Reliability-Based Structural Design Optimization

for Nonlinear Structures in OpenSees

by

Hong Liang
B.Eng., Tongji University, China, 1993
M.Eng., Tongji University, China, 1996

A THESIS SUBMITTED IN PARTIAL F U L F I L L M E N T OF

THE REQUIREMENT FOR THE DEGREE OF
M A S T E R OF APPLIED SCIENCE
in

THE F A C U L T Y OF G R A D U A T E STUDIES
Civil Engineering

THE UNIVERSITY OF BRITISH C O L U M B I A

May 2005
Hong Liang, 2005

Abstract

The aspiration of this thesis is to provide a tool for engineers in making rational decisions
based on the balance between cost and safety. This objective is accomplished by merging
the optimization and reliability analyses with sophisticated finite element models that
predict structural response. In particular, two state-of-the-art reliability-based design
optimization approaches are implemented in OpenSees, a modern and comprehensive
finite element software that has recently been extended with reliability and response
sensitivity analysis capabilities. These new implementations enable reliability-based
design optimization for comprehensive real-world structures that exhibit nonlinear
behaviour.
This thesis considers the problem of minimizing the initial cost plus the expected cost
of failure subject to reliability and structural constraints. This involves reliability terms in
both objective and constraint functions. In the two implemented approaches, the
reliability analysis and the optimization evaluation are decoupled, although they are not
bi-level approaches, thus allowing flexibility in the choice of the optimization algorithm
and the reliability method. Both solution approaches employ the same reformulation of
the optimization problem into a deterministic optimization problem. The decoupled
sequential approach using the method of outer approximation (DSA-MOOA) applies a
semi-infinite optimization algorithm to solve this deterministic optimization problem. An
important feature of the DSA-MOOA approach is that a convergence proof exists in the
first-order approximation. The simplified decoupled sequential approach (DSA-S)
utilizes an inequality constrained optimization algorithm to solve the deterministic
optimization problem. The DSA-S approach is demonstrated to result in a consistent
design, which lacks the convergence proof but requires less computational time than the
DSA-MOOA approach.
The gradients of the finite element response with respect to model parameters are
needed in reliability-based design optimization. These gradients are obtained using the
ii

direct differentiation method, which entails the derivation and implementation of

analytical derivatives of the finite element response. The potential negative effect of
response gradient discontinuities due to sudden yielding events is stressed in the thesis.
The problem is remedied through the use of the smooth material model and a section
discretization scheme.

Object-oriented programming is utilized when extending

optimization and sensitivity capabilities to OpenSees. The superior extensibility and

maintainability features of this approach are emphasized.
A numerical example involving a nonlinearfiniteelement analysis of a three-bay, sixstorey building is presented in the thesis to demonstrate new implementations in
OpenSees. Three cases are studied: a linear pushover analysis using

elasticBeam

elements, a nonlinear pushover analysis using beamWithHinges elements, and a nonlinear

pushover analysis using dispBeamColumn

elements with fibre sections. This thesis also

touches on practical experiences by comparing two implemented approaches, two

gradient computation methods, and linear and nonlinear analyses. The experience of
speeding up the convergence procedure by removing inactive constraints and scaling the
involved functions is also discussed.

iii

Contents

List of Tables

vii

List of Figures

viii

Conventions and Symbols

ix

Acknowledgements

Introduction

1.1 Reliability-Based Optimization Problems

1.2 Solution Algorithms

1.3 Thesis Organization

17

Finite Element Reliability Analysis

20

2.1 First-Order and Second-Order Reliability Methods

22

2.2 Monte Carlo and Importance Sampling

25

2.3 Gradient of the Failure Probability

26

Optimization Theory and Algorithms

28

3.1 Inequality Constrained Optimization Problem

28

3.1.1 First-Order Optimality Conditions

3.1.2

The Polak-He Algorithm

28
32

3.2 Semi-infinite Optimization Problem

34

3.2.1 First-Order Optimality Conditions

35

3.2.2

36

Method of Outer Approximation Algorithm

The OpenSees Software

40

4.1 Nonlinear Finite Element Analysis

41

4.2 Reliability Analysis

44
iv

Response Sensitivity for Nonlinear Structures

48

5.1 Finite Difference Method

48

5.2 Direct Differentiation Method

49

5.3 Object-Oriented Implementation in OpenSees

51

5.4 Continuity of Response Gradients

53

5.4.1

Smooth Material Model

54

5.4.2

Section Discretization Scheme

55

Implementation of Reliability-Based Design Optimization

57

6.1 Problem Reformulation

57

6.2 DSA-MOOA Approach

61

6.2.1

B1 - Inner Approximation

6.2.2

B2 - Constraints Expansion

66

6.2.3

B3 - Outer Approximation

68

6.3 DSA-S Approach

64

71

6.4 Object-Oriented Implementation in OpenSees

74

Numerical Examples and Case Studies

79

7.1 Six-Story Ductile Moment Resisting Frame

79

7.1.1

Case 1: Elastic Analysis using elasticBeam Element

81

7.1.2

Case 2: Nonlinear Analysis using beam WithHinges Element

88

7.1.3

Case 3: Nonlinear Analysis using

dispBeamColumn

Element and

Section

Fiber

95

7.2 Practical Experience from Case Studies

103

7.2.1

Comparison of Two Optimization Approaches

103

7.2.2

Comparison of Two Gradient Computation Methods

104

7.2.3

Comparison of Linear and Nonlinear Analyses

105

7.2.4

Active and Inactive Constraints

107

7.2.5

Acceleration of Convergence Procedure by Proper Scaling

108

Conclusions

110

8.1 Summary of Maj or Findings

110

8.2 Further Studies

113

Bibliography

115

Appendix A: Detailed Software Implementation

120

A.l Calling Fortran Routines from C++

A.2 Building LSSOL.LIB

120
122

A. 3 Extending OpenSees with RBDO Capacity

123

Appendix B: User's Guide to Optimization Analysis

129

B. l RBDO Modeling

129

B.2 Analysis Tools

131

B.3 Analysis Extension and Results

132

vi

List of Tables

Table 7.1 Vertical loads and lateral loads

81

Table 7.2 Definition and initial values of design variables for Cases 1 and 2

83

Table 7.3 Statistics of random variables in Case 1

83

Table 7.4a Results from RBDO analysis for Case 1

85

Table 7.4b Results from RBDO analysis for Case 1 (continued)

85

Table 7.5 Comparison of computational time for Case 1

87

Table 7.6 Statistics of random variables for Case 2

89

Table 7.7a Results from RBDO analysis for Case 2

92

Table 7.7b Results from RBDO analysis for Case 2 (continued)

92

Table 7.8 Comparison of computational time for Case 2

95

Table 7.9 Definition and initial values of design variables for Case 3

97

Table 7.10 Statistics of random variables for Case 3

98

Table 7.1 la ResultsfromRBDO analysis for Case 3

100

Table 7.1 lb Results from RBDO analysis for Case 3 (continued)

100

Table 7.1 lc Results from RBDO analysis for Case 3 (continued)

100

Table 7.12 Comparison of computational time for Case 3

102

Table 7.13 Comparison of linear and nonlinear cases

106

Table A.la New and modified classes for extending RBDO

126

Table A. lb New and modified classes for extending RBDO (continued)

127

Table A.lc New and modified classes for extending RBDO (continued)

128

vii

List of Figures

Figure 2.1 MPP searching algorithm in finite element reliability analysis

23

Figure 3.1 Local and global optimal points

29

Figure 3.2 Constrained optimization problems

30

Figure 4.1 Principal OpenSees objects

41

Figure 4.2 Element, section and material relationship

42

Figure 4.3 Fibre-section examples

43

Figure 4.4 Elastic-perfectly plastic material (FyP =0) and SteelOl material

44

Figure 4.5 Software framework for reliability analysis in OpenSees

45

Figure 5.1 The framework of response sensitivity in OpenSees

51

Figure 5.2 Bi-linear steel material model smoothed with circular segment

54

Figure 6.1 Flow chart of DSA-MOOA approach

62

Figure 6.2 ju solutions for inner approximation using the Polak-He algorithm

65

Figure 6.3 Reliability constraints set y/ (x)

67

Figure 6.4 Flow chart of DSA-S approach

72

Figure 6.5 New objects for optimization analysis in ReliabilityDomain

75

Figure 6.6 Software framework for optimization analysis in OpenSees

76

Figure 6.7 Interaction between optimization, reliability, and finite element module

77

Figure 7.1 Ductile moment-resisting frame model

80

Figure 7.2 Equal displacement principle

82

Figure 7.3 Structural responses for Case 1

86

Figure 7.4 beamWithHinges element

88

Figure 7.5 Evolution of the total expected cost for objective functions for Case 2

93

Figure 7.6 Structural responses for Case 2

94

Figure 7.7 Typical fibre sections for columns and beams

96

Figure 7.8 Structural responses for Case 3

101

viii

Conventions and Symbols

<p(-) denotes the standard normal probability density function.

O(-) denotes the standard normal cumulative distribution function.
Superscript denotes the transpose of a matrix.
T

Vectors and matrices denotes by roman letters are in bold.

Xj denotes they'-th component of a vector x.
(x,y) denotes the inner product of vectors x and y.
||x|| denotes the norm of vector x and is defined by ||x|| = ( x , x ) " .
2

a+ = max{0, a} for any a.

ix

Acknowledgements

I wish to express my deep appreciation to my supervisor Dr. Terje Haukaas for his
willingness to guide me through the challenging path towards a master degree. His
approach to research and teaching will always be a source of inspiration. He leads me
into the world of reliability and sensitivity analysis. His serious work manner and
optimistic life attitude will keep inspiring me on my personal and professional
development.
I am grateful to Dr. Johannes Ovrelid Royset for his patient and detailed explanation of
the optimization theory. His kind help makes my research work smooth and possible.
I would like to thank Dr. Sigi Stiemer for being willing to serve in my thesis
committee. His encouragement, and constructive, criticism are much appreciated. His
course gives me a lot of helps understanding the nonlinear and plastic theory of
structures.
I am thankful to my parents and brother for their encouragement and support. In
particular, I am indebted to Ling Zhu, who takes care of me as a wife, discusses my
research as a classmate and continuously gives me strength and confidence as a soul
mate.

To Ling and my parents

xi

Chapter 1

Introduction

The primary objective of structural engineering is a structural design that represents an

optimal balance between cost and safety. Traditionally, this problem has been addressed
through experience, trial and error, and ad-hoc comparisons of different designs. In these
approaches, a comprehensive exploration of design alternatives is not performed, and
uncertainties are not accounted for in a refined and consistent manner. In recent decades,
the optimization theory has been developed to find the optimal design in the
mathematical framework of minimizing an objective function subject to constraints. The
intention of this thesis is to implement, demonstrate, and improve state-of-the-art
algorithms for finding safe and optimal designs, as well as apply these implementations
to real-world structures exhibiting nonlinear behaviour.
A number of reliability-based design optimization (RBDO) approaches has been
developed, such as response surface methods as well as gradient-free and gradient-based
algorithms. These may be utilized to minimize the total volume or the total expected cost
of the structure subject to structural reliability constraints, to maximize the structural
safety subject to a given structure cost, or simply to achieve a target structural reliability.
The approaches adopted in this thesis are to employ sophisticated structural models, as
well as advanced reliability methods, to account for uncertainty.
In simulations of structural behaviour, unavoidable uncertainties are present in the
material, geometry, and load parameters, as well as in the model itself and the analysis
procedures. These uncertainties can be significant in determining the performance of a
structure and must be accounted for to ensure safe and reliable structures. Reliability
methods have been devised to estimate the probability of response events for random
structural properties and loads (Ditlevsen & Madsen, 1996). Reliability methods such as
first-order reliability methods and importance sampling are employed in this thesis to
evaluate structural reliability in R B D O analyses.

The finite element method is currently the leading-edge approach for numerical
simulations of structural behaviour. It is of considerable interest to incorporate
sophisticated finite element models into the RBDO analysis. Furthermore, this thesis
addresses the need for implementation of state-of-the-art optimization techniques in a
finite element code that is in widespread use. Flexible software architecture is required to
accommodate the extensive interaction between optimization, reliability, and finite
element modules of the software. OpenSees - open system for earthquake engineering
simulations - (McKenna et al., 2004), is ideal for this purpose. This is an object-oriented,
open-source software that is freely available from http://opensees.berkeley.edu. It serves
as the computational platform for the prediction of structural and geotechnical responses
for the Pacific Earthquake Engineering Research Center (PEER). Recently, OpenSees
was extended with reliability and response sensitivity analysis capabilities (Haukaas &
Der Kiureghian, 2004). This allows reliability analyses to be conducted in conjunction
with static and dynamic inelastic finite element analyses, with random material,
geometry, and load parameters.
A novelty of this thesis is the use of the object-oriented programming approach to
develop a library of software components (tools) for optimization analysis. This approach
provides a software framework that is easily extended and maintained. Indeed, the
decoupling optimization approaches considered in this thesis take advantage of the
object-oriented approach, in which solution algorithms for reliability and optimization
problems are readily substituted by future developed solution algorithms.

1.1

Reliability-Based Optimization Problems

A specific structural design is characterized by particular values of design variables. By

definition, the values of the design variables are assumed to be at the discretion of the
designer. They typically represent geometrical dimensions and material strengths, and are
collected in the vector x. The random variables of the structural problem, such as
material, geometry, and load parameters, are collected in a separate vector v. It is noted
2

that d e s i g n v a r i a b l e s m a y represent d i s t r i b u t i o n p a r a m e t e r s o f the r a n d o m v a r i a b l e s . F o r

instance, the designer m a y w i s h to optimize the dimensions o f a girder cross-section.
H o w e v e r , the d i m e n s i o n s are uncertain d u e t o imperfect w o r k m a n s h i p , etc. H e n c e , only
the m e a n o f t h e structural d i m e n s i o n s i s at t h e d i s c r e t i o n o f t h e d e s i g n e r . S i m i l a r l y , i n
some

cases the designer m a y have

control over the dispersion o f the probability

distribution o f a r a n d o m variable through tolerance specifications tothe manufacturer o f

the c o m p o n e n t .
R B D O c a n b e c l a s s i f i e d i n t o t h r e e b r o a d p r o b l e m c a t e g o r i e s : (1) m i n i m i z a t i o n o f t h e
structural cost o r v o l u m e subject to constraints o n structural properties a n d / o r reliability;
(2) m a x i m i z a t i o n o f s t r u c t u r a l r e l i a b i l i t y s u b j e c t t o c o n s t r a i n t s o n c o s t a n d / o r s t r u c t u r a l
p r o p e r t i e s ; a n d (3) m i n i m i z a t i o n o f t h e d i s c r e p a n c y b e t w e e n s t r u c t u r a l r e l i a b i l i t y a n d
s p e c i f i e d target reliability subject to structural constraints a n d p o s s i b l e cost constraints.
I n t h e first c a t e g o r y o f p r o b l e m s , t h e c o s t m a y i n c l u d e t h e e x p e c t e d c o s t o f f a i l u r e i n
a d d i t i o n t o t h e i n i t i a l c o s t . T h e t o t a l c o s t i s w r i t t e n a s c, =c0+ cfp^,
initial cost o f the d e s i g n , c

is the present cost o f future failure,

where c
1

a n d pj-

p r o b a b i l i t y o f f a i l u r e . T h e r e l i a b i l i t y o f t h e s t r u c t u r e i s d e f i n e d a s l-pf.

is the
is the

T h e failure

probability is obtained f r o m either a c o m p o n e n t reliability p r o b l e m o r a s y s t e m reliability

p r o b l e m , a s d e s c r i b e d i n C h a p t e r 2. It s h o u l d b e n o t e d t h a t t h e c o s t a n d p r o b a b i l i t y o f
failure generally d e p e n d o n design variables: c = c , ( x ) = c ( x ) + c ( x ) / ? ^ ( x ) .
0

Several

p r o b l e m s are identified i n this category:

. x = a r g m i n j c (x) + (x)p (x)

0

=argmin{

C[

c (x) + cf(x)pf(x)
0

| f(x)<0

I f(x)<0,

(x)<p

Pf

is obtained for the case of continuous compounding by Sexsmith (1983): c = c

where c

f f u t u r e

(1.1)

(1.2)

f f u t u r e

u l(i + u),

is the future cost, / is the real interest rate (excluding inflation), and U is the rate of

occurrence of the Poisson process that describes the hazard.

arg mm

Co( ) + ^ k( )Pk( )
x

| W^O,

p (x)<p
k

where x* is the optimal design, f is the vector of structural constraints, p

target probability, K denotes the number of failure modes, and c

(1.3)

denotes the

and p are the cost of

k

failure and the probability of failure of the k failure mode, respectively. A n example of a
th

structural constraint is / - d -d ,
0

where d is a structural dimension and do is the

prescribed upper bound of d.

The view adopted in this thesis is that the problem in Eq. (1.1) is the fundamental
problem in RBDO. Eq. (1.1) seeks to minimize the total expected cost, which implicitly
includes structural reliability, in light of various constraints to ensure an esthetical and
functional design. This strategy finds its analogy in the field of decision analysis
(Benjamin & Cornell, 1970), where rational decisions are made based on the expected
utility of the decision alternatives. Consequently, the optimal balance between cost and
safety is achieved. Theoretically, no constraint on the reliability is needed, provided that
the cost of failure is appropriately defined. However, defining the appropriate cost of
failure is a key problem in modern RBDO. The cost of failure potentially includes the
value of human life and other intangible costs. For this reason it is useful to introduce Eq.
(1.2), which includes a reliability constraint to ensure that the design conforms to
minimum safety requirements.
Eq. (1.3) addresses problems where multiple failure modes are possible, each with
individual failure costs. This is different from Eqs. (1.1) and (1.2), where multiple failure
modes may be present in the system reliability 1 - pj-, but where the failure cost is
associated with the global system failure. A weakness of the formulation in Eq. (1.3) is
the implicit assumption of independence between failure modes, which is rarely satisfied
in practice.
Simplified versions of the problems in Eqs. (1.1) to (1.3) are obtained by minimizing
only the initial design cost:

x*=argmin{

c ( x ) | f(x)<0 }

(1.4)

x*=argmin{ c ( x ) | f (x) < 0, p (x)<p

0

(1.5)

These problems are frequently addressed in engineering practice because they avoid the
need for assessing the failure cost. In fact, E q . (1.4) denotes the well-known deterministic
(non-RBDO) design optimization problem for which uncertainty is not accounted.
Relative to E q . (1.4), the problem in Eq. (1.5) introduces a safety constraint for which the
reliability analysis is required. This is, conceptually, the simplest R B D O problem.
The second category of R B D O problems is identified as
x*=argmin{
x*=argmin{

x*=argmin{ p (x)
f

| f (x) < 0 }

p (x)
f

(1.6)

p (x) | f(x)<0, c ( x ) < c

f

| f (x) < 0, c (x) + c (x)p (x)<c

0

(1.7)

(1.8)

where c is the prescribed upper bound o f the cost. A n extended set o f problems is
formulated by replacing

p (x)
f

with

max p ( x ) ;
k

namely,

the maximum failure

probability over all failure modes. The problems in Eqs. (1.6) to (1.8) then turn into minmax type problems.
The problem in E q . (1.6) seeks to maximize the reliability given structural constraints.

While this is sometimes referred to as the inverse reliability problem in the literature, we
reserve this term for a problem introduced below. In Eqs. (1.7) and (1.8) the initial cost
and the total expected cost are introduced as constraints. Hence, these two equations are
counterparts to Eqs. (1.5) and (1.2), respectively. However, although Eqs. (1.7) and (1.8)
represent the "flipped" version of Eqs. (1.5) and (1.2), they are not equivalent problems.
That is, the optimal design achieved by addressing Eqs. (1.8) and (1.2) is generally
different.
The third category of R B D O problems contains what is referred to as inverse reliability
problems (Der Kiureghian et al., 1994; L i & Foschi, 1998). Here, the discrepancy

between structural reliability and prescribed target reliability is minimized, subject to

constraints:
x*=argmin{ ^ ( x ) - ^
x * = a r g m i n { \p (x)-p \
f

x * = a r g m i n { \p (x)-p \

| f (x) < 0 }

(1.9)

| f(x)<0, c (x)<c
0

(1.10)
<c }

(1.11)

In addition, an extended set o f problems is formulated by replacing pj- (x) -

with

| f (x) < 0, c (x) + c (x)p (x)

m a x ^ (x) - p |, where p (x) is the failure probability o f failure mode k. Ideally, the
k

value o f the objective function o f Eqs. (1.9) to (1.11) at the design point is zero. This
would imply that target reliability is achieved and that constraints are satisfied. However,
it may not be possible to achieve the reliability

l-P/

for given structural/cost

constraints. In fact, Eqs. (1.9) to (1.11) are related to the problems in Eqs. (1.6) to (1.8),
which seek maximization o f reliability rather than convergence to a target reliability. For
instance, E q . (1.6) results in design variable values that minimize failure probability,
while E q . (1.9) potentially results in design variable values that provide a less safe
design, but that complies with the prescribed reliability 1 - pj-.
In this thesis, the first category o f R B D O problems is considered. This choice is
founded on our belief that the principles o f rational decision-making should form the
basis for R B D O . Moreover, we address the problem for a single failure event in E q . (1.2).
Ideally, the problem in E q . (1.1) should be addressed. However, the difficulties in
obtaining the "true" cost o f failure make this problem less practical. The problem in E q .
(1.3) is not considered in this thesis, because only one failure cost is assumed.

1.2

Solution Algorithms

In this thesis it is assumed that the problem in E q . (1.2) is defined in terms o f a finite
element model. Specifically, the design variables x and the random variables v are

specified in terms of the input parameters of the finite element model. In this situation, a
number of challenges are present when attempting to solve Eq. (1.2):
Challenge 1. The failure probability p (x)
f

must be computed using reliability

methods that are coupled with the finite element analysis. This type of analysis is
termed the finite element reliability analysis and may be challenging in itself. A
number of reliability methods exists, all approximate. The choice of method
influences the choice of optimization algorithm and its behaviour. The failure
probability is a nonlinear function of x, regardless of whether the limit-state
function is linear. Moreover, the failure probability may not be continuously
differentiable.
Challenge 2. The structural response may be nonlinear, which is the case under
consideration in this thesis. The nonlinearities of the structural response and the
failure probability cause the objective function and constraint functions to be
nonlinear as well.
Challenge 3. The objective function, constraint functions, and the limit-state
function are implicit functions expressed by structural responses from the finite
element analysis.
Challenge 4. The most effective algorithms to solve Eq. (1.2) are gradient-based.
That is, they require the gradient of the objective function and constraint functions
with respect to x to be computed accurately and efficiently. Unless a
reformulation technique is employed, the gradient of the failure probability and
possibly thefiniteelement response must be computed. The gradient computation
may be both analytically problematic and computationally costly. Additionally,
inaccuracies in the gradients lead to convergence problems in the optimization
analysis.
Challenge 5. In this thesis, we typically consider problems including 10-100 design
variables and 10-500 random variables. It is imperative that the solution

algorithms provide feasible computation times as the number of variables grows.

This is known as the high-dimensional problem.
We conclude that candidate algorithms to address Eq. (1.2) must be assessed according to
their efficiency, accuracy, and convergence robustness. In recent years, significant
research efforts have been assigned to solve this RBDO problem. Consequently, a
number of algorithms is available. In the following we briefly review the main
approaches and comment on their performance relative to the challenges listed above,
before arriving at the approaches adopted in this thesis.
A broad class of algorithms used to address Eq. (1.2) is termed gradient-free. Their key
characteristic is that derivatives (gradients) of the objective function and constraint
functions are not required in the analysis. Instead, the algorithms evaluate the objective
and constraint functions at a number of trial points in the space of design variables.
Several classes of gradient-free algorithms are available to solve RBDO problems. One
example is genetic algorithms (Itoh & Liu, 1999; Thampan & Krishnamoorthy, 2001).
Analogous to principles of natural genetics, these algorithms are evolutionary
computation techniques. The key to genetic algorithms is the representation of a design as
a set of binary numbers. A set of designs is termed a population. Thefirstpopulation is
established randomly, through a user-defined number of candidate solutions. The
objective of the procedure is to improve the population (the set of designs) in a manner
similar to genetic operations in real nature. Three genetic operators (reproduction,
crossover, and mutation) are used to create a new population (generation). During this
process, the candidate solutions with the better designs are selected as parents to produce
the next generation with improved design. The more genetic operations are present, the
better solutions are achieved. Usually, the user stops the optimization at a predefined
maximum number of operations.
In the following we address the advantages and disadvantages of generic algorithms
based on their performance relative to the aforementioned challenges:
1. The failure probability is computed using any available finite element reliability
methods.
8

2. Genetic algorithms are applicable to complex and nonlinear RBDO problems.

3. Genetic algorithms have the ability to couple with the finite element analysis.
4. Genetic algorithms avoid the convergence problems caused by inaccurate or
discontinuous response gradients because of their gradient-free properties.
5. Genetic algorithms are computationally expensive and converge slowly since they
require many more objective function evaluations than some alternative
techniques. Their computational effort increases rapidly as the number of design
variables increases. Therefore, genetic algorithms are not suitable for highdimensional problems.
The optimal solutions are not highly accurate because the representation of design
variables as binary numbers does not allow high accuracy. Genetic algorithms
have been suggested to narrow down the optimal solution region, followed by a
traditional optimization method to identify a more precise optimum.
Genetic algorithms are able to address problems with continuous and discrete
design variables. An example of a discrete design variable is the cross-section
type within a class of steel cross-sections. The problem with discrete design
variables is not considered in this thesis.
The response surface method is an alternative to the gradient-free algorithms. The use
of response surface methods is generally contemplated in problems where the objective
function, and possibly the constraints, is not explicitly defined. This is indeed the case in
Eq. (1.2). Even when cost functions are explicitly defined, a reliability analysis is
required to obtain the probability of failure. In the response surface method, the value of
the function in question is evaluated using a number of realizations of the design variable
vector x. An explicit algebraic continuously differentiable expression is then used to fit to
these points in the space of design variables. The reformulated optimization problem is
then solved using standard nonlinear gradient-based optimization algorithms. The
gradients are readily found because response surfaces are simple algebraic expressions.
An important advantage of the response surface method is that sensitivities of the
9

functions themselves are unnecessary. Hence, the response surface method may be
termed quasi

gradient-free.

