Escolar Documentos
Profissional Documentos
Cultura Documentos
Kirsch
Structural
Optimization
Fundamentals and Applications
Springer-Verlag
Berlin Heidelberg NewYork
London Paris Tokyo
Hong Kong Barcelona Budapest
ISBN-13:978-3-540-55919-1
e-ISBN-13:978-3-642-84845-2
DOl: 10.1007/978-3-642-84845-2
This work is subject to copyright. All rights are reserved, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data
banks. Duplication of this publication or parts thereof is permitted only under the provisions of
the German Copyright Law of September 9, 1965, in its current version, and permission for use
must always be obtained from Springer-Verlag. Violations are liable for prosecution under the
German Copyright Law.
Springer-Verlag Berlin Heidelberg 1993
The use of general descriptive names, registered names, trademarks, etc. in this publication does
not imply, even in the absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general use.
Typesetting: Camera ready by authors
6113020-543210 - Printed on acid-free paper.
To my wife, Ira
Uri Kirsch holds the Sigmund Sommer Chair in Structural Engineering at the
Technion - Israel Institute of Technology. He brings to this book a background of
over twenty years' experience in research and development related to structural
optimization. He has been involved as a consultant in various engineering design
projects and has taught graduate courses on design optimization at several
universities.
Dr. Kirsch received his B.Sc., M.Sc. and D.Sc. degrees from the Technion. He
was Fulbright Research Scholar (1970 - 1971) at the University of California in
Los angeles and served as Visiting Professor at Case Western University, the
University of Waterloo, Carnegie-Mellon University, Virginia Polytechnic
Institute and State University, Heriot-watt University and the University of Essen.
He was Carnegie Fellow (1989) in the United Kingdom.
Dr. Kirsch has published over sixty papers on structural optimization. He is the
author of the book "Optimum Structural Design", published by McGraw-Hill,
New York (1981) and Maruzen, Tokio (1983). He is also a co-author of several
books.
Dr. Kirsch is a member of Editorial Boards of four international journals and has
been a member of various international committees on structural optimization. He
also acted as chairman of the Division of Structural Engineering and Vice Dean in
the Department of Civil Engineering at the Technion.
Preface
This book was developed while teaching a graduate course at several universities in
the United States. Europe and Israel. during the last two decades. The purpose of
the book is to introduce the fundamentals and applications of optimum structural
design. Much work has been done in this area recently and many studies have been
published. The book is an attempt to collect together selected topics of this
literature and to present them in a unified approach. It meets the need for an
introductory text covering the basic concepts of modem structural optimization. A
previous book by the author on this subject ("Optimum Structural Design".
published by McGraw-Hill New York in 1981 and by Maruzen Tokyo in 1983).
has been used extensively as a text in many universities throughout the world. The
present book reflects the rapid progress and recent developments in this area.
A major difficulty in studying structural optimization is that integration of
concepts used in several areas. such as structural analysis. numerical optimization
and engineering design. is necessary in order to solve a specific problem. To
facilitate the study of these topics. the book discusses in detail alternative problem
formulations. the fundamentals of different optimization methods and various
considerations related to structural design. The advantages and the limitations of the
presented approaches are illustrated by numerous examples.
Most of the material in the book is general and covers a wide range of
applications. However. the presentation is concentrated on well established and
developed areas of structural optimization. The text is intended to both the student
and the practising structural engineer. Previous knowledge of optimization methods
is not required; the reader is expected to be familiar with basic concepts of matrix
structural analysis and structural design. but the necessary material on structural
analysis is included in the book.
Chapter 1 deals with the problem statement. Some typical characteristics of
structural optimization are discussed and the role of automated numerical
optimization in the overall design process is outlined. The background material of
structural analysis needed in the rest of the book is presented. terminology used
throughout the text is defined. and the general problem is mathematically
formulated. Various problem formulations are introduced and some typical
formulations are discussed in detail and compared.
Chapter 2 presents general optimization methods used in the solution of various
structural optimization problems. Some optimization concepts are flfSt introduced.
and unconstrained minimization and constrained minimization methods are then
Preface
Contents
Problem Statement
1.1
1.1.1
1.1.2
1.1.3
1.1.4
1.2
1.2.1
1.2.2
1.3
1.3.1
1.3.2
1.3.3
1.3.4
1.4
1.4.1
1.4.2
Inttoduction
Automated Structural Optimization
Structural Optimization Methods
Historical Perspective
Scope of Text
Analysis Models
Elastic Analysis
Plastic Analysis
General Formulation
Design Variables
Constraints
Objective Function
Mathematical Formulation
Typical Problem Formulations
Displacement Method Formulations
Force Method Formulations
Exercises
1
1
3
5
8
9
9
20
25
25
27
30
31
44
44
45
52
Optimization Methods
57
2.1
2.1.1
2.1.2
2.2
2.2.1
2.2.2
2.3
2.3.1
2.3.2
2.3.3
2.3.4
2.4
2.4.1
2.4.2
2.4.3
Optimization Concepts
Unconstrained Minimum .
Constrained Minimum
Unconstrained Minimization .
Minimization Along a Line
Minimization of Functions of Several Variables
Constrained Minimization: Linear Programming
Inttoduction
Problem Formulation
Method of Solution
Further Considerations
Constrained Minimization: Nonlinear Programming
Sequential Unconstrained Minimization
The Method of Feasible Directions
Other Methods
Exercises
58
58
59
66
66
72
80
80
80
85
94
97
98
110
116
120
XII
Contents
Approximation Concepts
125
3.1
3.1.1
3.1.2
3.1.3
3.2
3.2.1
3.2.2
3.2.3
3.2.4
3.2.5
General Approximations .
Design Sensitivity Analysis .
Intermediate Variables
Sequential Approximations
Approximate Behavior Models
Basic Displacement Approximations
Combined Displacement Approximations
Homogeneous Functions .
Displacement Approximations along a Line
Approximate Force Models
Exercises
126
126
133
139
145
146
150
160
164
169
175
Design Procedures
179
4.1
4.1.1
4.1.2
4.2
4.2.1
4.2.2
4.3
4.3.1
4.3.2
4.3.3
4.3.4
4.4
4.4.1
4.4.2
4.4.3
4.5
4.5.1
4.5.2
4.6
4.6.1
4.6.2
4.7
4.7.1
4.7.2
4.7.3
4.7.4
4.8
4.8.1
4.8.2
4.8.3
180
180
191
197
197
206
210
210
217
.220
223
.225
.225
233
241
.246
.246
250
.254
255
259
.262
.262
.264
.267
275
278
278
280
.280
.284
293
.299
1 Problem Statement
1.1 Introduction
1.1.1 Automated Structural Optimization
The motivation of optimization is to exploit the available limited resources in a
manner that maximizes utility. The object of optimal design is to achieve the best
feasible design according to a preselected measure of effectiveness. A growing
realization of scarcity of the raw materials resulted in a demand for light weight and
low cost structures. This demand emphasizes the need for weight and cost
optimization of structures.
The Design Process. The structural design process may be divided into the
following four stages [99]:
1 Problem Statement
Itemtive procedures for the four stages are often required before the final solution is
achieved. The portion of the structural design process that can be optimized
automatically has been considembly increased in recent years. The significant
progress in this field is a result of developments in structural analysis,
optimization methods and automated design procedures.
1.1 Introduction
1 Problem Statement
1.1 Introduction
1 Problem Statement
a. the classical literature dealing with the least weight layout of highly idealized
frameworks;
b. optimum design of structural components based on weight-strength analysis or
structural index methods; and
c. minimum weight optimum design of simple structural systems based on the
plastic collapse or limit analysis design philosophy.
1.1 Introduction
systems. Working within the elastic design philosophy, it was shown that the
minimum weight optimum design of elastic statically indeterminate structures
could be stated as a nonlinear mathematical programming problem in a design
variables space.
Mathematical programming (MP) techniques were shown to be an effective tool
for design of numerous civil, aeronautical, and space structures. This promising
tool was new and much development required to establish the methodology. Indeed,
these structural synthesis concepts were considered by many researchers to be a
revolutionary change in the traditional approach to design. In the 1960's, enough
computational experience had been documented to indicate that MP techniques
applied to structural design were limited to only a few dozens design variables.
Thus, despite the generality of MP, applications were limited to relatively small
structures.
In the late 1960's an alternative approach, called Optimality Criteria (OC), was
presented in analytical form by Prager and coworkers [108, 109] and in numerical
form by Veokayya and coworkers [141]. Although this approach was largely
intuitive, it was shown to be most effective as a design tool. Its principal
attraction was that the method was easily programmed for the computer, was
relatively independent of problem size, and usually provided a near-optimum design
with a few structural analyses. This last feature represented a remarkable
improvement over the number of analyses required for MP methods to reach a
solution. Consequently, much research effort was devoted to OC methods during
the early and mid 1970's.
MP was attractive due to its generality and rigorous theoretical basis. On the
other hand, OC had no clear theoretical basis and would, on occasion, lead to
nonoptimum designs. However, OC offered a solution for a variety of practical
design problems. The main obstacles to the implementation of efficient MP
methods prior to 1970 were associated with the large problem size (large numbers
of design variables and constraints) and the need to repeat the structural analysis
many times. Much effort has been devoted to solve these problems since the mid
1970's. The introduction of approximation concepts, via reduction of the problem
size and the construction of high quality explicit approximations for the
constraints, has led to the emergence of MP based structural optimization methods
that are computationally efficient
During the late 1970's and 1980's, development continued in both OC and MP
approaches to structural optimization. The dual MP formulation was interpreted as
a generalized OC method, and was presented as a basis for coalescing of the two
approaches. Approximation concepts have been used for efficient solution of the
optimization problem, and have been combined with the dual formulation to create
new tools.
In recent years, the range of applicability of structural optimization has been
widened and much progress has been made in various topics associated with this
area. Efficient techniques for derivative calculation have been developed, and
problems with complex analysis model and various types of constraints and
objective function have been investigated. Extensive research and development is
continually being done on such topics as shape and layout optimization,
decomposition of large scale problems, optimal control of structures and
application of genetic algorithms. The significant progress in these advanced topics
1 Problem Statement
emphasizes the need for a deeper insight and understanding of the fundamentals of
structural optimization.
There are clear indications that optimum structural design methodology is
coming of age. It has matured significantly and has grown in the last three decades
from an abstract concept to a practical tool which is currently serving the quest for
better structural designs. Although structural optimization has not yet achieved the
near universal acceptance level enjoyed by fmite element analysis methods, a ftrlO
knowledge and experience base exists for the further development of rather general
and efficient capabilities.
constraints are related to the common stress and displacement limitations under
service loads, and plastic collapse constraints under overload conditions, as well as
technological constraints on the design variables. Static linear structural response
and fmite (discrete parameter) optimization problem formulations are considered. A
single criterion objective function is assumed, representing the cost or the weight
of the structure, and numerical methods of structural optimization are used to solve
the design problem.
The remainder of this chapter deals with general analysis tools and various
formulations of optimal design problems. General optimization methods that can
be used to solve various optimal design problems are discussed in Chap. 2.
Approximation concepts, which are essential in optimal design of practical
structures, are presented in Chap. 3. Various problem formulations, optimization
methods and approximate analysis models are combined to introduce design
procedures in Chap. 4. The special problems associated with such integration are
discussed, emphasizing the physical aspects and the engineering considerations.
Finally, optimal design of the structural geometry and topology, that can greatly
improve the design, is demonstrated.
a. Methods for elastic analysis of framed structures, such as beams, frames and
trusses. The basic relations are first presented; the force method, the
displacement method and the virtual load method, which have widely been used
in structural optimization, are then briefly described.
b. Plastic analysis methods for framed structures. Only the static approach, which
is often used in optimal design formulations, is presented.
