
1.00 3.32 1.25 1.44 0.50 1.3753 0.3125 1.6878
1.50 3.32 1.25 1.44 0.50 1.3566 0.3144 1.6709
2.00 3.32 1.25 1.44 0.50 1.3396 0.3173 1.6569
10.00 0.50 3.30 1.25 1.44 0.50 1.4513 0.3355 1.7868
1.00 3.30 1.25 1.44 0.50 1.4374 0.3356 1.7731
1.50 3.30 1.25 1.44 0.50 1.4244 0.3360 1.7604
2.00 3.30 1.25 1.44 0.50 1.4122 0.3366 1.7488

avg
= 3.33
=
16.2092 3.6807 19.8899

min
= 3.30

max
= 3.38




160

5.6 Optimum Load Factors Table and Discussion of Results
The product of a calibration process is the load factor table for the load
combinations considered. A calibration example using Reliability Based Cost
Optimization is illustrated in this chapter and the optimum load factors are presented in
Table 53 considering both component and system safety except load combination with
earthquake load. In earthquake plus live and dead load combination case the safety index
calculated corresponds to system safety index. Therefore, only optimum load factor
considering the system safety index is provided in Table 53 for D+E+L combination
case. In load combinations other than the case involving earthquake, the system safety
index is approximated as the component safety index plus one in optimization process.
The aim of this example is to show the implementation of the approach studied.
Available data were gathered, but when there were not enough data assumed values had
to be used, such as the marginal wind load to gravitational load cost ratio,
W
. Therefore,
the recommended values are based on these assumed values. In next chapter effects of the
parameters are studied in order to see the sensitivity of the optimum load factors.
The optimum total cost values show that the expected total costs differ
considerably from the current code values; in D+W case the sum of total expected cost,
TCF is reduced 3%, in D+W+L case the reduction in TCF is about 55% and in D+E+L
case the reduction is about 30%. The nominal load ratio ranges used are also assumed,
and results show that the range and number of design points used do affect the outcome.


161
Therefore, in order to improve results, it is believed that the ranges are needed to be set
clearly, and number of design samples used to represent the design space has to be
chosen to represent real design situations. In the illustration here the samples are assumed
to be equally likely. If a database containing the bridge design considering the possible
nominal loads to dead load ratios could be collected, weighting factors would be used.
This would give more applicable results for the specification.


Table 53 Optimum Load Factors
Combination Case
D

L

W

E

W

E
D+L 1.25 1.75 ---- ---- ---- ----
D+W 1.25 ---- 1.49 ---- 1.60 ----
D+W+L 1.25 1.35 0.99 ---- 1.10 ----
D+E+L 1.25 0.50 ---- ---- ---- 1.44
Component Safety System Safety








162
6.0 SENSITIVITY ANALYSIS

It is important to see the effect of a change in assumed parameters of optimum
loads factors. It is not always possible to verify some of the parameters or statistical data,
so the recommendations should reflect any possible changes in these parameters. In order
to see the effect of the assumed parameters on the load factors obtained herein a
sensitivity analysis is run. Basically, the effects are studied by changing the assumed
values, and behavior of the calculated load factors due to these changes are investigated.
In this chapter sensitivity of the optimized load factors calculated in last chapter
to any change in parameters, such as, real interest rate, j, failure cost, g and initial cost
change C
I
, etc. are studied.

6.1 Effect of Initial Cost Slope Change, C
I
on K
G

The initial cost change; C
I
is assumed as 2% for a 25% change in live load
factor,
L
. In order to visualize the effect of any data different than the data assumed in
this study, the gravitational load cost factor, K
G
values are calculated for different values
of initial cost slope change namely 0.5,1, 2, 3, 4% with the design space as =0.5, 1.0,
1.5, 2.0 and averaged along the design space. Calculated K
G
values are tabulated in Table
54.



163

Table 54 K
G
values for changing C
I

C
I
= 0.50% 1.00% 2.00% 3.00% 4.00%

0.50 0.049 0.097 0.194 0.291 0.389


1.00 0.034 0.069 0.137 0.206 0.274
1.50 0.030 0.059 0.118 0.177 0.236
2.00 0.027 0.054 0.109 0.163 0.217
Average 0.035 0.070 0.140 0.209 0.279


Increasing the initial cost change also increases the K
G
value linearly. Therefore
the results of the gravitational load cost factor depends on the assumed value for the
initial cost change.

6.2 Effect of Gravitational Load Cost Factor, K
G
on Deduced Failure Cost Factor,
g
The deduced values of failure cost factor linearly affects the failure cost factor, g.
The relationship between K
G
and g can be seen clearly in Equation (5-28). The values of
g are tabulated in for different values of K
G
.






164

Table 55 Deduced g values for different K
G
values
K
G
g
0.10 25.0
0.12 30.0
0.15 37.5
0.20 50.0


6.3 Effect of Design Life, T on Deduced Failure Cost Factor, g
In AASHTO Specifications lifetime of the bridges are assumed as 75 years. The
failure cost factor, g is calculated by using Equation (5-23) for different design life, T
values namely 5,10,25,50,75,100 and 200 years. K
G
value is assumed to be 0.12 in
deducing g value.

