Professional Documents
Culture Documents
Anti-Optimization of
Structures Under Uncertainty
Isaac Elishakoff
Florida Atlantic University, USA
Makoto Ohsaki
Kyoto University, Japan
Published by
Imperial College Press
57 Shelton Street
Covent Garden
London WC2H 9HE
Distributed by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN-13 978-1-84816-477-2
ISBN-10 1-84816-477-7
Printed in Singapore.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
v
This page intentionally left blank
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Preface
vii
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
are no two structures with identical material and geometric properties, boundary
conditions, or loading and, therefore, uncertainty analysis is called for.
In reality, this deterministic versus non-deterministic attitude is a fallacy. To
justify this undiplomatic statement, we need deeper insight into the essence of de-
terministic analysis. In it, after careful evaluation of displacements, strains, and
stresses (often with many significant digits since computer outputs consist of
numbers), the engineers compare the latter against the allowable stress; i.e., the ex-
perimentally measured yield- or ultimate-level multiplied by some fudge coefficient
called the safety factor (as in better safe than sorry). Use of this factor in text-
books on engineering mechanics is invariably motivated by insufficient knowledge
(dont we always have to make analytic assumptions of various sorts?), or by un-
certainty (yes, deterministic analysts do pay lip service to the property they claim
to neglect!) in loads, geometric characteristics, or scatter in material properties.
Thus, deterministic analysis, strictly speaking, is not as deterministic as ad-
vertised. Within it, uncertainty is introduced as dessert or, as in Bolotins words
(Bolotin, 1961), via the back door.
By contrast, in non-deterministic design, uncertainty figures as a legitimate
ingredient throughout the whole design process, specifically during the thinking,
analysis and design stages.
One may conclude, therefore, that non-deterministic analysis is more honest than
its so-called deterministic counterpart as it places its cards on the table from the very
beginning. As for the latter, in view of the above reasoning, it would be appropriate
to dub it pseudo-deterministic rather than deterministic. If experience enables
us to establish the precise value of a safety factor, then the pseudo-deterministic
unknowns are nothing less than the lumped, integral equivalent of the inherently
present uncertainty in engineering problems.
Naturally, the advent of the safety factor was at the time a major breakthrough
in design. Despite its obvious arbitrariness, it allowed us to put up a wall against
failure. At the same time, this dogmatism implies that if the actual stress turns out
to be below the allowable level, then there can be no failure, and the structure will
be safe throughout its exploitation.
But how can we establish the safety factor, except by experience or trial-and-
error (one could sarcastically ask at this stage: is this a trial or an error?). According
to Norton (2000), choosing the safety factor is often a confusing proposition for the
beginning engineer, and not only for the beginning one, we may add. According to
Bruhn (1975),
Preface ix
As we observe, even with the manner of explanation of the safety factor, this
concept is quite critical within the semi-deterministic framework for it utilizes ad-
jectives like arbitrary (Bruhn, 1975) and confusing (Norton, 2000).
Freudenthal (1968) emphasized:
Preface xi
This worst-case design was never adopted by the nuclear industry because of
its extreme conservatism, as Drenick informed one of us (I.E.). He, accordingly,
became very pessimistic about the likelihood of the engineering profession ever
adopting it. He was however very encouraging, and so kind as to write the foreword
to the book by Ben-Haim and Elishakoff (1990) in which he stated perhaps with
some measure of exaggeration: Their approach is novel and highly welcome. In my
opinion, it is inevitable that it, and its extensions, will dominate the future practice
of engineering.
In consequence of the above-mentioned book, it was realized by one of us (I.E.)
that in order to make this worst-case scenario research practical, it needed an infu-
sion of new blood. (It was realized that unknown-but-bounded uncertainty analysis
is a defacto as anti-optimization process, the reverse of searching for the best so-
lution.) It became also apparent that it is impractical to content oneself with the
least favorable static or dynamic response, or with the worst possible buckling load.
Rather a structure should be so designed as to minimize the first and maximize the
second in other words, adherence to the time-honored precept of making the best
out of the worst. It is gratifying to read that this sentiment is now shared by other
investigators, for example, Ben-Tal and Nemirovski (2002):
under discussion (Einstein, here too, comes handy: you do not really understand
something unless you can explain it to your grandmother).
Indeed, would not it be an engineering sin, as it were, not to improve the worst
scenario and leave it intact especially if one has to save expenditure? Only very
rich companies, or those in positions of exceptional responsibility, can afford exclu-
sive reliance on anti-optimization.
One argument against optimization is that if it is applied to a specific loading
condition, the result may prove unsatisfactory under a different condition. In such a
case, it suffices to add the latter and repeat the procedure for the new circumstances.
Alternatively, one can introduce uncertainty and assign bounds so as to obtain the
optimal solution under the worst-case scenario.
Another criticism concerns optimization against nonlinear buckling. As ex-
plained in detail in Chap. 5, optimization against buckling does not always yield
imperfection-sensitive structures, the resulting sensitivity being in fact lower than
that of a non-optimal design (Ohsaki, 2002c). However, the buckling load of a
perfect system is drastically increased by optimization, and its counterpart for an
imperfect system is little affected by sensitivity changes.
The combined optimistic-pessimistic approach can be characterized as follows:
optimists build ships; pessimists build lifeboats. Both are needed for safe sailing.
Recall the overly optimistic view on the Titanic, which was claimed to be so robust
that even God couldnt sink it! As a result, an insufficient number of lifeboats was
on board during its maiden voyage. This complacency cost over 1,500 lives.
We advocate a combination of healthy doses of optimism (represented by opti-
mization) and pessimism (anti-optimization).
The above discussion gave rise to the notion that uncertainty analysis is re-
ducible to one of the three vertices of the uncertainty triangle (Elishakoff, 1990,
1998b), shown below. These three approaches do not represent three non-interesting
magistrata!
Although the first paper co-authored by one of us (Ben-Haim and Elishakoff,
1989b) was presented at two conferences in 1988 and appeared in 1989, we had been
able, on an earlier occasion (Elishakoff, 1983, p. 42), to compare the probabilistic
and worst-case designs on a simple example included on the Statistical Methods
in Elasticity course at the Technion - Israel Institute of Technology since 1973.
The monograph by Ben-Haim and Elishakoff (1990) did not correlate probabilistic
and convex modelings. Only later, in the paper by Elishakoff, Cai and Starnes
(1994a), was a direct comparison made between probabilistic and anti-optimization
techniques on the nonlinear boundary-value problem (see also its generalization by
Qiu, Ma and Wang (2006a)). It turned out that if the probabilistic information is
available and the required reliability is not excessively high, anti-optimization may
be conservative. However, if near-unity reliability is required, the two approaches
tend to yield close or coincident results.
It was also realized that these seemingly competing approaches are nevertheless
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Preface xiii
Theory of
Probability and
Randon Processes
Anti-Optimization with
Fuzzy Sets Set Theoretical Aproach
(Least Favorable Responses)
Still, we should be thankful that the innocent number , for example, is not
declared a random variable with mean value 3.14... and zero variance! One should
refrain from considering all of them as random variables unless they have sufficiently
large variability to justify the non-deterministic approach. Likewise, one should
abandon the idea (attractive for beginners) of treating all parameters as uncertain,
even if some of them have sufficiently small variability.
In conclusion, one has to maintain a delicate and healthy balance in modeling a
system. Bolotin (1969), for example, is against both underestimation and overesti-
mation of uncertainty analysis. We hope that this book will serve as a beacon for
practicing engineers and researchers, and help them reevaluate their thinking and
practices on uncertainty analysis.
The authors are indebted to the Japan Society for Promotion of Science (JSPS)
established by the late Emperor Showa (Hiro-Hito) in 1932. The JSPS, long before
the modern trends of globalization, made it possible for non-Japanese researchers
to spend time in the Land of the Rising Sun with their host counterparts, thereby
enabling both sides to elevate their horizons through collaboration. I.E. is thankful
to Kyoto University which provided the hospitable atmosphere for the joint research.
We record our appreciation to the co-authors of some joint papers that found
their way into our monograph. These are Professor Dan Givoli of the Department of
Aerospace Engineering, Technion - Israel Institute of Technology, for his kind agree-
ment for us to reproduce the shortened version of the paper Stress concentration
at nearly circular hole with uncertain irregularities, by D. Givoli and I. Elishakoff,
1992; Professor Nobuhiro Yoshikawa of the Institute of Industrial Science, Univer-
sity of Tokyo for his kind agreement for us to reproduce the contents of the paper
Worst case estimation of homology design by convex analysis, by N. Yoshikawa,
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Preface xv
Contents
Preface vii
1. Introduction 1
1.1 Probabilistic Analysis: Bad News . . . . . . . . . . . . . . . . . . 1
1.2 Probabilistic Analysis: Good News . . . . . . . . . . . . . . . . . 10
1.3 Convergence of Probability and Anti-Optimization . . . . . . . . 13
xvii
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
3.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3.2 A simple example . . . . . . . . . . . . . . . . . . . . . . 51
3.3.3 General procedure . . . . . . . . . . . . . . . . . . . . . . 52
3.4 Ellipsoidal Model . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.1 Definition of the ellipsoidal model . . . . . . . . . . . . . 57
3.4.2 Properties of the ellipsoidal model . . . . . . . . . . . . . 58
3.5 Anti-Optimization Problem . . . . . . . . . . . . . . . . . . . . . 61
3.6 Linearization by Sensitivity Analysis . . . . . . . . . . . . . . . 63
3.6.1 Roles of sensitivity analysis in anti-optimization . . . . . 63
3.6.2 Sensitivity analysis of static responses . . . . . . . . . . 64
3.6.3 Sensitivity analysis of free vibration . . . . . . . . . . . 67
3.6.4 Shape sensitivity analysis of trusses . . . . . . . . . . . . 68
3.7 Exact Reanalysis of Static Response . . . . . . . . . . . . . . . . 69
3.7.1 Overview of exact reanalysis . . . . . . . . . . . . . . . . 69
3.7.2 Mathematical formulation based on the inverse of the
modified matrix . . . . . . . . . . . . . . . . . . . . . . . 70
3.7.3 Mechanical formulation based on virtual load . . . . . . . 74
Contents xix
Contents xxi
Bibliography 343
Index 387
Author Index 391
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 1
Introduction
May the best yeve ever seen, Be the warst yell ever see. (Scottish
blessing)
It is not enough to do your best; you must know what to do, and then
do your best. (W. Edwards Deming)
Its only the man who can look at the same problem from many dif-
ferent aspects that will make a true leader. (Takao Fujisawa)
Apparently, the first study in which probabilistic methods were applied to structures
was conducted by Mayer (1926). Since then, probabilistic methods have developed
beyond the age of adolescence (Cornell, 1981).
Usually, probabilistic reliability studies involve assumptions on the probability
densities, whose knowledge regarding the relevant input quantities is central. Given
this, the probabilistic properties of the output quantities can be determined. As
Wentzel (1980) stresses, probabilistic methods are
1
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Introduction 3
methods? The reply to this question is best given in terms of the design of a
structure. In a deterministic design process, one requires
y (1.1)
where is an actual stress that is assumed to be positive in tensile state, and y is
an yield stress. Here, uncertainty enters the picture in the sense that we may not
know the loads precisely, or there may be imprecision in measuring the geometric
parameters of the cross-section, or we may have built an imperfect mechanical model
to describe the behavior of the structure. Accordingly, a required safety factor k req
is introduced, and Eq. (1.1) is replaced by
y
(1.2)
kreq
Once the structure is designed, one can introduce the actual safety factor
y
kact = (1.3)
max
where max is the maximum actual stress occurring in the structure. Thus the
design requirement can be formulated as kact kreq , or, in other words, the actual
safety factor should not be less than the required one.
Can one quantify that the actual safety factor of the uncertainty is not hidden,
but is directly introduced into the scene? Let us attempt to answer this question.
Let the random force p, acting on a tension-compression element (bar) with cross-
sectional area a, have a Weibull (1951) distribution with the following probability
distribution function " k #
p p0
FP (p) = 1 exp , k > 0, w > p0 , p p0 (1.4)
w p0
For p < p0 , FP (p) 0. Equation (1.4) is called the 3-parameter Weibull distribu-
tion, which is a generalization of the exponential distribution. It has been developed
originally to describe failure strength of metals. More recently it has been utilized
in connection with fracture, as well as life distribution of mechanical compounds.
For the special case p0 = 0, the distribution is called 2-parameter Weibull distri-
bution; in this case k = 1 yields the exponential distribution, whereas k = 2 is
associated with Rayleigh distribution. The shape parameter k < 1 is typical of
wearing phenomena, whereas k > 1 is typical of aging effects. The 3-parameter
Weibull distribution is useful in describing phenomena for which some minimum
value p0 exists for the random variable P , so that P takes on values greater than
or equal to p0 . In the following, we denote random variables by upper-case, and
lower-case notation is reserved for their possible values.
We are interested in the reliability of the structure, i.e., the probability of the
structure performing its intended mission satisfactorily. Such a performance is
identified with relationship holding Eq. (1.1) true. Thus, the reliability becomes:
R = Prob( y )
= Prob(P/a y )
(1.5)
= Prob(P y a)
= FP (y a)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Thus, we have
" k #
y a p 0
R = 1 exp (1.6)
w p0
How can we define the safety factor in the context of probabilistic design? One
natural way is to relate it to some characteristic load, say the average load. The
latter equals
E(P ) = p0 + (w p0 )(1 + 1/k) (1.7)
where ( ) is the Gamma function. The variance of the load is
V (P ) = (w p0 )[(1 + 2/k) 2 (1 + 1/k)] (1.8)
The central safety factor n is defined as the ratio of the yield stress to its average
counterpart E(P )/a:
y a y a
n= = (1.9)
E(P ) p0 + (w p0 )(1 + 1/k)
Let us design the structure probabilistically. Probabilistic design requires that
the reliability be not less than a codified value r:
Rr (1.10)
Thus, in view of Eq. (1.6), we obtain
" k #
y a p 0
1 exp r (1.11)
w p0
The design value of the cross-sectional area adesign is found from the equality R = r,
and reads
p0 + (w p0 )[ln 1/(1 r)]1/k
adesign = (1.12)
y
Substitution of adesign for a in Eq. (1.9) enables us to write the central safety factor
explicitly in terms of the codified required reliability r:
p0 + (w p0 )[ln 1/(1 r)]1/k
n= (1.13)
p0 + (w p0 )(1 + 1/k)
For the set of parameters w = 3p0 , k = 4, we have
1 + 2[ln 1/(1 r)]1/4
n= (1.14)
1 + 2(1.25)
Therefore, from Eq. (1.14), n = 1.2314 for r = 0.9; n = 1.3971 for r = 0.99;
n = 1.5082 for r = 0.999; n = 1.5942 for r = 0.9999; n = 1.6653 for r = 0.99999;
n = 1.7263 for r = 0.999999, etc. As we see, a probabilistic model allows us to
associate the required reliability directly with the safety factor.
Let us see now how a small error can affect the reliability calculations. Say that
the actual values are w1 , p1 and k1 , while those used in the analysis are w, p0 and
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Introduction 5
k. The actual reliability Ract and the actual probability of failure Pf,act are related
as
Pf,act = 1 Ract (1.15)
The actual reliability is given by Eq. (1.6) with the actual values substituted:
" k #
y a p 1 1
Ract = 1 exp (1.16)
w1 p 1
and the actual probability by
" k #
y a p 1 1
Pf,act = exp (1.17)
w1 p 1
If the design of the structure has been performed using the values w, p0 and
k, the appropriate value of the cross-sectional area is given by Eq. (1.12). To
calculate the actual probability of failure corresponding to the design value adesign ,
we substitute the expression adesign into Eq. (1.17) and have
" k #
y adesign p1 1
Pf,act = exp (1.18)
w1 p 1
or
" k1 #
p0 p1 + (w p0 )[ln 1/(1 r)]1/k
Pf,act = exp (1.19)
w1 p 1
Let us consider some particular cases. In the simplest case p1 = p0 , w = w1 ,
but k1 6= k,
" k1 /k #
1
Pf,act = exp ln (1.20)
1r
Since r is the required reliability, 1 r is recognized in Eq. (1.20) as the allowed
probability of failure Pf,all . Thus, Eq. (1.20) can be rewritten as:
" k1 /k #
1
Pf,act = exp ln (1.21)
Pf,all
0.00002
0.000015
Pf , act
0.00001
0.000005
0
0.9 0.95 1 1.05 1.1
k1 / k
Fig. 1.1 Actual probability of failure as a function of the ratio k1 /k (solid line); for k1 = k it
coincides with the required one (dotted line); for k1 /k > 1 it is less than the required one; for
k1 /k < 1 it may well exceed the allowed value, resulting in a detrimental state.
probability of failure is less than the allowable one. However, when k1 /k > 1, the
effect of a small error in evaluating k may be detrimental: If k1 /k = 0.95, the
actual probability is about five times the allowed one; if k1 /k = 0.93, the actual
probability is approximately ten times as large as the one which was permitted! We
conclude that the probability of failure is too sensitive a parameter to make do with
imprecise input characteristics, and that accurate determination of the probabilistic
characteristics of input must be an integral part of a rigorous probabilistic analysis.
Regarding the effect of a small deviation in the probability density on the relia-
bility estimate, we consider a bar with cross-sectional area a subjected to a load P
which is a random variable with the probability density fP (p), p being a possible
value of the load. The material of the bar is assumed to be perfectly elastic in com-
pression and has a yield stress in tension y . We also assume that the bar cannot
lose its stability, or undergo any other form of failure. To recapitulate, its reliability
is defined as the probability of the stress = P/a not exceeding the yield stress:
R = Prob( y ) (1.24)
Let us consider the situation in which the data is suggestive for the analyst to
assume the probability density of the load in the form of the log-normal variable
1 (ln p b)2
exp , for p > 0
fP (p) = pc 2 2c2 (1.25)
0, otherwise
where b and c characterize the probability density. The mean value E(P ) and the
variance V (P ) are expressed as
1 2
E(P ) = exp b + c
2 (1.26)
2 2
V (P ) = exp(2b + c )[exp(c ) 1]
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Introduction 7
since ln P is a normal variable with mean b and variance c2 . Let us assume now
that the actual probability density differs slightly from the true one and reads, for
p > 0:
() 1 (ln p b)2
fP (p) = exp {1 + sin[2(ln p b)]} (1.29)
pc 2 2c2
where is a constant belonging to an interval [1, 1]. For p < 0, the probability
()
density vanishes. It can be shown that fP (p) is indeed a probability density, non-
negative and satisfies the equality
Z
()
fP (p)dp = 1 (1.30)
To prove this property, we have to demonstrate that
Z
fP (p) sin[2(ln p b)]dp = 0 (1.31)
To do this, we make a substitution ln p = t. Thus, the integral to be calculated
reads
Z
1 (t b)2
I= exp sin[2(t b)]dt (1.32)
c 2 2c2
and further substitution t b = u leads to
Z
1 u2
I= exp 2 sin(2u)du (1.33)
c 2 2c
which vanishes since the integrand is an odd function of u. Thus, the function
defined in Eq. (1.29) represents a probability density of some random variable,
()
denoted by P () . Obviously, if = 0, fP is equal to fP as per Eqs. (1.25) and
(1.29). Stoyanov (1987, pp. 8991) demonstrates that for any k = 1, 2, . . . , we have
E(P () )k = E(P k ) (1.34)
()
i.e., the perturbed random variable P has the same moments as those of an un-
perturbed variable P .
Let us calculate now the true reliability associated with f () :
Z y a
()
R = Prob(P y a) = fP (p)dp (1.35)
0
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
0.02
0.015
Pf , act
0.01
0.005
0
1 0.5 0 0.5 1
Fig. 1.2 Relation between and Pf,act for c = 0.1, Pf,all = 0.01 (r = 0.99).
0.002
0.0015
Pf , act
0.001
0.0005
0
1 0.5 0 0.5 1
Fig. 1.3 Relation between and Pf,act for c = 0.2, Pf,all = 0.001 (r = 0.999).
or
Z y a
1 (ln p b)2
R= exp {1 + sin[2(ln p a)]}dp (1.36)
0 pc 2 2c2
Introducing again the new variable of integration ln p = t, we reduce reliability to
Z ln y a
1 (t b)2
R= exp {1 + sin[2(t b)]}dt (1.37)
0 c 2 2c2
and with t b = u, we have
Z ln y ab
1 ln y a b 1 u2
R = + erf + exp 2 sin(2u)du (1.38)
2 c c 2 0 c 2 2c
Evaluation of this integral yields
r
1 ln y a b 2 2 c2
R = + erf + e
2 c c 2 2
(1.39)
2c2 i(b + ln[y a]) 2c2 + i(b + ln[y a])
erfi + erfi
2c 2c
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Introduction 9
0.0002
0.00015
Pf , act
0.0001
0.00005
0
1 0.5 0 0.5 1
Fig. 1.4 Relation between and Pf,act for c = 0.2, Pf,all = 0.0001 (r = 0.9999).
Introduction 11
At this juncture it is instructive to introduce the central safety factor as the following
ratio
y
n= (1.48)
E()
along with the coefficient of variation of the stress
d
= (1.49)
E()
Dividing the numerator and denominator of Eq. (1.47) by E(), we obtain
n 1 + 3
R= (1.50)
2 3
This formula relates reliability, safety factor, and the coefficient of variation of the
stress. Expressing the safety factor from Eq. (1.50), we have
1
n = 1 + 2 3 R (1.51)
2
Probabilistic design is based on the requirement
Rr (1.52)
where r is the required, codified reliability, and the required safety factor nr is
obtained from Eq. (1.51) by setting R = r:
1
nr = 1 + 2 3 r (1.53)
2
This is a simple formula connecting the required safety factor nr with the required
reliability r and coefficient of variation of the stress.
We can deduce several useful conclusions:
(i) The safety factor, so often criticized by practitioners and researchers alike, is
actually a powerful concept, which can be given a probabilistic interpretation.
(ii) Probability theory strips the mystery from the safety factor, which instead of
being assigned at the will of the designer, or out of the sky as it were, gains an
analytical framework for its determination.
(iii) If one can quantify the required reliability, say through legislation, one can
quantify the safety factor as well.
(iv) If the required reliability r is greater than 0.5, i.e., if we do not tolerate nearly
half of our products being defective (and, hopefully, we dont!), the required
safety factor exceeds unity. This is in agreement with all textbooks and designs,
where the required safety factor is in excess of unity.
(v) The higher the coefficient of variation of the stress, the higher the required
safety factor. This matches ones anticipation, as it should be. Moreover,
Eq. (1.53) tells us exactly how much the safety factor must be, depending on
the magnitude of uncertainty . Were the coefficient of variation negligibly
small, i.e., = 0+ , the safety factor would be nr = 1+ . This implies that
the truly deterministic design is valid only when tends to zero. This fact
clearly indicates that purely deterministic design is contained as a particular
case in probabilistic design.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Introduction 13
lead to the same result! Not only is there no antagonism between them, but they
both tell us the same thing!
This implies that we can interchangeably utilize either of the two approaches!
Those who prefer the probabilistic sophistification, must be allowed (does anyone
need permission, really?) to continue development and usage of probabilistic me-
chanics. Those who prefer simplicity will stick to the anti-optimization technique.
Naturally, this begs the question: Is one of the approaches preferable to the
other? It appears that the anti-optimization approach, as the simpler of the two,
has a certain advantage. Moreover, as one of the respondents stressed in the poll
conducted by Elishakoff (2000b):
At the same time, it should be borne in mind that in its pure form, without
proper improvements in the structure via optimization, anti-optimization is likely
to yield ultra-conservative results.
This is why this monograph opts for a combination of anti-optimization and
optimization.
Another natural question may pop up: if the probability approach is so good as to
provide the same answer as its intellectual competitor, anti-optimization, why was
it necessary to talk, in the previous section, about it being (sometimes) bad?
It is important to have a reply to this question, because the variously named
non-probabilistic approaches are presented as sound alternatives to the probabilistic
description. It appears that this happens because probability theory often uses
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
random variables which stretch from minus infinity to plus infinity. Such is the case
with the best-known discussion the normal or Gaussian; do the exponential and
Gamma distributed random variables take on values from zero to infinity? Are any
of these random variables, as mental constructs, good for describing geometric or
material characteristics?
Whereas the exponential and Gamma distributions with their only positive val-
ues can be utilized for describing positive quantities, it is doubtful that there exists
any physical parameter that can take on values beyond physical limits. Likewise,
use of the normal distribution to describe positive quantities appears to be doubtful,
if not outright abnormal (pun intended). At best, these distributions are convenient
analytical approximations.
The above implies that engineers ought to use truncated distributions, i.e., dis-
tributions with bounded support. It is gratifying that awareness of this fact grows
amongst stochastic analysts. For example, Grigoriu (2006) mentions: All distri-
butions with bounded support in (0, ) are consistent with available information.
Ma, Leng, Meng and Fang (2004) write: Bounded random parameters ... are more
reasonable for engineering structures than the unbounded Gaussian random ones.
In their study, Minciarelli, Gioffre, Grigoriu and Simiu (2001) write: Unless oth-
erwise indicated, all uncertainty variables . . . will be assumed to have truncated
normal distribution.
Cai and Wu (2004) note:
Likewise, Cai and Lin (2006) write: Physically realistic random procedures are
bounded, and they may deviate far from being Gaussian. At the end of their paper
these authors are even more forceful: Physically realistic stochastic processes must
be bounded. (see also Cai (2003)).
Ang and Tang (1975) realized this fact quite long time ago:
Introduction 15
In their study, Simiu and Heckert (1996) analyzed the wind speed model and
fitted an approximate extreme value distribution. They concluded that the reverse
Weibull distribution is more appropriate than those of Gumbel or Frechet. Their
conclusion is supported by the physical fact that non-tornadic extreme winds are
expected to be bounded:
Kanda (1994) also showed that extreme winds are best fitted by distributions
with limited tails; see also Walshan (1994). Holmes (2002) notes that extreme winds
have a physical upper limit (which may differ for different storm types); hence,
they have a bounded distribution with data from nearly all stations in Australia
using various fitting methods. He reached a bounded generalized extreme value
(GEV) distribution in over 80% of the cases. He stresses: The approach ... in
substituting one third of the observation by unbounded, randomly generated
normal variables is not convincing.
Thus, the pragmatic engineer cannot agree with Lindley (1987) who states:
Probability is the only satisfactory description of uncertainty. Neither can one
agree with the notion implied by the title of Talebs 2001 book that we are fooled
by randomness. (Readers are also advised to read insightful articles on the different
sides of the issue by Klir (1989, 1994).)
Berleant and Goodman-Strauss (1998) used interval analysis to bound the re-
sults of arithmetic operations on random variables of unknown dependency. Thus,
randomness and boundedness may pragmatically interact with each other. Kim,
Ovseyevich and Reshetnyak (1993) and Elishakoff, Cai and Starnes (1994a) com-
pared probabilistic and anti-optimization approaches. Ferson and Ginzburg (1996)
advocate for application of different methods to describe ignorance and variability.
We opt in this book for the anti-optimization, or worst-scenario approach be-
cause it is simpler than the probabilistic one, not because we are against as Kosko
(1994) calls it the probability monopoly. Moreover, it yields the same results as
a more complicated probabilistic methodology (according to Hans Hoffmann, the
ability to simplify means to eliminate the unnecessary so that the necessary may
speak). We just add a spice to the approach: anti-optimized, worst-scenario re-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 2
Good, better, best. Never let it rest until your good is better and
your better is best! (Anonymous)
2.1 Introduction
In the design process of structures in various fields of engineering, the design vari-
ables such as the cross-sectional geometries, nodal locations, material properties,
etc., are modified by the designer to improve structural performance within the
limited cost for production and construction. Hence, the whole process of struc-
tural design may be regarded as that of seeking the best set of design variables
under given design requirements.
17
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Find the best design iteratively modifying the design variables by trial-and-error
process based on the experience of the designer, under requirements on the struc-
tural responses and mechanical properties, e.g., the stresses and displacements
against static/dynamic design loads, eigenvalues of free vibration, liner/nonlinear
buckling loads, etc.
A design satisfying all the constraints (feasible design) can be found automatically,
and efficiently, while simultaneously minimizing the objective function, such as
the total structural volume, if the problem is appropriately formulated so that
the set of feasible designs is nonempty.
The optimization tool helps the decision making of the designer, i.e., it is not an
automatic design tool that gives a negative impression to the structural engineers
and designers. If the optimal design is not acceptable to the designer, the upper-
level parameters, such as the geometry and material properties of the structure,
can be modified, or additional constraints can be assigned for the optimization
problem, to obtain a more reasonable optimal solution. Therefore, the designers
can spend more time on the synthetic jobs, rather than structural analysis, if
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Design parameter
Design load
(geometry, topology)
Response
(stress, displacement) Design variable
optimization tools are effectively used for decision making. The conventional
design process and optimization-based design process are illustrated in Fig. 2.1.
Optimization is very helpful for designing complex structures, for which even an
experienced designer cannot easily find a feasible design.
Even if the optimal solution cannot be used directly in design practice, the optimal
design gives insight into a better design.
If we start with a feasible and nearly optimal design found by an expert, the
design cannot be worse after optimization, and usually a better design can be
found.
Furthermore, in view of decision making,
The trade-off relation between the structural cost and responses can be made
clear, if optimization is carried out several times by modifying the input param-
eters such as the cost coefficients and upper bounds of responses.
The most positive ways of using optimization may be
To find new structural systems and shapes by optimization (Bendse and Sig-
mund, 2003).
To generate innovative structure or material that cannot be found without op-
timization, e.g., compliant mechanisms. Figure 2.2 shows an example of a com-
pliant bar-joint structure utilizing snapthrough behavior (Ohsaki and Nishiwaki,
2005).
Output node
Output
displacement
where subject to means under constraints on, and mechanical performances include
member stresses, nodal displacements, etc., for which the upper and lower bounds
are given.
The bounds for the mechanical performances are determined from the require-
ments given by the codes (regulations) in practical design process. The total struc-
tural volume (or weight) is usually assigned as the objective function, because it is a
critical requirement to reduce the weight of the aerospace and mechanical structures.
For the long-span structures in architectural and civil engineering, the reduction of
the self-weight generally leads to reduction of the design loads, and hence leads to
lower cost. It is also important that, for any acceptable definition of the objective
function, a solution satisfying all the constraints, called feasible design, is obtained
after optimization as discussed in Sec. 2.2.
If the concept of minimum weight is not under consideration, the following
alternative formulation may be valid:
Maximize mechanical performance subject to upper bound for cost (weight, vol-
ume).
Note that the two formulations stated above lead to similar optimal solutions if the
problem parameters are appropriately assigned.
Consider, for instance, the problem of minimizing the structural volume of a
truss under constraints on stresses and displacements against static loads. The de-
sign variables are the cross-sectional areas of the members. Let = (1 , . . . , m )>
and U = (U1 , . . . , Un )> denote the vectors of member stresses and nodal displace-
ments that are called state variables, respectively, where m is the number of mem-
bers and n is the number of degrees of freedom. All vectors are column vectors, and
a subscript is used for indicating a component of a vector throughout this book.
The vector of member cross-sectional areas are denoted by A = (A1 , . . . , Am )> ,
and let Li denote the length of the ith member.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where the upper and lower bounds are denoted by the superscripts ( )U and ( )L ,
respectively.
In Problem (2.1), the stress of the ith member is written as a function of A only
as i (A). However, for a general statically indeterminate structure, the responses
are defined implicitly through the vector of nodal displacements U (A), which is
also a function of A. The responses may also depend explicitly on A, e.g., the axial
force Ni of member i is defined as i (U )Ai .
Next, we formulate the problem in a more general form. Let x = (x1 , . . . , xm )>
denote the vector of design variables representing the cross-sectional areas, nodal
coordinates, etc. Suppose the objective function such as the total structural volume
is defined as an explicit function of x as F (x). Then, the structural optimization
problem is formulated as follows:
where Gi and Hi are generally nonlinear functions of x and U (x), and N E and N I
are the numbers of equality and inequality constraints, respectively. If the design
variables x can take continuous values, then the structural optimization problem
(2.2) is classified as a nonlinear programming problem, for which various algorithms
such as sequential quadratic programming and method of feasible directions are
available. On the other hand, if the variables take integer values or are selected
from a list or a catalog of available values, then the problem is classified as a mixed
integer nonlinear programming problem with continuous state variables U (x) (see
Sec. 2.4.2 for a detailed classification of optimization problems).
The standard approach to nonlinear programming problem is based on the gra-
dient information of the objective and constraint functions. Suppose the vector U
of state variables is defined in a general form by the state equations such as the
equilibrium (stiffness) equations as
Local optimal solution: The solution that has the smallest objective value
among those in its neighborhood feasible solutions is called a local optimal solu-
tion.
Global optimal solution: The solution that has the smallest objective value
among those of all the feasible solutions is called a global optimal solution.
Active/inactive constraints: An inequality constraint that is satisfied with
equality is said to be active. The constraint is said to be inactive if it is sat-
isfied with strict inequality.
Convex set: Let x1 and x2 denote two solutions belonging to a set I. If x1 +
(1 )x2 is contained in I for any 0 1, then the set I is said to be a
convex set.
Convex function: Let x1 and x2 denote two solutions belonging to a convex set
I. If F (x1 ) + (1 )F (x2 ) F (x1 + (1 )x2 ) for any 0 1, then F (x)
is said to be a convex function.
Convex programming problem: If the objective function is convex, and the
feasible region is a convex set, then the optimization problem is called a convex
programming problem, for which the solution satisfying the conditions of local
optimality is also a global optimal solution.
Positive semidefinite matrix: An n n symmetric matrix S is positive semidef-
inite, if all the eigenvalues of S are non-negative, which is equivalent to the non-
negativeness of the bilinear form b> Sb for any n-vector b. Furthermore, S is
positive definite if b> Sb is positive for any b 6= 0.
The constraints (2.7d) are called side constraints, bound constraints, or box con-
straints, which are treated independently of the general inequality constraints (2.7c)
in most of the optimization algorithms.
If the variables take only integer values, the problem is called an integer pro-
gramming (IP) problem. Since the optimal combination of the variables is found
by IP, it is equivalently called combinatorial optimization problem. Among various
formulations of IP, the problem with 01 variables only is called a 01 programming
problem. A problem that has real and integer variables is called a mixed integer
programming (MIP) problem.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Consider next the case where all the variables can take real values. If F (x),
Gi (x) and Hi (x) are all linear functions of x, the problem is called a linear pro-
gramming (LP) problem, for which the global optimal solutions can be found by
well-established methods such as the simplex method and interior point method. If
at least one of F (x), Hi (x) and Gi (x) is a nonlinear function of x, the problem is
called a nonlinear programming (NLP) problem.
For LP and NLP, the Lagrangian for Problem (2.7) is given as
E I
N
X N
X
L(x, , ) = F (x) + j Gj (x) + j Hj (x) (2.8)
j=1 j=1
j 0, j Hj = 0, (j = 1, . . . , N I ) (2.9b)
with the constraints (2.7b) and (2.7c), where the conditions (2.9b) are called a
complementarity conditions.
Consider a simple problem called quadratic programming (QP) problem, where
the objective function is a quadratic function of x and we have only linear equality
constraints as
minimize x> Dx + c> x (2.10a)
>
subject to C x b = 0 (2.10b)
where b = (b1 , . . . , bn )> and c = (c1 , . . . , cm )> are constant vectors, C is an n m
constant matrix, D is an m m constant matrix, and we assume rank C < m.
If D is positive definite, then Problem (2.10) is called convex QP. From the
KKT conditions, the optimal solutions are explicitly written as
1
x = D 1 (C > + c),
2 (2.11)
= (CD1 C > )1 (2b + CD1 c)
In this book, we often formulate an anti-optimization problem of maximizing
(minimizing) a linear objective function under a quadratic constraint on x as
maximize c> x (2.12a)
>
subject to x Dx b (2.12b)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where b is a positive constant, and D is a positive definite weight matrix. The anti-
optimal solutions for Problem (2.12) are explicitly derived from the KKT conditions
as
1 1 CD 1 C
x= D c, = = (2.13)
2 2 b
where is the Lagrange multiplier for the constraint (2.12b), and = corre-
sponds to the optimal solution that minimizes c> x.
The optimization problem containing constraints such that the variable matrix is
positive semidefinite is called a semidefinite programming (SDP) problem (Wolkow-
icz and Vandenberghe, 2000; Ohsaki and Kanno, 2007). The objective function and
other constraints are linear in the standard form of SDP problem. Since the con-
straints of positive semidefiniteness of matrices include linear and convex quadratic
constraints, the SDP is an extension of LP and convex QP.
Let X O indicate that the symmetric matrix X is positive semidefinite. The
inner product X Y of the n n matrices X = (Xij ) and Y = (Yij ) is defined as
n X
X n
X Y = Xij Yij (2.14)
i=1 j=1
Then the following relation is obtained from Eqs. (2.19) and (2.20):
m
Rj X Rj b xi
+ = 0, (j = 1, . . . , N A ; k = 1, . . . , N P ) (2.21)
pk i=1
x i p k
By multiplying b
xi /pk to the both sides of Eq. (2.17) and taking summation
over i, we obtain
A
m N m
X F b xi X X Rj b xi
+ j = 0, (k = 1, . . . , N P ) (2.22)
i=1
x i pk j=1 i=1
x i p k
The following relation is derived by multiplying j to Eq. (2.21) and taking sum-
mation over j corresponding to the equality constraints and the active inequality
constraints:
A
N
X NA m
Rj X X Rj b xi
j + j = 0, (k = 1, . . . , N P ) (2.23)
j=1
p k j=1 i=1
x i p k
is derived. We can see from Eq. (2.24) that the derivative of the optimal objective
value with respect to pk can be obtained without computing the derivatives of the
variables and the Lagrange multipliers with respect to the problem parameter.
Differentiation of Eq. (2.17) with respect to pk leads to
Xm NA
X 2 2 NA
X
2
F R j b
x l F 2 Rj
+ j + + j
xi xl j=1 xi xl pk xi pk j=1 xi pk
l=1
(2.25)
XNA bj
Rj
+ = 0, (i = 1, . . . , m; k = 1, . . . , N P )
j=1
x i p k
The derivatives of b1 , . . . ,
bN P and x
b1 , . . . , x
bm with respect to pk are computed
A
from a set of m + N linear equations (2.21) and (2.25), where the terms in the
parentheses of the first term on the left-hand-side of Eq. (2.25) consist of the Hessian
of the Lagrangian.
F2
F1
Fig. 2.3 Pareto optimal set in objective function space for N F = 2; gray region: feasible region.
where the side constraints are included in the inequality constraints (2.26b) for
simplicity.
Contrary to the single-objective problem, generally there is no solution that si-
multaneously minimizes all objective functions. Consider two feasible solutions x(1)
and x(2) satisfying all the constraints (2.26b) and (2.26c). Dominance of solution
is defined as
The set of Pareto optimal solutions is called a Pareto optimal set, or Pareto set for
brevity. The thick line in Fig. 2.3 illustrates the set of Pareto optimal solutions in
the objective function space for the case of N F = 2, where the gray region is the
set of feasible solutions. For example, the solution A is a Pareto optimal solution,
because there is no solution in the region surrounded by the two dotted lines, where
both of the two objective functions are improved.
The approaches to selecting the most preferred solution from the Pareto optimal
set are classified as a priori articulation of preference and a posteriori articulation of
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
preference (Marler and Arora, 2004). In the a priori approach, the objective func-
tions are combined or transformed to constraints to formulate a single-objective
problem. The most popular and convenient approach for this purpose is the lin-
ear weighted sum approach, where the objective functions F1 (x), . . . , FN F (x) are
combined as
F
N
X
F (x) = wi Fi (x) (2.27)
i=1
with positive weight coefficients w1 , . . . , wN F . This approach is suitable for the case
where the Pareto optimal solutions form a smooth convex curve or hyper-surface
in the objective function space, and the preference of the designers or the decision
makers are rather fixed. Note that the single-objective problem for minimizing F (x)
should be solved iteratively modifying the weight coefficients wi , if the solution is
not acceptable to the decision maker.
In the a posteriori approach, on the other hand, a set of Pareto optimal solutions
is first generated, and the most preferred solution is selected by the decision maker
from the set based on additional preference functions or trade-off relations among
the objective functions. Even for a problem with continuous variables and smooth
differentiable functions, an NLP approach is not suitable for the purpose of gener-
ating many Pareto optimal solutions with enough diversity, because it generates a
single optimal solution for the given set of problem parameters. The genetic algo-
rithm (GA) is one of the most suitable method for this purpose (Goldberg, 1989;
Ohsaki, 1995), because it is classified as a population-based approach that has many
solutions at each step of iteration called generation. On the other hand, the sim-
ulated annealing (SA) and tabu search (TS) are categorized as single-point-search
heuristics that have only one solution at each step of optimization (see Sec. 2.6 for
details of SA for single-objective problems). Recently, SA has been extended to mul-
tiobjective programming problems (Whidborne, Gu and Postlethwaite, 1997). TS
has also been shown to be very effective with multiobjective problems (Baykasoglu,
Owen and Gindy, 1999; Ohsaki, Kinoshita and Pan, 2007).
surface method (RSM), has been widely applied to structural optimization problems
due to its simplicity of implementation (Roux, Stander and Haftka, 1998).
Let F (x) denote the response quantity to be approximated by a quadratic func-
tion of the variables x = (x1 , . . . , xm )> as
m
X m X
X m
F (x) ' F a (x) = c0 + c i xi + cij xi xj + (2.28)
i=1 i=1 j=i
where c0 , ci , cij , . . . are the coefficients. Suppose the values of F (x(k) ) at the set
of data points x(k) (k = 1, . . . , N D ) are known. The coefficients are determined by
minimizing the square error e given as
D
N
X
e= (F a (x(k) ) F (x(k) ))2 (2.29)
k=1
The differential coefficients of e with respect to cij vanishes as
ND
X Xm Xm X m
e c0 + (k) (k) (k) (k) (k)
=2 c i xi + cij xi xj F (x(k) ) xi xj = 0 (2.30)
cij i=1 i=1 j=i
k=1
Similar equations are obtained for differentiation with respect to ci . Thus, the
coefficients are found by solving a set of linear equations based on the well-known
method of least squares.
Suppose the variables are normalized to the range 1 xi 1 (i = 1, . . . , m).
The sampling values for xi may be given by the two-level model (1, 1), three-level
model (1, 0, 1), etc. In the three-level model, however, the number of data points is
3m if the set of all possible combinations is to be considered. The number increases
exponentially as m is increased. Therefore, the method called design by experiment
(Myers and Montgomery, 1995) is effectively used to select the data points of a
moderately small number, while a priori estimation of the approximation error is
minimized to obtain the set of data points.
2.6 Heuristics
which are generally nonlinear functions of the design variables. Thus, the problem
turns out to be a mixed-integer nonlinear programming (MINLP) problem (Floudas,
1995).
It is very easy to solve a combinatorial optimization problem by using the inte-
ger programming approaches, e.g., branch-and-bound method and branch-and-cut
method (Horst and Tuy, 1990), if the number of variables is small and the sub-
problem formulated by relaxing some integer variables to real variables is convex.
However, the computational cost increases as an exponential function of the prob-
lem size, and it is not possible to solve a practical problem within a practically
admissible computational time. Furthermore, for the structural optimization prob-
lem, the relaxed nonlinear programming problems are usually non-convex problems,
for which the global optimality is not guaranteed.
Recently, with the rapid development of computer hardware and software tech-
nologies, we can carry out structural analysis many times to obtain optimal solu-
tions. Another important point is that the global optimality need not be strictly
satisfied in the practical design process. Heuristic approaches (or heuristics for sim-
plicity) have been developed to obtain approximate optimal solutions within rea-
sonable computational time, although there is no theoretical proof of convergence
(Reeves, 1995). The most popular approach is the genetic algorithm (GA) (Gold-
berg, 1989), which can be categorized as a multipoint search or population-based
method that has many solutions at each iterative step called generation. Since com-
putational cost for evaluating the objective and/or constraint functions at each step
can be very large for structural optimization problems, a multipoint strategy may
not be appropriate especially for optimization/anti-optimization of large structures.
Therefore, single-point search heuristics such as simulated annealing (SA) (Aarts
and Korst, 1989) and tabu search (TS) (Glover, 1989) can be effectively used. In
the following, we present basic algorithms of single-point search heuristics.
where D1 is the Manhattan norm (L1 -norm), D2 is the Euclidean norm (L2 -norm),
and D is the Chebychev norm (L -norm). For an integer variable defined by a
binary bit string, the humming distance is defined by the number of different bits,
e.g., the humming distance between the binary numbers 1010110 and 1001101 is 4.
However, the definition of a neighborhood for heuristics is less strict than those
based on distances. For example, the following operations can be used for generating
a neighborhood solution:
Step 1 Assign the initial solution J (0) , and set the iteration counter k = 0.
Step 2 Randomly generate a neighborhood solution J of the current solution J (k) ,
or select the best solution J from randomly generated neighborhood solutions
based on the criterion defined by the values of the objective function F (J ) and
the constraint function Hj (J ).
Step 3 Update the solution as J (k+1) = J or reject J as J (k+1) = J (k) in
accordance with the given strategy.
Step 4 Output the best solution satisfying the constraints and terminate the pro-
cess, if the solution converged, or the stopping criteria are satisfied; otherwise,
let k k + 1 and go to Step 2.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
We can use either specified or randomly generated values as the initial solution.
An optimal solution obtained by another optimization approach can be used, with
possible modification, as the initial solution; e.g., the nearest solution with integer
variables from the optimal solution with real variables can be used.
For a constrained optimization problem, the constraints can be incorporated to
the objective function by using the penalty function approach. For example, for
a problem with inequality constraints Hj (J ) 0 (j = 1, . . . , N I ), the following
penalty function can be used to penalize the solution with violated constraints:
I
N
X
Fe (J ) = F (J ) + j [max(Hj (J ), 0)]2 (2.32)
j=1
where j is a sufficiently large penalty coefficient. It is seen from Eq. (2.32) that
no penalty is given if the constraint is satisfied as Hj (J ) 0, and the penalty
proportional to the square of Hj (J ) is given if the constraint is violated. The
penalty function of this type is called the exterior penalty function.
On the other hand, the interior penalty function can be defined by the logarith-
mic barrier function as
NI
X
e
F (J ) = F (J ) j log(Hj (J )) (2.33)
j=1
where the penalty coefficient j should be sufficiently small. However, the interior
penalty function approach cannot be used for the case where a solution can be
infeasible during the search process, because log(Hj (J )) is not defined for Hj (J ) >
0. Therefore, only the feasible solutions should be searched, which is very difficult
for a heuristic approach to optimization of large-scale complex structures.
The approach that is the reverse of the greedy method, which starts from a solution
satisfying all the constraints and reduces the objective value consecutively, is called
the stingy method. The greedy method and the stingy method work fairly well if
F (J ) and Hj (J ) are monotonic functions of J. However, the algorithms strongly
depend on the initial solution, and will often reach a local optimal solution.
Step 1 Randomly generate the initial solution J (0) . Initialize the temperature
parameter as T (0) = T0 , where T0 is the specified value, and set the iteration
counter k = 0.
Step 2 Let J denote a randomly generated neighborhood solution, and define
F = F (J ) F (J (k) ). If F < 0, let J (k+1) = J ; otherwise, accept J with
the probability P defined by the temperature parameter T (k) as
F
P = exp (k) (2.34)
T
Step 3 Decrease T (k) to T (k+1) based on the prescribed rule, e.g., the Metropolis
rule T (k+1) = T (k) , where is the specified parameter that is slightly less than
1.
Step 4 Set k k + 1 and go to Step 2 if the termination condition is not satisfied;
otherwise, output the best solution satisfying the constraints, and terminate the
process.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Optimum design
Shape optimization
Cross-sectional Topology Geometry
optimization optimization optimization
Fig. 2.4 Classification of optimization of trusses and frames with respect to design variables.
P P
P P
Feasible region
Ceiling
Wall
Cable
H
Rod
P = 1000 N
W = 1000 mm
( 10 4 )
8.0
6.0
5.0
30 40 50 60 70
Angle of cable (deg.)
Fig. 2.8 Relation between the angle of the cable and the optimal total structural volume.
W W
H H
P P
(a) (b)
given as
c 50, 50 r (2.35)
Fig. 2.10 Optimal cross-sectional areas with four units for H/W = 0.2.
Table 2.1 Relation between the number of units and the optimal structural vol-
ume.
Number of units 2 3 4 5 6
Optimal structural volume (105 ) 1.620 1.533 1.530 1.560 1.607
where Y is the vector describing random parameters of the system, and E[C] and
E[Cf ] are the mean values of C and the expected cost Cf due to failure with regard
to the jth criterion, respectively. In his review article, Schueller (1998) emphasizes
that ... irrespective of the route of optimization which is followed, its treatment
within efficient software environment is an indispensable requirement.
A pertinent question arises concerning the assignment of the acceptable proba-
bility of failure. Freudenthal (1956), anticipating the reliability-based optimization,
stressed the following about acceptable risk, which ought to be evaluated:
Hence, the cost and probability of failure should be accurately defined for reliable
formulations of probabilistic or reliability-based optimization (see further discussion
in Chap. 10).
It appears that we ought to describe all (that are known to us, obviously) approaches
to optimization of structures, so as not to leave an impression that we include only
what we think is important in this field.
The first study on fuzzy sets was apparently pioneered by a British-American
philosopher, Max Black (19091988) over 70 years ago. He started by quoting the
famous British philosopher Bertrand Russell (18721970), who stated that Vague-
ness and accuracy are important notions, which are very necessary to understand.
Black (1937) introduces the essence of his argument as follows:
Further,
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
The word fuzzy has the meaning vague, not clear, and is
used ... in contrast to the precision normally expected of mathe-
matics ... with fuzzy sets, a mathematical model can be made of
properties or quantities that are imprecisely defined, such as every-
day statements.
According to Kosko (1992), fuzzy theory holds that all things are matters of de-
gree ... Fuzzy theory reduces black-white logic and mathematics to special limiting
cases of gray relationships.
In applications of the fuzzy set theory, one first identifies the so-called universal
set (Klir, St. Clair and Yuan, 1997), which is a set that consists of all the members
that are of interest in the problem at hand. For example, if one is dealing with
classifying loads applied at a structure by various criteria, then the universal set is
composed of all the values that the loads can take on.
Every subset A of a set X can be uniquely represented by a function, called
characteristic function A (x), defined as
1, if x A
A (x) = (2.42)
0, if x /A
for any x X . Examples of characteristic functions are shown in Fig. 2.11. The
function c (x) in Fig. 2.11(a) corresponds to a crisp set or a classical non-fuzzy set
defined by 0 x x0 , where c (x) = 1 for 0 x x0 ; otherwise, c = 0. However,
membership of the load in a fuzzy set may allow some uncertainty; therefore, it is
a matter of degree (Kosko, 1992; Klir, St. Clair and Yuan, 1997). One can state
that the degree of membership constitutes the degree of compatibility of introduced
operating concepts with the fuzzy set. Figure 2.11(b)(d) illustrates three types
of characteristic functions, where the function value 0 F (x) 1 indicates the
corresponding value of x is possibly the member of the set; hence, the characteristic
function is also called the membership function.