10
I Problem Statement
(1.1)
or
(1.3)
(1.4)
Q=CT
(1.5)
in which
Displacement method. Substituting (1.3), (1.4) and (1.5) into (1.1) yields
(1.6)
Denoting the system stiffness matrix by K, where
(1.7)
tThe following symbols have been used throughout this text: bold letters represent
matrices or column vectors; superscripts T represent transposed matrices or vectors;
subscripts d denote diagonal matrices.
11
Kr=R
(1.8)
(1.15)
Rearranging this equation and substituting
(1.16)
(1.17)
(1.18)
gives the force method compatibility equations
12
1 Problem Statement
(1.19)
F N= ~
in which F is the system flexibility matrix, N is the vector of redundant forces and
~ are displacements corresponding to redundants.
[..fia
..fia
01 -..fia
-1] {~:} =..fi {10}
..fia 0 A3
10
~
where E is the modulus of elasticity and ai is the the cross-sectional area of the
ith member. The compatibility relations (1.4) are
Choosing members 3 and 4 as redundants, then from (1.9), (1.10), (1.12) and
(1.13) we have
dimensions throughout this text are arbitrary, unless certain dimensions are
specified.
100
20
The inverse of CB is
C-1 =[...fi
0]
1
-1
...fi2...fi
2 ...fi
-+-+-
100 lit
E
llz
OJ
2 ...fi
-+al
a2
...fi
-+lit
llz
{~}=_ 20~100
2...fi
al
a2
a4
-+-+-
al
2
al
13
14
1 Problem Statement
forces.
(1.20)
(1.21)
the compatibility equations (1.20) become [see (1.19)]
F N=S
(1.22)
(1.23)
A =Ap+ANN
(1.24)
15
Sp DN and Dp are functions of both the geometry and members' cross section, and
~ elements of AN and Ap depend only on the geometry of the structure.
Example 1.2. To illustrate solution by the force method, consider the simple
continuous beam shown in Fig. 1.2a. The beam has a constant flexural rigidity EI,
and the object is to find the forces and the rotations at the supports. The structure
is statically indeterminate to the second degree, so that two redundant forces must
be determined. The chosen redundants, Nb N 2 , and the corresponding primary
structure (a cantilever beam) are shown in Fig. 1.2.
The required coefficients, computed in the primary structure, are
(a)
(b)
AN11
AN11
C~
at
ANII
AHll
(c)
Fig. 1.2. Continuous beam example: a. Loads and redundants. b. External loading
on the primary structure, c. Unit redundants on the primary structure.
16
1 Problem Statement
~ _{~PI}_~ {26}
p -
Ap
~p2
48EI
~o ={~}
97
ApI} { 2P }
={ Ap2 = -Pi/2
{a}
NI}=.!.. { 69}
{N2
56-64
~} Ptl {10} i 2
={ D2 =8EI 13 + 2EI
AI} {2P }
A= {A2 = -Pi/2 +
56
-64
56
3U
K=
stiffness matrix, whose elements Kij represent the force in the ith
in which
coordinate due to unit displacement in the jth coordinate (Kij are computed in the
restrained structure and i and j are coordinates corresponding to displacement
degrees of freedom); r = the vector of unknown displacements; RL = the vector of
forces corresponding to the unknown displacements in the restrained structure; Ro
=the vector of extemalloads corresponding to the unknown displacements (Ro =0
if there are no loads acting in the direction of degrees of freedom). Defining the
load vector R =Ro - RL then (1.25) becomes [see (1.8)]
Kr=R
(1.26)
17
(1.29)
Ro
AL
R=Pl {-3}
={_p~}
48
ALl} ={PIl}
={AL2
Pl/8
Ar =[~11
Ar2i
-43
[3 00]
Ar12] = 2EI
12 I
~22
.
3
EI
I
[5 I] {Ii} =
1 5
r2
-3}
PI {
48-43
Pl2 { 7}
{r2Ii} = 384EI
-53
18
1 Problem Statement
IP
Pi
~r_"""",==t==::;;;;--i~~'
...I~Cf1= =U!Ol. . l.QI. !;P: !:/: :f: 1::1=p:;;;;;!:.I1i'~jOJOOf:Il~':""'oI.,l
.I..,"Uy..I..i!lL.JI"i~~
~
::::::==' _
A2 (
Alt.~
0-
__
;:::>
i_ _
rl
'2
~_____I_.5_i_ _ _ _~_____i____~
_I
'1-
'1
(a)
RI.\
A 1.2 (
R,.2
~L_~lLP_...:....D..l~~%,j':J':J':J,:r':r,:rtj(j:~III'I' I'IIDII
:t;~:C,1
II
I:I I:ICII~I~
1:'I::1:11:'
tAI.\
I:'
(h)
Am
C~
............
t Arll
--- i
1\21
1\11
1)
1.0
K~'!.
KI2
Am
C~
t Ar12
'"~
--------
1.0
- ;Yr1 ----- ~
/"
fe)
Fig. 1.3. a. Continuous beam example, b. Loads on the restrained structure, c.Unit
displacements on the restrained structure.
A=
D= LTj(Aj)lgj(Xj )
19
(1.30)
Aj
A;9dl j
(1.31)
A;9
jth member due to a virtual load Qi= 1.0 applied in the ith direction; lj is the
member length; E is the modulus of elasticity; and glXj ) is a function of Xj'
representing the cross-sectional properties of the jth member (for example, crosssectional area in truss elements, or moment of inertia in beam elements). The
displacements expression (1.30) is used particularly in problems where only a
small number of displacements are to be considered.
It should be noted that (1.30) is based on the assumption that a single force
(such as axial force or bending moment) is sufficient to describe the response
behavior of each member. However, this fonnulation can be extended to the more
general case of multiple force members.
Assuming the common case where
(1.32)
then (1.30) becomes
D= L
TjlXj =T l/X
(1.33)
in which matrix T consists of the vectors T j and the vector 1/X consists of the
elements l/Xj. Writing (1.33) for the displacement degrees of freedom
r=T l/X
(1.34)
(1.35)
where matrix P is defined by P=ST. It has been noted that in truss structures the
elements of S are independent of the members' sizes. If the truss is statically
detenninate, the forces Aj and
the elements of matrix T depend only on the truss geometry. For indetenninate
structures, where the force distribution depends on the members' sizes, the elements
of matrix T are usually implicit functions of the cross-sectional dimensions.
20
1 Problem Statement
Example 1.4. Consider the four-bar truss shown in Fig. 1.1, with the
following forces obtained for the given cross-sectional areas Xj =1.0
AT = {8.28, 8.28,0, -8.28}
The forces A are implicit functions of the areas Xj and satisfy the conditions of
equilibrium and compatibility. The forces AQ, due to unit loads in the directions
of rl and r2' must satisfy only the conditions of equilibrium and may be calculated
for a statically equivalent system where some redundant forces will be arbitrarily
set to zero. Assuming, for example, A2 =0 for QI = 1.0 and A4 =0 for Q2 = 1.0,
the resulting forces are given by
Q T
(A ) =
100[4.14
E
4.14
1/ Xl]
[I/X
0 4.14] 1/ X2
4.1400
1/ X4
2.07 o
0.5 0.5]
[ 4.14
E
0
1
4.14 4.14 o
0"=- [
r=
0
2.07 o -2.07 VX3 100 -0.5 0.5
0
-1
0
-4.14 -4.14 0
o VX 4
-8.28
21
...
t:
'"
E
Yield stress
Curvature
Strain
(b)
(a)
While the exact calculation of the collapse load of a structure requires the
solution of a costly nonlinear system of equations, it is possible to obtain a
conservative estimate of that load by assuming an elastic-perfectly-plastic
behavior. The material is assumed to deform in the idealized manner shown in Fig.
1.4. The strain and stress in Fig. 1.4a are proportional to one another up to the
yield stress, at which the strain increases without any further increase in stress. In
members subjected to bending the idealized relation shown in Fig. I.4b, between
the bending moment and curvature at a section, is assumed. The curvature and
bending moment are assumed to be proportional to one another up to the fully
plastic moment Mpl. At the fully plastic moment a plastic hinge is formed, and
the curvature (rotation at the hinge) is increased without any increase in the
moment The rotations at the cross section before Mpl is reached are considered to
be relatively small and the equilibrium equations are referred to the undeformed
geometry of the structure. It is assumed that plastic hinges are concentrated at
critical sections with ductility being unlimited. In addition, the loads applied to the
structure are assumed to increase proportionally.
Existing methods for plastic analysis are based on either the kinematic approach or
the static approach [55]. The latter approach, which is often used in optimal design
formulations, is briefly described herein.
Static Approach. According to the static theorem of plastic analysis [55] , the
bending moment distribution at collapse is such that the corresponding load factor
is the largest statically admissible multiplier, i.e.,
A= max(AJ
(1.36)
22
1 Problem Statement
A ~ max
J
L C/cjMj=AR"
(1.37)
= t .... nE
(equilibrium equations)
(1.38)
j=l
1. ... .J
(yield conditions)
(1.39)
4Ml
max
+ 2M2
2M2 + 4M3
-Mpl~ M j ~ Mpl
AP
A?
I.
!
1
= APt
= 2APt
j= 1, 2. 3
2AP
2
2\:
.1.
!
3
Do
.1
(b)
Bending-moment
23
3 ::!
max A
-3
A-+ max
-Mpl
~( ~i A-~M2)
-Mpl
-Mpl
~(~i A-~M2)
M2
This formulation could be obtained also from (1.24). where all bending moments
are expressed in terms of the chosen redundant force M 2 = N. A graphical
presentation of this problem in the space of A and M2 is shown in Fig. 1.6. The
solution is
A=3Mpl
Pi
24
1 Problem Statement
AI /{2 + Az + A3 /{2 = AP
At /{2
-A3/{2 = 0
At ~Apl
Az ~2Apl
A3 ~1.5Apl
(AP - Az )/ {2 ~ Apl
Az ~2Apl
(AP-Az)/..fi ~1.5Apl
A = 3.414Apl l P
25
26
1 Problem Statement
27
Other examples for this class of variable include the location of supports in a
bridge, the length of spans in a continuous beam, and the height of a shell
structure. Although many practical structures have geometry which is selected
before optimization, geometrical variables can be treated by most optimization
methods. In general, the geometry of the structure is represented by continuous
variables.
Cross-Sectional Design Variables. Cross-sectional dimensions are the
simplest design variables. The cross-sectional area of a truss member, the moment
of inertia of a flexural member, or the thickness of a plate are some examples of
this class of design variable. In certain cases a single design variable is adequate to
describe the cross section, but a more detailed design with several design variables
for each cross section may be necessary. For example, if the axial buckling of
members is considered, the cross-sectional dimensions which define the area and the
moment of inertia can be taken as design variables. It is often useful to choose
quantities other than the obvious physical ones as design variables. In the above
example, instead of the cross-sectional dimensions, we may use the area and the
moment of inertia as variables. Such transformation of variables may simplify the
problem formulation and can also yield considerable advantage in the solution.
In practical design, cross-sectional variables may be restricted to some discrete
values. Such are the rolled steel members, which are produced in distinct sizes with
unevenly spaced cross-sectional properties. In such cases the design variable is
permitted to take on only one of a discrete set of available values. However, as
discrete variables increase the computational time, the cross-sectional design
variables are usually assumed to be continuous.
1.3.2 Constraints
Any set of values for the design variables represents a design of the structure.
Clearly, some designs are useful solutions to the optimization problem, but others
might be inadequate in terms of function, behavior, or other considerations. If a
design meets all the requirements placed on it, it will be called afeasible design.