Table 56 Failure cost factor, g for different design life, T values

L
T=5 T=10 T=25 T=50 T=75 T=100 T=200
0.5 38.78 35.43 32.37 31.14 30.81 30.69 30.63
1.0 42.13 37.86 34.04 32.51 32.09 31.95 31.87
1.5 39.08 35.00 31.35 29.89 29.50 29.36 29.28
2.0 36.79 32.90 29.43 28.05 27.67 27.54 27.46
Average g= 39.19 35.30 31.80 30.40 30.02 29.89 29.81


The failure cost factor, g values for different design life, T values are tabulated in
Table 56. The deduced g values are averaged along the
L
=L
n
/D
n
, and these averaged g


165
values indicate a decrease with increasing design life, T. This decrease becomes
insignificant if design life, T is over 25 years.

6.4 Effect of Real Interest Rate, j on Deduced Failure Cost Factor, g.
In Chapter 5, the during the optimization analysis a real interest rate of 3% was
assumed. Since the loads such as live and wind loads show the time variation, the
discount is done to yearly failure costs to present worth through the interest rate, j. In
order to visualize how this assumption affects the failure cost factor, g deduced from the
previous code, the optimization analysis are repeated for values of 0, 1, 2, 3, 4 and 5%
and results are tabulated in Table 57.

Table 57 The effect of real interest rate, j on failure cost factor, g (D+W combination)

L
j=0.01 j=0.02 j=0.03 j=0.04 j=0.05
0.5 27.18 29.13 30.81 32.27 33.58
1.0 27.84 30.12 32.09 33.82 35.37
1.5 25.49 27.64 29.50 31.13 32.59
2.0 23.88 25.91 27.67 29.22 30.60
Average g= 26.10 28.20 30.02 31.61 33.03


The results show that the real interest rate, j does not have an important impact on
the deduction of the failure cost factor, g. The failure cost factor, g is increases from 26
for 1% of interest rate to 33 for 5% of interest rate.



166
6.5 Effect of Gravitational Load Cost Factor, K
G
on Optimum Load Factors
In order to see how the initial cost slope for gravitational loads, the analysis is
repeated for different values of K
G
namely, 0.07, 0.14, 0.21 and 0.28 which corresponds
to a 1, 2, 3 and 4 percent initial cost change, CI for a 25% change in live load factor.
The discrete design points of
W
= W
n
/D
n
={0.5 1.0 2.0 5.0} is used, and the initial cost
slope for wind is assumed to be equal to gravitational cost slope, and the failure cost
factor is chosen to be equal to deduced value of 30 to see the trend. The optimum load
factors along with the safety index ranges and averages are tabulated in Table 58.

Table 58 Effect of K
G
on optimum load factors and optimum ranges for D+W case
K
G

D

W
(
optimum
)
min
(
optimum
)
max
(
optimum
)
0.06 1.25 1.88 3.14 3.50 3.30
0.12 1.25 1.73 2.94 3.34 3.12
0.18 1.25 1.64 2.82 3.24 3.01
0.24 1.25 1.58 2.73 3.17 2.93
0.36 1.25 1.49 2.60 3.07 2.81


The results in Table 58 indicate that in the combination of dead and wind load a
change in K
G
affects the optimum wind load factor calculated. As the K
G
increases the
marginal gravitational cost factor increases, therefore increasing the load factors become
more expensive, so the wind load factor decreases by increasing K
G
. Consequently, as it
becomes more expensive to improve the structures safety the optimum safety level
decrease to reach the balanced solution.


167

6.6 Effect of Design Life, T on Optimum Load Factors
Design life, T also affects the optimum load factor and safety index values,
therefore the optimum analysis are run for D+W case with design life values of 5,
10,25,50,75,100,200 and 300. The design life of 5 years may correspond to an evaluation
code where remaining lives are short. The failure cost factor, g is assumed as 30 and
initial cost slopes for gravitational and wind loads are assumed to be 0.12.

Table 59 Effect of T on optimum load factors and optimum ranges (D+W combination)
T
D

W
(
optimum
)
min
(
optimum
)
max
(
optimum
)
avg
5 1.25 1.37 3.21 3.62 3.40
10 1.25 1.49 3.18 3.58 3.36
25 1.25 1.62 3.12 3.51 3.29
50 1.25 1.70 3.03 3.42 3.20
75 1.25 1.74 2.95 3.35 3.13
100 1.25 1.75 2.89 3.29 3.07
200 1.25 1.77 2.72 3.13 2.90
300 1.25 1.78 2.60 3.04 2.80


As the design life increases the optimum wind load factor increases, on the other
hand the optimum safety index values decrease. The safety index values correspond to
the design life safety index values. Let us consider design lives of 75 years and 100 years;
for same load factors 100-year design life gives lower safety index, because the wind
load is a time dependent random variable and increasing the time of exposure lowers the
safety. Therefore the balance between safety and cost is attained at lower safety levels.


168
For an optimum wind load factor of 1.74, the average optimum safety index is 3.13 for
75-year design life, when 100-year design life is considered for an optimum wind load
factor 1.75, the average optimum safety index is 3.07.