As an example of a response surface method, Gasser and Schueller (1998) approximate

the failure probability using an exponential function dependent on the design variables.
This approximation makes the failure probability numerically continuous. Torczon and
Trosset (1998) use an algebraic function to approximate the objective function in the
optimization problem. The approximate functions are constructed from numerous
evaluations of the objective function. Eldred et al. (2002) employ approximate functions
for both the objective function and the limit-state function.
Response surface methods have the following advantages and disadvantages relative to
the challenges listed above:
1. The failure probability is computed using any available finite element reliability
method.
2. Response surface methods are numerically robust since the reformulated failure
probabilities, objective functions, or limit-state functions are explicit and
continuous and can be solved using standard nonlinear optimization algorithms.
However, more sampling points are required for approximating the highly
nonlinear functions in RBDO problems.
3. Response surface methods have the ability to couple with the finite element
analysis.
4. Response surface methods can solve complex problems involving discontinuous
functions in optimization since continuous functions are used to replace the
original discontinuous functions.
5. The accuracy of the optimal solution depends on the accuracy of the
approximation to the original problem. The efficiency of the optimization process
relies on the computational time of constructing response surface, which is
dependent on design variables. The numerical effort increases tremendously with

10

the increase in the number of design variables. Therefore, response surface

methods are not suitable for high-dimensional problems.
Response surface methods are likely to find a global minimizer rather than simply
identifying a local minimizer.
Although gradient-free algorithms and response surface methods are viable
alternatives, gradient-based algorithms are the most efficient in solving Eq. (1.2). In fact,
traditional nonlinear optimization techniques are gradient-based. A variety of gradientbased algorithms is proposed in the literature. These algorithms are categorized into
nested bi-level approaches, mono-level approaches, and decoupled sequential
approaches. Enevoldsen and Sorensen (1994) solve Eq. (1.2) using the nested bi-level
approach. This is the "brute force" approach, in which the standard nonlinear
optimization algorithm is employed and the failure probability is evaluated each time the
values of the objective function and of the reliability constraint are needed. (Hence the
name

nested bi-level.)

A noteworthy feature of the work by Enevoldsen and Sorensen

(1994) is that thefirst-orderreliability method (FORM) is utilized to evaluate the failure

probability. This is a reasonable approximation unless the limit-state function is strongly
nonlinear. However, in nonlinear finite element reliability analyses, which are the case
under consideration in this thesis, strong nonlinearities may be present. The nested bilevel approach requires the objective function and constraints to be continuous and to
havefirst-orderderivatives with respect to optimization variables. An advantage of using
FORM analysis is that the gradient of the failure probability with respect to design
variables is available, as shown in Chapter 2.
The nested bi-level approach has the following advantages and disadvantages relative
to the challenges listed above:
1. Failure probability is practically limited to the FORM analysis.
2. Additional computational costs are required to determine the optimal design for
highly nonlinear functions in RBDO problems.
3. The nested bi-level approach has the ability to couple with finite element analysis.
11

4. The gradient discontinuity problem may cause non-convergence or slow

convergence.
5. The computation time is large since reliability analyses must be performed in each
step of the optimization analysis. Therefore, the nested bi-level approach is not
suitable for high-dimensional problems.
Madsen and Friis Hansen (1992) and Kuschel and Rackwitz (2000) develop the monolevel approach to improve the nested bi-level approach. Since the FORM analysis on the
inner reliability level is actually an optimization problem itself (see Chapter 2), it is
substituted by its first-order optimality conditions. In this manner, the reliability problem
becomes an additional constraint in the RBDO problem. The reformulated optimization
problem is then solved using standard nonlinear optimization algorithms in the
augmented space of design variables. The mono-level approach has the following
features:
1. The failure probability is computed using the FORM analysis. Extensions to
second-order reliability methods are possible but are rather expensive.
2. Additional computational costs are required to determine the optimal design for
the highly nonlinear functions in RBDO problems.
3. The second order derivative of the limit-state function is necessary even when
usingfirst-orderoptimization algorithms. This makes the RBDO for nonlinear
structures impossible since the second order derivative of structural response with
respect to model parameter is rarely available in the finite element analysis.
4. The gradient discontinuity problem may cause non-convergence or slow
convergence.
5. The reformulated optimization has a greater number of design variables than the
original one because of additional constraints from the reliability analysis. The
reformulated problem may be too big to be solved for high-dimensional problems.

12

Standard optimization algorithms without any links to an external code are used
since the inner reliability level is eliminated.
The explicit transformation between the original space and the standard normal
space of random variables is required in the mono-level approach.
.

The mono-level approach is only applicable for the component reliability problem
and separable series systems, which do not have correlation between different
failure modes.

A "direct" mono-level approach was developed by Chen et al. (1997) and generalized
by Wang and Kodiyalam (2002) and Agarwal et al. (2003). In this approach design
variables are defined as the mean values of some random variables. Furthermore, FORM
reliability analysis is employed. Under the assumption of uncorrelated random variables,
a direct relationship is established between the design variables and the approximation
point in FORM analysis (this will later be termed

the most probable failure point)

for

fixed target reliability. The reformulated optimization problem is a deterministic

optimization problem not requiring a reliability analysis. It is solved using a standard
nonlinear optimization solver. This direct mono-level approach is appealing for several
reasons:
1. The failure probability evaluation is not required in the reformulated deterministic
optimization analysis.
2. Additional computational costs are required to determine the optimal design for
the highly nonlinear functions in RBDO problems.
3. Only first-order derivatives of the structural response with respect to design
variables are required. They are available from the finite element analysis.
4. The gradient discontinuity problem may cause non-convergence or slow
convergence.
5. The direct mono-level approach is efficient in the absence of a reliability analysis
in the optimization analysis. The computational cost is approximately the same as
13

in the deterministic optimization analysis (without adding additional constraints).

Therefore, this approach is applicable to high-dimensional problems.

The transformation between the original space and the standard normal space of
random variables does not have to be explicit.
The direct mono-level approach can only deal with mutually independent random
variables. A further study of correlated random variables is required.

Finally, we arrive at decoupled sequential approaches, some of which are considered to

be state-of-the-art solution algorithms. One decoupling approach is developed by Du and
Chen (2002) and Agarwal and Renaud (2004) to improve the nested bi-level approach.
Reliability and optimization analyses are performed separately and sequentially. That is,
a deterministic optimization is executed to obtain a new design without re-computing the
reliability. Then, the failure probability is updated by performing a reliability analysis for
the new design. The process is repeated until a consistent design is obtained. The term
decoupled is used because the analyst may freely select different methods for the

reliability and optimization problems. This is different from the mono-level approach
where FORM reliability analysis is implicit. The decoupled bi-level approach has the
following advantages and disadvantages:
1. The failure probability is computed using any available reliability methods since
the reliability analysis is decoupled from the optimization analysis.
2. Additional computational effort is required to determine the optimal design for
highly nonlinear functions in RBDO problems.
3. The decoupled bi-level approach has the ability to couple with finite element
analysis.
4. The gradient discontinuity problem may cause non-convergence or slow
convergence.

14

5. The decoupled approach is more efficient than the nested bi-level approach
because the number of reliability evaluations is significantly reduced. Therefore,
this approach is applicable to high-dimensional problems.

It is easy to code and to combine the reliability analysis with any optimization
software without having to reformulate the problem.
The true local optimal solution cannot be guaranteed because the failure
probability is always computed for a previous design.

Further improvements of the decoupled approach are developed by Kirjner-Neto et al.

(1998), Der Kiureghian and Polak (1998), and Royset et al. (2001a). Most notable of
these improvements is a decoupled sequential approach utilizing the method of outer
approximations (Polak, 1997) to solve several RBDO problems. Royset et al. (2001b,

2002, & 2004a) further develop the methodology to solve Eq. (1.2). In this approach, Eq.
(1.2) is reformulated as a semi-infinite optimization problem (Polak, 1997). This
reformulated problem has proven to be identical to the original problem, when FORM
analysis is used to compute the failure probability. Moreover, a heuristic scheme is
implemented to improve the reliability estimate. The term semi-infinite comes from the
fixed number of design variables and the infinite number of reliability constraints in the
reformulated problem. This approach has the following advantages and disadvantages:
1. The failure probability is computed using any available reliability methods since
the reliability analysis is completely decoupled from the optimization analysis.
2. Additional computational costs are required to determine the optimal design for
the highly nonlinear functions in RBDO problems.
3. The decoupled sequential approach has the ability to couple with the finite
element analysis.
4. The gradient discontinuity problem may cause non-convergence or slow
convergence.

15

5. This approach is more efficient than the nested bi-level approach. However, this
approach requires an infinite number of reliability constraints to achieve the
"true" optimal solution. This is not attainable in practice. Usually, the user stops
the optimization at a predefined accuracy and obtains an approximate solution.
Therefore, this approach is applicable to high-dimensional problems.
The method of outer approximations has proofs of convergence (Kirjner-Neto et
al., 1998). This implies that there is a convergence proof for the decoupled
sequential approach when the limit-sate function is linear in the space of random
variables.
In this thesis we implement the decoupled sequential approach in OpenSees and apply
it to structures that exhibit nonlinear behaviour. This approach is termed the decoupled
sequential approach by the method of outer approximations (DSA-MOOA). We also

implement a simplified decoupled sequential approach (DSA-S) by combining the

problem reformulation of DSA-MOOA with thefindingsof Du and Chen (2002) and
Agarwal and Renaud (2004). In the DSA-S approach, the reformulated optimization
problem is an inequality constrained optimization problem (Polak, 1997) with a single
reliability constraint, as opposed to the infinitely many constraints of the DSA-MOOA
approach. The DSA-S approach has the same properties as the DSA-MOOA approach
when facing the aforementioned challenges 1-5. However, while the DSA-S approach is
more efficient than the DSA-MOOA approach, it lacks the convergence proof of the
DSA-MOOA approach.
In additional to the above methods, Royset and Polak (2004b) develop the decoupled
sequential approach by sample average approximations to solve the RBDO problem in

Eq. (1.2). Failure probabilities are here computed using Monte Carlo or importance
sampling. Thefirst-orderderivative of the failure probability with respect to design
variables is obtained analytically and computed using Monte Carlo or importance
sampling. The original RBDO problem is reformulated as an inequality constraint
optimization problem and solved using standard optimization algorithms. The number of
samples increases as the design approaches the optimal design point. Royset and Polak
16

(2004b) prove that the optimization algorithm converges with the optimal design when
the number of samples approaches infinity. This approach has the following advantages
and disadvantages:
1. Failure probabilities are computed using Monte Carlo sampling and importance
sampling.
2. Additional computational costs are required to determine the optimal design for
the highly nonlinear functions in RBDO problems.
3. This approach has the ability to couple with the finite element analysis.
4. The gradient discontinuity problem may cause non-convergence or slow
convergence.
5. The computational cost is higher than in the DSA-MOOA approach since the
sample average approximations use Monte Carlo sampling or importance
sampling to compute the values and gradients of failure probabilities. The number
of sampling points increases as the design nears the optimal stage. Therefore, this
approach is still applicable to high-dimensional problems.

The sample average approximations have proofs of convergence even if the limitsate function is nonlinear in the space of random variables.

1.3

Thesis Organization

Following the introduction, the fundamentals of finite element reliability analysis and
optimization theory are introduced in Chapters 2 and 3, respectively. Chapter 2 reviews
the concept of finite element reliability and describes FORM, the second-order reliability
method, Monte Carlo sampling, and importance sampling. Chapter 3 presents the
inequality constraint optimization problem and the semi-infinite optimization problem, as
well as their correspondingfirst-ordernecessary optimality conditions. The Polak-He
algorithm and the method of outer approximation algorithm are also described in this
chapter.
17

Chapter 4 describes the finite element software, OpenSees, which is extended with
reliability analysis capability. The element, section, and material objects used in this
thesis are emphasized in this chapter. The reliability analysis module adds the
ReliabilityDomain

and ReliabilityAnalysis objects to the original OpenSees. The former

defines the random variables and their correlation. It also maps the values of the random
variables into the finite element model. The latter includes eight analysis types and
several analysis tools.
Chapter 5 introduces the sensitivity analysis capability in OpenSees. Two main
response sensitivity methods, the finite difference method and the direct differentiation
method, are two analysis tools in OpenSees. The direct differentiation method is stressed
in the chapter by briefly describing its equation derivation and several implementation
issues. Chapter 5 ends with a section on the continuity of response gradients. The
potential negative effects of discontinuous structural responses are observed and
remedied using two methods: the smooth material model and the section discrimination
scheme.
Chapter 6 presents the problem reformulation and algorithms of the DSA-MOOA and
DSA-S approaches. The reformulated optimization problem is identical to the original
RBDO problem when the limit-state function is linear in the space of random variables.
The applications of the method of outer approximation and the Polak-He algorithms are
described in detail in this chapter. The optimization capability of OpenSees is extended
through the addition of several objects defining all functions involved in the optimization
problem. Moreover, two analysis types (DSA-MOOA analysis and DSA-S analysis) and
several analysis tools are added.
Chapter 7 presents a numerical example involving a nonlinear finite element analysis
of a three-bay, six-storey building to demonstrate the new implementations in OpenSees.
Three cases are studied: a linear pushover analysis using elasticBeam elements, a
nonlinear pushover analysis using beam WithHinges elements, and a nonlinear pushover
analysis using dispBeamColumn elements with fibre sections. The case studies focus on
convergence performance and computational time. This chapter also presents practical
18

experience by comparing two implemented approaches, two gradient computation

methods, and linear and nonlinear analyses. Speeding up the convergence procedure by
removing inactive constraints and scaling the functions involved are also discussed.
Chapter 8 summarizes the major findings of this thesis and points to areas of future
study. Appendix A offers the detailed software implementation used in the thesis, while
Appendix B contains a user guide to the new implementations.

19

Chapter 2

Finite Element Reliability Analysis

Since the introduction of limit-state design in the early 1980s, reliability methods have
been masked by partial safety coefficients in prescriptive design code requirements.
Direct usage of reliability methods has only been observed in special projects, such as
offshore structures and nuclear power plants. This paradigm is changing with the
introduction of performance-based engineering. Over the past several decades, analytical
and numerical models have drastically improved the engineers' ability to predict
structural performance. However, such predictions can only be made in a probabilistic
sense. Unavoidable uncertainties are present in model parameters and in the numerical
model itself. Reliability analysis and probabilistic methods are therefore rapidly
becoming required tools in engineering practice. This chapter presents reliability analysis
in conjunction with nonlinear finite element models to make probabilistic predictions of
structural response. Such analysis is required to solve the reliability-based design
optimization (RBDO) problem in Eq. (1.2).
The name finite

element reliability

stems from the combination of advanced reliability

methods withfiniteelement analysis to estimate the probability of exceeding prescribed

structural response criteria. As mentioned in Chapter 1, the structural model is defined in
terms of vector v of random variables and vector x of deterministic design variables. The
structural response of interest is collected in a vector d, which depends upon random and
design variables: d=d(x,v). A failure criterion is prescribed in terms of the response
quantities d by means of a limit-state function g(d). Conventionally, a negative outcome
of the limit-state function, g < 0, is defined as a failure, while a positive outcome, g > 0,
is defined as safe. A typical limit-state function employed in this thesis is expressed by a
threshold d and a response quantity d in the form
g(d(x,v)) = d-d(x,v)

20

(2.1)

W e observe that when the response d exceeds the threshold d, the limit-state function
becomes negative, as required by the syntax rules. It is emphasized that response
quantities d may represent displacements, stresses, strains, etc.
The reliability problem described by a single limit-state function is termed the
component reliability problem. If failure is prescribed by the joint state o f several limitstate functions, the reliability problem is referred to as a system reliability

problem.

In the component reliability problem, the probability o f failure p (x) is defined by the
f

integration o f the joint probability density function (PDF) / ( v ) o f the random vector v
over the failure domain in the space of random variables:
p (x)=
f

jjfWdv

(2.2)

g(d(x,v))<0

The failure probability depends on the design variables x. In addition, we note that the
failure probability does not change i f the limit-state function is arbitrarily scaled using a
finite positive number. This is important for later developments.
Analytical solutions to E q . (2.2) are generally not available, and approximate methods
are employed to evaluate the failure probability. In these reliability methods, it is
common to transform the problem into the standard normal space. That is, the original
random variables v are transformed into a vector u o f uncorrelated standard normal
random variables. The Nataf model (Liu & Der Kiureghian, 1986) is an example o f such
a transformation. T is denoted as the transformation for a given design vector x and
x

replace the random vector v by T~ (u). W e then obtain the equivalent limit-state function
x

g(d(x, v)) = g(d(x,r

_l

(u))). The joint P D F o f u is the joint standard normal P D F (p(u).

Hence, the reliability problem is re-defined in the standard normal space as:
p (x)=
f

J]Wu)rfu

(2.3)

g(d(x,u))<0

For system reliability problems with limit-state functions g , ke {l,2,...,K], the failure
k

domain is specified as a series system, a parallel system, or a general system. In a series

21

system, failure occurs when any components (limit-state functions) fail. That is, the
failure domain is defined by the union of failure domains of components:
Ug*(d(x,u))<0

(2.4)

Conversely, the failure domain of a parallel structural system is defined by the

intersection of failure domains of all components:

r|g*(d(x,u))<0

(2.5)

For a general system, the definition of the failure domain involves both union and
interaction operations.

2.1

First-Order and Second-Order Reliability Methods

The first-order reliability method (FORM) approximates the integration boundary

g(d(x,u)) = 0 using a hyperplane in the standard normal space. The ideal point to
linearize the limit-state function is denoted as u* and is the point on the hyperplane
closest to the origin:
u*(x) = argmin{||u|| | g(d(x,u)) = 0}

(2.6)

u* is the point in the failure domain with the highest probability density and is therefore
termed the most probable

point

(MPP). Often, Eq. (2.6) is defined with the inequality

constraint g(d(x,u)) < 0 in place of the equality constraint. This is acceptable, as long as
the origin is in the safe domain. This is the case for most practical problems, where the
failure probability is much less than 0.5.
Searching for the MPP in Eq. (2.6) is an optimization problem in itself. This
optimization process requires the first-order derivative of the limit-state function to be
continuous in the standard normal space u. One effective algorithm for searching for the
MPP is the iHLRF-algorithm (Hasofer & Lind, 1974) (Rackwitz & Fiessler, 1978)
22

(Zhang & Der Kiureghian, 1997), which is gradient-based and employs a line search
scheme. In addition, Eq. (2.6) with an inequality constraint formulation can be solved
using the Polak-He algorithm (Polak, 1997) or using standard nonlinear optimization
algorithms such as NLPQL (Schittkowski, 1985) or NPSOL (Gill et al., 1998). The latter
algorithms, however, are not specialized for the present case of one constraint.
Initialize variables u, x
Transform u-> v
Evaluate g(d(x,v))

Reliability
Analysis
Module

Evaluate

dd/dv

dg _ dg dd dv
du dd dv du

Finite
Element
Analysis
Module

Take a step
= u' + step size x search direction

.Convergence check.
JYeT

Post processing
p,(x)3>(Figure 2.1 MPP searching algorithm in finite element reliability analysis
Haukaas and Der Kiureghian (2004) employ the iHLRF algorithm and the Polak-He
algorithm to find the MPP for finite element reliability problems. The outline of the
search algorithm is shown in Figure 2.1. The search algorithm requires a transformation
between the original v-space and the standard normal u-space. The value of limit-state
function g and the gradient dgldu are evaluated in this algorithm. They are used in
finding a search direction and a step size. The figure also shows the interaction between
the search algorithm and the finite element code. The finite element analysis module is
23

repeatedly updated with new realizations of the random variables v. In return, the finite
element analysis module produces the response d and the response gradients dd/b\. The
need for response derivatives is due to the need for the gradient of the limit-state function
in the standard normal space. The chain rule of differentiation yields:
dg_ dg_ddd^_
du dd dv du
=

where dg 13d is easily found because g is a simple algebraic expression in terms of d,

dd/dx are the response gradients, and dv/du is the Jacobian matrix from the probability
transformation.
The distance from the origin to the MPP in the standard normal space is called the
reliability index /?:
B(x) = \\u(x)\\

(2.8)

The FORM approximation of the failure probability reads:

p (x)*<b(-fi(x))

(2.9)

where O(-) is the standard normal cumulative distribution function.

The second-order reliability method approximates the limit-state function using a
quadratic surface that passes through the MPP. Breitung (1984) offers a simple formula
to compute the component failure probability based on such a quadratic surface
approximation:
m-l

^(^^^(Ax^no + AxK/x))-"

(2.10)

7=1

where Kj(x),j-\,...,m-\

are the principal curvatures of the limit-state surface

g(d(x,u)) = 0 at the MPP.

24

2.2

Monte Carlo and Importance Sampling

Monte Carlo sampling is an alternative method used to approximate the failure

probability p {\).
f

Monte Carlo sampling generates a family of simulated points Ui,

u/v, which are statistically independent standard normal random variables. The indicator
function /(u,.) corresponding to each simulated point is established using the following
rule: /(u,) = 1 whenever g(d(x,u)) < 0, and 7(u.) = 0 otherwise. Monte Carlo sampling
(

gives an approximation of the component failure probability p (x) by:

f

P/W-TFZAu,)

(2.11)

The quality of the solution of Monte Carlo sampling is measured using the coefficient
of variation (c.o.v.) of the probability estimate: c.o.v. = ^(1- p (x))/(N- p (x)). Monte
f

Carlo sampling is stopped at the user-defined number of samplings, or when the

coefficient of variation achieves a specified target, for example 2%.
In order to estimate small failure probabilities accurately, crude Monte Carlo sampling
requires a large number of simulations. This is because the sampling distribution of crude
Monte Carlo sampling is centered at the mean point, while failure events are inclined to
occur in the tail region of the probability distributions.
Importance sampling improves the efficiency of Monte Carlo sampling by centering
the sampling distribution near the failure domain. Usually, this centre is selected as the
MPP from FORM analysis. Importance sampling generates a family of simulated points
ui,

uw with the probability density h(u). All simulated points are statistically

independent standard normal random variables. h(u) is a joint PDF and nonzero in the
failure domain. Importance sampling gives an approximation of component failure
probability p (x) using:
f

, l I

25

^ A

(2.12)

where <p(-) is the standard normal PDF (Ditlevsen & Madsen, 1996). The coefficient of
variation of failure probability in importance sampling is defined by:

(2.13)

c.o.v.

Importance sampling is efficient and requires fewer simulations than Monte Carlo
sampling since the sampling distribution is centered on the MPP, where failure
realizations are frequently encountered.

2.3

Gradient of the Failure Probability

When gradient-based algorithms are employed in RBDO, it is important to note whether

the derivative of the failure probability can be computed and whether it is continuous.
The answers to these questions are not straightforward. In fact, the "brute force"
application of gradient-based algorithms, such as the mono-level approach explored by
Enevoldsen and Sorensen (1994), requires the gradient of failure probability with respect
to design variables to be available. In FORM analysis, an analytical formulation of
dp /dx is possible. Eq. (2.9) is differentiated to obtain:
f

dp,
ox

ox

=^[i -

=- 44- ^p)=- 4<piP)

d

ox

a
ox op

ox

(2.i4)

The derivative of the reliability index is:

dP _
dx

1 dg
|V g| dx

(2.15)

where V g is the by-product of FORM analysis and dg/dx = (dg / dd\dd / dx) is readily
u

obtained by utilizing response sensitivities from the finite element code (Hohenbichler &
Rackwitz, 1986) (Bjerager & Krenk, 1989). However, the derivative of the failure
probability in Eq. (2.14) cannot be proven to be continuous. This is because in a
26

reliability problem, the MPP may jump to a different location on the limit-state surface
g=0 due to an infinitesimal perturbation of the design variables. This jump leads to a kink
on the function p (x) in the x-space. In effect, the gradient of the failure probability in
f

the x-space is discontinuous.

Having demonstrated that the derivative of the failure probability in FORM is possible
to obtain but is not continuously differentiable, we move to the case of sampling analysis.
Formulae for the sensitivity of failure probability from sampling became available only
recently. In Royset and Polak (2004b), indicator functions 7(u.) in Eqs. (2.11) and (2.12)
(

are replaced by the joint standard normal cumulative distribution function. Then, the
reformulated failure probability is differentiated. The sensitivity of the failure probability
is expressed through the joint standard normal PDF and evaluated using Monte Carlo or
importance sampling.
It is important to note that not all gradient-based algorithms that address the problem in
Eq. (1.2) require the gradient dp Idx. In fact, as will be shown in subsequent chapters,
f

the algorithms implemented in this thesis circumvent the problem of actually computing
this gradient by using an augmented design variable to take place of the failure
probability.

27

Chapter 3

Optimization Theory and Algorithms

This thesis implements two decoupled sequential approaches to solve the reliability-based
design optimization problem expressed in Eq. (1.2). These approaches profit from the
advantages of both mono-level and bi-level approaches that were discussed in Chapter 1.
The optimization problem in Eq. (1.2) is reformulated to enable the use of the method of
outer approximation (MOOA) for semi-infinite inequality optimization problems, or the
use of the Polak-He algorithm for ordinary inequality constraint optimization problems
(Polak, 1997). This chapter defines the fundamental concepts of the optimization theory,
which forms the basis for the subsequent problem reformulation and the corresponding
solution algorithms.