The restrictions that must be satisfied in order to produce a feasible design are
called constraints. From a physical point of view we may identify two kinds of
constraints:
a. Constraints imposed on the design variables and which restrict their range for
reasons other than behavior considerations will be called technological
constraints or side constraints. These constraints, which are explicit in form,
may derive from various considerations such as functionality, fabrication, or
aesthetics. Thus, a technological constraint is a specified limitation (upper or
lower bound) on a design variable, or a relationship which fixes the relative
value of a group of design variables. Examples of such constraints include
minimum slope of a roof structure, minimum thickness of a plate, or
maximum height of a shell structure.
b. Constraints that derive from behavior requirements will be called behavior
constraints. Limitations on the maximum stresses, displacements, or buckling
28
1 Problem Statement
(1.40)
(1.41)
where nil is the number of equalities. In many cases equality constraints can be
used to eliminate variables from the optimization process. thereby reducing their
number. The constraints (1.41) may represent the analysis equations or various
design considerations such as a desired ratio between the width of a cross section
and its depth. Such a simple and explicit constraint can easily be used to reduce
the number of independent variables. However. in certain cases the elimination
procedure may be complex and time consuming and some equality constraints must
be considered.
The constraints (1.40) and (1.41) may be linear or nonlinear functions of the
design variables. These functions may be explicit or implicit in X and may be
evaluated by analytical or numerical techniques. However. except for special
classes of optimization problems. it is important that these functions be
continuous and have continuous first derivatives in X .
Design Space. We may view each design variable as one dimension in a design
space and any particular set of variables as a point in this space. In cases with two
variables the design space reduces to a plane. In the general case of n variables. we
have an n-dimensional hyperspace.
Considering only the inequality constraints (1.40). the set of values of the design
variables that satisfy the equation gj{X) = 0 forms a surface in the design space. It
is a surface in the sense that it cuts the space into two regions: one where gj> 0
and the other where gj < O. The design space and the constraint surfaces for a
typical truss (example 1.7) are shown in Fig. 1.9. A design which satisfies all the
constraints is a feasible design. and the set of all feasible designs form the feasible
region. t
tIn all figures. the convention will be to hatch the feasible region, i.e., the acceptable
side of the constraints.
29
00
II
II
"'_
Ov>
3.0
Intersection
point
2.0
/design
Feasible
0",
1.0
region
- - 70
- "'0
1.0
0.5
Fig. 1.9.
1.5
Points within the feasible region [i.e., where #X) < 0, j 1,... , ng] are called
unconstrained designs. Points on the surface [i.e., feasible designs for which at
least one gJ{X) = 0] are called constrained designs. The subspace where two or
more constraints gj<X) = 0 is called an intersection. In a two dimensional space,
two constraints intersect in a point. The jth constraint is said to be active in a
design point for which gj(X) 0 and passive if gj(X) < O. If gj(X) > 0 the
constraint is violated and the corresponding design is infeasible .
The equality constraints hj(X) = 0, j = 1, ... , nit , introduce couplings between
the variables and may be thought of as surfaces in the n-dimensional design space.
The feasible design points are required to be located in the intersection of these
surfaces. The number of equality constraints nit must not exceed the total number
of (dependent and independent) variables nT' In a case with nlt=nT the variables
may, in principle, be determined as solutions to the equations hJ{X) = 0 and there is
no optimization problem in the proper sense. When nit > nT the system of
equations is overdetermined and either there are some redundant equality constraints
or the formulation is inconsistent.
aL ~ a
aU
XL~X ~Xu
(displacement constraints)
(stress constraints)
(side constraints)
(1.42)
(1.43)
(1.44)
30
1 Problem Statement
in which L and U are superscripts denoting lower and upper bounds, respectively;
and D and a are vectors of displacements and stresses, respectively. Both the
displacements and the stresses are in general nonlinear and implicit functions of the
design variables, given by the analysis equations. The lower and upper bounds are
usually preassigned parameters. An exception is a lower bound on stresses which
might depend on the design variables if buckling strength is considered.
(1.45)
The selection of an objective function can be one of the most important decisions
in the whole optimal design process. The mathematical formulation of the
objective function may be a very difficult task as, for instance, when important
aesthetical values are influenced by the design variables. In general, the objective
function represents the most important single property of a design, but it may
represent also a weighted sum of a number of properties. Weight is the most
commonly used objective function due to the fact that it is readily quantified,
although most optimization methods are not limited to weight minimization. The
weight of the structure is often of critical importance, but the minimum weight is
not always the cheapest. Cost is of wider practical importance than weight, but it
is often difficult to obtain sufficient data for the construction of a real cost
function. A general cost function may include the cost of materials, fabrication,
transportation, etc. In addition to the cost involved in the design and construction,
other factors such as operating and maintenance costs, repair costs, insurance, etc.,
may be considered. In cases where a general objective function is considered, the
result might be a "flat" function which is not sensitive to variations in the design
variables and the optimization process, practically, will not improve the design. In
most practical applications the objective function is indeed flat near the optimum
(Fig. 1.10). Thus, a near optimal solution, rather then the theoretical optimum, is
often sufficient
Another approach is to consider both the initial cost of the structure and the
failure costs which depend upon the probabilities of failure. The assumption is
that the failure cost is given by the damage cost associated with a particular failure
multiplied by its probability of occurrence. It is, however, recognized that
answering the moral question of what constitutes an appropriate failure damage
31
= 1... n,
(1.47)
32
1 Problem Statement
-15
~ (J ~
20
(side constraints)
(a)
(stress constraints)
(b)
(d)
-0.5 0.5
I~
100
33
100
100
(e)
Substituting (e) into (d), we obtain the stresses expressed in terms of the design
variables
Only constraints which may affect the design must be considered. Since a1 and a2
will always be positive, and a 3 negative, some constraints can be deleted and we
may consider only the stress constraints
al-20~O
a2-20~O
-a3-15~O
Substituting (f) into (g) we fmd the following explicit stress constraints
(g)
34
1 Problem Statement
o
11
."
-.
-I-
3.0
2.0
1.0
Objective function
~~-~r---~4--contours
Fig. 1.12. Design space and objective function contours, three-bar truss.
(h)
(i)
The locus of all points satisfyingj{X) =constant forms a surface. For each value
of the constant there corresponds a different member of a family of surfaces. Figure
1.12 shows the family of constant volume (or weight) contours, called objective
function contours. Every design on a particular contour has the same volume. It
35
can be observed that the minimum value of f(X) in the feasible region occurs at
point A, which represents the optimal design
min Z= 263.9
Xl = 0.788
It can be noted that only the constraint cr 1 -20
cr1 =20.
j = 1,... , ng
(1.49)
hiX, Y)= 0
= 1,... , nIl
(1.50)
36
1 Problem Statement
(1.51)
where L is the matrix of linking constants giving the predetermined ratios between
variables X and XI' In (1.51) the variables X are taken as a linear combination of
XI' In many cases in which only simple design variable linking is used, the matrix
L takes on a special form, in which each row contains only one nonzero element.
The reduced-basis concept further reduces the number of independent design
variables by expressing the vector XI as a linear combination of s basis vectors b j
giving
L yjbj =by
s
XI =
(1.52)
j=l
X=Lby=ty
where t is the matrix of prelinked basis vectors and y is the vector of a reduced set
of design variables.
xG
X4 Xs
.d77T
x7
X 13
:::n:::
:L
:::b..-
Xg
x~
Fig. 1.13
X 12 XII
X9
X3
Continuous beam.
XIO
37
Example 1.S. To illustrate variable linking and basis reduction, consider the
continuous beam shown in Fig. 1.13, with a given span I, six geometrical
variables (Xl> X 2 X 3 Xg X 9 XlO) representing the elements length. and seven
cross-sectional variables (X4' X S,X6 , X 7 , Xu. X12, XI3 ). Assuming symmetry,
the problem can be stated in terms of only seven independent variables Xlo X7 .
From the relation (1.51)
{jJ
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
{z}
(a)
X2 = (Xl + X 3 )fl.
Xs = (X4 + X6 )12
(b)
{JJ
1
0
0
0 0
0.5 0.5 0
0 0
1
0
0
0 0
0
0
1
0 0
0
0 0.5 0.5 0
0
0
0
1 0
0
0
0
0 1
Xl
X3
X4
X6
X7
(c)
38
1 Problem Statement
1
0 0
0
0
0.5 0.5 0
0 0
1
0
0 0
0
1
0 0
0
0
0
0 0.5 0.5 0
1 0
0
0
0
0 1
0
0
t=Lb= 0
1
0
0 0
0
0.5 0.5 0
0 0
1
0
0
0 0
1
0 0
0
0
0
0 0.5 0.5 0
0
0
0 1
0
(d)
and the original variables X are expressed in tenus of the reduced set of independent
variables y by (1.53).
Scaling or Design Variables. It is often desirable to eliminate wide
variations in the magnitudes of design variables and the value of constraints by
normalization. Design variables may be nonualized to order 1 by scaling. This
operation may enhance the efficiency and reliability of the numerical optimization
process.
Consider for example the variables XI and X2 , limited by the side constraints
To illustrate the effect of scaling on the objective function, consider the function
(Fig. 1.14a)
39
(b)
(a)
8.
Assuming
=(D/DU) - 1.0:s; 0
(1.55)
40
1 Problem Statement
This nonnalization does not affect the feasible region. The denominators of (1.55)
represent nonnalization factors which place each constraint in an equal basis. For
example, if the value of a stress constraint is -0.1 and the value of a displacement
constraint is -0.1, this indicates that each constraint is within 10% of its allowable
value. Without nonnalization, if a stress limit is 20,000, it would only be active
(within 10%) if its value was 19,999.9. This accuracy is difficult to achieve on a
digital computer. Also, it is not meaningful since loads, material properties and
other physical parameters are not known to this accuracy. Using nonnalization, the
constraint values are of the order of one, and do not depend on the units used.
Constraint Deletion Techniques. The number of inequality constraints in
optimal design problems may be very large, particularly in structures consisting of
many elements and subjected to multiple loading conditions. Constraint deletion
techniques [121] can be used to reduce the number of constraints. It is recognized
that, during each stage of an iterative design process, it is only necessary to
consider critical or potentially critical constraints. On the basis of analysis of the
structure, all the inequality constraints may be evaluated. Constraint deletion
techniques are then used to temporarily eliminate redundant and noncritical
constraints that are not likely to influence significantly the design process during
the subsequent stage. For each constraint type the most critical constraint value is
identified using regionalization and truncation techniques.
An example of the regionalization technique is that only the most critical stress
constraint in each region under each load condition is retained. The regionalization
idea works well provided the design changes made during a stage are small enough
that they do not result in a shift of the critical constraint location within a region.
A truncation technique, on the other hand, involves temporary deletion of
constraints for which the ratio of the stress to its allowable value is so low that the
constraint will clearly be inactive during the stage. Evidently, none of the
constrains included in the original problem statement are pennanently deleted
unless they are strictly redundant
Consider for example the nonnalized constraints (1.55). We can delete all but the
most critical (most positive) constraints. Alternatively, we may delete any
constraint whose value is less than (more negative than) some cutoff value gc' say
gc =-0.5 at the beginning of the optimization. This value can be reduced gradually
at the later stages.
Other techniques may be used. For example, we may delete first some
constraints that are more expensive to evaluate and solve the problem. These
constraints are then evaluated and if they are not violated the optimum is reached,
having avoided much costly computation. If these constraints are violated, we add
them to the constraint set and proceed from there.
Unlike the reduction in the number of design variables (by linking or basis
reduction) where the reduction is global in character, constraint deletion is a local
strategy.
Relative Minima. One difficulty in solving a nonlinear programming (NLP)
problem is that there can be multiple relative minimum points. A point is said to
be a relative (local) minimum if it has the least function value in its neighborhood,
but not necessarily the least function value for all X . Relative minima may occur
41
in NLP problems due to the nature of either the objective function or the
constraints, or of both. Consider for example a two-dimensional problem with
inequality constraints. It is evident that the minimum may be a point where the
constraints have no influence (Fig. LISa) and yet the problem has a relative
minimum . Relative minima may exist also in problems where the constraints are
active (Fig. USb). In both cases the multiple-optimum points are due to the
form of the objective function. A relative minimum that occurs due to the form
of the constraints is shown in Fig. USc.