For example, consider the uniform bar under tensile load, which is chosen as
variable x. Then X can be identified with the set of all nonnegative real numbers,
assuming the material strength of the bar is unbounded. Let N be a set of loads
that are real values and range from 10 kN to 20 kN. Then, if we use the framework
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
c F
1 1
0 x0 x 0 x0 x
F
F
1 1
0 xL xU x 0 xL xU x
Zadeh (1994) vehemently disagrees with these criticisms in his tellingly titled
paper Why the success of fuzzy logic is not paradoxical. The following definition
is proposed:
... the skeptics will find it hard to understand why they failed to
realize that fuzzy logic is a phase in a natural evolution of science
an evolution brought by the need to find an accommodation with
the pervasive impression of the real world.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Haim (1997), Elishakoff (1990, 1998c), Pal (1999), and Oberguggenberger (2004)
discuss various possible modeling techniques of uncertainty. Finally, a mention
should be made of random sets that combine the probabilistic and set-based models
(Matheron, 1975; Dubois and Prade, 1991; Bernardini, 1999; Tonon, 2004). Opti-
mization of uncertain structures in the context of the random sets was performed by
Tonon and Bernardini (1998). M oller and Beer (2008) studied optimization in con-
junction with non-probabilistic uncertainty modeling. Popular and highly readable
books on the fuzzy sets and logic are those by McNeill and Freiberger (1993), Kosko
(1993, 1999), and Sangalli (1998). Pertinent books that employ the fuzzy sets in
the structural analysis context are those by Blockley (1980), Yao (1985), Ayyub
(1998), Vick (2002), M oller and Beer (2004), Hanss (2005), and Fellin, Lessmann,
Oberguggenberger and Vieider (2005). Nikolaidis and Haftka (2001) revised theo-
ries of uncertainty for risk assessment when data are scarce. Tsompanakis, Lagaros
and Papadrakakis (2007) edited a definitive collection of papers on various methods
of uncertainty analysis in the optimization context. Nikolaidis, Ghiocel and Singhal
(2005) edited Engineering Design Reliability Handbook which provides much more
than the title promises; the handbook covers both the traditional probabilistic re-
liability as well as various non-probabilistic treatments of uncertainty.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 3
We are not suggesting that all computation should be carried out using
interval techniques but only that interval methods provide another set
of tools in applied mathematics... (Ramon E. Moore, 1979)
3.1 Introduction
47
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(i) A physics problem at Lockheed (ca. 1960): Q: Is the strange behavior of the
computer model due to round-off errors? A: After converting the program to run
in interval arithmetic with outward rounding, it was determined that round-off
error was very small in the original program. Result: the physicist took another
look at the model equations and found that there was a missing term.
(ii) A long controversy between research groups at MIT and Caltech concerned
whether the observed behavior of computer simulations was due to roundoff error
of defects in the mathematical model, in the case of computer solutions of the
Birkhoff-Rota complex partial differential/integral equations modeling the onset
of turbulence in wind-shears. My graduate student Jeffrey Ely wrote a program
for variable-precision interval arithmetic with outward rounding, and finally by
using about 300 decimal place (nearly 1000 bits) in outwardly rounded interval
arithmetic, was able to settle the controversy. By carrying enough bits, the model
realistically determined the onset of turbulence at a reproducible, finite time after
the initial appearance of the wind-shear. I have always found it odd to suppose
that anything we want to compute can be done carrying only some fixed number
of digits or bits. Is 40 bits enough? 80? A thousand? It depends on what we are
trying to compute...
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
The natural question arises: What is interval analysis? As Hayes (2003) informs
us, Interval analysis is not a new idea. Invented and reinvented several times, it
has never quite made it into the mainstream of numerical computing, and yet it
has never been abandoned or forgotten either. The first paper on interval analysis
apparently belongs to Rosalind Cicely Young (1931) who published a paper on an
algebra of many-valued quantities and gave rules for calculating with intervals. Paul
S. Dwyer (1951) described arithmetic with intervals, calling it range numbers. The
consequent three works appeared nearly simultaneously: by Mieczyslaw Warmus
(1956) in Poland, Teruo Sunaga (1958) in Japan, and Ramon E. Moore (1966) in
the United States.
There are several concepts for modeling uncertainty based on the non-probabilistic
approach. Ladev`eze, Puel and Romeuf (2006) introduced the concept of lack of
knowledge to take into account uncertainty of the parameters and modeling er-
rors. They proposed a method for reducing errors through additional information,
and applied it to static and free vibration problems. Oden, Babuska, Nobile, Feng
and Tempone (2005) classified the errors of numerical analysis into modeling error,
discretization error and error due to uncertainty. Soize (2005) classified the un-
certainty into data uncertainty and model uncertainty. Data uncertainty concerns
the structural parameters defining geometry, stiffness, material properties, and so
on, which can be successfully represented by the parametric probabilistic approach.
Model uncertainty, on the other hand, is related to approximation in analysis and
the details that are actually unknown. They applied their methods to dynamic
problems. Probabilistic and non-probabilistic approaches to anti-optimization are
compared by Elishakoff and Zingales (2003). Various models of uncertainty are
reviewed by Elishakoff and Fang (1995).
Let a = (a1 , . . . , aN )> denote the vector of uncertain parameters, such as nodal
loads and material properties. The types of the feasible regions or the bounds of
the uncertain parameters are classified as follows (Qiu, M uller and Frommer, 2001):
aLi ai aU
i , (i = 1, . . . , N ) (3.1)
3.3.1 Introduction
In the theory of structural analysis, we often encounter the situation that once the
problem is formulated and the governing differential equations are written down,
the exact solution is either unavailable or demands elaborate symbolic analysis
with special functions. In these circumstances, we usually resort to approximation
techniques like the RayleighRitz or Galerkin method. For the eigenvalues, some
enclosure methods are developed. Thus, intervals which contain the exact solution
of the formulated problem are to be found.
Archimedes, in the third century before the common era (circa 287212 B.C.E.)
derived such bounds for the transcendental number , showing that it belongs to
the interval
10 1
3 <<3 (3.7)
71 7
by approximating the circle with the inscribed and circumscribed 96-side regular
polygons.
In this section, we present an example of interval analysis of simple static prob-
lems.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
[12.5, 14.7]
[2.32, 2.41]
[12.5, 14.7]
M = [mL , mU]
N1
N2
R N1
N2
Next, we consider the uniform bar (type-1) as shown in Fig. 3.2, which is clamped
at its left end, and is subjected to two loads of uncertain magnitudes. Still, we can
state with confidence that the values that the loads can take lie within certain
bounds. Let the first load N1 lie between 17.4 kN and 21.5 kN, whereas the second
load N2 lies somewhere between 12.5 kN and 14.7 kN. The reaction R = N2 + N1
then lies in the interval
[12.5, 14.7] + [17.4, 21.5] = [12.5 + 17.4, 14.7 + 21.5] = [29.9, 36.2] kN (3.9)
N1
N2
R N1
N2
An interval X of a real number x is defined with the upper bound xU and the
lower bound xL as
X = [xL , xU ] = {x|xL x xU } (3.14)
The following operations are defined for the addition and subtraction of the variables
in the intervals X = [a, b] and Y = [c, d]:
Addition: X + Y = [a + c, b + d]
(3.15)
Subtraction: X Y = [a d, b c]
Note that the subtraction can be reduced to the sum as
X Y = X + (1) Y (3.16)
Likewise, the division of the non-interval variable x by an interval variable Y , pro-
vided that Y does not contain zero, is reduced to the multiplication
x/Y = x [1/y U, 1/y L ] (3.17)
In general, a single formula provides a definition of the four standard arithmetic
operations to intervals. Let us denote by any of the operations +, , and /
(Muhanna, Mullen and Zhang, 2005), and the division by an interval variable that
includes zero is prohibited. Then X Y is defined as
X Y = [min xi y j , max xi y j ]
(3.18)
for i {U, L}, j {U, L}, {+, , , /}
which is alternatively written as
[xL , xU ] [y L , y U ]
= [min(xL y L , xL y U , xU y L , xU , y U ), (3.19)
L L L U U L U U
max(x y , x y , x y , x , y )]
In other words, one has to compute the four possible combinations of the lower and
upper bounds. The next step is to choose the smallest of the four values as the
lower bound, and to take the largest of them as the upper bound. We are sure that
every possible combination of x y lies within these limits.
It must be stressed that the interval analysis processes information obtained for
input data to provide the interval for the output. Moore (2006) writes:
Suppose we have measured that input parameters fall within cer-
tain upper and lower limits, then we can use interval inputs for
them, and interval computation is designed for just that sort of
thing.
For example, if we measure a width as w = 7.2 0.1, length as
l = 14.9 0.1, and height as h = 405.6 0.02, then the volume
V = w l h is within the interval
([7.1, 7.3] [14.8, 15.0]) [405.4, 405.8]
(3.20)
= [42599.432, 44435.1] = 43517.266 817.834
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Theorem 2: Let D be a convex set in N -dimensional space, which has the property
of central symmetry with respect to some point a0 . There exists an ellipsoid
E(a0 , W ) D such that E(a0 , (1/N )W ) contains D, i.e.,
The proof of Theorem 1 is given in the book by Leichtweiss (1980), whereas the
proof of Theorem 2 is given by John (1998). For approximating sets, Chernousko
(1999) suggests using inner ellipsoid E L of the maximum volume and outer ellipsoid
E U of the minimal volume. The following theorems (Zaguskin, 1958) are relevant:
Theorem 3: For any bounded set D in N -dimensional space, there exists a unique
ellipsoid E U of the minimum volume containing D as E U D.
Theorem 4: For any closed convex set D in N -dimensional space, there exists a
unique ellipsoid E L of the maximum volume contained in D as E L D.
Here a01 and a02 are the N -dimensional vectors of the centers of ellipsoids, and W 1
and W 2 are the N N positive definite matrices. The Minkowski sum of these
ellipsoids reads
a = a1 + a2 S E(a01 , W 1 ) E(a02 , W 2 ),
(3.67)
a1 E(a01 , W 1 ), a2 E(a02 , W 2 )
where the set S is bounded, closed, and convex, and in general is not an ellipsoid.
Chernousko (1980) posed the following pertinent problems:
Problem 1 Find the ellipsoid E(aU , W U ) of the minimal volume containing the
sum S of two ellipsoids (3.67), i.e., an ellipsoid such that S E(aU , W U ),
det W U min.
Problem 2 Find the ellipsoid E(aL , W L ) of the maximal volume contained in
the sum S of two ellipsoids (3.67), i.e., an ellipsoid such that E(aL , W L ) S,
det W L max.
Chernousko (1980) solved both problems. For the details, see Chernousko (1999,
p. 143). Chernousko (2000) writes:
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Such reachable sets were derived by Chernousko (1988, 1994, 1999, 2005).
The apparently first applications of ellipsoidal analysis to applied mechanics was
the work by Borkowski (1983). Ellipsoidal analysis methods have been considered by
Attoh-Okine (2002, 2004), Au, Cheng, Tham and Zheng (2003), Ben-Haim (1992,
1999), Calafiore and El Ghaoui (2004), Chernousko (1980), Elishakoff (1991b, 1991c,
1994, 1995b, 1998a, 1998c, 2000b, 2004), Elishakoff and Ben-Haim (1990b), El-
ishakoff and Colombi (1993), Elishakoff, Gana-Shvili and Givoli (1991), Elishakoff,
Li, and Starnes (2001), Elishakoff, Lin and Zhu (1994e), Elseifi and Khalessi (2001),
Kurzhanski (1977), Kurzhanski and Valyi (1992, 1996), Lindberg (1992a, 1992b),
Pantelides and Booth (2000), Pantelides and Ganzerli (2001), Pantelides and Tzan
(1996), Qiu (2003, 2005), Qiu, Chen and Wang (2004a), Qiu, Ma and Wang (2006a),
Qiu, M uller and Frommer (2004e), Schweppe (1973), Tzan and Pantelides (1996a,
1996b), Vinot, Cogan and Lallement (2003), Wang, Elishakoff and Qiu (2008a),
Yoshikawa (2002, 2003), Yoshikawa, Nakagiri and Kuwazuru (1998b), Zhu and El-
ishakoff (1996), and Zuccaro, Elishakoff and Baratta (1998).
Ben-Haim (1985) observed that ellipsoidal analysis is a particular case of convex
analysis. A first application of convex analysis to applied mechanics are apparently
the works by Moreau (1966, 1976, 1979) and Panagiotopoulos (1976, 1985). The
comprehensive presentation of convex analysis in the applied mechanics context was
made by Ben-Haim and Elishakoff (1990).
In closing this section, it must be noted that in recent work by Kreinovich,
Neumaier and Xiang (2008) a combination of interval and ellipsoidal uncertainties
is treated. Specifically, they consider the interesting case in which the actual values
of uncertain variables belong to the intersection of interval and ellipsoid.
where gi is the specified positive parameter that represents the semi-axis of the
parameter ai . Note that a has been appropriately transformed so that the ellipsoid
has its center at the origin 0 and the direction of each semi-axis coincides with a
coordinate axis.
The anti-optimal solution of Problem (3.68) can be found from the stationary
condition of the Lagrangian Q defined by
" N 2 #
X ai
Q(a, ) = f (a) + 1 (3.69)
i=1
gi
where 0 is the Lagrange multiplier. The stationary condition of Q with respect
to a leads to
Q f 2ai
= 2 = 0, (i = 1, . . . , N ) (3.70)
ai ai gi
The details of the Lagrange multiplier approach can be found in Sec. 2.4.2.
If f (a) is a linear function of a, or has been linearly approximated with respect
to a as
XN
f (a) = ci a i (3.71)
i=1
Note that = corresponds to the minimum value of f (a). This way, the
anti-optimal solution can be obtained by a simple arithmetic operation.
If f (a) is a nonlinear function, then only locally anti-optimal solution can be
found by using a method of nonlinear programming. The global anti-optimality is
guaranteed when f (a) is a concave function.
Alternatively, the bound constraints can be assigned to the anti-optimization
problem as
maximize f (a) (3.75a)
subject to aLi ai aU
i , (i = 1, . . . , N ) (3.75b)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where aLi and aU i are the lower and upper bounds for ai .
Problem (3.75) is a linear programming problem, if f (a) is a linear function of
a. For this simple problem with bound constraints only, the solution is found at
ai = aLi or aUi depending on the sign of the coefficient for ai in f (a). For the case
where f (a) is a nonlinear function, the globally anti-optimal solution can be found
by a nonlinear programming approach, if f (a) is concave. On the other hand, if f (a)
is a convex or monotonic function of a, then the globally anti-optimal solution can
be found by enumerating the vertices of the feasible region defined by Eq. (3.75b)
(see Secs. 4.6 and 5.6 for examples of anti-optimization by vertex enumeration).
If ai 0 (i = 1, . . . , N ), the bounds for a are generalized as
" N # p1
X ai p
1, (p = 1, . . . , ) (3.76)
i=1
gi
y 4
3
x
Y 2 Node 2
1
X i
Node 1
Fig. 3.4 Coordinates, displacement numbers and node numbers of a plane truss member.
coordinates is given as
Ai E >
kli = dd (3.79)
Li
where
d = (1, 0, 1, 0)> (3.80)
and E is Youngs modulus.
From Eqs. (3.77) and (3.79), the member stiffness matrix k gi with respect to the
global coordinates is given as
Ai E > >
kgi = T dd T i (3.81)
Li i
and kgi of all the members are assembled to construct the n n global stiffness
matrix K, where n is the number of degrees of freedom of the structure.
Let m denote the number of members. For design sensitivity analysis, where
the cross-sectional areas are considered as design variables, the stiffness matrix is
a function of A = (A1 , . . . , Am )> , which is written as K(A). If the self-weight is
considered, the nodal load vector P = (P1 , . . . , Pn )> is also a function of the design
variable vector A, which is written as P (A). The displacement vector U (A) =
(U1 , . . . , Un )> is obtained by solving the stiffness equation as
K(A)U (A) = P (A) (3.82)
In the following, the argument A is omitted for brevity.
Differentiation of Eq. (3.82) with respect to Ai leads to
K U P
U +K = (3.83)
Ai Ai Ai
Then the sensitivity coefficient of U with respect to Ai is obtained from
U P K
K = U (3.84)
Ai Ai Ai
Therefore, the sensitivity coefficients of P and K are needed to compute those of
the displacements.
By differentiating Eq. (3.81) with respect to Ai , we obtain the explicit relation
kgi E > >
= T dd T i (3.85)
Ai Li i
Hence, for a truss, the sensitivity coefficient of K with respect to Ai is easily
obtained by an arithmetic operation, because K is an assemblage of k gi . Sensitivity
of P is also easily computed, when the self-weight is considered, because P is an
explicit function of A. Note that K is decomposed to a product of triangular and
diagonal matrices by Cholesky decomposition in the process of computing U from
Eq. (3.82). Hence, the computational cost of solving Eq. (3.84) is very small even
for the case where the sensitivity coefficients with respect to all m design variables
A1 , . . . , Am are to be computed.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
U1 U2
P1 P2
A1 A2
Consider, for example, a 2-bar structure as shown in Fig. 3.5 subjected to axial
forces P1 and P2 at nodes 1 and 2, respectively. The displacements of nodes 1 and
2 are denoted by U1 and U2 , respectively. The two bars have the cross-sectional
areas A1 and A2 , and the same length L. Then Eq. (3.82) is written as
E A1 + A2 A2 U1 P1
= (3.86)
L A2 A2 U2 P2
which is easily solved for U1 and U2 as
(P1 + P2 )L (P1 + P2 )L P2 L
U1 = , U2 = + (3.87)
A1 E A1 E A2 E
Consider A1 as the design variable, and suppose that P1 and P2 are independent
of A1 . Equation (3.83) is then written as
U1
E 1 0 U1 E A1 + A2 A2 A1
0
+ = (3.88)
L 0 0 U2 L A2 A2 U2 0
A1
By incorporating Eq. (3.87) into Eq. (3.88), we obtain
U1 U2 (P1 + P2 )L
= = (3.89)
A1 A1 A21 E
which agrees with the direct differentiation of U1 and U2 in Eq. (3.87) with respect
to A1 .
The sensitivity coefficients can also be computed by the finite-difference ap-
proach. Let Ai denote the m-vector whose ith component is Ai and the other
components are 0, where Ai is the small variation of Ai . The approximate sensi-
tivity coefficients by the central finite-difference approach is given as
U U (A + Ai ) U (A Ai )
' (3.90)
Ai 2Ai
The approximate sensitivity coefficients can also be found by the forward finite-
difference approach
U U (A + Ai ) U (A)
' (3.91)
Ai Ai
or the backward finite-difference approach
U U (A) U (A Ai )
' (3.92)
Ai Ai
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Although these finite-difference approaches are very simple, the number of response
analysis (solution process of Eq. (3.82)) for sensitivity evaluation with respect to
m design variables is m for the forward or backward finite-difference approach and
2m for the central finite-difference approach. Therefore, it is seen that the compu-
tational cost for sensitivity analysis is drastically reduced by direct differentiation
of the governing equations.
Uj
reanalysis
sensitivity analysis
Ai Ai +Ai
1996). Kirsch (2000) extended his method to the combined reanalysis for nonlinear
static analysis and eigenvalue analysis.
The virtual distortion method has been developed for static and dynamic analy-
ses of trusses (Putresza and Kolakowski, 2001). Deng and Ghosn (2001) developed
a pseudo force method for nonlinear analysis and reanalysis. Bickford (1987) pre-
sented a refined higher-order perturbation method for eigenvalues and eigenvectors
of a generalized eigenvalue problem. Chen, Yang and Lian (2000) compared several
reanalysis methods for eigenvalue analysis. Chen, Wu and Yang (2006) presented
a reanalysis method for eigenvalue analysis allowing large parameter modification.
Makode, Corotis and Ramirez (1999) presented a pseudodistortion method for static
nonlinear analysis of trusses.
From a mathematical point of view, structural reanalysis can be regarded as
a process of finding the inverse of the stiffness matrix of the modified structure.
Calculation of the inverse of a modified matrix has been discussed in many fields
of engineering and mathematics. Developments in general form of inverting modi-
fied matrices are summarized in the review article by Henderson and Searle (1981).
However, it is important to note, for application to the static analysis problem of
structures, that the inverse of the stiffness matrix of the initial structure is not
usually known; i.e., the matrix is only decomposed to a product of triangular and
diagonal matrices in the process of structural analysis. Kavlie, Graham and Powell
(1971) developed a stiffness-based method for computing the displacements of the
modified design based on the ShermanMorrisonWoodbury (SMW) formula (Sher-
man and Morrison, 1950; Woodbury, 1950). Akg un, Garcelon and Haftka (2001)
extended the SMW formula to the nonlinear reanalysis problem.
Ohsaki (2001) showed that the displacements of the modified structures are
found without computation of the inverse of the stiffness matrix of the initial system;
i.e., only the Cholesky decomposition is necessary. We present below a method of
exact reanalysis for the static responses of trusses.
KU 0 = P (3.105)
where K i is the nn stiffness matrix for unit cross-sectional area of the ith member.
Note that the member properties are defined by the matrices and vectors of the size
n for the convenience of presenting the matrix operation. Computation is carried
out, however, by element-size matrices and vectors.
Consider the case where the cross-sectional area of the kth member is increased
by Ak . Equation (3.105) for the modified truss is formulated as
(K + Ak K k )U 1 = P (3.107)
The objective here is to find the displacement vector U 1 of the modified truss for
the given load vector P . In the following, the superscripts ( )0 and ( )1 denote the
values of the initial and the modified structures, respectively.
There have been a variety of studies on computing the inverse of the modified
matrix from the known inverse of the initial matrix. For our purpose, however,
those formulas cannot be directly used, because the inverse of K is not usually
known; i.e., K is only decomposed for computing U 0 . Therefore, a more explicit
formula is needed for computing U 1 of the modified truss.
Let N = (N1 , . . . , Nm )> denote the vector of axial forces. The relation between
Ni and U 0 is written by using an n-vector ci as
0
0
Ni0 = Ai c>
i U (3.108)
Let E and Li denote Youngs modulus and the length of the ith member, and define
vector bi as
Li
bi = ci (3.109)
E
Then the relation between Ni0 and the n-vector of equivalent nodal forces F 0i of the
ith member is written as
F 0i = Ni0 bi (3.110)
Hence, bi represents the nodal load vector corresponding to unit axial force of the
ith member. The matrix K i is defined by using ci and bi as
K i = b i c>
i (3.111)
Note from Eqs. (3.109) and (3.111) that the rank of K k is 1. The inverse of the
modified matrix K + Ak K k , when rank K k = 1, is obtained from (Henderson
and Searle, 1981)
1 1
(K + Ak K k ) = K + Ak bk c> k
Ak K 1 bk c> 1 (3.112)
kK
= K 1 > 1
1 + Ak ck K bk
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where Eq. (3.108) is used, and k is the strain of member k. Note that U k and k k
can be computed through simple arithmetic operation, because K has been already
decomposed in the process of computing U 0 . This way, U 1 can be obtained without
using K 1 . The strain of the member with null cross-sectional area may also be
obtained from the displacements of the nodes connected to the member. Therefore,
Eq. (3.114) can be used for addition or removal of a member between two existing
nodes.
Next, we consider a case where the cross-sectional areas of q ( 2) members
are modified simultaneously. It is assumed without loss of generality that Ak (k =
1, . . . , q) are modified. The stiffness matrix of the modified truss is formulated using
Eqs. (3.109) and (3.111) as
q
X q
X Ak E
K+ Ak K k = K + bk b>
k
Lk (3.115)
k=1 k=1
= K + BDB >
where the ith column of the n q matrix B is bi , and D is a q q diagonal matrix
defied as
D = diag(A1 E/L1 , . . . , Aq E/Lq ) (3.116)
The inverse of the modified stiffness matrix is written as (Henderson and Searle,
1981; Sherman and Morrison, 1950; Woodbury, 1950)
(K + BDB > )1 = K 1 K 1 B(D 1 + B > K 1 B)1 B > K 1 (3.117)
By post-multiplying P to Eq. (3.117), the following relation is derived:
U 1 = U 0 K 1 B(D 1 + B > K 1 B)1 DB > U 0
(3.118)
= U 0 U (D1 + B > K 1 B)1 y
where the ith component of the q-vector y is 0i Li , and the ith column of the n q
matrix U is U i .
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Ak , Nkk
1 1
Ak , Nk
Fig. 3.9 Initial truss subjected to P and Fig. 3.10 Modified truss without virtual
k
virtual load k P . load.
k
loads P and k P shown in Fig. 3.9 should be the same as that of the modified
truss without the additional virtual loads. Hence, the pair of loads are transformed
back to the axial force k as shown in Fig. 3.10, and we obtain the following relation
from the equivalence of the strain of member k:
Lk k Lk k
(Nk + k N k ) = (Nk + k N k + k ) (3.129)
Ak E (Ak + Ak )E
which results in
Ak Nk
k = k
(3.130)
Ak N k Ak
k
Since N k is the axial force of member k against a pair of loads corresponding
to the unit axial force in member k, the following relation holds for a statically
indeterminate truss:
k
1 < N k < 0 (3.131)
and the denominator of Eq. (3.130) is positive for Ak Ak ; i.e., the value
of k for Ak + Ak 0 that includes removal of member k is obtained from
Eq. (3.130). Accordingly, the displacements, axial forces, strains, etc., can be com-
k
puted by applying the loads P + k P to the initial truss with cross-sectional areas
A. It is easily observed that k in Eq. (3.128) is equal to the coefficient for U k
in Eq. (3.118) derived from the formula of the inverse of a modified matrix. Hence,
the mathematical and mechanical formulations lead to the same result.
k
For a statically determinate truss, N k = 1 and the strain of member k of the
modified truss is obtained from Eq. (3.128) as
1 k
kk = (Nk + k N k )
Ak E
1
= (Nk k ) (3.132)
Ak E
Nk
=
(Ak + Ak )E
which corresponds to the fact that the axial force is independent of the cross-
sectional areas.
As an illustrative example, consider again the 2-bar structure in Fig. 3.5 in
Sec. 3.6.2, with the same parameter values in the example in Sec. 3.7.2. The cross-
sectional areas of the initial truss is A1 = A2 = 1, and A1 is increased to 2. Then
we have
1 1 1 1 1
P = , N1 = N2 = 1, N 1 = 1, 1 = = (3.133)
0 1+1 2
and the strains of the modified truss are obtained from Eq. (3.128) as
1 1
11 = 1 = , 12 = 1 (3.134)
2 2
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
which is consistent with the displacements in Eq. (3.125) obtained by the mathe-
matical formulation.
In closing this section, it should be noted that the exact reanalysis method can
be very effectively used for enumerating the responses of the trusses corresponding
to the vertices of feasible regions of the uncertain cross-sectional areas. Direct
application of the reanalysis method to truss topology optimization can be found
in Ohsaki (2001).
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 4
Hope for the best but prepare for the worst. (English proverb)
If you think the worst, you wont be far wrong. (Spanish proverb)
The worst is not always certain but its very likely. (French proverb)
If anything can go wrong, it will, and at the worst possible time.
(Popular version of Murphys law)
In this chapter, we first present a simple static anti-optimization problem so that the
analytically minded researcher may follow it with pen, pencil, and of course with the
head. Later we explore the apparently first static anti-optimization study by Boley
(1966a) devoted to the worst temperature distribution through the thickness of the
plate, and we present several anti-optimization problems including the evaluation
of the worst possible hole that produces the maximum stress concentration factor.
The worst distribution of prestress of a tensegrity structure is found by vertex
enumeration of the feasible region of the uncertain parameters.
77
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
P2 , U2
P1 , U1
H
x
W1 W2
y
with
1 4 + 2 1 4 + 6
c1 = , c2 = (4.3)
AE(2 + 3) 4 + 6 AE(2 + 3) 4 + 3 2
We assume uncertainty in P as
P = P0 + a (4.4)
0
where P = (P10 , P20 )> >
is the nominal load vector and a = (a1 , a2 ) represents the
uncertainty.
Suppose, for simplicity, U2 is chosen as the performance measure, and our pur-
pose is to find the worst value of a that maximizes U2 , which is assumed to be
positive. Consider the case where the bound of a is given by the quadratic inequal-
ity constraint
a> a D (4.5)
as the simplest ellipsoidal model presented in Sec. 3.4, where D is a prescribed
value. Then the anti-optimization problem is formulated in the following form of a
quadratic programming problem:
maximize U2 (a) (4.6a)
>
subject to a a D (4.6b)
It is seen from Eqs. (4.2) and (4.4) that U2 is a linear function of a. Therefore, the
inequality constraint (4.6b) is satisfied with equality at the anti-optimal solution,
which is found explicitly as follows.
The Lagrangian Q of Problem (4.6) is defined as
Q(a, ) = U2 (a) + (D a> a) (4.7)
where 0 is the Lagrange multiplier. The stationary condition of Q with respect
to a leads to
U2
2ai = 0, (i = 1, 2) (4.8)
ai
from which we obtain
1
a= c2 (4.9)
2
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
a2
Anti-optimal solution
1 1 a1
where Eqs. (4.2) and (4.4) have been used. Since the constraint (4.6b) is satisfied
with equality at the anti-optimal solution, the Lagrange multiplier is obtained as
follows by incorporating Eq. (4.9) into Eq. (4.5), and using 0:
r
1 c>
2 c2
= = (4.10)
2 D
Note that = corresponds to the optimum (minimum) value of U2 .
If EA = 1 and D = 1, for simplicity, then c2 = (0.4155, 2.2086)>, and the
anti-optimization problem is formulated as
maximize U2 = 0.4155a1 + 2.2086a2 (4.11a)
subject to a21 + a22 1 (4.11b)
From Eq. (4.11a), a2 is written in terms of U2 and a1 as
a2 = 0.4528U2 + 0.1881a1 (4.12)
The gray area in Fig. 4.2 is the feasible region defined by Eq. (4.11b), and each
dashed line represents the solution with a constant value of U2 . The anti-optimal
solution is found as a feasible solution that attains the maximum intercept of the
a2 -axis. Since = 1.1237 is obtained from Eq. (4.10), the worst load-case is found
from Eq. (4.9) as a = (0.1849, 0.9828)>.
If the bounds are given for a by an interval
aL a a U (4.13)
with the lower bound aL and upper bound aU , the anti-optimization problem turns
out to be a linear programming problem:
maximize U2 (a) (4.14a)
L U
subject to a a a (4.14b)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
a2
Anti-optimal solution
(1, 1) (1, 1)
a1
(1, 1) (1, 1)
Boley (1966a,b) pioneered a study on the bounds of the thermoelastic stresses and
deflections in a beam or a plate. Its derivations will be reproduced here briefly as
the first work of anti-optimization.
The normal thermoelastic stress in a beam of arbitrary cross-section, e.g.,
rectangular cross-section as shown in Fig. 4.4, is given by the formula (Boley and
Weiner, 1960)
PT MT y
= ET (y, z) + + (4.16)
A I
where T (y, z) is the temperature, which is a function of the coordinates (y, z) in
the cross-section, PT is the thermal force, MT is the thermal moment, defined,
respectively, as Z Z
PT = ET dA, MT = ET ydA (4.17)
A A
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
y y
c
x z
c
and A is the cross-sectional area, I is the moment of inertia about the z-axis, E
is Youngs modulus, and is the thermal coefficient. Although Eq. (4.16) is valid
whether principal axes are used or not, we will assume that (y, z) is the principal
coordinate system for simplicity.
In his paper, Boley (1966b) poses the following question: What is the maximum
value of that can be caused in a given beam by arbitrary temperature distribu-
tions? He presumes that the temperature T (y, z) is bounded in the interval
Tm T (y, z) TM (4.18)
where Tm and TM are the constants such that TM Tm . The problem is to
determine the constant k such that
|| k(Tm + TM ) (4.19)
for all possible temperature distributions. The coefficient k has a physical sense of
the factor of proportionality between the maximum possible stress and the maxi-
mum temperature excursion Tm + TM .
The following question arises: Why is it useful to know the coefficient k? It
turns out that the temperature excursion can often be straightforwardly estimated,
especially in problems in which the body surface temperature is known. There-
fore, Eq. (4.19) permits the easy estimation of the maximum normal stresses. For
simplicity, we will deal with the beam of the rectangular cross-section as shown in
Fig. 4.4, which was studied by Boley (1966b) in an earlier paper, presented at a con-
ference in 1964. Accounting that the product E is constant, we rewrite Eq. (4.16)
as follows:
Z Z
() 1 1 3 1
= T () + T (0 )d0 + T (0 )0 d0 (4.20)
E 2 1 2 1
where = y/c with c being the half-depth of the beam as shown in Fig. 4.4. For a
special case
T0 , for 0 < < 0 +
T = 2 2 (4.21)
0, elsewhere
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
with a symmetric result for negative values of . It represents the maximum value,
either positive or negative, which can be expected in the stress. A comparison of
Eqs. (4.19) and (4.26) reveals that
1 1 1
2 + 3 + , for < 1
kmax = 4 3 3 (4.28)
1, 1
for 0
3
A less accurate inequality for kmax is
3 , for 1 < || < 1
kmax 4 3
1 (4.29)
1, for 0 ||
3
or even
4
kmax (4.30)
3
The above formulas allow us to estimate the maximum stress that can develop
in a beam of rectangular cross-section under any temperature distribution that
satisfies the inequality (4.18). If the temperature is a positive interval value, such
that
0 T Tmax (4.31)
then
4
|| ETmax (4.32)
3
Boley (1966b) also derived a more refined bound for the case when the tem-
perature distribution is symmetric or antisymmetric with respect to y. For the
symmetric case, T () = T (), and Eq. (4.23) turns out to be
Z 1
= T () + T (0 )d0 (4.33)
E 0
The following formula for the maximum stress is obtained
||max E(Tm + TM ) (4.34)
Thus, in this case
kmax 1 (4.35)
instead of the inequality kmax 4/3 in Eq. (4.30). Clearly, by gaining additional
information on the symmetry in distribution of the temperature, a smaller upper
bound is derived as expected.
The situation is analogous if the information is obtainable on the antisymmetry
of the temperature distribution, T () = T (). Then the stress becomes
Z 1
= T () + 3 T (0 )0 d0 (4.36)
E 0
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
TM T () TM , T () = T () (4.38)
Boley (1966b) derives the following final expression for this case:
5
||max ETM (4.39)
2
In order to compare with Eq. (4.27), Eq. (4.39) is rewritten as
5
||max E(2TM ), T () = T () (4.40)
4
Therefore, for the antisymmetric case, the inequality for kmax results in
5
kmax (4.41)
4
for which the upper bound is again smaller than its counterpart 4/3 in Eq. (4.30).
In the second paper, Boley (1966a) derived the expression for kmax for the gen-
eral case of the cross-section. For example, for the solid circular cross-section and
arbitrary temperature distribution
1 11 15 1 1
kmax + + sin1 ' 1.428 (4.42)
2 16 4
The general formulations presented in Chap. 3 are readily used for simple static
responses such as nodal displacements and member stresses of trusses.
Pantelides and Ganzerli (1997) found anti-optimal solutions of trusses consider-
ing uncertainty in the direction and magnitude of nodal load vector using ellipsoidal
bounds on parameters as
XN 2
ai
1 (4.45)
i=1
gi
where a = (a1 , . . . , aN )> is the vector of uncertain parameters, and the value of
the semi-axis gi is defined by the minimum volume method as shown in Sec. 3.4.
The static response Z(a) that represents nodal displacement or member stress is
linearly estimated from the nominal values a0 = (a01 , . . . , a0N )> with ai = ai a0i
as
XN
Z(a0 )
Z(a0 + a) = Z(a0 ) + ai (4.46)
i=1
ai
using the sensitivity analysis in Sec. 3.6.2 or the perturbation approach in Sec. 4.4.
The worst parameter values are obtained from the stationary conditions of the
Lagrangian of the anti-optimization problem as demonstrated in Sec. 4.1.
The elastic responses such as member stress and nodal displacement are often
chosen as the objective function to be maximized. The compliance defined as the
external work (twice of the strain energy) is also often taken as the global per-
formance measure (see Secs. 9.69.8 for examples of anti-optimization for static re-
sponses considering stress, compliance and homology design, respectively, which are
demonstrated as the lower-level problem of hybrid optimizationanti-optimization).
Cherkaev and Cherkaev (2003) defined principal compliance as the maximum value
of the compliance against the admissible loading scenarios. Barbieri, Cinquini and
Lombardi (1997) used second-order approximation for anti-optimization of trusses.
Other studies on static anti-optimization include papers by Alotto, Molfino and
Molinari (2001), Attoh-Okine (2002, 2004), Babuska, Nobile and Tempone (2005),
Banichuk (1975, 1976), Banichuk and Neittaanm aki (2007), Barbieri, Cinquini and
Lombardi (1997), Ben-Haim (1992, 1999), Ben-Tal and Nemirovski (1998, 2002),
Chen, Lian and Yang (2002), Chen and Ward (1990), Cho and Rhee (2003, 2004),
Corliss, Foley and Keafott (2007), Dzhur, Gordeyev and Shimanovsky (2001), El-
ishakoff, Gana-Shvili and Givoli (1991), Hlav acek (2002a, 2007), Hlavacek, Chle-
boun and Babuska (2004), Kanno and Takewaki (2006b,c, 2007a,b), K oyl
uo
glu,
C akmak and Nielsen (1995), Kulpa, Pownuk and Skalna (1998), Liu, Chen and
Han (1994), Lombardi (1995, 1998), Lombardi and Haftka (1998), Lombardi, Mar-
iani and Venini (1998), McWilliam (2001), Mullen and Muhanna (1999), Muhanna
and Mullen (2001, 2005), Neumaier and Pownuk (2007a), Popova, Datcheva, Iankov
and Schanz (2003, 2004), Qiu (2003), Qiu, Chen and Song (1996c), Qiu, Chen and
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Wang (2004a), Qiu, Ma and Wang (2004d), Qiu, Wang and Chen (2006b), and
Skalna (2003, 2006).
Anti-optimization in the context of comparison of alternative structural models
was proposed by Haftka and his co-workers; see, e.g., the paper by Gangadharan,
Nikolaidis, Lee and Haftka (2003) and Lee, Haftka, Griffin, Watson and Sensmeier
(1994).
In this section, we follow Qiu, Chen and Song (1996c) for the perturbation method
to solve the static problem in the matrix, or finite element setting
KU = P (4.47)
where K is the stiffness matrix, P is the load vector, and U is the displacement vec-
tor. The perturbation theory deals with the behavior of a structure that is subjected
to a small perturbation in its variables. The perturbation may occur in the stiffness
matrix as K + K and/or in the load vector, which takes the form P + P . We
are interested in the change of behavior of the system under perturbations. The
perturbation terms K and P will represent for us the ever-present uncertain-
ties. We suppose that the solution U = K 1 P of the unperturbed structure are
known. Note that Eq. (4.47) is solved without computing K 1 ; however, the form
U = K 1 P is used for the expression purpose. Note that the sensitivity analysis
in Sec. 3.6.2 is equivalent to the first-order perturbation.
The solution in the presence of perturbation is denoted as U new . Equation (4.47)
is replaced by
(K + K)(U + U ) = P + P (4.50)
KU (1) = P KU (4.51a)
KU (2) = KU (1) (4.51b)
..
.
KU (n) = KU (n1) (4.51c)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
2P
1 4
P 2 (2) 3
1.5P
0.8 L
y
x
(4)
0.6 L
L U
U(1),1 = 0.00087, U(1),1 = 0.00093,
L U
U(1),2 = 0.00032, U(1),2 = 0.00034,
L U (4.68)
U(1),3 = 0.00091, U(1),3 = 0.00097,
L U
U(1),4 = 0.00033, U(1),4 = 0.00031
L U
where U(1),i and U(1),i are the lower and upper bounds for the first-order term
of Ui , respectively. As is seen, the uncertainty in interval displacements is very
c L U
small. Indeed, the central value U(1),1 = (U(1),1 + U(1),1 )/2 = 0.00090, whereas
U L
the uncertainty parameter U(1),1 = (U(1),1 U(1),1 )/2 equals 0.00003, with the
c
deviation parameter U(1),1 /U(1),1 being 0.03.
Other examples of interval analysis of structures were studied by Chen, Lian
and Yang (2002), Dimarogonas (1995), K oyl
uo
glu, C akmak and Nielsen (1995),
Kulpa, Pownuk and Skalna (1998), McWilliam (2001), Mullen and Muhanna (1999),
Neumaier and Pownuk (2007a, 2007b), Qiu, Chen and Song (1996c), Qiu and Gu
(1996), Qiu, Chen and Wang (2004a), Qiu, Ma and Wang (2004d), Qiu, Wang and
Chen (2006b), Rao and Berke (1997), and Wang and Li (2003). Ellipsoidal modeling
was employed by Kanno and Takewaki (2006a), Rao and Majumder (2008), Qiu,
Ma and Wang (2006a), Qiu, M uller and Frommer (2001), and Qiu, Wang and Chen
(2006b). Comparison of static response estimation by interval and ellipsoidal models
was conducted by Qiu (2003).
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
4.5.1 Introduction
Stress concentrations in the vicinity of holes of various profiles have been treated
in a number of publications. For a pertinent bibliography one may consult Neuber
(1958), Peterson (1974), and Savin (1961). Traditionally, the stress concentration
factors (SCF) around holes of ideal shapes have been considered. However, in prac-
tice the manufacturing process of structural parts inevitably produces imperfections
which may affect the actual value of the SCF. Irregularities in ideally shaped holes
have been modeled using a probabilistic description by Palmov (1964). The profile
of the hole as a random function was also treated by Sayles (1982). Lomakin and
Sheinin (1970, 1974) obtained stress concentration at a boundary of a randomly
inhomogeneous elastic body.
Lomakin (1968) developed a double perturbation scheme for a hole with a rapidly
oscillating boundary. The boundary was characterized as a random stationary Gaus-
sian function with zero mean and with a specified autocorrelation function. The
stress concentration factor was determined, as suggested by Palmov (1964), from
the condition that < E( ) holds in the region 0 2 with specified
probability. Here is the circumferential stress at the boundary (which turns
out to be a random function too), and E( ) is its mathematical expectation. If
one has a complete probabilistic description of the irregularity of the hole bound-
ary, Lomakins approach is an attractive method to examine its effect on the stress
concentration. Yet in many cases only a limited amount of information on the
irregularities may be available.
In this section, we present the results in Givoli and Elishakoff (1992) for stress
concentration at the boundary of a nearly circular hole in an infinite elastic plane
under uniform radial tension at infinity. We consider a realistic situation, where we
do not possess the full probabilistic information, but only have a partial knowledge
with regard to the shape of the imperfection around the hole boundary. We define
an admissible imperfection profile as one which is bounded in some sense, and which
may be required to have (depending on the boundedness restriction of the profile)
a finite number of harmonics in its Fourier expansion. Then we find the worst-case
possible imperfection profile that gives, among all admissible profiles, the maximum
SCF.
The technique used here to obtain the SCF for a circular hole with arbitrary ir-
regularities is the boundary perturbation method. Some early works employed other
methods to treat holes of arbitrary shapes (Ishida and Tagami, 1959; Averin, 1964);
however, those methods are less amenable for the anti-optimization performed as a
second step.
In Sec. 4.5.2, we find a first-order asymptotic approximation of the SCF at the
boundary of a nearly circular hole. The results are compared with the exact solution
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 4.6 Unit vectors on the ideal circular boundary B and on the imperfect boundary B .
circle B. The resulting boundary conditions are written as follows up to the second
order:
(0)
rr (R, ) = 0 (4.76a)
0
(1) h (0) (0)
rr (R, ) = r hrr,r (4.76b)
R
(2) h0 (1) (1) 1 (h0 )2 (0)
rr (R, ) = r hrr,r h2 rr,rr
(0)
+
R 2 2R2 rr
hh0 (0) hh0 (0)
+ r,r 2 r (4.76c)
R R
(0)
r (R, ) = 0 (4.76d)
0
(1) h (0) (0)
r (R, ) = hr,r (4.76e)
R
2 h0 (1) (1) 1 (0) (h0 )2 (0)
r (R, ) = hr,r h2 r,rr +
R 2 2R2 r
0 0
hh (0) hh (0)
+ ,r 2 (4.76f)
R R
In Eqs. (4.76a)(4.76f), all the functions on the right-hand-side are evaluated at
r = R.
Note that Eqs. (4.76a), (4.76d) and (4.76e) can be used to simplify Eqs. (4.76b),
(4.76c) and (4.76f) as
(1) (0)
rr (R, ) = hrr,r (4.77a)
0 2
h (0) (1) 1
(2)
rr (R, ) = hrr,r h2 rr,rr
(0)
(4.77b)
R 2
2 h0 (1) (1) 1 (0) hh0 (0) hh0 (0)
r (R, ) = hr,r h2 r,rr + ,r 2 (4.77c)
R 2 R R
Equations (4.72a)(4.72c) and (4.76a)(4.76f) define a sequence of problems in do-
main . In what follows we solve the first two problems in this sequence; namely,
we find a first-order asymptotic approximation. The zeroth-order problem is ax-
isymmetric and is easily solved as
(0) R2
rr = 1 2 (4.78a)
r
(0) R2
= 1 + 2 (4.78b)
r
(0)
r = 0 (4.78c)
Using this solution as well as Eqs. (4.72a), (4.72c), (4.76b) and (4.76e), we conclude
(1)
that ij should satisfy
(1) (1)
rr = r = 0 at r (4.79a)
(1) 2
rr (R, ) = h() (4.79b)
R
(1) 2 0
r (R, ) = h () (4.79c)
R
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(1)
The solution for can be written in the form
Z
(1) 2 X 2
= [Brn (r, , 0 )h(0 ) + Bn (r, , 0 )h0 (0 )]d (4.80)
R n=0 0
where Brn and Bn are the kernels; see Givoli and Elishakoff (1992) for details.
Now Eqs. (4.78c) and (4.80) provide us with an approximate solution on the
circular boundary B. We use Eqs. (4.71) and (4.75) again to obtain a solution on
the actual boundary B :
(0) (0) (1)
ij (r , ) = ij (R, ) + [h()ij,r (R, ) + ij (R, )] + O(2 ) (4.81)
which leads to an expression for on B . However, the SCF is defined via the
circumferential stress tt on B rather than via ; see Fig. 4.6 for the difference
between t and e . To find tt , we transform the stresses according to the formula
where is the angle between er and n, and cos = nr , sin = n . Then, from
Eqs. (4.73a), (4.73b), (4.78c) and (4.80)(4.82), the stress concentration factor
along B is obtained as
X Z 2
tt (r , )
' =1 h() + [Brn (R, , 0 )h()
2 2 R n=0 0
(4.83)
n 0 0 2
B (R, , )h ()]d + O( )
which leads to
" Z 2 #
1X
= 2 1 g() + n cos n( )g()d (4.84a)
n=0 0
h()
g() = (4.84b)
R
1
0 = , n = 2n 1 for n 1 (4.84c)
2
It is easy to verify that if g() is constant, corresponding to an ideal circle, then
= 2. Moreover, even when g() = cos or g() = sin , then Eq. (4.84a) yields
the constant value = 2. This implies that, for these boundary profiles, the effect
of the deviation from the ideal circle on the SCF does not manifest itself in the
first-order perturbation term.
We now compare our result to that of Tunguskova (1970), who presented an exact
expression for the SCF around a hole in the shape of a hypotrochoid (Lockwood,
1961). In cartesian coordinates, this curve is defined by the parametric equations
where k is a non-negative integer and 0 < k < 1. The polar coordinates on this
curve can easily be written as a function of the parameter t, and be expanded in
powers of as
sin(k + 1)t
tan = tan t + O(2 ) (4.86a)
cos2 t
r = R + R cos(k + 1)t + O(2 ) (4.86b)
One also can see that cos(k + 1)t = cos(k + 1) + O(). Comparing Eq. (4.86b) with
Eq. (4.69) and using Eq. (4.84b), we obtain
g() = cos(k + 1) (4.87)
Figure 4.7 describes a hypotrochoid with k = 7 and = 0.01.
Tunguskovas exact stress concentration factor around the hole is
2(1 2 k 2 )
= (4.88)
1 2k cos(k + 1)t + 2 k 2
with a maximum value of
2(1 + k)
max = (4.89)
t 1 k
On the other hand, we obtain as follows from our asymptotic solution (4.84a)
after substituting the function g() given in Eq. (4.87) and using the orthogonality
property of the trigonometric functions:
= 2 [1 + 2k cos(k + 1)] (4.90)
with a maximum value
max = 2(1 + 2k) (4.91)
t
Note that Eqs. (4.90) and (4.91) can be obtained from Eqs. (4.88) and (4.89),
respectively, by using Taylor series expansion up to the first order in .
In Fig. 4.8, we compare the exact solution (4.89) (solid line) and the asymptotic
solution (4.90) (dashed line) for k = 7 and = 0.01. The agreement between the
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 4.8 A comparison between Tunguskovas exact SCF (solid line) and the present asymptotic
SCF (dashed line) for the hypotrochoid of Fig. 4.7.
two solutions is excellent. An easy calculation shows that the relative difference
between the maximum values of and is
max max 2(k)2
= (4.92)
max 1 + k
This relative error as a function of k is plotted in Fig. 4.9. With the choice of
k = 7 and = 0.01, the relative error amounts to only 0.9%. Increasing to
0.035 we have an error of about 10%. This may raise the question of whether
our first-order asymptotic solution (4.84a) is satisfactory. We claim that in many
practical situations, it is satisfactory, since k is very small. We are interested in
small irregularities, and an irregularity characterized by = 0.01 is quite usual in
practice. As for the value of k, we confine ourselves to the case where k is of order
10 at the most. If k is large, say 100, we may use Lomakins asymptotic solution
(Lomakin, 1968), which is valid only for this case of a rapidly oscillating boundary
profile. In this sense, our solution and that of Lomakins are complementary. For the
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
M
X
g() = (Am cos m + Bm sin m) (4.93)
m=0
We also assume that the irregularities are bounded in the L2 sense, namely,
Z 2
g 2 ()d (4.94)
0
where is a given constant. The set of admissible profiles g() consists of those
functions which are of the form Eq. (4.93) and satisfy Eq. (4.94).
We find g() which maximizes in Eq. (4.84a) at a given angle = 0 , which
will turn out later to be arbitrary. Substituting Eq. (4.93) in Eq. (4.84a) and setting
= 0 result in
XM
(0 )
I' =1+ 2(m 1)(Am cos m0 + Bm sin m0 ) (4.95)
2 m=1
M
X
2A20 + (A2m + Bm
2
) (4.96)
m=1
M
X
J ' 2A20 + (A2m + Bm
2
) + p2 =0 (4.97)
m=1
where p is considered as a new unknown. Now the problem is to find the coefficients
Am and Bm that will maximize I subject to the constraint J = 0.
This problem is easily solved by the Lagrange multiplier method. The unknowns
are A0 , Am , Bm (m = 1, 2, . . . , M ), p and the Lagrange multiplier . The necessary
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 4.10 The worst-case profile g() for the parameters M = 10 and = 10 4 .
Fig. 4.11 The worst-case profile g() for the parameters M = 5 and = 10 4 .
It is interesting
R 2 to note that the average of around the boundary of the hole,
(1/2) 0 ()d, is always 2, i.e., it is equal to the SCF for a circle.