3.1

Inequality Constrained Optimization Problem

This section introduces thefirst-orderoptimality conditions and the Polak-He algorithm

that builds upon them for solving the deterministic inequality constrained optimization
problem of the form
x* =argmin{

F(x)

\ f(x)<0

where x* is the design point, F(x) is the objective function,

(3.1)
andf[x)

is the maximum

value of the n -dimensional vector of constraints f(x) (Polak, 1997):

/(x) = max f(x)

3.1.1

(3.2)

First-Order Optimality Conditions

A candidate solution to any optimization problem must satisfy the optimality conditions
of the problem. These are generally

necessary

conditions, but they are not

sufficient

to

guarantee that the optimal point has been found. Moreover, optimality conditions can
28

only indicate whether a local minimum has been found. This is a fundamental problem
with any optimization algorithm; only through engineering insight and repeated searches
from different starting points, among other strategies, can we confidently state that the
global optimal point has been found. The problem is schematically illustrated in Figure
3.1, where it is shown that an objective function may have a local as well as a global
minimum. It is easy to imagine that the search algorithm may "get stuck" at the local
minimum, without realizing that another point is the actual solution to the optimization
problem. A repeated search with a new start point may reveal the global solution.
Objective function
Local minimum

Global minimum

Design variable
Figure 3.1 Local and global optimal points
Optimality conditions are not only convergence criteria for the optimization algorithm.
They are often used to construct search algorithms to solve the optimization problem.
This further motivates the following exposure of optimality conditions for different
optimization problems.
For pedagogical purposes, first consider a deterministic optimization problem without
constraints:
x* = argmin{ F(x)}

(3.3)

An optimality condition for this problem is clearly VF = 0, where VF = dF/dx is the

gradient of the objective function. This is equivalent to the requirement that a function
with one variable have a zero derivative at extremum points.
29

Next, consider a deterministic inequality constrained optimization problem with one

constraint:
x* =argmin{F(x) | /(x)<0}

(3.4)

where J[x) is the constraint. Note that a problem with the equality constraint f\x)=0 may
be reformulated into an inequality constrained problem with two constraints, f(x) < 0
and - /(x) < 0. In other words, we introduce two inequality constraints for every
equality constraint.
Contours of the
objective function

X2

=0

/>0

Figure 3.2 Constrained optimization problems

For the constrained optimization problem in Eq. (3.4), the gradient of F(x) need not
vanish at the solution point (refered to as the design point). Instead, two cases are
possible: (1) the constraint is not active at the design point, in which case
VF = 0

(3.5)

at the design point; (2) the constraint is active at the design point, in which case the
gradient of the objective function is proportional to the gradient of the constraint at the
design point. Figure 3.2, in which the minimization of an objective function with two
design variables is considered, clarifies this concept. The contours of the objective
30

function are shown as circles, while the constraint l i m i t / = 0 is shown as a line. The solid
arrows in the figure depict the gradients of objective functions and constraint functions at
certain points. A t the design point x* gradient vectors VF

and Vf

are clearly parallel,

although they point in different directions. This orthogonality at the design point between
the gradient of the objective function and the gradient of the constraint is written as

(3.6)

VF = - Vf

Mo

Mi

where ju and / i , are positive constants.

0

The two optimality conditions in Eqs. (3.5) and (3.6) are combined into one equation
by first defining an auxiliary function termed the "Lagrange function:"
Z(x) = / i F ( x ) + / i / ( x )
0

(3.7)

where ju and //, are denoted Lagrange multipliers. The method of Lagrange multipliers
0

is a traditional method of enforcing constraints in an optimization problem. This method

requires the derivatives of the Lagrange function with respect to the design variables to
be zero:
V X = / / V F + //, V / = 0
0

With the additional requirement that either fj.\ - 0 or f -0,

(3.8)
both of the above cases

(active and inactive constraint) are included. First, considering the case where the
constraint is active at the design point, E q . (3.8) can be turned into E q . (3.6) by l e t t i n g / =
0 and //, > 0 . Second, considering the case where the constraint is inactive, E q . (3.8) can
be turned into E q . (3.5) by setting ju\ = 0. In conclusion, the optimality conditions for the
problem in E q . (3.4) read
V I = 0 and / / , / = 0
where / / , / = 0 implies that either p

f<0,

>0,

Mo

and

(3.9)

or / must be zero. Additionally, we must have

A>0.

Turning to the case of multiple inequality constraints, the optimization problem reads

31

x* =argmin{ F(x) | f(x)<0 } .

(3.10)

where f(x) is the vector of constraints. In this case, one positive Lagrange multiplier is
introduced for each constraint. Consequently, the Lagrange function reads
L(x) = ju F(x) + //, /,(x) + fi Mx)
0

+ - + fi f (x) = fi F(x) + uf(x)

(3.11)

where u. is the n-dimensional vector of Lagrange multipliers. The optimality conditions

for this case are referred to as the Karush-Kuhn-Tucker conditions, which take the form
VL = 0 and uf = 0
again with f < 0, ju > 0, and
0

\i > 0

(3.12)

. These are the first-order optimality conditions for

the inequality constrained optimization problem in Eq. (3.1). We note that

VZ = p VF(x') + uVf(x*)
0

(3.13)

Additionally, we require that // + //, + fi H h /u = 1 for the reason of normalization,

0

and that all involved functions F and f are continuously differentiable (Polak, 1997).

3.1.2 The Polak-He Algorithm

The use of optimality conditions to obtain corresponding solution algorithms is addressed

by Polak (1997). The Polak-He algorithm is presented for the case of multiple objective
functions, where the objective is to minimize the maximum of the objective functions,
subject to multiple constraints. In this thesis we apply the algorithm to two cases: (1) one
objective function with one constraint; and (2) one objective function with multiple
constraints. We start with the more general latter case. The Polak-He algorithm based on
thefirst-orderoptimality conditions is introduced below.

Polak-He Algorithm
Parameters. Select parameters 0<a<l, 0</?<l, 0<S, 0<y.
Input Data. Input the initial design XQ.
32

Step 0. Set i' = 0.

Step 1. Compute the search direction vector h. (see below).
Step 2. Calculate the step size A, along the search direction h . using the Armijo rule
(

(see below).
Step 3. Update the design x

/ + 1

= x. + Ah
(

replace / by / + 1, and go to Step 1.

Step 1 computes the search direction vector h. by solving a quadratic sub-optimization

problem, which is the combination of the first-order optimality conditions in E q . (3.12).

where / ( x ) = max{0,f(x)}. E q . (3.14) minimizes a quadratic function dependent on

+

Lagrange

multipliers

and

\i, subject

to

the

linear

constraint

ju + p + ju + + // = 1. This problem is solved in a finite number o f iterations, and its

0

solutions are ju* and u.*. Then, the search direction vector h, is computed using
0

(3.15)
The Armijo rule used in Step 2 is a line search scheme, which applies a merit function
in the step-size selection. Ideally, the step size is determined at the minimum merit
function since this leads to an optimal rate o f convergence. The merit function in this case
is
M(x ,x
i

+ B%) = m x{F(x + B\)-F(x )- f(xXJ(^

a

+ B\)-f(xX}

(3.16)

where s is an integer to be introduced shortly. The step-size selection must meet the
requirement o f M(x ,x
i

+ B h )<aB 0 .
s

In E q . (3.16), the parameter s is an integer

starting from the initial value 0. The parameter s increases or decreases by unit steps until
an acceptable and maximum step size is found. Finally, the appropriate step size X is set
t

equal to the maximum value o f B and corresponding to the minimum merit function:
s

33

A,.=max{ fi' | M(x ,x +P h )<ap e

s

(3.17)

The inequality constrained optimization problem in Eq. (3.1) is a nonlinear

optimization problem. Theoretically, it requires an infinite number of iterations (step 1 to
3) to converge to an optimal solution. This is not applicable to practical problem solving.
Usually, the Polak-He algorithm is terminated when a user-defined precision s {e.g.,
10" ) is reached. When two adjacent solutions are very close (||x, -x .||<s), the merit
6

+1

function M(x .,x ) and/or the value of sub-optimization #(x.) in Eq. (3.14) is
(

(+1

sufficiently close to zero, the Polak-He algorithm is stopped, resulting in an approximate

optimal solution.
It can be shown that solutions generated by the Polak-He algorithm are optimal for the
inequality constrained optimization problem. Solution x, from the Polak-He algorithm
approaches the optimal solution x* as the number of iterations approaches infinity. Since
the Polak-He algorithm is developed from thefirst-orderoptimality conditions, the
optimality conditions in Eq. (3.12) are satisfied automatically at the point of optimal
solution. In the meantime, the sub-optimization problem in Eq. (3.14) also reaches the
minimum value, i.e., 9(x*) = 0.

3.2

Semi-Infinite Optimization Problem

This section describes thefirst-orderoptimality conditions and the MOOA algorithm for
solving the semi-infinite optimization problem of the form
x*=argmin{ y (x) | ^(x)<0 }
0

(3.18)

where y/ (x) is the objective function, and y/(x) is the maximum value of the n0

dimensional constraints (Polak, 1997). The objective and constraint functions y/ are
defined by
^(x) = max y/j(x)
34

(3.19)

and y/., j = { 0 , 1 , n } is given by

^.(x) = max^,(x,y)

(3.20)

where the function <f>. (x, y) is determined by the design vector x and the extra argument
y , which are all points in the domain Y . , i.e., y e Y . . The design vector x is finitedimensional, but the number o f functions ^ (-,y) is infinity because o f the infinite
7

number o f points y in the domain Y . . That is the reason for the term "semi-infinite." A n
example o f function ^ . ( x , y ) is the negative value o f limit-state function - g ( x , u ) , in
which x is the design vector, u is the random vector (namely point y ) , and domain Y is
y

the standard normal space. A s mentioned in Chapter 2, a positive outcome o f the limitstate function ( g > 0) is defined as safe; hence the constraints - g ( x , u ) < 0 ensure a safe
structure.

3.2.1

First-Order Optimality Conditions

First-order optimality conditions for semi-infinite optimization problems require the

involved functions <f>j to be continuously differentiable. Additionally, the domain Y , must
be bounded and closed. Similar to E q . (3.12), the necessary optimality conditions for
semi-infinite optimization problems are
0eJ>,V^.(x*)

and

7=0

f y/ (x) = 0
jMj

(3.21)

7=0

where x* denotes the optimal design, and Lagrange multipliers / / , , / / are positive0

defined. Additionally it is required that / / +

0

+ ju H
2

\- ju = 1 for the reason o f

n

normalization (Polak, 1997).

The optimality conditions in E q . (3.21) require that at least one constraint ^ ( x * ) be
active at the optimal design point. Therefore, the value o f this constraint must be zero

35

(y/j(x*) = 0), while the corresponding Lagrange multiplier must be larger than zero
(jUj > 0 ) . This situation is illustrated in Figure 3.2, where the constraint/= 0 is active at
the design point x*, and the objective function reaches minimum in the mean time.

3.2.2

Method of Outer Approximation Algorithm

The M O O A algorithm solves the semi-infinite optimization problem by repeatedly

solving a series o f inequality-constrained problems. In these inequality-constrained
problems, progressively more constraints are used to get a gradually more accurate
solution (Polak, 1997). A discretization method is used to discretize the domain Y , and
produce more constraints, in which case, the domain Y is discretized into ./V number o f
7

sub-domains Y

j<N

c: Y . . For example, a domain [0,1] is descretized as the sub-domain

{0, 1/4, 2/4, 3/4, 1} when 7V= 4. These approximate sub-domains Y

jN

are sequentially

constructed as the algorithm progresses. Then, the semi-infinite problem is approximated

as an inequality constrained problem with N constraints. When N equals infinity, the
solution o f the inequality-constrained problem is the optimal solution o f the semi-infinite
problem.
Using the above discretization method, we approximate Eqs. (3.19) and (3. 20) for a
natural number N= 1, 2, 3 , . . . as
^ ( x ) = max y/ (x)

(3.22)

if/ j (x) = max ^ , ( x , y )

(3.23)

j<N

Similar to the Polak-He algorithm, the approximation to the merit function in the
Armijo rule is denoted in the following:
M (x',x") =m a x { ^ ( x " ) - ^ ( x ) - ^ ( x ' ) , ^ ( x " ) - ^ ( x ' ) }
,

where y > 0 is a parameter, and y ( x )

w

= max{0,y/ (x)}.
N

36

(3.24)

In addition, a quadratic sub-optimization problem 0 (x) is constructed to obtain a

N

search direction h
e (x) = mmM (x,x + h)
N

(3.25)

where M (x,x + h) comes from the following definitions:

N

M (x,x + h) - maxjy/o (x,x + h)-y/ (x)- / (x) ,

N

0N

7V

fi (x,x

+ h)-y/ (x) }
N

(3.26)

fi (x,x + h) = max(/7 ,(x,x + h)

A

(3.27)

^ x ^

(3.28)

^ . ( x , x + h) = m a x j ^ x , y ) + ( V

The search direction h is defined as

7=0

The solutions to the sub-optimization problem i n E q . (3.25) are Lagrange multipliers

/ / , , / / , which satisfy the requirement /u +/u +/J +hfj.
0

= 1 . Then, the search

direction h is computed using E q . (3.29).

Method of Outer Approximations Algorithm

Input Data. Input the initial discretization number No, the initial design x , the
N

tolerances

= r -OA/N

<7

and sNJl = p - p , where 0 < p< 1.

k

Step 0. Set N = N .
0

Step 1. Inner approximation: F i n d a point y

j<N

in the domain Y . .

Step 2. Constraints expansion: Collect a l l appropriate points y

jN

constraint set using the corresponding constraint y/j (x , y . ).

N

<N

Step 3. Outer approximation: Find an approximate new design x^+i.

37

and update the

Step 4. Replace N with N + 1 and go to Step 1.

Step 1 employs the Polak-He algorithm to solve the following inequality constrained
optimization problem:
YJ,N =argmin {^,(x,y) }
The solution to Eq. (3.30) is a point y

(3.30)

in the domain Y . . This step is an inner

JN

optimization problem and is terminated when a user-defined tolerance cr is satisfied.

N

In Step 2, all points y , (j = \,,, k = l,---,N) from Step 1 are collected. If the
jk

requirement max{o, (j)j(x ,y )\>

k

</>j(x ,y )
k

jk

JJl

is satisfied, the corresponding

Ntk

constraint

is added to the constraint set of the semi-infinite optimization problem.

Step 3 computes an approximate new solution xyv+i of the following inequality

constrained optimization problem using the Polak-He algorithm:
=argmin{ y ,* i( ) I ^ . ( ) ^ }

N+t

(- )

31

which satisfies
0* (x )>-r

(J.32)

^ .(x )<r

(3.33)

+1

with 0 (-) and

N+i

^+,(0

+1

+ l

defined in Eqs. (3.25) and (3.23), respectively.

As a nonlinear optimization problem, the semi-infinite optimization problem requires

an infinite number of iterations (steps 1 to 4) to converge to an optimal solution. In
general, the MOO A algorithm is terminated when a user-defined precision

(e.g., 10")
6

is reached. In this thesis, when two adjacent solutions are very close (||x - x . | | < s), the
/+l

MOOA algorithm is stopped, resulting in an approximate optimal solution.

After an iteration of MOOA, Step 3findsa solution x

N+l

of the approximate inequality

constrained problem. As the discretization number /V increases, the MOOA algorithm

38

results in a gradually more accurate solution. The optimal solution x* is reached when N
equals infinity. A t the design point x * , the first-order optimality conditions in E q . (3.21)
are satisfied; namely, at least one constraint is active at that point, and the value of the
objective function reaches its minimum.

39

Chapter 4

The OpenSees Software

The OpenSees software framework (McKenna et al., 2004) serves as the computational
platform for research within the Pacific Earthquake Engineering Research (PEER) Center
and is rapidly gaining users internationally. Its source code, documentation, and
executable files are freely available on the web site

http://opensees.berkeley.edu.

OpenSees was originally designed to compute the response of nonlinear structural and
geotechnical systems using finite element techniques. Haukaas and Der Kiureghian
(2004) extended OpenSees with reliability and response sensitivity capabilities for
nonlinear finite element analysis. This chapter introduces nonlinear finite element
analyses and reliability analyses in OpenSees. The response sensitivity analysis is
discussed separately in Chapter 5.
The object-oriented programming approach was employed in the development of
OpenSees. The introduction of object-oriented programming has brought with it a
revolution in software development (Deitel & Deitel, 1998). This revolution is based on
the notion of standardized, interchangeable software components. These components are
called objects or, abstractly, classes. Objects are instantiated at run-time based on
specifications made by the developer in the corresponding classes. Each class, and hence
object, may contain member functions and member data. Detailed specification of the
member functions and data members is found in class interfaces. Class interfaces contain
key information necessary to understand an object-oriented software framework. Class
interfaces also facilitate the transparent nature of object-oriented programming. Their
structure is common to all object-oriented software. Armed with the knowledge of
universal syntax rules of the programming language such as C++, a user is able to
understand the software architecture of a specific object-oriented framework. Such
software design has extensibility and maintainability as its integral feature. The
programming language C++ is employed in this thesis for the purpose of object-oriented
programming.
40

4.1

Nonlinear Finite Element Analysis

This section introduces the OpenSees software framework for the nonlinear finite
element analysis as detailed in the

OpenSees Command Language Manual

(Mazzoni et

al., 2005). Element, section, and material objects used in this thesis are emphasized, as is
the fundamental knowledge needed for the case studies in Chapter 7.
OpenSees consists of a set of modules, which create a finite element model, specify an
analysis procedure, analyze the model, and output the results. A complete finite element
analysis involves four main types of objects, as shown in Figure 4.1. The

ModelBuilder

object establishes the finite element model by defining the nodes, elements, loads, and
constraints. The

Analysis

object performs simple static linear analyses or transient

nonlinear analyses. The structural response, such as the displacement history at a node or
the entire state of the model at each load step, is recorded and output by the
object. The Domain object stores the model created by the

ModelBuilder

Recorder

object and

provides model information for the Analysis and Recorder objects.

Holds the state ofthe model
at t i m e t a n d (J + dt)

ModelBuilder>

Domain

<(

Analysis
M o v e s the model from

Constructs the objects in

the model and a d d s

state at lime t to slate at

them to the domain.

time t + dt

Recorder
Monitors u s e r d e i n e d
parameters in the model
during the analysis

Figure 4.1 Principal OpenSees objects (Mazzoni et al., 2005)

Figure 4.2 shows the relationship among elements, sections, and materials. In general,
a structure is comprised of many elements. Each element is divided into several sections,
while each section consists of several materials, such as steel and concrete materials.
Elements, sections, and materials have their own properties, which are described in the
following paragraphs. Hence, the structures created by them can exhibit different types of
behaviour, such as linear elastic and nonlinear behaviour.
41

Figure 4.2 Element, section and material relationship (Mazzoni et al. 2005)
One type of element objects is the elastic beam-column element,

elasticBeamColumn,

which is described by the following main parameters: cross-section area A , Young's

Modulus E, and the second moment of area I. Moreover, there are basically two types of
nonlinear beam-column elements: the displacement-based element,
and force-based elements, including
beamWithHinges

beamWithHinges

and

dispBeamColumn,

nonlinearBeamColumn.

The

element follows the flexibility formulation and contains an elastic

interior and two plastic hinges at the each ends. The parameters to construct this element
are pre-defined sections at two ends, ratios of the hinge length to the total element length,
cross-sectional area A , Young's Modulus E, and the second moment of area I. The
parameters A , E, and / are used for the elastic interior, which has linear-elastic properties.
Two plastic hinges represent the inelastic regions, in which forces and deformations are
sampled at the hinge midpoints using mid-point integration. A

nonlinearBeamColumn

element spreads the distributed plasticity along the element and follows the force
formulation. A dispBeamColumn element is a displacement-based beam-column element,
which has distributed plasticity with linear curvature distribution. To describe these two
elements, pre-defined sections and a number of integration points along the element are

42

required. The integration along the element is based on the Gauss-Lobatto quadrature
rule, with two integration points at the element's ends.
A section object defines the stress resultant force-deformation response at a cross
section of a beam-column or of a plate element. There are three types of sections: elastic
section, defined by material and geometric constants; resultant section, which is a general
nonlinear description of force-deformation response (e.g. moment-curvature); and fibre
section, which is discretized into smaller regions for which the material stress-strain
response is integrated to produce resultant behaviour (e.g. reinforced concrete) (Mazzoni
et al., 2005). A fibre section has a general geometric configuration formed by sub-regions
of simpler, regular shapes (e.g. quadrilateral, circular, and triangular regions) called
patches. In addition, layers of reinforcement bars can be specified. The subcommands
patch and layer are used to define the discretization of the section into fibres. Individual
fibres, however, can also be defined using the fibre object. During generation, the fibre
objects are associated with material objects, which enforce Bernoulli beam assumptions
(Mazzoni et al., 2005). Two examples of fibre sections are shown in Figure 4.3 to
describe a circular section and a rectangular section.

Figure 4.3 Fibre-section examples (Mazzoni et al., 2005)

43

There are several material objects in OpenSees. This thesis uses the

uniaxialMaterial

object, which represents uniaxial stress-strain or force-deformation relationships

(Mazzoni et al., 2005). As a uniaxialMaterial object, the Elastic-Perfectly

Plastic

(EPP)

material constructs an elastic-perfectly plastic uniaxial material object. Several

parameters used in this object are tangent E, yield stress or force in tension FyP, and
yield stress or force in compression FyN. By setting FyP

= 0, the Elastic-Perfectly

Plastic

material can be used to simulate concrete material, as shown in Figure 4.4. The steel
material,

SteelOl,

is used to construct a uniaxial bilinear steel material object with

kinematic hardening and optional isotropic hardening (Mazzoni et al., 2005). This
material needs the following parameters: yield strength F , initial elastic tangent E, and
y

hardening ratio a . This material object is also illustrated in Figure 4.4.

stress or
force

stress or
force

FyV-

FyP

strain or
deform.

strain or
deform.
FyN

EPP

SteelOl

A-Fy

Figure 4.4 Elastic-perfectly plastic material (FyP =0) and SteelOl material
4.2

Reliability Analysis

This section focuses on the implementation of the reliability analysis based on the work
of Haukaas and Der Kiureghian (2004). A ReliabilityDomain
Domain

object, while a ReliabilityAnalysis

object was added to the

object was included in the Analysis object of

44

OpenSees. An overview of some of the objects/classes in the reliability analysis module

is shown in Figure 4.5.
ReliabilityAnalysis

ReliabilityDomain

randomVariable

probabilityTransformation

runFORMAnalysis

correlation

gFunEvaluator

randomVariablePositioner

gradGEvaluator

parameterPositioner
i

performanceFunction
I
modulatingFunction

searchDirection

filter

meritFunctionCheck

spectrum
T

reliabilityConvergencCheck

runSORMAnalysis
i
runSamplingAnalysis
i
runOutCrossingAnalysi
i
runFOSMAnalysis
i
runSystemAnalysis
i
runGFunVizAnalysis
i
runFragilityAnalysis

stepSizeRule
rootFinding

startPoint
1
findDesignPoint
i

randomNumberGenerator

ii
i
i

\
\
"Analysis Types'

findCurvatures
1

"Analysis Tools"
Figure 4.5 Software framework for reliability analysis in OpenSees
(Triangle symbol denotes the relationship between base class and
subclasses, while the diamond symbol denotes analysis tools)
Three categories of classes are present in Figure 4.5. First, the ReliabilityDomain
contains model data. A randomVariable object creates random variables in several ways,
by given the random variable type, the mean, the standard deviation, etc. A correlation
object

specifies

the

correlation

coefficient
45

between

random

variables. A

randomVariablePositioner

object and a parameterPositioner

object map random

variables and parameters into a finite element model. A performanceFunction object

defines the limit-state function, also called the performance function, using an expression.
This expression may include random variables, design variables, and the structural
response from the finite element analysis. To create stochastic ground motion input in
OpenSees, the discretized random processes is used as a time series object. A
modelatingFunction object, a filter object, and a spectrum object are used to simulate this

random process.
The next category in Figure 4.5 is the "analysis tools," used for reliability analysis in
OpenSees. This framework of analysis tools makes use of the concept of object-oriented
programming, which allows the "base classes" to promise feature that is implemented by
the "sub-classes." In this manner, several implementations are made available for each of
the analysis tools. For instance, a number of sub-classes are available to evaluate the
limit-state function, and OpenSees only executes the sub-class that is specified by the
user. This illustrates the extensibility feature of OpenSees: new algorithms to perform
various analysis tasks are implemented without modifying the software framework.
A probabilityTransformation object transfers random variables between the original
space and the standard normal space. The Nataf model is applied in the current
implementation. A gFunEvaluator object computes the value of limit-state functions for a
given realization of the random variables. A gradGEvaluator object evaluates the
gradients of limit-state functions with respect to the random variables. Currently, two
alternatives are available: the finite difference method and the direct differentiation
method. Both are described in Chapter 5. A serachDirection object computes the search
direction vector when finding the most probable point (MPP) in the algorithm. Current
implementations include the iHLRF algorithm, the Polak-He algorithm, the sequential
quadratic programming algorithm, and the gradient projection algorithm. A stepSizeRule
object obtains an appropriate step size along a search direction using the line search
scheme. A simple algorithm uses the fixed step size throughout the search, and the
alternative algorithm employs the Armijo line search algorithm. A rootFinding object is
46

used by the gradient projection search algorithm to visualize the limit-state surface. A
meritFunctionCheck object checks the value of merit function and determines the
suitability of a step size. A reliabilityConvergenceCheck object checks the convergence
when searching for the MPP. One criterion determines the closeness of the MPP to the
limit-state surface; another criterion determines how closely the gradient vector points
towards the origin in the standard normal space. A startPoint object provides the starting
point when searching for the MPP; it can also serve as the centre of the sampling density
in an importance sampling analysis. Usually, the analysis starts from the mean of the
random variables, the origin of the standard normal space, or user-defined values. A
findDesignPoint object searches for the MPP using a step-by-step search scheme. The
search direction is determined by the serachDirection object, and the trial point is
determined by computing the step size along this search direction using a line search
scheme. A randomNumberGenerator object is used in the sampling analysis. The
standard library function in the programming language C++ is used in this object. A
findCurvature object is required in the second-order reliability analysis. It finds the
curvatures of the limit-state surface at the MPP.
The third category in Figure 4.5 shows eight analysis types in the reliability module of
OpenSees. The users are required to specify some necessary analysis tools before
executing these analysis commands. OpenSees prints corresponding information to a file
or a computer monitor, thereby allowing the users to monitor the reliability analysis
process. Following a successful analysis, OpenSees outputs the results into a userspecified file. In this thesis, thefirst-orderreliability analysis (runFORMAnalysis) and
the importance sampling analysis (runSamplingAnalysis) are employed in case studies in
Chapter 7.