Example 1.9. The grillage shown in Fig. 1.16a is subjected to two
concentrated loads P = 10.0. The width of the rectangular cross sections is b =
12.0, the members' lengths are lx = 1.0 and ly = 1.4, and the depths of the
longitudinal and transverse beams (Xl and X 2 , respectively) are chosen as design
variables. Neglecting torsional rigidity of the elements, the force method of
analysis is considered, with N being the vertical contact force between beams at the
intersection (Fig. U6b). The single compatibility equation is
(a)
The bounds on stresses are a U=-01- = 1.0, and the allowable moments are given by
i = 1,2
(b)
where the subscripts i denote the beam numbers. The optimal design problem is to
(c)
(d)
-2xi ~(0.7N)~2Xi
(e)
where N is given in terms of Xl and X2 by (a). The topology of the structure can
be optimized indirectly by letting Xl =0 or X2 =0, resulting in elimination of the
longitudinal or transverse beams, respectively. In either case a statically
determinate structure is obtained (with N =0 or N = 10).
The design space is shown in Fig. 1.17. It can be noted that three relative
optimum points, representing three different topologies have been obtained. The
optimal design values for the variables and objective function are given in Table
1.1. This result is typical to many grillages, where local optima fall into three
categories:
42
1 Problem Statement
(a)
Global
minimum
(b)
Global
/minimum
Relative
minimum
(c)
Fig. 1.lS.
Relative minima.
~"/-<
l,
43
!"~', f
tP
;00-1(
;00-1(
l,
I_
.
~+-~+~~
l,
;00
I(
l,
;00
(b)
(a)
X*T
z*
1
2
3
(2.24. O)
(1.63. 1.28)
{O l.87}
80.6
144.7
125.7
44
1 Problem Statement
including many bridge supports, transmission towers, ship masts and roof
supports.
b. A finite-element code for truss analysis is easily written, so it is not necessary
to expend major effort on the analysis portion of the design program.
c. Truss structures can be created which span the range of complexity from very
simple to highly nonlinear. This type of structure provides excellent test cases
for the study of optimization techniques.
min
rL ~ r(X) ~ r U
(1.56)
where r(X) and O'(X) are given in terms of X by the analysis equations (1.26) and
(1.29)
r(X) =K-iR
O'(X) = S r = SK-IR
(1.57)
45
z =.f(X) -+ min
(1.58)
Kr=R
C1=Sr
In this formulation the analysis equality constraints are included in the problem
formulation. These constraints are satisfied at the optimum but not necessarily at
intermediate designs. before the optimum is reached. Since the stresses are given
explicitly in terms of the displacements. it is not necessary to consider C1 as
independent variables. Substituting the explicit stress-displacement relations into
the stress constraints. the problem (1.58) can be stated in terms of only X and r as
z =.f(X) -+ min
(1.59)
Kr=R
In this formulation the equilibrium conditions are the only equality constraints.
1.4.2 Force Method Formulations
Formulation in the Design Variables Space. Considering the constraints
(1.42) through (1.44). the problem (1.46) and (1.47) becomes: find X such that
Z =.f(X) -+ min
DLS D(X) SDu
(1.60)
46
I Problem Statement
where D(X) and O'(X) are given by the analysis equations. Assume the common
force-stress relations
(1.61)
Aj
W.{XJ O'j
where Wi is the ith modulus of section, which is a function of the ith design
variable Xi [in the case of b'Uss structures, the cross sectional area Xj is used instead
of W.{XJ in (1.61)]. The displacements and the stresses are then given in tenns of
X by (1.22) through (1.24)
D(X) = Dp + DNF-tl)
O'(X) = Wit (Ap + ANF-tl)
(1.62)
where Wit is a diagonal matrix of the reciprocals wj- t . In the above fonnulation
the elements of Dp, DN, F and l) are explicit functions of both the cross-sectional
and the geometrical variables. The elements of Ap and AN are given explicitly in
tenns of only the geometrical variables and the elements of W tl are explicit
functions of only the cross-sectional variables. The matrix F-t is usually a
nonlinear implicit function of all design variables.
Simultaneous Analysis and Design. Considering the integrated formulation
(1.48) through (1.50), the problem (1.60) and (1.62) can be stated as follows: find
X, N, D and 0' such that
Z=f(X)
min
DLSD SDu
O'L SO'S aU
XLSX SXU
FN
(1.63)
= l)
D = Dp + DNN
0' = Wit (Ap + ANN)
Since the displacements and the stresses are explicit functions of X and N, these
relations can be substituted into the constraints and the problem can be expressed
in tenns of only X and N as
Z=f(X) ~ min
DL S Dp + DNN S D U
a L SWi1(Ap+ANN)Sau
XLSX SXU
(1.64)
47
In this formulation the implicit compatibility conditions are the only equality
constraints.
Explicit Formulations. In some problems, for example in statically
determinate structures, it is not necessary to consider the compatibility conditions
(1.22). Neglecting the latter conditions in a general statically indeterminate
structure, the problem (1.64) can be formulated explicitly in terms of X and N as
Z=f{X)
min
(1.65)
Since the forces in Ap and AN satisfy the equilibrium conditions, any selection of
N also will result in a corresponding set of forces that will satisfy these conditions
but not necessarily the compatibility conditions.
Considering the equilibrium conditions (1.1) as equality constraints, the problem
(1.65) can be formulated in terms of X and A. The virtual-load method may be
used in this case to obtain the displacement expressions (1.33). The resulting
explicit problem is to fmd X and A such that
Z=f{X)
min
DL ~ T 1/X ~ DU
(1.66)
C A=R
The optimal solutions of problems (1.65) and (1.66) are identical. However, the
following differences can be observed in the problem formulation:
-
The number of variables in problem (1.66), where all members' forces are
considered as variables, is larger. In problem (1.65) only the redundant forces
are considered as variables.
In problem (I.66) the equilibrium conditions are considered as additional
equality constraints. As noted earlier, any solution of the problem (1.65) will
satisfy these conditions.
48
1 Problem Statement
L liXi =lTX
II
Z=
i=l
(1.67)
These are typical relations in truss structures where Xi are the cross-sectional areas
and li are the members' lengths. Substituting (1.67) into (1.65) and considering
only stress and side constraints, the problem becomes: fmd X and N such that
Z = (l'X -+ min
(1.68)
XL:5:X:5:XU
In this formulation (J~ and (J~ are diagonal matrices of bounds on stresses, and
the elements of Ap and AN are constant, computed in the primary determinate
structure. Since the objective function and all the constraints are linear functions
of the variables, this is a linear programming (LP) problem.
Alternatively, substituting (1.67) into (1.66) and considering only stress and side
constraints, the following LP problem is obtained: find X and A such that
Z =(l'x -+ min
(1.69)
C A=R
Since the elements of C are independent of the cross-sectional variables, they are
constant. The two LP problems (1.68) and (1.69) are equivalent but, as noted
earlier, the number of variables is larger and the equilibrium conditions are
considered as additional equality constraints in problem (1.69). Both formulations
have been used extensively in various optimal design applications and will be
discussed throughout this text . In problems of optimal plastic design (Sect.
4.1.1), only equilibrium and yield conditions are considered, and both formulations
can be viewed as simultaneous analysis and design (SAND).
In the above presentation, explicit formulations of optimal design problems have
been obtained by neglecting the implicit analysis equations. Explicit exact
formulations can be achieved in simple systems and elements, or in statically
determinate structures where the stresses and displacements are given explicitly in
terms of the design variables. It will be shown now that such formulations are
possible also for some statically indeterminate structures. It has been shown that,
in general, the design variables are chosen as the independent ones and the behavior
(dependent) variables are determined by the analysis equations. Alternatively, if
49
compatibility conditions are neglected, both design and behavior variables are
assumed as independent variables [formulations (1.65) through (1.69)]. Fuchs [37]
proposed to choose cross-sectional design variables of the basic statically
determinate structure and the remaining redundant forces as independent variables.
To illustrate this possibility, consider again the basic relations (1.12) and (1.14),
where AR =N [see (1.18)]
(1.70)
FdRN
=G FdB AB
(1.71)
(1.72)
(1.73)
where F R is a vector of the diagonal elements of F dR, and N;i is a diagonal
matrix of the inverse elements of N. The chosen independent variables are the
elements of F dB and N, and the corresponding dependent variables are the elements
of AB and FR, given explicitly by (1.72) and (1.73) [instead of using the implicit
compatibility equations (1.22) to calculate N]. Thus, the implicit optimal design
problem (1.60) can be formulated now explicitly as follows: find F dB and N such
that
(1.74)
in which the relations (1.72) and (1.73) are substituted as necessary. It should be
noted that this formulation is not general and involves several limitations,
including:
- it is suitable only for certain cross-sectional design variables;
- it is not suitable for problems with linking of variables and multiple loading
conditions;
- it might be effective for problems where the force method of analysis is
assumed; in problems where the displacement method of analysis is used, the
number of variables might be considerably larger.
Example 1.10. In this example, various problem formulations are
demonstrated for the four-bar truss shown in Fig. 1.1. Assume the four crosssectional areas as design variables XT = {Xl' X2 X3 X4 } ,the volume of material
as an objective function
50
1 Problem Statement
(a)
(J
= 100
0.5
0.5]
[ 0
-0.5 0.5
-1
0
{rJ
'i
(d)
0.5 0.5]
L
E
O
1
Ii.
u
[
(J <-<(J
- 100 -0.5 0.5
-1
0
{rJ-
(e)
The assumed variables and constraints for the formulations (1.56) through (1.59)
are summarized in Table 1.2.
Force methodformulations. For the force method of analysis, the redundant forces
NT={A3 A4 } are given implicitly in terms of X by (1.22)
v'2(
100
E
~+l..
)+~
Xl X3
Xl
2
v'2
Xl
Xl
-+-
2 v'2
-+Xl Xl
2v'2
20v'2
Xl
{~}=_I~
--+-+X
Xl
Xl
(f)
40
Xl
51
(h)
(i)
Substituting (h) and (i) into the stress constraints (b) yields
Analysis
Method
Displacement
Force
Force
(explicit)
Formulation
(1.56),(1.57)
(1.58)
(1.59)
(1.60),(1.62)
(1.63)
(1.64)
(1.68)
(1.69)
(1.74)
Variables
~
X, r, cr
X, r
X
X, A, cr
X,N
X,N
X,A
X"Io X2,A 3,A 4
Constraints
Analysis
rules
(b)
(b),(c),(d)
(c),(d)
(c)'(~J
(b)
(f),( h),( i)
(b)'(f),(h),(i) (fl,(j)
(k)
(l),(m)
(p)
52
1 Problem Statement
(k)
[..fia
..fia
01 -..fia
-1] {~}= ..fi{10}
..fia 0 A3
10
(I)
(m)
..fi
..fi
X3
Xl
=[All
X4
A41
..fi
X2
Xl
X2
{~:}
(n)
Substituting (g), (i) and (n) into (b) gives the explicit constraints
aLS;
a (Xl, X 2 , A 3 , A4) s; aU
(p)
The assumed variables and constraints for the various force method formulations
are summarized in Table 1.2 .
Exercises
A. In exercises 1.1, 1.2 and 1.3, assume the displacement method of analysis.
Formulate the optimal design problem:
Exercises
100
53
100
1------1-- - - I
Fig.
1.18.
Sketch the feasible region and the objective functions contours. Find graphically
the optimal solution and the active constraints at the optimum.
1.1 The symmetric truss shown in Fig. 1.18 has two design variables: the crosssectional area XI and the angle X2 The bounds on stresses are aU = 20.0, 01- = 15.0. The upper bound on the vertical displacement at the free node is
rf = 0.02
xf
54
1 Problem Statement
100
Fig.
1.19.
i= 1,2
The bounds on stresses are aU = -01- = 20.0. The constraints are related to the
stresses in sections A, B, C, D and the objective function represents the volume of
the beam. The modulus of elasticity is 30,000.