Figure 4.10 is the graph of the profile function g() for the parameters M = 10
and = 104 . The angle 0 was chosen to be . Any other choice for 0 would just
shift the plot horizontally. The maximum value of g, namely, the maximum value
of imperfection divided by R, is 0.015 at 0 . Figure 4.11 shows the corresponding
distribution of around the hole boundary. The maximum SCF is 2.38, namely,
19% higher than the SCF associated with the ideal boundary. Figure 4.12 is the
graph of g() for the parameters M = 5 and = 104 . Here, the maximum value
of g is 0.010, and the maximum corresponding SCF is now 2.12, or 6% larger than
that for a circle.
It is instructive to compare the SCF obtained by the worst profile with the one
obtained by a hypotrochoid, as found by Tunguskova (1970). The maximum stress
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 4.12 The distribution of the SCF around the worst-case profile.
For M = 10 and = 104 , the minimum SCF is 1.62. The corresponding profile
is the same as the one in Fig. 4.10, but with an opposite sign. The maximum
obtained for this profile is 2.30.
Up to this point, the set of admissible imperfection profiles g() was defined
via the bound constraint (4.94) and the requirement that g() has a finite number
of harmonics. The latter requirement has several deficiencies. First, it is a very
strong requirement which causes many otherwise reasonable profiles to be excluded.
Second, it may be impractical for the designer to determine how many harmonics
are needed in the analysis. Third, and perhaps most important, the SCF result
obtained above (see Eqs. (4.101) and (4.103)) grows without limit as the number
of harmonics M increases. This is expected since Eq. (4.94) does not pose any
restriction on the curvature of the profile, and so the worst-case profile tends to a
saw-type function as M tends to infinity.
To overcome these deficiencies, we now replace Eq. (4.94) by a constraint on the
curvature of the profile, i.e.,
Z 2
2
(g 00 ()) d (4.108)
0
where is a given positive number. This enables us to go again through the deriva-
tion while totally eliminating the requirement that g() has a finite number of
harmonics. The consequence of this derivation for the maximum SCF is found to
be (cf. Eq. (4.103)) r
=2 1+2 (4.109)
where
X
(m 1)2
= = 0.3230 . . . (4.110)
m=2
m4
Thus, with the constraint of the form Eq. (4.108), we are able to sum the series
in Eq. (4.109), whereas the series in Eq. (4.101), which is based on the constraint
(4.94), does not converge as M tends to infinity.
Both constraints (4.94) and (4.108) can be viewed as special cases of the general
constraint Z
2
[0 g 2 + 1 (g 0 )2 + 2 (g 00 )2 + + L (g (L) )2 ]d (4.111)
0
where i are the given real constants. The maximum SCF in this general case is
again given by Eq. (4.109), where the expression (4.110) for is replaced by
X
(m 1)2
= (4.112)
+ 1 m2 + 2 m4 + + L m2L
m=2 0
The sum in Eq. (4.112) converges if and only if at least one of the coefficients i for
i 2 is nonzero. Thus, in order that no requirement on the number of harmonics of
g() would be needed, the second derivative and/or higher derivatives of g() must
be constrained.
The generalization of this section to the orthotropic case was given by Gana-
Shvili (1991).
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 4.13 A tensegrity structure; thick line: strut, thin line: cable.
4.6.1 Introduction
A tensegrity structure consists of cables and struts that can transmit only tensile
and compressive forces, respectively, and it is stiffened by prestresses in the self-
equilibrium state. Figure 4.13 illustrates a simple example of a tensegrity structure,
where thick and thin lines represent struts and cables, respectively. The term tenseg-
rity was invented by Buckminster Fuller as a contraction of tension and integrity,
although the inventor of the structural system is under controversy (Fuller, 1975;
Lalvani, 1996). By the rigorous definition, a tensegrity structure is free-standing
without any support, and the struts are connected by continuous cables and do not
have contact with each other (Motro, 1992).
Since the tensegrity structure is unstable in the absence of prestresses, the shape
and the stability at the self-equilibrium state strongly depend on the member forces
(Zhang and Ohsaki 2006; 2007). Therefore, there exist many difficulties in the
design of shape and forces of tensegrity structures, and many methods have been
presented for this process.
Tensegrity structures usually have several modes of self-equilibrium forces that
are governed by a set of linear equations with respect to the member forces, and
the member forces are defined as a linear combination of the self-equilibrium force
modes. Additional requirements exist for the member forces, because the cables
and struts can transmit only tensile and compressive forces, respectively. Further-
more, the upper bounds should be given for the cable forces so that the stress
remains within the yield stress divided by the safety factor. A bound for compres-
sive stress should also be given for a strut to prevent buckling. Since these bound
constraints are expressed as linear inequalities with respect to the coefficients of the
self-equilibrium force modes, the feasible region of the coefficients forms a convex
region.
The performance of the force design defined by the feasible or admissible set of
coefficients should be determined based on the mechanical properties such as the
eigenvalues of the tangent stiffness matrix and the displacements against external
loads. However, the member forces are very sensitive to the imperfection or variation
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
of the geometrical parameters such as the unstressed lengths of members and nodal
locations. Therefore, the anti-optimal solution of self-equilibrium forces is defined as
the member forces that minimize the performance measures under bound constraints
of the forces.
If those performance measures have monotonicity and/or convexity properties
with respect to the member forces, the anti-optimal solutions can be found by
searching the vertices of the feasible region of the member forces. In this section, we
present an anti-optimization method of the forces of tensegrity structures (Ohsaki,
Zhang and Elishakoff, 2008a). The lowest eigenvalue of the tangent stiffness matrix
is considered as the performance measure (Ohsaki, Zhang and Ohishi, 2008b). The
properties of the objective functions are investigated, and the anti-optimal solution
is found by the enumeration of the vertices of the convex feasible region.
The nodes are classified into n free nodes and nf fixed nodes (supports); i.e.,
n = n + nf . Suppose the nodes are numbered such that the free nodes precede the
s
fixed nodes. Then C s is divided into the m n matrix C and the m nf matrix
C f corresponding to the free and fixed nodes, respectively, as
C s = (C, C f ) (4.114)
Let s = (s1 , . . . , sm )> denote the vector of member forces. In the state of self-
equilibrium, the equilibrium equation of a pin-jointed structure in three-dimensional
space is written as (Schek, 1974)
Ds = 0 (4.118)
where the 3n m equilibrium matrix D is defined by
C>Hx
D = C > H y L1 (4.119)
C>H z
DRi = 0 (4.124)
s = 1 g 1 + + q g q
(4.125)
= G
si = b >
i , (i = 1, . . . , m) (4.126)
The self-equilibrium equation with respect to member forces s in Eq. (4.118) can be
rewritten in the following form with respect to the nodal coordinates (Zhang and
Ohsaki, 2006):
Ex + E f xf = 0
Ey + E f y f = 0 (4.127)
f f
Ez + E z = 0
K = KE + KG (4.129)
with
where I 3 is the 3 3 identity matrix, and the ith diagonal component K ii of the
diagonal matrix K is the stiffness of the ith member; i.e., K ii = Ai Ei /Lii where Ai ,
Ei and Lii are the cross-sectional area, Youngs modulus and the length of member
i, respectively.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
force modes are fixed, and the variation of the member forces is investigated in the
space of the coefficients .
We first design the nominal value s0 = G0 of the self-equilibrium forces. The
lower and upper bounds sL and sU , respectively, are appropriately given for s0 to
assign the constraints as
sL s 0 s U (4.143)
where the signs of the forces for cables and struts are to be considered; i.e., the
lower bounds for the cables and the upper bounds for the struts are 0.
The vertices of the feasible region satisfying the constraint (4.143) are enumer-
ated in the space of . Denote by 0 the mean value of at all the vertices called
the center of the feasible region (Ohsaki, Zhang and Ohishi, 2008b), for brevity,
which is conceived as the nominal value of the uncertain coefficient vector for the
self-equilibrium force modes. Note that the nominal value need not be the center
of the feasible region, and can be defined arbitrarily, e.g., based on the preference
of the designer.
Next we assign the range of uncertainty of s as an interval
s0 s s s0 + s (4.144)
where si is the maximum possible increase or decrease of si . Note that the range
is given for the force vector s, although the independent parameters for the forces
are .
In the following anti-optimization problem, the deviation of the forces from the
desired values is chosen as one of the performance measures. Then the deviation e
of from the desired value 0 is defined as
e = ( 0 )> G> G( 0 ) (4.145)
which is a convex function of .
We consider the following three performance measures so that the anti-
optimization problem is formulated as minimization of a concave function, or max-
imization of a convex function, and the anti-optimal solution exists at a vertex of
the feasible region defined by linear inequalities:
(1) Lowest eigenvalue 1 of K after constraining the rigid-body motions, which is a
concave function of , and is to be minimized in the anti-optimization problem.
(2) External work called compliance against the specified load after constraining
the rigid-body motions, which is a convex function of , and is to be maximized
in the anti-optimization problem.
(3) Deviation e of the forces from the desired value, which is a convex function of
, and is to be maximized in the anti-optimization problem.
For example, the anti-optimization problem of minimizing 1 is formulated as
minimize 1 () (4.146a)
subject to s0 s s s0 + s (4.146b)
s = G (4.146c)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Step 1 Specify the geometry, topology, and material property of the tensegrity
structure.
Step 2 Construct the equilibrium matrix D from Eq. (4.119) and compute the self-
equilibrium force modes g 1 , . . . , g q by the singular value decomposition (4.123).
Step 3 Assign bound constraints (4.143) for the nominal value of the self-
equilibrium forces.
Step 4 Generate the list of the vertices of the feasible region of satisfying (4.143).
Step 5 Find the center 0 of the feasible region that is regarded as the nominal
value of .
Step 6 Specify s, and generate the list of the vertices of the feasible region of
for the anti-optimization problem satisfying (4.144).
Step 7 Assign the support conditions, and compute the tangent stiffness matrix
K from Eq. (4.129) at each vertex.
Step 8 Compute the lowest eigenvalue 1 from Eq. (4.131), the compliance W from
Eq. (4.137), and the force deviation e from Eq. (4.145) at each vertex of the
feasible region satisfying Eq. (4.144), then find the three vertices, each of which
corresponds to the anti-optimal solution that minimizes 1 , maximizes W , or
maximizes e.
In the following numerical example, the vertices of the feasible region are enu-
merated by cdd+ (Fukuda, 1999; Avis and Fukuda, 1996) based on the efficient
procedure called reverse search. cdd+ can enumerate the vertices and the asso-
ciated active constraints of the region defined by linear inequality and equality
constraints.
22
400
5 10 9
300
20 17 18
14
21
4
3 y 200
7
8
12
19 15 16 13 100
1 6 0
2
0 100 200 300
11 x
(a) diagonal view (b) plan view
Fig. 4.14 Tensegrity grid.
the units of length and force are mm and N, respectively, which are omitted for
brevity. The rank r of the equilibrium matrix D is 59. Therefore, the structure
has q = 63 59 = 4 self-equilibrium force modes denoted by g 1 , . . . , g 4 with the
coefficients = (1 , . . . , 4 )> .
The lower bound sLi and the upper bound sU i for the axial forces of the struts
and cables, respectively, are 10000 and 10000. Therefore, the maximum absolute
values of the strains are 0.001 for the struts and 0.01 for the cables in the process
of designing the nominal values of the self-equilibrium forces. The vertices of the
feasible region have been enumerated by cdd+ (Fukuda, 1999; Avis and Fukuda,
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
6.8
Eigenvalue
6.6
6.4
6.2
6
0 20 40 60 80 100 120 140
Vertex number
Fig. 4.16 Lowest eigenvalue of the tangent stiffness matrix of each solution corresponding to the
vertex of the feasible region.
5.92
5.915
Compliance
5.91
5.905
5.9
0 20 40 60 80 100 120 140
Vertex number
Fig. 4.17 Compliance against the specified load of each solution corresponding to the vertex of
the feasible region.
1996) to find 74 vertices. The mean values are calculated for the coefficients of all
the vertices to obtain the nominal value 0 = (0.4198, 0.5832, 3.2567, 0.4294)>.
The maximum and minimum forces of the cables of the nominal self-equilibrium
forces are 8507.8 and 1359.6, respectively, whereas those for the struts are 3847.6
and 8290.3, respectively. The value of 1 at the center is 6.4812. The lowest
eigenvalue 1 of K is positive at all the vertices after constraining the rigid-body
motions; i.e., the structure is stable at any set of self-equilibrium forces in the
feasible region for designing the nominal values.
The worst-case scenario under constraint (4.144) is next investigated. Let si =
500 for all members. Figure 4.16 shows the values of 1 at the 138 vertices of the
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
2600
2400
Force deviation
2200
2000
1800
1600
0 20 40 60 80 100 120 140
Vertex number
Fig. 4.18 Force deviation of each solution corresponding to the vertex of the feasible region.
region defined by Eq. (4.144). The worst value of 1 is 6.0948, which is about 94%
of the nominal value 6.4812.
The compliance W under static loads is next investigated. The vertical loads of
10 are given asymmetrically in the negative direction of z-axis at nodes 3, 12 and
20 indicated in Fig. 4.14(a). To exclude the rigid-body motions of the structure,
the displacements are constrained at nodes 11 and 19 in x-direction, at nodes 6 and
11 in y-direction, and at nodes 6, 10, 11, 18 and 19 in z-direction. Note that the
loads are assumed to be sufficiently small to investigate the stiffness against small
disturbance by linear analysis. The values of compliances at the vertices of the
feasible region are plotted in Fig. 4.17. As is seen, the compliance does not strongly
depend on the forces.
Finally, the deviation e of forces from the nominal value is plotted in Fig. 4.18.
As is seen, the deviation is widely distributed in the region of uncertainty; however,
the distribution depends on the definition of s. Although we have considered
only three performance measures, the method based on vertex enumeration is very
effective for the case where many performance measures are to be evaluated in the
process of force design.
Here it is worthwhile to mention that some investigators have conducted studies
on determining the worst possible stochastic responses due to incompletely known
probabilistic characteristics of the input quantities. These works include those of
Abbas and Manohar (2005, 2007), Chen, Lian and Yang (2002), Deodatis, Graham
and Micaletti (2003a, 2003b), Deodatis and Shinozuka (1989, 1991), Iyengar (1970),
Iyengar and Dash (1976), Iyengar and Manohar (1987), Papadopoulos, Deodatis and
Papadrakakis (2005), Sarkar and Manohar (1998), Shinozuka (1987), and Wall and
Deodatis (1994).
The references mentioned in Sec. 6.6.4, pertaining to earthquake engineering,
also fall into this category.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 5
Anti-Optimization in Buckling
5.1 Introduction
It is well known that the buckling loads of thin-walled structures such as plates
and cylindrical shells are drastically reduced in the presence of initial geometrical
imperfection, and the magnitude of reduction of the buckling load strongly depends
on the shape of the imperfection. Therefore, extensive research has been carried
out for anti-optimization for finding the worst imperfection, which minimizes the
buckling load under constraint on the norm of the imperfection.
In the general theory of stability of elastic conservative systems, the buckling is
characterized as a critical point of equilibrium state, which is classified to limit
point and bifurcation point. The details of the classification of critical points
may be consulted in Thompson and Hunt (1973, 1984). Worst imperfections have
been found for various types of critical points including unstable-symmetric bifur-
cation point (Ikeda and Murota, 1990), stable-symmetric bifurcation point (Ohsaki,
113
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
K1
K2 L2
2002b), multiple bifurcation points (Ho, 1974; Peek and Triantafyllidis, 1992), and
hilltop branching point (Ikeda, Oide and Terada 2002; Ohsaki and Ikeda 2007a,
2007b). Since the buckling loads generally coincide as a result of optimization
against buckling (Masur, 1984; Ohsaki, 2000; Olhoff and Rasmussen, 1977), it is
very important to investigate worst imperfection for coincident critical points.
In the course of finding the worst imperfection, an anti-optimization problem
is formulated and solved explicitly or numerically. For the structures where the
prebuckling deformation is negligible for the evaluation of the buckling load, anti-
optimization problems can be formulated in a simple form of linear buckling analysis
(de Faria and de Almeida, 2003a).
Lindberg (1992a, 1992b) investigated the problem with multimodal dynamic
buckling with convex bounds in the Fourier coefficients of the imperfection. Pan-
telides (1996a) investigated worst imperfection in geometrical and material proper-
ties for stiffening elements using the ellipsoidal bounds. Takagi and Ohsaki (2004)
defined worst imperfection using the eigenvalues of the tangent stiffness matrix. El
Damatty and Nassef (2001) found worst imperfection of a ribbed shell by using a
genetic algorithm.
Consider a rigid vertical bar with length L as shown in Fig. 5.1 supported by a
horizontal spring with extensional stiffness K1 and a rotational spring with stiffness
K2 L2 . The bar is subjected to a vertical load P , and the rotation of the bar is
denoted by . The dotted and solid lines in Fig. 5.1 are the shapes before and after
deformation, respectively.
The total potential energy is written as
1 1
= K1 (L sin )2 + K2 L2 2 P L(1 cos ) (5.1)
2 2
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Consider the case where the bounds are given for a by interval as
aL a a U (5.10)
with the lower bound aL and upper bound aU . Since P c is a linear function of a1
and a2 , as observed from Eqs. (5.5) and (5.6), the anti-optimization problem turns
out to be a linear programming problem
minimize P c (a) (5.11a)
L U
subject to a a a (5.11b)
The anti-optimal solution exists at a vertex of the feasible region, and the obvious
anti-optimal solution is found as (a1 , a2 ) = (aL1 , aL2 ), because P c (a) is a monotoni-
cally increasing function of a.
is a constant matrix defined by p. Hence, the linear buckling load factor b , which
is often called simply the linear buckling load, is found as the smallest positive value
of satisfying
(K E + b K G0 )b = 0 (5.15)
where b is the linear buckling mode. Note that the linear eigenvalue problem
(5.15) generally has negative buckling load factors, because the structure may buckle
against the loads applied to the reverse direction and the symmetric matrix K G0 is
indefinite. However, the direction of the load is generally fixed, and we can assume
that the buckling load factor takes a positive value. The ith positive smallest value
of b is called the ith linear buckling load factor; however, we are interested in the
lowest positive buckling load only.
The buckling load factor represents the safety against static loads when plastic
failure is not considered. Therefore, in the following, the linear buckling load factor
is minimized in an anti-optimization problem. As demonstrated with the simple
example in Sec. 5.2, the anti-optimization problem can be formulated with either
the ellipsoidal or linear bound of the uncertain parameter vector a representing
variability of load, shape, material property, etc. For the ellipsoidal bound with
weight matrix W , the anti-optimization problem is formulated as
minimize b (a) (5.16a)
>
subject to a W a 1 (5.16b)
The details of the ellipsoidal model can be found in Sec. 3.4.
The anti-optimal solution to Problem (5.16) can be obtained explicitly if the
buckling load is linearized using sensitivity coefficients. Sensitivity analysis of the
eigenvalue of vibration has been presented in Sec. 3.6.3 as the simple case of the
eigenvalue sensitivity. For linear buckling load, differentiation of Eq. (5.15) with
respect to ai leads to
K E b b b K G0 b b
+ KE + K G0 b + b + b K G0 =0 (5.17)
ai ai ai ai ai
Since K E is positive definite and the buckling load b is positive, b is normalized
in view of Eq. (5.15) as
b> K G0 b = 1 (5.18)
Premultiplying b> to Eq. (5.17) and using Eqs. (5.15) and (5.18), we obtain the
sensitivity coefficient
b K E K G0
= b> + b b (5.19)
ai ai ai
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
It is seen by comparing Eq. (5.19) with Eq. (3.96) in Sec. 3.6.3 that the sensitivity
coefficients of the buckling loads can be computed in a similar manner as those of the
eigenvalues of free vibration. However, the geometrical stiffness matrix K G0 for the
unit load factor corresponding to the load pattern vector p implicitly depends on a i
through the axial forces or stresses against p. Therefore, sensitivity analysis of the
responses against the static load p is needed to compute the sensitivity coefficients
of the linear buckling load (see Sec. 3.6.2 for sensitivity analysis of static responses).
When the bounds of the uncertain parameters are given by interval, and b (a) is
linearized with respect to a, the anti-optimization problem turns out to be a linear
programming problem. For the case without linearization of b (a) and with linear
constraints for a, the anti-optimal solution can be found using vertex enumeration
of the feasible region as shown in Sec. 5.6.
y
P P P P
5000 9 10 11 12
5000 5 6 7 8
6000 1 2 3 4
small, the buckling load is evaluated by linear buckling analysis, whereas the buck-
ling load of an imperfect structure is evaluated by nonlinear path-tracing analysis,
for which the displacement increment approach is used. The effect of yielding is not
considered here to investigate the elastic properties of the frame, thereby simplifying
the anti-optimization process.
Let i () denote the maximum absolute value of the stress of the flange at the
two ends of the ith beam-column element. The maximum value of i () throughout
the frame is denoted by m () as
m () = max i () (5.21)
i
third- and fourth-order differential coefficients of the total potential energy with
respect to the generalized displacements in the directions of buckling modes.
(iii) The worst-case design and the corresponding worst imperfection that maximize
the increase of the maximum stress are found by linear analysis under specified
loads, and compared with the worst imperfection obtained by the nonlinear
path-tracing analysis.
Fig. 5.3 Buckling modes of optimal solution under linear buckling constraint.
Ri
their original locations and the imperfection vectors R = (R1 , . . . , Rf )> and =
(1 , . . . , mc )> , R and are regarded as the uncertain parameters in the anti-
optimization problem. The maximum load m of the imperfect frame in Table 5.1
is defined by the stress constraints (5.21) and (5.22), as the load factor when the
maximum stress m computed by path-tracing analysis for specified R and reaches
the upper bound. Hence, the anti-optimization problem is stated as
Since the maximum loads of imperfect systems are highly nonlinear functions
of the imperfection parameters R and , it is very difficult to obtain the anti-
optimal solution by a gradient-based nonlinear programming approach. Therefore,
the worst imperfection is searched for at the upper and lower bounds of the column
out-of-straightness i = 1/500, whereas the three cases Ri = 1/500, 0, 1/500 are
searched for story inclination. Hence, the anti-optimization problem is formulated
as a mixed integer nonlinear programming problem.
From f = 3, mc = 12, and symmetry of the frame, the total number of solutions
is 33 212 /2 = 55296. Therefore, it is possible to find the anti-optimal solution
by the enumeration process. Incremental path-tracing analysis has been carried
out for each set of imperfection parameters to find the maximum load. The worst
imperfection found by this enumeration is shown in Fig. 5.5 and in the first row
of Table 5.2. In solution of the table, the first three columns give the story incli-
nation R1 , R2 , R3 , and the remaining show the out-of-straightness of the columns
1 , . . . , 12 , where the column numbers are given in Fig. 5.2. The signs + and
correspond to 1/500 and 1/500, respectively. As is seen from Fig. 5.5, the worst
imperfection is similar to a sway mode (Mode 1 in Fig. 5.3) and is asymmetric with
respect to the y-axis.
The maximum load of the imperfect system with worst imperfection is 76.067,
which is about 76% of the maximum load 100 of the perfect system. The relations
between m and the load factor obtained by path-tracing analysis for the perfect
frame and the frame with worst imperfection are plotted in Fig. 5.6. Note that no
drastic decrease due to geometrical nonlinearity is observed in the imperfect system.
For a larger frame with more members, it is very difficult to find the anti-optimal
solution by enumeration. Therefore, in order to present a general method, we next
find approximate anti-optimal solutions by simulated annealing (see Sec. 2.6.3 for
details of the method). The initial temperature parameter is 0.02, and the temper-
ature is multiplied by 0.99 at each iterative step. The iteration is terminated after
500 steps. The results from the randomly generated ten different initial solutions
are listed as SA-1, . . . , SA-10 in Table 5.2. It is seen that a good approximate
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Enumeration 76.067 + + + + + + + + +
SA-1 76.306 0 + + + + + +
SA-2 76.216 + + + + + + + + +
SA-3 76.154 + + + + + + + + +
SA-4 76.306 0 + + + + + +
SA-5 76.454 0 + + + + + +
SA-6 76.638 + + + + +
SA-7 77.397 + + + + +
SA-8 76.995 + + 0 + + + + + + +
SA-9 76.398 + + + + + + + +
SA-10 76.763 + + + + + +
100
Perfect
80
Load factor
60 Imperfect
40
20
0
0 0.5 1 1.5
Maximum stress
Fig. 5.6 Relations between maximum stress m and load factor of the perfect and imperfect
systems.
solution with m = 76.154 has been found by the SA with random start, where
the total number of analysis is 500 10 = 5000, which is less than 1/10 of the
enumeration.
c = c1 c1 + c2 c2 + c3 c3 (5.24)
exists in the potential energy . Let U c denote the value of the displacement
vector at the critical point of the perfect system. For Mode 2, the deformation
corresponding to the displacement vector U = U c + c2 and U = U c c2 have
different values of . However, the asymmetry is very small, because the maximum
load for the imperfections (c1 , c2 , c3 ) = (0, 1, 0) and (0, 1, 0) are almost the same
as those shown in Table 5.3.
Let ,i denote the differential coefficient of with respect to the generalized
displacement in the direction of ci . The third- and fourth-order differential coeffi-
cients are computed as
,222 = 8.7 104 , ,112 = 1.3 104 , ,233 = 6.6 104 ,
,1111 = 3.9 102 , ,2222 = 5.1 102 , ,3333 = 8.5 101 , (5.25)
,1122 = 2.9 102 , ,1133 = 5.7 102 , ,2233 = 3.3 102
where the vanishing terms are not shown, e.g., 111 = 333 = 0 due to symme-
try (see Ohsaki and Ikeda (2007a) for details of computation of the higher-order
differential coefficients, where the symbolic computation software package Maple 9
(Maplesoft, 2004) has been extensively used).
To investigate the effect of third-order interaction, let q denote the generalized
coordinate in the direction of the asymmetric bifurcation mode c2 :
U = U c + qc2 (5.26)
Then can be expanded as
1 1 1
= c + ,2 q + ,22 q 2 + ,222 q 3 + ,2222 q 4 (5.27)
2 6 24
where c is the value of at the critical point.
Since ,2 = 0 and ,22 = 0 are satisfied from the equilibrium and the criticality
conditions, the leading term of governing the symmetry is ,222 . Suppose q = 100
corresponding to maximum out-of-straightness of the column about 1/100 in the
direction of c2 , which is larger than the expected value in design practice. In this
case, the orders of the third- and fourth-order differential coefficients ,222 and
,2222 , respectively, are 102 and 106 . Therefore, the third-order term is negligibly
small compared to the fourth-order term, and the asymmetry of c2 is very small;
hence, the third-order interaction can be neglected.
The fourth-order interaction should also be investigated for this almost sym-
metric system. Since the fourth-order interaction terms ,1122 , ,1133 and ,2233
among the three modes are all positive and their absolute values are in the same
order as those of 1111 , ,2222 and ,3333 , no increase of imperfection sensitivity
will be observed for this frame (Thompson and Hunt, 1984).
In order to further investigate the effect of mode interaction, maximum loads of
imperfect systems of non-optimal frames are computed. If the cross-sectional areas
of the braces of the optimal frame in Table 5.1 are replaced by 0, i.e., if we remove
the braces, the lowest buckling mode is a sway mode, and the three lowest buckling
loads are 22.53, 35.89 and 53.43, respectively, which do not coincide. The maximum
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 5.7 Worst imperfection for maximizing the stress ratio I / P by linear analysis.
We next consider the worst-case design of the 3-story 3-span frame. Let P and
I
denote the maximum absolute value of the stresses of the perfect and imperfect
systems, respectively, obtained by linear analysis, where in general, I > P holds.
The worst imperfection is defined so as to maximize the ratio I / P due to the
imperfection defined by the vector consisting of the out-of-straightness of each
column element, which is assumed to be a continuous variable with an upper bound
of 1/500 and a lower bound of 1/500. The nodal locations are computed from the
base to the top by summing up the inclinations of column elements multiplied by
the element lengths. The vector A of the cross-sectional areas is also considered
as design variable vector to find the worst-case design, where the symmetry of
the frame is incorporated. The constraint is given for the linear buckling load as
b 100.
Anti-optimization to find the worst-case design is carried out by the optimiza-
tion software package IDESIGN Ver. 3.5 (Arora and Tseng, 1987), where sequential
quadratic programming is used. The cross-sectional areas of the worst-case design
are listed in Table 5.4, and the corresponding worst imperfection is shown in Fig. 5.7.
The value of the objective function I / P at = 100 is 1.0912; i.e., the maximum
stress increases about 9.1% due to the worst imperfection. Therefore, if the maxi-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
100
80 Perfect
Load factor
60
Imperfect
40
20
0
0 0.5 1 1.5
Maximum stress
Fig. 5.8 Relations between maximum stress ratio m / P and load factor by path-tracing analysis
of the worst-case design.
mum stress is specified, the ratio of the maximum load of the imperfect system to
that of the perfect system is 1/1.0912 = 0.91642 due to the worst imperfection.
Geometrically nonlinear analysis has been carried out to verify the validity of
using linear analysis for anti-optimization. The maximum loads of the imperfect
system defined by the upper-bound stress m = P at = 100 found by geomet-
rically nonlinear analysis is = 81.189. Therefore, the reduction of the maximum
load is 19%, and the reduction ratio is magnified twice due to geometrical non-
linearity. Figure 5.8 shows the relation between the maximum stress ratio m / P
and the load factor of the perfect and imperfect systems. The initial stiffnesses are
almost the same, and a slightly nonlinear behavior is observed near the maximum
load factor.
In this section, we briefly summarize the anti-optimization method for linear buck-
ling loads of multi-parameter systems utilizing the convexity of the stability region
that is derived from the convexity of the reciprocal value of the linear buckling load.
Consider the case where uncertainty exists in the load vector P , which is pa-
rameterized by
XN
P = a i pi (5.32)
i=1
where a = (a1 , . . . , aN )> is the vector of uncertain parameters, which take the
role of coefficients for the load pattern vectors pi , and N is the number of loading
patterns.
Consider the situation where the total load vector P is multiplied by the load
factor , which is increased to find the buckling (critical) load factor c for each
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
specified parameter vector a. For this multi-parameter loading condition, the set of
critical points in the space of a forms a surface that is called the stability boundary
(Huseyin, 1975). The boundary degenerates to a point for the conventional case
of proportional loading with single load pattern. If the prebuckling deformation is
negligibly small, and the buckling load is estimated by linear buckling analysis, the
anti-optimization problem can be solved efficiently, as follows, utilizing the convexity
of the stability region (de Faria, 2000; de Faria and Hansen, 2001a, 2001b).
De Faria and de Almeida (2003a) presented an anti-optimization method of a
plate with variable thickness. Although they presented detailed investigation for
general cases, we briefly present the results below based on the concavity of the
reciprocal value of the buckling load.
We first derive the concavity property of the lowest eigenvalue of the tangent
stiffness matrix at an equilibrium state, which is assumed to be a linear function
of a. Let 1I and 1II denote the lowest eigenvalues of the tangent stiffness matrices
K(aI ) and K(aII ), respectively, for two sets of parameter values aI and aII . The
interpolation between K(aI ) and K(aII ) is given by the parameter a = aI + (1
)aII (0 1). The eigenvector associated with the lowest eigenvalue of K(a )
is denoted by . Based on the Rayleigh principle, the following relations hold:
> K I > K II
>
1I , 1II (5.33)
>
Therefore, the following inequality is derived for the lowest eigenvalue 1 of K(a ):
> [K I + (1 )K II ]
1 =
>
> K I > K II (5.34)
= >
+ (1 )
>
1I + (1 )1II
It is seen from Eq. (5.34) that the lowest eigenvalue 1 (a) of K(a) is a concave
function of the load parameters a. Hence, the feasible (stable) region S in the
space of a satisfying 1 0 is convex.
Consider next the linear buckling problem, where the tangent stiffness matrix
K(a) is divided into the linear stiffness matrix K E and the geometrical stiffness
matrix K G = K G0 . Note that K E is independent of P , and K G is a linear
function of P , hence, K G is a linear function of a. Then the problem of finding the
linear buckling load factor b is formulated as
(K E + b K G0 ) = 0 (5.35)
Suppose we are interested in the positive smallest buckling load factor and define
b as
1
b = (5.36)
b
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
P2
P1
H H
y
1.5
a2
0.5
0
0 0.2 0.4 0.6 0.8 1 1.2
a1
Fig. 5.10 Stability boundary (solid line) of the 2-bar truss; : anti-optimal solution.
problem of b (a). Since the feasible region defined by Eq. (5.41) is convex, the
anti-optimal solution that maximizes the convex function b (a) exists at a vertex
of the convex polyhedron (5.41).
Let V denote the set of vertices of the region (5.41). The anti-optimal solution
is found by solving the following problem:
maximize b (a) (5.42a)
subject to a V (5.42b)
The solution of Problem (5.42) can be found by vertex enumeration, if the number
of uncertain load parameters is moderately small.
For example, consider a 2-bar truss as shown in Fig. 5.9, where H = 1000 mm.
The two members consist of the same cross-section, where the cross-sectional area
is 300 mm2 , and the moment of inertia is 7.5 105 mm4 . Youngs modulus is
2
200 kN/mm . Each member is divided into two elements to consider member
buckling. In the following, the units of length and force are mm and kN, respectively.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
0.03
0.025
0.02
0.015
0.01
0.005
0
1
0.8
0.6
0
0.2
a2
0.4
0.4
0.6 0.2
a1
0.8
10
0.8 0.028
0.026
0.024
0.022
0.6 0.02
0.018
0.016
a2
0.014
0.012
0.4 0.01
0.008
0.006
0.004
0.2 0.002
0
0 0.2 0.4 0.6 0.8 1
a1
Fig. 5.12 Contour lines of the reciprocal value of linear buckling load b = 1/b .
a2 as
P1 = 10000a1, P2 = 10000a2 (5.45)
The reciprocal value b = 1/b of the linear buckling load is computed for each set
of (P1 , P2 ) for 0 a1 1 and 0 a2 1 to obtain the plot in Fig. 5.11, where
the contour lines are also plotted. It is confirmed from Fig. 5.11 that b is a convex
function of (a1 , a2 ).
Suppose the constraint is given as
a1 + a2 0.8, a1 0, a2 0 (5.46)
Then the feasible region is a triangle as shown in the thick straight lines in Fig. 5.12,
and the worst value of b is about 0.0187, which is attained at the vertex (a1 , a2 ) =
(0.8, 0).
5.7.1 Introduction
It has been pointed out by many researchers that imperfection sensitivity generally
increases as a result of coincidence of bifurcation points. Thompson and Hunt
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(1973) showed that the critical load of a latticed column drastically decreases due
to interaction of the critical loads associated with local and global buckling modes.
However, very large imperfection beyond the practically acceptable level for civil
engineering was considered therein. Ohsaki (2002a, 2002c) demonstrated that a
building frame with practical proportions does not have severe reduction of buckling
loads due to mode interaction (see Sec. 5.5 for details).
Thompson and Schorrock (1975) derived imperfection sensitivity laws for the
hilltop branching point, often abbreviated as hilltop point, that has a bifurcation
point at a limit point. The critical loads of imperfect systems are expressed as
a piecewise linear function of the imperfection parameter. Therefore, imperfec-
tion sensitivity is even decreased by the coincidence of critical points from the
imperfection-sensitive structure with a simple bifurcation point.
Ohsaki (2000) demonstrated that optimization of shallow shell-type truss against
buckling often leads to a hilltop point with multiple bifurcation. Ikeda, Oide and
Terada (2002) verified the imperfection sensitivity laws using random imperfections.
Ohsaki (2003b) derived imperfection sensitivity law for the case where an asymmet-
ric bifurcation point exists at a limit point. Ikeda, Ohsaki and Kanno (2005) derived
imperfection sensitivity laws for hilltop points of symmetric systems based on group
theoretic approach.
Ohsaki and Ikeda (2006) derived a piecewise linear imperfection sensitivity law
for the hilltop point with many symmetric bifurcation points. Imperfection sen-
sitivity of degenerate hilltop points was investigated by Ohsaki and Ikeda (2009).
The worst imperfection for hilltop points is presented in Ikeda, Oide and Terada
(2002) and Murota and Ikeda (1991). In this section, we verify the results in Ohsaki
and Ikeda (2006) using random imperfections. Interested readers may consult with
Ohsaki and Ikeda (2007a, Chap. 11) for details on random imperfection. Anti-
optimization problems are formulated to derive the worst imperfection. Numerical
examples are given for an arch-type truss.
where i is normalized as
>
i i = 1, (i = 1, . . . , n) (5.49)
Suppose 1 > 0 is satisfied at the initial state = 0. The first critical point as is
increased from 0 is defined by 1 = 0 with the critical load factor = c .
Consider the hilltop point where m 1 bifurcation points exist at a limit point;
i.e., m eigenvalues 1 , . . . , m vanish at the critical point. Let qi denote the gener-
alized displacement in the direction of ci as
n
X
c
U =U + qi ci (5.50)
i=1
where U c is the displacement vector at the critical point. The coordinates defined
by q1 , . . . , qm are called active coordinates, whereas those by qm+1 , . . . , qn are called
passive coordinates.
The increment of from c is denoted by as
= c (5.51)
>
Then the total potential energy is written in terms of q = (q1 , . . . , qn ) and as
V (q, ). In the following, differentiation of V with respect to qi is denoted by the
subscript ( ),i , and all variables are evaluated at the critical point without explicit
notation of the superscript ( )c . Differentiation of V with respect to is denoted
by the subscript ( ), .
where ckj is the jth component of ck . In the following, the component of a vector
is denoted by a subscript as h = (hi ), q e = (eqi ).
The vector h is divided into the components hB = (h1 , . . . , hm1 )> correspond-
ing to the bifurcation modes and hL = hm to the limit point mode. The vector q e
eB = (e
is similarly divided into q q1 , . . . , qem1 )> and qeL = qem . Let F B denote the
(m 1) n matrix for which the ith row is equal to c> i . Then, from Eq. (5.61),
we obtain
hB = F B Bd (5.62)
The (m 1) (m 1) matrix G is defined so that the (i, j)-component is equal
to V,ijm . Then, from Eq. (5.60),
eB = G1 F B Bd
q (5.63)
is derived. Hence, from Eq. (5.59), we have
Cm = d> B > F B> (G1 )> F B Bd (5.64)
For simplicity, we write
e = V,mmm B > F B> (G1 )> F B B
G (5.65)
Then Eq. (5.58) is written as
c>
m Bd 1 p >e
c (d) = d Gd || (5.66)
V,m V,m
Finally, the anti-optimization problem for finding the worst imperfection pattern
that minimizes c is formulated as
minimize c (d) (5.67a)
N
X
subject to d2i = 1 (5.67b)
i=1
where ci is the coefficient, and the following orthogonality conditions are satisfied
between a bifurcation mode and the limit point mode:
c> c
i Bm = 0, (i = 1, . . . , m 1) (5.69)
The accuracy of this assumption is investigated in the numerical examples. Note
that N = n holds in this case.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Let D B denote the n (m 1) matrix for which the ith column is ci , and
divide c into the components cB = (c1 , . . . , cm1 )> and cL = cm , corresponding to
the bifurcation modes and the limit point mode, respectively. By defining
>
e DB
G = D B> G (5.70a)
em = c> c
m Bm (5.70b)
em cL 1 B> B
c (c) = c G c || (5.71)
V,m V,m
cB> cB = , (0 1) (5.73)
Since G is a real symmetric matrix, the minimum of the second term in Eq. (5.71)
is obtained as (1/V,m ) ||, where is the maximum eigenvalue of G . Hence,
using Eq. (5.72b), is obtained by minimizing c () as
em 1
c () = 1 || (5.74)
V,m V,m
p
H
3 4 4 3
4 4 4 4 4 4
2 2
1 1 x
L L L L L L L L
y
Load factor
3
0
0 20 40 60 80 100
Displacement
Fig. 5.14 Relation between load factor and vertical displacement at the center node.
0.00015
duplicate
0.0001
Eigenvalues
0.00005
limit point
0
0 20 40 60 80 100
Displacement
Fig. 5.15 Relation between eigenvalues and vertical displacement at the center node.
Mode 1 c1
Mode 2 c2
Mode 3 c3
Mode 4 c4
Fig. 5.16 Four member buckling modes c1 , . . . , c4 and limit point mode c5 .
4.775
4.770
Maximum load
4.765
4.760
4.755
0.02 0.04 0 0.02 0.04
Imperfection parameter
Fig. 5.17 Relation between c and for imperfection in the direction of sum of five modes; solid
line: linear estimation using sensitivity coefficient, +: critical load found by path-tracing analysis.
/ 2
/ 2
Fig. 5.19 Worst imperfection in the space of buckling modes as the anti-optimal solution of
Problem (5.72).
4.8
4.75
Maximum load
4.7
4.65
4.6
4.55
0 0.5 1 1.5 2
Norm of imperfection
Fig. 5.20 Relation between c and imperfection norm for small random imperfection = 0.1;
solid and dotted lines: linear estimation using Problem (5.67) and Problem (5.72), respectively,
+: critical load found by path-tracing analysis.
The worst objective value of Problem (5.72) is 0.11231, which is divided by the
norm 1.0478 to obtain the sensitivity 0.10719 with respect to the imperfection pa-
rameter . The estimation by linear approximation using this sensitivity coefficient
is plotted by dotted line in Fig. 5.20, where is the result of path-tracing analysis
for various norms of imperfections in the direction of the worst imperfection. The
+ mark shows the value of c for small random imperfection corresponding to
= 0.1. Similar results for moderately large imperfection = 1.0 are shown in
dotted line in Fig. 5.20. As is seen, the critical loads of imperfect systems can be
linearly approximated by sensitivity coefficients with good accuracy in the range of
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
4.8
Maximum load
4.6
4.4
4.2
4
0 5 10 15 20
Norm of imperfection
Fig. 5.21 Relation between c and imperfection norm for moderately large random imperfection
= 1.0; solid and dotted lines: linear estimation using Problem (5.67) and Problem (5.72),
respectively, +: critical load found by path-tracing analysis.
Fig. 5.22 Worst imperfection in the general direction as the anti-optimal solution of Prob-
lem (5.67).
small imperfection. The solution of Problem (5.72) gives a good lower bound for
imperfections defined by a moderately smooth shape as shown in Fig. 5.18.
On the other hand, the worst imperfection mode in the general direction ob-
tained by solving Problem (5.67) is as shown in Fig. 5.22, which is more irregular
than the mode in Fig. 5.19. Therefore, in order to incorporate it into practical
situations, the imperfection mode should be smoothed by restricting the shape as a
combination of possible imperfection shapes during fabrication and manufacturing.
The sensitivity coefficient with respect to the imperfection parameter correspond-
ing to this worst imperfection is 0.25822. The solid lines in Figs. 5.20 and 5.21
are the critical loads by linear approximation using the sensitivity coefficient. The
result of path-tracing analysis is indicated by . As is seen, the worst imperfec-
tion defines a conservative lower bound of the critical load for the specified norm of
smooth imperfection shown in Fig. 5.18.
If we express the worst imperfection mode of Problem (5.67) by ci (i = 1, . . . , 5),
the coefficients are obtained as 0.0, 0.41847, 0.0, 0.10663, 0.071310 for which the
norm is 0.43769, far less than 1. Therefore, the worst mode has components of the
passive coordinates, and we can conclude that the worst imperfection cannot be
obtained by considering the active coordinates only.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 6
Anti-Optimization in Vibration
It was the best of times, it was the worst of times... (Charles Dickens,
A Tale of Two Cities)
... singular praise it is to have done the best things in the worst
times... (Inscription in Harold Church, Staunton, for Sir Robert
Shirley, Baronet)
6.1 Introduction
The eigenvalues of vibration and responses to dynamic loads play important roles
in the design process of various fields of engineering. For eigenvalue problems, the
uncertainty in structural parameters can be introduced in a similar manner as the
static problems presented in Chap. 4. However, for the responses to dynamic loads
such as seismic excitations, the uncertainty of loads in time domain should be con-
sidered. Thus, the anti-optimization problem becomes very difficult to formulate.
Chen and Wu (2004a) applied interval parameters to anti-optimization consid-
ering dynamic responses, where the nonlinear responses are approximated by sensi-
tivity analysis. The beams under uncertain excitation was studied by Sadek, Sloss,
Adali, and Bruch (1993). Anti-optimization for earthquake excitation was studied
by Baratta, Elishakoff, Zuccaro and Shinozuka (1998) and Zuccaro, Elishakoff and
Baratta (1998).
Modares, Mullen and Muhanna (2006) presented an interval formulation for the
anti-optimization of eigenvalues of vibration, where linear perturbation by positive
semidefinite matrix representing the uncertainty in elastic modulus is considered and
145
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
1 M0 2 M0
K1 K2
the monotonicity of the eigenvalue is used. Interval matrix problem for eigenvalue
analysis was presented by Qiu, Chen and Elishakoff (1995a, 1996b) (see Chap. 7 for
basic properties of interval matrices).
Let K and M denote the stiffness matrix and mass matrix, respectively, of a
finite-dimensional structure. The eigenvalue problem of undamped free vibration is
formulated as
Ki = i M i , (i = 1, . . . , n) (6.1)
where i and i are the ith eigenvalue and eigenvector, respectively, and n is the
number of degrees of freedom. i is orthonormalized as
>
i M j = ij , (i, j = 1, . . . , n) (6.2)
where ij is the Kronecker delta. An anti-optimization problem of free vibration is
formulated to find the worst-case scenario for minimizing the lowest eigenvalue.
Consider a 2-bar structure as shown in Fig. 6.1 that has lumped mass M0 at
nodes 1 and 2. The extensional stiffnesses of members 1 and 2 are denoted by K1
and K2 , respectively. The stiffness matrix K and mass matrix M are defined as
K1 + K2 K2 M0 0
K= , M= (6.3)
K2 K2 0 M0
The lowest eigenvalue and eigenvector are obtained as
q
1
1 = K1 + 2K2 K12 + 4K22
2M0
p (6.4)
M0 K1 + K12 + 4K22
1 =
2K2
where
q
2 = 2K12 + 8K22 2K1 K12 + 4K22 (6.5)
The sensitivity coefficients of the eigenvalue can be computed generally using the
formulations in Sec. 3.6.3. However, for this simple example, the sensitivity coeffi-
cients c1 and c2 are obtained by directly differentiating 1 with respect to K1 and
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
K2 as
!
1 1 1 K1
c1 = = p >0
K1 M0 2 2 K12 + 4K22
! (6.6)
1 1 2K2
c2 = = 1 p 2 >0
K2 M0 K1 + 4K22
Therefore, as expected, 1 is a monotonically increasing function of K1 and K2 .
Consider uncertainty in the stiffnesses as
K1 = K10 + a1 , K2 = K20 + a2 (6.7)
where K10 and K20 are the nominal values and a1 and a2 are the uncertain param-
eters. If the bound of a = (a1 , a2 )> is given as follows by the quadratic inequality
constraint
a> a D (6.8)
which is a simple form of an ellipsoidal bound, then the anti-optimization problem
is formulated as
minimize 1 (a) (6.9a)
>
subject to a a D (6.9b)
Since 1 is a nonlinear function of a1 and a2 , the anti-optimal solution cannot be
obtained explicitly by applying the Lagrange multiplier approach, as was the case
for the static response with load uncertainty in Sec. 4.1.
Let c = (c1 , c2 )> , and denote by c0 the value of c at (K1 , K2 ) = (K10 , K20 ).
Then 1 (a) is linearized at a = 0; i.e., at (K1 , K2 ) = (K10 , K20 ), to reformulate the
problem to a linearized problem:
minimize 1 (0) + c0> a (6.10a)
>
subject to a a D (6.10b)
In a similar manner as the simple example of static anti-optimization problem in
Sec. 4.1, the anti-optimal solution can be obtained as
r
1 1 c> 2 c2
a = c2 , = = (6.11)
2 2 D
Note that = corresponds to the optimal solution that maximizes 1 .
Consider the case, for simplicity, where M0 = 1, K10 = 2, K20 = 1, and D = 1.
Then Problem (6.10) is explicitly written as
minimize 0.5858 + 0.1464a1 + 0.2928a2 (6.12a)
subject to a21 + a22 1 (6.12b)
>
and the anti-optimal solution is found to be a = (0.4472, 0.8944) with =
= 0.1637.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
If 1 (a) is not linearized, and the bounds are given for a by the interval
aL a a U (6.13)
with the lower bound aL and upper bound aU , then the anti-optimization problem
has a nonlinear objective function and the linear bound constraints as
minimize 1 (a) (6.14a)
L U
subject to a a a (6.14b)
Since 1 is a monotonically increasing nonlinear function of a1 and a2 , the anti-
optimal solution for Problem (6.14) exists at a vertex of the feasible region, and the
obvious anti-optimal solution is found as a = aL .
where ai is the Fourier coefficient for the vibration mode with circular frequency
i . Let i (t) denote the response corresponding to the ith mode that is an explicit
function of t. Then the response is written as a linear function of ai as
N
X
u(t) = ai i (t) (6.17)
i=1
Consider the case where the displacement, velocity, and acceleration of the base
excitation is linearly bounded with the specified upper bound denoted by ( )U as
| U
ub | u b | u U
b , |u
U
b , |ub | ub (6.18)
Then the constraints are written explicitly with respect to a, and the anti-
optimization problem for finding the worst excitation for maximizing |u(t)| at a
discretized time step, or maximizing the maximum value of |u(t)|, |u(t)| or |
u(t)|
during the specified time period, is formulated as a linear programming problem.
If the ellipsoidal bound is given for a = (a1 , . . . , aN )> , the problem turns out to be
a quadratic programming problem.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
The bounds in Eq. (6.18) can be more generally written as time-varying func-
tions:
| U
ub | u b | u U
b (t), |u
U
b (t), |ub | ub (t) (6.19)
Michalopoulos and Riley (1972) investigated a two-degree-of-freedom system and
found that the worst excitation tends to be a bang-bang type function in control
theory.
A global measure can also be used for defining the bound of the input force
(Tzan and Pantelides, 1996a). For example, an energy bound can be given as
Z T
[fb (t)]2 dt E U (6.20)
0
where E U is the specified upper bound, and T is the duration of the motion.
This equation is not reproduced here. Bulgakov (1946) writes the maximum
maximorum as the sum of the convergent series:
Z t1 Z t2
d d
mj () = Sj (lp ) j ( ) (lp )d + j ( ) (+lp )d +
0 d t1 d
(6.22)
The signs must be chosen so that both terms will be positive. Bulgakov (1946)
expresses mj () as mj () = Aj l , with
Aj = Sj j (0) 2j (t1 ) + 2j (t2 ) (6.23)
where Aj has a physical sense of influence coefficients. Finally, the maximum de-
viation in the response xj (t) caused by all disturbing actions combined is expressed
in the following form:
r
X
Mj () = Aj l (6.24)
=1
Special cases of this problem have been discussed in his book (Bulgakov, 1939)
and paper (Bulgakov, 1940). Bulgakovs solution is referred to in the Russian liter-
ature as the method of accumulation of disturbances. Bulgakovs approach was ex-
tended by Balandin (1993), Bulgakov and Kozovkov (1950), Gnoenskii (1961), Ko-
rneev and Svetlitskii (1995), Krichevskii and Ulanov (1973), Mironovkii and Slaev
(2002), Naishul and Svetlitskii (1956), Nguyen (1962, 1965), Roitenberg (1958),
Svetlitskii (1958, 1967, 2003), Svetlitskii and Prokhorov (1970), and Svetlitskii and
Stasenko (1967). Analogous solutions were also developed in the West, especially in
the theory of control (see, e.g., Michalopoulos and Riley (1972), Sage (1968)) and
is referred to as the bang-bang optimal control problem.