47

Chapter 5

Response Sensitivity for Nonlinear Structures

Response sensitivity is essential in reliability-based design optimization (RBDO) when a

gradient-based approach is employed. Reliability analysis requires the gradient of the
limit-state function with respect to the random variables when searching for the most
probable point. The optimization analysis requires the gradient of the objective function
with respect to the design variables to satisfy thefirst-orderoptimality conditions. When
it comes to the optimization of real-world structures, where thefiniteelement method is
employed, the RBDO algorithms require finite element response sensitivities to be
available.
Two response sensitivity methods are described in this chapter: the finite difference
method (FDM) and the direct differentiation method (DDM). The derivations below
follow Haukaas and Der Kiureghian (2004, 2005). The negative effects of the response
gradient discontinuity are remedied using two methods, smooth material models and
section discretization, which make possible the RBDO for nonlinear structures.

5.1

Finite Difference Method

The FDM consists of perturbing the values of model parameters, re-evaluating structural
responses, and finally obtaining a finite difference estimate of the gradient vector. A
typical equation of the FDM is
dF(G) _ F(0 + Ad) - F{6 - AO)
86

2A0

where 6 is the model parameter representing a material, geometry, or load parameter,

F(9) is the response function evaluated by the finite element analysis, and AO is the
perturbation. This form of the FDM requires two additional executions of the response
function for each parameter.
48

A simplified FDM is the forward finite difference method, which is similar to Eq.
(5.1), but requires only one additional execution of the response function for each
parameter. It has the form:
dF(O)F(0

80

&e)-F(e)
A0

Eqs. (5.1) and (5.2) indicate that the FDM is not efficient because of additional
executions of the response function for each derivative. In addition, the accuracy of the
gradient depends on the selection of the perturbation factor of each parameter, which is a
challenging task.

5.2

Direct Differentiation Method

The DDM consists of analytically differentiating the response equations and

implementing them inside the finite element code. Therefore, the DDM computes
response sensitivities alongside the response estimate without additional response
computations. The DDM differentiates the equilibrium with respect to a genetic model
parameter 0. For inelastic static analyses, the equilibrium equation requires internal
forces

P"'

to equal external loads

P '
EX

at a pseudo-time

t.
n

(5-3)

= p r

where F' "' is a function of the displacement vector u, and the subscript denotes a
n

quantity at time t . By differentiating Eq. (5.3) with respect to 9 and rearranging, we

n

obtain:
T

3u
" 80

ap;*' a p ; ,
(

80

80

'"

f,xed

where K = 8P' "' I8\x is the algorithmically consistent stiffness matrix, and
n

the desired displacement sensitivity for reliability and optimization analyses.

49

8u 180
n

8u /80
n

is

is

solved after every convergence of the equilibrium Eq. (5.3). The computation of du 180
n

is efficient since Eq. (5.4) is linear with respect to 8xx l80.

n

The derivative of internal force for the fixed current displacement appears in Eq. (5.4)
and is assembled over all elements based on the strain-displacement matrix B, the stress
vector a, and the element stiffness matrix k.
dPf |
QQ

_ M f

~ J\
KJ

lufixed

lu fixed

R
+

a^,

r .
D

QQ

r . j^.

lu fixed "

r r >

QQ

n n
K

~
\e

fixed J

UX

W - - V

where U denotes assembly over all elements, and Q. denotes the domain of each
el

element. The differentiation of the element integration is also required in the assembly.
When parameter 0 represents material properties, the derivative of the internal force is
assembled from derivatives of stress at each material point for the fixed current strain.
Then, Eq. (5.5) can be simplified as follows:
^ i ^ = u k : ^ A - d *

(5.6)

Two important issues are considered in the implementation of the DDM. First, the
response sensitivity 8u /80 must be solved at each load increment for inelastic
n

materials since sensitivity computations require history variables and their derivatives to
be computed and stored at each increment. In each sensitivity computation, the material
routine is called twice because the derivative of the stress 8a/80\

flxed

in Eq. (5.6) is

conditioned upon fixed strain at the current increment only. The first call obtains the
derivative of the stress conditioned upon fixed stain, and the second call computes and
stores unconditional derivative of the history variables when displacement and strain
sensitivities are obtained.
Second, 8a I80 \

E fixed

8V /80\
nl

flxed

must be assembled over all inelastic material points to compute

since all components of the displacement sensitivity 8u/d0 are

generally not zero. The strain sensitivity 8s 180 in the finite element domain is not zero,
and thus da I'80\

z fixed

is also not zero at all material points after the first load step
50

because deldO enters the derivatives of the history variables. For further details, see
Haukaas and Der Kiureghian (2004).

5.3

Object-Oriented Implementation in OpenSees

The FDM and the DDM are implemented in OpenSees based on the object-oriented
software architecture. Figure 5.1 shows the framework of response sensitivity in
OpenSees. The

finiteDifferenceGradGEvaluator

utilizes the

GfunEvaluator

object to

compute the value of limit-state functions for perturbed parameters and then to calculate
the response gradient using the FDM. The OpenSeesGradGEvaluator

makes use of the

DDM implementation in OpenSees. Two new classes are applied to sensitivity

computations: the
algorithm

sensitivity

algorithm

and the

sensitivity

integrator.

The

sensitivity

computes response sensitivities with respect to all desired parameters.

Currently, two options are available for the sensitivity

algorithm:

the

computeAtEachStep

gradGEvaluator

finiteDifferenceGradGEvaluator

OpenSeesGradGEvaluator

GFunEvaluator
Sensitivity Integratoi

dynamic static

Sensitivity Algorithm

:omputeAtEachStep computeByCommand

Figure 5.1 The framework of response sensitivity in OpenSees

(Triangle symbol denotes the relationship between base class and
subclasses, while the diamond symbol denotes analysis tools)
51

option is used in all dynamic analyses and all inelastic analyses, while the
computeByCommand

option is used in elastic static analyses. The sensitivity

integrator

assembles the right-hand side of Eq. (5.4) for each parameter for either static analysis or
dynamic

analysis.

The DDM computes response sensitivities at each equilibrium state of an inelastic

static analysis. After the finite element analysis converges at a pseudo-time t, the
sensitivity algorithm

matrix. The

is used to compute response sensitivities and update the tangent

sensitivity

algorithm

iterates over all parameters for which response

sensitivities are desired, and performs the following operations: First, parameter 9 in the
finite element domain is "activated" to obtain correct contributions from element, section,
material, node, and load objects. Second, the

sensitivity integrator

assembles the right-

hand side of Eq. (5.4) and collects contributions from the objects of the finite element
domain. Next, Eq. (5.4) is solved to obtain the displacement sensitivity du/89, and the
results are stored in the node objects. Finally, all material objects are called by the
sensitivity

integrator

to store unconditional derivatives of history variables through

element and section objects. Stain sensitivities are computed based on the displacement
sensitivity by using ordinary kinematic equations.
As part of the finite element reliability analysis, the reliability module maps random
variables and design variables to the finite element module and receives structural
responses and response sensitivities from the finite element module. The mapping
procedure updates the value of model parameters in the finite element model each time
new random variables and design variables are available. It also identifies parameters to
ensure that correct contributions to response sensitivity computations are assembled.
Three member functions are used in the classes containing desired model parameters: a
member function identifying the parameters, a member function updating the value of the
parameters, and a member function

activating

the parameters for response sensitivity

computations.
Two objects are involved in the mapping procedure: the
object (as part of the reliability analysis) and an object
52

(e.g.,

randomVariablePositioner

a material object) in the

finite element domain. The randomVariablePositioner object makes use o f the object
from the finite element domain as its data member. The detail mapping procedure is
described as follows. First, the setParameter method in the finite element object creates a
link between relevant random variables and the parameter in the finite element object.
Then, when the reliability analysis updates the random variables i n the finite element
model,

the update

method

of

the

randomVariablePositioner

object

and the

updateParameter method o f the finite element object are called upon update the values o f
the model parameters using new random variables from the previously created link.

5.4

Continuity of Response Gradients

In the nonlinear finite element analysis it is common to employ material models with
sudden transitions from elastic to plastic response. A s discussed' in Haukaas and Der
Kiureghian (2004), this may lead to discontinuities in response sensitivities. This also has
an adverse effect on the convergence to the most probable point in reliability analysis. In
this thesis we emphasize the potential negative effect o f the optimization analysis. In fact,
the effect o f gradient discontinuities due to sudden transitions from elastic to plastic
response is dramatically detrimental to the performance o f the optimization algorithm.
This is because (1) the proof o f convergence o f the optimization algorithms requires
continuously differentiable limit-state functions; (2) the discontinuities in the gradient
may cause the algorithm to stall; and (3) the abrupt changes in the gradient may cause i l l conditioning (even though it is theoretically acceptable) and hence slow convergence to a
solution. This leads to the conclusion that the issue o f gradient discontinuities is even
more important in the R B D O analysis than in the search for the most probable point in
the stand-alone reliability analysis. It is stressed that the nonconvergence or slow
convergence problems are expected since the assumption o f continuous differentiability
is violated. In fact, all standard nonlinear programming algorithms w i l l experience
difficulties when applied to such inappropriate problems.

53

In the following subsections, two remedies are introduced to solve discontinuity

problems in response sensitivities.

5.4.1

Smooth Material Model

To avoid gradient discontinuities at yielding, the original bi-linear material model is

replaced with a smoothed version developed by Haukaas and Der Kiureghian (2004). The
bi-linear material model contains an elastic range and a plastic range. Unloading is
assumed to be elastic. The tangent stiffness is E in the elastic range and aE in the plastic
range, where 0 < a < 1. The yield strength Fy identifies at the transition between elastic
and plastic states. In the smooth material model, a circular segment is used to smooth the
transition between the elastic and plastic response state. The tangent stiffness of the
stress/force A

original bi - linear model

smoothed material model

Figure 5.2 Bi-linear steel material model smoothed with circular segment

54

circular segment coincides with those of the elastic and plastic ranges at intersection
points. The smoothed stress-stain curve and its tangent stiffness are thus continuous. The
circular segment starts from y-Fy in the stress-strain curve, where the parameter is
0 < y < 1. Figure 5.2 illustrates how the bi-linear steel material model is smoothed with a
circular segment. Haukaas and Der

Kiureghian (2004) also show that the overall response

does not change significantly as a result of smoothing. They recommend the selection of
parameter y > 0.8 to avoid results that differ significantly from those obtained with the
bi-linear model.
It is demonstrated in Haukaas and Der Kiureghian (2004) that the smooth material
model leads to continuity in response sensitivities. To compute response sensitivities
using the DDM

approach, Haukaas and Der Kiureghian (2004) differentiate the

DDM

equations for the smooth material model and implement them in OpenSees. They also
present several examples to show that the smooth material model successfully avoids the
response sensitivity discontinuity problem in the reliability analysis.

5.4.2

Section Discretization

S c h e m e

The section discretization scheme discretizes the element cross-section into smaller
regions or fibres. The uniaxial material response for each fibre is integrated to produce
approximately smooth behaviour. The use of this section discretization scheme makes the
nonlinear structural response "approximately continuously differentiable" to meet the
requirement of the RBDO algorithms.
The section discretization scheme takes advantage of the fibre section object in
OpenSees. The fibre section is ideal for defining a reinforced concrete section: 1-2 top
and bottom fibres of the concrete cover using normal strength concrete, 10-20 side fibres
of the concrete cover using normal strength concrete, 10-20 fibres of the concrete core
using higher strength confined concrete, and several layers of reinforced bars. Examples
of suchfibresection are shown in Figure 4.3.
55

The section discretization scheme makes the structural response approximately

continuously differentiable. During the finite element analysis, individual fibres pass the
yield point and enter the plastic range one by one. Each fibre is only a small part of the
whole cross section. The discontinuity of one fibre affects the property of the whole
section, but the effect becomes progressively smaller as the number of fibres in the whole
section increases. Thus, theoretically, the discretized cross section possesses continuous
properties when the section is discretized by an infinite number of fibres. Practical studies
in this work suggest that approximately 20 fibres are sufficient to avoid convergence
problems. It is emphasized, however, that a smooth material model must be employed for
the reinforcing steel, since this is represented by a single fibre layer that is not affected by
an overall increase in the number of fibres in the cross-section.

56

Chapter 6

Implementation of Reliability-Based Design

Optimization

Two decoupling approaches for reliability-based design optimization (RBDO) problems

are implemented in this chapter: a decoupled sequential approach using the method of
outer approximations (DSA-MOOA) and a simplified decoupled sequential approach
(DSA-S). The two approaches employ the same problem reformulation, which is
presented below. The algorithms that are used in the DSA-MOOA approach and the
DSA-S approach are also detailed. Finally, OpenSees is extended in light of these
optimization capabilities through object-oriented programming.

6.1

Problem Reformulation

The RBDO problem in Eq. (1.2) minimizes the initial design cost plus the expected cost
of failure subject to reliability and structural constraints. Let x be the vector of design
variables. Then, this problem takes the form (see Chapter 1)
x*=argmin{ c (x) + c (x)p (x)
0

| f (x) < 0, p (x)<p

f

(6.1)

where c (x) is the initial cost of the design, c (x) is the present cost of future failure,
0

p (x) denotes the probability of failure for one failure event, f (x) are deterministic
f

constraints, and p is the upper bound on the probability of failure.

f

The solution algorithm for Eq. (6.1) requires the functions involved, c (x), c (x), and
0

f (x), to be continuous, and the constraint set f (x) to be closed and bounded. Since the
failure probability p (x) is involved in both the objective and constraint functions, the
f

failure probability is also required to be continuous. Royset et al. (2002) have proven that
this is the case in realistic design problems.

57

The problem in Eq. (6.1) is computationally difficult since the failure probability,
which depends on the design variables, is defined in terms of a high-dimensional integral
over the domain of random variables. Royset et al. (2002) replace the failure probability
p (x) with parameter a, which is updated during the optimization analysis to develop a
f

computationally feasible problem. This parameter is included in an augmented design

vector x = (x,a) and appears as an added constraint of a reformulated optimization
problem, which reads:
x = argmin|c (x) + c (x)a | f(x) < 0, p {x) = a, 0 < a < p ^ j
0

(6.2)

Royset et al. (2002) have proven that Eqs. (6.1) and (6.2) have identical global optimal
solutions when some assumptions are satisfied.
Because the gradient of the true failure probability is unavailable, it is problematic that
the failure probability still appears among the constraints. This is addressed by making
use of concepts from the first-order reliability method (FORM). As outlined in Chapter 2,
the FORM estimate of the failure probability is p/= 0(-/?), where the reliability index 6 is
the minimum distance from the origin to the limit-state surface g=0. To this end, the
equality constraint p (x) = a in Eq. (6.2) is replaced by constraint y/ - 0, where ip is
f

the negative value of the minimum of limit-state function g within a hyper-sphere of

radius - <X> (a). Hence, the problem is now reformulated into the form
1
_

x = arg min{ c (x) + c (x)a | / , (x) < 0, y/(x) = 0,0<a<p}

0

(6.3)

where
^(x) = - min {g(x, u)}
ueB(0,r)

58

(6.4)

with B ( 0 , r ) = {u | ||u||<r}, where r = - 0 '(a).

It should be stressed that this

reformulation is motivated by the desire to cast the optimization problem in a semiinfinite form, thereby allowing it to be solved using the method of outer approximations
(DSA-MOOA). This is because the constraint y/ = 0 in fact represents an infinite number
of constraints, one for each point within the hyper-sphere. In this thesis we also
demonstrate that a simplified approach (DSA-S) can be used to solve the problem by only
including one constraint to enforce y/ = 0.
If a FORM approximation of the failure probability is acceptable, then the
reformulation of the constraint p (x) = a in terms of the function y/ is correct should
f

the parameter r be equal to - O

- 1

(a). However, when the limit-state function g(x,u) is

prescribed in terms of response quantities from a nonlinear finite element analysis, then
the limit-state function is nonlinear. To account for such nonlinearity, a correction factor /
is introduced: r = - O

- 1

(a) t. The start value of the auxiliary variable t is unity, and it is

updated during the optimization analysis.

Eq. (6.3) is a semi-infinite optimization problem with equality constraints. As
described in Chapter 3, the available algorithms address problems with inequality
constraints. Hence, a more suitable problem formulation is obtained by converting the
equality constraint into an inequality constraint:
x = argmin(c (x) + c ( x ) a | f ( x ) < 0 , ^ ( x ) < 0 , 0 < a < p
0

(6.5)

Royset et al. (2002) have proven that replacing the equality constraint in Eq. (6.3) by an
inequality constraint does not alter the solution. This proof assumes that the failure cost is
positive, and that the origin in the standard normal space is in the safe domain. The

The solution algorithm requires that the solution domain of Eq. (6.3) remainfixed.This is obviously not

the case because r varies in the optimization process. In the computer implementation, this problem is
solved by applying the u = ru transformation, where u is the solution domain that remains a ball of unit
radius, namely u e B(0,1).

59

former assumption is trivially satisfied, and the latter one is generally satisfied due to the
high reliability of structures.
Eq. (6.5) is the final reformulation of the original problem. If the limit-state function is
linear or i f the F O R M approximation of the failure probability is acceptable, then t is set
at 1 and the reformulated problem in Eq. (6.5) has the same solution as the original
problem in Eq. (6.1). This is proven by Royset et al. (2002). Moreover, the method of
outer approximation (MOOA) algorithm to solve the semi-infinite problem in Eq. (6.5)
has convergence proofs (Polak, 1997). Hence, we are guaranteed to find a converged
solution for our final approximate problem in Eq. (6.5).
However, the first-order approximation of the reliability problem may be a poor
assumption when the limit-state function is highly nonlinear. That is, in nonlinear finite
element reliability problems, the parameter a from the first-order approximation does not
equal the probability of failure p (x)
f

from a more precise reliability analysis (for

example, importance sampling). To be able to deal with these cases, we solve the final
approximate problem through updating the value of the parameter t. The parameter t
starts from unity and is updated during the optimization analysis to account for the
nonlinearity in the limit-state function. In this manner, approximate solutions are
obtained with increasing accuracy as the algorithm proceeds. Specifically, the parameter /
is updated by multiplying with the correction factor 0~ (a)/0~ (p (x)).
l

The philosophy

behind this update is that i f p (x)>a, then the constraint y/ < 0 in the final approximate
f

problem allows the limit-state surface {u | g(x,u) = 0} to come too close to the origin in
the u-space, thus requiring the radius of the ball associated with y/ to be increased. The
increase of the ball radius is obtained by increasing t. If p^ (x)<a, then the limit-state
surface is required to be too far away from the origin in the u-space by the constraint
y/ < 0, and the size of the ball must be reduced (i.e., t is reduced).

60

6.2

DSA-MOOA Approach

In this thesis we employ solution algorithms that are termed "decoupled" and
"sequential." The justification for the decoupled characterization is that any reliability
method can be used to obtain a more precise estimate of the failure probability than the
FORM analysis. The justification for the sequential characterization is that the
optimization analysis in Eq. (6.5) and the reliability analysis in Eq. (6.4) are solved
repeatedly and in sequence to address the bi-level problem in Eq. (6.1). The reliability
constraint is updated for each optimization analysis in Eq. (6.5). It is stressed that the
decoupled approach allows flexibility in the choice of optimization algorithm and
reliability computation method.
In this section we present the implementation of the DSA-MOOA approach developed
by Kirjner-Neto et al. (1998) and Royset et al. (2002). It makes use of the problem
formulation in Eq. (6.5). Figure 6.1 presents a flow chart of the DSA-MOOA algorithm,
which consists of iterations at several levels. Upon the initialization of design x and
parameters a and t, the top level iteration includes three tasks:

Al. Update design vector x and auxiliary variable a.

A2. Compute failure probability p (x) using a method that is more accurate than
f

FORM.

A3. Update parameter t.

This procedure is repeated until the design vector x, the auxiliary design variable a, and
the parameter t stabilize. Usually, 5 to 15 iterations (Al to A3) are needed to reach a
satisfactory solution.
Task A l consists of obtaining updated values of x and a by solving the semi-infinite
optimization problem in Eq. (6.5). The user can choose any suitable algorithm for this
purpose. The MOOA algorithm is employed in the current OpenSees implementation. In
this algorithm, the ball B(0,1) is discretized into a finite number of points, ui, u ,
2

in the following manner: approximate the constraint

61

y/ < 0

UN,

by N constraints

g(x,-3> \a)-t -u )>0, j-\,2,...,N

and solve the resulting standard nonlinear

inequality constrained problem. In task A l , the MOOA algorithm consists of three

iteratively performed tasks (B1-B3):

o'

Pf

p, \,

i = 0, j = 0 t =l
0

a =p ,
0

x =(x ,a )

Al - Update design x and variable a

BI - Inner approximation
<0]

min{g(x,u)|||u|| -[--\a)tf
2

^ (x) = -g(x,-<D-'(a)fu )
y

>

B2 - Constraints expansion
if y/j(x)> p

-p

1 VOO

B3 - Outer approximation
min j? (x) + c (x)alf(x) < 0,y/(\,a) < 0,0 < a < p \
(x,a)

7=7 + 1

Yes
X/+i

i+i

A2 - Compute failure probability p

ta ,p(x )
M

A3 - Update parameter /
-1,))
t =tp~\a )/0- (pAx
l

Optimal design
Figure 6.1 Flow chart of DSA-MOOA approach

62

B l . Inner approximation: Obtain the minimum value of the limit-state function

within the ball B(0, - <J>~' (a) t) and the corresponding random vector u using the Polak7

He algorithm. In this thesis, the negative value of the limit-state function at the vector u j
is denoted as y/. Terminate the Polak-He algorithm when tolerance - a

Here, o

= 0.1/ N

is a user-defined series, a

is satisfied.

> 0 as N -> oo.

B2. Constraints expansion: Update the constraints by accumulatively storing

solutions Uj from task Bl for which solution y/j exceeds p - p ,
j

where p is a user-

defined parameter usually set at 0.5. In this manner, the number of constraints
represented by y/ < 0 evolves during these iterations. We have observed that these
constraints are simply a collection of points Uj for which the limit-state function is
required to stay positive in task B3.
B3. Outer approximation: Solve the constrained optimization problem in Eq. (6.5)
using the Polak-He algorithm. The number of constraints in this problem is equal to the
number of structural constraints, plus the N constraints added by the previous item and
the single constraint a<p .
f

The result is a new augmented design vector x = (x,a).

According to proofs presented by Polak (1997) for the MOO A algorithm, an "exact"
solution is found if the discretization number N approaches infinity.

Tasks B l , B2, and B3 are repeated until the optimality conditions are satisfied
according to a user-defined precision tolerance. Typically, 75 to 150 iterations are
required to find the optimal solution. These tasks are described in further detail in
subsections 6.2.1-6.2.3, in which we focus on the connections between the particular
problems discussed in this chapter and the general algorithms discussed in Chapter 3.
One important advantage of DSA-MOOA approach is reiterated, namely that the
reliability and optimization calculations are decoupled, thus allowing flexibility in the
choice of optimization algorithm in task A l and reliability computation method in task
A2. In addition to the MOO A algorithm, Polak (1997) provides a pre-defined
discretization scheme to solve the semi-infinite problem in task A l . Similarly, the user is

63

free to estimate the failure probability in task A 2 by selecting a suitable computational

reliability method such as the second-order reliability method, Monte Carlo sampling, or
importance sampling. This selection depends on the desired accuracy in satisfying the
probability constraint p (x) < p .
f

F o r example, an importance-sampling scheme with a

required 1-2% coefficient o f variation of the sampling result is an appropriate choice i f

the user wants a high degree o f confidence that the reliability constraint is satisfied.
Typically, users require the structural reliability to be very high. In effect, the failure
probability p (x) is very small. The D S A - M O O A approach may experience numerical
f

difficulties caused by the potential difference in orders o f magnitude between a and the
other design variables x. F o r this reason, in our implementation the parameter
6 = -0~'(a)

is used in place o f a. With reference to E q . (2.9), we note that b is a

substitute for the reliability index B, in the same way that a is a substitute for the failure
probability p (\).

In conclusion, the optimization in D S A - M O O A approach is over the

design vector (x, b).

6.2.1

B l - Inner Approximation

Given design x and parameters a and t, task B l solves the following reliability problem:

Uj - u* = argmin|g(x,u) | ||u|| - r < 0 j

2

(6.6)

where r = -<b~\a)-t. This problem is equivalent to E q . (6.4). The result is a random

vector

u*,

or

u,

(for /

discretization

point).