B. In exercises 1.4 - 1.8 assume the force method of analysis. Formulate the
optimal design problem:
a. in the space of redundant forces and design variables [SAND, see (l.64)];
b. in the design variables space [see (l.60) - (l.62)];
c. in the LP form (l.68);
d in the LP form (1.69).
1.4 Solve the truss of exercise l.2, assuming N I
and N z =redundant force in member 4.
1.5 Solve the beam of exercise l.3, assuming NI = bending moment over the
interior support under PI' and N z
bending moment over the interior support
underPz
X,
I:
Fig.
lP' =8
I
A
1.20.
~P2
BC
I zs; I
:I:
= 16
I
D
100
i2
'I'
=200
X2
~
100
:1
Exercises
10.0
5.0--
8)
CD
CD
55
100
Fig.
100
.1.
100
.1
1.21
1.7 Assume the symmetric grillage shown in Fig 1.16 with the following data:
P
= 10.0,
Lx
= 100,
Ly
= 150.
= =
where Xh X2 are design variables representing the cross-sectional areas of the two
beams. The following relationships have been assumed:
56
1 Problem Statement
Fig.
1.22.
Modulus of section
Mcmnentofine~a
i= 1,2
Choose the ve~cal interaction force in the intersection of the two beams as the
redundant force and neglect torsional rigidity of the elements.
2 Optimization Methods
58
2 Optimization Methods
with large numbers of variables and constraints are often based on approximation
concepts. Some of the latter concepts will be presented in Chap. 3.
Nonlinear programming (NLP) problems can be divided into unconstrained and
constrained problems. In unconstrained optimization we do not consider constraints
and the problem is much easier to solve. Occasionally it is possible to eliminate
some or all of the constrains from a problem. Some optimization concepts related
to unconstrained minimum and constrained minimum are discussed in Sect 2.1.
Unconstrained minimization along a line in the design space, which is common to
many optimization algorithms, and methods for unconstrained minimization of
functions of several variables are described in Sect. 2.2. Methods for constrained
optimization are discussed in Sects. 2.3 (linear programming problems) and 2.4
(nonlinear programming problems).
f(X*) ~ft..X)
(2.1)
for all X in that region. Assume the Taylor series expansion of I about X* up to
quadratic tenns
(2.2)
r,
where
Vf"T, and H* are computed at X*. In (2.2), vr is the vector of first
derivatives, or the gradient vector (G) of I, and H is the matrix of second
derivatives, or the Hessian matrix, given by
(2.3)
(2.4)
59
vf* = 0
(2.6)
(2.7)
for all L1X -:F- O. Equation (2.7) in this case is a sufficient condition for a local
minimum of f{X) at X*. Both (2.6) and (2.7) ensure that a point is a relative
minimum. These conditions involve derivatives but not the value of the function.
It should be noted that if we add a constant to f{X), or if we multiply f{X) by any
positive constant, the minimum point X* is unchanged although the value j(Xj of
the function is altered .
The Hessian H* is a positive definite matrix if the condition (2.7) is satisfied for
every L1X. It is a positive semi-definite matrix if
(2.8)
and it is indefinite if q is positive for some vectors L1X and negative for others.
The matrix H* is positive definite if and only if all its eigenvalues are positive; it
is positive semi-defmite if and only if all eigenvalues of H* are nonnegative; and
it is indefinite if some eigenvalues are positive and some others are negative.
Several methods can be used for checking the form of H*. One way is by using
its principal minors. A principal minor Hi is a square sub-matrix of H of order i
whose principal diagonal lies along the principal diagonal of matrix H. The
matrix H is positive-definite if the determinants of all the principal minors located
at the top left corner of the matrix are positive numbers. The matrix H is positive
semi-defmite if the determinants of all principal minors are non-negative. If H is
positive semi-definite but not positive-defmite, then the determinant of at least one
of the principal minors is zero and higher order derivatives of f(X) are needed to
establish sufficient conditions for a minimum.
2.1.2 Constrained Minimum
Lagrange Multipliers. Consider the problem of minimizing a function
subject to equality constraints. At the optimum, the differential change in the
objective functionf(X), in terms of the differential change in X, must still vanish
60
2 Optimization Methods
(2.9)
where the differential changes dXb dX2> ... , dXn are related through the constraints.
Considering only a single constraint
h{X) =0
(2.IO)
oh
oh
oh
dh=-a dX1 +
- dX2 ++a1X2
aIX,. dX,.=O
1X1
(2.ll)
Of
oh
-+A-=O
oXi
i = 1, ... , n
oXi
(2.13)
Thus, we have a system of n+I equations [(2.1O) and (2.13)] and unknowns (X
and A). In the case of multiple equality constraints
= 1,... , nit
(2.14)
L Ajhj{X)
,.~
C\>{X, A) = f{X)+
(2.15)
j=l
i = I, ... ,n
(2.16)
j = l, ... ,n"
(2.17)
61
....
Vf+
I, AjVh j =0
(2.18)
j=l
=1,..., nit
(2.19)
#X)~O
= 1,... , n,
(2.20)
hj(X. Sj)=gj(X)+SJ=O
(2.21)
,(X.
s.
",
A)=f+I,Aj(gj+SJ)
j=l
(2.22)
i= 1..... n
(2.23)
i= 1..... n,
(2.24)
=l, .... n,
(2.25)
Equations (2.25) ensure that the inequalities gj ~ 0 are satisfied. Equations (2.24)
state that either Aj or Sj is zero, which implies that either the constraint is active
(gj 0) and must be considered in testing (2.23), or it is inactive (Aj 0).
Equations (2.23) require that Vf lie in the subspace spanned by those Vlj which
correspond to the active constraints.
62
2 Optimization Methods
L AiVgi =0
J
Vf+
(2.26)
i=l
Equations (2.26) are based on the conditions (2.23), considering only the active
constraints. With this definition, the conditions (2.24) can now be excluded.
To avoid situations in which the conditions (2.26) are satisfied and yet X is not
a local minimum, we require
j= 1,... , J
(2.27)
Equations (2.26) and (2.27) are the Kuhn-Tucker (KT) conditions for a relative
minimum [88]. Define a cone as a set of points such that if Vg is in the set, AVg
is also in the set for A~ o. The set of all nonnegative linear combinations
(2.28)
forms a convex cone. The KT conditions require that -Vf be within the convex
cone comprised by the active constraint normals Vgi (j = 1,... , I). These are
necessary conditions for a point to be a relative minimum, but they are not
sufficient to ensure a relative minimum. (This can be seen for example by the case
of point A in Fig. USb). In convex programming problems discussed in the next
subsection, the KT conditions are necessary and sufficient for a global minimum.
A two-dimensional geometric interpretation of the conditions is shown in Fig.
2.1. In the case of Fig. 2.1a, - Vf is not within the cone formed by Vgi' The point
is not optimal because/may be decreased without violating the constraints. KuhnTucker conditions are not satisfied, since we cannot find nonnegative Ai for which
-Vf is expressed as a linear combination of Vgi' In the case of Fig. 2.1b the
conditions are satisfied and the point is optimal. It can be seen that - Vf is within
the cone of Vgi' and hence we cannot make any move that reduces/ in the feasible
domain. We can find nonnegative Ai for which -Vf is expressed as a linear
combination of Vgi.
To illustrate the physical meaning of the Lagrange multipliers Ai' consider the
problem
Z=/(X, b)
min
(2.29)
j
= 1,... , n,
(2.30)
63
Vg j cone
(b)
(a)
Fig. 2.1.
8.
where bj are some given parameters. The KT conditions for this problem are given
by (2.26) and (2.27). It can be shown that at the optimum
dZ
-=-1...
db.
J
J
(2.31)
That is, Aj is the marginal price that we pay in terms of an increase in the objective
function for making the constraints more difficult to satisfy. This explains why at
the optimum all the Lagrange multipliers have to be non-negative. A negative
Lagrange multiplier would indicate thatj(X) can be reduced by making a constraint
more difficult to satisfy which is irrational.
In summary, the KT conditions can be used to check whether or not a given
point is a candidate minimum. Practical application of KT conditions as a test for a
minimum usually requires the solution of simultaneous linear equations for the Aj.
A procedure to compute Aj is discussed elsewhere [48]. The KT conditions provide
also the basis for some of the constrained NLP methods.
z =Xl + X2 ~
min.
gl ==xf+xi-8~O
g2 ==-XI-2.5~O
g3 == -X2 - 2.5 ~ 0
64
2 Optimization Methods
-3.0
-2.0
-1.0
-1.0
-2.0
-3.0
Active constraints
gl andg2
gl andg3
g2 and g3
X'
{-2.5, -1.323}
{-1.323,-2.5}
{-2.5, -2.5}
-3.823
-3.823
-5.0
In case C, the solution XTis infeasible because it does not satisfy the constraint gl
SO. In case A Al = 0.37S, ~ = -O.S90, A3 = 0, and in case B, Al = 0.37S, A2 =
0, A3 = -O.S90, both being nonoptimal solutions. The optimum is XT = -{2.0,
2.0} , Z = -4, Al = 0.25, and A2 = A3 = 0, only the constraint gl S 0 being
active. Modifying the constraint
xl+x~ SS
such that
Xf +X~ SS.1
we find the new optimum XT = -{2.0125, 2.0125}, Z = -4.025. At this point [see
(2.31)]
dZ == lJZ
dbt
llht
= -4.025-(-4) =-o.25=-AI
S.1-S.0
a!CX2 )+(l-a)!CXI
65
(X 2 )
x
(b)
Convex Functions and Convex Sets. The nature of the objective function
and the feasible region can be determined using the definitions of convex function
and convex set. A function f(X) is said to be convex if, on the line connecting
every pair of points Xl and X 2 in its domain of definition, the value of the
function is less than or equal to a linear interpolation offtX l ) andf(Xz), i.e.,
0< a < 1
(2.32)
The function is strictly convex if the strict inequality holds. A convex function is
illustrated in Fig. 2.3a. If a functionf is convex, then (1) is concave. A linear
function is both convex and concave, but neither strictly convex nor strictly
concave. A function may be neither convex nor concave (see Fig. 2.3b).
~-------------------- XI
(a)
(b)
66
2 Optimization Methods
A set of points is called convex if the line segment joining any two points Xl
and X2 is contained entirely within the set. Mathematically, the set is convex if for
all Xl and X 2 in the set, and 0 < a < I, the point Y;;;; aX I + (1 - a)X2 is also
in the set (Fig. 2.4). The set may be bounded or unbounded. The functionftX)
defined on a convex set is convex if and only if the Hessian matrix is positive
semi-definite or positive-definite at all points in the set.
A convex programming problem for minimization is one with a convex
objective function ftX) and convex inequality constraint functions gj(X). In this
case, the feasible domain formed by a single inequality constraint can be shown to
be convex. Furthermore, the intersection of convex domains is convex. Thus, if
the individual domains gj(X) ~ 0 are convex, the domain that is defined by all of
them is also convex. A problem with equality constraints is convex if the hJ(X) are
linear and if f(X) and gj(X) are convex. The intersection of linear equality
constraints is convex, since a single linear equality constraint is a convex domain.
Since all linear functions are convex, a linear programming problem is always a
convex programming. A nonlinear equality constraint always defmes a nonconvex
feasible region for the problem.
For convex programming problems the Kuhn-Tucker necessary conditions are
also sufficient. The significance of the above definitions is that in a convex
programming problem any local minimum is a global one. However, it is often
difficult to ascertain whether the functions in a given problem are convex.
Problems which are not convex programs may still have only a global minimum
or they may be solved for their relative minima, which provide useful information.
Most optimal design problems cannot be shown to be convex. However, some
of the approximate problems presented later in Chap. 3 are convex.
67
finite number of operations. and we often attempt to find only an estimate of the
minimum. In this section. the golden section method and polynomial fitting
techniques. which are commmly used. are described.