Later works, in the context of free and forced vibrations, are also associated
with evaluation of bounds albeit without uncertainty aspects. The reader may like
to consult, for example, works by Hu and Schiehlen (1996), Sharuz (1996), Sharuz
and Krishna (1996), and Schiehlen and Hu (1997).
6.4.1 Introduction
Behavior, vibration, and stability of viscoelastic structures have been dealt with in
a number of monographs (Bland, 1960; Christensen, 1982; Fl ugge, 1975). In these
studies, material properties of the structure have been fixed at some deterministic
parameters. However, it is well known that the viscoelastic properties of structures
exhibit a large scatter (Koltunov, 1976). This scatter is usually accounted for by
considering the material properties as random variables. Cozzarelli and Huang
(1971) and Huang and Cozzarelli (1972) were apparently the first investigators
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
to include material uncertainty in their analyses. Hilton, Hsu and Kirby (1991)
extended the elastic-viscoelastic analogies to the stochastic case due to random
linear viscoelastic material properties. Both Gaussian and beta distributions were
considered for modeling the uncertainty in the data.
In probabilistic analyses, the needed probabilistic information for analysis was
postulated as given. For example, Huang and Cozzarelli (1973) utilized a log-normal
density or a truncated log-normal density. However, extensive experimental data
are needed to substantiate the probability densities with regards to the data.
More often, the necessary data are simply lacking, or only partial information is
available about the parameters. In these circumstances, the usefulness of the results
of probabilistic modeling based on incomplete data may be questionable.
Shinozuka (1987) studied the response variability in the stochastic context. Up-
per bound results were derived in two cases in which the spectral density function of
the stochastic field has assumed limiting shapes. The importance of such response
variability studies is immediately understood if one recognizes that it is rather dif-
ficult to estimate experimentally the autocorrelation function or, equivalently, the
spectral density function of the stochastic variation of material properties.
In this section, we present the non-probabilistic, convex modeling (Ben-Haim
and Elishakoff, 1990; Elishakoff and Ben-Haim, 1990a) for dealing with material
uncertainty and attendant response variability for viscoelastic structures. In par-
ticular, the problem of forced vibrations of viscoelastic beams is studied. First, the
analytical solution by Inman (1989) is generalized for a deterministic set of vari-
ables, describing material properties. Next, these variables are treated as varying
in a solid ball in the four-dimensional space, thus modeling the scatter in material
properties. The response uncertainty is directly related to the material uncertainty.
The least favorable response needed for the design of the structure is determined.
where IV
is the fourth-order derivative of n with respect to x, and (, ) is the
n
inner product defined by
Z l
(, ) = (x)(x)dx (6.31)
0
Because of the orthogonality property (6.28), we are left with
Z t
d2 a m dam (IV
m , m ) (IV , m )
2 +c + E0 I am + I g(t )am ( ) m am d
dt dt (m , m ) 0 (m , m )
(f, m )
= , (m = 1, 2, . . . ) (6.32)
(m , m )
or, more specifically,
Z t
d2 a m c dam 4 4 1 1
2
+ + m am (t) + m g(t )am ( )d = fm (t),
dt dt 0 E0 (6.33)
g(t) = 0 for t < 0
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where
(f, m )
fm (t) = (6.34)
(m , m )
We assume zero initial conditions. Taking the Laplace transform of Eq. (6.33)
yields
c 4 1
s2 + s + m 4
Am (s) + m G(s)Am (s) = Fm (s) (6.35)
E0
We now follow Golla and Hughes (1985) and Inman (1989) and take the following
analytical form for G(s):
s2 + s
G(s) = 2 (6.36)
s + s +
Some restrictions will be imposed later on the values of the parameters , ,
and . Now, G(s) can be represented as
s
G(s) = 1 + 2 2 (6.37)
s + s + s + s +
so that the positive kernel function g(t) is given as
aeat bebt
g(t) = (t) + e(1/2)t sinh Dt (6.38)
ab D
where (t) is the Dirac delta function, and
2
a = + D, b = D, D = (6.39)
2 2 4
In addition, we assume that the discriminant D is nonnegative. Substitution of
Eq. (6.36) into Eq. (6.35) results in
4
2 c 4 m s2 + s 1
s + s + m + 2
Am (s) = Fm (s) (6.40)
E0 s + s +
Equation (6.40) corresponds to an equation with the fourth degree of s; hence, it
can be written as two second-degree equations of the form
1
m1 m2 2 c1 c2 k1 k2 Am (s) F m (s)
s + s+ = (6.41)
m2 m3 c2 c3 k2 k3 Zm (s) 0
where Zm (s) is an artificial variable; for the general issue of artificial variables, one
may consult studies by Ahrens (1990), McTavish (1988), McTavish, Hughes, Soucy
and Graham (1992), and Ottl (1987). The matrices in Eq. (6.41) are supposed to be
symmetric to simplify the analysis. The artificial variable Zm (s) is chosen in such a
way that the transfer functions in Eqs. (6.40) and (6.41) are coincident. Equation
(6.40) reads
"
4 4
4 c 3 4 c m 2 c 4 m
s + + s + m + ++ s + + m + s
E0 E0
# (6.42)
4 1 2
+ m am (s) = Fm (s)(s + s + )
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
2 2
following values have been taken: 1 = 1 N/m , 1 = 1 s1 , 1 = 1 N/m s,
1 = 1 s2 . The scatter is modeled as the vector belonging to some set Z(P )
taken as a four-dimensional solid ball given by
where P is the radius of the sphere. It should be noted that the derivation of radius
P assumes that the parameters , , and are of similar order of magnitude
and that the variations of these values are also of similar magnitude. On the other
hand, McTavish (1988) gives a numerical example (taken from Bagley and Torvik
(1983)) that demonstrates that these parameters have multiple orders of magnitude
difference. However, this may be simply remedied by choosing 1 , 1 , 1 and
1 such that an approximate distribution may be represented by a sphere in the
appropriately transformed coordinates. We are interested in the maximum possible
value the function f (, , , ) may take in this solid ball. This will result in the
least favorable response the system may experience when the parameters vary in
the solid ball defined by Eq. (6.79). We will also be interested in the minimum
value of f (, , , ). This will be identified with the best possible response.
Let 1 and 2 denote the maximum and minimum of the function f (, , , )
within the solid ball, given in Eq. (6.79):
Since Z(P ) is convex, the linear function f (, , , ) reaches its maximum and
minimum values on the boundary; i.e., on the set
Therefore, the final results for the least favorable response (denoted as LF R)
and for the most favorable response (indicated as M F R) are
p
LF R = 1 (P ) = f0 + P > (6.86a)
p
M F R = 2 (P ) = f0 P > (6.86b)
where f0 = f (0 , 0 , 0 , 0 ). Equation (6.76) can be put in the following form, in
conjunction with Eq. (6.58):
1
f (, , , ) = (eit + eit ) (6.87)
2
where asterisk denotes complex conjugate. The function , with term m = 1,
retained is
B1 2q0 2l
1 = , B1 = ( 2 + i) (6.88)
C1
where
(E0 )I 4 c
4
C1 = + + 2
L
(6.89)
E0 I 4 c (E0 )I 4 c 2
+ + i + +
L L
In the following, all the functions and their derivatives are evaluated at
(, , , ) = (0 , 0 , 0 , 0 ). The derivatives of are calculated as
1 4q0 2 I 4 2
= 2 ( + i) (6.90a)
C1 l
4q0 i 2 2 E0 I 4 ic
= C 1 ( + i) + + (6.90b)
C12 l
1 4q0 i 4
= 2 ( 2 + i) I (6.90c)
C1 l
4q0 E0 I 4 ic
= 2 C1 ( 2 + i) 2 + + (6.90d)
C1 l
Equations (6.86a) and (6.86b) become
!1/2
0 2 2 2 2
LF R = f0 P + + + (6.91a)
!1/2
2 2 2 2
M F R = f0 + P + + + (6.91b)
In the case of the timewise linearly increasing excitation given in Eq. (6.67), the
least favorable response as a function of non-dimensional time T reads, in view of
Eq. (6.73), as
!1/2
am 2 am 2
LF R(T ) = am (T ) + P + (6.92a)
!1/2
am 2 am 2
M F R(T ) = am (T ) P + (6.92b)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
LFR LFR
MFR MFR
(a) (c)
LFR
LFR
MFR MFR
(b) (d)
Fig. 6.2 Dependence of the least favorable response (LF R) and most favorable response (M F R)
(a) (a) (b) (a)
upon uncertainty radius P : (a) a0 = 34.35109 N/m2 , 0 = 7.8125105 s2 ; (b) a0 = 2a0 ,
(b) (a) (c) (a) (c) (a) (d) (a) (d) (a)
0 = 0 ; (c) 0 = 0 , 0 = 20 ; and (d) 0 = 20 , 0 = 20 .
= T norm of f
Fig. 6.3 Absolute value = > of the norm of gradient of the function f (, , , ) as a
function of parameters and .
The value of the viscous damping is fixed at c(a) = 5983.2 kg/(m s), corre-
sponding to the coefficient = c/(21 ) = 0.2. As is seen in Fig. 6.2(a), the
amplitude of vibration for the nominal structure is 1.424 103 . For the beam
without viscoelastic effects, the response of the structure evaluated through the use
of the expression
4q0
Y2 = (6.95)
|EI(/l)4 2 + ic|
is 1.979103 m3/2 . This is 28% more than the nominal response of the viscoelastic
beam. The least favorable response linearly increases with P , whereas the most
favorable response linearly decreases with P . At P = 1000, for example, LF R =
1.293 103 m3/2 ; i.e., about 9% larger than the nominal response. Figure 6.2(b)
(b) (a)
is associated with the same data as in Fig. 6.2(a) except that now a0 = 20 .
(c) (a)
In Fig. 6.2(c), the value 0 = 0 , but the nominal value of is twice as much
(c) (a)
as in Fig. 6.2(a), namely, 0 = 20 . In Fig. 6.2(d), the data are the same as in
(d) (a) (d) (a)
Eq. (6.94) except 0 = 20 and 0 = 20 .
Figure 6.3 depicts the behavior p of the absolute value of the gradient of the
function f (, , , ), namely, = > defining the least and most favorable
responses in Eqs. (6.86a) and (6.86b). As is seen for large values of and small
values of , the values of can be significant, contributing to larger values of the
least favorable response. On the other hand, for materials with small values of
and relatively large values of , the effect of uncertainty will be less pronounced.
Figure 6.4 is associated with the response to timewise linearly increasing excita-
tion. It should be noted that the scatter in the responses stems from the constant
term in Eq. (6.71). Remarkably, the viscoelastic beam and its elastic counterpart
share the same slope in the asymptotic response, when t . In these circum-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
LFR
MFR
Fig. 6.4 Response variability region in the beam subjected to timewise linearly increasing exci-
(3) (3)
tation (0 = E0 , 0 = 10 s2 ); response variability region is hatched.
LFR
MFR
Fig. 6.5 Response variability region in the beam subjected to timewise linearly increasing exci-
(4) (3) (4) (3)
tation (0 = 20 , 0 = 20 ); response variability region is hatched.
stances, however, use of the nonlinear governing equations is necessary, since the
response becomes too large to justify employing the linearized analysis. The data
(3) (3)
in Fig. 6.4 are the same as in Fig. 6.2(a) except 0 = E0 and 0 = 10 s2 ; in
addition, the uncertainty radius is fixed at P = 2. The upper curve represents
the LF R, whereas the lower curve represents the M F R. The hatched area is a
region of the response variability (or uncertainty). Equation (6.71) in conjunction
with Eqs. (6.77) and (6.88) was utilized for obtaining the M F R and the LF R. It is
noteworthy that the contribution of the viscoelastic effects to the equivalent viscous
damping ceq in Eq. (6.72) through ec is much more important than that of the purely
viscous damping for the numerical data chosen in Fig. 6.4. Therefore, the time shift
predicted by the second term in Eq. (6.71) is non-negligible. In Fig. 6.5, the data
(4) (3) (4) (3)
were changed to 0 = 20 and 0 = 20 with the same value of uncertainty
radius. As is seen, the response variability is reduced in comparison with Fig. 6.3.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
LFR
MFR
Fig. 6.6 Response variability region in the beam subjected to timewise linearly increasing exci-
(5) (3) (5) (3)
tation (0 = 0 /2, 0 = 0 /2); response variability region is hatched.
(5) (3)
In Fig. 6.6, the data are the same as in Fig. 6.2(a) except that now 0 = 0 /2
(5) (3)
and 0 = 0 /2.
The response uncertainty region is now increased. This is understandable since
the relative measure of uncertainty (RM U ) could conveniently be defined as
P
RM U = (6.96)
Q
where P is the radius of input uncertainty, and Q is the measure of the nominal
location of the uncertain vector as
s
2 2 2 2
0 0 0 0
Q= + + + (6.97)
1 1 1 1
where 1 , 1 , 1 and 1 are defined in Eq. (6.78).
With increasing RM U , one would anticipate the increase in the response un-
certainty. This qualitative observation is in agreement with the obtained results.
Moreover, the method presented in this section provides a rigorous way to quantify
the response variability, if scant information on variability is available.
6.5.1 Introduction
There is a variety of literatures on modeling earthquake excitations. Modern anal-
ysis is based on the recognition that this excitation is an uncertain process. In an
overwhelming majority of the studies, the uncertainty is modeled as a random pro-
cess, either stationary or nonstationary with various approximations and attendant
models. The works by Drenick (1970, 1977b), Drenick, Novomestky and Bagchi
(1989), Drenick and Yun (1979), Elishakoff and Pletner (1991), and Shinozuka
(1970) have been the exceptions in the literature dedicated to uncertainty modeling
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
'
''
Fig. 6.7 Collection of m = 7 points representing the recorded observations, to be included in the
N = 3 dimensional ellipsoid.
be the coordinate vectors of the points, collecting any relevant parameters to identify
the characters of the phenomenon at the observation Pi in the reference frame with
origin O, which is given, e.g., as follows for N = 3:
1 0 0
(e1 , e2 , e3 ) with e1 = 0 , e2 = 1 , e3 = 0 (6.116)
0 0 1
The proposed procedure is conducted basically in two phases:
The second phase (B) is straightforward, as it will be shown later. Phase (A) is
founded on the solution of two auxiliary problems as follows.
'
''
Fig. 6.8 Two parallel planes 0 and 00 ; is the unit vector orthogonal to both.
and only if
x> 0 >
i , xi
00
(6.117)
are satisfied simultaneously. By applying Eq. (6.117) for i = 1, . . . , m, one obtains
that the strip includes all points if and only if
X > 0 1, X > 00 1 (6.118)
where
>
1 x1
.. ..
1 = . , X = . (6.119)
1 x>
m
The total width of the strip is given by 0 00 . Therefore, the first problem denoted
by Problem I turns out to be
find SN 1 , 0 and 00 that minimize 0 00 (6.120a)
subject to X > 0 1 (6.120b)
> 00
X 1 (6.120c)
with SN 1 being the unit sphere in EN .
This problem is approached in two steps:
Step I.1: For the given SN 1
find ( 0 , 00 ) that minimize 0 00 (6.121a)
> 0
subject to X 1 (6.121b)
X > 00 1 (6.121c)
It is trivially given as
0 = max x> >
i = xr ,
00
= min x> >
i = xs (6.122)
i i
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Step I.2: In order to find the smallest strip containing all points, consider the
following problem:
find 1 that minimizes 01 001 (6.123a)
> 01
subject to X 1 1 (6.123b)
> 001
X 1 1 (6.123c)
>
1 1 =1 (6.123d)
where 01 and 001 are implicitly related to 1 by Eq. (6.122). The above
problem is a nonlinear programming problem, mainly because of the constraint
>1 1 = 1.
Any constrained search procedure can be applied to find the optimal vector
1 . The simplest, and also effective, as tested in this investigation, is a random
search. The only difficulty lies in generating unit random vectors 1 s satisfying
the constraint (6.123d). The following procedure is suggested. Let be a generic
unit random vector and 1 the unit vector that minimizes 01 001 . Let ni
and si (i = 1, . . . , N ) be any random numbers in [0, 1], i = 2ni the random
angle in [0, 2] of and SGi the sign of the ith component of with
SGi = +1 if si 0.5
(6.124)
SGi = 1 if si > 0.5
The components of are chosen through the following steps:
(a) RN = 1,
(b) RN i = sin2 (p
i )RN i+1 (i = 1, . . . , N 1) and i = 2ni ,
(c) i+1 = SGi+1 Ri+1 Ri (i = N 1, N 2, . . . , 1).
Problem I is thus solved after Steps I.1 and I.2.
Problem II : Suppose 1 is found, and the first minimal strip is obtained, with width
10 100 = d1 . Let 1 , . . . , k (k < N ) be given unit vectors, mutually orthogonal,
i.e., >i j = ij , and consider the following Problem II:
are orthogonal unit bases for the entire EN , where the first is a known base and the
second is known up to k . Note that
>
e1
1 ..
B = . (6.127)
e>
N
One has to infer B k+1 from B k . Let 0 = k , and in accordance with the Gram
Schmidt orthogonalization method, one sets
1 = 11 k + k2
2 = 12 k + 22 1 + k3
3 = 13 k + 23 1 + 33 2 + k4 (6.128)
..
.
N k = 1N k k + 2N k 1 + 3N k 2 + + kN k+1
where, by orthogonalization
11 = (k2 )> k
(k )>
12 = (k3 )> k , 22 = (3 )> 1
1 1
(k >
4 ) 1 ( k )>
13 = (k4 )> k , 23 = ( 1 )> 1
, 33 = ( 4 )> 2
2 2
.. ..
. .
(k >
N k+1 ) 1 (k >
N k+1 ) 2
1N k = ( kN k+1 )> k , 2N k = ( 1 )> 1
, 3N k = ( 2 )> 2
,
N k (k )>
N k = (
N k+1
)>
N k+1
N k+1 N k+1
(6.129)
After the orthogonal vectors i have been found, one can normalize k+1
i as
k+1 i
i = , (i = 1, . . . , N k) (6.130)
| i |
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Note that each i for i > 0, and consequently each k+1 i , is orthogonal to j
(i = 1, . . . , k). In fact, every i turns out to be a linear combination of vector k
and vectors ki . All these vectors are assumed to have been built up orthogonal
to 1 , . . . , k in the previous step. Hence, one can write for suitable aij
NX
k+1
i = ai0 k + air kr (6.131)
r=1
and
NX
k+1
> >
i j = ai0 k j + air kr j = 0 (6.132)
r=1
with
N
X k N
X k
(k+1 )> k+1 = ik rk ( k+1
i )> k+1
r (6.135)
i=1 r=1
where, by orthonormality of s,
N
X k
(k+1 )> k+1 = (ik )2 (6.136)
i=1
Step A.1: Solve Problem I, and find 1 , 01 , 001 . After this step a first strip
between hyperplanes 0 and 00 is determined.
Step A.2: Solve Problem II with k = 2 in the form (6.137), and find k+1 , 0k+1 ,
00k+1 .
Step A.3,. . . , A.N : Iterate step A.2 for k = 3, . . . , N . After Step A.N, N cou-
ples of hyperplanes are individuated; each couple defines a minimal strip in EN
containing all m points.
By iterating these N steps, the minimal parallelepiped is found.
x1 100 300 700 400 200 400 0 100 300 700 400 200 750 1380
x2 500 550 450 470 500 500 500 200 500 800 600 800 1370 250
x3 800 800 600 600 400 400 200 200 100 100 900 900 1600 260
Note that the above minimal ellipsoid may not be the absolute minimum, since it
is minimal under the assumption that it is related to the minimal parallelepiped.
The ellipsoid so found is in general included between the ellipsoid contained in P
and the one containing P.
Fig. 6.9 Minimal parallelepiped, ellipsoid contained in it, and principal planes of the ellipsoid.
based on the consideration that the maximum structural response predicted by the
preferred method ought to be minimal. Prior to analytical treatment, the existing
data ought to be enclosed by the minimum-volume hyper-cube of volume V1 that
contains all experimental data. The available data also have to be enclosed by the
minimum-volume ellipsoid V2 . If the response R(V1 ) based on V1 is smaller than the
response R(V2 ) based on V2 , then one has to prefer interval analysis. If, however,
R(V1 ) is in excess of R(V2 ), one is recommended to utilize ellipsoidal analysis. If
R(V1 ) and R(V2 ) equal each other, or are in close vicinity, then two approaches
are equally valid. For further details and examples of data for 7-bar planar truss
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 6.11 Ellipsoid obtained by homothetic enhancement of ellipsoid 1 to include all points
of observation.
structure, or for 60-bar space truss structure, one is referred to the above papers.
Let P(E0 ) be the set of functions defined in interval (0, T ) and possessing the given
energy E02 . Then the problem reduces to
maximize u(t) = kukT = E0 HT , t (0, T ) (6.155a)
subject to a P(E0 ) (6.155b)
The following relation was proved by Drenick:
Z T Z T
HT2 = max h2 (t )d = h2 (t0 )d, t (0, T ) (6.156)
0 0
As T , the base accelerogram ac (t) yielding the maximum peak response, and
coined by Drenick (1973) in a later paper as the critical excitation, has a shape
coincident with the impulse response function reversed with respect to time, and is
given by
E0
ac (t) = h(t) (6.157)
H
The maximum peak response is given by
kuk = uc (0) = E0 H (6.158)
where uc (t) denotes the response due to acceleration ac (t). Some criticism was
expressed by Drenick (1973) himself, mainly about the circumstance that aseismic
design based on critical excitation could be ... far too pessimistic to be practical.
Yet Drenick (1973) investigated the ratio of the maximum response due to criti-
cal excitation, to the maximum response of actually recorded seismic accelerograms.
Response spectra were derived on the basis of both the critical excitation and a num-
ber of accelerograms from the set of recorded earthquakes, with attendant compar-
ison of results. It was shown that for the range of structures with natural period
from 0.5 to 1.2 sec and damping ratio in the range of 0.02 and 0.10, the ratio of
the critical to the experimental response was almost invariably in the neighborhood
of 2. By extending the class of possible accelerograms from only the recorded ones
to linear combinations of these, the above ratio decreases to the range of values
between 1.3 and 1.6.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
is the number of retained terms. One simple example of such a space is the Fourier
expansion. As a result, every accelerogram can be represented by a linear combi-
nation of harmonic functions. Let us assume that the duration of the earthquake
is known, and that it is possible (from geophysics, for instance) to find a bound for
the energy of the excitation, as in Drenicks problem
Z T
a2 (t)dt E02 (6.160)
0
If the base functions ai (t) are orthogonal and all scaled to the same energy, every
function of the type (6.159) satisfying the constraint (6.160) will possess coefficients
ci satisfying the following inequality:
N
X
c2i 1 (6.161)
i=1
Any possible function that satisfies Eq. (6.160) is referred to as an admissible ac-
celerogram. Thus, admissible functions can be mapped in the Euclidean space of
N -dimensional vector c = (c1 , . . . , cN )> . Admissible accelerograms are contained
in the unit sphere S0 centered at the origin.
The values of the maximum and minimum displacements of the structure are
defined for a(t) varying within S0 . Let ui (t) denote the response function under
ai (t). The response under any a(t) expressed by Eq. (6.159) are given as follows
due to the linearity of system:
N
X
u(t) = ci ui (t) (6.162)
i=1
Thus, with matrix notation, the problem is set as follows for any fixed instant t in
(0, T ):
find max c> u(t) (6.163a)
c
>
subject to c c 1 (6.163b)
>
where u(t) = (u1 (t), . . . , uN (t)) . Since Problem (6.163) is to maximize a linear
function within a convex domain, the anti-optimal solution exists on the boundary
of S0 , and the problem is reformulated as
find max c> u(t) (6.164a)
c
>
subject to c c = 1 (6.164b)
which differs from Problem (6.163) in that the inequality in Problem (6.163) is
replaced by an equality. Let us now introduce notations
r(c) = c> u, f (c) = c> c 1 (6.165)
At the solution point, with 0,
f = r for c> u max. (6.166)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
N=200
N=100
N=50
N=25
t
Fig. 6.12 Influence of N and the number of harmonics on the bound (6.170).
and
f = r for c> u min. (6.167)
where is the Lagrange multiplier (see Sec. 2.4 for details).
Hence, with Eq. (6.164b), the optimal and anti-optimal solutions for the fixed
time instance t is given as
c(t) = (t)u(t) (6.168)
with
1
2 (t) = (6.169)
u(t)> u(t)
The maximum displacement at any instant t in (0, T ), is finally given by
q
umax (t) = umin (t) = u(t)> u(t) (6.170)
A numerical example has been carried out by assuming that admissible functions
are continuous. All ai (t)s were taken as harmonic functions, as in the Fourier
expansion:
A0 sin(i t) for i m
ai (t) =
A0 cos(i t) for m < i N
(6.171)
2i 2 2 2
i = , A0 = E0
T T
with N even, and m = N/2.
In Fig. 6.12, the bound (6.170) is plotted for different values of m = 25, 50,
100, 200 for t ranging from 0 to 4 sec. The figure shows how the refinement of
the functional space of the excitation makes the bounds increase progressively, and
a convergence is attained for m = 100. Figure 6.13 depicts a sample maximizing
function for T = 30 and m = 100. These bounds are of the type of Drenicks critical
excitation. It appears to be highly questionable that such functions can be accepted
as credible accelerograms.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
b)
t
MAXIMAL ACCELEROGRAM AT T = 15
Fig. 6.13 Upper and lower bounds and sample maximal accelerogram for T = 30 sec: (a) bounds
(6.170); (b) umax of any harmonic component.
c2
Sd
c0
c1
So
subject to c c = 1, c>
>
0 c=a (6.182b)
where all constraints appear in the form of equalities. In order to solve Prob-
lem (6.182), we consider
L(c, r1 , r2 ; t) = c> u(t) + r1 (c> c 1) + r2 (c>
0 c a) (6.183)
where r1 and r2 are Lagrange multipliers. We require that the gradient of the
Lagrangian with respect to c vanishes as
L = u(t) + 2r1 c + r2 c0 = 0 (6.184)
Hence,
1
c = c(t) = (r2 c0 + u(t)) (6.185)
2r1
Substituting Eq. (6.185) into the second constraint c>
0 c = a in (6.182b) for S0 , we
obtain
1
c> c = 2 (u(t)> u(t) + r22 c> >
0 c0 + 2r2 c0 u(t)) = 1 (6.186)
4r1
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Table 6.2 Values of epicentral distance, local intensity, PGA, duration, and norm at
various sites.
Epicentral Local
Site distance intensity PGA Duration Norm
(km) (MCS) (g/10) (sec) (cm sec 3/2 )
6.6.3 Application
In order to have a quantitative estimate of the proposed procedure, consider a
particular area, namely, the Campania region in southern Italy, where a number of
accelerograms have been recorded during the Campania-Lucania earthquake which
took place 23 November, 1980. The magnitude of the event was estimated to be
6.5, the epicentral intensity was set at 7.5 degree the MercalliCancaniSieberg
(MCS) intensity scale. The event was rather non-typical, mainly because of its
duration that was up to almost 2 min at some sites. It should be noted that only
the duration of the strong motion will be considered here, which varies from about
52 sec to more than 86 sec. Information on the local characters of ground motion
was derived from the direct inspection of the recorded accelerograms, limiting the
analysis, for simplicity, to only the NS component. The considered records are
summarized in Table 6.2. Recorded ordinates of all accelerograms are converted to
cmsec2 . All earthquakes are preliminarily reduced to the same energy norm, given
in Eq. (6.154) as the one recorded in Torre del Greco, the closest site to Naples, the
town with the greatest intensity in the region, so that every record possesses energy
E0 = 79.46 cm sec3/2 .
The analysis was carried out including m = 400 sine and cosine waves for Torre
del Greco and Sturno, while m is increased to 800 for the accelerograms recorded in
Brienza and Bagnoli Irpino that exhibit power spectra scattered on a wider range
of frequencies. A plot of a typical accelerogram with the approximation resulting
from the associated Fourier expansion and its power spectrum is given in Fig. 6.15.
Note that the plot of Fourier approximation in Fig. 6.15(b) looks almost the same
as the original accelerogram Fig. 6.15(a).
Let fij be the ith coefficient of the Fourier expansion of the jth accelerogram
(i = 1, . . . , m; j = 1, . . . , 4). The central distribution of coefficients, collected in the
vector c0 , were obtained for each of the four analyzed sites, through the following
steps:
Step 1: Normalize fij , for every j, to unit modulus through division by a factor
Fj = |f j |, with f j = (f1j , . . . , fmj )> , and let cij = fij /Fj .
Step 2: Determine, for every site, the best values of the parameters j and j that
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
POWER SPECTRUM
t S
(a)
a*
(c)
t
(b)
Fig. 6.15 Torre Del Greco-Campania Earthquake of Nov. 23, 1980; (a) recorded accelerogram;
(b) Fourier approximation; (c) power spectrum.
give the optimal fit of the ordinates of the KanaiTajimi-type spectrum given in
Eq. (6.172) with the values of the spectrum corresponding to the normalized
coefficients cij from the recorded accelerograms on the corresponding sites.
Step 3: Calculate, for every value of , the target spectrum s0,ij , e.g., the spectrum
in Eq. (6.172), as follows:
m
X
s20,ij = f (; 1 , 1 , 2 , 2 ), s20,ij = 1 (6.197)
i=1
Step 4: Let
ci+m,j
ij = , (i = 1, . . . , m) (6.198)
cij
and, finally, calculate the central target spectra at the considered sites as follows:
s0,ij
c0,ij = q sign fij (6.199)
2
1 + ji
Step 5: Calculate, for every site, the quadratic scatter of the coefficients obtained
by directly processing the accelerogram and those calculated by the target spec-
trum
Xm
j2 = (cij c0ij )2 (6.200)
i=1
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 6.16 Processed power spectrum from recorded accelerograms and fitted central spectrum, in
the location of Torre del Greco.
In Fig. 6.16, the processed spectrum at Torre del Greco and the central spectrum
are plotted. In Table 6.3, the final values of j2 are listed illustrating that the
inequality j2 4 holds.
Assuming that the central spectrum is site-dependent, but that independence
holds for the quadratic scatter, it is possible to infer from the data that j2 ' 0.26
for the area under examination. After introducing this value in the results discussed
in Sec. 6.6.2, one obtains the bounds plotted in Figs. 6.17 and 6.18 for 0 = 10
100 sec1 .
It is also possible to perform a comparison with the previous, spectrum-free
bound, showing that the new bound is approximately half of the one constrained
only by the upper bound by Drenick.
The response displacement spectrum by the present procedure is finally calcu-
lated, for a value of the damping coefficient = 0.05. The comparison with the
envelope of the same spectra calculated with reference to the processed accelero-
grams shows that the present results, although better founded and less sensitive to
uncertain parameters than recorded accelerograms, are not too conservative, but
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Sn (cm)
PRESENT BOUNDS
5 SPECTRA OF RECORDED EARTHQUAKES:
TORRE DEL GRECO
4 BRIENZA
STURNO
CALITRI
3
BAGNOLI IRINO
0(sec-1)
20 40 60 80 100
Fig. 6.17 Response spectra of displacement: Comparison of present bounds with the envelope of
spectra from recorded accelerograms normalized to the same norm as Torre del Greco.
Sn (cm)
SHINOZUKAS BOUND
DRENICKS BOUND
5
PRESENT BOUNDS
0(sec-1)
20 40 60 80 100
Fig. 6.18 Response spectra of displacement: Comparison of present bounds with Drenicks and
Shinozukas bounds.
yield a magnification by a factor of about two in the design forces, in the whole
range of natural periods of the structures; this result can be directly incorporated
into the safety factor.
Following Shinozukas (1970) model, and the assumption given in Eq. (6.169),
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
with the parameters presented in Table 6.3, one obtains the following bound for the
maximum displacement over the whole duration of the excitation:
Z
1
Iei = |H()|X()d (6.201)
2
where
1
H 2 () = (6.202)
(02 2 )2 + 42 02 2
and
X() = K[g(|1 , 1 ) + g(|2 , 2 )] (6.203)
with K chosen so that
Z
1
X 2 ()d = E02 (6.204)
2
From inspection of Figs. 6.17 and 6.18, it is possible to infer that Problem (6.164)
yields the bound in the norm substantially close to that obtained via the Drenicks
original approach, apart from the range of higher frequencies, due to the truncation
of the expansion (6.171) by a finite number of harmonics (say, for i > 50 sec1 ).
Moreover, the spectrum-compatible bound in Problem (6.177) is not significantly
different from Shinozukas bound, although it yields slightly closer bounds in the
range of low frequencies. The opposite happens in the range of higher frequencies.
The ratio of the ordinates of the spectrum-compatible bound to the correspond-
ing ordinates of the envelope of spectra related to actual seismic records, normalized
to the same norm as Torre del Greco, vary in the range 2.0 2.5.
the approach in this section, as in Drenicks (1970) one, the bound involves also the
transient response starting with homogeneous initial conditions.
It should be emphasized that the present convex modeling combines positive
features of the approaches developed by Drenick (1970) and Shinozuka (1970), and
considerably reduces the estimates of the maximum possible response.
One should stress that the real structures are generally not SDOF-systems where
the maximum stresses are governed by the maximum displacement of a single mode.
The generalization of this method to multi-degree-of-freedom realistic structures is
possible through casting it in a state-space form as a vector differential equation.
One should also note that the maximum stress at two distinct positions of the
structure will be produced by the different worst-case combinations of excitations.
Obviously, one is interested in the worst response at the critical location of the
structure. Determining such a location may be a non-trivial task. Then, if the
structure is represented as the multi-degree-of-freedom system, one must explore
the maximum responses of each of the masses, and the maximum of these maxima
will constitute a critical response.
Analogously, a distributed system must be discretized through a fine mesh and
the maximum responses at each nodal point must be determined. Then the location
with maximum response amongst critical responses at each node will constitute the
globally critical response. The attractive feature in the present method is the fact
that one does not guess the critical location, but rather determine it through a
discretization scheme combined with convex optimization procedure for each nodal
point of the mesh.
Note that the present approach is restricted to linear systems. The generalization
to realistic nonlinear systems appears to be necessary for practical applicability of
this method. (This appears to be doable if and when the funding agencies are
more receptive to high-risk high-payoff methodologies.) In this context, convex
models have been applied to the nonlinear structures through treating the nonlinear
transfer function between the uncertain quantities and the output of the nonlinear
transformation, as a numerical code utilized by Ben-Haim and Elishakoff (1989b)
and the automatic differentiation procedure. Another alternative is the use of the
nonlinear programming schemes. For the application of the latter method the reader
may consult the paper by Li, Elishakoff, Starnes and Shinozuka (1996).
To sum up, it appears that the time is ripe to critically revisit existing stochastic-
ity concept-based approaches in earthquake engineering and examine the possibility
of utilizing alternative approaches. One such alternative approach was attempted
in this section, although to an SDOF system in the linear setting. Related sub-
jects, although in different settings were pursued by Abbas and Manohar (2005,
2007), Ben-Haim, Chen and Soong (1996), Elishakoff (1991b), Elishakoff and Plet-
ner (1991), Ganzerli, Pantelides and Reaveley (2000), He and Zhang (1997), Iyengar
and Manohar (1987), Lignon and Jezequel (2006), Ma, Leng, Meng and Fang (2004),
Manohar and Sarkar (1995), Moustafa (2009), Pantelides and Tzan (1996), Philap-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
pacopoulos (1980), Schmitendorf, Jabbari and Yang (1994), Srinivasan, Corotis and
Ellingwood (1992), Srinivasan, Ellingwood and Corotis (1991), Takewaki (2002a,
2002b, 2006b), Tzan and Pantelides (1996b), Wu and Soong (1996), Yoshikawa
(2002, 2003), Zhai and Xie (2007), and possibly others. Finally, we mention a
definitive monograph by Takewaki (2006a) and references therein. It generalizes
earlier investigations on convex modeling of earthquakes.
Shinozuka (1970) pioneered the determination of the maximum response in
stochastic settings. Further stochastic anti-optimization-based studies include those
by Deodatis, Graham and Micaletti (2003a, 2003b), Deodatis and Shinozuka (1989,
1991), Manohar and Sarkar (1995), Sarkar and Manohar (1998), Shinozuka (1987),
Shinozuka and Deodatis (1988), and others.
6.7.1 Introduction
In this section, we apply the hybrid optimization and anti-optimization to aeroe-
lastic problems, closely following the article by Zingales and Elishakoff (2001).
Optimization of structures with aeroelastic constraints has been dealt with by
several authors. A partial list of works includes those by Ashley (1982), Bishop,
Eastep, Striz and Venkayya (1998), Librescu and Bainer (1983), Livne and Mineau
(1997), Plaut (1971), Pierson and Genalo (1977), Ringertz (1994), Shirk, Hertz
and Weisshaar (1986), and Turner (1982). However, the uncertainty in elastic
moduli or in the material properties in conjunction with the aeroelastic optimization
is a relatively new topic. It has been addressed in the probabilistic setting by
Kuttekeuler and Ringertz (1998) and Allen and Maute (2004). The stochastic finite
element method for reliability of plates in supersonic flow was introduced by Liaw
and Yang (1990). In the future one would anticipate an increased utilization of the
stochastic finite element method in conjunction with aeroelastic phenomena. On
the other hand, it is instructive to quote Shinozuka (1987):
where A(x) is the cross-sectional area, I(x) is the moment of inertia, and is the
mass per unit length of the beam. The gas-flow interaction is represented by the
last term in Eq. (6.207), where k is the exponent of the polytropic thermodynamic
transformation, p is the pressure of the nondisturbed gas flow, c is the speed
of the shock waves in the nonperturbed airstream, and U is the relative velocity
between the gas flow and the system. The boundary conditions associated with the
simply supported beam read
2 w(x, t)
w(x, t) = = 0 at x = 0 and L (6.208)
x2
where L is the length of the beam. The initial conditions are
w(x, t)
w(x, t) = = 0 at t = 0 (6.209)
t
The governing equation for the deflection w(x, t) is not solvable exactly for arbitrary
choices of the functions 1 (x), 2 (x), b (x), and h (x). Hereinafter, the Bubnov
Galerkin approach will be utilized. We approximate the solution as
N
X
w(x, t) = i (x)fi (t) (6.210)
i=1
with N denoting the number of retained terms in the expansion, and i (x) designat-
ing known comparison functions satisfying all boundary conditions in Eq. (6.208).
We substitute Eq. (6.210) into Eq. (6.207). Naturally, the expression Eq. (6.210)
does not satisfy Eq. (6.207). The result of the above substitution, therefore, differs
from zero. We denote it by . We require that the inner product vanishes:
(, j ) = 0, (j = 1, . . . , N ) (6.211)
where
Z L
(, j ) = (x)j (x)dx (6.212)
0
This procedure leaves us with a set of ordinary differential equations:
N
X N
X N
X
kp
K f (t) + M f (t) + U N f (t) = 0, ( = 1, . . . , N ) (6.213)
=1 =1
c =1
The coefficients appearing in Eq. (6.213) are given as follows:
Z L
d2 d2 (x)
K = (x) 2 E(x)I(x) dx,
0 dx dx2
Z L
M = A(x) (x) (x)dx, (6.214)
0
Z L
d (x)
N = dx
0 dx
The solution of the system in Eq. (6.213) is furnished in the form
f (t) = A eit , ( = 1, . . . , N ) (6.215)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
one requires the determinant of the coefficient matrix of the system in Eq. (6.216)
to vanish:
det[ 2 M + U N + K ] = 0 (6.218)
leading to the approximate eigenfrequency equation for .
Equation (6.215) suggests that once the frequency takes a complex value, the
dynamic instability called flutter occurs. The critical condition appears when the
discriminant
For values larger than Ucr , the discriminant takes a negative sign, and the roots are
complex conjugate. For the case of the beam simply supported at both ends, the
following set can be utilized for the comparison functions:
x
(x) = sin , ( = 1, 2, . . . ) (6.228)
L
which represent the exact mode shapes of the associated uniform beam. Note that
in these circumstances,
yielding
p
22 22 ( 22 22 ) det[M ]
Ucr = = (6.230)
2 d3 2N12
Observe that Eq. (6.230) formally coincides with expression (4.104) in the mono-
graph by Bolotin (1963) for the uniform beam. The difference lies in the definition
of the circular natural frequencies .
Recall that E1 and E2 are uncertain variables, leading to variability of the
flutter velocity. The following question arises: How should one take into account
the variability of the elastic modulus on Ucr ?
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 6.19 Shape and location of the admissible uncertainty region C and of the 2 ellipses.
where
12 411 det[M ]
d11 = 2 ,
4N12
21 2 412 det[M ]
d12 = 2 , (6.237)
8N12
2 422 det[M ]
d22 = 2 2
4N12
One recognizes that the expression of the square of critical velocity is a quadratic
form in variables E1 and E2 . In matrix form, Eq. (6.236) can be rearranged as
2
Ucr = E > DE (6.238)
with
E1 d11 d12
E= , D= (6.239)
E2 d21 d22
Accounting for the physical meaning of the right-hand-side in Eq. (6.235), we can
conclude that the matrix D is positive definite. Therefore, Eq. (6.236), conveniently
rewritten in the form
2
E > DE Ucr =0 (6.240)
represents an ellipse in the plane (E1 , E2 ) with the center coinciding with the origin
of the coordinate system. We thus obtain a family of homothetic ellipses correspond-
2
ing to different values of Ucr if the latter is treated as an independent parameter.
2
Naturally, the larger values of Ucr correspond to larger semi-axes of the ellipse in the
2
plane. For small values of Ucr the ellipse in Eq. (6.236) and the region in Eq. (6.231)
are disjoint. For large values of it they have a common area. Therefore, there exist
2 2 2 2 2
two values of Ucr , denoted by Ucr,worst and Ucr,best with Ucr,worst < Ucr,best , for
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
which the two convex domains in Eqs. (6.231) and (6.236) share a single point. One
of the values corresponds to the case when one ellipse contains the other one en-
tirely, whereas the other point is associated with the case when they are on different
sides of the common tangent.
2 2 2
Values of Ucr falling outside of the interval [Ucr,worst , Ucr,best ] do not represent
feasible critical velocities, because the set (E1 , E2 ) will not satisfy Eq. (6.231). Note
2 2
that the two values of velocities, Ucr,worst and Ucr,best , respectively, represent the
smallest and the largest solutions allowed by Eq. (6.231). To determine these values,
we need to examine closely the relative spacing of the critical velocity ellipse and the
region of uncertain variation of the elastic modulus. According to the interpretation
of the parameters E1 and E2 as some effective elastic modulus, we shall confine our
analytical derivation to the arc of the ellipse in Eq. (6.236) contained in the first
quadrant of the plane (E1 , E2 ). It can be shown (see Zingales and Elishakoff (2001))
that, under some assumptions about the functions i (x), the critical velocity ellipse
in Eq. (6.240) is always oriented with its major axis in the second and fourth
quadrant of the plane (E1 , E2 ), as shown in Fig. 6.19.
2
Once the orientation of the ellipse, representing the critical velocity Ucr with
respect to the region C of the uncertain parameters E1 and E2 has been identified,
one can solve analytically the anti-optimization problem. The following sections
will be concerned with the solution of Problem (6.232) for various practical shapes
of the uncertainty region.
Before proceeding further, let us pose the following question: Why do we need
to consider various shapes of the uncertainty region? To answer this question, we
visualize that the results of the experimental measurements yield an ensemble of
functions Ei (x). These functions are then decomposed to sets of pairs (E1i , E2i ).
Each of these pairs forms a point in the plane (E1 , E2 ). We then are interested
in determining the region that contains all of these points. A convex hull of these
points is naturally such a set, forming a polygonal uncertainty region. Also, it
is easily visualized that different researchers and engineers may approximate the
uncertainty region via differing means. It makes sense, therefore, to treat different
possibilities.
(a) (b)
(c)
2
Fig. 6.20 Regions of uncertainty: (a) rectangular region of uncertainty, max 2
and min ellipses,
(b) polygonal region of uncertainty with particular location of vertices P 3 and P6 , (c) singular
location for rectangular region of uncertainty; extremal that does not coincide with one of the
vertices.
results in lower values of e. The particular choice of 1 (x) and 2 (x) in Fig. 6.20
has been made to enable an easier visualization of the ellipses corresponding to
2 2
extremal values min and max .
The objective of the next section is to derive approximations of the admissible
region for E1 and E2 by a continuous smooth curve for analytical purposes.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(a)
Eq.(6.244),
(b)
Fig. 6.21 Ellipsoidal region of uncertainty: (a) ellipse, (b) ellipse corresponding to the extremal
values of the coefficient 2 .
(E1,best , E2,best ) that satisfy Eqs. (6.244a) and (6.244b) and correspond, respec-
tively, to the worst and the best critical velocities Ucr,worst and Ucr,best. In the
following, we denote the coordinates of the points Pworst and Pbest , respectively, as
2 2
(e1,worst , e2,worst ) and (e1,best , e2,best ). Then Ucr,worst , and Ucr,best are given by
2
Ucr,worst = (d011 e21,worst + 2d012 e1,worst e2,worst + d022 e22,worst )b20 h60 ,
(6.246)
2
Ucr,best = (d011 e21,best + 2d012 e1,best e2,best + d022 e22,best )b20 h60
2
The obtained solution for the anti-optimized critical flutter velocity Ucr,worst shall
be utilized in the next section to obtain the best possible value of a certain objective
function involving it.
Fig. 6.22 Alternative location of the design box with respect to the constraint g 2 0.
Specifically, for the sake of determinacy, z is fixed at 10. The optimization problem
is then stated as
minimize W (b0 , h0 ) (6.253a)
subject to g1 (b0 , h0 ) = min b0 h30 U0 0 (6.253b)
g2 (b0 , h0 ) = 10h0 b0 0 (6.253c)
(b0 , h0 ) (6.253d)
The optimization problem (6.253) will be solved by employing geometrical con-
straints. Note that if in Eq. (6.253c) the entire region of variation is on the right side
of the line g2 (b0 , h0 ) = 0, then no acceptable solution exists. This case is represented
by the design region 1 drawn with a dotted line in Fig. 6.22. If, however, the region
is located on the left side of the line g2 (b0 , h0 ) = 0, the entire region is feasible
except for the additional constraint Eq. (6.253b) (see domain 3 in Fig. 6.22). If
the line g2 (b0 , h0 ) = 0 passes through the region, only the upper part (see the filled
part of the region 2 bounded by solid lines in Fig. 6.22) is feasible. We denoted
by Q1 and Q2 , respectively, the intersection points of the line g2 (b0 , h0 ) = 0 with
the region 2 that are closest and farthest from the origin O. Depending on the
location of the region 2 with respect to the origin, several possibilities arise. If the
line g2 (b0 , h0 ) = 0 crosses the edges AD and CD, the intersection points have the
coordinates
Q1 = (b0L , 1/10b0L), Q2 = (h0U , 1/10h0U) (6.254)
as shown in Fig. 6.22. Other possibilities of crossing the region by the line g2 = 0 are
elucidated in Figs. 6.23(a)(c), where the coordinates of the intersection points are
also indicated. To obtain the solution of the optimization problem in Eq. (6.253),
let us examine closely the critical velocity constraint g1 (b0 , h0 ) 0. If the curve
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(a)
(b)
(c)
Fig. 6.23 Feasible regions: (a) and (b) feasible design region e for specified velocity U0 , (c)
alternative location of the design box with respect to constraints g 1 0 and g2 0 for specified
velocity U0 .
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(a)
(b)
e (b) in corre-
Fig. 6.24 Solution of the optimization problem: (a) on the vertex of the region ,
spondence of points Q1 and Q2 in Eq. (6.254).
defined by
g1 (b0 , h0 ) = min b0 h30 U0 = 0 (6.255)
lies below the feasible region , then the constraints (6.253b) and (6.253c) are not
active for the specified value of the velocity U0 . In this case, the solution of the
optimization problem coincides with the closest vertex of the hatched region to the
origin O. This vertex, denoted hereinafter by Qopt , is either the point Q1 , as shown
in Figs. 6.22 and 6.23(b), or point A, as shown in Figs. 6.23(a) and (c), respectively,
if the straight line g2 (b0 , h0 ) = 0 intersects the edge AD or AB. However, if the
line in Eq. (6.255) is entirely above the hatched region , then no solution of the
optimization problem exists for the specified value of the velocity U0 . Thus, we
identify an interval [U0,min , U0,max ], where the specified velocity U0 must lie, so that
the inequality Ucr,worst U0 may hold. The extreme values U0,min and U0,max of
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(a)
(b)
whose smallest solution is = 5.07 106 . The values of e1 and e2 are then
e1 = 1.06 105 and e2 = 2.98 105 . In Fig. 6.25(a), the ellipse corresponding to
the minimum 2 and tangent to the region of uncertainty in Pworst = (e1 , e2 ) is
depicted with a solid line. We assume that the design variables b0 and h0 are in the
region
6 cm b0 30 cm, 0.5 cm h0 2.5 cm (6.268)
represented in Fig. 6.25(b). The lines corresponding to the extremal of the
interval [U0,min, U0,max ] are depicted, respectively, with dashed and dot-dashed
lines. The interval where the constraint g1 (b0 , h0 ) 0 is active is represented
by [1.9 103 , 1.02 106 ]. We set U0 = 900 m/s.
The region e of variation of the design variables b0 and h0 is depicted as a
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
6.7.9 Conclusion
We discussed the flutter of a beam with uncertain elastic modulus that is simply
supported at both ends in a stream of gas flow. The structural model has a variable
cross-section and variable elastic modulus along the axis of the beam to reflect the
realistic situation. We addressed the following question: How can the optimization
methods in the presence of the bounded uncertainty, which is neither probabilistic
nor fuzzy, be utilized? The parameters describing the elastic modulus were modeled
as variables belonging, respectively, to a polygonal hull of experimental points, or
to an ellipsoidal region. Hybrid optimization and anti-optimization proved to be an
effective tool for dealing with optimum design in the presence of uncertainty. We
first solved the problem to find an expression for the worst possible combination of
the uncertain parameters involved in one of the constraints of the design variables.
As a second step we optimized the design variables under the anti-optimized version
of one of the constraints. Note that the further studies on flutter problems in non-
probabilistic setting were conducted by Wang and Qiu (2009) and Khodaparasi,
Marques, Badcock and Mottershead (2009).
Suzuki (1997), Qiu, Chen and Elishakoff (1995a, 1996a, 1996b) Qiu, Chen and Jia
(1995b), Qiu, Elishakoff and Starnes (1996d), Qiu and Hu (2008), Qiu, Hu, Yang
and Lu (2008a), Qiu, Ma and Wang (2004c), Qiu, M uller and Frommer (2004e),
Qiu and Wang (2003, 2005, 2006), Sadek, Sloss, Adali, and Bruch (1993), Srini-
vasan, Corotis and Ellingwood (1992), Srinivasan, Ellingwood and Corotis (1991),
Venini (1998), Wang and Qiu (2009), Yamaguchi, Kogiso and Yamakawa (2007),
and Yoshikawa and Fr yba (1997).