The corresponding

constraint

y/j - -g(x, rUj) is the minimum of the limit-state function within a ball o f radius r.
The Polak-He algorithm described in Chapter 3 is used to solve E q . (6.6), an inequality
constraint optimization problem with a single constraint. The fact that there is only a
single constraint simplifies the optimization process. The values and gradients of
functions Fand/in

E q . (3.1) are

64

F(x, u) = g(x, u),

VF(x, u) = Vg(x, u)

/(x,u) = / (x,u) = ||u|| - r ,

2

(6.7)

V/,(x,u) = 2u

(6.8)

The first step searches for the direction vector h, by solving a quadratic optimization
problem #,.(x) in Eq. (3.14), subject to linear constraint

=1. By setting

and substituting Eqs. (6.7) and (6.8) into Eq. (3.14), this sub-optimization
problem can be simplified as
0,(x) = - mm{]f (x, u,) -

\]f (x, u,.) - f(x, u,) + + /, (x, u,)]

(6.9)
^[Vg(x,u,.) + /i [2u,.-Vg(x,u,.)]|

Eq. (6.9) can be solved by V0 /V//, = 0, and has the following solution:
!

b/-(x,u )-/(x,u ) +/;(x,u,.)]-lvg(x,u )[2u . - Vg(x,u,)]

i

<

rf=

"

-[2u,-Vg(x,u,)]

( 6

1 Q

0,

Mi

=1

Mi

Figure 6.2 n solutions for inner approximation using the Polak-He algorithm
x

whenever the right-hand side of Eq. (6.10) has a value in [0,1]. Otherwise, the solution of
Eq. (6.9) is either //* =0 or p.* = 1, whichever yields the lowest value for the objective

65

function i n E q . (6.9). Three types o f solutions are illustrated i n Figure 6.2. Finally, the
search direction is
h,. = -i[a-A/1')-Vg(x,ul) + 2 ^ u ]

(6.11)

The second step finds an appropriate step size X along the search direction h . using
t

the Armijo rule. Then, the random vector u,

+1

is replaced by u

/ + 1

= u, + A,h,. and is used

as the input data for the next iteration.

For fask B 1 , the Polak-He algorithm requires an infinite number o f iterations before it
converges to the optimal solution u*. However, we terminate the Polak-He algorithm
when user-defined tolerance -<J =-0.l/N

is satisfied. N starts from 1 and goes to

infinity, so a

starts from the larger tolerance 0.1 and goes to 0 as Af increases. Since

high accuracy is only needed when approaching the design point, a high tolerance in the
beginning o f task B1 is acceptable and can save computational time.

6.2.2

B2 - Constraints Expansion

Task B 2 first collects the u and y/j(x) from the inner approximation (task B l ) into a
y

matrix. For N discretization points, we have an JV-column matrix, i n which each column
contains random variables u and y/j (x) o f the form
y

u,

u,

_^i(x) ifj(x)

(6.12)

y/ (x)_
N

Second, task B 2 assembles the reliability constraints set y/(x), which includes the
constraint y/ (x) at the origin ( u = 0 ) in the whole assembling procedure. The reason to
0

include the constraint

at the origin

is to make sure

^ ( x ) = - g ( x , 0) < 0 , i.e.,
0

g(x, 0) > 0 , which is the requirement o f problem reformulation in subsection 6.1. Task
B 2 updates reliability constraints by accumulatively storing solutions u

66

when y/j(x)

exceeds p - p .
j

p e (0, 1) is a user-defined parameter set at 0.5 in the current

implementation. p - p
J

is a positive series and approaches zero when j approaches TV".

Finally, the reliability constraint set y/(x) has the following form:

y/(x) =

u =0

u,

^ (x)<0
0

y/ (x)>p -p
i

x / {x)>p -p"
J

iy (x)>0

(6.13)

In this manner, the number of constraints represented by ^(x) evolves with the
increase in the number of discretization points. These constraints are simply a collection

failure domain
g<0,

safe domain \
g>0,

V j

<0

X
X

Figure 6.3 Reliability constraints set y(x)

67

Vj>0

of points u for which the limit-state function is negative in task B2 but is required to
y

stay positive in task B3. Hence, the result of task B2 is an expanded constraints set ip(x),
which enters task B3 for outer approximation. From the above description, we know that
the number of columns of matrix ^(x) is less than or equal to N+l. Figure 6.3 illustrates
the procedure used to assemble reliability constraints set y(x) by collecting all of the
points in the failure domain in the standard normal space.

6.2.3

B3 - Outer Approximation

Given the objective function c (x) + c (x)a, m deterministic constraints f(x)<0, n

0

reliability constraints y/(x) < 0, and an extra constraint a< p , task B3 solves Eq. (6.5)
f

and updates design x and parameter a. This task is also solved using the Polak-He
algorithm. As opposed to task Bl, the number of constraints in task B3 is greater than
one. In fact, the number of constraints increases during the analysis. According to Polak
(1997), in the Polak-He algorithm a quadratic sub-optimization problem with linear
constraints must be solved to obtain the search direction h . This is currently addressed
(

by linking the quadratic programming software LSSOL (Gill et al., 1986) to OpenSees.
For afixednumber (q=m+n+\) of constraints, Eq. (6.5) is an inequality-constrained
problem. If the Polak-He algorithm is applied to task B3, functions F and / i n Eq. (3.1)
have the following form:
Obj etive function

F(x) = c (x) + c (x)a

0

Deterministic constraints / , (x) = / (x), , f (x) = f (x)

_

Reliability constraints

Extra constraint

m+l

(x) = \p (x), f

m+n+x

m+2

(x) = xp (x), , f
x

m+n

(6.14)

(x) = y _ (x)
n

(x) = a - p

Thefirststep searches for the direction vector h. by solving a quadratic optimization

(

problem

68

0, = -mm{Morm)+

+ Mj[f(*X -f(x,0]+^|KvF(x,.) + / i V f ( x , . ) | | (6.15)

2

where / ( x ) = max / (x), / ( x ) .= max {0,/(x)}. Note that 0, is a quadratic problem

+

7=IU

dependent

q]

on Lagrange multipliers (p ,p ,---,p )

0

subject to linear constraints

p + p + + p =1. Eq. (6.15) cannot be solved in the same way as Eq. (6.9) in the
0

inner approximation, so it is instead solved using the quadratic programming software

LSSOL (Gill et al., 1986) or the Matlab O P T M toolbox (Matlab, 1999). LSSOL is used
in the current implementation of OpenSees since it is a collection of F O R T R N 77
subroutines and is faster than the equivalent "Quadprog.m" in the Matlab code. In this
thesis, the constrained least-squares problem "LS2 Type Objective Function" is used and
stated in the following form:
1 II

minimize F(u) = g \i + b - G u |
(6.16)

subject to

- |

f -

where \i = [p , p , ,p ] , general constraints C = [l, 1, ,l] , vectors L and U are

T

?+l

the lower and upper bounds for all the variables and general constraints, respectively:
L = [0, 0, -,l]

r
+1

and U = [l, 1, ,1]^ . The constraints in Eq. (6.16) are equivalent to

|

the constraints in Eq. (6.15):

0< <l,

j = 0,\,-,q

Mj

(6.17)
Mo +M + + M
l

We refer to G as the least-square matrix and to vector b as the vector of observations,

here b = 0. The objective function in Eq. (6.16) is equivalent to the objective function in
Eq. (6.15), given the following vector g and matrix G :
r

69

/(!, ) -/ (x,.)

g'

G =

dx,

ax,.

5x,.

da,

5a,

(6.18)

dFCxJ
da,

LSSOL can solve Eq. (6.15) in a finite number of iterations and find the solution to
u* =[/u* , //,
0

,//* ]

Finally, the search direction is

h,. = - i L ; V F ( x , ) X^V/ .(x,.)j

+

(6.19)

7=1

The second stepfindsan appropriate step size X along the search direction h,. using
t

the Armijo rule. Then, a new design is found by x,. = x. + X.h. and used as input data
+1

for the next iteration.

Like the inner approximation, the Polak-He algorithm may require an infinite number
of iterations before it converges to the optimal solution x* for Eq. (6.5). In our analysis,
we terminate the Polak-He algorithm when
-cr <0 (x )<O
N

N+l

N+l

0<^ (x )<o-
+l

+l

(6.20)

(6.21)

with 0 (-) and W (-) defined in Eqs. (3.25) and (3.23), respectively. The definition of
+1

N+l

<J =0.\/N

is the same as for the inner approximation, which starts from larger

tolerance and goes to 0 as N increases. This is reasonable here since the accuracy of the
semi-infinite optimization algorithm gradually increases with the increase in the
discretization number and since high accuracy is only required when approaching the
design point.
In task B3 we need to evaluate three functions and their gradients with respect to the
augmented design variable vector x: the objective function, structural constraint
functions, and reliability constraint functions.

70

d[c (x) + c (x)a]

df(x)

dx

dx

dy/(x) _ dy/(x) dd
dx

dd

dx

(6.22)
d[c (x) + c (x)a]

8f(x)

da

da

dy/(x) _ dy/(x) dd
da

dd

da

where y/ is the negative value of the limit-state function g, dy/ldd is easily found
because g is a simple algebraic expression in terms of d, and dd/dx are response
gradients. Again, the existing FDM or DDM implementations in OpenSees are used to
obtain the required response gradients.
In conclusion, the MOOA algorithm solves a series of inequality-constrained problems
in tasks BI and B3. As the discretization number N increases, the MOOA algorithm
results in a gradually more accurate solution x. The optimal solution x* is reached when
/V equals infinity. At the optimal point, the value of the objective function reaches its
minimum, and the first-order optimality conditions in Eq. (3.21) are satisfied. Therefore,
one of reliability constraints is active and equal to zero at the optimal point. This means
that the reliability analysis finds the most probable failure point (MPP)

at the optimal

design point, shown in Figure 6.3.

6.3

DSA-S Approach

This section develops a DSA-S approach by combining the problem .reformulation in

Section 6.1 with the findings of Du and Chen (2002) and Agarwal and Renaud (2004).
However, the reformulated optimization problem is defined as an inequality constrained
optimization problem with a single reliability constraint as opposed to infinite constraints
in the DSA-MOOA approach. In this study, a deterministic optimization problem in Eq.
(6.5) isfirstsolved to find a new design x. Second, a reliability analysis is performed to
update the reliability constraint based on the new design x. This sequential iteration is
repeated until a consistent design is obtained.
Convergence cannot be proven mathematically for the DSA-S approach, since the
failure probability used for the deterministic optimization analysis is obtained from the
71

last iteration. However, the DSA-S approach is still attractive because a consistent design
is obtained at a considerably lower computational cost. The computational time of
discretization in the DAS-MOOA approach is saved since there is only one reliability
constraint. Another advantage of the DSA-S approach is the decoupling of the reliability
and optimization calculations. It is flexible in selecting optimization algorithms and
reliability computation methods.
X

Pf

0>

= 0, j = 0, f = l
0

^ 1 '

^2

x =(x ,a )

a =p ,
0

C I - Update design x and variable a

D l - Deterministic optimization analysis
min^ (x) + c (x)a|f(x) < 0,y/(x,a) < 0,0 < a < p j
0

x,a

D 2 -Reliability constraint update

min{g(x,u)|||u|| -[--\a)t]
2

<o}

^(x) = - g ( x , - 0 ( a ) / u )
_l

u, ^(x)

Yes
X

/+l'

i+l

C 2 - Compute failure probability p (x)

i

t a ,p {x )
n

C 3 - Update parameter t
t =t,*-\a )/<l>- (j> (x ))
l

i - i +1

No

Yes
Figure 6.4 Flow chart of DSA-S approach

72

Optimal design

Figure 6.4 presents aflowchart of the DSA-S approach, which consists of iterations at
several levels. Upon the initialization of design x and parameters a and t, the top level
iteration includes three tasks:
CI. Update design vector x and auxiliary variable a.
C2. Compute failure probability p (x)

using a method that is more accurate than

FORM.
C3. Update parameter t.

Tasks C2 and C3 are same as tasks A2 and A3 in the DSA-MOOA approach. The top
level iteration is repeated until design vector x, auxiliary design variable a, and parameter
t stabilize. Usually, 5 to 15 iterations (CI to C3) are needed to reach a consistent
reliability based design.
Task CI updates values of x and a in two steps: deterministic optimization analysis and
reliability constraint update. These are described in tasks Dl and D2, respectively. In the
current implementation, the Polak-He algorithm is employed to solve these two tasks.
D l . Deterministic optimization analysis: The inequality constrained optimization
problem in Eq. (6.5) is solved using the Polak-He algorithm. The constraints in this
problem include several structural constraints, one reliability constraint ipCx) < 0, and

an extra constraint a<p .

f

The result is a new augmented design vector x = (x,a).

During the first iteration, the random variables are set equal to their mean values, and
parameter a is set as p .
f

The Polak-He algorithm is terminated when user-defined

tolerance s is satisfied. Task Dl is similar to task B3 in the DSA-MOOA approach with

only one reliability constraint and fixed high tolerance. Task Dl also needs to solve a
quadratic sub-optimization problem with linear constraints using the quadratic
programming

software LSSOL

(Gill

et

al.,

1986) or Matlab

OPTM

toolbox

"Quadprog.m " (Matlab, 1999).

D2. Reliability constraint update: Obtain the minimum value of the limit-state
function within the ball B(0, - O (a) t) and the corresponding random vector u using
-1

73

the Polak-He algorithm. Then, the reliability constraint is updated with the new random
vector u. Terminate the Polak-He algorithm when the user-defined tolerance s is
satisfied. Task D2 is similar to task BI in the DSA-MOOA approach with fixed high
tolerance.
T a s k s D l a n d D 2 are repeated until both design variables a n d r a n d o m variables are
consistent.

In other

u -u .||<
/+1

and/or

words,

i>i ,
E

the calculations

where

are terminated

when

||x,. -x,.|| < s,

+1

is a predefined positive parameter a n d

is the

m a x i m u m n u m b e r o f iterations. T y p i c a l l y , 3 to 10 iterations ( D l t o D 2 ) are required to

find

6.4

an optimal solution.

Object-Oriented Implementation in OpenSees

I m p l e m e n t a t i o n o f R B D O procedures f o r finite element applications poses a n u m b e r o f

challenges to the software developer. First, the software must b e able to incorporate n e w
d e v e l o p m e n t s a n d features as research progresses. S e c o n d , the interaction b e t w e e n the
finite element, reliability, a n d optimization procedures m u s t b e robust a n d efficient. F o r
i n s t a n c e , it i s r e q u i r e d t h a t r e a l i z a t i o n s o f r a n d o m v a r i a b l e s a n d d e s i g n v a r i a b l e s b e
r e p e a t e d l y m a p p e d i n t o t h e finite e l e m e n t m o d e l . A d d i t i o n a l l y , t h e r e s p o n s e a n d r e s p o n s e
gradients must b e accurately a n d efficiently c o m p u t e d i n the finite element m o d e l a n d
returned to the reliability a n d optimization algorithms. O p e n S e e s has turned out to b e a n
ideal software
software

platform

f o r this p u r p o s e . T h i s i s m a i n l y

d u e to the object-oriented

a r c h i t e c t u r e that t h r o u g h o u t t h e e v o l u t i o n o f O p e n S e e s h a s k e p t

focus o n

maintainability and extensibility.

T o implement the R B D O ,

five

while t w o n e w analysis types

added to the original
optimization

work

costFunction,

and

n e w objects were a d d e d into the

(runDSA-MOOAAnalysis

ReliabilityAnalysis

adds

and

ReliabiltyDomain,

runDSA-SAnalysis)

were

object. A s s h o w n i n F i g u r e 6.5, the present

designVariable, designVariablePositioner, objectiveFunction,

constraintFunction

objects to the

14

ReliabilityDomain.

designVariable

object defines design variables by given their start points. A designVariablePositioner

object is used to map the design variables into structural properties in the finite element
model. The objectiveFunction, costFunction, and constraintFunction objects are userdefined expressions.
subsequently

A costFunciton object is used to create functions that are

combined into an expression

for an objectiveFunction

object.

constraintFunction object may be expressed using various quantities including design

variables and structural response quantities from an OpenSees finite element analysis.

ReliabilityDomain

designVariable

costFunction

constraintFunction

obj ecti veFunction

designVariablePositioner

Figure 6.5 New objects for optimization analysis in ReliabilityDomain

(The diamond symbol denotes analysis tools)
In Figure 6.6, two new "analysis types" added to OpenSees are named runDSAMOOAAnalysis and runDSA-SAnalysis. They are the orchestrating algorithms introduced
in this thesis. The runDSA-MOOAAnalysis is the top level of the D S A - M O O A approach
and is responsible for obtaining the optimal design by orchestrating tasks A l to A 3 . The
runDSA-SAnalysis is the top level of the DSA-S approach and is responsible for
obtaining the optimal design by orchestrating tasks CI to C 3 .
Three

new

"analysis

tools"

are

also

implemented:

NonlinSinglelneqOpt,

NonlinMultilneqOpt, and LinMultilneqOpt. These are so-called base classes that promise
features but do not contain actual implementations. Any number of sub-classes may be
implemented to perform the promised features. This illustrates the extensibility feature of
OpenSees: new algorithms to perform various analysis tasks can be implemented without
having to modify the software framework. The base class NonlinSinglelneqOpt promises
to solve tasks B l and D2. The sub-class implemented for this base class is named
PolakHeNonlinSinglelneqOpt.

Similarly, the base class NonlinMultilneqOpt promises to

75

solve

tasks

B3

and

Dl.

The

current

subclass

implementation

is

PolakHeNonlinMultilneqOpt. As mentioned above, in tasks B3 and D l a quadratic sub-

optimization problem with linear constraints is solved to find the search direction. This is
fulfilled by the base class LinMultilneqOpt. Currently, subclass LSSOLLinMultilneqOpt
is available in the implementation of OpenSees.
ReliabilityAnalysis

runD S A-MOO AAnaly sis

runDSA-SAnalysis

NonlinSinglelneqOpt

NonlinMultilneqOpt

PolakHeNonlinSinglelneqOpt

PolakHeNonlinMultilneqOpt

evaluateFun evaluateGradFun

evaluateFun LinMultilneqOpt evaluateGradFun

LSSOLLinMultilneqOpt
Figure 6.6 Software framework for optimization analysis in OpenSees
(Triangle symbol denotes the relationship between base class and
subclasses, while the diamond symbol denotes analysis tools)
The category of "analysis tools" also contains classes such as evaluateFun and
evaluateGradFun. The evaluateFun object evaluates the values of objective functions,

76

cost functions, and constraint functions. The evaluateGradFun object evaluates the
gradients of objective functions, cost functions, and constraint functions.
Optimization module
evaluateGradFun

evaluateFun

d[c (x) + c (x)a] dfjjx)

[c (x) + c (x)a],fj(x)
0

dx

' dx

NonlinMultilneq ^^
Opt
Pf>f:

dy/
~dx~

u,x

Reliability
module

runDSA-MOOAAnalysis /
runDSA-SAnalysis

Pf

u,x

. dd dd
d,,

Finite element
. module

du dx

Figure 6.7 Interaction between optimization, reliability, andfiniteelement module

Of particular interest in this implementation is the interaction between the above
algorithms and the existing analysis tools in OpenSees. Implementations were already
available to map random variables u and design variables x into thefiniteelement model
and the reliability module, as well as to obtain response and response sensitivities. Figure
6.7 schematically shows the interfaces among the optimization algorithm, the finite
77

element module, and the reliability module i n OpenSees. G i v e n design vector x and
random vector u, the finite element module can compute response d and response
sensitivities dd/du

and dd/dx with respect to random variables and design variables,

respectively. The reliability module can return failure probability p ,

f

limit-state function

value g , and gradients dg I du with respect to random variables and dg I dx with respect
to design variables. Function

i n Figure 6.7 is the negative value o f limit-state function

g. Inside o f the optimization module, evaluateFun and evaluateGradFun objects can

provide the values and gradients o f the objective functions, cost functions, and constraint
functions. Response sensitivity methods F D M and D D M in Chapter 4 are used to
compute such gradients.

78

Chapter 7

Numerical Examples and Case Studies

In this chapter, the new implementations in OpenSees are demonstrated by performing

reliability-based design optimization (RBDO) of a real-world structure. The structure
under consideration is a six-storey reinforced concrete building that serves as a design
example in the Canadian Concrete Design Handbook (Cement Association of Canada,
1995). Three structural models are considered: (1) a linear elastic model, where the
elements are modelled using the

elasticBeam

element of OpenSees; (2) a nonlinear

model, where the elements are modelled using the

beamWithHinges

element of

OpenSees; and (3) a nonlinear model, where the elements are modelled using the
nonlinearBeamColumn

element of OpenSees with fibre-discretized cross-sections. The

response of the building is assessed using the static "pushover" analysis. The limit-state
function is specified in terms of the total drift of the structure. The objective is to
minimize the total expected cost of the

design, given structural constraints and reliability

constraints. Moreover, this study compares the convergence performance and the
computation time for the algorithms and structural models presented herein. In particular,
convergence problems may occur in the algorithms that address the inner and outer
approximation problems described in the previous chapter. For example, a scaling of the
involved functions is required when the Polak-He algorithm is used to address these
problems. We also observe that the computational time is increased when redundant
(inactive) constraints are added. Such experience from the hands-on optimization analysis
is valuable for users of the developed software and is reported below.

7.1

Six-Storey Ductile M o m e n t Resisting

F r a m e

Consider a six-storey reinforced concrete frame building located in Vancouver, Canada

(Cement Association of Canada, 1995). The building has seven bays with 6m spacing in
the North-South (NS) direction and three bays (two office bays with 9m spacing and a
79

central 6m corridor bay) in the East-West (EW) direction. The interior columns are all
500 x500mm, while the exterior columns are all 450 x450mm. The beams of both NS and
EW frames are 400mm wide x 600mm deep for the first three storeys and 400 x550mm
for the top three storeys. Concrete with mean strength f = 30MPa is used throughout,
c

and the reinforcement has mean yield strength f - 400MPa. The Canadian Concrete
y

Design Handbook (1995) specifies that the frame is designed as a ductile moment
resisting frame with R = 4.0, where R is the ductility force modification factor that
reflects the capacity of a structure to dissipate energy through inelastic behaviour.

H6

H5
o

H4

P3

P6

P6

P3

P2

P5

P5

P2

P2

P5

P5

P2

PI

P4

P4

PI

PI

P4

P4

PI

PI

P4

P4

PI

OO

O
>/->

H3

H2 J

HI
o
co

9,000

6,000

9,000

Figure 7.1 Ductile moment-resisting frame model

80

This thesis aims to optimize the design of the columns and beams of this ductile
moment resisting frame. For this purpose, we consider linear and nonlinear pushover
analyses of the second EW frame. The finite element model and the applied loads are
illustrated in Figure 7.1.
The load case of "l.Oxdead load + 0.5xlive load + l.Oxearthquake load" is considered
in the analysis. We consider dead loads and live loads as deterministic. The lateral loads
from ground motion have lognormal distribution. Their means and coefficients of
variation are shown in Table 7.1.

Table 7.1 Vertical loads and lateral loads (c.o.v. indicates the coefficient of variation)
Loads
H
H

Pi
Pi
Pi

P4
Ps

P6

7.1.1

Mean
28490 kN
48950 kN
70070 kN
89100 kN
109780 kN
131890 kN
108000 kN
105000 kN
96000 kN
184000 kN
178000 kN
182000 kN

c.o.v.
0.15
0.15
0.15
0.15
0.15
0.15
N/A
N/A
N/A
N/A
N/A
N/A

Type
lognormal
lognormal
lognormal
lognormal
lognormal
lognormal
N/A
N/A
N/A
N/A
N/A
N/A

Description
random lateral load on floor 1
random lateral load on floor 2
random lateral load on floor 3
random lateral load on floor 4
random lateral load on floor 5
random lateral load on roof
deterministic vertical load
deterministic vertical load
deterministic vertical load
deterministic vertical load
deterministic vertical load
deterministic vertical load

Case 1: Elastic Analysis using elasticBeam Elements

In earthquake engineering it is common to assess the structural capacity using inelastic

pushover analysis. As a reference, a linear elastic analysis is also performed. In this
thesis, the elasticBeam element of OpenSees is used for the linear elastic analysis. This
type of element contains a linear elastic material model, without any yield limit. The
81

"equal displacement principle" shown i n Figure 7.2 is employed to compute the total
inelastic displacement d , subject to the lateral seismic force V. The solid line in the
e

figure denotes the inelastic response. The corresponding linear system, signified by the
dashed line, applies equivalent lateral seismic force V = VxR and results in the same
e

displacement d .
e

Correspoinding linear system

dJR

Figure 7.2 Equal displacement principle

In

the linear

case,

12 design

x = (b ,h ,b ,h ,b^,h ,b ,h ,b ,h ,b ,h ),
x

variables

are collected

in

the vector

as defined in Table 7.2. A total of 48 random

variables describing the loading and material properties are collected i n the vector
\ = (H ,H ,H ,H ,H ,H , ),
i

42

where H\, H , H , H , H , and H are the

2

equivalent lateral loads from the first storey to the roof, respectively. E\ to 4 2 represent
the modulus of elasticity of the concrete material for all 42 elements. W e assume that all
random variables are lognormally distributed with the means and coefficients of variation
listed i n Table 7.3. Random variables H\ to H$ are correlated with the correlation
coefficient o f 0.7, while random variables E\ to 4 2 are correlated with the correlation
coefficient of 0.7.
82

Table 7.2 Definition and initial values of design variables for Cases 1 and 2
Variable

Initial Value
0.45x0.45m

Jl2

0.45x0.45m

6 x hi

0.50x0.50m

/l

0.50x0.50m

bX
2

64 X

bs x /z5
be* he

0.40x0.60m
0.40x0.55m

Description
width and depth of exterior columns offirstthree stories
width and depth of exterior columns of top three stories
width and depth of interior columns offirstthree stories
width and depth of interior columns of top three stories
width and depth offirstthree stories' beams
width and depth of top three stories' beams

Table 7.3 Statistics of random variables in Case 1 (c.o.v. indicates the coefficient of
variation, and cc. indicates the correlation coefficient)
Variable

Mean

4x28490 kN

0.15

4x48950 kN

0.15

#3

4x70070 kN

0.15

4x89100 kN

0.15

4x109780 kN

0.15

He

4x131890 kN

0.15

E\ ~ E42

24648 MPa

0.15

c.o.v. cc.