In general. it is assumed thatj{a) is a unimodal lunction. That is. a minimum
exists and it is unique in the interval of interest. For functions that are not
unimodal. we locate only a local minimum point. Most solution procedures can be
divided into two phases:
a. The location of the minimum point is bracketed and the initial interval of
uncertainty is established.
b. The interval of uncertainty is refmed by eliminating regions that cannot cmtain
the minimum. This is done by computing and comparing function values in
the interval of uncertainty.
Golden Section Method. Using this method. it is assumed that the function I
is unimodal. but it need not have continuous derivatives. Define f1l'St the Fibonacci
sequence by
10
=1
11 = 2
= 2.3 ...
That is. any number is obtained by adding the previous two numbers. so the
sequence of numbers is 1.2.3.5.8. 13.21 .... The sequence has the property
lor n-+ oo
This ratio between two successive numbers. as n becomes large. is called the
Golden Ratio.
(2.35)
Assuming that
(2.36)
then the mlD1mUm lies between
uncertainty. I, is given by
aq
and
a q .2
68
2 Optimization Methods
/(0.)
~~
________
aq_2
(I-P)
ab
CXa
PI
________
>J-"E
I
L-~a
at.
Initial bounds
__
(1-
P)
PI
First update
Fig. 2.S.
= Cl q = L
q
Cl u
5(1. 618)q
ClL
= Cl q-2 =
j=O
L 5(1.618)q
q-2
j=O
q-l
(2.37)
where Clu and of- are upper and lower limits on the interval of uncertainty.
The object now is to reduce the interval of uncertainty. Assume two function
values within the interval 1 symmetrically located at a distance of fY from either
end, including the known value at Cl q_l' The new interval of uncertainty, 13/, is
determined such that either the left or the right portion of 1 is eliminated.
Considering the left portion of the interval, it can be seen from Fig. 2.5
(2.38)
giving, 13 = 1/1.618 = 0.618 (the second root is not meaningful). Thus, the two
points, Cl,. and Clb, are located at a distance of 0.6181 or 0.382/ from either side of
the interval.
The solution procedure for reducing the interval is as follows:
a. For the given Clq-2, Clq-lt Clq [see (2.35)] and 5 (a chosen small step size in
the interval 1 is calculated by (2.37).
Cl),
69
The method is most reliable and it is easily programmed for solution on digital
computers. On the other hand, it requires a relatively large number of function
evaluations.
(2.39)
which has an easily determined minimum point At the minimum of q(a) we have
to satisfy
dq
do.
or
= b + 2ca = 0
(2.40)
b
(2.41)
2c
The constant coefficients b and c (a is not needed) can be determined by computing
11./2./3' the value off(a) at three different 0. values, ah 0.2' 0.3, and solving the
equations
0. = - -
It = a + ba1 + cat
fz = a + ba2 + ca~
13 = a + ba3 + ca~
If we use 0.1 =0, 0.2 =dO., and 0.3
equations (2.42) become
(2.42)
It =a
fz = a + bda + c(.1a)2
13 = a + 2Ma + 4c(da)2
(2.43)
70
2 Optimization Methods
/(01.)
.6.01.
3ft - 13
c - ~/3~+_h~I_-....,,2,..=.f.~2
(2.44)
2(l\o.l
2l\o.
Substituting (2.44) into (2.41), the approximated value of a. corresponding to the
minimum value of q(o.) is
a=ft
b = 4/2 -
a. *
(2.45)
13+ft>/2
2
(2.47)
This means that the value of 12 must be below the line connecting 11 andl3 (see
Fig. 2.6). 0.* is computed by (2.45) only ifl2 <11 and 12 <13.
The Cubic Fitting. If the derivatives of I with respect to a. are readily
computed, a two-point cubic interpolation can be used.j{o.) is approximated by
C(o.) =a + bo. + co.2 + do. 3
(2.48)
71
(2.49)
(2.50)
d=_I_
(L\a)2
L\a
(ex)
A =0
ex*
D.ex
B = D.ex
(2.52)
72
2 Optimization Methods
da
(2.53)
Substituting (2.52) into (2.53), we may find that only the plus sign in (2.51) is
required
73
improved and new algorithms are emerging. The methods discussed in this section
fall into the following categories:
vr.
vr.
Direct Search. Direct search methods are based on comparison of the value of
the objective function without the need of using derivatives. These methods are
usually reliable. easy to program. can deal effectively with nonconvex and
discontinuous functions. and in many cases can work with discrete values of the
design variables. The price paid for this generality is that these methods often
require many function evaluations to achieve the optimum. Therefore. they are
most useful for problems in which the function evaluation is not expensive. and
will be competitive for relatively small optimization problems.
quadratic function
q = XT a X + XT b+ c
(2.54)
(2.55)
74
2 Optimization Methods
at or before the nth step, regardless of the starting point. Note that the order in
which the directions are used is immaterial to this property.
Consider again the Taylor series expansion of a general function f about its
minimum X* [see (2.2)]. It has been shown [see (2.6)] that
= 0 at the
minimum. As AX approaches zero. higher-order terms of the series can be
neglected and f(X) approaches the quadratic form. Moreover. if U* is positive
definite. the approximate quadratic function has its minimum at X*. Applying
quadratically convergent methods to a general function for which the Taylor series
is dominated by the quadratic terms near the minimum. then a rapid convergence in
the neighborhood of X* is expected.
vr
Conjugate directions method. Among the various methods based on the concept
of conjugate directions. powell's method [107] is one of the most efficient. reliable
and successful. The method requires that the functionf(X) be unimodal. that is. it
has only one minimum. The differentiability requirement onf is implicit in the
exact line searches of the method.
Consider the two-dimensional example shown in Fig. 2.8. The function is first
minimized in each of the coordinate directions (directions S 1 and S 2).
Minimization is then performed in the pattern direction. formed by a line through
the initial and the last points of the previous step (direction S3). One of the
coordinate directions is discarded now (direction Sl) and the pattern direction is
included in the next minimization. A new pattern direction. S4. is generated and
again one of the coordinate directions. S2. is replaced. The above steps are repeated
until convergence. It can be shown that the pattern directions S3 and S4 are
conjugate. A single iteration of this simple version of Powell's method is:
a. Choose an initial point X and n initial independent directions (e.g . the
coordinate directions). Sq. q = 1.2... n .
b. Set Y f- X .
c. Find
to minimize f(X + o.S q) and set X f- X + o.*Sq for q = 1. 2 ... n.
d. Set S,,+l f- X - Y. find 0.* to minimize f(X + o.S,,+I). and set X f- X +
o.*S,,+1 .
e. Replace Sq f- Sq+l for q = 1.2... n .
f. Repeat from step b.
0:
Since the conjugate directions are not uniquely defined. different sets of n
independent mutually conjugate directions can be found. The various ways for
generating such directions form the basis for different methods all of which are
quadratically convergent.
Some difficulties may arise in the practical application of Powell's method. One
problem is that each step is required to be a minimizing step in the given direction.
Computing the exact minimum in each iteration step may require much
computational effort. Usually. the functions to be minimized are not quadratic, and
the number of iterations becomes large. For quadratic functions, we require three
function evaluations per step and the number of steps is n2 , thus the total number
of function evaluations is 3n2 For nonquadratic functions this number may
75
become 5n3 or more function evaluations, which makes the method prohibitive in
problems with a large number of variables.
Another problem is that the method can come to a halt before the minimum is
reached. Both this failure and the previously mentioned inefficiency are because the
Sj may become dependent or almost dependent. The original set of Sj is, of course,
independent, and in theory each of the directions that are generated should be a
linear combination of all the preceding Sj' unless some ex = O. One way to
overcome this difficulty is to reset the directions to the original coordinate vectors
whenever there is some indication that the directions are no longer productive.
Powell [107] proposed a procedure for such modified directions.
Gradient methods. Gradient methods are usually more efficient than direct
search methods. The price paid for this efficiency is that gradient information must
be supplied and that these methods often perform poorly for functions which have
discontinuous fIrSt derivatives. While in direct-search methods only the values of
the objective function are used to determine the direction of minimizations in the
individual steps, gradient methods use information available by computing the
gradient vector G ofj[see (2.3)]. Geometrically, G is normal to the tangent plane.
That is, the direction of G coincides with that of greatest rate of change of j, the
direction of steepest ascent. Therefore, -G is the direction of steepest descent.
----~~----------------------~~------------------------XI
76
2 Optimization Methods
Given a point Xq , the best direction to move to reduce the function value would
seem to be the one in which the function decreases most rapidly, namely,
(2.56)
where G q is the gradient vector at point X q The solution is obtained by
computing successively G q [by (2.3)], a.. (by minimization along a line), and
Xq+l [by (2.56)] until the minimum is found. The method, called the method of
steepest descent, can be useful in some problems, but better directions than that of
-Gq can often be found.
One difficulty which may arise even in problems of low dimensionality is that
the process of solution can be very slow. For functions with significant
eccentricity the method gradually settles into a steady n-dimensional zigzag. The
method may work well for mildly distorted hyperspheres (Fig. 2.9). It can be
shown that successive directions of steepest descent are orthogonal to one another.
Another problem is concerned with computation of the gradient vector. Solving by
gradient methods, one has to consider the following three situations:
a. Expressions for apaXj can be derived but are time consuming to compute .
b. The gradient vector exists everywhere, but expressions for iJj(i)Xj are impractical
or impossible to derive.
c. The gradient vector is not defined everyWhere.
In the frrst and second circumstances, consideration can be given to finite difference
approximations of the partial derivatives. The third situation is more difficult,
since the use of finite difference at a point along a discontinuity will be
meaningless.
where
GTG
Sq =-G q + G T q G q
q-l
q-l
q-l
77
~------------------------------~Xl
(2.57)
which can also be written
(2.58)
Premultiplying by H;1 and designating X as the new value Xq+lt we obtain
(2.59)
78
2 Optimization Methods
Variable metric method. The variable metric method of Davidon [21] Fletcher and
Powell [28] is based on replacing the inverse of the Hessian H;1 by an
approximate matrix Jq The method eliminates the need for evaluating second
derivatives and performing matrix inversions. Yet the method is quadratically
convergent, and the matrix Jq which is improved at each iteration converges to
H;I. It is a conjugate direction method, and it can be viewed also as a quasiNewton or Newtonlike method. Quasi-Newton methods are used when only first
derivatives are available, thus Newton's method cannot be implemented directly,
but the spirit of this type of method is preserved.
Using the variable metric method, we start with a given initial Xo and derme an
initial positive definite matrix Joo The identity matrix I usually serves this
purpose, namely, Jo=I. An initial direction So is determined by So = -Jo Go
(where Go = Vfo ) and the qth iteration step is performed as follows:
b. Compute Jq+l
79
Y q = Gq+1 - Gq
In applying the method, care must be taken to ensure that J q is not updated with
data arising from poor approximations to Cl q In practice the algorithm is a
powerful method and difficulties seldom arise, except on very badly eccentric
functions.
As with gradient methods, the computation of Vf by finite difference can be
considered for the variable metric method. Stewart [133] developed an estimate of
the incremental size in finite difference that will produce maximum accuracy. With
Stewart's modifications, this method becomes competitive with powell's method
for situations in which formulas for derivatives are not available or are impractical
to compute.
Goldfarb-Shanno (BFGS) [42] is to update the Hessian rather than its inverse. The
solution procedure is as follows. For the initial design Xo. the Hessian of the
objective function is estimated by a symmetric positive definite matrix Ho. (i.e.
H o, =I), and the gradient vector, Go = Vfo, is calculated. The qth iteration step is
performed as follows:
a. Calculate the norm of the gradient vector IIG qll. If IIG qll < then stop the
iterative process.
b. The search direction Sq is determined by solving the linear system of equations
[see (2.57)]
where
80
2 Optimization Methods
YqY:
=_G--'-qG
.......~_
(GqS q )
=US q
It can be shown that the BFGS update formula keeps the Hessian approximation
positive defmite if exact line search is used. Numerical difficulties can arise since
the Hessian can become singular or indefinite due to approximate line search or
round-off errors. Some modifications can be used to overcome this problem.
a. The exact global optimum is reached in a finite number of steps and there are
no local optima.