6.7 Effect of Real Interest Rate, j on Optimized Load Factors and Safety Index
Since the process is a cost optimization the effect of the real interest rate, j on the
optimized load factors are expected. The interest rates, 1,2,3,4,5 and 6% are used to run
the optimization process for the case of K
G
and K
W
are equal to 0.12. Although the
change in j has an effect on g, it has shown in Section 6.4 this effect is not that
significant, such as for a j=1% the average g is 26.10 and for j=0.05 g increase to 33.03.
Therefore, g is assumed as 30 for all j values. The results are shown in Table 60. The
results given in Table 60 indicated that the real interest rate, j has an adverse effect on the
optimized safety index and the load factors.

Table 60 Optimized safety index, and load factors,
D
and
W
for different j values
(D+W combination)
j
D

W
(
optimum
)
min
(
optimum
)
max
(
optimum
)
avg
0.01 1.25 1.85 3.10 3.47 3.27
0.02 1.25 1.78 3.02 3.40 3.19
0.03 1.25 1.73 2.94 3.34 3.12
0.04 1.25 1.68 2.88 3.29 3.06
0.05 1.25 1.64 2.82 3.24 3.01
0.06 1.25 1.60 2.77 3.20 2.96




169
6.8 Effect of Failure Cost Factor, g on Optimized Load Factors and Safety Index
The deduced valued of the failure cost factor, g shows a variation along the design
space and an average value is adapted as the failure cost factor. Therefore it is important
to see the behavior of the optimized load factors for different values of failure cost factor,
g. The optimization analysis are run for g values of 863.50, 1727 and 3454, for the case
of K
W
/K
G
=3.0 considering system safety. The optimized system safety index, and load
factors are tabulated for D+W+L combination case in Table 61. As expected with
increasing failure cost factor, g, the safety index values also increase along with higher
optimized load factors. As the g increased from 863.5 to 1727, the wind load factor
increases by 0.11, and average safety index increase by 0.13. As the g increases from
1727 to 3454, the increase amounts are about same as the previous case. Increasing the g
factor makes failure cost curve to shift up and its slope to increase, so the optimum is
reached at higher safety index values; therefore the safety index increases along with load
factors. This is due to trade-off between safety and cost; increasing consequences of
failure has an impact on the optimized safety index level, the balance is obtained by
higher safety index and load factors.
The effect of deduced g on optimum load factors is also studied for D+E+L
(
E
=10.0) combination case as well and results are tabulated in Table 62. As in D+W+L
case increasing g value increases the average safety index value and the optimum load
factor. The optimum earthquake load factor increases by about 0.2 with doubling the g


170
value, and average safety index increases by about 0.16, where in D+W+L case wind
load factor increases by 0.11 and average safety index increases by 0.13 by doubling g.

Table 61 Optimized safety index and load factors for different g values (D+W+L
combination)
g
D

W

L

optimum

average
TCF
863.50 1.25 0.99 1.35 3.44-4.38 4.07 12.858
1727.00 1.25 1.10 1.35 3.64-4.48 4.20 13.008
3454.00 1.25 1.23 1.35 3.82-4.59 4.33 13.172


Table 62 Optimized safety index and load factors for different g values (D+E+L
combination)
g
D

E

L

optimum

average
TCF
863.5 1.25 1.24 0.50 3.14-3.23 3.17 17.514
1727.0 1.25 1.44 0.50 3.30-3.38 3.33 19.890
3454.0 1.25 1.66 0.50 3.45-3.52 3.48 22.633


6.9 Effect of Separation of Time Variant and Invariant Parts on Probability of
Failure Evaluation
An example is presented in order to illustrate the effect of the separation of
variables into time variant and invariant parts on the probability of failure calculations. A
reliability analysis is run for statistical data given in Table 63 and the data given for time
dependent, Q is for annual maximum. The limit state functions are (G=R-Y) for non-
separated and (G=R-XQ) for separated case. The probability of failure, safety index and
annual probability of failure vs. time are depicted in Figure 14. The difference between


171
the separated and nonseparated case especially in annual probability failures is clear with
the nonseparated case giving more conservative results.

Table 63 Statistical data for separation example
R.V. Distr. BIAS COV
R Logn. 1.12 0.10
Y Logn 0.33 0.59
X Logn 1.00 0.23
Q Logn 0.33 0.59


Figure 14 Effect of variable separation on probability of failure calculations




172
6.10 Effect of Separation of Time Variant and Invariant Parts on Optimum Load
Factors
In this section the effect of separation of the time variant and invariant random
variables on optimum load factors is studied. The optimization analyses are run for D+W
case considering the wind load without separation. In ANSI-A58 annual wind load effect
is given with a bias of 0.33 and a COV of 0.59. Optimum wind load factor is found using
the statistical data and load factors along with the safety level is given for
W
=3.0 are
given in Table 64. Also the results using the separation are given in the same table for
comparison. The results show that if not separated, the results are more conservative and
higher optimum load factors are obtained. Optimum wind load factor increase from 1.60
to 1.68 when no separation is applied.

Table 64 Comparison of separation of time variant and invariant random variables
Separated
D

W

opt
(
opt
)
avg.
TCF
Yes 1.25 1.60 3.77-4.20 3.96 4.2365
No 1.25 1.68 3.85-4.22 4.03 4.2538



6.11 Effect of COVs on Optimum Load Factors
The statistical data used is also studied for changes. The COVs effect on the
optimum load factors is studied considering changes in COVs of resistance, R, dead


173
load, D, wind speed, V and wind load modeling factor, X
W
. The deduced g value used is
1727 for all the analysis in this section.
The effect of resistance Rs COV on optimum wind load factor is studied for the
D+W case. The COV is varied from 0.05 to 0.15. The value used in calibration process
above is 0.10 and the wind load factor obtained is 1.60. As expected increasing the COV
requires higher wind load factor to minimize the sum of total cost factors. Also the
increase in COV leads to a decrease in average optimum safety indices.