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 7
In this chapter, we deal with anti-optimization in the context of the finite element
method, which is apparently the most universal method currently for structural
analysis. We first discuss the interval analysis of multi-degree-of-freedom systems.
Then we deal with interval finite element analysis of shear frame.
7.1 Introduction
211
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
eling of uncertainty was studied in the book by Ben-Haim and Elishakoff (1990), in
which a complete literature survey for continuous systems and several illustrations
for the applications in mechanics are presented. Elishakoff and Colombi (1993)
developed a hybrid stochastic-convex approach to solve some actual space-shuttle
problems.
Interval analysis is one of the tools for incorporating uncertainty in structural
analysis. We presented in Sec. 3.3 the basic properties of interval analysis using
simple mathematical and structural models. However, for practical design pro-
cess, response quantities, in the form of internal forces, stresses, displacements and
strains, are needed to evaluate the safety and serviceability of structures with many
degrees of freedom. Interval analysis for multidimensional structures is studied in
Muhanna and Mullen (2001) and Muhanna, Zhang and Mullen (2007). Moens and
Vandepitte (2005) summarized the interval finite element methods for static and
eigenvalue analyses. In this chapter, we present some results of FEM-based interval
analysis of static responses.
i.e.,
AH B = {A1 b|A A, b B} (7.5)
On the other hand, the matrix inverse A1 of a regular interval square matrix
A is defined as
A1 {A1 |A A} (7.6)
Then the following relation holds:
AH B A1 B (7.7)
where the equality holds if A is diagonal.
The problem of finding S(A, B) is a combinatorial problem that cannot be
solved in a practically acceptable computational time for a large-scale problem.
Therefore, our purpose here is to find the narrowest possible interval vector X
satisfying
AH B = S(A, B) X (7.8)
for which many algorithms exist such as interval Gaussian elimination, that may
break down for a large-scale problem (Nickel, 1977), GaussSeidel iteration (Gay,
1982), and its generalization (Neumaier, 1982; Rump, 1992). The typical approach
utilizes a regular n n transfer matrix R for converting the problem Ax = b to a
problem of finding the fixed point g(x) = x of an operator g(x):
g(x) = x R(Ax b) = Rb + (I RA)x (7.9)
where I is the n n identity matrix (see Muhanna, Zhang and Mullen (2007) for
details of this approach).
As an example, consider the case when uncertainty exists only in some compo-
nents of A as
2 [1, 0] 1.2
A= , B= (7.10)
[1, 0] 2 1.2
By computing x for the four cases (A12 , A21 ) = (1, 1), (1, 0), (0, 1), and
(0, 0), we have
0.4 0.3 0.6 0.6
x = A1 b = , , , (7.11)
0.4 0.6 0.3 0.6
Therefore, AH B is obtained as
H [0.3, 0.6]
A B= (7.12)
[0.6, 0.3]
On the other hand, we can compute A1 for the four cases (A12 , A21 ) =
(1, 1), (1, 0), (0, 1), and (0, 0) as
1 2/3 1/3 1/2 1/4 1/2 0 1/2 0
A = , , , (7.13)
1/3 2/3 0 1/2 1/4 1/2 0 1/2
Therefore, the interval of A1 is obtained
as
1 [1/2, 2/3] [0, 1/3]
A = (7.14)
[0, 1/3] [1/2, 2/3]
and A1 B results in
[0.2, 0.8]
1
A B= AH B (7.15)
[0.8, 0.2]
Hence, the relation (7.7) is satisfied.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
U1 U2
P1 P2
K1 K2
For a general finite element analysis problem, the explicit form of the displace-
ment vector cannot be obtained. Muhanna, Mullen and Zhang (2005) noted that
the accuracy of the interval vector of the displacements depends on the order of the
sequence of arithmetic operations in the solution process by finite element analysis.
They presented an element-by-element approach to delay the operation involving
the coupling of the interval variables, where the penalty terms are used to ensure
the compatibility among the element displacements as described below.
In the standard procedure of finite element analysis, the element stiffness matri-
ces are assembled to formulate the global stiffness matrix. If we write the stiffness
matrix of member i of the 2-bar structure as
Ki Ki
Ki = (7.23)
Ki Ki
then the four components of the matrix are treated as independent variables in the
interval analysis, although they depend on Ki only. If we write K i as
Ki 0 1 1
Ki = (7.24)
0 Ki 1 1
then the number of independent variables in interval analysis is reduced to two, and
the variables exist in a diagonal matrix.
Another source of overestimation of the intervals of displacements is the coupling
of elements through the common nodes; e.g., the term K1 + K2 appears due to the
coupling of the bars at the center node 1 of the 2-bar structure in Fig. 7.1. The
Lagrange multiplier approach can be used to delay the coupling in finite element
analysis (Muhanna and Mullen, 2001). The analysis problem is formulated as a
minimization problem of the total potential energy
1
= U > KU U > P (7.25)
2
under constraints
CU = V (7.26)
where C and V are the specified matrix and vector, respectively, to represent the
compatibility conditions between the elements. Note that the structure of K in
Eq. (7.25) is different from that of the conventional global stiffness matrix; it has
a block-diagonal form of the element stiffness matrices. Accordingly, the nodal
displacements of different elements at the same node are regarded as independent
variables. Although the definitions and sizes of K, U , and P are different from
those for conventional analysis, the same notations are used for simplicity. Let n e
and m denote the number of degrees of freedom of each element and the number of
elements, respectively. Then the size of K is ng ng with ng = ne m, and U and
P also have ng components. For example, for the 2-bar structure in Fig. 7.1, U is
written as
(1) (1) (2) (2)
U = (U0 , U1 , U1 , U2 )> (7.27)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where the superscript ( )(i) denotes the value in the ith bar, and the fixed support
is denoted by node 0. The compatibility conditions are given as
(1)
U
0(1)
1 0 0 0 U 1 0
(2) = (7.28)
0 1 1 0 U1 0
(2)
U2
The stationary conditions for with incorporated constraints (7.26) using the
vector Q of Lagrange multipliers lead to the following system of equations:
K C> U P
= (7.29)
C 0 Q V
The reader may consult Sec. 2.4 for details of the Lagrange multiplier approach.
Suppose the interval form K of K is expressed as K = DS, where D is the
uncertain diagonal matrix, and S is a deterministic matrix as demonstrated in
Eq. (7.24) for an element stiffness matrix. Then the first equation of Eq. (7.29) is
written in an interval form
DSU = P C > Q (7.30)
where P is assumed to be deterministic, for simplicity. Using V = 0 for gen-
eral compatibility conditions without forced displacements, the interval form of the
second equation in Eq. (7.29) is written as
CU = 0 (7.31)
>
which is premultiplied by C and added to Eq. (7.30) to obtain
D(S + C > C)U = P C > Q (7.32)
Eq. (7.32) can be rewritten as
DHU = P C > Q (7.33)
>
with H = S + C C. Hence, U is obtained as
U = H 1 D1 (P C > Q) (7.34)
1
where D is the exact inverse of the diagonal interval matrix D.
Note that Q has the physical meaning as internal force vector. Therefore, for
a statically determinate structure, for which the internal loads do not depend on
the stiffnesses of the elements, the vector P C > Q is deterministic, correspond-
ing to the deterministic (nominal) value of Q. For the interval of displacement
vector, suppose D is formulated so that each interval stiffness parameter appears
only once. In this case, the exact interval U can be successfully computed from
Eq. (7.34), if the structure is statically determinate. However, for a statically in-
determinate structure, the use of deterministic (nominal) value of Q in Eq. (7.34)
leads to underestimation of the interval U. By contrast, the use of Eqs. (7.31) and
(7.32) for computing the interval Q results in overestimation of U. Muhanna and
Mullen (2001) presented an iterative approach to obtain a sharp interval hull for U.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
5
K5 /2 K5 /2
4
K4 /2 K4 /2
3
K3 /2 K3 /2
2
K2 /2 K2 /2
1
K1 /2 K1 /2
those of the interval element stiffness matrices. Elishakoff, Ren and Shinozuka
(1997) applied Fuchs method to statically determinate and indeterminate beams
with random stiffness.
According to the BernoulliEuler beam theory, the displacement field v(x) in
the transverse direction of the eth element is related to the loading field p(x) by
[(EI)e (x)v(x),xx ],xx = p(x) (7.36)
where x is the coordinate along the beam axis, (EI)e (x) is the bending stiffness
field, and ( ),x denotes differentiation with respect to x. The assumptions and the
differential relationship of the BernoulliEuler beam theory are assumed to be valid
even when (EI)e (x) possesses some uncertainties. This means that (EI)e (x) is such
that these assumptions are not violated.
In what follows, the uncertain field (EI)e (x) is modeled using set-theoretical
interval approach, where (EI)e (x) is assumed to be bounded from below and above
with constant envelope as
(EI)Le (EI)e (x) (EI)U
e , (0 x Le ) (7.37)
where (EI)Le and (EI)U e are the lower and upper bounds for (EI)e (x), and Le
is the length of element e. The central value of the bending stiffness becomes
(EI)0e = [(EI)Le + (EI)Ue ]/2.
For a shear frame in Fig. 7.2, the rotation at each end of the column element
is fixed. Therefore, the uncertain displacement field V(x) of an element is approx-
imated as a linear combination of the uncertain vector of two-degree-of-freedom
transverse nodal displacements Ve multiplied by the independent vector of cubic
interpolation functions N (x) = (N1 (x), N2 (x))> :
V(x) = N > (x)Ve (7.38)
where
3 2 2 3 2
N1 = 1 3
x + 2 x3 , N 2 = 3 x2 2 x3 (7.39)
Le Le Le Le
Following the standard finite element methodology, the interval stiffness matrix
Ke of the beam element with (i, j)-component Ke,ij is evaluated as
Z Le
d2 Ni (x) d2 Nj (x)
Ke,ij = (EI)e (x) dx (7.40)
0 dx2 dx2
where (EI)e (x) is the interval bending stiffness field. Owing to the uncertainty in
the bending stiffness, all components of Ke are uncertain, yet bounded from above
and below, as follows, as a result of the interval model:
L U
Ke,ij Ke,ij Ke,ij (7.41)
Let Re be defined as
12(EI)e
Re = (7.42)
L3e
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
with its lower and upper bounds ReL and ReU , respectively. The lower bound K Le and
the upper bound K U e of K e are found, as follows, from minimization/maximization
of the integral in Eq. (7.40)
for each i and j independently
using
Eq. (7.37):
ReL ReU ReU ReL
K Le = , KU e =
(7.43)
U L L U
Re Re Re Re
We consider the eigenvalues 1 and 2 (1 < 2 ) of K e . The independent
computation for each Ke,ij in Eq. (7.43) gives a four-dimensional rectangular prism
L U
with sides defined by the coordinates Ke,ij and Ke,ij , within which K e lies for
any realization of (EI)e (x), and leads to conservative intervals compared to the
intervals where the interdependency is taken into account. For that reason, working
with Eq. (7.43) to obtain the bounds for the eigenvalues of the element interval
matrix leads to very conservative results. Moreover, the smallest eigenvalue that is
obviously equal to zero for any realization of a beam element without constraining
rigid-body motions would not become zero, but become an interval value containing
zero. Thus, the eigenvalues are first computed from the integral of Eq. (7.40)
directly, which leads to
Z Le
72
1 = 0, 2 = (EI)e (x) 6 2 2
dx (7.44)
0 L e (L e 4L e x + 4Le x )
Hence, the lower and upper bounds L2 and U 2 , respectively, on the second eigen-
value 2 can be obtained from minimization and maximization of 2 in Eq. (7.44),
respectively, as
L2 = 14ReL 12ReU, (7.45a)
U L
2 = 12Re + 14Re
U
(7.45b)
0
Then, the central value 2 and the radius 2 of 2 are computed as
L + U
02 = 2 2
= ReL + ReU
2 (7.46)
U L
2 = 13(Re Re )
We next evaluate the displacements of a general multi-story shear frame sub-
jected to static loads. Let K and P denote the global stiffness matrix and the
applied nodal loads, respectively. We assume that K is an interval matrix, but
P is a deterministic vector, for simplicity. Then the nodal displacement vector U
is also an interval vector denoted by U. One can write down the interval nodal
displacements as a function of interval eigenvalues as
U = K1 P = (Q1 )> J 1 Q1 P (7.47)
where J 1 is a diagonal matrix with 2/ 2 for all the diagonal
components, and
1 1
1 1
Q= (7.48)
1 1
1
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
12(EI)0e /L3e = 1.01. From Eq. (7.46), we obtain L2 = 0.98 and U 2 = 3.06. The
bounds for the global nodal displacements are given in Table 7.2, which covers all
possible displacements for realizations of the shear frame within the bounds on Re .
In addition to such a relation, non-dimensional quantities wU1 , . . . , wU5 , which are
called coefficients of interval uncertainty and defined as the ratio of the radius to the
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
U1 5 15 1 1
U2 9 29 1 0
U3 12 41 0 0
U4 14 50 2 2
U5 15 55 3 7
Table 7.2 Lower and upper bounds for the global nodal displacements of
the shear frame with interval stiffness properties.
absolute value of the central value, are obtained as shown in Table 7.3 for U1 , . . . , U5 .
Note that in Table 7.3 means that the coefficient cannot be computed, because
the central value vanishes.
In this section, we follow the basic ideas of interval analysis of beams and frames
subjected to pattern loads (K oyl
uoglu, C
akmak and Nielsen, 1995); however, the
notations are modified to be consistent with other parts of the book. Furthermore,
the concepts are presented in a simpler form, and only new numerical examples are
given.
Consider a BernoulliEuler beam subjected to a uniformly distributed load p,
which exists in the interval [p0 p, p0 + p]. Based on the conventional finite
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
2.2
2.1
U2
2.0
1.9
1.8
0.9 1.0 1.1
U1
Fig. 7.3 Bounds of nodal displacements of the 2-bar structure in Fig. 7.1.
p1 p2
1 (1) 2 (2)
2 4
1 3
Fig. 7.4 A cantilever beam consisting of two elements subjected to distributed loads.
The nominal values of the loads are P10 = P20 = 1, and the load uncertainty is given
by P1 = P2 = 0.1, which are independent of each other. Using Eq. (7.58), the
intervals of U1 and U2 are obtained as [0.9,1.1] and [1.8,2.2], respectively.
The displacements for the loads at the vertices of its feasible region (0.9,0.9),
(0.9,1.1), (1.1,0.9), (1.1,1.1) are obtained as (0.9,1.8), (1.0,2.1), (1.0,1.9), (1.1,2.2),
respectively. Since the nodal displacements are linear functions of the nodal loads,
the existable region of the nodal displacements (U1 , U2 ) is a convex region as indi-
cated in the gray area of Fig. 7.3, which verifies that the narrowest intervals of U 1
and U2 are [0.9,1.1], and [1.8,2.2], respectively.
Consider next the cantilever beam as shown in Fig. 7.4 consisting of two beam
elements, where the displacement numbers are defined in the lower figure. The
beam is subjected to uniformly distributed loads p1 and p2 on members 1 and 2,
respectively. Suppose the length L and the bending stiffness EI are 1, for simplicity,
for the two members. Then the inverse of the stiffness matrix is obtained as
2 3 5 3
1 3 6 9 6
K 1 = (7.60)
6 5 9 16 12
3 6 12 12
The load vectors corresponding to the nominal values p01 = p02 = 1 of distributed
loads are given as
6 6
0 1 1
, p2 =
0 1 1
p1 = (7.61)
12 0 12 6
0 1
When the radii p1 and p2 of uncertainty are equal to 0.1, the upper bound
P U = P 0 + P and the lower bound P L = P 0 P are obtained as
1.1000 0.9000
0.0167 0.0167
PU = L
0.5500 , P = 0.4500
(7.62)
0.0750 0.0917
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Since all components of K 1 are positive, the upper and lower bounds U U and U L
are easily computed from Eq. (7.58) as
0.7959 0.6208
1.3167 1.0166
UU = L
2.2584 , U = 1.7416
(7.63)
1.5167 1.1499
We next use Eq. (7.57) for the same nominal values and radii of p1 and p2 . Then
from the displacements corresponding to the loads (p1 , p2 ) = (1.1, 1.1), (1.1,0.9),
(0.9,1.1), and (0.9,0.9), the following upper and lowers bounds are obtained:
0.7792 0.6375
1.2833 1.0500
UU = L
2.2000 , U = 1.8000
(7.64)
1.4667 1.2000
Therefore, narrower bounds are obtained by evaluating the uncertainty in the dis-
tributed loads; i.e., evaluation after converting to nodal loads leads to a conservative
bound, because the number of independent variables increases from two to four as
a result of conversion.
Chapter 8
8.1 Introduction
227
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
are not certain, and as far as they are certain, they do not refer to reality. How
much more does this statement appear to apply to engineering, where numerous
uncertain factors are encountered almost in every problem? The pertinent question
arises of how to model this uncertainty. It is nowadays widely accepted that there
are three competing methods of describing uncertainty, as discussed in the preface of
this book; namely, the probabilistic method, the fuzzy-sets based approach, and the
method which is referred to as anti-optimization (Elishakoff and Ben-Haim, 1990a)
or as the guaranteed approach by Kurzhanski (1977) and by Chernousko (1994).
Each of these approaches has generated numerous research papers and monographs.
Yet the systematic comparisons among these three approaches are still missing.
Naturally, there may be a strong argument put forward against such a comparison,
in the first place. Indeed, these three approaches provide judgments of different
kinds: The probabilistic approach furnishes the reliability of the structure, namely,
the probability that the stresses, strains or displacements of the structure will not
take values over some safe threshold. The fuzzy-sets based approach yields the
safety margin via the concept of a membership function. Anti-optimization (or the
guaranteed approach) provides the least favorable, maximum displacements which
can be contrasted with the threshold values. If the maximum response does not cross
the threshold, we can guarantee that the failure will not occur. This corresponds to
the statement that if the least favorable, or an anti-optimized, response is below the
failure boundary, a successful performance will be secured. This particular thought,
namely, that three approaches provide different types of answers, precluded the very
attempt to compare them for a long time.
The first numerical comparison has been performed independently by Elishakoff,
Cai and Starnes (1994a) and Kim, Ovseyevich and Reshetnyak (1993) in different
but complementary contexts: The former study dealt with nonlinear static buck-
ling, whereas the latter investigated linear dynamic response. In the paper by
Elishakoff, Cai and Starnes (1994a), for certain regions of variations of the pa-
rameters, the design values of the buckling loads obtained via applications of two
alternative approaches were found to be extremely close to each other. In the paper
by Kim, Ovseyevich and Reshetnyak (1993), the design value corresponding to the
mean plus three times standard deviation, was adopted. Elishakoff and Li (1999)
studied a benchmark static problem, leading to direct contrast of the probabilis-
tic and non-probabilistic analyses of uncertainty. General understanding on the
limitation of probabilistic methods is summarized in Elishakoff (2000a). Langley
(2000) presented a unified framework of combined probabilistic and possibilistic
(deterministic) approaches.
In Secs. 8.2 and 8.3, probabilistic and anti-optimization approaches are compared
for impact buckling of uniform columns possessing single and multiple uncertain
initial imperfections following the studies by Elishakoff and Zingales (2000) and
Zingales and Elishakoff (2000), respectively.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
P(t) P(t)
h
w
We confine ourselves to the case where the initial imperfection w0 (x) has a multi-
plicative representation
x
w0 (x) = h sin (8.7)
L
where h is the magnitude of initial imperfection. Following Elishakoff (1978a) we
will use the non-dimensional quantities
x w0 w P
= , = 1 t, u0 = , u= , = (8.8)
L Pcl
where is the non-dimensional axial coordinate, is the non-dimensional time,
u0 () is the non-dimensional initial displacement, u(, t) is the non-dimensional
additional displacement, and is the non-dimensional axial load. In Eq. (8.8)
2 r EI r
I EI 2
1 = , = , Pcl = (8.9)
L mJ J L2
where 1 is the fundamental circular natural frequency of the ideal column, i.e., of
a column with neither initial imperfections nor axial load, is the radius of inertia,
and Pcl is the classical Euler buckling load. The differential equation (8.3) reduces
to the following non-dimensional one:
2 2 2
4u 2 u 4 u 2 u0
+ + = (8.10)
4 2 2 2
We introduce the non-dimensional initial imperfection g as
h
g= (8.11)
The additional displacements are expressed in a separable form analogous to
Eq. (8.7), namely,
u(, ) = e() sin() (8.12)
where e() is a time-dependent function. Substituting Eqs. (8.7) and (8.12) in
conjunction with Eq. (8.11) into Eq. (8.10) yields the following ordinary differential
equation with respect to the function e():
d2 e()
+ (1 )e() = g (8.13)
d2
The solution satisfying the initial conditions (8.6) is obtained as
g
(cosh(r) 1), > 1
1
1 2
e1 () = g , =1 (8.14)
2
g
(1 cos(r)), <1
1
where
p
r= | 1| (8.15)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
mean unsafe regions, or, as defined by Hoff, as regions of impact buckling. Here-
inafter we denote random variables with capital letters, whereas the possible values
they take on are designated with lower-case notation. Reliability R() at a pre-
selected non-dimensional time is defined as a probability that the event described
in Eq. (8.19) takes place:
R() = Prob(d Z(, ) c) (8.21)
Define V () as
V () = Z(1/2, ) (8.22)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
V () = Ga() (8.24)
the reliability varies between zero and unity. Generally, in engineering practice, the
codified reliability rc is not taken to be unity. Rather, it is a pre-selected quantity in
extremely close vicinity of unity. For rc tending to unity from below, the inequality
in Eq. (8.37) is valid. Hence the argument in the second term in Eq. (8.34) is
negative. Therefore, it vanishes, and we are left with
1
1 c(1 4 /M )
R() = 1 (8.38)
2 1 cosh(r) 4 /M
In order for the reliability to tend to unity from below, it is necessary and sufficient
that
c(1 4 /M )
lim = 2 (8.39)
d cosh(r) 4 /M
where d is the design value of the cross-sectional radius. For this we obtain the
following equation:
c(1 4d /M )
p = 2 (8.40)
cosh( M/4d 1) 4d /M
c = 0.5
0.8 1 = 0.1
2 = 0.6
0.6
R()
0.4
0.2
R = 0.133
0 1 2 3 4 5 6
Fig. 8.2 Reliability versus non-dimensional time; initial imperfections with uniform probability
density (P = 0.5Pcl ), Eqs. (8.25c) and (8.30).
c = 0.5
0.8 1 = 0.1
2 = 0.6
0.6
R()
0.4
0.2
Fig. 8.3 Reliability versus non-dimensional time; initial imperfections with uniform probability
density (P = Pcl ), Eqs. (8.25b) and (8.30).
whereas for P < Pcl this equation takes the following form:
p
2 [4d /M cos( M/4d 1)]
c(d , t, E, m, L) = (8.44)
1 4d /M
The results of sample calculations are presented in Figs. 8.28.4 portraying relia-
bility R() versus non-dimensional time for P < Pcl , P = Pcl , and P > Pcl , respec-
tively, for c = 0.5, 1 = 0.1, 2 = 0.6. In the following, we assume the mechanical
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
c = 0.5
0.8 1 = 0.1
2 = 0.6
0.6
R()
0.4
0.2
0 0.5 1 1.5 2
Fig. 8.4 Reliability versus non-dimensional time; initial imperfections with uniform probability
density (P = 2Pcl ), Eqs. (8.25a) and (8.30).
3.5
Probability of initial failure
3
2.5
fG (g) 2
1.5
0.5
Fig. 8.5 Geometrical representation of the reliability at the initial time instant.
Boundary c (cm)
3 II
1 I
c = 0.6
4
M = 3.86
0 2 4 6 8 10
Radius (cm)
Fig. 8.6 Design curve c = c(d ), t = 0.2 with G having a uniform probability density and unity
reliability requirement, Eq. (8.44).
the values c = 0.5 and the uppermost value that G may take, namely, 0.6, equals
2(0.6 0.5) = 0.2, which constitutes the probability of failure Pf () at the initial
time = 0 denoted by Pf (0) as shown in Fig. 8.5. This means that the reliability
curves in Figs. 8.28.4 have an initial value equal to R(0) = 1 Pf (0) = 0.8. This
probability is equivalent to the failure probability of the item at delivery.
For the actual load less than the classical buckling load, the variation of the total
displacement with time is harmonic. When approaches value /r, the absolute
value of the total displacement attains its maximum. At this time, the reliability
of the system reaches its minimum. Values of larger than /r yield a constant
value of reliability achieved at the latter time instant
R() = R(/r) for /r (8.46)
In Fig. 8.2 the value /r equals 4.1. Therefore, the reliability in Fig. 8.2 remains
constant and equals 0.15. Figure 8.6 shows the relation between the failure boundary
and the cross-sectional radius d with other parameters fixed. This curve enables us,
by specifying the failure boundary c, to obtain the smallest required cross-sectional
radius yielding the prescribed reliability for the prescribed external load. One must
observe that with values of c less than 2 = 0.6, it is not possible to design the
column in a manner so as to attain a unity or near-unity reliability. In fact, if
approaches infinity, the column gains an infinite frequency of vibration and the
displacement at the center remains fixed at the initial imperfection value. Naturally,
if the failure boundary happens to be less than or equal to 1 , i.e., the minimum
value that the initial imperfection can assume, there is no finite value of d that
will ensure extremely high reliability, since the structure buckles at the outset of
the experiment. Likewise, if c [1 , 2 ) the reliability that is arbitrarily close to
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
8
6 5
Boundary c
4
4
2
3
Time t (sec)
2 2
4
6 1
Radius (cm)
8
10
Fig. 8.7 Design surface c = c(d ), with G having a uniform probability density and unity relia-
bility requirement, Eq. (8.44).
unity cannot be obtained, since the non-zero probability of failure occurs. On the
other hand, when c 2 , the initial probability of collapse is absent, and the design
with near-unity reliability is feasible.
Figure 8.7 depicts the surface = (c, t) for a prescribed value of the external
load and unity reliability. With increase of the time interval [0, t], where safe perfor-
mance is required, the value of the failure boundary allowing unity reliability should
be greater. Naturally, values of d are larger than the ones obtained from such a
surface for the pre-selected time instant t corresponding to unity reliability. This
behavior is elucidated in Fig. 8.8, where the design curves are presented for various
values of the design times t1 = 0.3 sec, t2 = 0.8 sec, and t3 = 2.5 sec. Figure 8.9
represents the cross-section of the surface in Fig. 8.7 with the plane t = 0.5 sec
(solid line), contrasted with other design curves, which are obtained with smaller
values of the required reliability, namely, rc = 0.9 or 0.8, at the same time instant.
It is seen that, by requiring less reliability, the design value of the cross-sectional
radius decreases, as expected, for the same pre-selected value of the safety bound c.
We note that in Figs. 8.6 and 8.8 the abscissa axes are divided into two sub-
regions corresponding to different behaviors of the system. We denote the value 4 M
by dc . If the design radius d satisfies d dc , then the load ratio P/Pcl satisfies
P/Pcl 1, corresponding to region I. In this region, the displacement max z(, )
is an exponential function of the time t. The design relationships, conveniently
expressed as the failure boundary function depending on the variables d , t, E, m
and L as c(d , t, E, m, L), takes the form of either Eq. (8.42) or Eq. (8.43). The
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Boundary c (cm)
Dashed line t 0.8 sec
5
Dot-Dashed line t = 2.5 sec
4
2 II
1 I
0 2 4 6 8 10
Radius (cm)
Fig. 8.8 Design curves c = c(d ), with G having uniform probability density and unity reliability
requirement for different safe time instants, Eq. (8.44).
Solid line rc = 1
II
2
1 I
0 2 4 6 8 10
Radius (cm)
Fig. 8.9 Comparison of design curve, with G having uniform probability density and different
codified reliabilities, t = 0.5, Eq. (8.44).
argument of the hyperbolic function in Eq. (8.42), with the aid of Eq. (8.41), takes
the form
s p r
M (d )4 M 2 E
1 =t (8.47)
(d )4 d L m
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
By inspecting Eqs. (8.42) and (8.47), we conclude that the failure boundary function
in Eq. (8.42) is a monotonically decreasing function of the cross-sectional radius d
in region I as shown in Figs. 8.68.8. On the other hand, if d > dc ; i.e., region II
in Figs. 8.68.8, then the displacement function is represented by a non-monotonic
trigonometric function Eq. (8.17c). In this case, the function c(d , t, E, m, L) is
given in Eq. (8.44) with the attendant argument of the cosine function in Eq. (8.47).
Inspecting the expression of the function c(d , t, E, m, L), we note that its maximum
value is reached for the first time when the expression in Eq. (8.47) reaches . If,
for a combinations of design radius d and time t, the expression in Eq. (8.47) has
a value larger than , then the displacement in Eq. (8.17c) has already reached
its maximum value at a time t < t. Therefore, in order for the column to perform
satisfactorily with the required reliability until the specified time t, the trigonometric
function in Eq. (8.44) must acquire the value cos() = 1, and Eq. (8.44) becomes
d /M + 1
c = 2 (8.48)
d /M 1
Interestingly, an asymptotic behavior of the design curves may be observed from
Figs. 8.68.8. The asymptotic value of the cross-sectional radius d corresponds to
an infinite value of the first natural circular frequency of vibration 1 in Eq. (8.9).
In these circumstances the column behaves statically at every time instant t. Unity
reliability R() is achieved by setting the failure boundary c equal to the maximum
value of the initial imperfection g. The above discussion is confirmed analytically
by calculating the limit of the expression as
lim c = (2 1 )rc + 1 (8.49)
d
which, in case the specified reliability rc = 1 , will yield 2 as the final result.
c = 0.5
0.8 1 = 0.1
2 = 0.6
0.6
R()
0.4
R = 0.233
0.2
0 1 2 3 4 5 6
Fig. 8.10 Reliability versus non-dimensional time, with G having truncated exponential proba-
bility distribution (P = 0.5Pcl ), Eqs. (8.25c) and (8.53).
c = 0.5
0.8 1 = 0.2
2 = 0.6
0.6
R()
0.4
0.2
Fig. 8.11 Reliability versus non-dimensional time, with G having truncated exponential proba-
bility distribution (P = Pcl ), Eqs. (8.25b) and (8.53).
value of reliability R(). We are interested in the design of a column with near-
unity reliability. It may be found from Eq. (8.53) that the necessary and sufficient
condition to achieve this value reads
c
2 (8.56)
a()
where the function a() assumes one of the three expressions (a), (b) and (c) in
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
c = 0.5
0.8 1 = 0.1
2 = 0.6
0.6
R()
0.4
0.2
0 0.5 1 1.5 2
Fig. 8.12 Reliability versus non-dimensional time, with G having truncated exponential proba-
bility distribution (P = 2Pcl ), Eqs. (8.25a) and (8.53).
Solid line r0 = 1
6
Dashed line r0 = 0.83
5 Dot-Dashed line r0 = 0.75
Boundary c (cm)
3 II
2
I
1
0 2 4 6 8 10
Radius (cm)
Fig. 8.13 Comparison of design curves for different codified reliabilities, t = 0.5 with imperfections
possessing truncated exponential distribution, Eq. (8.44).
Eq. (8.25) depending upon the value of the external load P . Specifying the appro-
priate expression for a() and choosing the time interval [0, t] required for successful
performance, Eq. (8.56) reduces to either Eq. (8.42), Eq. (8.43) or Eq. (8.44), re-
spectively, for P > Pcl , P = Pcl or P < Pcl . Thus we conclude that, although
the imperfections possess an exponential distribution curve, the curve in Fig. 8.6
obtained for uniformly distributed imperfections can be used for the column design
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
The reliability R() of the column is given through Eq. (8.27) with FG (c/a())
represented by Eq. (8.60).
In order to design a column with an extremely high reliability requirement, it is
sufficient that the argument c/a() 2 , resulting in
Z c/a()
c
lim FG = lim AfG (g)[hg 1 i hg 2 i ]dg
c/a()2 a() c/a()2
Z c/a()
= lim AfG (g)[hg 1 i hg 2 i ]dg
c/a()2 1
Z 2
= AfG (g)dg (8.61)
1
=1
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where c is the failure boundary. Bearing in mind Eq. (8.63) in conjunction with
Eq. (8.25), we obtain
a()2 c (8.65)
We note that the design condition in Eq. (8.65) coincides formally with Eq. (8.56),
when the required reliability approaches unity from below. This proves the equiva-
lence between the results furnished by interval analysis and probabilistic one, when
rc attains values extremely close to unity.
8.2.7 Conclusion
A one-dimensional impact-buckling problem has been studied in this section in
order to contrast the probabilistic and non-probabilistic analyses of uncertainty.
Simplicity of the example allows us not to lose sight of the forest for the trees of
analytical and numerical derivations. In particular, this example illustrates that
probabilistic analysis and non-probabilistic interval analyses are compatible with
each other. In probabilistic analysis, one postulates the knowledge of the probability
density function. However, such information is often unavailable. To model this
situation, we constructed two alternative models of probability density utilizing
a random variable with either uniform or non-uniform truncated density. Both
yielded the same value for the design variable, if the required reliability is extremely
close to unity. This result leads to the general conclusion that irrespective of the
density involved, if it extends between the lower possible value 1 and the upper
possible one 2 , the same designs will follow. This could be interpreted as good
news for the probabilists. Researchers in the field of stochasticity could argue
that the unavailability of the probability density, except the information on its
boundedness, may be immaterial. Thus any assumption on the probability density
will be equally efficient. If, however, the experimental data is available, then an
appropriate density should be utilized. We maintain that a design point at the
bound of the distribution provides a near-unity probability. This seems blatantly
transparent; yet this fact appears not to be emphasized by the probabilistic analysts.
Even though the knowledge of the probability density function is not needed in
interval analysis, one must face the issue on how to properly define bounds. Except
in special cases where there exist well defined bounds, the selection of the bounds
will likely rely on a decision, perhaps based on data analysis or physical arguments.
Remarkably, the same designs furnished by the probabilistic analysis are also
obtained through the use of the interval analysis. This suggests the following idea:
Since the initial imperfection is an uncertain value, and often no data is available
to justify the particular choice of the density, why not employ the simplest possible
approach? Interval algebra is indeed such a method. Although in this particular
case interval analysis may appear simple, it provides the same result as derived by
more sophisticated and algebraically lengthier probabilistic analysis. It is concluded,
therefore, due to engineering pragmatism, that in some circumstances the non-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
8.3.1 Introduction
In Sec. 8.2, a direct comparison was performed between the widely popular prob-
abilistic methods and interval analysis, which represents a simplest version of the
anti-optimization technique. Immediately a natural question arises: How can we
compare these two competing approaches in the case of the vector uncertainty?
This problem is investigated in this section.
Whereas in a single-dimensional case uncertainty can be identified as a variable
belonging to an interval, the two-or-more-dimensional case presents more possibil-
ities. For example, in a two-dimensional case, the uncertain vector is identified
with two coordinates varying in a rectangular region, which enables one to use in-
terval analysis in a vector setting. It is intuitively understood that the narrowest
interval shall be chosen to characterize the one-dimensional uncertainty; for the
two-dimensional case one should seek a rectangle of minimum area so that the fur-
ther evolution of the system will be more closely bracketed. Yet, this may not be a
best representation of the available data whose scatter must be modeled. Indeed, in
some cases, the use of regions other than the rectangle may result in an even smaller
area enclosing all available data. The possibility arises, for example, of enclosing
the data by the minimum area ellipse, whose area may turn out to be smaller than
that of the minimum-area rectangle. Along these thoughts, in addition to interval
analysis, ellipsoidal modeling was developed (Schweppe, 1973; Chernousko, 1994;
Kurzhanski and Valyi, 1992) for uncertainty analysis (see Sec. 3.4 for details of
ellipsoidal model). Interestingly, these two lines of thoughts (on intervals and ellip-
soids) intersected in extremely few works, and essentially have been developed in
parallel, mostly without knowledge about the developments in the other field.
It appears that the ellipsoidal framework has some advantages over interval anal-
ysis in the sense that it deals with a smooth, convex boundary of the enclosed data
with associated straightforward analytical or numerical treatment. Yet the ellip-
soidal model may suggest that the components are functionally dependent. This
tacit assumption may be unjustified in some circumstances. Hence the independent
data may be better justified as enclosed by a rectangular region. Interval and ellip-
soidal modeling are particular cases of convex modeling (Ben-Haim and Elishakoff,
1990; Ben-Haim, 1996). In fact, convex description of uncertainty is richer than the
ellipsoidal one: In addition to the ellipsoids per se, it includes sets with functions
with envelope bounds, or those with bounded integral rectangles, or those with
bounded integral rectangles of its derivatives and so on.
The case where the uncertain variables do not belong to a convex set may be
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
a=5
P(t) P(t)
h2 h
1
Initial imperfections
L/2
L
dealt with using the methods of nonlinear programming (Li, Elishakoff, Starnes and
Shinozuka, 1996; Luenberger, 2003) (see Sec. 2.4.2 for details of nonlinear program-
ming). All these analyses share the main ideas of anti-optimization, namely, of the
desire to determine least favorable responses, in order to guarantee the successful
performance despite the presence of uncertainty.
In this section, the two-dimensional uncertain imperfections are considered in
the context of the column impact problem (Zingales and Elishakoff, 2000), whose
one-dimensional counterpart was discussed in Sec. 8.2.
We are interested in finding the reliability of the system; namely, the probability
that the total displacement will remain in a safe region, in accordance with Hoffs
criterion given by Eq. (8.19) in Sec. 8.2.
The reliability is defined as the probability that the maximum displacement
Y = max Z(, ) remains in a safe region. For simplicity we assume that a is an
odd number. Hence
Y max Z(, ) = Z(1/2, ) = V1 () + V2 () (8.81)
2 2
f G 1 G2
1
0 1.5
g2
1
1
1.5
g1
2
respectively, as shown in Fig. 8.15. Moreover, we assume, for the sake of simplicity,
that j > 0, j > 0 (j = 1, 2). Marginal probability densities read
1
fG1 (g1 ) = [hg1 1 i hg1 2 i ],
2 1
(8.88)
1
fG2 (g2 ) = [hg2 2 i hg2 2 i ]
2 1
with
(j)
Z Z ca
2
(i)
()g2
a ()
R() = FY, (c, ) = 1
fG1 G2 (g1 , g2 )dg1 dg2 (8.90)
By substituting Eqs. (8.88) and (8.89) into Eq. (8.90) and performing integration,
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
2.75
3 4 5
2.25
B C
1.75
g2 B*
A*
A D
1.25
1 2
.75
0.75 1 1.25 1.5 1.75 2 2.25 2.5
g1
Fig. 8.16 Geometrical representation of the first term in reliability expression, Eq. (8.97).
2.75
3 4 5
2.25
B*
l C* C
B
1.75
g2
A*
A D
1.25
1 2
.75
0.75 1 1.25 1.5 1.75 2 2.25 2.5
g1
Fig. 8.17 Geometrical representation of the second term in reliability expression, Eq. (8.100).
3.25
B*
2.25
3 4 5
2.75
B C* C
g2 D*
1.75
A D A*
1.25 1 2
0.75
0.75 1 1.25 1.5 1.75 2 2.25 2.5
g1
Fig. 8.18 Geometrical representation of the third term in reliability expression, Eq. (8.103).
3.5
B*
2.5
3 4 5
g2
D*
B C C*
2
1.5 A*
A D
1 2
1 1.25 1.5 1.75 2 2.25 2.5
g1
Fig. 8.19 Geometrical representation of the fourth term in reliability expression, Eq. (8.104).
the region 3 . The shaded area ABC A , multiplied by the above constant value of
the probability density function, represents the reliability of the system. This area
is obtained as a difference between the area of the triangle AB A and the area of
the triangle BB C . The coordinates of the intersection point C are
!
(j)
c a 2 2
C = (i)
, 2 (8.98)
a1
whereas the expression of the coordinates of B coincides with Eq. (8.95). The side
lengths for the triangle AB A are then given by Eq. (8.96) and the ones of the
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
triangle BB C read
(i) (j)
c a 1 1 c a 2 2
BB = (j)
2 , BC = (i)
2 (8.99)
a2 a1
The area SBB C equals
(i) (j)
1 (c a1 1 a2 2 )2
SBEC = (BB )(BC ) = (i) (j)
(8.100)
2 2a a 1 2
and is recognized as the second term in Eq. (8.91). Note that the area of the triangle
AB A coincides with the first term of Eq. (8.91).
Analogously, the geometrical meanings of the third and fourth terms involved in
Eq. (8.91) may be deduced by examining Figs. 8.18 and 8.19, respectively. In fact,
Fig. 8.18 deals with the case that the broken line belongs to region 4 . In this case,
the reliability is given by the area ABC D D, and is obtained as the algebraic sum
SBC D DA = SAB A SBEC SDD A (8.101)
The first two terms of this expression have been already identified. In order to
obtain the last term in Eq. (8.101), we determine the coordinates of point D as
!
(i)
c a 1 2
D = 2 , (j)
(8.102)
a2
Note that the coordinates of the point A are given in Eq. (8.95). The area of the
triangle DD A is so obtained as
(i) (j)
1 (c a1 2 a2 1 )2
SDD A = (DD )(DA ) = (i) (j)
(8.103)
2 2a a 1 2
which coincides with the third term in Eq. (8.91). To determine the geometric
interpretation of the fourth term in Eq. (8.91), let us consider the case when the
broken line belongs to region 5 as shown in Fig. 8.19. The reliability is given as the
product of the area of the rectangle ABCD and the density (2 1 )1 (2 1 )1 .
The area itself equals (2 1 )(2 1 ). Hence, the reliability is unity as expected.
On the other hand, in order to identify the fourth term in Eq. (8.91), we represent
the area ABCD as the following algebraic sum:
SABCD = SAB A SBB C SDD A + SCC D (8.104)
The first three terms in Eq. (8.104) are given by their respective counterparts in
Eq. (8.91). The fourth term in Eq. (8.104) may be easily obtained by inspection of
Eqs. (8.98) and (8.102) as
(i) (j)
1 (c a1 2 a2 2 )2
SCC D = CC )(CD ) = (i) (j)
(8.105)
2 2a a 1 2
As is seen, each term in the reliability expression has an appropriate geometrical
meaning. Naturally, if the broken line lies in region 1 , then one immediately
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
c = 1.2(2+2) 1=2.39
1.2 = 0.5 2=2.0
1
0.8 5 4 3
R
0.6
0.4
0.2
4 3
0 1 2 3 4 5
Fig. 8.20 Reliability versus non-dimensional time, initial imperfections with uniform probability
density (D = [1.2, 2] [1.4, 2]), Eqs. (8.80a) and (8.91).
c = 1.2(2+2) 1=1.640
1.2 = 1 2=1.327
3=1.133
1
4=0.796
0.8 5 4 3 2 1
R
0.6
0.4
0.2
4 3 2 1
0 0.5 1 1.5 2 2.5
Fig. 8.21 Reliability versus non-dimensional time, initial imperfections with uniform probability
density (D = [1.2, 2] [1.4, 2]), Eqs. (8.80b) and (8.91).
deduces that the reliability vanishes. Likewise, if the broken line passes through
region 5 , the corresponding reliability is unity.
The results of sample calculations are portrayed in Figs. 8.208.24, with the
same mechanical and geometrical parameters used in Sec. 8.2, and the second term
of imperfection in Eq. (8.66) is given by a = 5, which is an odd number. The cases
(i) (j)
correspond to the five independent expressions for the coefficients a1 and a2 in
Eq. (8.80), which correspond, respectively, to < 1, = 1, 1 < < 25, = 25,
> 25. For a specified value of the failure boundary c and the different values of the
ratio P/Pcl , the reliability functions are depicted as functions of a non-dimensional
time . For the case < 1 in Fig. 8.20, buckling occurrence is not a certain event:
The structure may or may not buckle depending upon the parameters of the system.
Hence, the reliability does not necessarily tend to zero with the increase of time .
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
c = 1.2(2+2)
1.2 = 2
1
0.8 5 4 3 2 1
R
0.6
1=0.471
0.4 2=0.743
3=0.982
0.2 4=1.203
4 3 2 1
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Fig. 8.22 Reliability versus non-dimensional time, initial imperfections with uniform probability
density (D = [1.2, 2] [1.4, 2]), Eqs. (8.80c) and (8.91).
c = 1.2(2+2) 1=0.0697
1.2 = 25 2=0.0550
3=0.0500
1 4=1.0351
0.8 5 4 3 2 1
R
0.6
0.4
0.2
4 3 2 1
0 0.02 0.04 0.06 0.08 0.1
Fig. 8.23 Reliability versus non-dimensional time, initial imperfections with uniform probability
density (D = [1.2, 2] [1.4, 2]), Eqs. (8.80d) and (8.91).
c = 1.2(2+2) 1=0.0680
1.2 = 26 2=0.0538
3=0.0489
1 4=1.0343
0.8 5 4 3 2 1
R
0.6
0.4
0.2
4 3 2 1
0 0.02 0.04 0.06 0.08 0.1
Fig. 8.24 Reliability versus non-dimensional time, initial imperfections with uniform probability
density (D = [1.2, 2] [1.4, 2]), Eqs. (8.80e) and (8.91).
10
8 Ma = 1.65
2
4
M = 3.68
6
Boundary c
4
I II III
2
Ma 2
4
M
0
0 2 4 6 8 10 12
Radius d (cm)
Fig. 8.25 Design curve c = c(d ) at t = 0.2 for uniform probability density function and unity
reliability requirement (P = 3000 kg, D = [1.2, 2] [1.4, 2]), Eq. (8.91).
6
3
Boundary c 4
4 2
0
Time t (sec)
2 1
4
Radius d (cm)
6
0
Fig. 8.26 Design surface c = c(d , t) for uniform probability density function and unity reliability
requirement (P = 3000 kg, D = [1.2, 2] [1.4, 2]), Eq. (8.91).
10
Solid Line rc = 1
Dashed Line rc = 0.9
8
Dotted Line rc = 0.8
Boundary c
4
I II III
2
0
0 2 4 6 8 10 12
Radius d (cm)
Fig. 8.27 Comparison of design curves; uniform probability density function and different codified
reliabilities at t = 0.5 (P = 3000 kg, D = [1.2, 2] [1.4, 2]), Eq. (8.91).
Figure 8.25 depicts the failure boundary c versus the radius d for the prescribed
time t = 0.5 and unity reliability. As the dependence c = c(d ) has an asymptote at
c = 4, when 2 = 2 and 2 = 2, this numerical result can be derived from analytical
considerations. In fact, when d tends to infinity,
(3)
lim aj (d , t) = 1, (j = 1, 2) (8.109)
d
(3) 1 + 4d /M
a1 (d , t) = (8.113)
1 4d /M
As may be observed from Fig. 8.27, different values of the required reliability rc
lead to appropriate asymptotic values. In fact, let us design the system for high
reliability requirement as shown in region 4 in Fig. 8.18. The expression of the
failure boundary as a function of reliability, with the aid of Eq. (8.82), is given by
q
(i) (j) (i) (j)
c = 2 a1 + 2 a2 2a1 a2 (1 rc )(2 1 )(2 1 ) (8.114)
Evaluation of the limit value of Eq. (8.115) for an infinite value of d reads
p
lim = 2 + 2 (1 rc )(2 1 )(2 1 ) (8.115)
d
Equation (8.115) allows us to obtain the asymptotic value of the failure boundary
function c(d , t) by specifying the reliability value.
The results of this section clearly demonstrate that for high values of required
reliability, the design radii d of the cross-section are extremely close to the values
obtained by non-probabilistic analysis. The following question can be asked: Do the
design values of the cross-sectional radius d depend upon the particular form of the
probability density function fG1 G2 (g1 , g2 )? One can anticipate, generally speaking,
that the designs depend on the probabilistic inputs. Yet, as it will be elucidated,
the design in the extremely high reliability region is practically independent of the
input probabilistic information.
located in the first quadrant of the plane (g1 , g2 ), where A represents a normalization
coefficient depending upon the specific expression of the probability density function
fG 1 G2
(g1 , g2 ) that extends over the entire plane.
The reliability expression with z = c substituted into Eq. (8.90), and taking into
account Eq. (8.116), read,
(j)
Z Z ca
2
(i)
g2
a1
R() = dg2 AfG 1 G2
(g1 , g2 )[hg1 1 i hg1 2 i ]
[hg2 1 i
hg2 2 i ]dg1
Z
(8.117)
= [hg2 1 i hg2 2 i ]dg2
(j)
Z ca2 g2
(i)
a1
AfG 1 G2
(g1 , g2 )[hg1 1 i hg1 2 i ]dg1
Denoting the inner integral in Eq. (8.117) by I(g2 ), Eq. (8.117) can be rewritten as
Z
R() = I(g2 )[hg2 1 i hg2 2 i ]dg2
Z 2
= I(g2 )dg2 (8.118)
1
(j)
Z 2 Z ca2 g2
(i)
a1
= dg2 AfG 1 G2
(g1 , g2 )[hg1 1 i hg1 2 i ]dg1
1
In a high reliability range as shown in region 4 in Fig. 8.18, c lies in the vicinity
of a critical value c given as
(i) (j)
c = a 1 2 + a 2 2 (8.119)
The highest reliability requirement will be obtained when the line in Eq. (8.94)
tends to pass through the point B. This corresponds to value c = c substituted in
Eq. (8.94). Thus we calculate the limit
(j)
Z 2 Z ca2 g2
(i)
a1
lim R() = lim dg2 AfG 1 G2
(g1 , g2 )[hg1 1 i hg1 2 i ]dg1
cc cc 1
(j)
Z 2 Z c a2 g2
(i)
a
= dg2 1
AfG 1 G2
(g1 , g2 )[hg1 1 i hg1 2 i ]dg1
1
Z 2 Z 2
= dg2 AfG 1 G2
(g1 , g2 )[hg1 1 i hg1 2 i ]dg1
1
(j)
Z 2 Z 2 +
a
2
(i) (2 g2 )
a
+ dg2 1
AfG 1 G2
(g1 , g2 )[hg1 1 i hg1 2 i ]dg1
1 2
(8.120)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
2.5
A2
2 A1
G0
1.5
g2 K
D
0.5
0 1 2 3 4
g1
Fig. 8.28 Domain of integration of the probability density function (g 10 = 2, g20 = 1.5, K = 1.0),
Eq. (8.122).
The inner integral vanishes, because integration is carried out in the interval
(j) (i)
[2 , 2 + (a2 /a1 )(2 g2 )]. Since 1 g2 2 , the upper limit of integration is
greater than 2 . However, in the interval beyond 2 , the integrand in the square
parenthesis vanishes. Reliability is then given only by the first term in Eq. (8.120),
yielding
Z 2 Z 2
lim R() = lim dg2 AfG 1 G2
(g1 , g2 )dg1 = 1 (8.121)
cc cc 1 1
which represents the integral of the probability density function in its entire domain.
In this case of unitary reliability, the design value of the cross-sectional radius d is
found from Eq. (8.119).
quadrant of the plane (g1 , g2 ). Also we assume that the straight line
(i) (j)
c = a1 ()g1 + a2 ()g2 (8.123)
(i)
is drawn in a generic position depending on the values of the function and a1 ()
(j)
a2 ().