39,45].
81
II
Z=
CjX j
~ min
(2.61)
j=l
subject to
L
II
ajjXj {
~. =. ~ )bj
i=I. .... m
(2.62)
= 1,... , n
(2.63)
j=l
Xj ~
where ajj, bj and Cj are constant coefficients. The notation (~, =,~) means that the
constraints might be either equalities or inequalities (~ or ~). This formulation is
general and covers a wide range of problems, as can be seen from the following
observations:
(2.64)
where X'l; ~ 0 and X\ ~ O. Thus the constraints (2.63) can always be satisfied;
however, the number of variables is increased. Other transformations can be
used to ensure nonnegativity of the variables (see for example Sect. 4.1.1). If
any specific variable Xl; has a lower bound limitation
(2.65)
82
2 Optimization Methods
L" CjXj
(2.68)
j=l
subject to
"
L
aijXj =b
j=l
i=I, ... ,m
(2.69)
j= 1,... , n
(2.70)
lt
(2.71)
(2.72)
(2.73)
(2.74)
j=l
The slack and surplus variables satisfy the nonnegativity constraints (2.63). Thus,
any vector X that satisfies the equalities (2.72) and (2.74) will satisfy also the
83
original inequalities (2.71) and (2.73) . In addition, since all the coefficients of the
new variables in the objective function equal zero, their contribution is zero and the
original function (2.61) is unchanged. Using the transformation (2.72) or (2.74) for
all inequalities, the original problem can be formulated in the standard form of
(2.68), (2.69) and (2.70).
If m = n in the standard form problem and none of the equations (2.69) are
redundant, there is only one solution to the system of equations. If m > n, then
there are redundant equations that can be eliminated or the system has no solution.
In all cases of interest m < n, the system possesses an infinity of solutions of
which we seek the one that minimizes Z and satisfies the nonnegativity
constraints Xj ~ O.
A number of standard definitions can be made now. A vector that satisfies the
equalities (2.69) represents a solution. If the nonnegativity constraints (2.70) are
also satisfied, this is afeasible solution. The optimal feasible solution is the
feasible solution which minimizes the objective function Z. A basic feasible
solution is a feasible solution with no more than m nonzero Xj' In other words, it
has at least (n-m) Xj that are zero. A nondegenerate basic feasible solution has
exactly m positive Xj'
3
2
\\\
3
Xl
2X l + X 2 = 4 max Z = Xl + X 2
(b)
(a)
3
(e)
Fig. 2.10.
Cd)
84
2 Optimization Methods
Geometrical Interpretation. Some observations can be made for twodimensional examples (Figs. 2.10 and 2.11). Considering only inequalities in the
original problem, the constraints define a feasible region bounded by portions of
some of the planes ailXI + aizX2 =bi . It can be seen that the optimum value of Z
is obtained at one of the vertices (extreme points). Each vertex represents a basic
feasible solution of the standard form problem. If the region is not bounded, then Z
may still take on its minimum values at a vertex or it may approach 00 (or -00). In
the latter case the solution is said to be unbounded. If the LP has a bounded
solution, it will always take on its minimum at a vertex; but it may take on this
same value all along an "edge" joining two vertices, or over the hyperplane
containing three vertices, etc. These observations can be stated rigorously as
theorems and are proven in most standard texts on LP [20,45].
In Fig. 2.10, LP problems with the constraints
2X I + X2~4
Xl + 3X 2 ~ 6
Xl, X2~ 0
and different cases of objective functions are demonstrated. The feasible region is
bounded and different optimal solutions are obtained. In Fig. 2.11 the constraints
are
2X I + X 2 ~ 4
Xl + 3X2~ 6
XI,X2~0
The feasible region is unbounded and it can be seen that the optimal solution
might be either at a vertex or it may approach 00
X2
Xl
(a)
Fig. 2.11.
QI"
~";~
(b)
2.3.3
85
Method of Solution
It can now be stated that if the standard form linear programming problem has a
bounded solution. the minimum of Z is attained at one of the basic feasible
solutions of the program. This is an extremely powerful result. since there is a
finite number of basic solutions; that is. there can be no more than the number of
ways m variables can be selected from a group of n variables. or
n!
(n-m)!m!
(2.75)
=(n)
m
Usually there are much fewer possibilities. since many cQmbinations are
infeasible. The simplex method. which will be described briefly here. is a powerful
computational scheme for obtaining basic feasible solutions. If a solution is not
optimal. the method provides a procedure for finding a neighboring basic feasible
solution which has an improved value of Z. The process is repeated until. in a
finite number of steps (usually between m and 2m). an optimum is found.
Rewrite (2.69) in the expanded form
auXI
a21XI
(2.76)
It is possible to place this system into a form from which at least one solution can
readily be deduced. We obtain this form by replacing the original system of
equations with another one of the same size which is a linear combination of the
equations of the original system. This procedure. called a pivot operation. produces
the new system which is said to be equivalent to the original one. Both systems
have exactly the same solutions. We may repeat operations on the system until
there will be m columns. each containing zeros and a single 1.0 . These can always
be arranged in the form
+a'I,III+1 XIII+! + .. +a'I,,, X" = b'l
+a'2, ...+1 X III+1+... +a'2,,, X" =
XIII
b'2
(2.77)
where the primes indicate that the a'ij and b'i are changed from the original system.
Such a system is said to be canonical or in a canonical form. One solution to the
system is always
j= 1.2..... m
(2.78)
(2.79)
86
2 Optimization Methods
This solution is a basic solution and the ftrst m of the Xj are basic variables. If all
the bj are nonnegative, then the solution given by (2.78) and (2.79) is a basic
feasible solution. The general procedure for solving LP problems consists in going
from one basis, or canonical form, to an improved one until the optimum is found.
In discussing the method of solution the questions that should be answered are:
a. What are the criteria for choosing an improved basic feasible solution?
b. How to go from one basic feasible solution to another one?
c. How to ftnd an initial basic feasible solution in a simple way?
Given a
system in canonical form (2.77) corresponding to a basic feasible solution, we can
pivot to a neighboring basic feasible solution by bringing some speciftc variable
X t into the basis in place of some Xs. The objective function Z can be expressed in
terms of the nonbasic variables by eliminating the basic ones
Z = ZO +
L"
C'j Xj
(2.80)
j=III+1
where ZO is a ftxed value and C'j are constant coefftcients of Z in its new form.
Note that ZO represents the value of Z for the current basic solution, since Xj =
o for j = (m+l), ... , n [see (2.79)]. The question now is how to choose X t and Xs.
If any speciftc coefftcient C't (corresponding to a nonbasic variable X,) in the
objective function of the canonical form (2.80) is negative, then the value of Z can
be reduced by increasing the variable X t from zero, while keeping all the other
nonbasic variables at zero. Therefore, to improve the solution, we can bring any
such X t into the basis. If more than one C'j is negative, we can choose among
them, and one possibility is to choose the t for which
C't
=m~n(c'j )
(2.81)
This choice selects the variable for which the objective function reduces at the
greatest rate. It should be noted that some other choice may produce a greater
improvement in Z. However, the above criterion is simple, and experience has
shown it to be efftcient. If min (cj) is not negative, then there is no variable that
can be increased to improve the objective function, and the optimality condition is
satisfted.
Bringing the chosen Xt into the basis, we have to determine now what variable
Xs will leave the basis. Going from one basic solution to another, we increase X t
from zero to some positive value during the process, while Xs becomes zero. If we
are to keep the solution feasible, we cannot increase X, from zero by more than it
takes to make the "frrst" Xs just go to zero, since to increase it further would make
Xs negative. Suppose we take the canonical system (2.77); then keeping all the
nonbasic variables zero except Xt, we can see the effect of increasing XI on the
basic variables from
Xi
= b'i - a'it{,
i= 1,... , m
t> m
87
(2.82)
If any specific a'n is positive, the largest value of X, before Xs becomes negative is
Xl --~
(2.83)
a'n
X, = m~n(b'i la'il )
for
a'il>O
(2.84)
In other words, since we wish to make X, basic (nonzero) in place of some other
variable Xs which will become nonbasic (zero), we must choose the row s for
which X. = 0 by (2.82), and all others are nonnegative. Hence row s must
produce the minimum value in (2.84). Note also that the chosen variable Xs has a
coefficient of 1.0 in the sth row.
In summary, once it has been detennined to bring the currently nonbasic variable
X, into the basis [using the criterion (2.81)], we compute s by the condition
(2.84)
b's
-,-=
aSI
. (b'i I ail
' )
m~n
for
a'il>O
(2.85)
Ysj
iSj
=-,a sl
(2.86)
(2.87)
Yij and iij are the coefficients of the new and current bases, respectively, in the ith
row and the jth column. The transformations (2.86) and (2.87) are carried out also
for the column of b'i and for the row of the objective function Z.
Example 2.2. As an example consider the canonical fonn
88
2 Optimization Methods
Xl-4X2+~3-4X4
1
3
--X2 +3X3 --X4 +Xs
4
8X2 -24X3 + 5X4
=3
=5
= Z-28
The objective function is expressed here in tenos of the nonbasic variables. This
can be done by eliminating the basic variables Xl and Xs. The basic solution is
Xs = 5
From the last row it can be observed that Z = 28. Since C'3 < 0, we can improve
the objective function by bringing X3 into the basis. To detenoine the variable s
to leave the basis we compute
1)
b's = min(l
=1
2 '3
2
a's3
Thus, s = 1 and the pivot is a'st = a'l3 (the underlined term). Using the
transfonoations (2.86) and (2.87), the new canonical system is
131
-Xl --X2 +X3 --X4
2
8
8
3
7
3
--Xl +-X2
--X4 +Xs
2
8
8
12Xl -X2
+2X4
3
=2
1
=2
=Z+8
X3 and Xs are the new basic variables, and the new basic solution is
The improved value of the objective function is Z = -8. Since C'2 < 0, the objective
function can be improved by bringing X2 into the basis. Thus we may find the
new basis and proceed with the above procedure until the optimum is found.
Special Cases. If we pivot on a row for which the right-hand side b'i is zero,
the value of the objective function or of the basic variables will not change. In
such a case, the value of the basic variable corresponding to the zero b'i is zero, and
we have a degenerate basic feasible solution. In the degenerate case, the objective
function is not improved even though a variable which has a negative coefficient in
the objective function is brought into the basis. In the optimization procedure that
will be described subsequently, we will still perfono such pivot operations, since
doing so may allow us to proceed to other nondegenerate solutions.
89
If all the coefficients a'it in any column t corresponding to a negative c', (not
necessarily that with the smallest cj) are negative. Xt can be increased indefinitely
without causing any variable to become zero [see (2.82)]. The result is that the
objective function can be reduced indefmitely and we have unboUllded solution.
Initial Basic Feasible Solution. A basic feasible solution is required as a
starting point. The object is to form an initial canonical system in which the
coefficients aij of the basic variables form a unit mattix. so that a simple basic
feasible solution can be found. If all the constraints in the original problem are
inequalities (~) with nonnegative bi then in order to get the standard form we have
to add slack variables to all the constraints. The coefficients of the slack variables
form in this case a unit matrix. thus they can be chosen as initial basic variables of
the canonical system. This is not the case if some of the constraints are equalities
or ~) inequalities. In order to start the solution with a unit matrix. we may add to
each of the constraints a new variable called artificial variable. The original system
of equations (2.76) in its standard form will become
(2.88)
Equations (2.88) are expressed in a canonical form where Xj.j =(n+l) ... (n+m)
are artificial variables. A basic feasible solution of this system is (bi ~ 0)
j= 1.2... n
Xj=O
X"+j =bj
= 1.2,... , m
(2.89)
The coefficients of the artificial variables in the original objective function are set
equal to zero.