Table 65 Resistance, R COV effect on optimum (D+W combination)
COV
D

W

optimum

average
TCF
0.05 1.25 1.53 3.74-4.40 4.02 4.184
0.08 1.25 1.56 3.75-4.29 3.99 4.210
0.10 1.25 1.60 3.77-4.20 3.96 4.236
0.12 1.25 1.65 3.79-4.09 3.93 4.271
0.15 1.25 1.76 3.84-3.94 3.90 4.346


The effect of dead load, D s COV on optimized solution is studied next. The
COV value used in calibration for D is 0.08. The COV value is changed around the value
used in calibration and results are tabulated in Table 66. The optimum wind load factor is
not affected with changing COV of D. The values found for different COV are nearly
identical to 1.60 obtained as the result of calibration.



174

Table 66 Dead Load, D COV effect on optimum (D+W combination)
COV
D

W

optimum

average
TCF
0.05 1.25 1.60 3.80-4.25 4.01 4.234
0.08 1.25 1.60 3.78-4.23 3.98 4.236
0.10 1.25 1.61 3.77-4.20 3.96 4.239


The effect of terms representing wind load is also studied. The effect of COV for
wind speed V is presented in Table 67 and the effect due to COV of wind load modeling
is presented in Table 68. The results show that the changes in Vs COV have quite a large
impact on the optimized wind load factor. For instance, when COV is decreased from
0.15 the value used in calibration to 0.13, the optimum wind load factor decreases to 1.42
and when it is raised to 0.17 the optimum load factor increases up to 1.79. The modeling
uncertainty also has an effect on the optimum wind load factors, but it is not as influential
as the wind speed COV. Changing the value used in calibration (0.27) to 0.20 causes a
decrease in the optimum wind load obtained to 1.43. If the COV is increased to 0.35 the
optimum wind load factor goes up to 1.75. A decrease in average optimum safety index
values is observed with increasing COV values for both random variables.
Results indicate that the optimum load factors are sensitive to COV of the wind
speed. Different COV values correspond to different sites. In order to decrease the
influence of COV on optimum load factors, a more detailed design wind speed map
might be used or the wind load factor might be multiplied with another site specific
factor.


175


Table 67 Wind Speed, V COV effect on optimum (D+W combination)
COV
D

W

optimum

average
TCF
0.13 1.25 1.42 3.80-4.25 4.01 4.108
0.14 1.25 1.51 3.78-4.23 3.98 4.171
0.15 1.25 1.60 3.77-4.20 3.96 4.236
0.16 1.25 1.70 3.75-4.17 3.94 4.304
0.17 1.25 1.79 3.74-4.14 3.92 4.373


Table 68 Wind Load Modeling, X
W
COV effect on optimum (D+W combination)
COV
D

W

optimum

average
TCF
0.20 1.25 1.43 3.78-4.36 4.03 4.113
0.25 1.25 1.52 3.76-4.31 4.00 4.180
0.27 1.25 1.60 3.77-4.20 3.96 4.236
0.30 1.25 1.63 3.74-4.26 3.97 4.259
0.35 1.25 1.75 3.73-4.20 3.93 4.351





176
7.0 CONCLUSIONS AND FUTURE RESEARCH RECOMMENDATIONS

7.1 Conclusions
An approach to calibrate structural design codes is presented in this dissertation.
Reliability Based Cost Optimization is applied to structural code calibration. Although
the reliability methods have matured there is still a debate on deciding the safety levels
when calibrating the structural design codes. This dissertation aimed to develop a
rational method to decide on the safety levels of the load combinations in structural
design code calibration.
A probabilistic total cost model is developed to represent the cost of the structural
design code over a design space. Initial and the expected failure costs are the two
contributors to the expected total cost. The expected failure cost is the multiplication of
the failure cost with the probability of failure. Since the time dependent loads are
involved in load combinations a load combination process namely, Ferry-Borges Method
is used to combine time dependent effects probabilistically. In order to avoid violating the
independence assumption in Ferry-Borges Model, the load effects are separated to its
time dependent and independent components, and Nested Reliability Analysis is used to
find the failure probability using expectation.
The trade-off between cost and safety is used for implicitly deciding on reliability
level of the load combination cases. The cost is an important aspect of the calibration
process, but because of not having enough supporting data and controversial intangible