0.7

0.7

Type
lognormal
lognormal
lognormal
lognormal
lognormal
lognormal
lognormal

Description
equivalent lateral load on floor 1
equivalent lateral load on floor 2
equivalent lateral load on floor 3
equivalent lateral load on floor 4
equivalent lateral load on floor 5
equivalent lateral load on roof
modulus of elasticity of concrete

The reliability problem for the frame is defined in terms of the limit-state function
g(d(x,v)) = 23.1x0.02-rf

roof

(7.1)

where 23.1m is the height of the frame, 0.02 is the maximum limit of the drift ratio, and
d

T00{

is the roof displacement.

In this thesis, our objective is to achieve a frame design that minimizes the total
expected cost, given specific constraints. For this purpose, we model the initial cost of
design and the cost of failure in terms of the total volume of the members. The cost of

83

failure is assumed to be five times the volume of the members. This leads to the
following objective function:

where L, represent the total length of the members in each of the six categories identified
in the design vector, while hi and hi are cross-sectional dimensions. The reliability
constraint is prescribed as p (x) < 0.00135, which implies a minimum reliability index
f

of 3.0. The structural constraints are prescribed to be 0 < 6 hi and 0.5 < 6,//z, < 2 to
ensure positive dimensions and appropriate aspect ratios, where / = [1, 2, 3, 4, 5, 6].
A stand-alone finite element reliability analysis using elasticBeam elements was
performed. For initial values of the design variables in Table 7.2 and mean realizations of
the random variables in Table 7.3, the lateral displacement at the roof level was 238mm.
The corresponding drift ratio was 238/23100 = 1.03%, which was less than the limit of
2%. Afirst-orderreliability analysis (FORM) resulted in a reliability index 8= 3.646 and
r

corresponding failure probability p (x ) = 0.000133, which is acceptable according to

f

the prescribed reliability constraint. The total expected cost of the initial design in terms
of volume was 54.062m .
3

The first optimization analysis was performed using the DSA-MOOA approach. This
approach starts from the semi-infinite optimization analysis (task Al), which iteratively
updates the constraints represented by ip and obtains improved designs for / = 1. These
were identified earlier as tasks B l to B3. In this case, the limit-state function was linear,
since a linear relationship exists between the random variables and the response quantity
d f. Convergence was achieved within 1 to 3 iterations for task B l , and within 1 to 10
roo

iterations for task B3. After discretizing the ball using the method of outer approximation
(MOOA) algorithm by 75 pointsafter 75 loops of task B l to B3the algorithm
repeatedly produced the same design. This was taken to indicate convergence. At the
optimal design there were five reliability constraints. The tolerance of this solution to the
"true" converged solution was cr =0.1/ 75 =1.78xl0" . The total expected cost was
2

84

reduced from 54.062m to 38.701m , while the failure probability was 0.00135, which
3

satisfied the reliability constraint. Then, an importance sampling analysis based on the
new design variables was performed in task A2 to get the "real" failure probability with a
2% coefficient of variation of the sampling result. The results were 38.71 lm for the total
3

cost and 0.00140 for the failure probability. This difference was acceptable, and the
RBDO was stopped after one top level of iteration (tasks A l and A2). This was due to the
linear nature of the problem.
The second optimization analysis was performed using the DSA-S approach. The
algorithm started with task CI, which sequentially completed a deterministic
optimization analysis and then found a new reliability constraint for t = 1. These were
identified earlier as tasks Dl and D2. Similar to the DSA-MOOA approach, the
convergences within tasks Dl and D2 was achieved quicklywithin 1 to 30 iterations for
task Dl and within 1 to 6 iterations for task D2since the limit-state function was linear.
We used the same tolerance as with the DSA-MOOA to judge whether convergence was
achieved (i.e., s = 0.1/75 =1.78xl0" ). The optimal design was achieved after three
2

loops of tasks Dl and D2. The DSA-S approach and the DSA-MOOA approach gave the
same design: the total expected cost was 38.701m and the failure probability was
3

0.00135. In the next task, C2, we got the same solution as in task A2 in the DSA-MOOA
approach using importance sampling with a 2% coefficient of variation of the sampling
result. We accepted this as the optimal design and terminated the analysis.
Table 7.4a Results from RBDO analysis for Case 1
t

1.000

bx

h\

bi

hi

bi

ht,

1.000

hs

0.45 0.45 0.45 0.45

0.5
0.5
0.5
0.5 , 0.4
0.6
0.236 0.471 0.213 0.426 0.358 0.716 0.294 0.588 0.311 0.622

Table 7.4b Results from RBDO analysis for Case 1 (continued)

t

bs

be

he

+c a

c +c p

54.062
38.701

54.064
38.711

0.4
0.55 0.000133 0.000141
0.261 0.522 0.00135 0.00140
85

Tables 7.4a and 7.4b show the results obtained from the two implemented approaches.
The presented results include the value of the 12 design variables, the auxiliary parameter
a, the failure probability p

from the importance sampling with a 2 % coefficient of

variation, and the total expected costs corresponding to a and p .

f

The first rows show

the values of the initial design, while the second rows show values of the optimal design.
The two approaches produced the same solution, although the D S A - M O O A approach
guaranteed convergence with a first order approximation, while the D S A - S approach did
not.

0.00

0.08

0.16

0.24

0.32

0.40

0.48

R o o f Drift (m)
Figure 7.3 Structural responses for Case 1 (load factor versus roof displacement)
at: (1) the mean point of the initial design; (2) the M P P of the initial design; (3)
the mean point of the optimal design; and (4) the M P P of the optimal design.

86

Figure 7.3 shows the structural response of four characteristic realizations of design
variables and random variables. The response at the mean realization of the random
variables for the initial (original) design is shown as the thin solid line. As expected, this
response is linear. At the most probable failure point (MPP) of the initial design, the
structural response is still linear, but reaches the 2% of drift limit 0.462m, as prescribed
by the limit-state function. The structural response at the mean realization of random
variables for the optimal design is shown as the thick solid line. The response is linear,
but has a larger displacement than in the initial design, which is consistent with the
failure probability of the structure increasing from 0.000133 to 0.00135. Finally, the
structural response at the MPP of the optimal design is linear and reaches 2% of the drift
limit. The optimal design has an acceptable reliability, but a lower total expected cost.
This serves as an indication of the usefulness of the RBDO approach.

Table 7.5 Comparison of computational time for Case 1

DSA-MOOA
DDM
FDM
555
8559

DSA-S
DDM
147

FDM
1227

with

dg/du

80

N/A

N/A

with

dg/dx

347

N/A

58

N/A '

Importance Sampling

112486

112607

It is of interest to compare the computational cost of two implemented approaches,

particularly with respect to which method is being used to compute the response
gradients, the direct differentiation method (DDM) or the finite difference method
(FDM). For this purpose, Table 7.5 lists the number of calls to the limit state function,
which is a key indicator when studying computation time. The DDM method is much
more efficient in computing gradients than the FDM method. For example, the DSAMOOA approach requires 555 single limit-state function calls, 80 limit-state function
calls together with the computation of dg/du, and 347 limit-state function calls together
87

with the computations of dg/dx. Thus, altogether 982 limit-state function calls are
required for the DDM method. On the other hand, the FDM method computes the
gradient using one extra limit-state function evaluation for each random variable and
design variable. Thus, the total number of limit-state function calls required for the FDM
method is 555, as well as 80 times the number of random variables (48) and 347 times the
number of design variables (12). Finally, 8,559 limit-state function calls are required,
which is much more than needed for the DDM method. The two approaches require a
similar number of simulations to compute the failure probability using importance
sampling, shown in the last row of Table 7.5.
The DSA-S approach, which requires 213 (147+8+58) limit-state function calls,
appears to be more efficient than the DSA-MOOA approach, which requires 982 limitstate function calls for the DDM method. The key reason for this is that only one
reliability constraint is maintained in the DSA-S approach, while the DSA-MOOA
approach expands reliability constraints step by step when discretizing the ball with more
points.

7.1.2

Case 2: Nonlinear Analysis using beamWithHinges Elements

In this section we perform a nonlinear pushover analysis by using the beam WithHinges
element of OpenSees. We consider the plasticity to be concentrated at over 10% of the
element length at the element ends. The elastic properties are integrated over the beam
plastic hinge

linear elastic

plastic hinge

-O
0 right node

left node 0o
0.1Z

0.81

Figure 7.4 beamWithHinges element

88

0.12

interior, which is considered to be linear elastic. Forces and deformations o f the inelastic
region are sampled at the hinge midpoints. A bi-linear or smooth uniaxial material is used
in the plastic hinge region to model the moment-rotation

relationship. A typical

beamWithHinges element used in this thesis is illustrated in Figure 7.4.

A s in the linear case, 12 design variables shown in Figure 7.2 are collected in the
vector x = (b ,h ,b ,h ,b ,h ,b ,h ,b ,h ,b ,h ).

48 random variables are collected in

the vector v = (H ,H ,H ,H ,H ,H ,E ,---,E )

to describe loading and material

42

properties. W e assume that all random variables are lognormally distributed with the
means and coefficients o f variation listed in Table 7.6. Random variables H\ to H$ are
correlated with the correlation coefficient o f 0.7, and random variables E\ to 42 are
correlated with the correlation coefficient of 0.7. The limit-state function and the
objective function are as defined in Eqs. (7.1) and (7.2). The reliability constraint and
structural constraints are as prescribed for the linear structure.

Table 7.6 Statistics o f random variables for Case 2 (c.o.v. indicates the coefficient of
variation, and c.c. indicates the correlation coefficient)
Mean

c.o.v.

H-

28490 k N

0.15

lognormal

lateral load on floor 1

48950 k N

0.15

lognormal

lateral load on floor 2

#3

70070 k N

0.15

lognormal

lateral load on floor 3

89100 k N

0.15

lognormal

lateral load on floor 4

109780 k N

0.15

lognormal

lateral load on floor 5

131890 k N

0.15

lognormal

lateral load on roof

11097 M P a

0.15

lognormal

modulus o f elasticity o f concrete

E\~ E42

c.c.

0.7

0.7

Type

Description

Variable

A stand-alone finite element reliability analysis was performed. The bi-linear material
model was employed to model the plastic hinges. The stiffness o f cross-section was
evaluated by EI = Ebb? / 1 2 , where b and h were the width and the depth o f the section,

89

and the value of E was smaller than in the linear case because of considering concrete
cracking. For all columns, the yield stain e = 0.84, and the strain hardening factor
y

a -0.5. For all the beams, the yield stain s =0.52, and the strain hardening factor
y

a - 0.3. At the mean realization of the random variables in Table 7.6 and with the initial
design in Table 7.2, the lateral displacement at the roof level was 131mm. The
corresponding drift ratio was 131/23100 = 0.57%, which was less than the limit of 2%. A
reliability analysis using the FORM resulted in a reliability index /? = 3.536 and the
corresponding failure probability p (\ ) = 0.000203, which satisfied the prescribed
f

reliability constraint. The total expected cost of the initial design was 54.080m .
The first optimization analysis was performed using the DSA-MOOA approach with
the bi-linear material model. As outlined previously, the algorithm starts with the semiinfinite optimization analysis (task Al). In this case, the convergence within task B3 was
achieved for the first few iterations, namely when the number of constraints represented
by ip was low. However, the algorithm in task B3, the PolakHeNonlinMultilneqOpt,
exhibited progressively slower convergence as the number of constraints increased. In
fact, this problem made the algorithm grind to a halt. The presence of gradient
discontinuities due to sudden yielding events of the bi-linear material models was taken
to be the reason for this problem.
As a remedy to the above problem, a smoothed version of the bi-linear model
introduced in Chapter 5 was substituted. A circular segment in a normalized stress-strain
plane that started at 80% of the yield strength, y - 0.8, was employed to smooth the bilinear material as illustrated in Figure 5.2. Remarkably, the analysis proceeded without
any of the convergence problems described above. Convergence was achieved within 1 to
6 iterations for task BI, and within 1 to 65 iterations for task B3. This led us to conclude
that the presence of a non-smooth response surface due to sudden yielding events was a
serious impediment to the performance of the algorithm. Similar problems were also
observed in the stand-alone reliability analysis. However, in our experience the problem
was significantly amplified in the optimization analysis context.
90

After discretizing the ball in the MOOA algorithm by 75 points, or after 75 loops of
tasks B l to B3, the algorithm repeatedly produced the same design. At the design point,
there were 12 reliability constraints. The tolerance of this solution to the "true" converged
point was a = 0.1/75 =1.78xl0" . The total cost was reduced from 54.080m to
2

37.108m , and the failure probability was 0.00135, which satisfied the reliability
3

constraint. In the next task, A2, importance sampling based on the new design variables
was performed to get the "real" failure probability with a 2% coefficient of variation of
the sampling result. The results were 37.127m as the total cost and 0.00146 as the failure
3

probability. The difference between the two failure probabilities (0.00135 from task A l
and 0.00146 from task A2) shows the nonlinearity of the structure. In task A3, parameter
/ was updated, and the top level of the iteration (tasks A l to A3) was repeated. After two
more loops of tasks A l to A3, the differences of failure probabilities between task A l and
A2 were reduced and accepted, and the RBDO was stopped. The final total cost was
37.197m , and the failure probability was 0.00135.
3

The second optimization analysis was performed using the DSA-S approach with the
smooth material model. The algorithm started with task CI. Convergence was achieved
within 1 to 109 iterations for task D l , and within 1 to 13 iterations for task D2. We used
the same tolerance to judge the consistent design

(i.e., e = 0.\/75

= 1.78xl0" ). The
5

consistent design was achieved after four loops of tasks Dl and D2. The DSA-S approach
and the DSA-MOOA approach gave the same design: the total cost was 37.108m and the
failure probability was 0.00135. In the next task, C2, the DSA-S approach produced the
same solution as task A2 in the DSA-MOOA approach using importance sampling with a
2% coefficient of variation. As in the DSA-MOOA approach, the top level of iteration
(tasks CI to C3) was repeated for two more loops, producing consistent designs. The
optimization procedure was stopped at the total cost of 37.197m and the failure
3

probability of 0.00135.
Tables 7.7a and 7.7b show the results obtained from the two implemented approaches.
The presented results include the value of 12 design variables, the auxiliary parameter a,
the failure probability p from the importance sampling with a 2% coefficient of
91

variation, and the total expected cost corresponding to a and p . The first rows show the
values of the initial design, while the following rows show values of the optimal design.
In each of these iterations, the parameter t is updated to account for nonlinearities in the
limit-state function. After the first iteration, the value of t was updated as
1.0xO (0.00135)/O"'(0.00146) = 1.0077. The analysis was carried out for two more
_1

iterations. No appreciable difference in the design was observed. In the last row of the
table, a and p converge to the same acceptable value, 0.00135. In effect, the objective
functions have reached the minimum value: 37.197m . Hence, the design variables in the
3

last row of the table can be accepted as the optimal design.

Table 7.7a Results from RBDO analysis for Case 2

t

1.000
1.0077
1.0089
1.0086

bx

hi

bi

/?4

hs

0.45 0.45 0.45 0.45

0.5
0.4
0.5
0.5
0.5
0.6
0.246 0.491 0.200 0.399 0.316 0.631 0.289 0.577 0.323 0.645
0.246 0.492 0.200 0.400 0.316 0.632 0.289 0.578 0.323 0.646
0.246 0.492 0.200 0.400 0.316 0.632 0.289 0.578 0.323 0.646

Table 7.7b Results from RBDO analysis for Case 2 (continued)

t

be

he

1.000
0.4
0.55
1.0077 0.246 0.492
1.0089 0.246 0.492
1.0086 0.246 0.492

c +c a

0.000203
0.00135
0.00135
0.00135

0.000219
0.00146
0.00137
0.00135

54.080
37.108
37.186
37.197

c +c p
0

54.085
37.127
37.189
37.197

It is observed that the two approaches obtain an improved design that is close to the
final solution already after the first top-level iteration. This can also be seen in Figures
7.5, where the total expected costs (objective function) are plotted as the function of the
iteration number. This phenomenon shows that the structure is not highly nonlinear and
that the implemented approaches are effective in dealing with the nonlinear problem.
92

2
Loops of Top Level

Figure 7.5 Evolution of the total expected cost for objective functions for
Case2

(l)c (x) + c (x)a and (2) c (x) + c (x)p(x)

0

Figure 7.6 shows the structural response for four characteristic realizations of design
variables and random variables. The response at the mean realization of random variables
of the initial (original) design is shown as the thin solid line. As expected, this response is
close to linear, because no significant damage (yielding) is anticipated for this realization.
At the MPP of the initial design, however, substantial yielding occurs. This is reasonable,
since this realization represents failure. Third, the structural response at the mean
realization of random variables for the optimal design is shown as the thick solid line.
This response has larger displacement than the initial design. Finally, the structural
response at the MPP of the optimal design is also shown. Again, significant nonlinearity
in the finite element response is observed. This response is similar to that of the initial
design, but it is not equal to it. This is reasonable, because the limit-state function is
93

0.00

0.08

0.16

0.24

0.32

0.40

0.48

R o o f Drift (m)

Figure 7.6 Structural responses for Case 2 (load factor versus roof displacement)
at: (1) the mean point of the initial design; (2) the MPP of the initial design; (3)
the mean point of the optimal design; and (4) the MPP of the optimal design.
altered by changes in the structural design. The apparent lower value of the stiffness at
the MPP of the optimal design is explained as follows: for the optimal design a greater
reduction of the stiffness is required to "achieve" failure (i.e., to obtain the MPP). Again,
we observe that the optimal design has an acceptable reliability and a reduced total
expected cost. This serves as an indication of the usefulness of the RBDO approach.
Table 7.8 compares the computational cost of the two implemented approaches. The
data in the table come from the first iteration, and are almost the same as the data from
the second and third iterations. The DSA-S approach, requiring 6,195 limit-state function
calls, appears to be more efficient than the DSA-MOOA approach, requiring 22,730
limit-state function calls using the FDM method. We have observed that the nonlinear
94

case requires significantly more effort than the linear case, which only requires 1,227 and
8,559 limit-state function calls. For the linear case, updates in design do not dramatically
change the corresponding M P P of the reliability analysis. There are only five reliability
constraints after 75 loops of tasks B l to B3. For the nonlinear cases, however, the M P P
of the reliability analysis clearly changes when a new design is found. In addition, in
nonlinear cases there are 12 reliability constraints after 75 loops of tasks B l to B3.

Table 7.8 Comparison of computational time for Case 2

D S A - M O O A by F D M

DSA-S by F D M

22730

6195

Importance Sampling

109436

109342

7.1.3

Case 3: Nonlinear Analysis using dispBeamColumn Elements and fibre

Sections
In this section we perform a nonlinear pushover analysis by using the dispBeamColumn
element and the fibre section of OpenSees. To describe a better curvature distribution
along the element, one original element was divided into four elements, with four
integration points along each element. Each column and beam section was discretized
into about 20 fibres to give an "approximately continuous" structural response. A l l of the
fibres were described using the bi-linear concrete

material. In this case, an

elasticPerfectlyPlastic material is used as the concrete material by setting FyP as 0, as

shown in Figure 4.4. For the reinforced bars of these sections, the smooth steel material
was used, as shown in Figure 5.2. Typical fibre sections of columns and beams are
illustrated in Figure 7.7.
In

this

nonlinear

case,

18

design

variables

are

collected in the

vector

x = ( 6 , , h , b , h , b , h , b , h , b , h , b , h , A , A , A , A , A , A ) . In addition to b and h
x

()

95

defined in Cases 1 and 2, this case has the area of steel bars A as design variables. The
definitions and initial values of b, h, and A are described in Table 7.9.

"

unconfined concrete
20 fibers

confined concrete
20 fibers
reinforced steel layer

unconfined concrete
2 fibers

Column Fiber Section

Beam Fiber Section

Figure 7.7 Typicalfibresections for columns and beams

78 random variables for the loading and material properties are collected in the vector
V=

C^l'"

' ' -^6 '

feel''"''

fees ' ^ eel"

*' ^ecS >fcl>'">fcU>

^ c l ' " " ' ^ c l 4 ' fy\'''"'

fy

14 ' ^ 1 ' " " " '

^ 1 4 )

We assume that all random variables are lognormally distributed with the means and
coefficients of variation listed in Table 7.10. The random variables are correlated with the
correlation coefficient of 0.7 in several groups. More specifically, we have eight random
variables for confined concrete strength f ' and eight random variables for modulus of
cc

elasticity of confined concrete E . They are assigned to eight types of columns: first
cc

three-storey columns and top three-storey columns on four axes A, B, C, and D. We also
have 14 random variables for unconfined concrete strength f ' and 14 random variables
c

for modulus of elasticity of unconfined concrete E . They are assigned to eight types of
c

columns and six types of beams:firsttwo-storey beams, middle two-storey beams, and
top two-storey beams. In addition, we have 14 random variables for steel bars strength
/

and 14 random variables for modulus of elasticity of steel E assigned to eight types

of columns and six types of beams.

96

Table 7.9 Definition and initial values of design variables for Case 3
Variable

Initial Value

b\x h\

0.45x0.45m

A\

0.003m

0.45x0.45m
2

0.003m

63 x hi

0.50x0.50m

A3

0.003m

64X

0.50x0.50m

hi,

0.003m

65X

0.40x0.60m
0.0024m
0.40x0.55m
0.0024m

hs

64X

Description
width and depth of exterior columns offirstthree stories
half of the area of reinforced bars of exterior columns of
first three stories
width and depth of exterior columns of top three stories
half of the area of reinforced bars of exterior columns of top
three stories
width and depth of interior columns offirstthree stories
half of the area of reinforced bars of interior columns of first
three stories
width and depth of interior columns of top three stories
half of the area of reinforced bars of interior columns of top
three stories
width and depth of exterior columns offirstthree stories
area of reinforced bars offirstthree stories' beams
width and depth of exterior columns of top three stories
area of reinforced bars of top three stories' beams

The limit-state function was as defined in the same way as in Eq. (7.1). The objective
function was described in terms of the total volume of the members. Because of the price
difference between two materials in the current market (the price of steel bars per cubic
meter was 100 times the price of the concrete per cubic meter), the volume of steel bars
was accounted for by using its equivalent concrete volume, which was equal to 100 times
the actual volume of the steel bars. Again, the cost of failure was assumed to be five
times the initial volume. This led to the following objective function:
(

(*,*, + 100 2 A,) L, + X L ( b , h , + 100 A,) L, )

+ ^(x)

5( X;

= 1

fM

+ 100 2 A,) L, + ZlsWt +

1 0 0

4) i
L

where Z, represents the total length of the members in each of the six categories identified
in the design vector. The reliability constraint was still prescribed as p (x) < 0.00135.
f

97

The structural constraints were prescribed as 0 < b , h and 0.5 < bilh < 2 to ensure
t

positive dimensions and appropriate aspect ratios, where / = [1, 2, 3, 4, 5, 6]. The
structural constraints for the area of steel bars were 0.01-6,72,< A < 0.02-b^
i

for

columns, where i = [1, 2, 3, 4], and 0.008 b,h, < A, < 0.02 b h for beams, where i - [5,
i

6], to ensure appropriate reinforced bar ratios.

Table 7.10 Statistics of random variables for Case 3 (c.o.v. indicates the coefficient of
variation, and cc. indicates the correlation coefficient)
Variable
H

#3

He
fcc\

fcc%

Ecc\ ''' EccS

fc \ "'

Eel

fy\ "'
E\

fc\A

"'E \4
C

fy\A
"'E\4

Mean
28490 kN
48950 kN
70070 kN
89100 kN
109780 kN
131890 kN

c.o.v.
0.15
0.15
0.15
0.15
0.15
0.15

cc

39MPa

0.15

0.7

lognormal confined concrete strength

9750 MPa

0.10

0.7

lognormal

30 MPa

0.15

0.7

lognormal unconfined concrete strength

15000 MPa

0.10

0.7

lognormal

400 MPa

0.15

0.7

lognormal steel bars strength

200000 MPa

0.05

0.7

lognormal modulus of elasticity of steel

0.7

Type
lognormal
lognormal
lognormal
lognormal
lognormal
lognormal

Description
lateral load on floor 1
lateral load on floor 2
lateral load on floor 3
lateral load on floor 4
lateral load on floor 5
lateral load on roof
modulus of elasticity of
confined concrete
modulus of elasticity of
unconfined concrete

A stand-alone finite element reliability analysis was performed. At the mean realization
of the random variables in Table 7.10, with the initial design in Table 7.9, the lateral
displacement at the roof level was 168mm. The corresponding drift ratio was 168/23100
= 0.73%, which is less than the limit of 2%. A reliability analysis by the FORM resulted
in a reliability index /? = 3.120 and the corresponding failure probability
98

p (x )
f

= 0.000903,

which satisfied the prescribed reliability constraint. The total

expected cost of the initial design was 130.753m , which was larger than Case 1 and 2,
3

since in this case we considered the area of reinforced bars.

With the experience of solving convergence problem in Case 2, we were confident in
solving the nonlinear problem using

dispBeamColumn

elements and discretized

fibre

sections. First, we performed an optimization analysis using the DSA-MOOA approach.