According to Eq. (8.85), in order to find the reliability, it is necessary to integrate
the probability density function below the straight line in Eq. (8.122), or over the
hatched area in Fig. 8.28 as
(j)
ZZ Z Z ca2 g2
(i)
a1 1
R() = fG1 G2 (g1 , g2 )dg1 dg2 = dg1 dg2 (8.124)
D K 2
The double integral in the latter equation will be calculated with the aid of the
following polar coordinate system whose origin is at the center of the circle:
g1 = g10 + cos , g2 = g20 + sin (8.125)
where is the polar distance and is the polar angle. In this coordinate system,
the reliability becomes
ZZ
R() = 1 2
dd (8.126)
Dc K
with
1 (i) (j)
f= (c a1 g10 a2 g20 ) (8.130)
K
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
In case of R() tending to unity from below, we see from Eq. (8.127)
0 or 1 2 (8.131)
Hence, the square-root terms in Eq. (8.129) must gain
(i) (j)
(a1 )2 + (a2 )2 f 2 0+ (8.132)
Thus
q
(i) (j)
f (a1 )2 + (a2 )2 (8.133)
The geometrical meaning of the condition stated in Eq. (8.133) can be shown,
expressing the system in Eqs. (8.127a) and (8.127b) back in Cartesian coordinates.
Thus, bearing in mind Eqs. (8.125) and (8.133), we write Eq. (8.127) in the form
q
(i) (j) (i) (i) (j)
a1 g10 + a2 g20 a1 g1 + K (a1 )2 + (a2 )2
g2 = (i)
(8.134a)
a1
(g1 g10 )2 + (g2 g20 )2 = K 2 (8.134b)
Substituting Eq. (8.134a) into Eq. (8.134b), we obtain a quadratic equation for the
unknown g1
q0 (g1 )2 2g1 q1 + q2 = 0 (8.135)
with coefficients
!
(i) 2
a
q0 = 1 + 1
(j)
a2
q
(i) 2 (i) (i) 2 (j) 2
2(a ) g 10 2a 1 K (a 1 ) + (a 2 )
q1 = g10 + 1
+
(j) 2 (j) 2
(a2 ) (a2 )
! ! q
(i) 2 (i) 2 2a
(i)
(a
(i) 2
) + (a
(j) 2
)
a a 1 1 2
q2 = (g10 )2 + 1
(j)
(g10 )2 + 1
(j)
K2 + (j)
g10 K
a2 a2 (a2 )2
(8.136)
of which the discriminant q12 4q0 q1 is identically zero. Hence, the two coinciding
solutions of Eq. (8.135) read
(i)
(1) (2) a1 K
g1 = g1 = g10 + q (8.137)
(i) (j)
(a1 )2 + (a2 )2
This implies that the straight line in Eq. (8.134a) and the circle in Eq. (8.134b)
share one common point. Thus, we conclude that the straight line is tangent to the
circular domain of the initial imperfections. Eq. (8.133) can then be interpreted as
the condition for this situation.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
1 1 = 1.628
2 = 2.490
rc = 0.9
0.8
rc = 0.7
0.6
rc
0.4
0.2
1 2
0 1 2 3 4 5 6
10
Solid Line rc = 1
Boundary c (cm)
4 I II III
2
0 2 4 6 8 10 12 14
Radius d (cm)
Fig. 8.30 Comparison of design curves for uniform probability density function over a circular
domain and different required reliabilities (t = 0.5, P = 3000 kg, g10 = 2, g20 = 1.5, K = 1.0),
Eq. (8.139).
7
6 5
Boundary c
5
4
4
3 3
t (sec)
2 2
4
6 1
Radius d (cm)
8
10
Fig. 8.31 Design surface c = c(d , t), uniform probability density function over a circular domain
and required unity reliability (P = 3000 kg, g10 = 2, g20 = 1.5, K = 1.0), Eq. (8.139).
= 2 1
q
(i) (j) (i) (j)
f a1 + a2 (a1 )2 + (a2 )2 f 2
= cos 1
(i) (j)
(a1 )2 + (a2 )2 (8.138)
q
(j) (j) (i) 2 (j) 2 2
f a 1 + a 2 (a 1 ) + (a 2 ) f
cos1
(i) (j)
(a1 )2 + (a2 )2
Once is determined from Fig. 8.29, one can calculate the value from Eq. (8.138).
The solution of Eq. (8.138) for f , in conjunction with Eq. (8.130), allows one to
obtain the required expression for the design curve c = c(d ), which reads
(i) (j)
c(d ) = a1 (d )g10 + a2 (d )g20
q (8.139)
K (i) (j)
+ (a1 (d ))2 (1 + cos()) + (a2 (d ))2
2
and is portrayed in Fig. 8.30 for specified time t = 0.5 and different values of rc .
Figure 8.31 represents the design surface for unity reliability as a function c =
c(d , t), allowing one to find the cross-sectional radius d for specified exploitation
time t and the failure boundary c.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
g1 = [1 , 2 ], g2 = [1 , 2 ] (8.140)
where 1 and 2 are the lower and upper bounds of z, respectively. Since g1 and g2
are positive, so are the additional displacements. Hence, in the safety requirement
condition d z c, only the condition z c is operative. The column design via
the anti-optimization technique identifies the worst possible reachable condition for
the uppermost bound 2 of the interval variable z. If 2 c, the system will remain
in the safe domain, otherwise it will fail. Expressing the argument of the functions
(i) (j)
a1 () and a2 () in terms of the cross-sectional radius d and the pre-selected
time t by means of Eq. (8.108) (see Sec. 8.2), we obtain the formal design relation
(i) (j)
2 = 2 a1 (d , t) + 2 a2 (d , t) = c (8.142)
Comparing Eq. (8.142) with Eq. (8.108), we conclude that the former matches the
design condition in case of unity reliability requirement given in Eq. (8.108). If
the uncertainty region coincides with the domain represented in Fig. 8.15, then the
design surface c(d , t) represented by Eq. (8.142) coincides with Eq. (8.77). Thus,
the anti-optimization method and the probabilistic one with the highest reliability
requirement lead to the same design values for the cross-sectional radius d .
where (g10 , g20 ) is the center of the circle, and K is the radius as shown in Fig. 8.32.
The displacement at the center of column reads
(i) (j)
z = g1 a1 () + g2 a2 () (8.144)
where g1 and g2 are uncertain variables belonging to the set defined by Eq. (8.143),
which is convex.
The Lagrangian for the problem of maximizing z under constraint (8.143) is
given as
(i) (j)
= a1 g1 + a2 g2 + [(g1 g10 )2 + (g2 g20 )2 K 2 ] (8.145)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
3
S
2.5
A
2
G0
1.5
g2 K
d
1 B
0.5
0 1 2 3 4
g1
Fig. 8.32 Initial imperfection amplitudes modeled by convex variables: anti-optimization design
(g10 = 2, g20 = 1.5, K = 1.0), Eq. (8.152).
where is the Lagrange multiplier (see Sec. 4.1 for details of the Lagrange multiplier
approach). Necessary conditions for the extremal read
(i)
= a1 + 2(g1 g10 ) = 0 (8.146a)
g1
(j)
= a2 + 2(g2 g20 ) = 0 (8.146b)
g2
= (g1 g10 )2 + (g2 g20 )2 K 2 = 0 (8.146c)
The solutions of Eqs. (8.146a)(8.146c), denoted by g 1 and g2 , corresponding
to maximum z read
(i)
Ka1
g 1 = g10 + q (8.147a)
(i) (j)
(a1 )2 + (a2 )2
(j)
Ka2
g 2 = g20 + q (8.147b)
(i) (j)
(a1 )2 + (a2 )2
Note that Eq. (8.147a) coincides with Eq. (8.137) which represents the abscissa
of the common point between the circular domain of the random variables in
Eq. (8.122) and the straight line in Eq. (8.123).
Let us investigate the geometrical meaning of Eqs. (8.146a)(8.146c) and the
results in Eqs. (8.147a) and (8.147b). Eliminating parameter from Eqs. (8.146a)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
representing a line orthogonal to S and tangent to the circle at point A. One may
obtain the solution of the system in Eq. (8.146) by solving
(i) (j)
a1 g1 + a 2 g2 T = 0 (8.150a)
2 2 2
(g1 g20 ) + (g2 g20 ) = K (8.150b)
where the line in Eq. (8.150a) is perpendicular to the line in Eq. (8.149). The
parameter T is a free parameter that must be chosen so that the straight line in
Eq. (8.150a) constitutes a tangent to the circle in Eq. (8.150b). It may be deduced by
inspection of Eq. (8.147a) that the parameter T coincides with the failure boundary
value c.
The value of T is obtainable as follows by solving Eqs. (8.150a) and (8.150b):
q
(i) (j) (i) (j)
T = c = a1 (d , t)g10 + a2 (d , t)g20 + K (a1 (d , t))2 + (a2 (d , t))2 (8.151)
Chapter 9
In this chapter, as the main purpose of this book, hybrid approaches of optimization
with anti-optimization are presented for static and buckling problems. We first
present the framework of the problem and the methodologies for solving the two-
level optimization problem with anti-optimization. Several applications are shown
for buckling optimization, static problems considering stress, displacements and
compliance. The design methods for flexible structures are also shown, and finally,
an example of force identification of a tensegrity structure is presented.
9.1 Introduction
273
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
P2 , U2
P1 , U1
H
x
W1 W2
y
problems with loading uncertainty and a vibration isolator design with a specified
band of excitation frequency. Lanzi and Giavotto (2006) used a multiobjective
genetic algorithm for buckling optimization with uncertainty. El Damatty and
Nassef (2001) optimized a ribbed shell by using a genetic algorithm.
Structural optimization considering the worst-case scenario is closely related to
the so-called robust design. The purpose of robust design is to achieve the design
that is least sensitive to the uncertain parameters (Ben-Haim, Fr
yba and Yoshikawa,
1999; Au, Cheng, Tham and Zheng, 2003). Various methodologies have been de-
veloped based on the intuitively defined robustness function.
Structural optimization under uncertainty has been extensively studied in the
framework of reliability-based design (Palassopoulos, 1991; Kogiso, Shao, Miki and
Murotsu, 1997; Hajela and Vittal, 2006). There are many methods for reliability-
based structural optimization; see, for instance, Youn, Choi and Park (2003), Khar-
manda, Olhoff and El-Hami (2004), Cheng, Xu and Jiang (2006), and Kim and Choi
(2008). Optimization considering uncertainty can also be formulated by the fuzzy
set theory. Pantelides and Ganzerli (2001) showed that similar designs can be found
by fuzzy set and convex model (see Sec. 2.9 for fuzzy-set based optimization).
Consider again the asymmetric 2-bar truss as shown in Fig. 9.1, which has been
used in Sec. 4.1, subjected to static loads P1 and P2 , where H = 1, W1 = 3,
and W2 = 1. Let Ai and Li denote the cross-sectional area and the length of the
ith member, respectively. Youngs modulus is denoted by E. The horizontal and
vertical displacements are denoted by U1 and U2 , respectively.
We consider a simple optimization problem to find the design variable vector
A = (A1 , A2 )> for minimizing the total structural volume
V = A 1 L1 + A 2 L2 (9.1)
under displacement constraint
U2 U2U (9.2)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where U2U is the specified upper bound for U2 , which is assumed to be positive.
The relation between the load vector P = (P1 , P2 )> and the displacement vector
U = (U1 , U2 )> is written as
2
U1 = [(A1 + 2 2A2 )P1 ( 3A1 2 2A2 )P2 ]
A1 A2 E(2 + 3)
(9.3)
2
U2 = [( 3A1 2 2A2 )P1 + (3A1 + 2 2A2 )P2 ]
A1 A2 E(2 + 3)
Uncertainty is introduced for P as
P = P0 + a (9.4)
0
where P = (P10 , P20 )> is the vector of nominal values, and a = (a1 , a2 )> is the
uncertain parameter vector. Since U2 depends on A and a, it is written as U2 (a, A).
We are interested in designing the truss, i.e., selecting the values of A1 and A2 ,
so that the displacement U2 under the worst possible load scenario satisfies the
constraint (9.2).
Let ab (A) denote the worst load scenario for maximizing U2 . Since the worst load
scenario depends on the design variable vector A, the value of U2 corresponding
b is denoted as U
to a b2 (A) = U2 (b a(A), A). The lower and upper bounds for Ai are
denoted by ALi and AU i , respectively. Then the optimization problem is formulated
as
minimize V = A 1 L1 + A 2 L2 (9.5a)
b2 (A)
subject to U U2U (9.5b)
ALi Ai AU
i , (i = 1, 2) (9.5c)
Consider the case where the bound of a is given by the quadratic constraint
(ellipsoidal bound) as
a> a D (9.6)
where D is the specified upper bound. Then the lower-level anti-optimization prob-
b2 (A) is formulated as
lem for finding U
find b2 (A) = max U2 (a, A)
U (9.7a)
a
Since the constraint (9.7b) is satisfied with equality at the anti-optimal solution, we
obtain r
1 c> c
= (9.10)
2 D
Hence, by incorporating Eqs. (9.8)(9.10) into Eqs. (9.3) and (9.4), the load
vector P = P 0 + ab(A) and the displacement U b2 (A) corresponding to the worst-case
scenario are written as explicit functions of A, and the two-level optimizationanti-
optimization problem can be reformulated as a single-level optimization problem.
However, in general, such an explicit reformulation is not possible. Therefore,
many algorithms for solving the two-level problems have been developed.
Here, the feasible domain C for a is defined by the inequality constraints Ck (a) 0
(k = 1, . . . , N C ). Note that if a simple approach is used, Problem (9.14) should be
solved N I times at each step of modification of x in the upper-level problem (9.13).
Alternatively, the two-level problem can be formally written as
The readers can consult with Sec. 6 of Gurav (2005), Lombardi and Haftka (1998),
etc., for the details.
In this method, the lower-level anti-optimization problem should be solved many
times at each iterative step of design optimization, if constraints are given for many
response quantities as Hj (a, x) 0 (j = 1, . . . , N I ). Since the anti-optimization
problem for finding the worst value of Hj (a, x) can be solved simultaneously, the
use of parallel computation will be very effective for reducing the computational
cost (Gurav, Goosen and van Keulen, 2005a). Although this cycle-based approach
is simple and easy to implement, this approach does not rapidly converge if the
anti-optimal solution strongly or nonlinearly depends on x.
Lombardi and Haftka (1998) applied this algorithm to the minimum weight
design of a composite laminated structure with displacement constraints, where the
thickness and angle of each ply are chosen as design variables, and the uncertainty
is introduced for the load components, which are expressed by the multipliers ai
(i = 1, . . . , N ) for the load pattern vectors p1 , . . . , pN . The displacement vector
against pi is denoted by U i = (U1i , . . . , Uni )> , where n is the number of degrees of
freedom. Then the total displacement vector, considering the uncertainty in the
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
bL Ba bU (9.17)
where B is a constant matrix, and bL and bU are constant vectors of the lower and
upper bounds. Then the anti-optimization problem for finding the worst value of
the displacement component Uj is formulated as
N
X
find bj (x) = max
U ai Uji (9.18a)
a
i=1
subject to bL Ba b U
(9.18b)
where Uj has been assumed to be positive for simplicity. Then the cycle-based
method is applied to find the optimal distributions of the thickness and ply-angle.
It has been observed from their numerical experiments that the convergence of
the cycle-based method is very good if a is bounded with linear inequalities as in
Eq. (9.18b).
aLi ai aU
i , (i = 1, . . . , N ) (9.19)
problems (9.13) and (9.14) are combined into a single-level minimization problem
as
Although this problem formulation is simple, the total number of constraints (9.20b)
is 2N N I ; hence, the computational cost increases as an exponential function of
the number N of the uncertain parameters, if all the vertices of the feasible region
must be enumerated.
There are several studies based on this enumeration approach (Ganzerli and
Pantelides, 2000) (see Sec. 4.6 for anti-optimization of a tensegrity structure by
enumeration, and Sec. 9.6 for hybrid optimizationanti-optimization of a truss under
stress and displacement constraints).
(2006) used this approach for optimizing the truss under constraint on strain energy
(compliance) considering uncertain geometrical parameters.
9.5.1 Introduction
There has been extensive controversy over the effectiveness of optimization under
nonlinear buckling constraints. Thompson and Supple (1973) cautioned the danger
of optimization due to interaction of critical modes at buckling:
As one can see by careful reading of their comments, they did not unconditionally
reject the idea of optimization against buckling. However, the terms if and so defined
have been omitted for some reason by some researchers to state that optimization
against buckling is unconditionally meaningless. The safety against buckling of
structures can be effectively improved by optimization if uncertainties in geometry,
material, etc., are appropriately incorporated. Furthermore, optimization does not
always enhance imperfection sensitivity, and for some structures, sensitivity is even
decreased by optimization.
For plates and frames, for which the prebuckling deformation is negligibly small,
the buckling loads can be accurately estimated by linear buckling analysis. Angular-
ply laminas are extensively studied for optimization under buckling constraints with
uncertain initial imperfection (Adali, Richter and Verijenko, 1997). Laminated
plates are also optimized considering uncertainty in material properties (Adali, El-
ishakoff, Richter and Verijenko, 1994a) and external loads (Adali, Lene, Duvaut and
Chiaruttini, 2003).
The convexity of the stability region described in Sec. 5.6 can be effectively
used for optimization under the worst linear buckling load (de Faria and Hansen,
2001a,b). De Faria and de Almeida (2003b) presented the optimization method of
a plate with variable thickness. They extended their methodology to optimization
for fundamental frequency considering geometrical stiffness under the worst loading
scenario (de Faria and de Almeida, 2006).
For flexible structures such as spherical shells and latticed domes in civil and
architectural engineering, the effect of prebuckling deformation should be incor-
porated for accurate estimation of the buckling load. Ohsaki (2000, 2002c, 2005)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Ui
optimized shallow space trusses to find that imperfection sensitivity does not always
increase as a result of optimization, but is often reduced. Furthermore, optimization
of a shallow shell-type structure often leads to a coincident critical point, called the
hilltop branching point, as a coincidence of bifurcation point(s) and a limit point,
which is not always imperfection-sensitive (Ohsaki and Ikeda, 2007a). Therefore,
the critical load considering its reduction due to initial imperfection can be effec-
tively increased by optimizing the perfect system.
However, to be more comfortable with optimization against buckling, we can
directly incorporate the reduction of the critical load due to imperfection into the
process of optimization (Ramm, Bletzinger and Reitinger, 1993; Reitinger, Blet-
zinger and Ramm, 1994; Reitinger and Ramm, 1995). In this process, the worst-
case scenario should be taken into account to evaluate the lower bound of the crit-
ical load for the given magnitude of imperfection. Since optimization may lead to
imperfection-sensitive structures with multiple (coincident) buckling loads, if some
bifurcation points coincide, the imperfection sensitivity of such structures has been
extensively investigated (Ohsaki, 2002a,b; Ikeda, Ohsaki and Kanno, 2005; Ohsaki
and Ikeda, 2006). Ohsaki, Uetani and Takeuchi (1998b) presented an optimization
method for symmetric systems considering reduction of the maximum load due to
the worst imperfection. El Damatty and Nassef (2001) optimized a ribbed shell by
using a genetic algorithm. Ohsaki (2003a) defined the worst load case of flexible
structures by the strain energy stored at the deformation with the specified norm,
and optimized trusses under constraint such that the initial undeformed state is
recovered after releasing the loads corresponding to the worst-case scenario.
Perfect
Perfect
Imperfect Imperfect
Imperfect Imperfect
q q
(a) Unstable-symmetric bifurcation point (b) Stable-symmetric bifurcation point
Fig. 9.3 Classification of bifurcation points and associated equilibrium paths of perfect and im-
perfect systems; : bifurcation point.
c
b
(q, x) Optimization
b
(q, x)
c
q q
seen, for the unstable-symmetric bifurcation, the critical point of an imperfect sys-
tem turns out to be a limit point, which may be far below the bifurcation load of
the perfect system. However, for the stable-symmetric bifurcation point, the critical
point disappears due to the presence of imperfection.
For structures exhibiting unstable-symmetric bifurcation point, the critical load
reduced due to the worst initial imperfection with specified bound may be used
instead of c in Eq. (9.22b). However, the maximum load cannot be defined by a
critical load for the case of stable-symmetric bifurcation. In this case, the maximum
load may be defined by the upper-bound constraints on the responses such as stresses
and displacements. For example, consider a truss and let = (1 , . . . , N m )>
denote the vector of member stresses, where N m is the number of members. The
maximum load can be defined with the upper and lower bounds denoted by iU and
iL , respectively, of i as
m = max{|iL i () iU , i = 1, . . . , N m } (9.23)
Hence, the optimization problem may be defined as follows under constraint on the
maximum load m (x):
minimize V (x) (9.24a)
m
subject to m (x) (9.24b)
xLi xi xU
i , (i = 1, . . . , m) (9.24c)
m
where is the specified lower bound for the maximum load (see Sec. 5.5 for anti-
optimization of the linear buckling load of a braced frame using this definition of
the maximum load).
The maximum load of a structure exhibiting stable bifurcation can be alterna-
tively defined by the displacement constraint (Ohsaki, 2002b). Constraints on the
antisymmetric components of displacements can also be given for the case of un-
stable bifurcation point, because the antisymmetric deformation rapidly increases
with decreasing load after reaching the bifurcation load (Pietrzak, 1996).
The optimization problem can be alternatively formulated under the condition
such that the bifurcation point is stable; i.e., the structure is stabilized by opti-
mization so that the perfect system has the bifurcation path (initial postbuckling
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
path) with increasing load factor as illustrated in Fig. 9.4. Design sensitivity anal-
ysis of the responses on the postbuckling path is presented by Godoy and Taroco
(2000). Schranz, Krenn and Mang (2006a) investigated the characteristics of the
bifurcation path based on Koiters perturbation approach (Koiter, 1945), and pre-
sented a numerical approach in Schranz, Krenn and Mang (2006b) to convert an
imperfection-sensitive structure into an imperfection-insensitive one by placing ad-
ditional supports. Bochenek (2003) investigated the effect of the stiffness of spring
support on the shape of the postbuckling path in the loaddisplacement space.
Let b (q, x) denote the load factor along the bifurcation path of the perfect
system, which is a function of the generalized antisymmetric component q of the
displacement. Note that q = 0 corresponds to the bifurcation point. The curvature
b (x) of the bifurcation path at the bifurcation point is defined by
b 2 b (q, x)
(x) = (9.25)
q 2 q=0
are the specified values of generalized displacement in the direction of the bifurcation
mode. By enforcing the constraint (9.27c), the structure with a non-decreasing
bifurcation path is obtained, and the limit point (snapthrough) in the bifurcation
path is also avoided. Bochenek (2003) formulated similar problems considering
multiple buckling loads.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(5) (6) H
(12)
5 6
(3) (4) H
(9) (10)
(11)
3 4
(1) (2) H
(7) (8)
y
1 2
x
W W
4000
3000
Load factor
2000
1000
0
0 100 200 300 400
Horizontal displacement of node 7
Fig. 9.6 Relation between and U for = 0.01; solid line: = 1.0; dashed line: = 0.01.
for the truss with uniform cross-sectional area; i.e., = 1.0, and is 3842.1 for
= 0.01. Therefore, the truss with a given total structural volume has a larger
buckling load if the diagonal members have smaller cross-sectional areas than the
vertical and horizontal members.
An imperfection is given in the x-direction of the top node 7. Consider first the
case where = 0.01. The relation between P and the x-directional displacement
U of node 7 for = 1.0 is shown in the solid line in Fig. 9.6. As is seen, the
horizontal displacement suddenly increases as P approaches c (= 3678.0). Note
that P is a monotonically increasing function of U . For = 0.01, the equilibrium
path is plotted in a dashed line in Fig. 9.6. The load factor for = 0.01 is larger
than that for = 1.0 in the range of small U . However, gradually decreases
as U is increased after approaching the bifurcation load of the perfect system. In
fact, the value of for = 0.01 is smaller than that for = 1.0 for the range
U 160. Therefore, a stable bifurcation with smaller bifurcation load is preferred
to an unstable bifurcation with larger bifurcation load if the allowable displacement
is moderately large.
Figure 9.7 shows the equilibrium paths of imperfect systems for = 1.0, where
the solid and dashed lines correspond to the designs with = 1.0 and 0.01, respec-
tively. As is seen, the equilibrium path has a limit point for = 0.01, and the load
factor monotonically increases for = 1.0. The load factor for = 0.01 is smaller
than that for = 1.0 in the range U 185.
Optimal design is next found for = 1.0, where is chosen as the design
parameter. The allowable displacement is given as U U = 200; i.e., the maximum
load m is defined as the limit point load in the range U < U U or the load factor
m
at which U = U U is satisfied, and the specified maximum load factor is 4000.0.
The relation between the design parameter and the total structural volume for
m
satisfying = 4000.0 is plotted in Fig. 9.8. The optimal design that minimizes
the total structural volume exists at ' 0.027, and the corresponding optimal
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
4000
3000
Load factor
2000
1000
0
0 100 200 300 400
Fig. 9.7 Relation between and U for = 1.0; solid line: = 1.0; dashed line: = 0.01.
1400000
Total structural volume
1200000
1000000
800000
0 0.02 0.04 0.06 0.08 0.1
Design parameter
Fig. 9.8 Relation between design parameter and total structural volume.
objective value is about 9.0467 105 . This optimal solution with a small value of
agrees with the fact that a tower-type structure can be effectively stabilized by
adding very slender braces. Figure 9.9 shows the relation between and U for
designs satisfying m = 4000.0. As is seen, the maximum load is achieved at a
limit point for = 0.01 and at U = U U = 200 for = 1.0. Consequently, for the
optimal design = 0.027, has the maximum value at a limit point that satisfies
U = U U = 200.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
4500
4000
Load factor
3500
3000
0 100 200 300 400
Fig. 9.9 Relation between and U for designs satisfying m = 4000.0 ( = 1.0); solid line:
= 0.027; dashed line: = 0.1; dotted line: = 1.0.
1 (1) 2 (2) 3
2.0 m
(5) (6)
(8) (7) (9)
(10)
(3) (4) 6
y 4 5 P2
P1 P3
x
2.0 m 2.0 m
1 5.3536 6.0910
2 1.0000 1.0000
3 5.3536 5.3502
4 16.0000 15.3359
5 3.6464 4.9652
6 1.0000 1.0000
7 13.6421 13.6469
8 1.0000 1.0000
9 2.8284 7.5264
10 1.4142 1.0000
Since the stresses and displacements are linear functions of the loads, the maximum
and minimum values can be found by enumerating the eight sets of loads P V k (k =
1, . . . , 8) at the vertices of the feasible region defined by
Therefore, the optimization problem with the worst load scenario is formulated as
Xm
minimize V (A) = A i Li (9.31a)
i=1
subject to iL i (A, P V U
k ) i , (i = 1, . . . , m; k = 1, . . . , 8) (9.31b)
U6 (A, P V U
k ) U , (k = 1, . . . , 8) (9.31c)
ALi Ai , (i = 1, . . . , m) (9.31d)
where Li is the length of the ith member, and m is the number of members that is
equal to 10. In this formulation, we have 168 inequality constraints in addition to
the side constraints (9.31d) for Ai .
Alternatively, the maximum and minimum stresses and displacements for the
feasible set of loads defined by Eq. (9.30) can be found by combining the responses
against each load P1 , P2 , and P3 applied independently. Let ij denote the stress of
member i against the load Pj0 . From the definition (9.29) of load uncertainty, the
maximum imax and minimum imin of i are found as
X3
imax = (1 + 0.1 sign ij )ij ,
j=1
3
(9.32)
X
imin = (1 0.1 sign ij )ij
j=1
For example, if ij is negative, the maximum and minimum values of the stress of
the ith member to the jth load are (1 0.1)ij = 0.9ij and (1 + 0.1)ij = 1.1ij ,
respectively. Then the constraints (9.31b) can be written as
iL imin , imax iU , (i = 1, . . . , m) (9.33)
This way, the number of constraints is reduced to 21, including the displacement
constraint that can be defined in a similar manner as Eqs. (9.32) and (9.33).
Table 9.2 shows the optimal solutions without displacement constraint (Case 1)
and with displacement constraint (Case 2). It is seen by comparing the results in
Tables 9.1 and 9.2 that the optimal objective value significantly increases, especially
for Case 2, as a result of load uncertainty.
Table 9.3 shows the values of stresses at the eight vertices of the feasible region
of the optimal solution for Case 1, where U and L denote that the load takes
upper- and lower-bound values, respectively. For example, (U,L,U) corresponds to
the vertex defined by (P1 , P2 , P3 ) = (P1U , P2L , P3U ). The underline indicates that
the stress is equal to the lower or upper bound; i.e., the member is fully stressed.
As is seen, each member is fully stressed for at least one loading condition, except
members 8 and 9. Note that the cross-sectional area of member 8 is equal to the
lower bound and the stress constraint is satisfied with inequality. Since member 9
has a larger range of allowable stress, the stress is not equal to either the upper or
lower bound; i.e., the optimal solution under multiple loading conditions is not fully
stressed when the load uncertainty is considered and the bounds of the stresses are
not given uniformly.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
1 6.0932 14.3939
2 1.0000 1.0000
3 8.7550 10.4616
4 18.2481 17.6119
5 4.1565 4.2834
6 1.0000 1.0000
7 14.8966 11.9314
8 1.0000 7.0584
9 3.3680 12.7440
10 1.4142 1.0000
Member number (U,U,U) (U,U,L) (U,L,U) (U,L,L) (L,U,U) (L,U,L) (L,L,U) (L,L,L)
9.7.1 Introduction
It is well known that the interval analysis for anti-optimization of structures under
bounded uncertainty leads to a too conservative result if the dependency among the
uncertain variables are not appropriately incorporated; see Sec. 3.3 for details. One
strategy to avoid this unrealistic situation is to reduce the number of independent
uncertain parameters using Fourier transformation or modal decomposition where
the bounds are to be given for the coefficients of the modes, or simply by linking the
variables (Elishakoff, 1999). However, in these approaches, explicit bounds cannot
be assigned for each variable. Therefore, bounds should be given as additional
linear constraints, because the coefficient itself has no physical meaning related to
uncertainty in external loads, material parameters, geometrical properties, and so
on.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where i is the ith eigenmode of undamped free vibration, ci (t) is the modal re-
sponse that is a function of time, and n is the number of degrees of freedom.
The equivalent static load P i corresponding to i is defined as
P i = i M i (9.35)
where M is the mass matrix, and i is the ith eigenvalue of undamped free vibra-
tion.
Suppose a single mode, say mode i, dominates in the response. For example, for
a building frame, the response can be evaluated with good accuracy by using the
first mode only. In this case, the maximum response U max
i corresponding to mode
i is computed from
KU max
i = cmax
i Pi (9.36)
where K is the stiffness matrix, and cmax
i is the maximum modal response obtained
from the displacement response spectrum, which is multiplied by the participation
factor of mode i (see Sec. 6.6 for examples of response spectra).
Let R denote the representative response, such as stress or strain of the specified
member or element, which is a linear function of U , as
R = b> U (9.37)
where b is a constant vector. The maximum absolute value of the response corre-
sponding to mode i against the equivalent static load cmax
i P i for seismic design,
called design load for simplicity, is obtained as
Rimax = |b> U max
i | (9.38)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Even when a few modes dominate in the response of a structure, only those
dominant modes can be used for constructing the design loads. Suppose we use the
three modes 1, 2 and 3, for simplicity, and assume the response R is positive for
the maximum responses of all the modes. The worst possible static load pattern is
then given as
P worst = cmax
1 P 1 + cmax
2 P 2 + cmax
3 P3 (9.39)
max
where Ui,j is the jth component of U max
i . However, we can assign more explicit
constraints on cmax
i in Eq. (9.39) to give a set of design loads that consider uncer-
tainty in the combination of the coefficients; e.g.,
cL cmax
1 + cmax
2 + cmax
3 cU (9.41)
where cL and cU are the specified lower and upper bounds, respectively. More rig-
orous constraints can be incorporated based on the statistical data of maximum
responses of structures to a set of recorded seismic motions. Zhang, Ohsaki and
Uchida (2008) presented an approach to generating several design loads as combi-
nations of multiple modes.
The design load vector is alternatively defined by using only the most dominant
mode, say 1 , and adding the uncertain load deviation vector P dev for representing
the contribution of the higher modes:
P = cmax
1 P 1 + P dev ,
dev
(9.42)
Pj,L Pjdev Pj,U
dev
, (j = 1, . . . , n)
dev dev
where Pj,L and Pj,U are the specified lower and upper bounds for Pjdev , respectively.
However, also for this case, it is unrealistic that all Pjdev will take their maximum
or minimum at the same time instance during vibration. Therefore, we need to
incorporate additional constraints on Pjdev .
In the following, we present a general procedure for anti-optimization of static
responses with additional linear constraints. As an example, anti-optimization prob-
lem of a truss is formulated for maximizing the compliance (external work) under
uncertain static nodal loads. The optimum design problem under compliance con-
straint is solved by a simple heuristic approach called the greedy method.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
The truss is subjected to static nodal loads P , for which the nodal displacement
vector U is found from
U = CP (9.44)
where C is the flexibility matrix; i.e., C = K 1 with K being the stiffness matrix.
A constraint is given for the compliance W , defined by
subject to b> s
i P di , (i = 1, . . . , N ) (9.46b)
PiL Pi PiU , (i = 1, . . . , n) (9.46c)
where W U is the specified upper bound for W , and ALi and AU i are the lower and
upper bounds for Ai .
Since the constraint function of (9.47b) is not generally continuously differen-
tiable with respect to A, a simple heuristic approach called the greedy method is
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
used for design optimization (see Sec. 2.6 for details of heuristic approaches). The
design variable Ai is discretized as
Ai {jA | j = 1, . . . , N D } (9.48)
D
where A is the specified unit value of the cross-sectional area, and N is the
number of possible values of each cross-sectional area. The algorithm is summarized
as follows:
(0)
Step 1 Assign initial value Aj for Aj (j = 1, . . . , m), and set the iteration counter
k = 0.
Step 2 Find the anti-optimal solution P (k) and the corresponding worst value
c (A(k) ) of W , for A fixed at A(k) , by enumerating the values of W at all the
W
vertices of the feasible region defined by Eqs. (9.46b) and (9.46c).
Step 3 If Wc > W U , increase Aj (j {1, . . . , m}) by A for the member which
satisfies Aj < N D A and maximizes the ratio of the decrease of W c to the
c U
increase ALj of the structural volume. If W W , decrease Aj by A for
a member which satisfies Aj > A and minimizes the ratio of the increase of
c to the decrease of the total structural volume.
W
Step 4 Set k k + 1 and go to Step 3, if the termination condition is not satisfied.
Step 5 Output the best solution satisfying the constraints, and terminate the pro-
cess.
subject to D P1 P3 D (9.50b)
PiL Pi PiU , (i = 1, 2, 3) (9.50c)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
5 (6) 6
(5) (7)
y (11) (12)
4 (10) (13) 7
2
(9) (2) (3) (14)
1 3
(8) P2
(15)
(1) P1 P3 (4)
8 9
x
Node number x y
1 2000.0 1500.0
2 4000.0 2000.0
3 6000.0 1500.0
4 1000.0 1750.0
5 3000.0 2750.0
6 5000.0 2750.0
7 7000.0 1750.0
8 0.0 0.0
9 8000.0 0.0
(P1 , P2 , P3 ) D=6 D = 12 D = 18
140
130
Compliance
120
110
100
P1*
90
-6 -4 -2 0 2 4 6 8 10
P1
P4: From the symmetry of the truss, W (P , A) does not change by exchanging the
values of P1 and P3 .
From the properties P1 and P2, we can fix P2 at 10.0, because Eq. (9.50b)
does not include P2 . From P4, we can assume that P1 P3 . Then utilizing
the convexity of W with respect to P , the anti-optimal solutions are found by
enumerating the vertices (P1 , P2 , P3 ) = (10, 10, 10), (10D, 10, 10), (10, 10, 10+
D) and (10, 10, 10). Thus the number of vertices for which the responses are to
be evaluated is reduced from 12 to 4. The values of W at the vertices corresponding
to various values of D are listed in Table 9.5, along with the worst value among the
four vertices.
Variation of the compliance W is plotted in Fig. 9.12, for example, with respect
to P1 for P2 = P3 = 10.0. The value of W for P1 = 10 is 123.67, and W = 123.67 is
satisfied at P1 = P1 ' 5.833 as indicated in Fig. 9.12. Since P3 = 10.0, the lower
bound of P1 is 10.0 D from Eq. (9.50b). Hence, the anti-optimal value (worst load
scenario) of P1 corresponding to the maximum value of W is 10.0 for D < 15.833,
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
where A = 10. The worst load scenario is searched for at the vertices
(P1 , P2 , P3 ) = (10, 10, 10), (0, 10, 10), (10, 10, 0), and (10, 10, 10) for 20 3 = 8000
designs to find the exact optimal solution (A1 , A2 , A3 ) = (170, 90, 60), where
V = 2.8411 106 and W c = 98.391. The optimal solution is illustrated in Fig. 9.13,
where the width of each member is proportional to its cross-sectional area. Note
that the lower chords have large cross-sectional areas in the optimal solution. The
worst load for this optimal design is P1 = P2 = P3 = 10. Note that the number of
solutions to be enumerated can be reduced by using the current minimum value V
of the solution satisfying the compliance constraint. For example, in the process of
increasing A1 while fixing A2 and A3 , the process is terminated if V exceeds V ,
and A2 is next to be increased.
The optimal solution has also been found by the greedy method presented in
Sec. 9.7.2. The process is terminated in 30 iterations. Among the ten solutions
found from ten randomly generated initial solutions, seven solutions coincided with
the optimal solution (A1 , A2 , A3 ) = (170, 90, 60) found by enumeration. Therefore,
for this example, optimal solution can be easily found by the greedy method, for
which the total number of design evaluation is 30 10 = 300, that is far smaller
than the 8000 for the enumeration approach.
9.8.1 Introduction
A great deal of attention is paid to structure/control-integrated optimum design
for advanced structures, such as space structures, adaptive structures and smart
structures, which are required to fulfil complicated missions at high quality with
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Advanced structures
Structure
High quality
of performance
Mechanism Control
the aid of active control systems. The concept of advanced structures is illustrated
in Fig. 9.14 (Kida and Komatsu, 1993; Thomas, Sepulveda and Schmit, 1992).
The precise control on geometrical properties of these structures is stringently
demanded for high quality performance, and a sophisticated structural design
methodology considering the interaction with an active control system and mech-
anism should be devised. Homology design can be a candidate for such structural
optimization of advanced structures, utilizing the concept of homologous defor-
mation, in which a prescribed geometrical property at a part of the structure is
maintained before, during, and after the deformation. The resolution of a huge and
precise radio telescope was greatly enhanced by realizing homologous deformation,
which kept, the shape of the reflector structure parabolic in spite of the deformation
caused by its own weight (Morimoto, Kaifu, Takizawa, Aoki and Sakakibara, 1982;
von Hoerner, 1967). The control cost of the active system to adjust the shape of the
parabolic reflector is greatly reduced and the resolution quality is ensured easily.
Two formulations based on the finite element method have been derived for ho-
mology design in static problems (Hangai, 1990; Yoshikawa and Nakagiri, 1990).
However, the formulations were given in a purely deterministic manner, where the
external loadings must be specified. The advanced structures are to be made very
flexible owing to the lightweight requirement. Thus, the homologous deformation
under specified loading appears to be easily disturbed by the uncertain fluctua-
tion of loading. The sensitivity of homologous deformation with respect to such
uncertainties should be estimated prior to the application of homology design.
In this section, we closely follow the results of Yoshikawa, Elishakoff and Nak-
agiri (1998a). The formulation of homology design with an illustrative numerical
example of an 11-bar truss is given, using finite element sensitivity analysis, prior
to the discussion on uncertain loading. The uppermost deck subjected to uniformly
distributed vertical load as nominal loading is kept both straight and horizontal,
before and after the deformation by homology design. The methodology for the es-
timation of the disturbed homologous deformation is developed based on the convex
model of uncertain loadings (Ben-Haim and Elishakoff, 1990; Elishakoff, Lin and
Zhu, 1994e), in which uncertain variables are assigned to discretized nodal forces
and the existence domain of the uncertain variables is confined within the convex
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
hull of hyperellipse. The worst case of homology design is estimated on the bound-
ary of the convex hull as the point that maximizes the error index. The latter is
defined as the square of the Euclidean norm of displacements error from objective
homologous deformation. The validity of the proposed method for the worst-case
estimation of the homology design is demonstrated in a numerical example of the
11-bar truss after homology design.
KU = P (9.52)
H(U ) = 0 (9.53)
which consists of J equations with respect to the nodal displacements. The way of
representation is not unique even for the same homologous deformation. The success
of the objective homology design greatly depends on the manner of determining an
adequate trial design and neat representation of the homologous constraint. For
the sake of simplicity of discussion, we treat only linear homologous deformation,
the constraint for which is given by linear equations as follows with respect to the
nodal displacements:
CU = d (9.54)
pi = pi (1 + i ), (i = 1, . . . , m) (9.55)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
The effect of the design change is linearly approximated and the change of the nodal
displacement vector is expressed in the following linear form:
m
X
U =U+ U
i i (9.56)
i=1
where U is the displacement vector of the current design, and U i is the first-order
displacement sensitivity vector with respect to i , obtained by finite element sensi-
tivity analysis (see Sec. 3.6.2 for details of static sensitivity analysis). Substituting
the linearly approximated nodal displacement vector of Eq. (9.56) into the homolo-
gous constraint in Eq. (9.54), we obtain the governing equation of design variables
in the form
Xm
CU i i = d CU (9.57)
i=1
1 2 3 4
6 1500 mm
5 3000 mm 7
0.1 mm
The objective homologous deformation is not realized strictly by the above men-
tioned design modification, since the governing equation (9.58) is derived based on
the linear approximation of nodal displacements change. The deficient approxima-
tion can be compensated by the iterative update of the design variables, in which
the updated structure according to the solution of Eq. (9.58) is used as the trial
design for the next step. The iterative process is stopped when the homologous
constraint (9.54) is regarded to be satisfied, that is, the Euclidean norm of CU d
becomes sufficiently small.
Fig. 9.17 Realized homologous deformation and the final cross-sectional areas.
where vi is the vertical nodal displacement of node i, and vm is the average of vertical
nodal displacements on the uppermost deck, that is, vm = (v1 + v2 + v3 + v4 )/4.
The realized homologous deformation by changing the cross-sectional areas of all
the members is illustrated in Fig. 9.17. The vertical displacement of the uppermost
deck turns out to be 0.0516 mm. The manner of illustration of the deformation is
the same as that of Fig. 9.16. In Fig. 9.17, the cross-sectional area of the member
after homology design is indicated by the number attached to each member. The
increase in total weight is 58.7% in this example. Three update steps are needed
for the design variables to ensure a sufficiently small Euclidean norm of the vector
on the left-hand-side of Eq. (9.61); i.e., less than 0.001 mm in this case. One of
the transparent homology designs for this example is constructed so as to make the
cross-sectional area of vertical and outer horizontal members extremely large. This
method inevitably results in an extreme increase of the total weight.
where 0 is the Lagrange multiplier (see Sec. 4.1 for details of the Lagrange
multiplier approach). The uncertain variables in the worst case are so determined
as to satisfy the stationary condition of derived as
= 2(D q 2 W )a = 0 (9.69a)
a
= 1 q 2 a> W a = 0 (9.69b)
The stationary condition (9.69a) results in the eigenvalue problem formed by the
eigenvalue q 2 , which is rewritten as = q 2 and the eigenvector a. On the other
hand, the stationary condition (9.69b) gives the normalization condition for the
eigenvector. Generally, n eigenpairs are obtained by solving the eigenvalue problem
(9.69a) under the normalization condition (9.69b). In view of Eqs. (9.69a) and
(9.69b), the error index of Eq. (9.67) becomes
2 = a> Da = q 2 a> W a = = (9.70)
q2
as the magnification coefficient q is constant and the maximum value of the error
index, i.e., the worst case, is estimated by the maximum eigenvalue obtained
by solving Eq. (9.69a). The eigenvector corresponding to the maximum eigenvalue
gives the uncertain variable vector for the worst case.
a1 a2 a3 a4
0.1 mm
0.1 mm
displacements are linear functions of the nodal loads. The deformations of the 11-
bar truss in the two worst cases are illustrated in Figs. 9.18 and 9.19, respectively,
in a similar manner as Fig. 9.16.
hull by the Lagrange multiplier method. The numerical example using the 11-bar
truss structure shows the validity of the method by demonstrating the identification
of the worst case caused by uncertain variables confined within a hypersphere.
9.9.1 Introduction
The approaches to designing structures against external loads are classified into two
categories: stiff structure and flexible structure. In the stiff structure approach, the
structure is designed so as to have enough stiffness and strength against the static
and dynamic loads, and the deformation under such loads should be small enough.
In some situations, however, the structure may be flexible and large deformation
may be allowed so far as no failure or yielding occurs and the undeformed initial
state can be recovered after releasing the external loads.
Contrary to mechanical parts and structures, the loading condition for struc-
tures in the field of civil engineering is highly unpredictable. Therefore, it is not
practically acceptable to design such structures by considering only static loads and
proportional loading conditions with fixed load patterns. The magnitude of the
input energy by dynamic motions such as earthquake excitations and wind loads
can be characterized and bounded by the total energy applied to the structure.
Hence, the strain energy that can be stored may be used as the performance mea-
sure of elastic structures in the civil engineering field. This way, the stability of the
structure under dynamic seismic or wind loads with specified energy is assured (see
Sec. 6.6 for anti-optimization with energy bounds of seismic excitation). Note that
the energy dissipation capacity may be an important performance measure also for
the case where inelastic responses are allowed (Ohsaki, Kinoshita and Pan, 2007;
Pan, Ohsaki and Tagawa, 2007).
This section summarizes the optimizationanti-optimization approach developed
by Ohsaki (2003a) to designing elastic flexible structures. The structure is designed
allowing large deformation under the external loads with specified energy. A con-
straint is given for the storable strain energy under the specified norm of deforma-
tion. Furthermore, the structure should recover the undeformed state after releasing
the external loads. The criterion for asymptotic stability is used for formulating the
optimization problem. An illustrative example is shown using a 2-bar truss, and the
performances of the optimum design is verified by carrying out dynamic analysis of
a 24-bar dome-type truss.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
H(U)
Ui
Fig. 9.20 A quasi-convex function H(U ) that has a stationary point at the origin U = 0; U i : a
representative displacement component.
The specified value for D(U ) is denoted by D; i.e., the feasible domain for
U is given by D(U ) D. A tilde is used as H(Ue , A) to indicate that U is the
variable vector, and A is regarded as a parameter vector, whereas a hat is used
to indicate a function of A only. The worst (minimum) value H(A) b e , A)
of H(U
among the displacements satisfying D(U ) = D should not be less than the specified
lower bound H.
b
Hence, H(A) is defined as the anti-optimal objective values of the following
problem:
find b
H(A) e , A)
= min H(U (9.76a)
U
The undeformed initial state should be recovered by releasing the external loads
at any state with U in . Therefore, the condition (9.74) should be satisfied. Since
a strict inequality condition cannot be given for an optimization problem, Eq. (9.73)
is replaced by
e , A) 0
G(U (9.77)
Moreover, it is difficult to satisfy the equality constraint (9.76b) during the optimiza-
tion process. Therefore, the equality constraint is relaxed to inequality D(U ) D.
However, if the minimum value of G(U e , A) is simply searched for in a region
bounded by D(U ) D, an obvious and meaningless solution U = 0 will be found.
Therefore, the objective function G(U e , A) of the anti-optimization problem is re-
e
placed by G(U , A)/D(U ) to prevent convergence to U = 0.
The minimum value of G(U e , A)/D(U ) at the anti-optimal solution is denoted
b
by G(A), which is a function of the design variable vector A. The anti-optimization
problem is formulated as
e , A)
G(U
find b
G(A) = min (9.78a)
U D(U )
subject to D(U ) D (9.78b)
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
U1
h
U2
w w
b
If G(A) 0, then D(U ) > 0 leads to G(U e , A) 0 throughout the region
except U = 0, and H(U ) is a quasi-convex function that has the maximum value
at D(U ) = D. Since the minimum value of H(U ) along the boundary D(U ) = D is
b
H(A), which is obtained by solving Problem (9.76), H(U ) is a quasi-convex function
in the region including the origin U = 0 bounded by H(U ) H(A). b Therefore,
the undeformed state is recovered from any state U satisfying H(U ) H(A).b
Finally, the upper-level optimum design problem is formulated as
where H is the specified lower bound of the strain energy. Suppose that the objective
function V (A) is an increasing function of A, which is a natural assumption for the
b
case, e.g., V (A) is the total structural volume. If the strain energy H(A) is also an
increasing function of A, e.g., for the case where A is the vector of cross-sectional
area of a truss, the constraint (9.79b) is satisfied in equality at the optimal solution.
In the following numerical examples, optimization is carried out by using a
standard gradient-based nonlinear programming with line search, where the sensi-
tivity coefficients are found by using the finite difference approach. The lower-level
anti-optimization problems (9.76) and (9.78) are solved at each step of solving the
upper-level optimum design problem (9.79).
13
(23) (24)
9 (20) 12
10 (21) 11 (22)
1
z
x
Node Number x y z
the member numbers and node numbers, respectively. The specified strain energy
is H = 1.0 106 . The members are divided into three groups 1, 2 and 3, each of
which has the members {8, 9, 12, 13, 16, 17}, {4, 7, 10, 15, 18, 21}, {1, 2, 3, 5, 6, 11,
14, 19, 20, 22, 23, 24}, respectively, where the cross-sectional areas of the members
in group i have the same value Agi . Therefore, the number of design variables is
three.
Optimal solutions are found for various values of D. The optimization results
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Fig. 9.23 Displacement U I that has the minimum value of strain energy for specified displacement
norm D = 180.0; solid line: deformed shape; dotted line: undeformed shape.
are listed in Table 9.9. For D = 200.0, a solution satisfying G(A) b 0 could not
be found; i.e., the undeformed state cannot be recovered from the worst state cor-
responding to D = 200.0. The constraints on strain energy and the displacement
norm are satisfied in equality for three cases with D = 50.0, 100.0 and 180.0. G(A) b
has the minimum value on the boundary D(U ) = D for D = 50 and 180. For
b
D = 100, however, G(A) has the minimum at the point inside the region. The total
structural volume decreases as D is increased, which agrees with the intuitive obser-
vation that the amount of material may be reduced by allowing large deformation.
The cross-sectional areas of the members in group 1 that connect to the center node
have relatively large values if D is small. On the contrary, the cross-sectional areas
of the members in group 3 that connect to the supports have large values if D is
large.
The displacement U I at the anti-optimal solution of Problem (9.76) for D =
180.0 is shown in solid lines in Fig. 9.23 in real scale, where the dotted lines indicate
the undeformed shape. Note that U II at the anti-optimal solution of Problem (9.78)
has a similar shape as U I . It is observed from Fig. 9.23 that the vertical displace-
ment of the center node is very large with deformation of the members in group 1
only in this worst-case deformation.
Asymptotic stability of the undeformed state may be verified by one of the
following two approaches:
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
12000000
1000000
Strain energy
800000
600000
400000
200000
0
0 1 2 3 4
Time
Fig. 9.24 Time histories of strain energy from various initial conditions.
120
100
Norm of displacement
80
60
40
20
0
0 1 2 3 4
Time
Fig. 9.25 Time histories of norm of displacement vector from various initial conditions.