177
costs, it has not used in calibration process explicitly. Rather, cost has been addressed in
the past intuitively.
In order to overcome the problem of failure cost estimation, the most agreed load
combination case is used as a reference case and the failure cost value that the society
accepted is deduced by using the proposed cost model. The load combination case
involving earthquake loads already accounts for the system effects such as ductility. For
other load combination cases the system safety analysis are carried out herein by
increasing element safety by an assumed amount (
system
=
element
+1). Therefore the
calibration is carried out for element safety index of 3.5 and a system safety index of 4.5.
A consistent failure cost factor, g is deduced considering both component and system
safety.
The approach presented in this dissertation is illustrated for the AASHTO Bridge
specifications. Four load combination cases are selected and the load factors are
calibrated using Reliability Based Cost Optimization. The sensitivity of the results to the
assumed or deduced parameters is studied and the optimum load factors for both
considering component and system safety are presented as a result of the calibration
process. The system safety is preferred to conclude the load factors. In the total cost
model a full failure analysis can be considered only with the system safety, therefore
recommended load factors are based on system safety.
The optimum load factors lower the sum of expected total costs from the current
code values 3%, 55% and 30% for D+W, D+W+L and D+E+L, respectively. The


178
current code load factors give low safety index values for some design samples.
Therefore, the sum of expected failure costs and the sum of expected total costs are
considerably high and these samples dominate the optimization. In order to reach a
balance between safety and cost, the load factors increase and the sum of expected total
costs decrease.
The sensitivity analyses indicate that increasing failure cost ratio, g increases the
optimum load factors and optimum average reliability index. As the consequences of a
failure increase the balance between safety and cost is attained at higher level of safety
index and higher load factors.
The marginal cost slopes also affect the optimum load factors, if the marginal cost
slope is increased, then increasing a load factor becomes more costly, therefore the
balance between the safety and cost is achieved at lower load factor values.
The optimum load factors found depends on the number of design samples and
the design ranges used in calibration.

7.2 Future Research
In order to make the approach presented herein more accurate the cost model has
to be verified by using sample bridge designs. A detailed cost and sensitivity analysis of
sample bridge designs would help to apply the method efficiently. Collecting information
about marginal cost slopes data for wind, earthquake, etc. and the normalized nominal


179
load ranges (W
n
/D
n
, L
n
/D
n
, E
n
/D
n
etc.) with the weighting factors would improve the
accuracy of the model.
A simplification is used to consider system safety herein, namely, Nowaks
equation of
sys
=
component
+1. This equation was proposed by Nowak (1993) for
gravitational loads only. It would be appropriate to investigate if it holds also for other
load cases as well.
The FORM approximation using Ferry-Borges Method is not capable to consider
changes in limit state function in time, such as the case of dead, wind and live load
combination. In the AASHTO Specification, it is assumed to have no truck load on
bridges with speeds over 55 mph. Therefore during the analysis design points (star
values) for wind speed has to be checked whether it is over 55mph or not if so, the live
load has to be dropped from the limit state function. Unfortunately Rackwitz-Fiessler
Algorithm is not capable of directly handling such a change in limit state function.
Herein, a reduction in the range of wind load to live load nominal ratios is used to
account for the 55 mph speed constraint (See Section 5.5.2). It is believed that modifying
the method with the use of truncated wind speed data would help to find a solution. The
modification of Rackwitz-Fiessler Method to account for such a case is left for future
research.









APPENDIX


181
APPENDIX

A.1. Fundamental Reliability Problem
Consider a simple structural element as in Figure A-1, that has a probability
distribution f
R
of the ultimate strength R. The load affecting the element is S and has a
distribution function f
S
.

R
S

Figure A-1 Basic Structural Element

The probability of failure, P
f
can be expressed as
( ) ( ) ( )dx x f x F S R P P
S R f


= = 0 (A-1)
Equation (A-1) is simply the probability of load effect S being greater than the available
strength capacity. This is illustrated in Figure A-2.




182
f
R
(r), f
S
(s)
r, s
S R

Figure A-2 Fundamental Reliability Example

The integration in Equation (A-1) is called the convolution integral. The integrand
in this integral is the probability of s being at the interval x and x+dx times the probability
of R being less than x. When this integration is carried out through the domain, the result
is the probability of failure, P
f
.
The easiest case is when both R and S are independent and normally distributed
with means and standard deviations
R
,
S,

R
, and
S
, respectively. The limit state
function for this case is
S R g = (A-2)
Since both the variables are normally distributed the g function is also normally
distributed as in Figure A-3. So the mean
g
and standard deviation
g
are
2 2 2
S R g S R g
+ = = (A-3)



183
f
g
(g)
g
failure Region
g<0



Figure A-3 Fundamental Reliability Example

Failure occurs when g<0 and can be expressed as
[ ]
|
|
.
|

\
|
= < =
g
g
f
g P P

0 (A-4)
where is the standard normal cumulative distribution function. The term
g
g

was
named as safety index, by Cornell (1969), and it is simply the ratio of the mean value of
g function to its standard deviation. Safety index became the representation of the
probability of failure then and Equation (A-4) can be rewritten as
( ) =
f
P (A-5)
The general reliability case mostly involves more design variables, X
i.
A general
limit state function is represented as
( )
n
x x x g G , , ,
2 1
L = (A-6)


184
G=0 gives the failure surface and failure occurs when G<0. The probability of failure, P
f

is calculated as
( )
n n X f
dx dx dx x x x f P L L L
2 1 2 1
, , ,

= (A-7)
where f
X
is the joint probability distribution of random variables X. The joint probability
function, f
X
is not available most of the time, except when the x
i
s are independent where
Equation (A-7) becomes multiple integration of the individual distribution functions.
Even if this is the case multiple integration would be too cumbersome. Therefore some
analytical approximations are derived. General limit state function is expanded about
mean values
i
x as
( ) ( ) ( )( ) L +