The semi-infinite optimization analysis (task Al) converges quickly: within 1 to 5
iterations for task B l , and within 1 to 21 iterations for task B3. The results show that
discretized concrete fibre sections, together with the smooth steel material, can achieve a
"continuous" structural response. After discretizing the ball by 75 points, or after 75
loops of task B l to B3, the algorithm repeatedly produced the same design. At the
optimal design there were 17 reliability constraints. The tolerance of this solution to the
"true" converged point was a =0.1/75 = 1.78xl0' . The total cost was reduced from
2

130.753m to 85.221m , and the failure probability was 0.00135, which satisfied the
3

reliability constraint. Next, an importance sampling based on the new design variables
was performed to get the "real" failure probability with a 2% coefficient of variation (task
A2). The results were 85.327m for the total cost and 0.00160 for the failure probability.
3

This difference between the two failure probabilities (0.00135 from task A l and 0.00160
from task A2) shows the nonlinearity of the structure. The parameter / was updated in
task A3, and the top level of iteration (task A l to A3) was repeated. After two more loops
of tasks A l to A3, the RBDO was stopped when the differences in failure probabilities
between tasks A l and A2 were reduced to an accepted level. The final total cost was
85.663m , and the failure probability was 0.00135.
3

The second optimization analysis was performed using the DSA-S approach. The
approach began from task CI. Convergence was achieved within 1 to 88 iterations for
task Dl, and within 1 to 17 iterations for task D2. We used the same tolerance as in the
DSA-MOOA approach to judge the consistent design (i.e.,

e = 0.\/75

=1.78xl0' ). An
5

optimal design was achieved after five loops of tasks Dl and D2. The DSA-S approach
and the DSA-MOOA approach produced the same design. The total cost was 85.221m ,
3

99

and the failure probability was 0.00135. In the next task, C 2 , the D S A - S approach
produced the same solution as task A 2 of the D S A - M O O A approach using importance
sampling with a 2 % coefficient of variation. A s the D S A - M O O A approach, the top level
of iteration (tasks C I to C3) was repeated for two more loops and the designs were
consistent. The entire optimization procedure was stopped at the total cost of 85.663m
and failure probability of 0.00135.

Table 7.1 l a Results from R B D O analysis for Case 3

6,

hx

1.000

0.45

0.45

0.45

0.45

1.0177

0.224

0.449

0.181

1.0182

0.225

0.450

1.0182

0.225

0.450

/?4

hs

0.5

0.5

0.5

0.5

0.4

0.6

0.362

0.306

0.611

0.286

0.572

0.334

0.668

0.182

0.363

0.306

0.612

0.287

0.573

0.335

0.670

0.182

0.363

0.306

0.612

0.287

0.573

0.335

0,670

Table 7.1 l b Results from R B D O analysis for Case 3 (continued)

t

be

he

A,

1.000

0.4

0.55

0.003

0.003

0.003

0.003

0.0024

0.0024

1.0177

0.334

0.524

0.0010

0.0007

0.0019

0.0016

0.0029

0.0014

1.0182

0.335

0.526

0.0010

0.0007

0.0019

0.0016

0.0029

0.0014

1.0182

0.335

0.526

0.0010

0.0007

0.0019

0.0016

0.0029

0.0014

Table 7.1 l c Results from R B D O analysis for Case 3 (continued)

c

+c a

c +c p

1.000

0.000903

0.000941

130.753

130.778

1.0177

0.00135

0.00160

85.221

85.327

1.0182

0.00135

0.00136

85.652

85.655

1.0182

0.00135

0.00135

85.663

85.663

100

Tables 7.11a, 7.11b, and 7.11c show the results obtained from the two implemented
approaches. The presented results include the value of 18 design variables, the auxiliary
parameter a, the failure probability p

from the importance sampling with a 2%

coefficient of variation, and the total expected costs corresponding to a and p . The first
f

rows show the values of the initial design, while the following rows show the values of
the optimal design. In each of these iterations, parameter t was updated to account for
nonlinearities in the limit-state function. After the first iteration, the value of t was
updated as 1.0xO" (0.00135)/O" (0.00160) = 1.0177. The analysis was carried out for
1

two more iterations. No appreciable difference in the design were observed. In the last
row, a and p converge to the same acceptable value 0.00135. In effect, the objective
f

functions have reached the minimum value, 85.663m . Hence, the design variables in the
3

last row were accepted as the optimal design.

(1) Original mean point

(2) Original MPP
(3) Optimal mean point
- (4) Optimal MPP
0.00

0.08

0.16

0.24
Roof

0.32

0.40

0.48

Drift (m)

Figure 7.8 Structural responses for Case 3 (load factor versus roof displacement)
at: (1) the mean point of the initial design; (2) the MPP of the initial design; (3)
the mean point of the optimal design; and (4) the MPP of the optimal design.
101

Figure 7.8 shows the structural response of four characteristic realizations of the design
variables and random variables. The responses at the mean realization of the random
variables and at the MPP of the random variables for the initial (original) design are
shown as the thin solid line and the thin dashed line, respectively. The structural
responses at the mean realization of the random variables and at the MPP of the random
variables for the optimal design are shown as the thick solid line and the thick dashed
line, respectively. The figure shows similar properties to the nonlinear case, using
beamWithHinges

elements in Case 2.

Table 7.12 Comparison of computational time for Case 3

DSA-MOOA
DDM
FDM
5471
40415

DSA-S
DDM
1300

FDM
7618

with

dg/du

91

N/A

33

N/A

with

dg/dx

1547

N/A

208

N/A

Importance Sampling

109436

101848

Table 7.12 compares the efficiency of the FDM and DDM methods, as well as the
computation cost of the two implemented approaches, by measuring the number of calls
to the limit state function. We came to the same conclusion as in the linear case: using the
DDM method to compute the gradients is much more efficient than using the FDM
method, regardless of which optimization approach is adopted.
As shown in Table 7.12, the DSA-S approach, requiring 1,541 (1,300+33+208) limitstate function calls, appears to be more efficient than the DSA-MOOA approach,
requiring 7,109 (5,471+91+1547) limit-state function calls using the DDM method. This
nonlinear case requires much more computational effort than the linear case, 213 and 982
limit-state function calls, respectively. However, the number of limit-state function calls
is almost the same as that required in the nonlinear Case 2.
102

7.2

Practical Experience from Case Studies

This section describes in further detail the observations and practical experiences that
have been gained from the case studies presented above. Comparisons are made between
two implemented optimization approaches (DSA-MOOA and DSA-S), between the FDM
and the DDM methods, and between linear and nonlinear pushover analyses. We also
make the observation that the convergence of the optimization procedure is significantly
improved by removing inactive constraints or by properly scaling the functions involved.

7.2.1

C o m p a r i s o n of T w o Optimization

A p p r o a c h e s

Both the DSA-MOOA and the DSA-S approaches are gradient-based algorithms and
decoupled sequential optimization approaches. The reliability analysis and the
optimization analysis are decoupled in them, so the users have the flexibility to choose
any available reliability methods and optimization algorithms according to their
requirements. However, the two approaches have different behaviours with regards to
convergence performance and computational time.
The two approaches use the same problem reformulation. The original problem and the
reformulated problem are proved to be identical in the first-order approximation. The
DSA-MOOA approach considers the reformulated problem as a semi-infinite
optimization problem and solves it using the MOOA algorithm, which has a converged
solution when using an infinite number of reliability constraints. On the other hand, the
DSA-S approach considers the reformulated problem as an inequality constraint
optimization problem and solves it using the Polak-He algorithm, which can only find a
consistent design without the proof of convergence. According to case study results, the
two approaches can achieve the same solution if the analysis is stopped at the same
tolerance.
When the two approaches' convergence speed is compared, it can be seen that the
DSA-S approach is much faster than the DSA-MOOA approach: the former only needs
103

-20% of limit-state function calls of the latter. The DSA-S approach solves the final
optimization problem using a single reliability constraint, while the DSA-MOOA
approach expands the reliability constraints step by step by discretizing the ball with
progressively more points to achieve a gradually precise solution. Hence, 80% of the
computational time in the DSA-MOOA approach is used to deal with the discretization of
points and a progressively larger reformulated problem.
In conclusion, the DSA-S approach is effective and accurate enough. However, if this
approach fails in the converge procedure, the user has to rely on the DSA-MOOA
approach, which is reliable but slow.

7.2.2

C o m p a r i s o n of T w o Gradient C o m p u t a t i o n

M e t h o d s

Two methods of computing response sensitivities in OpenSees are used in case studies:
the FDM and the DDM. This section compares the two methods in light of three
requirements: consistency, accuracy, and efficiency.
Consistency refers to the computed sensitivities being consistent with the
approximations made in computing the response itself. In the DDM, consistency is
ensured through differentiating time- and space-discretized finite element response
equations (Haukaas & Der Kiureghian, 2005). The computation of the structural response
and the response gradient are both conducted in the finite element analysis. On the other
hand, the FDM simply computes the ratio of the structural response difference and
perturbation.
Accuracy is important to response sensitivity, since the convergence of reliability and
optimization algorithms depend on it. The sensitivities computed by ordinary finite
difference may not be sufficiently accurate to guarantee convergence of the solution
algorithms (Haukaas & Der Kiureghian, 2005). The DDM ensures better accuracy than
the FDM, since the DDM evaluates the exact derivatives of the approximate finite
element response.

104

Efficiency is an important requirement in computation, since sensitivities are

repeatedly computed in the solution algorithms. In the DDM, the response sensitivities
for each parameter are obtained from a linear equation upon convergence of the finite
element response (Haukaas & Der Kiureghian, 2005). The additional time used by the
gradient calculation in the DDM is less than the time required by another nonlinear finite
element analysis. Instead, the FDM method requires one more nonlinear finite element
analysis with perturbed parameter values for each random variable and design variable.
Hence, the computational time by the FDM equals the number of random/design
variables multiplied by the computational time for a single finite element analysis.
The DDM method requires a one-time consolidated effort to derive differentiation
equations and implement them in thefiniteelement program. However, once we have it,
the DDM method is more accurate and efficient than the FDM method.

7.2.3

Comparison of Linear and Nonlinear Analyses

Usually the users choose between linear or nonlinear pushover analyses according to their
requirements and their analysis ability. The comparison in this section shows the possible
results and computational cost for each selected case. This comparison can serve to guide
the users when making the decision about which analysis method to choose.
Table 7.13 presents the comparison between initial and optimal reliability indexes for
the three cases. All three initial reliability indexes are greater than 3.0 regardless of the
analysis model. This implies that the initial design is safe but may not be optimal.
Following the RBDO analysis, the reliability indexes go down to 3.0, which is the lower
bound of the reliability constraint.
The optimal total costs of nonlinear cases are lower than those for the linear case. This
is reasonable, since the linear analysis is based on the "equal displacement principle" and
results in equivalent results, while nonlinear analyses offer more "exact" results.
However, from the structural design point of view, the results of the linear analysis are
also acceptable.
105

Table 7.13 Comparison of linear and nonlinear cases

Initial reliability index

Optimal reliability index
Initial total cost
Optimal total cost
Number of top level of
iteration
Number of limit-state
function calls in first
iteration

Case 1:
Linear
3.646
3.0
1.0
0.715

Case 2: Nonlinear
(beam WithHinges)

Case 3: Nonlinear
(dispBeamColumn

3.536
3.0
1.0
0.688

3.121
3.0
1.0
0.651

982

7047

7109

+ fibre)

We have also compared computational costs for linear and nonlinear cases. In the
linear case, the DSA-MOOA and the DSA-S analyses require only one iteration to
achieve an acceptable design, while in the nonlinear cases three iterations are required. In
thefirstiteration the linear case calls 982 limit-state function calculations, which is about
15% of the number of limit-state function calls required by the nonlinear analysis (about
7,000 calls). Hence, the linear analysis is much more effective than the nonlinear
analysis. For the linear case, with the updating of the design the corresponding MPP of
the reliability analysis does not change dramatically. In addition, there are only five
reliability constraints after 75 loops of tasks B l to B3. On the other hand, the MPP of the
reliability analysis for the nonlinear cases changes apparently when a new design is
found. In addition, there are 12 to 17 reliability constraints after 75 loops of tasks B l to
B3.
In conclusion, nonlinear analyses produce "exact" and trustworthy optimal designs,
while the linear case is effective and its results are also acceptable. It is advisable to
conduct a linear analysis for an optimal design first. If the user really needs a more
"exact" design, the nonlinear analysis can start from the results of the linear analysis.

106

7.2.4

Active a n d Inactive

Constraints

There are three categories of constraints in the reformulated optimization problem: the
deterministic constraints f (x) < 0 , reliability constraints if/ < 0 , and constraints p

<p

A t the optimal design point, the limit-state function is g(d(x,u)) = 0 , which satisfies the
reliability constraints y/ = -g < 0 , making these reliability constraints active. In Tables
7.4, 7.7, and 7.11, all failure probabilities at the optimal design reach the upper bound
p =0.00135, so this constraint is also active.
f

However, by observing the optimal results, we find that some of the deterministic
constraints are not active. For example, in the nonlinear case using

dispBeamColumn

elements and fibre sections only two types of constraints are active in six types of
constraints. In this case, we define the following six constraints:

b>0
b/h<2

h>0
b/h> 0.5

* * P bh
min

where p

m i n

and p

m a x

(7.4)

As < p bh

max

are the lower and upper bounds of longitudinal reinforcement

ratios. Only two categories,

b/h> 0.5 and As>p bh,

min

are active. When all of the

inactive deterministic constraints are removed, the final results are the same as those of
the full constraints, but the computational time is reduced to about 60-80% of the original
time. The reduction in time stems from the reduced size of vector g and matrix G in the
L S S O L analysis. In the two implemented approaches, the time is saved in tasks B 3 and
Dl.
In summary, removing inactive constraints can speed up the optimization procedure.
Once the user finds that some inactive constraints are violated in the analysis, the user
can

stop the analysis and add the constraints back into the optimization to make sure that

the solutions are correct.

107

7.2.5

Acceleration of Convergence Procedure by Proper

Scaling

Both tasks B1/D2 and B3/D1 apply the Polak-He algorithm to solve an optimization
problem. The Polak-He algorithm requires the computation of the values and gradients of
the objective function, deterministic constraints, and reliability constraints. It has been
observed that the different ways to define these functions can affect the convergence
speed.
Without scaling, the objective function is about 50m to 130m , the deterministic
3

constraints are between 0.5 and 2, and the reliability constraints are about 0.2 to 1.0.
These values are not in the same order of magnitude. For the nonlinear analysis, task
B1/D2 requires about 1 to 10 iterations to converge, while task B3/D1 requires several
hundred or even thousand iterations to converge. This convergence speed is not
acceptable. If a scaling is applied (scaling all involved functionsobjective function,
deterministic constrains, and reliability constraints) to approximately the same order 10
= 1.0, the new convergence performance in task B1/D2 remains same, but the
computational cost of task B3/D1 is reduced to less than one hundred iterations (and
often less than 10 iterations). Originally, the functions involved in task B1/D2 had the
same order, so scaling did not affect these tasks. The benefit of task B3/D1 is apparent
because only 10% of the original computational time is required.
The gradients of objective and constraint functions cannot be scaled directly. It may not
be possible to conduct proper scaling for both the functions and their gradients. Scaling
the values of functions to 1.0 can approximately scale the gradient in the order of 10.
This is better than the value of function in the order 10 and the gradient in the order 10 .
2

The bigger the difference between the vector and the matrix cells in LSSOL, the more
"ill-conditioned" the problem becomes. This is a general problem that cannot be fixed
easily. It is recommended to do scaling in the beginning of the process, when defining all
of the functions.
As mentioned in Chapter 6, another scaling skill can be used to speed up the
convergence procedure. Parameter 6 = -0 (a) is used in place of parameter a. With
_1

108

reference to Eq. (2.9), parameter b is a substitute for the reliability index /?, in the same
way as parameter a is a substitute for failure probability p (x). This parameter
f

replacement avoids numerical difficulties caused by the difference in orders of magnitude

between a and other design variables x. Hence, the optimization in the DSA-MOOA
approach is over the design vector (x, b).
In summary, the convergence procedure in the RBDO can be accelerated by properly
scaling the functions involved and by using the reliability index to take the place of the
failure probability in the optimization. The benefit of this scaling is apparent, because it
only takes 10% of the original time.

109

Chapter 8
8.1

Conclusions

Summary of Major Findings

In this thesis we implement two reliability-based design optimization (RBDO)

approaches in the object-oriented software framework OpenSees. The total expected cost
is minimized in the optimization process subject to constraints on structural properties
and component structural reliability. The implementations comprise a merger between
reliability, optimization, and finite element techniques. This enables the RBDO of
comprehensive real-world structures that exhibit nonlinear behaviour. Our work provides
a tool for engineers in making rational decisions based on the balance between cost and
safety in engineering practice.
The fact that the failure probability in terms of a high-dimensional integral is involved
in the objective and constraint functions violates two basic requirements of the standard
nonlinear optimization solver: all involved functions must be evaluated in a finite time
and must be continuously differentiable. In this thesis, the first requirement is satisfied by
evaluating the probability of failure using efficient approximation methods such as the
first-order reliability method and importance sampling. The second requirement is
satisfied by making use of smooth material model and discretized cross-sections.
This thesis proposes two decoupled optimization approaches: the DSA-MOOA and the
DSA-S. These are efficient, robust, and versatile tools for solving RBDO problems. In
both of them, the required reliability and optimization calculations are decoupled, thus
allowing flexibility in choosing optimization algorithms and reliability computation
methods. The original optimization problem is reformulated as a deterministic
optimization problem, which is identical to the original problem in the first order case.
The DSA-MOOA considers the reformulated problem as a semi-infinite optimization
problem and solves it using the method of outer approximation. The DSA-S considers the
reformulated problem as an inequality constrained optimization problem and solves it
using the Polak-He algorithm. This simplified algorithm is demonstrated to provide the
same optimal design as the DSA-MOOA approach for the cases considered.
110

The solutions of the two approaches are based on thefirst-orderapproximation to the

failure probability. If the users accept this failure probability, the analysis is stopped,
resulting in an optimal design. However, for the problem with nonlinear limit-state
function, the new design and failure probabilities are updated using a higher-order
reliability method (such as importance sampling) to take into account the nonlinear
structural behaviour. The parameter t is employed to fulfill the nonlinearity
approximation process. Starting from unity, the parameter t is updated during the
optimization analysis to account for the nonlinearity in the limit-state function.
An effective, accurate, and consistent response sensitivity analysis is essential in
gradient-based optimization algorithms. For the finite element analysis, we need the
gradients of the structural response with respect to model parameters. Two gradient
evaluation methods, the finite difference method (FDM) and the direct differentiation
method (DDM), are employed in this thesis. The DDM method requires the derivation
and implementation of analytical derivatives of the finite element response. Once we
have the DDM implementation in thefiniteelement software, the DDM method is more
accurate and efficient than the FDM method, which requires an additional nonlinear finite
element analysis for each random variable and design variable.
The key difficulty resolved in the implementation is the negative effect of response
gradient discontinuities due to sudden yielding events. The possible response gradient
discontinuities for nonlinear structures cause non-convergence or slow convergence in
the optimization analysis as well as in the first order reliability analysis. Two remedies
are applied in this thesis: the smooth material model builds a "continuously
differentiable" response, and the section discretization scheme results in an
"approximately continuously differentiable" response. Hence, the requirement of
standard nonlinear optimization algorithms is satisfied and the nonconvergence problem
is avoided.
The new implementations results in a modern and comprehensive software, OpenSees,
with RBDO capacities. Object-oriented programming was utilized when extending
OpenSees' reliability, optimization, and sensitivity capabilities. The superior extensibility
111

and maintainability features of this programming type are emphasized. Originally,

OpenSees has employed four principal objectsModelBuilder,
Recorderto

Domain, Analysis,

and

perform the finite element analysis. OpenSees is then extended with the

ReliabilityDomain

and the Reliability

Analysis

object to perform reliability analysis. This

thesis further extends OpenSees with optimization capacities by adding several objects to
ReliabilityDomain
DSA-SAnalysis

and establishing two new analysis types,

DSA-MOOAAnalysis

and

objects. The reliability domain contains all functions involved in the

optimization problem and maps the design variables into the finite element model. The
analysis part includes two RBDO approaches and several analysis tools. The extended
OpenSees has the capacity to perform finite element analysis, reliability and sensitivity
analyses, and the optimization analysis for comprehensive real-world structures
exhibiting nonlinear behaviour.
A numerical example involving a nonlinear finite element analysis of a three-bay, sixstorey building is used to demonstrate the implementations/In particular, the need for a
continuously differentiable response with respect to thefiniteelement model parameters
is emphasized. The linear pushover analysis using

elasticBeam

elements does not

encounter any convergence problems in optimization. The nonlinear pushover analysis

using

beamWithHinges

elements cannot converge using the traditional bi-linear steel

material model. This issue is cured by using the smooth steel material model, in which a
circular segment starting at 80% of the yield strength is employed to smooth the bi-linear
material. The nonlinear pushover analysis using

dispBeamColumn

elements with

fibre

sections avoids the non-convergence issue by utilizing smooth steel materials and
discretized concretefibresections.
The observations and practical experiences are summarized following numerical case
studies. It was found that the DSA-MOOA and the DSA-S can achieve the same solution.
Yet, while the DSA-MOOA can theoretically prove its convergence, the DSA-S cannot.
When convergence speeds were compared, it was found that the DSA-S only needed
about 20% of limit-state function calls required by the DSA-MOOA. Thus, the DSA-S

112

can find an optimal design more effectively. However, if this algorithm fails in the
convergence procedure, the user must use the more reliable DSA-MOOA.
It was observed that the linear case required only 15% of limit-state function calls used
by nonlinear analyses. Nonlinear analyses produced more "exact" optimal designs, but
the linear case was more effective while also producing acceptable results. It is thus
suggested that the linear analysis be conducted first to find an optimal design. If a more
"exact" design is then required, the nonlinear analysis can begin from the results of the
linear analysis.
Some of the deterministic constraints were observed to be inactive in the optimization
process. Removing them can speed up the optimization procedure and save about 20-40%
of the original computational time. If the user finds that some inactive constraints are
violated in the analysis, the user can stop the analysis and add these constraints back into
the optimization again to make sure the solutions are correct.
Finally due to the use of the Polak-He algorithm, which only has linear convergence
properties, it is necessary to scale all involved functions properly (including the objective
function, deterministic constrains, and reliability constraints) to approximately the same
order 10= 1.0. The benefit of this scaling is apparent in the optimization loop, since only
10% of the original computational time is used. Another scaling is also suggested,
namely using the reliability index instead of the failure probability as the auxiliary
variable in the analysis. This is because the failure probability is too small and does not
appear in the same order of magnitude as the design variables. In summary, the
convergence in the RBDO can be accelerated by properly scaling the involved functions
and by using the reliability index instead of the failure probability.

8.2

F u r t h e r

Studies

A real-world structure is actually a general system reliability problem. Der Kiureghian

and Polak (1998)firstattempted to deal with the series structural system, and the RBDO
for series reliability system were finally solved in Royset et al. (2001a, 2001b, 2002, &
113

2004a). Based on the experience obtained in this thesis, the implementation of series
system problems in the reliability constraints can be achieved in the future work.
The analysis in this work is limited to static pushover finite element analysis. When
considering dynamic finite element analysis, time-variant reliability analysis must be
used to evaluate the failure probability. An available time-variant reliability analysis
method is the mean out-crossing reliability analysis. Furthermore, cyclic loading may
cause the degradation of the structural response. The application of RBDO to such
problems represents an important challenge for further work.
The definitions of initial costs and future costs are here made in terms of structural
volume or weight. More detailed cost computations are desirable. For instance, it is of
interest to include the present cost of future events. Such realistic considerations could be
an interesting future study.
In this thesis, we noticed the importance of proper scaling, which speeds up the
convergence procedure and avoids the nonconvergence problem. However, we only
scaled the values of the involved functions at the beginning of the analysis. In addition,
we did not know how to scale the gradient properly. It is thus suggested to develop an
automatic scaling scheme to scale both the value and the gradient, and scaling them at
each optimization step in future work.

114

Bibliography
Agarwal, H. and Renaud, J. E. (2004). "Decoupled methodology for probabilistic design
optimization." Proceedings of

th

ASCE Joint Specialty Conference on Probabilistic

Mechanics and Structural Reliability, July, Albuquerque, New Mexico.

Agarwal, H., Renaud, J. E., and Mack, J. D. (2003). "A decomposition approach for
reliability-based multidisciplinary design optimization." Proceeding of the
AIAA/ASME/ASCE/AHS/ASC

44

th

Structures, Structural Dynamics and Materials (SDM)

Conference, Norfolk, VA.

Benjamin, J. R. and Cornell, C. A. (1970). Probability, statistics, and decision for civil
engineers. McGraw-Hill, New York.

Bjerager, P. and Krenk, S. (1989). "Parameter sensitivity infirstorder reliability theory."

Journal of Engineering Mechanics, 115(7), 1577-1582.

Breitung, K. (1984). "Asymptotic approximation for multinomial integrals." J.

Engineering Mechanics, 110(3), 357-366.

Cement Association of Canada. (1995). Canadian Concrete Design Handbook. 2

nd

Edition, third printing November 2001.

Chen, X., Hasselman, T., and Neill, D. (1997). "Reliability-based structural design
optimization for practical applications." American Institute of Aeronautics and
Astronautics, AIAA-97-1403, 2724-2732.

Deitel, H. M. and Deitel, P. J. (1998). C++ How to program. Prentice Hall, Inc., Upper
Saddle River, NJ.
Der Kiureghian, A., Zhang, Y. and Li, CC. (1994). "Inverse reliability problem." Journal
of Engineering Mechanics, ASCE 120:1154-9.

Der Kiureghian, A. and Polak, E. (1998). "Reliability-based optimal design: A decoupled

approach." Reliability and Optimization of Structural Systems, A.S. Nowak (Ed.), Book
Crafters, Chelsea, Michigan.
115

Ditlevsen, O. and Madsen, H. (1996). Structural reliability methods. Wiley, New York,
New York.
Du, X. and Chen, W. (2002). "Sequential optimization and reliability assessment method
for efficient probabilistic design." ASME Design Engineering Technical Conference, 28

th

Design Automation Conference. September.