(i) Randomly assign an initial static equilibrium state with specified strain energy,
and carry out transient response analysis.
(ii) Assign initial velocities at the nodes so that the initial kinetic energy is equivalent
to the specified strain energy, and carry out transient response analysis.
Assuming that the target application of the proposed method is seismic design
of a latticed dome as shown in Fig. 9.25, the second approach is used for verification
purposes; i.e., the specified strain energy represents the maximum energy supplied
by the seismic motion.
Dynamic analysis for specified initial velocity has been carried out by using
MSC.Nastran Ver. 70.5 (Nastran, 2001). The solution type of Nastran is SOL129,
which is a transient response analysis with geometrical nonlinearity. The time
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
increment for integration with quasi-Newton iteration is 0.02 sec. Since the dome
structures in civil engineering usually have large masses attached to the nodes, the
member mass is neglected in the following analysis.
Initial velocities are given randomly only in z-direction of the internal nodes. A
random number ri is generated for each node, and let
N
X
ri
ri = sum
, rsum = rj (9.80)
r j=1
where N is the number of internal nodes, which is equal to 7 for this example.
Let mi denote the mass at the ith internal node. Then the initial velocity vi is
calculated from
s
XN
2H
vi = r i , M= (mj rj ) (9.81)
M j=1
9.10.1 Introduction
Assessment and maintenance of member forces are of great importance for pre-
stressed structures such as cable nets and tensegrity structures, which take advan-
tage of prestress to enhance their stiffness. Therefore, the distribution of forces
should be precisely identified and carefully adjusted in the construction process.
The member forces should also be monitored and maintained after construction
for the process of life-cycle management (see Sec. 4.6 for anti-optimization in force
design of tensegrity structures).
In this section, we present the methodology for the force identification of pre-
stressed pin-jointed structures based on measurements of the member forces and
nodal locations (Zhang, Ohsaki and Araki, 2004). An optimization problem is for-
mulated for determination of the optimal placement of the measurement devices
for the member forces, and a heuristic approach is presented for this combinatorial
optimization problem.
Because of the high ratio of stiffness to structural weight, prestressed structures
are considered to be the ideal structural form for lightweight structures, and they
attract much attention in theoretical studies as well as practical applications. We
mainly consider the prestressed structures, of which members are pin-jointed. Many
novel structural forms, such as cable nets, tensegrity structures, and tensile mem-
brane structures modeled by cable nets, fall into this category (Kanno and Ohsaki,
2005; Ohsaki and Zhang, 2006).
Since the prestressed pin-jointed structures are usually stabilized by the pre-
stresses in the state of self-equilibrium, their configurations and member forces are
highly interdependent (Kanno, Ohsaki and Ito, 2002; Vassart and Motro, 1999;
Zhang and Ohsaki, 2006). Furthermore, the distribution of member forces is of
great importance to the stiffness of the structures, because it has significant influ-
ence on the geometrical stiffness. In the design problem of this kind of structure,
member forces are usually properly assigned to satisfy the self-equilibrium equa-
tions, and more importantly, to have positive effect on its stiffness so as to have
higher capability of resisting external loads.
However, it is very difficult to achieve the member forces as expected in practical
applications, mainly due to (a) fabrication errors, (b) installation errors, (c) short
and long term creep of materials, and (d) some other unexpected environmental
influences such as temperature change, deformation of base and supporting struc-
tures, etc. Their capability to resist external loads may be significantly reduced or
even lost due to unexpected change of member forces. Moreover, some structures
make use of the interdependency between configurations and member forces to ac-
tively modify their configurations in order to satisfy certain requirements; see for
example the work by Shea, Fest and Smith (2002). Hence, it is extremely important
to accurately identify the current distribution of member forces.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
In the following, the components of vectors and matrices are indicated by subscripts,
e.g., as hx = (hxi ), H x = (Hij
x
). The length matrix L = (Lij ) is diagonal and given
as
q
Lii = (Hiix )2 + (Hiiy )2 + (Hiiz )2 , (i = 1, . . . , m) (9.85)
Let s denote the vector of member forces. In the state of self-equilibrium, the
equilibrium equation of a pin-jointed structure is written as (Schek, 1974)
Ds = 0 (9.86)
where the 3n m matrix D is called equilibrium matrix, which is defined by
> x
C H
D = C > H y L1 (9.87)
C>H z
The self-equilibrium equation (9.86) will be extensively used to identify the force
distribution of the structure in the state of self-equilibrium, based on measurement
data of the nodal coordinates and the forces of some members. It is easily observed
from Eq. (9.86) that the measurement errors in nodal coordinates in D and some
components of member forces in s have significant influence on the identification
accuracy of the member forces that are not measured, because the self-equilibrium
equation has to be exactly satisfied.
Suppose that the structure consists of h independent modes of member forces
that satisfy the self-equilibrium equation (9.86), where
h = m rank(D) (9.88)
For prestressed structures, h 1 always holds so that there exist forces in the
self-equilibrium state. As will be discussed in the following sections, at least
hindependent components of member forces need to be measured so as to determine
the forces of the remaining members based on the self-equilibrium equation.
Since C and C f are constant, and the measurement error of coordinates of fixed
nodes is not considered, Eq. (9.101) becomes
hx
= Ci (9.102)
xi
where C i is the ith column of C.
From H x = diag(hx ), we have
H x
= diag(C i ) (9.103)
xi
Since H y and H z do not depend on the x-coordinates, their sensitivity coefficients
with respect to xi vanish as
H y H z
= 0, =0 (9.104)
xi xi
x y z
The sensitivity coefficients of H , H and H with respect to yi and zi can be
formulated similarly to Eq. (9.103) and Eq. (9.104) as
H x H x
=0
=0
yi
y
ziy
H H
= diag(C j ) , =0 (9.105)
y i
zi
z
z
H
H
=0 = diag(C k )
yi zi
From Eq. (9.85), we have
L2 = (H x )2 + (H y )2 + (H z )2 (9.106)
Let t represent a variable xi , yi , or zi . The sensitivity coefficient of L with respect
to t is derived as x y z
L 1 x H y H z H
=L H +H +H (9.107)
t t t t
Since the sensitivity coefficient of L1 with respect to t is
L1 L
= L2 (9.108)
t t
the sensitivity coefficient of L1 with respect to t is written as follows by substi-
tuting Eq. (9.107) into Eq. (9.108):
L1 H x H y H z
= L3 H x + Hy + Hz (9.109)
t t t t
For the structure with given topology, the connectivity matrix C is a constant
matrix. From the definition of the equilibrium matrix D in Eq. (9.87), its sensitivity
coefficient with respect to a variable t is derived as
x 1
> H 1 > x L
C t L + C H t
D 1
> H y 1 L
= C L + C>H y (9.110)
t
t t
H z 1 L1
C> L + C>H z
t t
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
In SA, the initial solution and the cooling schedule play critical roles in find-
ing the optimal solution with good accuracy. Hence, it is more reliable to start
the SA from a better initial solution. Zhang, Ohsaki and Araki (2004) developed
several simple heuristic approaches, including the stingy method, to investigate the
identification accuracy of force distribution of prestressed pin-jointed structures,
where only the measurement errors of member forces are considered. It has been
demonstrated that the stingy method has relatively good accuracy with small com-
putational cost. Therefore, the stingy method can be effectively used for producing
a good initial solution, rather than random solutions, to reduce the computational
cost as well as to improve the accuracy in SA.
The stingy method is a basic heuristic approach, based on the local search,
to combinatorial optimization problems. For the problem considered here, it starts
from a complete set of the possible measurement members, and successively removes
the member with least contribution to the objective function from the current set
of solutions, under the condition that the removal does not lead to an infeasible
e
solution violating the requirement on rank deficiency of D.
The stingy method used in this example is summarized as follows:
Step 0: Set J = {1, . . . , m}, i.e., the forces of all members are to be measured,
and p = m.
Step 1 Let Ep0 and Epj denote the identification errors by the sets I and I {j},
respectively. Find the member k in I that minimizes Epj Ep0 .
Step 2 If |J | = p, then terminate the process; otherwise, set I I {k} and
|J | |J | 1 accordingly, because it has the minimum contribution in I to
reduction of the identification errors and go to Step 1.
This way, we can find the p measurement members with the approximately
minimal identification error for the optimization problem (9.112), which is used as
the initial solution for SA.
10
10
15 10 5 0 5 10 15
2
1
0
1
15 10 5 0 5 10 15
(c) side view
we have only 12 measurement devices, and the objective of this example is to find
12 cables, i.e., p = 12, to be measured so as to minimize the identification error.
The dotted lines in Fig. 9.27(a) show the cables to be measured, which are
determined by the stingy method only. The identification error in this case is
1.1988. As is seen, the members are not symmetrically located. We next use this
result as the initial solution for SA, and obtain the optimal locations of cables to be
measured as shown in the dotted lines in Fig. 9.27(b). The identification error in
this case is reduced to 1.0418, which is smaller than that of using the stingy method
only.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Chapter 10
Concluding Remarks
Would you tell me, please, which way I ought to go from here?
That depends a good deal on where you want to get to, said the
cat. (Lewis Carroll, Alice in Wonderland, Chapter 8)
Youve got to be very careful if you dont know where you are going,
because you might not get there. (Yogi Berra)
Optimal design became a rich and rewarding field of research. (Niord-
son, 2001)
In this chapter we try to summarize this book and some thoughts on the opti-
mization and probabilistic methods. It appears to us that total acceptance of the
ever-present uncertainty may infuse the old subject of optimization with the needed
new blood. The hybrid approach may help release both concepts from their con-
finement in the esoteric ivory tower and use them in the everyday quest for the
best and safest designs. To encourage further thought, the sections hereinafter
will be presented as questions in an attempt to provide replies in the form of first
approximations.
Ever since structural optimization came into being, researchers have been expressing
doubts as to its justification and/or practicality. The background can be described
as follows:
(1) Why should we look for the best design if an acceptable, hence, a good one, can
be found? In short, adherence to the American maxim If it aint broke, dont
fix it was recommended.
(2) In his review article Ashley (1982) describes a poll conducted among 46 people
about the merits and demerits of structural optimization. The responses were
327
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
as follows:
an engineer experienced in civil and aeronautical structures: One of the
reasons that I stopped work in optimization was my dismay... that there
were so very few applications.
a Dean of Engineering who has known the field for over a quarter of a
century: I do not recollect any applications.
a foremost specialist on synthesis with aeroelastic constraints: I am sorry,
but I dont really have any...
a recently retired senior design engineer, describing events at his aerospace
company:
For 15 years I beat my head against a stone wall... The end was:
formal optimization techniques were never used in aircraft design
(even to this day!). The company was forced to use [them] in its
subsequent ICBM and space programs.
an interview with a distinguished European aerodynamicist regarding wing
design:
Numerical optimization techniques were being used to some ex-
tent in prefeasibility studies, but [he was] unable to produce any
specific or referenceable material.
Did the situation change in the past quarter century? Venkataraman and Haftka
(2004) address this topic. They write:
Yet, in 2004 structural optimization practitioners hear the same re-
frain from structural analysts that was heard in 1971; a single struc-
tural analysis takes several hours or even several days to run on my
computer; how can I even afford to think about optimization?
However,
... computing speed and storage capacity increased one million fold
at a rate of about 100-fold increase every decade. In 1971, when
structural optimization was still struggling to deal with any real-
life problems, a one-million factor increase in computational speed
would have seemed to solve all problems. After all, engineers were al-
ready analyzing complex structures with detailed finite element (FE)
models, and the ability to do one million analyses is sufficient for
most optimization problems.
Others concluded that the paradox of the continuing difficulty of applying
optimization to real-life problems is, of course, due to increasing complexity of
the structural models that analysts consider adequate for predicting structural
response.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Venkataraman and Haftka also stress the progress made by structural optimiza-
tion:
Reservations regarding the various uncertainty analyses are not unlike their counter-
parts regarding the optimization approaches. Currently, the most utilized methods
for incorporating uncertainty are the probabilistic and fuzzy-sets based methodolo-
gies.
The book by Ben-Haim and Elishakoff (1990) devotes its first chapter to the
pros and cons of probabilistic modeling, including numerous pertinent quotations
in Sec.1.3; these will not be repeated here. Elishakoff (2000b) published a result of
the poll that he conducted for 12 years among more than 40 engineering researchers
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Soong (1988) notes: The gap between research and practice, it appears, is
widening... Sexmith and Nelson (1969) state: In spite of many notable contri-
butions... probabilistic concepts have not found their way into practice or design
codes in anything but a superficial way.
Cohn (1994) emphasizes:
... truly safe structures can only be assured by design and fab-
rication procedures which thoroughly examine and accommodate
the real causes of structural failures. The probability of failure
approach forms only a small part of these...
and Frangopol (2006). these volumes, except the latter one, are conveniently titles
Reliability and Optimization of Structural Systems.
Probabilistic modeling involves design with allowance for the aspect of reliability,
defined as the probability of successful performance of the structure. This means
that the engineer envisages some measure of the probability of failure, which is unity
minus the reliability. The main proponent and developer of probabilistic methods,
Alfred Freudenthal (1961) states: ... the notion that a finite (no matter how small)
probability of failure or at least of unserviceability is repulsive to a majority of
engineers. The persistent question arises: Has the situation changed since this
observation was made over half a century ago?
Vladimir Vasilievich Bolotin (see Elishakoff (2000b)) responds in the following
manner in the poll:
Grandori (1991) writes: ... the concept of structural safety will not leave the
realm of metaphysics unless we devise a method for justifying the choice of risk
acceptability levels.
Kalman (1994) is even harsher:
Schueller (1996) responds with the following statement, expressing guarded op-
timism:
It appears that this controversy, sometimes very hot, will continue in the fu-
ture. It is more pragmatic to pose a more practical question on how to model the
unavoidable uncertainty in engineering practice. Bolotin (see Elishakoff (2000b))
provides a useful insight:
It is this very methodology of finding the most favorable (best) responses along
with the least favorable (worst) ones that is adopted in this monograph. Bolotins
idea is overlaid with optimization study. Indeed, it may be overly conservative to
use the worst possible response without optimization: in fact, Kharitonov (1997)
tellingly titles his paper Interval uncertainty structure: conservative but simple.
Optimization minimizes overdesign, thereby becoming a natural ally in looking for
the worst-case scenario, which was dubbed here and elsewhere as anti-optimization.
Therefore, this book deals with optimization (Chap. 2), anti-optimization (Chaps.
38) and their hybrid (Chap. 9).
Sniedovich (2008a) and Burgman (2008) include Shakespeares name in the titles
of their respective papers. Burgman (2008) notes:
Leonhard Euler (17071783) stressed that For since the fabric of the universe
is most perfect and the work of a most wise Creator, nothing at all takes place
in the universe in which some rule of maximum or minimum does not appear.
(Tikhomirov, 1990).
In theoretical and applied mechanics, the idea of minimization of the upper
bounds of mechanical characteristics apparently occurs first in the 1877 book by
J. W. Strutt (Lord Rayleigh), Theory of Sound, in which a method of determining
the natural frequencies of discrete and continuous structures was presented. The
method now bears his name. Specifically, he addressed the well-known problem of
the vibrating string. For the fundamental mode he assumed
n
2x
w(x, t) = 1 cos t (10.1)
L
where is the sought circular frequency, 2L is the length of the string, x is the axial
coordinate (he chose x = 0 at the midpoint of the string). In the end, Rayleigh
arrived at the following formula:
2(n + 1)(2n + 1) T
2 = (10.2)
2n 1 L2
where T is the tension in the string, and is the mass density. For n = 1, the formula
yields = 2 L2 /T = 12, instead of the exact value = 2 ' 9.8696; n = 2 yields
a better result, = 10, whereas a minimum , or in Leissas (2005) words the best
possible result from Eq. (10.2) is 9.8990 when n = ( 6 + 1)/2 = 1.72474.
Clearly, Rayleigh carried out minimization of the upper bound of the fundamen-
tal natural frequency. Apparently he was also the first to use (see p. 287 in his
book) the combined term maximumminimum condition. (A detailed description
of Rayleighs method can be found in almost any text on vibration, and a careful
paper on its history was published by Leissa (2005)). Use of a function containing a
power-law term with an undetermined exponent was resurrected by Schmidt (1985),
and reviewed extensively by Bert (1987). Note that minimization of the maximum
response under a constraint on structural weight is equivalent to that of structural
weight under a constraint on the response.
Sniedovich (2006, 2007, 2008a, 2008b, 2008c) attributes the modern terminol-
ogy of the maximin principle to Abraham Wald (19021950) as follows: Rank
alternatives by their worst possible outcomes: adopt the alternative the worst out-
come of which is at least as good as the worst outcome of the others. Moreover,
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
am happy with the old terms worst case and best-case and
Maximin. I have been using them for more that 40 years.
Here the following personal anecdote seems to be instructive. Some years ago
one of us (I.E.) lectured at a certain university; then, the idea of combined anti-
optimization and optimization was propageted. In the questions-and-answers pe-
riod, one faculty member said: I dont like the term anti-optimization, in fact, had I
gotten your manuscript for review, I would have rejected it, just for this term! The
answer was that one of the leading optimization experts, Professor R. T. Haftka, had
endorsed it. The faculty member responded, If Haftka has embraced this notion, I
feel and am defeated! It is pleasing to record that this term has been adopted by
many researchers around the world.
The above correspondence with Sniedovich helped us to spell out the need
to discuss its various brothers and sisters. We hope that in this section we
have correlated the various origins, counterparts and equivalents of the term
optimization/anti-optimization hybrid. As can be seen, applied mechanics pro-
vided development of ideas from various analyses, leading to terms worst-case.
least favorable, best-case, most favorable, minmax or anti-optimization with
a view to determining the worst-case scenarios. Then, combined optimization seeks
to find the minimum of the worst response, or the least worst response. Hence, it is
worthwhile to make comprehensive survey of the problem formulations and method-
ologies for optimization considering maximum (worst-case) responses. Although the
idea is known independently in several fields of engineering, this book presents all
the necessary knowledge and approaches, as well as the philosophy underlining the
hybrid concept, and establishes a new direction based on computer technology.
In Chap. 4 we analyze the worst, or anti-optimized, responses in a static setting,
whereas Chap. 6 deals with analogous vibration problem. In the buckling context,
worst-case or anti-optimization entails finding the worst (or least favorable) re-
sponse, i.e., the minimum buckling load. Combination with anti-optimization leads
then to maximization of the minimum buckling load. Since the FEM is the central
numerical method in applied mechanics, it made sense to discuss, in Chap. 7, how
the anti-optimization is performed in a discretized environment. The topic is under
active development and the final word has not been said about it, luckily for us
researchers!
Chapter 8 correlates the probabilistic and anti-optimization approaches, showing
perhaps to the disappointment of separationists (those who divide and conquer
the specific fields of research) and hopefully, to the satisfaction of unionists (those
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
who try to find similarities in seemingly different fields) that the two approaches
lead to close or even the same results. Our general conclusion is in line with the
title of the paper by Bai and Andersland (1994) Stochastic and Worst Case Sys-
tem Identification are not Necessarily Incompatible; moreover, Bai and Andersland
claim:
Clearly the two approaches [stochastic and worst case I.E. and
M.O.] are based on different philosophies. In the literature they are
usually treated differently and separately, we believe that they are
not necessarily incompatible, and, in fact, may be complementary.
This study does not aim to make a bombastic statement like that
of de Finetti (1974), Probability does not exist. We can take
the liberty of suggesting to consider two possible scenarios: either
probability does not exist or it does. Consider first the former case:
if probability does not exist, then obviously, anti-optimization
method is of utmost importance, and it must be utilized in de-
sign. Consider now the latter case: if probability exists, then for
the structure to be acceptable it must possess high, near-unity
reliability; for the case of bounded random variations, the prob-
abilistic method tends to one furnished by the anti-optimization.
Thus, irrespective of the existence or nonexistence of probability,
the anti-optimization method must be incorporated.
Likewise, Takao Fujisawa (see Sato (2006)) mentions: Its only the man who
can look at the same problem from many different aspects that will make a true
leader.
It is noteworthy that stochasticians are tending more and more to invoke the
fact of physically realizable random variables and processes. To review the cur-
rent thinking of the stochasticians and the interested reader may like to consult
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
the papers by Cai (2003), Cai and Lin (2006), Grigoriu (2006), Ma, Leng, Meng
and Fang (2004), and Simiu and Heckert (1996) already quoted and discussed in
Sec. 1.3. It appears therefore that the prediction of increased appreciation, of the
anti-optimization method in the future, is not premature.
The combination of the probabilistic and ellipsoidal analyses was facilitated in
works by Elishakoff and Colombi (1993), Elishakoff, Lin and Zhu (1994e), Elishakoff
and Li (1999), and Qiu, Yang and Elishakoff (2008b). In this context Drenick (1995)
wrote in the correspondence with one of us (I.E.)
You mentioned that the problems you liked the best had both
probabilistic and non-probabilistic aspects. I agreed with you be-
cause I felt that the treatment of such mixed problems might
not run into the same unyielding resistance that is encountered by
the pure non-probabilistic type.
We would like to end with the provocative question in the above subtitle. Indeed,
with the abundance of texts on optimization and/or uncertainty why bother with
another?
Our answer is: because it combines these two cardinal ideas, thereby ejecting
them from their respective ivory-tower spheres and providing a friendly environment
for the much needed communication and interaction. In such circumstances, the
existential question posed by Laville (2000), Shall we abandon optimization the-
ory? can be answered in the negative, and the equally important one by Mongin
(2000) Does optimization imply rationality? in the affirmative.
The thesis that the optimization/anti-optimization hybrid saves both optimiza-
tion and uncertainty analysis from pure academism is an exaggeration. In fact, the
reliability-based design optimization (RBDO) is an established field answering the
above need of marrying uncertainty and optimization. As Chateauneuf and Aoues
(2007) note:
Moreover,
Note that the overwhelming majority of works on RBDO and SRBDO deal in
independent random variables. The need to allow for interdependence in stochastic
analysis is discussed in numerous papers; the interested reader is referred to the
papers of Ferson, Ginzburg and Akcakaya (1995) and Ferson, Nelsen, Hajagos,
Berleant, Zhang, Tucker, Ginzburg and Oberkampf (2004).
Ganzerli and de Palma (2007) write:
Bibliography
Aarts, E. and Korst, J. (1989). Simulated Annealing and Boltzmann Machines: A Stochas-
tic Approach to Combinatorial Optimization and Neural Computing (Wiley, Chich-
ester, U.K.).
Aarts, M. and Lenstra, J. K. (eds.) (1997). Local Search in Combinatorial Optimization
(John Wiley & Sons, Chichester, U.K.).
Abbas, A. M. and Manohar, C. S. (2005). Reliability-based critical earthquake load mod-
els, Part I: Linear structures, Part II: Nonlinear structures, Journal of Sound and
Vibration 287, pp. 865882:883900.
Abbas, A. M. and Manohar, C. S. (2007). Reliability-based vector nonstationary random
critical earthquake excitations for parametrically excited systems, Structural Safety
29, pp. 3248.
Abdullah, M. M., Richardson, A. and Hanif, J. (2001). Placement of sensors/actuators
on civil structures using genetic algorithms, Earthquake Engineering and Structural
Dynamics 30, pp. 11671184.
Adali, S. (1982). Minimum-maximum deflection and stress design of beams under harmonic
excitation by mathematical programming, Engineering Optimization 5, 4, pp. 223
234.
Adali, S. (1992). Convex and fuzzy modelings of uncertainties in the optimal design of
composite structures, in P. Pedersen (ed.), Optimal Design with Advanced Materials
(Elsevier, Amsterdam, Netherlands).
Adali, S., Elishakoff, I., Richter, A. and Verijenko, V. E. (1994a). Optimal design of sym-
metric angle-ply laminates for maximum buckling load with scatter in material prop-
erties, in J. Sobieszczanski-Sobieski (ed.), Proc. 5th AIAA/USAF/NASA/ISSMO
Symposium on Multidisciplinary Analysis and Optimization, Vol. 2, pp. 10411045,
PaperAIAA944365CP.
Adali, S., Lene, F., Duvaut, G. and Chiaruttini, V. (2003). Optimization of laminated
composites subjected to uncertain buckling loads, Composite Structures 62, pp.
261269.
Adali, S., Richter, A. and Verijenko, V. E. (1995a). Minimum weight design of symmetric
angle-ply laminates under uncertain loads, Structural and Multidisciplinary Opti-
mization 9, 2, pp. 8995.
Adali, S., Richter, A. and Verijenko, V. E. (1995b). Non-probabilistic modelling and design
of sandwich plates subject to uncertain loads and initial deflections, International
Journal of Engineering Science 33, pp. 855866.
Adali, S., Richter, A. and Verijenko, V. E. (1997). Minimum weight design of symmetric
angle-ply laminates with incomplete information on initial imperfections, Journal of
343
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 345
Ashley, H. and Zartarian, C. (1956). Piston theory: A new aerodynamic tool for the
aeroelastician, Journal of the Aeronautical Sciences 23, pp. 11091118.
Attoh-Okine, N. O. (2002). Uncertainty analysis in structural number determination in
flexible pavement design a convex model approach, Construction and Building
Materials 16, 2, pp. 6771.
Attoh-Okine, N. O. (2004). Uncertainty analysis in surface micromachined force gauges:
Convex model approach, Journal of Aerospace Engineering 17, 1, pp. 4044.
Au, F. T. K., Cheng, Y. S., Tham, L. G. and Zheng, G. W. (2003). Robust design of
structures using convex models, Computers & Structures 81, 2829, pp. 26112619.
Averin, A. M. (1964). On the bending of a cantilever plate weakened by a hole of a general
shape, in Proc. 4th USSR Conference on Plates and Shells (American Academy of
Sciences Press, Erevan, Armenia), pp. 167170, (in Russian).
Avis, D. and Fukuda, K. (1996). Reverse search for enumeration, Discrete Applied Math-
ematics 65, 13, pp. 2146.
Ayyub, B. (1998). Uncertainty Modeling and Analysis in Civil Engineering (CRC Press,
Boca Raton, FL).
Babuska, I., Nobile, F. and Tempone, R. (2005). Worst case scenario analysis for elliptic
problem with uncertainty, Numerische Mathematik 101, 2, pp. 185219.
Bagley, R. L. and Torvik, P. J. (1983). Fractional calculus: A different approach to the
analysis of viscoelastically damped structures, AMA Journal 21, 5, pp. 741748.
Bai, E.-W. and Andersland, M. S. (1994). Stochastic and worst-case system identification
are not necessarily incompatible, Automatics 30, 9, pp. 14911493.
Baitsch, M. and Hartmann, D. (2006). Optimization of slender structures considering
geometrical imperfections, Inverse Problems in Science and Engineering 14, 6, pp.
623637.
Balandin, D. V. (1993). Accumulation of perturbations in linear and non-linear shock-
driven systems, Journal of Applied Mathematics and Mechanics 57, 1, pp. 2127.
Banichuk, N. V. (1975). Game problems in the theory of optimal design, in Proc. IUTAM
Symposium on Optimization in Structure Design (Springer Verlag, Berlin), pp. 111
121.
Banichuk, N. V. (1976). Minimax approach to the structural optimization problems, Jour-
nal of Optimization Theory and Applications 20, 1, pp. 111127.
Banichuk, N. V. and Ivanova, S. Y. (2009). Optimization problems of contact problems,
Mechanics Based Design of Structures and Machines 37, 2, pp. 143156.
Banichuk, N. V. and Neittaanm aki, P. (2007). On structural optimization with incomplete
information, Mechanics Based Design of Structures and Machines 35, 1, pp. 7596.
Baratta, A. (1980). Analisi dei dati e simulazione per la sicurita antisismica delle
costruzioni, in Proc. National Conference L Ingegneria Sismica in Italia (Udine,
Italy).
Baratta, A. (1981). Optimized earthquake simulation for high-reliability aseismic design,
in Proc. 3rd International Seminar on Reliability of Nuclear Power Plants (Paris).
Baratta, A., Elishakoff, I., Zuccaro, G. and Shinozuka, M. (1998). A generalization of the
DrenickShinozuka model for bounds on the seismic response of a single-degree-of-
freedom system, Earthquake Engineering and Structural Dynamics 27, pp. 423437.
Baratta, A. and Zuccaro, G. (1992). High-reliability aseismic design, in Proc. 10th World
Conference in Earthquake Engineering (Madrid).
Barbieri, E., Cinquini, C. and Lombardi, M. (1997). Shape/size optimization of truss struc-
tures using non-probabilistic description of uncertainty, in S. Hern andez and C. A.
Brebbia (eds.), Computer Aided Optimum Design of Structures V (Computational
Mechanics Publications, Southampton, U.K.), pp. 163172.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 347
Boley, B. A. (1964). Upper and lower bounds in problems of melting and solidifying slabs,
Quarterly Journal of Mechanics and Applied Mathematics 17, pp. 253269.
Boley, B. A. (1966a). Bounds on the maximum thermoelastic stress and deflection in a
beam and plate, Journal of Applied Mechanics 33, 4, pp. 881887.
Boley, B. A. (1966b). Estimate of errors in approximate temperature and thermal stress
calculations, in H. Gortler and P. Sorger (eds.), Applied Mechanics, Proc. 11th Inter-
national Congress of Applied Mechanics, Munich, Germany, 1964 (Springer, Berlin).
Boley, B. A. and Weiner, J. H. (1960). Theory of Thermal Stresses (Wiley, New York).
Bolotin, V. V. (1961). Statistical Methods in Structural Mechanics (Gosstroiizdat,
Moscow), (First edition, in Russian, second edition, 1965).
Bolotin, V. V. (1963). Non-Conservative Problems in the Theory of Elastic Stability (Perg-
amon, New York), pp. 242257.
Bolotin, V. V. (1969). Statistical Methods in Structural Mechanics, Holden-Day Series
in Mathematical Physics 7 (Holden-Day, San Francisco, CA), (Translation of the
second, 1965 edition).
Borkowski, A. (1983). On application of ellipsoidal methods in rigid-plastic analysis and
design, in H. A. Eschenauer and N. Olhoff (eds.), Optimization Methods in Structural
Design (Wissenschaftsverlag, Mannheim, Germany), pp. 289294.
Borse, G. J. (1997). Numerical Method with MATLAB (International Thomson Publish-
ing).
Boyd, S. and Vandenberghe, L. (2004). Convex Optimization (Cambridge University Press,
Cambridge, U.K.).
Brach, R. M. (1968). Minimum dynamic response for a class of simply supported beam
shapes, International Journal of Mechanical Science 10, pp. 429439.
Brown, C. B. (1979). A fuzzy safety measure, Journal of Engineering Mechanics Division,
ASCE 105, 5, pp. 855872.
Brown, C. B., Johnson, J. L. and Loftus, J. J. (1984). Subjective seismic safety assessments,
Journal of Structural Engineering 110, 9, pp. 22122233.
Brown, C. B. and Yao, J. T. P. (1987). Resent civil engineering applications of fuzzy sets,
Mathematical Modelling 9, 6, pp. 491492.
Bruhn, E. F. (ed.) (1975). Analysis and Design of Missile Structures (Tri-Stable Offset
Company, Cincinnati, OH).
Budiansky, B. and Hutchinson, J. W. (1964). Dynamic buckling of imperfection-sensitive
structures, in H. G ortler and P. Sorger (eds.), Applied Mechanics, Proc. 11th Inter-
national Congress of Applied Mechanics (Springer, Berlin), pp. 636651.
Bulgakov, B. V. (1939). Applied Theory of Gyroscopes (Moscow), (in Russian).
Bulgakov, B. V. (1940). Fehleranhaefung bei Kreiselapparaten, Ingenieur-Archiv 11, pp.
461469, (in German).
Bulgakov, B. V. (1946). On the accumulation of disturbances in linear systems with con-
stant coefficients, Doklady Akademii Nauk SSSR LI(5), pp. 339342, (in Russian).
Bulgakov, B. V. and Kozovkov, N. T. (1950). On the accumulation of disturbances in linear
systems with variable coefficients, Prikladnaya Matematika i Mekhanika (Applied
Mathematics and Mechanics) 14, pp. 712, (in Russian).
Burgman, M. (2008). Shakespeare, Wald and decision making under uncertainty, Decision
Point 23, p. 10.
Burns, S. (ed.) (2002). Recent Advancements in Optimal Structural Design (ASCE Press,
Raton, VA).
Cai, G. Q. (2003). Generation of bounded stochastic process, in W. Q. Zhu, G. Q. Cai
and R. C. Zhang (eds.), Advances in Stochastic Structural Engineering (CRC Press,
Boca Raton, FL).
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 349
perturbation method for predicting the natural frequency bounds of intelligent truss
structures with uncertain-but-bounded parameters, Key Engineering Materials 347,
pp. 569574.
Cheng, G., Xu, L. and Jiang, L. (2006). A sequential approximate programming strategy
for reliability-based structural optimization, Computers & Structures 84, pp. 1353
1367.
Cheng, Y. S., Au, F. T. K., Tham, L. G. and Zeng, G. W. (2004). Optimal and robust
design of doking blocks with uncertainty, Engineering Structures 26, pp. 499510.
Cherkaev, A. V. and Cherkaev, E. (1999). Optimal design for uncertain loading conditions,
in V. Berdichevsky, V. Jikov and G. Papanicolaou (eds.), Homogenization (World
Scientific, Singapore), pp. 193213.
Cherkaev, A. V. and Kucuk, I. (1999). Optimal structures for various loadings, in Proc.
3rd World Congress of Structural and Multidisciplinary Optimization (Buffalo, N.
Y.).
Cherkaev, E. and Cherkaev, A. V. (2003). Principal compliance and robust optimal design,
Journal of Elasticity 72, pp. 7198.
Cherkaev, E. and Cherkaev, A. V. (2008). Minimax optimization problem of structural
design, Computers & Structures 86, 1314, pp. 14261435.
Chern, J. M., Dafalias, J. F. and Martin, J. B. (1973). Structural design for bounds on
dynamic response, Journal of Engineering Mechanics Division, ASCE 99, pp. 261
270.
Chernousko, F. L. (1980). Optimal guaranteed estimates of uncertainties by means of
ellipsoids, IIII, Tekhnicheskaya Kibernetika Part I, issue 2, 311: Part II, issue 4,
311: Part III, issue 5, 511 (in Russian), (translated into English as Engineering
Cybernetics).
Chernousko, F. L. (1988). Estimation of the Phase State of Dynamic Systems (Nauka
Publishing House, Moscow), (in Russian).
Chernousko, F. L. (1994). State Estimation of Dynamic Systems (CRC Press, Boca Raton,
FL).
Chernousko, F. L. (1999). What is ellipsoidal modeling and how to use it for control and
state estimation? in I. Elishakoff (ed.), Whys and Hows in Uncertainty Modelling
(Springer, Vienna), pp. 127188.
Chernousko, F. L. (2000). Ellipsoidal bounds of reachable sets of dynamical systems with
matrices subjected to uncertain perturbations, Journal of Optimization Theory and
Application 104, 1, pp. 119.
Chernousko, F. L. (2005). Ellipsoidal state estimation for dynamical systems, Nonlinear
Analysis 63, pp. 872879.
Chiang, W. L., Dong, W. M. and Wong, F.-S. (1987). Dynamic response of structures with
uncertain parameters : A comparative study of probabilistic and fuzzy set models,
Probabilistic Engineering Mechanics 2, pp. 8291.
Cho, M. and Rhee, S. Y. (2003). Layup optimization considering free-edge strength and
bounded uncertainty of material properties, AIAA Journal 41, 11, pp. 22742282.
Cho, M. and Rhee, S. Y. (2004). Optimization of laminates with free edges under bounded
uncertainty subject to extension, bending and twisting, International Journal of
Solids and Structures 41, pp. 227245.
Choi, K. K. and Kim, N.-H. (2004). Structural Sensitivity Analysis and Optimization 1,
Linear Systems (Springer, New York).
Christensen, R. M. (1982). Theory of Viscoelasticity (Academic Press, New York).
Chryssanthopoulos, M. K. and Poggi, C. (1995). Stochastic imperfection modeling in shell
buckling studies, Thin-Walled Structures 23, 14, pp. 179200.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 351
Bibliography 353
Bibliography 355
Yeager (eds.), Proc. 1994 Conference on Noise Control Engineering (New York),
pp. 519524.
Elishakoff, I., Eliseeff, P. and Glegg, S. (1994b). Convex modelling of material uncertainty
in vibrations of viscoelastic structures, AIAA Journal 32, pp. 843849.
Elishakoff, I. and Fang, J. J. (1995). Diagnosis of local uncertain modification in the
boundary condition of a rectangular plate via convex modeling, Computer Methods
in Applied Mechanics and Engineering 124, 4, pp. 303319.
Elishakoff, I. and Ferracuti, B. (2006a). Four alternative definitions of the fuzzy safety
factor, Journal of Aerospace Engineering 19, 4, pp. 281287.
Elishakoff, I. and Ferracuti, B. (2006b). Fuzzy sets based interpretation of the safety factor,
Fuzzy Sets and Systems 157, 18, pp. 24952512.
Elishakoff, I., Gana-Shvili, J. and Givoli, D. (1991). Treatment of uncertain imperfec-
tions as a convex optimization problems, in L. Esteva and S. E. Ruiz (eds.), Proc.
6th International Conference on Applications of Stochastics and Probability in Civil
Engineering, Vol. 1 (Mexico City), pp. 150157.
Elishakoff, I., Haftka, R. T. and Fang, J. (1994c). Structural design under bounded
uncertaintyoptimization with anti-optimization, Computers & Structures 53, 6,
pp. 14011405.
Elishakoff, I. and Hasofer, A. M. (1996). Detrimental and serendipitous effect of human
error on reliability of structures, Computer Methods in Applied Mechanics and En-
gineering 129, pp. 17.
Elishakoff, I. and Li, Q. (1999). How to combine probabilistic and anti-optimization meth-
ods? in Whys and Hows of Uncertainty Modeling: Probability, Fuzziness, and Anti-
Optimization, CISM Courses and Lectures, No. 388 (Springer, Vienna), pp. 319339.
Elishakoff, I., Li, Y. W. and Starnes, Jr., J. H. (1994d). A deterministic method to pre-
dict the effect of unknown-but-bounded elastic moduli on the buckling of composite
structures, Computer Methods in Applied Mechanics and Engineering 111, 12, pp.
155167.
Elishakoff, I., Li, Y. W. and Starnes, Jr., J. H. (1996). Imperfection sensitivity due to
the elastic moduli in the RoordaKoiter frame, Chaos, Solitons & Fractals 7, 8, pp.
11791186.
Elishakoff, I., Li, Y. W. and Starnes, Jr., J. H. (2001). Non-Classical Problems in the
Theory of Elastic Stability (Cambridge University Press, Cambridge, U.K.).
Elishakoff, I., Lin, Y. K. and Zhu, L. P. (1994e). Probabilistic and Convex Modelling of
Acoustically Excited Structures, Studies in Applied Mechanics 39 (Elsevier, Amster-
dam, Netherlands).
Elishakoff, I. and Pletner, B. (1991). Analysis of base excitation as an uncertain function
with specified bounds on it and its derivatives, in Structural Vibration and Acoustics,
Proc. ASME Design Technical Conference (Miami, FL), pp. 177183.
Elishakoff, I., Ren, Y. J. and Shinozuka, M. (1997). New formulation of FEM for deter-
ministic and stochastic beams through generalization of Fuchs approach, Computer
Methods in Applied Mechanics and Engineering 144, 3-4, pp. 235243.
Elishakoff, I., van Manen, S., Vermeulen, P. G. and Arbocz, J. (1987). First-order second-
moment analysis of the buckling of shells with random initial imperfections, AIAA
Journal 25, 8, pp. 11131117.
Elishakoff, I., Wang, X. J. and Qiu, Z. P. (2008). Comparison of interval analysis and
convex analysis, in R. L. Muhanna and R. L. Mullen (eds.), Proc. 3rd International
Workshop on Reliable Engineering Computing (Georgia Institute of Technology, Sa-
vannah, GA), pp. 269288.
Elishakoff, I. and Zingales, M. (2000). Probability vs anti-optimization in an applied me-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
chanics problem: Part I, scalar uncertainty, Journal of Applied Mechanics 67, pp.
472483.
Elishakoff, I. and Zingales, M. (2003). Contrasting probabilistic and anti-optimization
approaches in an applied mechanics problem, International Journal of Solids and
Structures 40, pp. 42814297.
Elseifi, M. (1998). A New Scheme for the Optimum Design of Stiffened Composite Panels
with Geometric Imperfections, Ph. D dissertation, Department of Aerospace Engi-
neering, Virginia Polytechnic Institute and State University, Blacksburg, VA.
Elseifi, M. A., G urdal, Z. and Nikolaidis, E. (1999). Convex/probabilistic models of un-
certainties in geometric imperfections of stiffened composite panels, AIAA Journal
37, 4, pp. 468474.
Elseifi, M. A. and Khalessi, M. R. (2001). A convex set based optimization algorithm
for structural probabilistic analysis, in 42nd AIAA/ASME/AHS/ASCE Structures,
Structural Dynamics, and Materials Conference and Exhibit.
Eschenauer, H. A. (2003). Optimization procedures in structural design, Applied Mechanics
Reviews 56, 4, pp. R19R26.
Eschenauer, H. A. and Olhoff, N. (2001). Topology optimization of continuum structures:
A review, Applied Mechanics Reviews 54, 4, pp. 331390.
Fang, J. J., Smith, S. A. and Elishakoff, I. (1998). Combination of anti-optimization and
fuzzy-set-based analysis for structural optimization under uncertainty, Mathematical
Problems in Engineering 4, pp. 187200.
Fatemi, J., Mooij, E., Gurav, S. P., Andreykiv, A. and van Keulen, F. (2005). Re-entry
vehicle design optimization with integrated trajectory uncertainty, in AIAA/CIRA
13th International Space Planes and Hypersonics Systems and Technologies Confer-
ence, AIAA-2005-3386 (Capua, Italy).
Fellin, W., Lessmann, L., Oberguggenberger, M. and Vieider, R. (eds.) (2005). Analyzing
Uncertainty in Civil Engineering (Springer, Berlin).
FEMA (2000). Recommended Seismic Design Criteria for New Steel Moment-Frame Build-
ings, FEMA-350, Federal Emergency Management Agency, Washington D.C.
Ferson, S. (2002). Ramas Risk Calc 4.0 Software: Risk Assessment with Uncertain Numbers
(Lewis Publishers, Boca Raton, FL).
Ferson, S. and Ginzburg, L. R. (1996). Different methods are needed to propagate igno-
rance and variability, Reliability Engineering and System Safety 54, 2, pp. 133144.
Ferson, S., Ginzburg, L. R. and Akcakaya, R. (1995). Whereof one cannot speak: when
input distributions are unknown, (Available at http://www.ramas.com.whereof.pdf).
Ferson, S., Nelsen, R. B., Hajagos, J. G., Berleant, D. J., Zhang, J., Tucker, W. T.,
Ginzburg, L. R. and Oberkampf, W. L. (2004). Dependence on probabilistic mod-
eling, DempsterShafer theory and probability bounds analysis, in Sandia Report
SAND 20043072.
Feynman, R. F. (1988). What Do You Care What Other People Think? (Bantham Books,
New York).
Floudas, C. A. (1995). Nonlinear and Mixed-Integer Optimization (Oxford University
Press, New York).
Fl
ugge, W. (1975). Viscoelasticity (Springer-Verlag, Berlin).
Fox, R. L. and Kapoor, M. P. (1970). Structural optimization in the dynamic response
regime: A computational approach, AIAA Journal 8, 10, pp. 17981804.
Frangopol, D. M. (1995). Reliability-based optimum structural design, in C. Sundararajan
(ed.), Probabilistic Structural Mechanics Handbook: Theory and Industrial Applica-
tions (Chapman & Hall, New York), pp. 352387.
Frangopol, D. M. (1998). Probabilistic structural optimization, Progress in Structural En-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 357
Bibliography 359
234.
Hilton, H. H. and Feigen, M. (1960). Minimum weight analysis based on structural relia-
bility, Journal of Aerospace Science 27, 9, pp. 641652.
Hilton, H. H., Hsu, J. and Kirby, J. S. (1991). Linear viscoelastic analysis with random
material properties, in Y. K. Lin and I. Elishakoff (eds.), Stochastic Structural Dy-
namics 1: New Theoretical Developments (Springer-Verlag, Berlin), pp. 83110.
Hirota, K. (1991). Fuzzy concept very clear in Japan, Asahi Evening News, August 20,
Special Supplement .
Hlavacek, I. (2002a). Post-buckling range of plates in axial compression with uncertain
initial imperfection, Applied Mathematics 47, pp. 25444.
Hlavacek, I. (2002b). Worst scenario approach for elastoplasticity with hardening and
uncertain input data, Zeitschrift f
ur angewandte Mathematik und Mechanik (ZAMM)
82, pp. 671684.
Hlavacek, I. (2007). Von K arman equations with uncertain input data and worst scenario
method, Zeitschrift f ur angewandte Mathematik und Mechanik (ZAMM) 87, pp.
747761.
Hlavacek, I., Chleboun, J. and Babuska, I. (2004). Uncertain Input Data Problem and the
Worst-Case Scenario Method (North-Holland, Amsterdam, Netherlands).
Hlavacek, I., Plesek, J. and Gabriel, D. (2006). Validation and sensitivity study of an
elastoplastic problem using the worst scenario method, Computer Methods in Applied
Mechanics and Engineering 195, pp. 763774.
Ho, D. (1974). Buckling load of nonlinear systems with multiple eigenvalues, International
Journal of Solids and Structures 10, pp. 13151330.
Hoff, N. J. (1965). in G. Herrmann (ed.), Dynamic Stability of Structures (Pergamon Press,
New York), pp. 744.
Hoh, S. and Hagaki, H. (1989). Fuzzy-Bayesian analysis to structural reliability, in Proc.
ICOSSAR89, San Francisco (ASCE Press, New York), pp. 17711774.
Holmes, J. D. (2002). Discussion of Classical extreme value model and prediction of
extreme wind, by Janos Galambos and Nicholas Macri, Journal of Structural En-
gineering 125, 7, pp. 792794.
Horn, R. A. and Johnson, C. R. (1990). Matrix Analysis, reprint edn. (Cambridge Univer-
sity Press, Cambridge, U.K.).
Horst, R. and Tuy, H. (1990). Global Optimization (Springer-Verlag, Berlin).
Hu, B. and Schiehlen, W. (1996). Eigenvalue, frequency response and variance bounds of
linear damped systems, European Journal of Mechanics, A/Solids 15, pp. 617646.
Huang, W. N. and Cozzarelli, F. A. (1972). Steady creep in a beam with random material
parameters, Journal of the Franklin Institute 294, 5, pp. 323338.
Huang, W. N. and Cozzarelli, F. A. (1973). Damped lateral vibration in an axially creep-
ing beam with random material parameters, International Journal of Solids and
Structures 9, pp. 765788.
Huseyin, K. (1975). Nonlinear Theory of Elastic Stability (Noordhoff, Leyden, Nether-
lands).
Ikeda, K. and Murota, K. (1990). Critical initial imperfection of structures, International
Journal of Solids and Structures 26, 8, pp. 865886.
Ikeda, K., Ohsaki, M. and Kanno, Y. (2005). Imperfection sensitivity of hilltop branching
points of systems with dihedral group symmetry, International Journal of Non-linear
Mechanics 40, pp. 755774.
Ikeda, K., Oide, K. and Terada, K. (2002). Imperfection sensitive strength variation of crit-
ical loads at hilltop bifurcation point, International Journal of Engineering Science
40, pp. 743772.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 361
Ilyushin, A. A. (1956). The law of plane sections at high supersonic velocities, PMM;
Journal of Applied Mathematics and Mechanics 20, pp. 733755, (in Russian).
Inman, D. (1989). Vibration analysis of viscoelastic beams by separation of variables and
modal analysis, Mechanics Research Communications 16, 4, pp. 213218.
Ishida, M. and Tagami, S. (1959). On the tension of an infinite strip containing a hole of
an arbitrary profile, in Proc. 9th Japan National Congress for Applied Mechanics,
pp. 5154.
Iyengar, R. N. (1970). Matched inpurs, Report 47, Series Y, Center of Applied Statistics,
Purdue University, West Lafayette, IN.
Iyengar, R. N. and Dash, P. K. (1976). Random vibration analysis of stochastic time-
varying systems, Journal of Sound and Vibration 45, pp. 6989.
Iyengar, R. N. and Manohar, C. S. (1987). Nonstationary random critical seismic excita-
tions, Journal of Engineering Mechanics 113, 4, pp. 529541.
Jensen, H. (2000). An interval analysis approach to deal with earthquake response of uncer-
tain systems, in Proc. 12th World Conference on Earthquake Engineering (Auckland,
New Zealand).
Jensen, H. (2001). Structural optimal design of systems with imprecise properties: a pos-
sibilistic approach, Advances in Engineering Software 32, 12, pp. 937948.
Jensen, H. and Maturana, S. (2002). A possibilistic decision support system for impre-
cise mathematical programming problems, International Journal of Production Eco-
nomics 77, 2, pp. 145158.
John, F. (1998). Extremum problems with inequalities as subsidiary conditions, in
K. Friedrichs, O. Neugebauer and J. J. Stocker (eds.), Studies and Essays Presented
to R. Courant on His 60th Birthday (Wiley, New York).
Kalinowski, A. and Pilkey, W. D. (1975). Design for incompletely prescribed loading,
Journal of Engineering Mechanics Division, ASCE 101, 4, pp. 505510.
Kalman, R. E. (1994). Randomness reexamined, Modeling, Identification and Control 15,
3, pp. 141151.
Kam, T.-Y. and Brown, C. B. (1983). Updating parameters with fuzzy entropies, Journal
of Engineering Mechanics 109, 6, pp. 13341343.
Kanda, J. (1994). Application of an empirical extreme value distribution to load models,
Journal of Research at the National Institute of Standards and Technology 99, pp.
413420.
Kanno, Y. and Ohsaki, M. (2005). Minimum principle of complementary energy for non-
linear elastic cable networks with geometrical nonlinearity, Journal of Optimization
Theory and Application 126, 3, pp. 617641.
Kanno, Y., Ohsaki, M. and Ito, J. (2002). Large-deformation and friction analysis of
non-linear elastic cable networks by second-order cone programming, International
Journal for Numerical Methods in Engineering 55, pp. 10791114.
Kanno, Y., Ohsaki, M. and Katoh, N. (2001). Sequential semidefinite programming for
optimization of framed structures under multimodal buckling constraints, Interna-
tional Journal of Structural Stability and Dynamics 1, 4, pp. 585602.
Kanno, Y. and Takewaki, I. (2006a). Confidence ellipsoids for static response of truss
with load and structural uncertainty, Computer Methods in Applied Mechanics and
Engineering 196, pp. 393403.
Kanno, Y. and Takewaki, I. (2006b). Robustness analysis of trusses with separable local
and structural uncertainties, International Journal of Solids and Structures 93, 9,
pp. 26962669.
Kanno, Y. and Takewaki, I. (2006c). Sequential semidefinite program for maximum robust-
ness design of structures under load uncertainty, Journal of Optimization Theory and
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Application 130, 2, pp. 265287, also Kyoto University, BGE Research Report 04-05,
2004, Available at http://www.archi.kyoto-u.ac.jp/bge/RR/.