+ =
j j i
j i
i i
i
i
x x x x
x x
g
x x
x
g
x g G
2
2
1
(A-8)
The mean and variance of limit state function can be estimated by (Ang and Tang
(1975))
( )
) , (
, , ,
1 1
2
2 1
j i
j
n
i
n
j i
g
n
x x COV
g
x
g
x x x g G

= =

L
(A-9)
This formulation does not use probability distribution of the design variables x
i

therefore unless the random variables are normal variables and the limit state function is
linear, accurate results cannot be obtained. Also, the safety index calculated is not
invariant for different formulations of performance function. For example, limit state
function in Equation (A-2) can also be expressed as g=ln(R/S), but solutions differ when


185
Equation (A-9) is used although both formulations correspond to same performance
function; if load applied is higher than the resistance failure occurs.
Hasofer and Lind (1974) proposed a method to deal with the invariance problem,
and used design points (most probable points) instead of mean values. In this method the
random variables are converted to reduced variables with zero mean and standard
deviation of 1
( ) n i
x x
X
i
i i
i
, , 2 , 1 L =

(A-10)
This method is an iterative procedure since design points are not known in
advance. The iteration can start with the assumption of mean values as the design points
and calculate the reduced variables as
i
x
i i
i
x x
X

*
*
(A-11)
Then the derivative of the limit state function with respect to
i
x is calculated at
*
i
x . Next the direction cosines,
*
i
are calculated by using

=
|
|
.
|

\
|

|
|
.
|

\
|

=
n
i i
i
i
x
g
x
g
1
* 2
*
*
(A-12)
So the updated design point is

i i
x x i
x
*
1
*
= (A-13)


186
in terms of safety index, . Next using updated
*
i
x values rewrite limit state equation
g(X)=0 and solve it for . By using calculated, evaluate the design points in reduced
space by

* *
i i
x = (A-14)
and repeat above mentioned calculations till the convergence criteria is satisfied.
The procedure outlined above works well when the variables are normally
distributed but if the variables are non-normal the results dont reflect the actual
distributions of the variables.
Rackwitz and Fiessler (1978) proposed to use transformation to normal variables
by finding mean and standard deviation that would give same values for cumulative and
probability distribution as the original distribution at the design points. The normalized
standard deviation and mean value is given by
( ) [ ] { }
( )
( ) [ ]
N
x i i i
N
x
i i
i i N
x
i i
i
x F x
x f
x F

* 1 *
*
* 1

=
(A-15)
Then the rest of the procedure can be used as outlined in the Hasofer-Lind
Method. In calculation of safety index, , Rackwitz and Fiessler algorithm (Figure A-4)
which uses FORM approach along with Ferry-Borges Model to combine time dependent
loads (See Chapter 2) is used in this text.
The distribution type that the collected data fits better is used to define the random
variables. Distribution types used in this text are normal, lognormal and extreme


187
distribution types I and II and probability distribution functions, PDF and cumulative
distribution functions, CDF of these distributions are given in Table A-1.
An example to illustrate the accuracy of the FORM is presented next. A structural
element with resistance, R is subjected to one time independent load, S
1
and a time
dependent load, S
2
. Resistance, R has a mean value of 2.30 and a COV of 0.10, load, S
1

has a mean value of 1.0 and a COV of 0.08. The time dependent load, S
2
has a mean
value of 0.33 and a COV of 0.59 for annual maximum. All three random variables are
distributed normally, and the limit state function is given as g=R-S1-S2. The safety index
values are evaluated for 10 years and the result are presented in Figure A-5 along with the
Monte Carlo Simulation results. The results for both FORM and Monte Carlo Simulation
are pretty close to each other. FORM analysis are pretty accurate and very efficient
cputime wise and preferred in this study.


188

Figure A-4 Rackwitz-Fiessler Algorithm (Rakcwitz and Fiessler (1978))



189


Table A-1 Probability distributions
Type PDF CDF
Normal
( )
(
(

|
.
|

\
|
=
2
2
1
exp
2
1


x
x f
( )
|
.
|

\
|
=

x
x F
Log-Normal
( )
(
(

|
|
.
|

\
|
=
2
ln
2
1
exp
2
1


x
x
x f
( )
|
|
.
|

\
|
=

x
x F
ln

Extreme Type I ( )
( ) ( )
[ ]
n n n n
u x u x
n
e e x f

=

exp ( )
( )
[ ]
n
u x
e x F

=

exp
Extreme Type II
( )
(
(

|
.
|

\
|
|
.
|

\
|
=
+ k
n
k
n
n
x
v
x
v
v
k
x f exp
1
( )
(
(

|
.
|

\
|
=
k
n
x
v
x F exp
is mean, is standard deviation, un is the characteristic largest value of the variate X, n is an inverse measure of
dispersion of X , vn is the characteristic largest value for Extr. Type II and k is the shape parameter, 1/k is a measure of
dispersion and
2
2
1
ln =