Eldred, M. S., Giunta, A. A., Wojtkiewicz, S. F., and Trucano, T. G. (2002).

"Formulations for surrogate-based optimization under uncertainty." Proceedings of the
9

th

AIAA/ISSMO Symposium on Multidisciplinary

Analysis and Optimization, Paper

AIAA-2002-5585, Atlanta, Georgia.

Enevoldsen, I. and Sorensen, J. (1994). "Reliability-based optimization in structural
engineering." Structural Safety, 15(3), 169-196.

Gasser, M. and Schueller, G. (1998). "Some basic principles in reliability-based optimization (RBO) of structures and mechanical components." Stochastic programming
methods and technical applications, K. Marti and P. Kail (Eds.), Lecture Notes in

Economics and Mathematical Systems 458, Springer-Verlag, Berlin, Germany.

Gill, P.E., Hammarling, S.J., Murray, W., Saunders, M.A. and Wright, M.H. (1986).
User's Guide to LSSOL. A Fortran Package for Constrained Linear Least-Squares and
Convex Quadratic Programming. Technical Report SOL 86-1. Califorlia: Stanford

Optimization Laboratory, Stanford.

Gill, P., Murray, W., Saunders, M., and Wright, M. (1998). User's guide to NPSOL 5.0:
A Fortran package for

nonlinear programming. Report No. SOL-86-1, System

Optimization Laboratory, Stanford University, Stanford, California.

Gobbo, B. (1999). "Calling Fortran routines from a C++ program in Unix."
http://pccosrvl.cern.ch/compass/software/offline/software/fandc.html.

Hasofer, A. M. and Lind, N. C. (1974). "Exact and invariant second-moment code

format." Journal of Engineering Mechanics, 100( 1), 111-121.

116

Haukaas, T. and Der Kiureghian, A. (2004). Finite Element Reliability and Sensitivity
Methods for Performance-Based Engineering. Report No. PEER 2003/14. California:

Pacific Earthquake Engineering Research (PEER) Center, University of California,

Berkeley.
Haukaas, T. and Der Kiureghian, A. (2005). "Parameter sensitivity and importance
measures in nonlinearfiniteelement reliability analysis." Accepted for publication, ASCE
Journal of Engineering Mechanics.

Hohenbichler, M. and Rachwitz, R. (1986). "Sensitivity and importance measures in

structural reliability." Civil engineering systems, 3, 203-209.
Itoh, Y. and Liu, C. (1999). "Multiobjective optimization of bridge deck maintenance."
Case Studies in Optimal Design and Maintenance Planning if Civil Infrastructure
Systems, D.M. Frangopol (Ed.), ASCE, Reston, Virginia.

Kirjner-Neto, C , Polak, E., and Der Kiureghian, A. (1998). "An outer approximations
approach to reliability-based optimal design of structures." J. Optimization Theory and
Application, 98(1), 1-17.

Kuschel, N. and Rackwitz, R. (2000). "A new approach for structural optimization of
series system." Proceedings

8th Intern. Conf. On Applications

of Statistics and

Probability (ICASP) in Civil Engineering Reliability and Risk Analysis, R.E. Melchers

and M.G. Stewart (Eds.), Sydney, Australia.

Li, H. and Foschi, R.O. 1998. "An inverse reliability method and its application."
Structural Safety. 20 (3) 257-270.

Liu, P.-L. and Der Kiureghian, A. (1986). "Multivariate distribution modes with
prescribed marginals and covariances." Probabilistic engineering mechanics, 1(2), 105112.
Madsen, H. and Friis Hansen, P. (1992). "A comparison of some algorithms for
reliability-based structural optimization and sensitivity analysis." Reliability and

117

Optimization of Structural Systems, Proceedings IFIP WG 7.5, R. Rackwitz and P. Thoft-

Christensen (Eds.), Springer-Verlag, Berlin, Germany.

Matlab Mathworks, Inc. (1999). Matlab reference manual, Version 5.3, Release 11.
Math-Works, Inc., Natick, Massachusetts.
Mazzoni, S., McKenna, F., Fenves, G.L. & Scott, M.H. (2005). "OpenSees command
language manual." http://opensees.berkeley.edu/. Pacific Earthquake Engineering
Research Center, University of California, Berkeley, CA.
McKenna, F., Fenves, G.L. & Scott, M.H. (2004). "OpenSees: Open System for
Earthquake Engineering Simulation." http://opensees.berkeley.edu/. Pacific Earthquake
Engineering Research Center, University of California, Berkeley, CA.
Polak, E. (1997). Optimization. Algorithms and consistent approximations. SpringerVerlag, New York, New York.
Rackwitz, R. and Fiessler, B. (1978). "Structural reliability under combined load
sequences." Computers and structures, 9, 489-494.
Royset, J., Der Kiureghian, A., and Polak, E. (2001a). "Reliability-based optimal design
of series structural systems."/. Engineering Mechanics, 127(6), 607-614.
Royset, J., Der Kiureghian, A., and Polak, E. (2001b). "Reliability-based optimal
structural design by the decoupling approach." J. Reliability Engineering and System
Safety, 73(3), 213-221.
Royset, J.O., Der Kiureghian, A. & Polak, E. (2002). Reliability-based Design
Optimization of Series Structural Systems. Roport No. UCB/SEMM-2002/15. California:

Department of Civil and Environment Engineering, University of California, Berkeley.

Royset, J.O., Der Kiureghian, A. & Polak, E. (2004a). "Optimal Design with
Probabilistic Objective and Constraints." Journal of Engineering Mechanics, submitted.
Royset, J.O. and Polak, E. (2004b). "Reliability-based optimal design using sample
average approximations." Probabilistic engineering mechanics, 19 (2004), 331-343.

118

Schittkowski, K. (1985). User's guide to nonlinear programming code, handbook to

optimization program package NLPQL. University of Stuttgart, Stuttgart, Germany.

Sexsmith, R. G. (1983). "Bridge risk assessment and protective design for ship collision."
IABSE Colloquium Copenhagen 1983 - Ship Collision with Bridges and Offshore
Structures, Preliminary Report, V42, 425-433, Copenhagen, Denmark

Thampan, C. and Krishnamoorthy, C. (2001). "System reliability-based structural

configuration optimization of trusses." J. Structural Engineering, 127(8), 947-955.
Torczon, V. and Trosset, M. (1998). "Using approximations to accelerate engineering
design optimization." Proceedings of the 7th AIAA/USAF/NASA/ISSMO

Symp. on

Multidisciplinary Analysis and Optimization, AIAA Paper 98-4800, St. Louis, Missouri.

Wang, L., Kodiyalam, S. (2002). "An efficient method for probabilistic and robust design
with non-normal distribution." Proceeding of the 43

rd

AIAA/ASME/ASCE/AHS/ASC

Structures, Structural Dynamics, and Materials Conference. April.

Welch, B. B. (2000). Practical programming in Tel and Tk. Prentice Hall, Inc., Upper
Saddle River, New Jersey, 3 edition.
rd

Zhang, Y. and Der Kiureghian, A. (1997). Finite element reliability methods for inelastic
structures. Report No. UCB/SEMM-97/05, Dept. of Civil

and Environmental

Engineering, University of California, Berkeley, Berkeley, California.

119

Appendix A: Detailed Software Implementation

This appendix contains detailed software implementations of the reliability-based design

optimization (RBDO). These implementations include ways to call Fortran routines from
C++, methods of building LSSOL.LIB, and prerequisites to extending OpenSees with the
RBDO capabilities.

Al:

Calling Fortran Routines from C++

In the Polak-He algorithm, a Fortran 77 program (LSSOL) is used to solve a suboptimization problem and to find the search direction. Therefore, a technique to call
Fortran routines from OpenSees (in C++) is required. This section introduces several
methods to implement this mix-program technique, focusing especially on how to pass
and return variables and arrays between C++ and Fortran.
1. The extern "C" directive is used to declare the external Fortran subroutine LSSOL in
C++.
#ifdef_WIN32
extern "C" void LSSOL(int *m, double *c, double *A, double *obj, double *x);
#else
extern "C" void lssol_( int *m, double *c, double *A, double *obj, double *x);
#endif
where m is an input variable, c is an input one-dimensional array, and A is an input
two-dimensional array. They are passed from C++ to Fortran, obj is a returned
variable and x is a returned one-dimensional array. Both of them are returned from
Fortran to C++. Note that variables and arrays listed above are only examples and are
not complete.
2. When C++ calls Fortran, the reference to Fortran symbols is specified in lowercase
letters, since C++ is a case sensitive language, but Fortran is not.
120

3. C++ passes the variables by value, while Fortran passes them by reference. It is
necessary to specify in the C++ that the Fortran subroutines expect call-by-reference
arguments using the address-of operator & (Gobbo, 1999). An example of passing the
variable m = 10 from C++ to Fortran and returning the variable obj - 20.0 from
Fortran to C++ is below:
// Define the variable m and obj
m=10;
obj = 0.0;
// Call LSSOL
#ifdef_WIN32
LSSOL(&m, &obj);
#else
lssol_(&m, &obj);
#endif
As a returned variable, obj = 20.0 is then used directly in C++.
4. C++ passes arrays using pointers, while Fortran passes them using references. In
addition, C++ stores arrays in a row-major order, whereas Fortran stores arrays in a
column-major order. Finally, the lower bound for C++ is 0, but for Fortran it is 1
(Gobbo, 1999). For instance, given an array "fun," the Fortran array element fun(l,l)
is the same as the C++ array element fun[0][0]; the Fortran array element fun(6,8)
corresponds to the C++ array element fun[7][5]. An example of passing onedimensional array c = [l.l 1.2] and two-dimensional array A

"2.1 2.3"
from
2.2 2.4_

C++ to Fortran and returning one-dimensional array x = [3.1 3.2] from Fortran to
C++ is below:
// Preparing input data for LSSOL
c[0]= 1.1;
c[l]= 1.2;
A[0] = 2.1;
A[l] = 2.2;
A[2] = 2.3;
x[0] = 0.0;
x[l] = 0.0;
// Call LSSOL
#ifdef WIN32
121

A[3] = 2.4;

LSSOLfe, A, x);
#else
lssol_(c, A, x);
#endif
As a returned array, x[0] = 3.1 and x[l] = 3.2 is then used directly in C++.

All

Building LSSOL.LIB

LSSOL is complied using Intel(R) visual Fortran compiler for Windows, standard
edition,

which

is

freely

available

from

http://wwwdntel.com/software/products/compilers/dow

the

Intel

website
Compiling

LSSOL requires the following procedure:

1. Download and install "Microsoft Visual Studio.NET" and 'Tntel(R) Software
Development Tools."
2. Run "Microsoft Visual Studio.NET." Set the project type, template, and name of
project workspace in the following way:
File * New > Project...
Project Types: Intel (R) Fortran Projects
Templates: Static library
Name:
LSSOL
3. Set the library wizard when you see "Welcome to the Fortran static library wizard."
Make sure that the option "Prevent the insertion of linker directives for defaults
libraries" is not selected.
Library Settings: Additional features
Prevent the insertion of linker directives for defaults libraries
4. Copy LOSSL files to folder ".ALSSOL," and add LSSOLfilesinto the workspace.
Solution Explorer - LSSOL
LSSOL Source Files Add Add Existing Item ... Add all files
5. Set Fortran libraries.
122

LSSOL * Properties
> Fortran
> Libraries:
Use Common Windows Libraries:

Yes

Use Portlib Library:

Yes

Disable Default Library Search Rules:

No

6. Compile LSSOL through "Build -+ Build LSSOL." The result is a library

LSSOL.LIB (Size: 1,337 KB), which is saved in the folder "..\LSSOL\debug."
7. Backup four LIB files below from 'Tntel(R) Software Development Tools." The
following libraries are required when compiling OpenSees.
IFCONSOL.LIB (Size:
IFWIN.LIB

A3:

7 KB)

(Size:

30 KB)

LIBIFCORE.LIB (Size:

955 KB)

LIBIFPORT.LIB (Size:

412 KB)

Extending OpenSees with RBDO Capacity

Install Tcl/Tk (which is needed to run OpenSees)

1. Uninstall any previous versions of Tel
2. Download the installation file for Tcl/Tk
3. Run the installation by running the "exe" file downloaded in step 2
4. Install Tel in the folder C:\Program FilesYTcl
5. Restart the computer

Install CVS software (required to download OpenSees from the CVS repository)
1. Downloadfilescvs-1 -11 -5.zip and cvslOgin.bat
2. Unzip cvs-1-1 l-5.zip and run the installation file cvs-1.1 L5.exe
3. Restart the computer

123

Download OpenSees code from the "official" CVS repository in Berkeley

1. Open a DOS window (e.g.; Start > Programs > Accessories > Command Prompt)
2. "cd" into the folder where you have put the "cvslogin.bat" file
3. Execute "cvslogin" command
4. Note that steps 2 and 3 above can be replaced by issuing the following commands:
set CVS_RSH=ssh
set CVSROOT=:pserver:anonymous@opensees.berkeley.edu:/usr/local/cvs
cvs login
5. When prompted, provide the password "anonymous"
6. Go to the directory where you want to put the OpenSees code
7. Give the command "cvs checkout OpenSees"
Later, when updating the code with the most recent changes in the CVS repository, you
can follow steps 1 to 6 and then give the command "cvs -q update -d" (-q is used to
suppress output, -d is used to check out any new directories). It may be a good idea to do
this "directory by directory" in the SRC directory. The command "cvs diff' is used to list
differences between local files and the CVS repository files. When doing updates, the
following abbreviations are used to identify the actions taken for each file:
M ~ local copy has been modified
P merged changes on server with the local copy
C conflict with what's on server and the local copy
U check a new file that is not part of local copy

Compile the "official" OpenSees version

1. Make sure the "include path" for tcl.h is correct in the projects damage, database,
domain, element, material, recorder, reliability, and openSees by doing the

following:
a) Right-click on the project and choose "Settings > C/C++ > Preprocessor"
124

b) Select "Settings for: All Configurations"

c) In "Additional include directories" the last statement should be "c:\Program
Files\tcl\include"
2. Include tcl84.lib in the openSees project by doing the following:
a) Right-click on the project and choose "Settings > Link > Input"
b) Select "Settings for: All Configurations"
c) In "Additional library paths," include "c:\Program Files\tcl\lib" (make sure to
have a comma between the different paths)
Press F7 to compile. If many error messages appear, try to press F7 again to "clean
up."

Add new and/or improved files from the UBC team

Put all files listed in Tables A.la, A.lb, and A.lc into their respective directories. For
new classes, remember to include the files in the appropriate project according to the
"location offile"provided in the Tables A.la, A.lb, A.lc,.

How to identify the difference between local files and the "official" version at
Berkeley

There are two ways of identifying the difference between the files that have been
modified by the "UBC team" and the official Berkeley files:
1. Download the files, include them in local OpenSees version, and use the "diff"
feature of CVS to see the differences. (Give the command "cvs diff in the relevant
directory.)
2. Search for the text string "UBC Team." All UBC team modifications are marked with
this stamp.

125

Include the F o r t r a n l i b r a r y " L S S O L . L I B "

1. Copy the following files into the directory: Win32/lib: I F C O N S O L . L I B , I F W I N . L I B ,
L I B I F C O R E . L I B , and L I B I F P O R T . L I B .
2. Copy the file L S S O L . L I B into the Win32/lib/debug and Win32/lib/release directories.
3. A d d the L S S O L library to the project settings by doing the following:
a) Right-click on the "opensees" project and choose "Settings > L i n k > General"
b) Select "Settings for: A l l Configurations"
c) In "Object/library modules" add the filename L S S O L . L I B

Table A . l a N e w and modified classes for extending R B D O (Classes that do not exist in
the " o f f i c i a l " version are marked with *)
Project

L o c a t i o n of file

Files
classTags.h

OpenSees Source
Header
Reliability analysis/types

commands.cpp
commands.h
DSA_MOOAOptimizationAnalysis.cpp*
DSAMOOAOptimizationAnalysis.h*
DSA_SOptimizationAnalysis.cpp*
DSA_SOptimizationAnalysis.h*

domain/components ReliabilityDomain.cpp
ReliabilityDomain.h
ConstraintFunction.cpp*
ConstraintFunction.h*
CostFunction.cpp*
CostFunction.h*
DesignVariable.cpp*
DesignVariable.h*
DesignVariablePositioner.cpp*
DesignVariablePositioner.h*
ObjectiveFunction.cpp*
ObjectiveFunction.h*
126

Table A. lb New and modified classes for extending RBDO (continued) (Classes that do
not exist in the "official" version are marked with *)
Location of file
Files
Project
Reliability analysis/designPoint NonlinSinglelneqOpt.cpp*
NonlinSinglelneqOpt.h*
PolakHeNonlinSinglelneqOpt.cpp*
PolakHeNonlinSinglelneqOpt.h*
NonlinMultilneqOpt.cpp*
NonlinMultilneqOpt.h*
PolakHeNonlinMultilneqOpt.cpp*
PolakHeNonlinMultilneqOpt.h*
LinMultilneqOpt.cpp*
LinMultilneqOpt.h*
LSSOLLinMultilneqOpt.cpp*
LSSOLLinMultiIneqOpt.h*
analysis/gFunction GFunEvaluator.cpp
GFunEvaluator.h
OpenSeesGFunEvaluator.cpp
OpenSeesGFunEvaluator.h
analysis/sensitivity GradGEvaluator.fi
FiniteDifferenceGradGEvaluator.cpp
FiniteDifferenceGradGEvaluator.h
OpenSeesGradGEvaluator.cpp
OpenSeesGradGEvaluator.h
FEsensitivity
SensitivityAlgorithm.cpp
TclReliabilityBuilder.cpp
tcl
Element
Information.cpp
TclElementCommands. cpp
dispBeamColumn DispBeamColumn2d.cpp
beamWithHinges
BeamWithHinges2d_bh.cpp*
BeamWithHinges2d_bh.h*
TclBeamWithHingesBuilder.cpp
elasticBeamColumn ElasticBeam2d_bh.cpp*
ElasticBeam2d_bh.h*
TclElasticBeamCommand.cpp
127

Table A.lc New and modified classes for extending RBDO (continued) (Classes that do
not exist in the "official" version are marked with *)
Location of file
Project
Material uniaxial

section

Files
SteelO lepsy.cpp*
Steel01_epsy.h*
SmoothSteelOlepsy.cpp*
SmoothSteel01_epsy.h*
ElasticPPMaterial_Fy.cpp*
ElasticPPMaterial_Fy.h*
SmoothElasticPPMaterial_Fy.cpp*
SmoothElasticPPMaterialFy.h*
TclModelBuilderUniaxialMaterialCommand.cpp
FiberSection2d.cpp
RCFiberSection2d.cpp*
RCFiberSection2d.h*
TclModelBuilderSectionCommand.cpp

128

Appendix B: User's Guide to Optimization Analysis

This appendix contains the user guide to new implementations of the reliability-based
design optimization (RBDO). It is a complement to the user guide to reliability and
sensitivity analyses in Haukaas and Der Kiureghian (2004). The optimization commands
used in this section have the same format as in Haukaas and Der Kiureghian (2004). An
example of a command is:
commandName argl? arg2? arg3? <arg4? ...>

A question mark after an argument indicates that an integer or a floating-point number

should be provided; otherwise, a character string is given. Optional arguments are
enclosed in angular brackets (Haukaas & Der Kiureghian, 2004). Note that all mentioned
tasks (A1-A3, B1-B3, C1-C3, and D1-D2) in this appendix are described in detail in
Chapter 6.

Bl:

RBDO Modeling

This section describes how to define design variables and functions involved in the
RBDO analysis. The object mapping design variables into the finite element domain is
also introduced.
A design variable object defines design variables by giving their start points through
the following command:
design Variable tag? startPt?

The tag argument indicates the identification number of the design variable. These
objects must be ordered in a consecutive and uninterrupted manner. The startPt
argument allows the user to specify a value for the design variable to be used as the start
point in the search for the design point (Haukaas & Der Kiureghian, 2004).
129

A design variable positioner object is used to map the design variables into structural
properties in thefiniteelement model through the following command:
designVariablePositioner tag? -dvNum dvNum? (...parameter identification...)

The tag argument indicates the identification number of the design variable positioner.
The dvNum argument indicates the identification number of the pre-defined design
variable. The parameter identification alternatives in the command are exactly the same
as in the random variable positioner command in Haukaas and Der Kiureghian (2004).
A constraint function object defines constraint functions using user-defined
expressions through the following command:
constraintFunction tag? "expression"

The tag argument indicates the identification number of the constraint function. The
expression must be enclosed in double quotes and can be any analytical expression that

can be evaluated by the Tcl interpreter (Welch, 2000). This function may be expressed by
various quantities including random variables, design variables, structural response
quantities from an OpenSees finite element analysis, and parameters defined in the Tcl
interpreter (Haukaas & Der Kiureghian, 2004). The syntax used in this command is the
same as that in the performance function command in Haukaas and Der Kiureghian
(2004). An example of syntax for design variables is {d_l}, which means thefirstdesign
variable.
A cost function object defines cost functions using user-defined expressions through
the following command:
costFunction tag? "expression"

The tag argument indicates the identification number of the cost function. The
expression has the same properties as that in the constraint function command. However,

only design variables and parameters defined in the Tcl interpreter are employed as
quantities in this expression.

130

An objective function object defines objective functions through combing the

previously defined cost functions. Currently, the standard type of objective function is
available for this object in OpenSees. The corresponding command reads:
objectiveFunction tag? -type standard -costFunctions costNuml? costNum2?

The tag argument indicates the identification number of the objective function. A
standard objective function is created in the following way: objective function = 1 cost
st

function + failure probability x 2 cost function, where the failure probability is passed
nd

from the ReliabiltyDomain each time the objective function object is called.

B2:

Analysis Tools

Before a RBDO analysis is executed the user must create an aggregation of necessary
analysis components or tools. Which analysis components are needed depends on the
analysis type. The order in which the tools are provided is of importance, since some
tools make use of other tools. The user will be notified by an error message if
dependencies are violated (Haukaas & Der Kiureghian, 2004).
A nonlinSinglelneqOpt object is created to be responsible for solving nonlinear single
inequality constrained optimization problems. This object promises to solve the tasks Bl
and D2. The corresponding command reads:
nonlinSinglelneqOpt PolakHe -alpha argl? -beta arg2? -gamma arg3? -delta
arg4?

This type of optimization problem is solved by the Polak-He algorithm. In the PolakHe algorithm, argl denotes the parameter alpha (default = 0.5), arg2 denotes the
parameter beta (default = 0.8), arg3 denotes the parameter gamma (default = 2.0), and
arg4 denotes the parameter delta (default = 1.0).

A linMultilneqOpt

object is responsible for solving linear multi-inequality

constrained optimization problems. The corresponding command reads:

131

linMultilneqOpt LSSOL

In order to find the search direction for tasks B3 and D l , a quadratic sub-optimization
problem with linear constraints must be solved. This sub-optimization problem is fulfilled
by the linMultilneqOpt object. Currently, a Fortran 77 program LSSOL is called to solve
this problem and is the available implementation of this object in OpenSees.
A nonlinMultilneqOpt object is created to be responsible for solving nonlinear multiinequality constrained optimization problems. This object promises to solve tasks B3 and
D l . The corresponding command reads:
nonlinMultilneqOpt

PolakHe -alpha argl? -beta arg2? -gamma arg3? -delta

arg4?

This type of optimization problem is solved using the Polak-He algorithm. A quadratic
sub-optimization problem with linear constraints must be solved in this object to find the
search direction. Therefore, a linMultilneqOpt object must be created before the
nonlinMultilneqOpt object can be instantiated, argl to arg4 are user-defined parameters
used in the Polak-He algorithm and have the same definition as the parameters in the
nonlinSinglelneqOpt object.

B3:

Analysis Execution and Results

Two analysis types are available in the optimization module of OpenSees. This section
describes the corresponding commands to execute them. Required analysis tools must be
specified prior to using any of these commands. During the course of a RBDO analysis,
status information may be printed to a file or to a computer monitor. The complete results
from a successful analysis are printed to an output file whose name is specified by the
user, as show below (Haukaas & Der Kiureghian, 2004).
A DSA-MOOA analysis object is the top-level of the DSA-MOOA approach and is
responsible for obtaining the optimal design by orchestrating tasks A l to A3. This object
is executed using the following command:
132

runDSAMOOAOptimizationAnalysis outputfilename -betaO argl? -targetCost

arg2? -maxIterOuter

arg3? -maxlterlnner

arg4? -numSimulation

arg5? -

targetCOV arg6?

The order of arguments is arbitrary, argl denotes the lower bound of failure
probability (default = 3.0), arg2 denotes the target total expect failure cost, arg3 denotes
the maximum number of iterations on top level (A1-A3), and arg4 denotes the maximum
number of iteration in task B3. arg5 and arg6 are input parameters necessary for
importance sampling in task A2. arg5 denotes the maximum number of simulations
(default = 10 ), while arg6 denotes the target coefficient of variation (default = 2%). The
6

nonlinSinglelneqOpt, HnMultilneqOpt, and nonlinMultilneqOpt objects must be created

before the DSA_MOOAAnalysis object is created. The results in the output file are selfexplanatory, including the optimal design as well as reliability index and total expected
failure cost.
A DSA-S analysis object is the top level of the DSA-S approach and is responsible for
obtaining the optimal design by orchestrating tasks CI to C3. This object is executed
using the following command:
runDSASOptimizationAnalysis outputfilename -betaO argl? -targetCost arg2? maxIterOuter arg3? -maxlterlnner arg4? -numSimulation

arg5? -targetCOV

arg6?

The input data and the results of this analysis type are exactly same as those in the
DSA-MOOA analysis, expect for arg3 and arg4. arg3 denotes the maximum number of
iterations of top level (C1-C3), while arg4 denotes the maximum number of iterations in
taskDl.

133