Kanno, Y. and Takewaki, I. (2007a). Maximum robustness design of trusses via semidefi-
nite programming, in Y. Tsompanakis, N. D. Lagaros and M. Papadrakakis (eds.),
Structural Design Optimization Considering Uncertainties (Taylor & Francis, Lon-
don), pp. 471498.
Kanno, Y. and Takewaki, I. (2007b). Worst-case plastic limit analysis of trusses under
uncertain loads via mixed 0-1 programming, Journal of Mechanics of Materials and
Structures 2, pp. 245273.
Kapur, K. C. (1975). Optimization in design by reliability, AIIE Transactions 7, 2, pp.
185192.
Kaveh, A. (2004). Structural Mechanics: Graph and Matrix Methods, 3rd edn. (Research
Studies Press, Somerset, U.K.).
Kavlie, D., Graham, H. and Powell, G. H. (1971). Efficient reanalysis of modified struc-
tures, Journal of Structural Division, ASCE 97, ST1, pp. 377392.
Keafott, R. B. (1996). Interval computations: Introduction, uses, and resources, Euromath
Bulletin 2, 1, pp. 95112.
Khachaturian, N. and Haider, G. S. (1966). Probabilistic design of determinate structures,
in Proc. Specialty Conference (ASCE, New York), pp. 623641.
Kharitonov, V. (1997). Interval uncertainty structure: conservative and simple, in H. G.
Natke and Y. Ben-Haim (eds.), Uncertainty: Models and Measures (Akademie Ver-
lag, Berlin), pp. 231243.
Kharmanda, G., Olhoff, N. and El-Hami, A. (2004). Recent developments in reliability-
based design optimization, in Proc. WCCM VI in conjunction with APCOM04
(Tsinghua University Press & Springer-Verlag, Beijing).
Khodaparasi, H. H., Marques, S., Badcock, K. J. and Mottershead, J. E. (2009). Estima-
tion of flutter boundaries in the presence of structural uncertainty by probabilistic
and fuzzy methods, in N. Maia (ed.), Proc. International Conference on Structural
Engineering Dynamics (Ericeira, Portugal).
Kida, T. and Komatsu, K. (1993). Vibration control of space structures, Science of Ma-
chine 45, 12, pp. 12731279, (in Japanese).
Kildibekov, I. G. (1971). Nonlinear vibrations of circular cylindrical shells under axial
loading and acoustic pressure, in W. V. Bolotin and A. Chyras (eds.), Problemi
Nadezhnosti v Stroitel noi Mekhanike (Problems of Reliability in Structural Me-
chanics) (Vilnius, Lithuania), pp. 5863.
Kim, C. and Choi, K. K. (2008). Reliability-based design optimization using response
surface method with prediction interval estimation, Journal of Mechanical design
130, 12, Paper No. 121401.
Kim, S.-E. and Chen, W.-F. (1996). Practical advanced analysis for braced steel frame
design, Journal of Structural Engineering 122, 11, pp. 12661274.
Kim, T.-U. and Sin, H.-C. (2001). Optimal design of composite laminated plates with
the discreteness in ply angles and uncertainty in material properties considered,
Computers & Structures 79, pp. 25012509.
Kim, Y. V., Ovseyevich, A. I. and Reshetnyak, Y. N. (1993). Comparison of stochastic and
guaranteed approaches to the estimation of the state of dynamic systems, Journal
of Computer and System Sciences International 31, 6, pp. 5664.
Kirkpatrick, S., Gelatt, C. D. and Vecchi, M. P. (1983). Optimization by simulated an-
nealing, Science 220, 4598, pp. 671680.
Kirsch, U. (1989). Optimal topologies of truss structures, Applied Mechanics Reviews 42,
pp. 223239.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 363
Kirsch, U. (1994). Efficient sensitivity analysis for structural optimization, Computer Meth-
ods in Applied Mechanics and Engineering 117, pp. 143156.
Kirsch, U. (2000). Combined approximation: A general reanalysis approach for structural
optimization, Structural and Multidisciplinary Optimization 20, pp. 97106.
Kirsch, U. and Liu, S. (1995). Exact structural reanalysis by a first-order reduced basis
approach, Structural Optimization 10, pp. 153158.
Kirsch, U. and Liu, S. (1996). Structural reanalysis for general layout modification, in
Proc. 6th AIAA/NASA/ISSMO Symposium on Multidisciplinary Analysis and Op-
timization (AIAA), pp. 11421149.
Kirsch, U. and Rubinstein, M. F. (1972). Structural reanalysis by iteration, Computers &
Structures 2, pp. 497510.
Klir, G. J. (1989). Is there more to uncertainty than some probability theorists might have
us to believe? International Journal for General Systems 15, pp. 347378.
Klir, G. J. (1994). On the alleged superiority of probabilistic representation of uncertainty,
IEEE Transaction on Fuzzy Systems 2, pp. 2731.
Klir, G. J., St. Clair, U. H. and Yuan, B. (1997). Fuzzy Set Theory: Foundations and
Applications (Prentice Hall, Upper Saddle River, NJ), pp. 50,63,75.
Kogiso, N., Shao, S., Miki, M. and Murotsu, T. (1997). Reliability analysis and reliability-
based design of composite laminated plate subjected to buckling, in Proc. 38th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials
Conference and Exhibit, Vol. Part 2 (Kissimee, FL).
Kogiso, N., Shao, S. and Murotsu, T. (1996). Sensitivity analysis of buckling load of
the simply supported laminated plate, Transaction of JSME 62A, pp. 834841, (in
Japanese).
Koiter, W. T. (1945). On the Stability of Elastic Equilibrium, Ph. D. dissertation, Delft Uni-
versity of Technology, Delft, Netherlands, english Translation, NASA, TTF-10833,
1967.
Koltunov, M. A. (1976). Creep and Relaxation (Visshaya Shkola Publishing House,
Moscow), (in Russian).
Komkov, V. and Dannon, V. (1987). Design optimization for structural or mechanical
systems against the worst possible loads, Acta Applicandae Mathematicae 8, pp.
3764.
Korneev, A. B. and Svetlitskii, V. A. (1995). Analysis of nonstationary dynamic processes
in the presence of mixed information on random perturbation, Journal of Machinery
Manufacture and Reliability 3, pp. 612.
Kosko, B. (1990). Fuzziness vs. probability, International Journal of General System 17,
pp. 211240.
Kosko, B. (1992). Neural Networks and Fuzzy Systems: A Dynamical Systems Approach
to Machine Intelligence (Prentice Hall, Englewood Cliffs, NJ), p. 3.
Kosko, B. (1993). Fuzzy Thinking: The New Science of Fuzzy Logic (Hyperion, New York).
Kosko, B. (1994). The probability monopoly, IEEE Transaction on Fuzzy System 2, 1, pp.
2233.
Kosko, B. (1999). The Fuzzy Future: From Society and Science to Heaven in Chip (Har-
mony Books, New York).
Kocvara, M. and Stingl, M. (2006). The worst-case multiple load FMO problem revisited, in
M. P. Bendse, N. Olhoff and O. Sigmund (eds.), IUTAM Symposium on Topological
Design Optimization of Structures, Machines and Materials: Status and Perspectives
(Springer, Berlin), pp. 403411.
K
oyluoglu, H. U., C akmak, H. S. and Nielsen, S. R. K. (1995). Interval algebra to deal
with pattern loading and structural uncertainties, Journal of Engineering Mechanics
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 365
Bibliography 367
McTavish, D. J., Hughes, P. C., Soucy, Y. and Graham, W. B. (1992). Prediction and mea-
surement of modal damping factors for viscoelastic space structures, AIAA Journal
30, 5, pp. 13921399.
McWilliam, S. (2001). Anti-optimization of uncertain structures using interval analysis,
Computers & Structures 79, pp. 421430.
Melosh, R. J. and Luik, R. (1968). Multiple configuration analysis of structures, Journal
of Structural Division, ASCE 94, ST11, pp. 25812596.
Michalopoulos, C. D. and Riley, T. A. (1972). Response of discrete linear systems to forcing
functions with inequality constraints, AIAA Journal 10, 8, pp. 10161019.
Michell, A. G. M. (1904). The limits of economy in frame structures, Philosophical Maga-
zine Sect. 6 8, 47, pp. 589597.
Minciarelli, F., Gioffre, M., Grigoriu, M. and Simiu, E. (2001). Estimation of extreme
wind effects and wind load factors: influence of knowledge uncertainties, Probabilistic
Engineering Mechanics 16, 4, pp. 331340.
Minsky, M. (1987). The Society of Mind (Simon and Schuster, New York).
Mironovkii, L. S. and Slaev, V. A. (2002). Optimal test signals as a solution of the gener-
alized Bulgakov problem, Automation and Remote Control 63, 4, pp. 568577.
Modares, M., Mullen, R. L. and Muhanna, R. L. (2006). Natural frequencies of a structure
with bounded uncertainty, Journal of Engineering Mechanics 132, 12, pp. 1363
1371.
Moens, D. and Vandepitte, D. (2005). A survey of non-probabilistic uncertainty treatment
in finite element analysis, Computer Methods in Applied Mechanics and Engineering
194, pp. 15271555.
Moens, D. and Vandepitte, D. (2006). Recent advances in non-probabilistic approaches
for non-deterministic finite element analysis, Archives of Computational Methods in
Engineering 13, 3, pp. 389464.
M
oller, B. and Beer, M. (2004). Fuzzy Randomness: Uncertainty in Civil Engineering and
Computational Mechanics (Springer, Berlin).
M
oller, B. and Beer, M. (2008). Engineering optimization under uncertainty: Capabilities
of non-traditional models, Computer Methods in Applied Mechanics and Engineering
86, 1112, pp. 10241041.
M
oller, B., Grab, W. and Beer, M. (2000). Fuzzy structural analysis using -level opti-
mization, Computational Mechanics 26, pp. 547565.
Mongin, P. (2000). Does optimization imply rationality? Syntheses 124, 12, pp. 73111.
Moore, R. E. (1962). Interval arithmetic and automatic error analysis in digital computing,
Ph. D. dissertation, Stanford University.
Moore, R. E. (1966). Interval Analysis (Prentice-Hall, Englewood Cliffs, NJ).
Moore, R. E. (1979). Methods and Applications of Interval Analysis (SIAM, Philadelphia,
PA).
Moore, R. E. (2006). Introductory remarks on reliable engineering computing, Reliable
Computing 12, pp. 405408.
Moore, R. E., Kearfott, R. B. and Cloud, M. J. (eds.) (2009). Introduction to Interval
Analysis (SIAM, Philadelphia, PA).
Moreau, J. J. (1966). Quadratic programming in mechanics: One-sided constraints, SIAM
Journal of Control 4, pp. 153158.
Moreau, J. J. (1976). Application of convex analysis to the treatment of elastoplastic
systems, in P. Germain and B. Nayroles (eds.), Applications of Methods of Functional
Analysis to Problems in Mechanics (Springer, Berlin).
Moreau, J. J. (1979). Application of convex analysis to some problems of dry friction, in
H. Zorski (ed.), Trends in Applications of Pure Mathematics to Mechanics, Vol. 2
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
(Pitman, London).
Morimoto, M., Kaifu, N., Takizawa, Y., Aoki, K. and Sakakibara, O. (1982). Homology de-
sign of large antenna, (Mitsubishi Electric Corporation), pp. 495502, (in Japanese).
Morse, P. M. and Feshbach, H. (1953). Methods of Theoretical Physics (McGraw-Hill, New
York), pp. 10521060.
Moses, F. and Kinser, D. E. (1967). Optimum structural design and failure probability
constraints, AIAA Journal 6, 6, pp. 11521158.
Motro, R. (1992). Tensegrity systems: the state of the art, International Journal of Space
Structures 7, 2, pp. 7583.
Motro, R. (2003). Tensegrity Structural Systems for the Future (Butterworth-Heinemann,
Oxford, U.K.).
Mourelatos, Z. P. and Zhou, J. (2005). Reliability estimation and design with unsufficient
data based on possibility theory, AIAA Journal 43, 8, pp. 16961705.
Moustafa, A. (2009). Critical earthquake load inputs for multi-degree-of-freedom inelastic
structures, Journal of Sound and Vibration 325, 3, pp. 532544.
Muhanna, R. L., Kreinovich, V., S olin, P., Chessa, J., Araiza, R. and Xiang, G. (2006).
Interval finite element methods: New directions, in REC 2006: Reliable Engineering
Conference (Georgia Institute of Technology, Savannah, GA).
Muhanna, R. L. and Mullen, R. L. (2001). Uncertainty in mechanics problems interval
analysis approach, Journal of Engineering Mechanics 127, pp. 557566.
Muhanna, R. L. and Mullen, R. L. (2002). Short review for the latest development
of interval FEM, Available at: http://www.gtrep.gatech.edu/rmuhanna/IFEM-
review.pdf.
Muhanna, R. L. and Mullen, R. L. (2005). Interval methods for reliable computing, in
E. Nikolaidis, D. Ghiocel and S. Singhal (eds.), Engineering Design Reliability Hand-
book, pp. 12.112.24.
Muhanna, R. L., Mullen, R. L. and Zhang, H. (2005). Penalty-based solution for the in-
terval finite-element methods, Journal of Engineering Mechanics 131, 10, pp. 1102
1111.
Muhanna, R. L., Zhang, H. and Mullen, R. L. (2007). Interval finite elements as a basis
for generalized models of uncertainty in engineering mechanics, Reliable Computing
13, 2, pp. 173194.
Mullen, R. L. and Muhanna, R. L. (1999). Bounds of structural response for all possible
loading combinations, Journal of Structural Division, ASCE 125, 1, pp. 98126.
Mullen, R. L. and Muhanna, R. L. (2002). Efficient interval methods for finite element
solution, in J. N. Almhana and V. C. Bhavsar (eds.), High Performance Computing
Systems and Applications (IEEE Press, New York), pp. 161168.
Mungan, I. (2001). Structural engineering and structures from antiquity to the present,
in Proc. IASS Symposium 2001 (International Association for Shell and Spatial
Structures, Nagoya, Japan), pp. 13.
Murota, K. and Ikeda, K. (1991). Critical imperfection of symmetric structures, SIAM
Journal on Applied Mathematics 51, 5, pp. 12221254.
Murotsu, Y., Shao, S. and Watanabe, A. (1994). An application to reliability-based opti-
mization of redundant structures, Structural Safety 16, pp. 133143.
Myers, R. H. and Montgomery, D. C. (1995). Response Surface Methodology: Process and
Product Optimization using Design Experiments (Wiley, New York).
Naishul, A. B. and Svetlitskii, V. A. (1956). Determination of region of possible solutions
of the system of linear differential equations, Journal of Applied Mathematics and
Mechanics 20, 1, pp. 144147, (in Russian).
Nakagiri, S. (1995). A note on finite element interval analysis (Part 1: Linear stiffness
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 369
aid of electronic modelling, Automatics and Telemechanics 23, 9, pp. 11301139, (in
Russian).
Nguyen, T. B. (1965). On accumulation of disturbances, Izvestiya Akademii Nauk SSSR
Mekhanika (Communications of the USSR Academy of Sciences, Mechanics) 4, pp.
97100.
Nickel, K. (1977). Interval-analysis, in D. A. H. Jacobs (ed.), The State of the Art in
Numerical Analysis (Academic Press, London).
Nikolaidis, E., Chen, S., Cudney, H. and Haftka, R. T. (2004). Comparison of probability
and possibility for design against catastrophic failure under uncertainty, Journal of
Mechanical Design 126, pp. 386394.
Nikolaidis, E., Ghiocel, D. M. and Singhal, S. (eds.) (2005). Engineering Design Reliability
Handbook (CRC Press, Boca Raton, FL).
Nikolaidis, E. and Haftka, R. T. (2001). Theories of uncertainty for risk assessment when
data is scarce, International Journal for Advanced Manufacturing Systems 4, 1, pp.
4956.
Niordson, F. I. (2001). Early numerical computations in engineering, Applied Mechanics
Reviews 54, 6, pp. R17R19.
Nishihara, O., Asami, T. and Kumamoto, H. (1999a). Design optimization of dynamic
vibration absorber for minimization of maximum amplitude magnification factor:
Consideration of primary system damping by numerical exact solution, Dynamics &
Design Conference 1999, B, pp. 365368, (in Japanese).
Nishihara, O., Asami, T. and Watanabe, S. (1999b). Exact expressions of a dynamic
vibration absorber for minimization of maximum amplitude response: Case 2, hys-
teretic damping absorber, Dynamics & Design Conference 1999, B, pp. 361364,
(in Japanese).
Norton, R. L. (2000). Machine Design (Prentice-Hall, Upper Saddle River, NJ).
Nowak, A. S. and Szerszen, M. M. (eds.) (1998). Reliability and Optimization of Structural
Systems (University of Michigan, Ann Arbor, MI).
Nowak, A. S. and Szerszen, M. M. (eds.) (2000). Reliability and Optimization of Structural
Systems (University of Michigan, Ann Arbor, MI).
Oberguggenberger, M. (2004). Introductory remarks: Mathematical models of uncertainty,
Zeitschrift f
ur angewandte Mathematik und Mechanik (ZAMM) 84, pp. 661665.
Oberguggenberger, M. and Russo, F. (2001). Fuzzy, probabilistic and stochastic modelling
of an elastically bedded beam, in G. de Cosman, T. L. Fine and T. Seidenfell (eds.),
Proc. 2nd International Symposium on Imprecise Probabilities and Their Applica-
tions (Shaker Publishing BV, Cornell University, Ithaca, NY), pp. 293300.
Oden, J. T., Babuska, I., Nobile, F., Feng, Y. and Tempone, R. (2005). Theory and
methodology for estimation and control of errors to modeling, approximation, and
uncertainty, Computer Methods in Applied Mechanics and Engineering 194, pp.
195204.
Oden, J. T., Belytschko, T., Babuska, I. and Hughes, T. Y. R. (2003). Research direc-
tions in computational mechanics, Computer Methods in Applied Mechanics and
Engineering 192, pp. 913922.
Ohsaki, M. (1995). Genetic algorithm for topology optimization of trusses, Computers &
Structures 57, 2, pp. 219225.
Ohsaki, M. (1997). Simultaneous optimization of topology and geometry of a regular plane
truss, Computers & Structures 66, 1, pp. 6977.
Ohsaki, M. (2000). Optimization of geometrically nonlinear symmetric systems with coin-
cident critical points, International Journal for Numerical Methods in Engineering
48, pp. 13451357.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 371
Ohsaki, M. (2001). Random search method based on exact reanalysis for topology opti-
mization of trusses with discrete cross-sectional areas, Computers & Structures 79,
6, pp. 673679.
Ohsaki, M. (2002a). Imperfection sensitivity of optimal symmetric braced frames against
buckling, International Journal of Non-Linear Mechanics 38, pp. 11031117.
Ohsaki, M. (2002b). Maximum loads of imperfect systems corresponding to stable bifur-
cation, International Journal of Solids and Structures 39, pp. 927941.
Ohsaki, M. (2002c). Structural optimization for specified nonlinear buckling load factor,
Japan Journal of Industrial and Applied Mathematics 19, 2, pp. 163179.
Ohsaki, M. (2003a). Optimum design of flexible structures under constraints on strain
energy and asymptotic stability, Computer Methods in Applied Mechanics and En-
gineering 192, pp. 44874496.
Ohsaki, M. (2003b). Sensitivity analysis of an optimized bar-spring model with hill-top
branching, Archive of Applied Mechanics 73, pp. 241251.
Ohsaki, M. (2005). Design sensitivity analysis and optimization for nonlinear buckling
of finite-dimensional elastic conservative structures, Computer Methods in Applied
Mechanics and Engineering 194, pp. 33313358.
Ohsaki, M. and Arora, J. S. (1993). A direct application of higher order parametric pro-
gramming techniques to structural optimization, International Journal for Numeri-
cal Methods in Engineering 36, pp. 26832702.
Ohsaki, M., Fujisawa, K., Katoh, N. and Kanno, Y. (1999). Semi-definite programming
for topology optimization of trusses under multiple eigenvalue constraints, Computer
Methods in Applied Mechanics and Engineering 180, pp. 203217.
Ohsaki, M. and Ikeda, K. (2006). Imperfection sensitivity analysis of hill-top branching
with many symmetric bifurcation points, International Journal of Solids and Struc-
tures 43, 16, pp. 47044719.
Ohsaki, M. and Ikeda, K. (2007a). Stability and Optimization of Structures Generalized
Sensitivity Analysis, Mechanical Engineering Series (Springer, New York).
Ohsaki, M. and Ikeda, K. (2007b). Worst imperfection of an arch-type truss with multiple
member buckling at limit point, Journal of Structural and Constructional Engineer-
ing, Architectural Institute of Japan , 611, pp. 7177, (in Japanese).
Ohsaki, M. and Ikeda, K. (2009). Imperfection sensitivity of degenerate hilltop branching
points, International Journal of Non-linear Mechanics 44, pp. 324336.
Ohsaki, M. and Kanno, Y. (2003). Form-finding of cable domes with specified stresses by
using nonlinear programming, in Proc. IASS-APCS Symposium 2003 (International
Association for Shell and Spatial Structures, Taipei).
Ohsaki, M. and Kanno, Y. (2007). Semidefinite programming for engineering applications,
in J. S. Arora (ed.), Optimization of Structural and Mechanical Systems (World
Scientific, Singapore), pp. 541567.
Ohsaki, M. and Katoh, N. (2005). Topology optimization of trusses with stress and local
constraints on nodal stability and member intersection,, Structural and Multidisci-
plinary Optimization 29, pp. 190197.
Ohsaki, M., Kinoshita, T. and Pan, P. (2007). Multiobjective heuristic approaches to
design of frames with standard sections, Earthquake Engineering and Structural Dy-
namics 36, 11, pp. 14811495.
Ohsaki, M., Nakamura, T. and Isshiki, Y. (1998a). Shape-size optimization of plane trusses
with designers preference, Journal of Structural Engineering 124, 11, pp. 13231330.
Ohsaki, M. and Nishiwaki, S. (2005). Shape design of pin-jointed multistable compliant
mechanism using snapthrough behavior, Structural and Multidisciplinary Optimiza-
tion 30, pp. 327334.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 373
theories in structural design, Mechanical Systems and Signal Processing 15, 3, pp.
499511.
Pantelides, C. P. and Tzan, S.-R. (1996). Convex models for seismic design of structures
1: Analysis, Earthquake Engineering and Structural Dynamics 25, 9, pp. 927944.
Papadopoulos, V., Deodatis, G. and Papadrakakis, M. (2005). Flexibility-based upper
bound on the response variability of simple beams, Computer Methods in Applied
Mechanics and Engineering 194, 1216, pp. 13851404.
Papoulis, A. (1970). Maximum response with input energy constraints and matched filter
principle, IEEE Transaction on Circuit Theory CT-17, 2, pp. 175182.
Parimi, S. R. and Cohn, M. Z. (1975). Optimal criteria in probabilistic structural design,
in Z. Mr oz and A. Sawczuk (eds.), Proc. IUTAM Symposium on Optimization in
Structural Design (Springer Verlag, Berlin).
Parimi, S. R. and Cohn, M. Z. (1978a). Optimal solutions in probabilistic structural design:
Part I, theory, 2, 1, pp. 4772.
Parimi, S. R. and Cohn, M. Z. (1978b). Optimal solutions in probabilistic structural design:
Part II, applications, 2, 1, pp. 7392.
Park, J. and Sandberg, I. W. (1991). Universal approximation using radial-basis-function
networks, Neural Computation 3, pp. 246257.
Parkinson, C. N. (1959). Parkinsons Law on the Pursuit of Progress (John Murray, Lon-
don).
Parnes, R. and Beltzer, A. I. (1986). A higher-order boundary perturbation method for
asymmetric dynamic problems in solids I. general formulation, International Jour-
nal of Solids and Structures 22, 11, pp. 11771187.
Peek, R. and Triantafyllidis, N. (1992). Worst shapes of imperfections for space trusses with
many simultaneously buckling modes, International Journal of Solids and Structures
29, 19, pp. 23852402.
Peterson, R. E. (1974). Stress Concentration Design Factors, Charts and Relations Useful
on Making Strength Calculations for Machine Parts and Structural Elements (Wiley,
New York).
Philappacopoulos, A. J. (1980). Critical excitation for linear and non-linear structural
systems, Ph. D. dissertation, Polytechnic Institute of New York.
Pierson, B. L. and Genalo, L. J. (1977). Minimum weight design of a rectangular panel
subjected to a flutter speed constraint, Computer Methods in Applied Mechanics and
Engineering 10, pp. 4562.
Pietrzak, J. (1996). An alternative approach to optimization of structures prone to insta-
bility, Structural Optimization 11, pp. 8894.
Plaut, R. H. (1970). On minimizing the response of structures to dynamic loading, Journal
of Applied Mathematics and Physics 21, pp. 10041006.
Plaut, R. H. (1971). Structural optimization of a panel flutter problem, AIAA Journal 9,
1, pp. 182184.
Pochtman, Y. M. and Khariton, L. E. (1976). Optimal design of structures with reliabil-
ity taken into account, in Stroitelnaya Mekhanikai Raschet Sooruzhenii (Structural
Mechanics and Structural Analysis), No. 6, pp. 815, (in Russian).
Popova, E. D., Datcheva, M., Iankov, R. and Schanz, T. (2003). Mechanical models with
interval parameters, in K. Guerlebeck, L. Hempel and C. Koenke (eds.), Proc. 16th
International Conference on the Applications of Computer Science and Mathematics
in Architecture and Civil Engineering (Bauhaus University, Weimar, Germany).
Popova, E. D., Datcheva, M., Iankov, R. and Schanz, T. (2004). Sharp bounds for strains
and stresses of uncertain mechanical models, in I. Lirkov et al. (ed.), Lecture Notes
in Computer Science 2907 (Springer, Berlin), pp. 262269.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 375
Qiu, Z. P., Elishakoff, I. and Starnes, Jr., J. H. (1996d). The bound set of possible eigen-
values of structures with uncertain-but-non-random parameters, Chaos, Solitons &
Fractals 7, 11, pp. 18451857.
Qiu, Z. P. and Gu, Y. (1996). Extension of convex models and its improvement on the
approximate solution, Acta Mechanica Sinica 12, 4, pp. 349357.
Qiu, Z. P. and Hu, J. X. (2008). Two non-probabilistic set-theoretical models to predict
the transient vibration of cross-ply plates with uncertainty, Applied Mathematical
Modelling 32, 12, pp. 28722887.
Qiu, Z. P., Hu, J. X., Yang, J. L. and Lu, Q. S. (2008a). Exact bounds for the sensitivity
analysis of structures with uncertain-but-bounded parameters, Applied Mathematical
Modelling 32, 6, pp. 11431157.
Qiu, Z. P., Ma, L. H. and Wang, X. J. (2006a). Ellipsoidal-bound convex model for the
non-linear buckling of a column with uncertain initial imperfections, International
Journal of Non-linear Mechanics 41, 8, pp. 919925.
Qiu, Z. P., Ma, Y. and Wang, X. J. (2004c). Comparison between non-probabilistic interval
analysis method and probabilistic approach in static response problem of structures
with uncertain-but-bounded parameters, Communications in Numerical Methods in
Engineering 20, 4, pp. 279290.
Qiu, Z. P., Ma, Y. and Wang, X. J. (2004d). Interval analysis method for vibration of com-
posite laminate with uncertain parameters, Journal of Beijing University of Aero-
nautics and Astronautics 30, 7, pp. 682685.
Qiu, Z. P., M uller, P. C. and Frommer, A. (2001). Ellipsoidal set-theoretic approach for
stability of linear state-space models with interval uncertainty, Mathematics and
Computers in Simulation 57, 1-2, pp. 4559.
Qiu, Z. P., M uller, P. C. and Frommer, A. (2004e). The new non-probabilistic criterion of
failure for dynamical systems based on convex methods, Mathematical and Computer
Modelling 40, 12, pp. 201215.
Qiu, Z. P. and Wang, X. J. (2003). Comparison of dynamic response of structure with
uncertain-but-bounded non-probabilistic interval analysis, International Journal of
Solids and Structures 40, 20, pp. 54235439.
Qiu, Z. P. and Wang, X. J. (2005). Two non-probabilistic set-theoretical models for dy-
namic response and buckling failure measures of bars with unknown-but-bounded
initial imperfections, International Journal of Solids and Structures 42, 34, pp.
10391054.
Qiu, Z. P. and Wang, X. J. (2006). Interval analysis method and convex models for im-
pulsive response of structures with uncertain-but-bounded external loads, Acta Me-
chanica Sinica/Lixue Xuebao 22, 6, pp. 265276, (in Chinese).
Qiu, Z. P. and Wang, X. J. (2008). Set-Theoretic Convex Methods for Problems in Struc-
tural Mechanics with Uncertainties (Science Press, Beijing), (in Chinese).
Qiu, Z. P., Wang, X. J. and Chen, J. (2006b). Exact bounds for the static response set of
structures with uncertain-but-bounded parameters, International Journal of Solids
and Structures 43, pp. 65746593.
Qiu, Z. P., Yang, D. and Elishakoff, I. (2008b). Combination of structural reliability and
interval analysis, Acta Mechanica Sinica 24, pp. 6167.
Qiu, Z. P., Yang, D. and Elishakoff, I. (2008c). Probabilistic interval reliability of structural
systems, International Journal of Solids and Structures 45, pp. 28502860.
Rackwitz, R. (2001). Reliability analysis : A review and some perspectives, Structural
Safety 23, pp. 365395.
Rackwitz, R., Augusti, G. and Bori, A. (eds.) (1995). Reliability and Optimization of
Structural Systems (Chapman & Hall, London).
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 377
Royset, J. O., Der Kiureghian, A. and Polak, E. (2003). Algorithms for variability-based
optimal design, in S. Jendo and K. Doloinski (eds.), Reliability-Based Design and
Optimization, Lecture Notes 16, Institute of Fundamental Technological Research
(Warsaw, Poland), pp. 211240.
Rozvany, G. I. N. (1976). Optimal Design of Flexural Systems (Pergamon Press).
Rozvany, G. I. N. (1997). Aims, scope, basic concepts and methods of topology opti-
mization, in G. I. N. Rozvany (ed.), Topology Optimization in Structural Mechanics,
CISM Courses and Lectures, No. 374 (Springer, Vienna).
Ruiz, P. and Penzien, J. (1969). Probabilistic Study of the Behavior of Structures During
Earthquakes (Earthquake Engineering Research Center, University of California,
Berkeley).
Rump, S. M. (1992). On the solution of interval linear systems, Computing 47, pp. 337
353.
Russell, B. (1985). The Philosophy of Logical Atomism (Open Court, LaSalle, IL).
Rustem, B. and Howe, M. (2002). Algorithms for Worst-Case Design and Applications to
Risk Management (Princeton University Press, Princeton, NJ).
Rzhanitsyn, A. R. (1947). Determination of Safety Factor in Construction, Vol. 8 (Stroi-
telnaya Promyshlennost (Building Industry)), (in Russian).
Sadek, I. S., Sloss, J. M., Adali, S. and Bruch, Jr., J. C. (1993). Nonprobabilistic mod-
elling of dynamically loaded beams under uncertain excitations, Mathematical and
Computer Modelling 18, pp. 5967.
Sage, A. P. (1968). Optimum Systems Control (Prentice Hall, Englewood Cliffs, NJ).
Sakata, S., Ashida, F. and Zako, M. (2003). Structural optimization using Kriging ap-
proximation, Computer Methods in Applied Mechanics and Engineering 192, pp.
923939.
Salle, J. L. and Lefschetz, S. (1961). Stability by Liapunovs Direct Method (Academic
Press, New York).
Sanayei, M., Onipede, O. and Babu, S. R. (1992). Selection of noisy measurement locations
for error reduction in static parameter identification, AIAA Journal 30, 9, pp. 2299
2309.
Sangalli, A. (1998). The Importance of Being Fuzzy and Other Insights from the Border
between Math and Computers (Princeton University Press, Princeton, NJ).
Sarkar, A. and Manohar, C. S. (1998). Critical seismic vector random excitations for
multiplly supported structures, Journal of Sound and Vibration 212, 3, pp. 525
546.
Sarma, K. and Adeli, H. (2002). Life-cycle cost optimization of steel structures, Interna-
tional Journal for Numerical Methods in Engineering 55, pp. 14511462.
Sato, M. (2006). The Honda Myth: The Genius & His Wake (Vertical, Inc., New York).
Savin, G. N. (1961). Stress Concentration Around Holes (Pergamon, Oxford, U.K.).
Savoia, M. (2002). Structural reliability analysis through fuzzy number approach, with
application to stability, Computers & Structures 80, 12, pp. 10871102.
Saxena, V. (2003). Interval Finite Element Analysis for Load Pattern and Load Combina-
tion, M. Sci. Thesis, Georgia Institute of Technology, Atlanta, GA.
Sayles, R. S. (1982). The profile as a random process, in T. R. Thomas (ed.), Rough
Surfaces (Longman, London), pp. 91118.
Schek, H. J. (1974). The force density method for form finding and computation of general
networks, Computer Methods in Applied Mechanics and Engineering 3, pp. 115134.
Schiehlen, W. and Hu, B. (1997). Amplitude bounds of linear vibration responses,
Zeitschrift f
ur angewandte Mathematik und Mechanik (ZAMM) 76, pp. 453454.
Schmidt, R. (1985). Towards resurrecting the original Rayleigh method, Industrial Math-
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 379
Simiu, E. and Heckert, A. (1996). Extreme wind distribution fails: A peak over threshold
approach, Journal of Structural Engineering 122, 5, pp. 539547.
Skalna, I. (2003). Methods for solving systems of linear equations of structural mechanics
with interval parameters, Computer Assisted Mechanics and Engineering Science
10, pp. 281293.
Skalna, I. (2006). A method for outer solution of systems of linear equations depending
linearly on interval parameters, Reliable Computing 12, 2, pp. 107120.
Smallwood, D. O. (1972a). An extension of transient vibration technique using least favor-
able responses, Tech. Rep. Report SC-RR-72 0735, Sandia National Laboratories,
Albuquerque, NM.
Smallwood, D. O. (1972b). A transient vibration test technique using least favorable re-
sponses, Tech. Rep. Report SC-DR-72 0897, Sandia National Laboratories, Albu-
querque, NM.
Smallwood, D. O. (1973). A transient vibration test techniques and the method of least
favorable responses, in The Shock and Vibration Bulletin, Vol. 43 (Naval Research
Laboratory, Wasington D.C.), pp. 151164.
Smallwood, D. O. (1989). Personal communication to Isaac Elishakoff, March 23.
Sniedovich, M. (2006). Eureka! info-gap is worst case analysis (maximum) in disguise!
Working Paper No. MS-2-06, Department of Mathematics and Statistics, The Uni-
versity of Melbourne, Australia.
Sniedovich, M. (2007). The art and science of modeling decision-making under severe
uncertainty, Decision Making in Manufacturing and Services 1, 12, pp. 111136.
Sniedovich, M. (2008a). From Shakespeare to Wald: Modeling worst-case analysis in the
face of severe uncertainty, Decision Point 22, pp. 89.
Sniedovich, M. (2008b). The mighty Maximin, Working Paper No. MS-02-08, Depart-
ment of Mathematics and Statistics, The University of Melbourne, available at:
http://maximin.moshe-online.com.
Sniedovich, M. (2008c). Walds maximin model: a treasure in disguise! Journal of Risk
Finance 9, 3, pp. 287291.
Sniedovich, M. (2009). Personal communication, January 10.
Sobieszczanski-Sobieski, J., James, B. and Dovi, A. (1985). Structural optimization by
multilevel decomposition, AIAA Journal 23, pp. 17751782.
Soize, C. (2005). A comprehensive overview of a non-parametric probabilistic approach of
model uncertainties for predictive models in structural dynamics, Journal of Sound
and Vibration 288, 3, pp. 623652.
Sokal, A. and Bricmont, J. (1998). Intellectual Impostures (Profile Books, London), pub-
lished in the USA as Fashionable Nonsense.
Soong, T. T. (1988). Stochastic structural dynamics: Research versus practice, in S. V. Sri
Namachchivaya, H. H. Hilton and Y. K. Wen (eds.), Stochastic Structural Dynamics
(University of Illinois, Urbana-Champaign, IL), pp. 289297.
Sophie, Q. C. (2000). Comparing Probabilistic and Fuzzy Set Approaches for Design in the
Presence of Uncertainty, Ph. D. dissertation, Virginia Polytechnical Institute and
State University, Blackburg, VA.
Srensen, J. D. and Frangopol, D. M. (eds.) (2006). Advances in Reliability and Optimiza-
tion of Structural Systems (Taylor & Francis, London).
Spunt, L. (1971). Optimum Structural Design (Prentice-Hall, Englewood Cliffs, NJ).
Srinivasan, S., Corotis, R. and Ellingwood, B. (1992). Generation of critical stochastic
earthquakes, Earthquake Engineering and Structural Dynamics 21, pp. 275288.
Srinivasan, S., Ellingwood, B. and Corotis, R. (1991). Critical base excitations of structural
systems, Journal of Engineering Mechanics 117, 6, pp. 14031442.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Stolpe, M. and Svanberg, K. (2001). An alternative interpolation scheme for minimum com-
pliance topology optimization, Structural and Multidisciplinary Optimization 22, pp.
116124.
Stoyanov, J. (1987). Counterexamples in Probability (Wiley, Chichester, U.K.).
Stroud, W. J., Krishnamurthy, T. and Smith, S. A. (2001). Probabilistic and possi-
bilistic analyses of the strength of a bonded joint, in Proc. 42nd AIAA/ASME/
ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference and
Exhibit, Paper-AIAA-2001-1238 (Seattle, WA).
Strutt, J. W. (1877). Theory of Sound (Macmillan, London), (reprinted Dover Publica-
tions, New York, 1945, pp. 112113).
Sunaga, T. (1958). Theory of interval algebra and its application to numerical analysis, in
RAAG Memoirs, Vol. 2 (Gakujutu Bunken Fukyu-kai, Tokyo), pp. 2944, (Available
at http://www.cs.utep.edu/interval-comp/sunaga.pdf).
Suzuki, K. and Otsubo, H. (1995). The structural optimization under uncertainty, in Com-
putational Mechanics 95, Theory and Applications, Proc. International Conference
on Computational Engineering Science (Hawaii).
Svanberg, K. (1984). On local and global minima in structural optimization, in E. Atrek,
R. H. Gallagher, K. M. Ragsdel and O. C. Zienkiewicz (eds.), New Directions in
Structural design, Vol. 34 (John Wiley & Sons, Chichester, U.K.), pp. 327341.
Svanberg, K. (1994). On the convexity and concavity of compliances, Structural Optimiza-
tion 7, pp. 4246.
Svetlitskii, V. A. (1958). Determination of initial data of a perturbed motion, in G. A.
Nikolaev (ed.), Raschety na Prochnost v Mashinostroenii (Strength Analysis in Me-
chanical Engineering), pp. 234245, (in Russian).
Svetlitskii, V. A. (1967). Determination of maximally possible reactions in dynamic sys-
tems under excitations that are bounded by their modulus, Inzhenernyi Zhurnal
Mekhanika Tverdogo Tela (Engineering Journal of Mechanics of Solids) 3, pp. 161
165, (in Russian).
Svetlitskii, V. A. (2003). Statistical Dynamics and Reliability Theory for Mechanical Sys-
tems (Springer, Berlin).
Svetlitskii, V. A. and Prokhorov, V. N. (1970). Vibrations of a beam under the moving
random force that is bounded by its modulus, Izvestiya VUZOV, Mashinostroenie
(Proceedings of the Institution of Higher Learning, Mechanical Engineering) 6, pp.
2529, (in Russian).
Svetlitskii, V. A. and Stasenko, I. V. (1967). Beam on elastic foundation under the loading
that is bounded in modulus, Prikladnaya Mekhanika (Applied Mechanics) 3, 7, pp.
6569, (in Russian).
Switzky, H. (1964). Minimum weight design with structural reliability, in Proc. AIAA 5th
Annual Structural and Materials Conference, pp. 316322.
Takagi, J. and Ohsaki, M. (2004). Design of lateral braces for columns considering critical
imperfection of buckling, Int. J. Structural Stability and Dynamics 4, 1, pp. 6988.
Takewaki, I. (2002a). Robust building stiffness design for variable critical excitations, Jour-
nal of Structural Engineering 128, 12, pp. 15651574.
Takewaki, I. (2002b). Seismic critical excitation method for robust design: A review,
Journal of Structural Engineering 128, 5, pp. 665672.
Takewaki, I. (2006a). Critical Excitation Methods in Earthquake Engineering (Elsevier,
Amsterdam, Netherlands).
Takewaki, I. (2006b). Probabilistic critical excitation method for energy input rate, Journal
of Engineering Mechanics 132, 9, pp. 9901000.
Taleb, N. (2001). Fooled by Randomness: The Hidden Role of Chance in the Markets and
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 381
Tzan, S.-R. and Pantelides, C. P. (1996b). Convex models for seismic design of struc-
tures 2: Design of conventional and active structures, Earthquake Engineering and
Structural Dynamics 25, 9, pp. 945963.
Utkin, L. V. and Gurov, S. V. (1996). A general formal approach for fuzzy reliability
analysis in the possibility context, Fuzzy Sets and Systems 83, pp. 203213.
van Dyke, M. (1975). Perturbation Methods in Fluid Mechanics (Parabolic Press, Stanford,
CA), pp. 1314.
van Keulen, F., Damveld, H., Elishakoff, I. and Toporov, V. (2001). Bounded-but-unknown
uncertainties in design optimization, in Proc. AIAA Structural Dynamics Conference
(CD-ROM), AIAA Paper 2001-1242.
Vanderplaats, G. N. (1991). Efficient structural optimization techniques, in Computer
Application in Civil and Building Engineering, Proc. 4th International Conference
(Kozo System Inc, Tokyo), pp. 151158.
Vanderplaats, G. N. (1993). Thirty years of modern structural optimization, Advances in
Engineering Software 16, pp. 8188.
Vanderplaats, G. N. and Yoshida, N. (1985). Efficient calculation of optimum design sen-
sitivity, AIAA Journal 23, pp. 17981803.
Vanderplaats, G. N. (1999). Numerical Optimization Techniques for Engineering Design
(Vanderplaats Research & Development, Inc., Colorado Springs, CO).
Vanmarcke, E. H., Diaz-Padilla, J. and Roth, D. A. (1972). Reliability-based optimum
design of simple plastic frames, Tech. Rep. R72-46, Department of Mechanical En-
gineering, Massachusetts Institute of Technology.
Vassart, N. and Motro, R. (1999). Multiparametered formfinding method: application to
tensegrity systems, International Journal of Space Structures 14, 2, pp. 147154.
Venini, P. (1998). Robust control of uncertain structures, Computers & Structures 67, pp.
165174.
Venkataraman, S. and Haftka, R. T. (2004). Structural optimization complexity: What
has Moores law done for us? Structural and Multidisciplinary Optimization 28, pp.
375387.
Venkayya, V. B. (1993). Introduction: Historical perspective and future direction, in M. P.
Kamat (ed.), Structural Optimization: Status and Promise (AIAA Press, Washing-
ton D.C.), pp. 110.
Venter, G. and Haftka, R. T. (1996). A two species genetic algorithm for designing
composite laminates subjected to uncertainty, in Proc. 37th AIAA/ASME/ASCE/
AHS/ASC Structures, Structural Dynamics, and Materials Conference and Exhibit:
Part 3 (Salt Lake City, UT).
Venter, G. and Haftka, R. T. (1999). Two species genetic algorithm for design under
worst case condition, in Proc. 3rd World Congress of Structural and Multidisciplinary
Optimization (Buffalo, N. Y.), (Also Evolutionary Optimization (Internet Journal),
2(1), 119).
Venter, G., Haftka, R. T. and Starnes, Jr., J. H. (1998). Construction of response surface
approximation for design optimization, AIAA Journal 36, 12, pp. 22422249.
Vick, S. G. (2002). Degrees of Belief: Subjective Probability and Engineering Judgement
(ASCE Press, Reston, VA).
Vinot, P., Cogan, S. and Lallement, G. (2003). Approche non-probabiliste de fiabilite base
sur les mode les convexes, Mecanique & Industries 4, 1, pp. 4550, (in French).
Vitali, F., Blanc, G., Larque, P., Duplay, J., Morvan, G., Nakagiri, S. and Suzuki, K.
(1999). Finite element interval analysis of external loads identified by displacement
input with uncertainty, Computer Methods in Applied Mechanics and Engineering
168, 1, pp. 6372.
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 383
Wolfram, S. (1996). The Mathematica Book, 3rd edn. (Cambridge University Press, New
York).
Wolkowicz, R. and Vandenberghe, L. (eds.) (2000). Handbook of Semidefinite Programming
Theory, Algorithms, and Applications (Kluwer, Dordrecht, Netherlands).
Wood, K. L., Antonnson, E. K. and Beck, J. L. (1990). Representing imprecision on engi-
neering design comparing fuzzy and probabilistic calculus, Research in Engineering
Design 1, 34, pp. 187203.
Woodbury, M. A. (1950). Inverting modified matrices, Memorandom Report 42, Statistical
Research Group, Princeton, NJ.
Wu, T. S. and Boley, B. A. (1965). Bounds in melting problems with arbitrary rates of liq-
uid removal, Technical Report 24, Department of Civil Engineering and Engineering
Mechanics, Columbia University, New York.
Wu, Z. and Soong, T. T. (1996). Design spectra for actively controlled structures based
on convex models, Engineering Structures 18, 5, pp. 341350.
Xu, C. (1989). Fuzzy optimization of structures by the two-phase method, Computers &
Structures 31, 4, pp. 575580.
Yamaguchi, T., Kogiso, N. and Yamakawa, H. (2007). Optimal interplanetary trajectories
for impulsive deflection of potentially hazardous asteroids under velocity increment
uncertainties, Transactions of the Japanese Aerospace Sciences 55, pp. 432438.
Yang, J. N. (1994). Application of reliability methods to fatigue, quality assurance and
maintenance, in G. I. Schueller, M. Shinozuka and J. T. P. Yao (eds.), Structural
Safety and Reliability (Balkama, Rotterdam, Netherlands), pp. 320.
Yao, J. T. P. (1985). Safety and Reliability of Existing Structures (Pitman, Boston, MA).
Yau, W. (1975). Optimal design of simply supported beams for minimum upper bound of
dynamic response, Journal of Applied Mechanics 41, 1, pp. 301302.
Yeh, Y. C. and Hsu, D. S. (1990). Structural optimization with fuzzy parameters, Com-
puters & Structures 3, 4, pp. 575580.
Yoshikawa, N. (2002). Convex approach to estimate the worst excitation, in H. A. Mang,
F. G. Rammerstorfer and J. Eberhardsteiner (eds.), Proc. 5th World Congress on
Computational Mechanics (Vienna).
Yoshikawa, N. (2003). Structural integrity analysis via convex model of uncertainty for
worst-case ground motion caused by earthquakes, Journal of Society of Material
Science of Japan 52, 1, pp. 1015.
Yoshikawa, N., Elishakoff, I. and Nakagiri, S. (1998a). Worst case estimation of homology
design by convex analysis, Computers & Structures 67, pp. 191196.
Yoshikawa, N. and Fr yba, L. (1997). Interval analysis of dynamic railway track on random
foundation, in Y. Naprstek and J. Minsther (eds.), Dynamics and Vibrations, Proc.
Engineering Mechanics, Vol. 2, pp. 227232.
Yoshikawa, N. and Nakagiri, S. (1990). Homology design of frame structure by finite el-
ement method, in Theoretical and Applied Mechanics, Proc. 42nd Japan National
Congress for Applied Mechanics, Vol. 42 (University of Tokyo Press), pp. 4351.
Yoshikawa, N., Nakagiri, S. and Kuwazuru, O. (1998b). Structural synthesis for worst case
mitigation based on convex set of uncertainty, in N. Shiraishi, M. Shinozuka and
W. K. Wen (eds.), Structural Safety and Reliability (Balkema, Rotterdam, Nether-
lands), pp. 693696.
Youn, B. D., Choi, K. K. and Park, Y. H. (2003). Hybrid analysis method for reliability-
based design optimization, Journal of Mechanical Design 125, 2, pp. 221232.
Young, R. C. (1931). The algebra of many-valued quantities, Mathematische Annalen 104,
pp. 260290, (Available at http://www.cs.utep.edu/interval-comp/young.pdf).
Youssef, N. A. N. and Popplewell, N. (1978). A theory of the greatest maximum response
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Bibliography 385
Index
387
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
Index 389
Author Index
391
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization
L
u, Z-Z. 225 Mitra, S. K. 301
Luenberger, D. G. 18, 23, 248 Modares, M. 145, 208
Luik, R. 69 Moens, D. 45, 212
Molfino, P. 85
Ma, J. 208 Molinari, G. 85
Ma, L. H. xii, 61, 89, 144 M
oller, B. xiii, 43, 45, 46
Ma, X. P. 14, 189, 339 Mongin, P. 339
Ma, Y. 86, 89, 209 Montgomery, D. C. 30, 31, 279
Maglaras, G. 45 Mooij, E. 325
Majumder, L. 89 Moore, R. E. 4749, 52, 53, 56, 87
Makode, P. V. 70 Moreau, J. J. 61
Malozemov, V. N. 336 Morimoto, M. 299
Manasyan, A. A. 336 Morrison, W. J. 70, 72
Mang, H. A. 284 Morse, P. M. 92
Manohar, C. S. 112, 189, 190 Morvan, G. 225
Maplesoft 125, 139 Moses, F. 39
Mariani, M. 85, 325 Motro, R. 102, 109, 316
Marler, T. 30 Mottershead, J. E. 208
Marques, S. 208 Mourelatos, Z. P. 45
Marti, K. 40 Moustafa, A. 189
Martin, J. B. 336 Muhanna, R. L. 5254, 85, 89, 145, 208,
Masur, E. F. 114 211213, 215, 216, 225
Matheron, G. xiii, 46 Mulkay, E. 225
Matsumoto, M. 225 Mullen, R. L. 5254, 85, 89, 145, 208,
Maturana, S. 45 211213, 215, 216, 225
Maute, K. 40, 190, 330, 331 M
uller, P. C. 49, 61, 89, 209
Mayer, G. S. 54 Mungan, I. vii
Mayer, M. 1 Murota, K. 113, 134
Maymon, G. 227 Murotsu, T. 144, 274
McNeill, D. 46 Murotsu, Y. 340
McTavish, D. J. 153, 158 Murray, W. 289
McWilliam, S. 85, 89 Myers, R. H. 30, 31, 279
Melosh, R. J. 69
Meng, G. A. 14, 189, 339 Naishul, A. B. 150
Micaletti, R. 112, 190 Nakagiri, S. 61, 208, 225, 299
Michalopoulos, C. D. 149, 150 Nakamura, T. 27, 36
Michell, A. G. M. 18 Nakayama, H. 279
Miki, M. 274 Nassef, A. O. 114, 274, 281
Minciarelli, F. 14 Nastran 314
Mineau, D. 190 Natke, H. G. 208, 325
Minsky, M. 338 Nayak, A. 272
Mironovkii, L. S. 150 Neittaanm aki, P. 85
January 28, 2010 16:28 World Scientific Book - 9.75in x 6.5in optimization