|
|
.
|

\
|
+ =
2
2
2
1 ln











190

Figure A-5 Comparison of FORM results with Monte Carlo Simulation












BIBLIOGRAPHY


192
BIBLIOGRAPHY
AASHTO (1994). LRFD Bridge Design Specifications. American Association of State
Highway and Transportation Officials, Washington, DC.
AISC (1986). Manual of Steel Construction- Load and Resistance Factor Design.
American Institute of Steel Construction, Chicago.
Aktas, E., Moses, F., Ghosn, M. (2000). Calibration of Load Factors using Reliability
Based Cost Optimization Proceedings of the 8
th
ASCE Joint Specialty Conference on
Probabilistic Mechanics and Structural Reliability, University of Notre Dame, Notre
Dame. CD-ROM Proceedings (Kareem, Haldar, Spencer and Johnson eds.) paper
PMC2000-042, 6 pages.
Aktas, E., Moses, F., Ghosn, M. (2001a). Structural Design Codes for Multiple Threats:
A Reliability and Cost Optimization Approach Proceedings of the 8
th
International
Conference on Structural Safety and Reliability, Newport Beach, CA. To be appear on
CD-ROM proceedings, A. A. Balkema Publ, Rotterdam, The Netherlands.
Aktas, E., Moses, F., Ghosn, M. (2001b). Cost and Safety Optimization of Structural
Design Specifications Reliability Engineering and System Safety, Vol. 73, No. 3, pp.
205-212.
Ang A. HS., De Leon, D. (1997). Determination of Optimal Target Reliabilities for
Design and Upgrading of Structures Structural Safety, Vol. 19, No. 1, pp. 91-103.
Ang, A. H-S., Tang, W. (1975). Probability Concepts in Engineering Planning and
Design, Volume I- Basic Principles. John Wiley & Sons, New York.
API RP 2A-LRFD (1989). Recommended Practice for Planning, Designing and
Constructing Fixed Offshore Platforms- Load and Resistance Factor Design. American
Petroleum Institute.
Arora, J.S. (1989). Introduction to Optimum Design. McGraw-Hill Book Co., New York.
ASCE (1996). Minimum Design Loads for Buildings and Other Structures, ANSI/ASCE
7-95.American Society of Civil Engineers, New York.
Barker, R.M., Puckett, J.A. (1997). Design of Highway Bridges, Based on AASHTO
LRFD, Bridge Design Specifications. John Wiley & Sons, Inc., New York.
Belk, C.A., Bennett, R.M. (1991). Macro Wind Parameters for Load Combination,
Journal of Structural Engineering , Vol. 17, No. 9, pp. 2742-2759.


193
Cornell, C. A. (1969). A Probability-Based Structural Code. Journal of the American
Concrete Institute. Vol. 66, pp. 974-985.
Der Kiureghian, A. (1990). Bayesian Analysis of Model Uncertainty in Structural
Reliability Reliability and Optimization of structural Systems, Proceedings of the 3
rd

IFIP WG 7.5 Conference, Berkeley, CA. A. Der Kiureghian, P. Thoft-Christensen eds.,
Springer-Verlag, New York.
Ditlevsen, O. (1997). "Structural Reliability Codes for Probabilistic design- a debate
paper based on elementary reliability and decision analysis concepts" Structural Safety,
Vol. 19, No. 3, pp. 253-270.
Ellingwood, B. (1994). "Probability-based Codified Design: Past Accomplishments and
Future Challenges", Structural Safety Vol. 13, pp. 159-176, 1994.
Ellingwood, B., Galambos, T.V., MacGregor, J.G., Cornell, C.A. (1980). Development of
a Probability Based Load Criterion for American National Standard A58. National
Bureau of Standards Special publication No. 577, Washington, D.C.
Ellingwood, B., Kanda, J. (1991). Formulationof Load Factors Based on Optimum
Reliability Structural Safety, Vol. 9, pp. 197-210.
Ellingwood, B., Rosowsky, D. (1996). Combining Snow and earthquake Loads for
Limit State Design ASCE Journal of Structural Engineering, Vol. 122, No. 11, pp. 1364-
1368.
Ersoy, U. (1991). Reinforced Concrete. Middle East Technical University, Ankara.
Ferrito, J. M. (1984). Economics of Seismic Design for New buildings, Journal of
Structural Engineer, Vol. 110, No. 12, pp. 2925-2938.
Forssell, C. (1924). "Economy and Construction" Sunt Fornoft Vol. 4, pp. 74-77 (in
Swedish). Translated in English by N. C. Lind, in Structural Reliability and Codified
Design" SM Study No. 3, N. C. Lind Ed. University of Waterloo, Waterloo, Ontario,
Canada (1970).
Freudenthal, A. M. (1947). The Safety of Structures, ASCE Transactions, Vol. 112, pp.
125-159.
Freudenthal, A. M. (1956). Safety and Probability of Structural Failure. Transactions
ASCE, 121, pp. 1337-1375.
Freudenthal, A. M., Garrelts, J. M., Shinozuka, M. (1966). The Analysis of Structural
Safety. Journal of the Structural Division, ASCE. Vol. 92, No. ST1, pp. 267-325.


194
Galambos, T. V., Ravindra, M. K. (1978). Load and Resistance Factor Design Journal
of the Structural Division, ASCE, ST9, pp. 1335-1336.
Ghosn, M., Moses, F., Wang, J. (2001). Design of Highway Bridges for Extreme Events,
Draft Final Report for TRB, NAS-NRC. Not published.
Gies, J. (1963). Bridges and men. Doubleday & Company, Inc. New York.
Hoogenboom P.C.J., Kasbergen, C. (1998). A Consistency Check for Safety Factors
Proceedings of the 17
th
International Conference on Offshore Mechanics and Arctic
Engineering, Lisbon Portugal, Paper No. 1352, CD-ROM.
Hwang, E-S., Nowak, A. S. (1991). Simulation of Dynamic Load for Bridges, ASCE
Journal of Structural Engineering, Vol. 117, No. 5, pp. 1413-1434.
Kanda J. (1996). Normalized Failure Cost as a Measure of Structure Importance
Nuclear Engineering and Design, Vol. 160, pp. 299-305.
Kanda J. and Ellingwood B. (1991). Formulation of Load Factors Based on Optimum
Reliability Structural Safety, Vol. 9, pp. 197-210.
Kirsch U. (1993). Structural Optimization, Springer-Verlag, New York.
Lee, J-C., Pires, J. A., Ang, A. H-S.(1998). Optimal Target reliability and Development
of Cost-Effective Aseismic Design Criteria for a Class of R.C. Shear-Wall Structures
Lind N. C. (1976). Approximate Analysis and Economics of Structures. Journal of
Structural Division, ASCE, Vol. 102, No. 6, pp. 1177-1196.
Lind, N. C. (1970). "Deterministic Formats for the Probabilistic Design of Structures"
SM Study No. 1 M. Z. Cohn, Ed. University of Waterloo, Waterloo, Ontario, Canada.
Lind, N. C. (1977). Formulation of Probabilistic Design J. of Engineering Mechanics
Division, ASCE, Vol. 103, No. EM2, pp. 273-284.
MATLAB (1999). Optimization Toolbox Users Guide Version 2, The MathWorks Inc.
Melchers, R. B. (1999). Structural Reliability Analysis and Prediction, 2
nd
Ed. John
Wiley & Sons, New-York.
Moses, F. (1969) "Approaches to structural reliability and optimization" SM Study No. 1
M. Z. Cohn, Ed. University of Waterloo, Waterloo, Ontario, Canada.
Moses, F. (1989). Effects on Bridges of Alternative Truck Configurations and Weights.
Final Report, TRB, National Research Council, Washington, DC.


195
Moses, F., Kinser, D. E. (1967). "Optimum Structural Design with Failure Probability
Constraints" AIAA Journal, Vol. 5, No. 6, pp. 1152-1158.
Moses, F., Stevenson, J. D. (1970). "Reliability-Based Structural Design" Journal of the
Structural Division, ASCE, Vol. 96, No. 2, pp. 221-244.
NEHRP (1988). Natinal Aearthquake Hazards Reduction Program, 1988.Recommended
Provisions for Seismic Regulations for New Buildings and other Structures , Federal
Emergency Management Agency, FEMA 302, Building Safety Council, Washington DC.
Nowak A. S., Lind Niels C. D. (1995). ProbabilityBased Design Codes. Probabilistic
Structural Mechanics Handbook, Theory and Industrial Applications C. R. Sundararajan
edition, Chapman & Hall, New York.
Nowak, A. S. (1993). Calibration of LRFD Bridge Design Code. Final Report, NCHRP,
12-33, Transportation Research Board, Washington, D.C.
Rackwitz, R. (2000). Optimization- the basis of code-making and reliability
verification Structural Safety, Vol. 22, pp. 27-60.
Rackwitz, R., Fiessler, B. (1978). Structural Reliability under Combined Random Load
Sequences Computers and Strcutures, Vol.9, pp. 489-494.
Ravindra M. K., Lind N. C. (1971). Optimization of a Structural Code. Report No. 72,
Solid Mechanics Division, University of Waterloo, Waterloo, Ontario, Canada.
Rosenblueth, E., Esteva, L. (1971). Reliability optimization in isostatic structures.
Journalof the Engineering Mechanics Division, ASCE, Vol. 97, No. 6, pp. 1625-1640.
Schittowski, K. (1985). NLQPL: A FORTRAN-Subroutine Solving Constrained
Nonlinear Programming Problems Annals of Operations Researchm Vol. 5, pp. 485-
500.
Turkstra, C. J. (1970). Theory of Structural Design Decisions. SM Study No. 2. N. C.
Lind Ed. University of Waterloo, Waterloo, Ontario, Canada.
Turkstra, C. J., Madsen, H.O. (1980). Load Combination in Codified Structural Design
Journal of Structural Division (ST12), ASCE Vol.106, pp. 2527-2543.
Viscusi, W. K. (1993). The Value of Risks to Life and Health, Journal of Economic
Literature, Vol. 31, Issue 4, pp. 1912-1946.
Wen, Y. K., Kang, Y. J. (1997). Minimum life-cycle cost design criteria, Advances in
Structural Optimization Proceedings of the US-Japan Joint Seminar on Structural
Optimization, ASCE, New York, pp. 192-203.


196
Wen, Y.K. (1977). Statistical Combination of Extreme Loads, Journal of Structural
Engineering, ASCE, Vol . 103, No. 5, pp. 1079-1095.
Wen, Y.K. (1981). A Clustering Model for Correlated Load Processes, Journal of
Structural Engineering, ASCE, Vol . 107, No. 5, pp. 965-983.

S-ar putea să vă placă și