The algorithm for solving the problem with artificial variables consists of two
phases: phase I to find a basic feasible solution. if one exists. in which all
artificial variables equal zero. and phase II to compute the optimal solution. In
phase I we defme an "artificial" objective function
=L
"+111
Z'
Xj
(2.90)
j=,,+l
which is to be minimized. If the minimum of Z' is zero. then all the artificial
variables have been eliminated from the basis and a new basic feasible solution is
available which contains only the variables of the original system Xj' j = 1. 2 ..
n. Then the artificial variables and objective function (X) can be dropped and we
proceed and solve phase II. namely. the problem with the original variables and
objective function. If the minimum of X is greater than zero, then no basic
feasible solution to the original problem exists.
90
2 Optimization Methods
b.
c.
d.
e.
f.
= m~n(c'j )
(2.91)
a.t
for
a'it>O
C't < 0,
find the
(2.92)
illj
II
a'.
y.=-
91
(2.93)
i;#s
(2.94)
s is the row in which XII has a coefficient of 1.0, and Yij and
iij
are the
2XI+ X2 S;4
Xl + 3X2 S; 6
Xl + X 2 =Z-+ max
2XI + X 2 +X3
=4
X I +3X 2
+X4 =6
-Xl -
X2
X3
X4
1
0
0
0
1
0
=Z -+ min
or in a tableau fonn
XI
1
-1
1t
Xa
1
3
-1
b
4
6
Since all the constraints are (S;) inequalities with nonnegative bit artificial
variables are not required and we start directly with phase II of the solution. The
initial basic feasible solution is X3 4, X4 6, Xl X2 0, and Z =0, which is
the vertex {O,O} in Fig. 2.10a. Since c't C'2 -I, either Xl or X2 can enter the
basis. Choosing X, = Xl we compute the variable to be eliminated from the basis
by
= = =
= =
2 ' ~}=2
I
min{i
namely, X3 is eliminated from the basis. The next pivot is au 2, the underlined
term in the tableau. The denote the basic variables, 1t shows the new basic
92
2 Optimization Methods
variable, and .u the variable to leave the basis. Performing the pivot operations we
fmd the new canonical system
Xl
1
0
0
X2
1/2
X3
1/2
-1/2
1/2
-1/2
X4
Z+2
.u
1t
The new basic solution is Xl =2, X4 =4, X2 =X3 =0 and Z = -2, which is the
vertex {2,O} in Fig. 2.10a. The new pivot is an =
and after transformation
we have the new canonical form
sa,
o
*
3/5
-1/5
2/5
-1/5
2/5
1/5
6/5
8/5
Z+14/5
As all e'j ~ 0, this is the optimal solution: Xl =6/S = 1.2, X2 = 8/S = 1.6, X3 =
X4 =0, and Z = -14/5 = -2.8. (Note that for the original problem, Z = 2.8). It
represents the vertex {1.2, 1.6} in Fig. 2.10a. In this example we see how the
solution proceeds from one vertex, or basic solution, to a neighboring and
improved one until we fmd a vertex in which the objective function is better than
in all its neighboring vertices. We do not have to check points inside the feasible
region during the solution.
Example 2.4. Find the optimal solution of the LP problem
5X l
=20
In order to have an initial basic solution we add two artificial variables, X6 and
X7.The new problem is
93
X2
-4
-1
6
5
Xl
5
1
1
-6
Artificial
X4
-2
-1
1
3
X3
11
5
-7
-18
Xs
1
1
5
-2
X7
0
1
0
0
X6
1
0
0
0
1t
20
8
Z
Z'-28
!!
Initial values of the objective functions are Z' = 28, Z = O. In phase I only Z is
minimized, but transformations are made also for Z. Preforming the next two
pivots, we have
Artificial
Real variables
Xl
5/13
-12/13
48/13
12/13
X2
-4/13
7/13
50/13
-7/13
X2
-3/8
Xl
1/2
-3/2
12
0
-1
0
X3
1
0
0
0
Xs
1/13
Bill
72/13
-8/13
20/13
4/13
Z+140/13
Z'-4/13
1t
X4
-1/8
-3/8
2
0
X3
1
0
0
0
X7
0
1
0
0
X6
1/13
-5/13
7/13
18/13
!!
Real variables
1t
X4
-2/13
-3/13
-1/13
3/13
Artificial
Xs
0
1
0
0
X6
1/8
-5/8
4
1
X7
-1/8
13/8
-9
1
3/2
1/2
Z+8
Z'+O
!!
Since Z' = 0 and the artificial variables are not in the basis, phase I is completed.
Now we can drop the artificial variables and the row of Z'. Z is the objective
function in phase II, and after performing the pivoting we have
Xl
-1/7
-12/7
72/7
X2
0
1
0
X3
1
0
0
X4
-2/7
-3/7
11/7
Xs
3/7
8/7
8/7
12/7
4/7
Z+60/7
X2 =4n
X3 = 12n
Xl =X4 =Xs = 0
Z = -60n
94
2 Optimization Methods
Z=
..
CjX j
(2.95)
-+ min
j=l
subject to
..
aijXj
=1, ....
(2.96)
= 1,... , n
(2.97)
~ bi
j=l
This can be done by expressing all inequalities in (~) form. Note that we do not
require bi ~ 0, and that the equality constraints can also be expressed as inequalities.
For example, the constraint
L
II
akjXj
(2.98)
=bk
j=l
-L
II
II
akjXj
~bk
j=l
~ -bk
akjXj
(2.99)
j=l
The problem (2.95), (2.96) and (2.97) is called the primal problem, for which a
dual problem, with variables Ai' can be stated as follows
L biAi -+ max
WI
cjI =
(2.100)
i=l
subject to
L
WI
aijAi $, Cj
=1. ....
(2.101)
i=l
i= 1... m
(2.102)
95
in which the coefficients aij. bit and Cj are identical to those of the primal problem.
If. for example. the primal problem is
The solutions of the two problems are identical. namely. min Z = max cII. If one
of the two problems is solved. we can find the solution of the other. Since the
computational effort in solving LP problems is a function of the number of
constraints. it is desired to reduce this number. The number of constraints in the
dual problem is equal to that of variables in the primal and vice versa, thus we can
solve the problem with the smaller number of constraints. For example. if we have
a primal problem with two variables (n=2) and 20 inequality constraints (m=20).
then 20 surplus variables are needed to convert the problem into a standard form.
At each stage we must deal with a 20 x 20-basis matrix. The dual problem.
however. has two inequality constraints and 20 variables. This leads to a 2 x 2basis matrix and to a great computational advantage.
If the primal problem is given in the standard form: find XT (Xl ..... X.. ) such
that
Z=
..
..
CjX j
~min
(2.103)
j=l
aijXj
= hi
i=I ..... m
(2.104)
j=l
= 1..... n
then the corresponding dual problem is: fmd AT=(A1 ..... A.".) such that
j
(2.105)
96
2 Optimization Methods
(2.106)
i=l
L
III
aijAi
i=l
~ Cj
j=I, ... , n
(2.107)
(2.108)
~ unconstrained in sign
A primal problem in any of the forms (2.95) through (2.97) or (2.103) through
(2.105) may be changed into any other by using the following devices:
A. = A(minZ)
,
Mi
(2.109)
Equation (2.109) is valid only in the range in which changes in bi do not change
the basis of the optimal solution but only the value of the basic variables.
Table 2.1 Primal-dual correspondence.
Primal Quantity
L"
L biAi ~max
L ~
L =
III
CjXj
~min
i=l
j=l
III
aijAi
Cj
aijAi
Cj
i=l
III
Xj
unconstrained in sign
L"
L"
i=l
aijXj =bi
j=l
j=l
aijXj
~bi
Ai unconstrained in sign
97
(2.110)
98
2 Optimization Methods
provides acceptable results for the problem of interest. Some of the more
complicated algorithms, considered best by the theoreticians, are found to be less
reliable for problems that are not carefully formulated. On the other hand,
algorithms like the feasible directions method (Sect. 2.4.2) and the sequential LP
(Sect. 3.1.3) are considered "poor" by the theoreticians, but usually perform
reliably and efficiently in a practical design environment.
Using constrained minimization methods, the design variables are modified
successively during the design process by moving in the design space from one
design point to another. Most methods consist of the following four basic steps:
99
designs away from the constraint surfaces until final convergence. One
drawback is that we have to start the solution always with a feasible design.
c. The augmented Lagrange multiplier method, where Lagrange multipliers are
incorporated into the optimization strategy to reduce the iII-conditioning often
encountered in SUMT.
The use of penalty function was fmt suggested by Courant [19]. Further work in
this area is based on developments introduced mainly by Carroll [15] and Fiacco
and McCormick [26]. The methods presented herein are widely used in structural
design and have some practical advantages. The algorithms are general and suitable
for various optimization problems. For problems of moderate complexity, the
unconstrained formulations for constrained problems are usually simple and
convenient to apply, provided an adequate minimization algorithm is available. In
addition, the methods might work well with approximate behavior models
discussed in Chap. 3. On the other hand, they may not be as efficient for some
problems as the direct methods.
An Exterior Penalty-Function. Consider a general nonlinear programming
problem where the equalities are excluded: fmd X such that
Z=j(X)
min
j
= 1,..., n,
(2.111)
The idea behind penalty-function methods is simple. Rather than trying to solve
the constrained problem, a penalty term that takes care of the constraints is added to
the original objective function I and the problem is transformed to the
minimization of a penalty function 'P(X, r)
".
'P(X,r)=I+rL, (gjr
(2.112)
j=l
The factor r performs the weighting between the objective function value and the
penalty term, and it is often called the penalty parameter or response lac tor. The
surfaces of 'P(X, r) are correspondingly termed response sUrfaces. The bracket
operator <gj> means
if gj > 0
(2.113)
if gj SO
100
2 Optimization Methods
'I' = f +
",
rj
< gj
101
(2.114)
>2
j=l
SO
f=XI2+xi
~min
gl ==4-XI -4X2 ~O
g2 == 4 - 4XI - X2
1.28. The solution for r = 1 is X =0.77, '1'. = 1.23,/ = 1.18, which is outside
the feasible region. Variations of/, '1'., and X with r are plotted in Fig. 2.13. It
can be observed that, for r > 1, changes in the values of X are small. For r = 10,
for example, X = 0.797,/ = 1.27 and '1'. = 1.275, which is very close to the true
optimum.
>It
>It(r = 10)
20
18
16
14
12
10
8
6
4
2
0.2
0.4
0.6
0.8
1.0
1.2
1.4
102
2 Optimization Methods
X' f'
,It"
1.2
f'
0.8
X'
0.6
0.4
0.2
~--~----------------------------------------~r
10
Fig. 2.13.
j=l
(2.115)
gj
but other forms of the penalty term can also be chosen. The idea is again to
minimize (2.115) for a sequence of values of r, instead of solving the constrained
problem (2.111). Since, at an interior point, all the terms in the sum are negative,
a positive choice of r will result in a positive penalty term to be added to f. As a
boundary of the feasible region is approached, some gj will approach zero and the
penalty term will approach 00. Reducing successively the parameter r, the
constrained minimum off is approached. It can be shown, however, that as with
the exterior penalty-function method. the closer to the constrained optimum the
103
minimum of 'II is forced to lie, the more eccentric the function becomes. Thus, it
is necessary again to minimize'll sequentially.
The solution process is as follows. Initial values of X and r are chosen. X
must lie in the feasible region, namely, all the constraints gj(X) ~ 0 are satisfied.
For the given r, the function 'I'(X, r) is minimized to obtain X, and the
convergence criterion of X to the optimum is checked. If it is not satisfied, the
parameter r is reduced r f- cr (c < 1), and starting from X the function 'I'(X, r)
is minimized again. The steps of reducing r and minimizing 'I'(X, r) are repeated
until the convergence criterion is satisfied. The following points should be
considered in practical application of the method:
, == fmin('i-l)- fmin(rj)
(2.116)
fmin(rj)
,.
Another
and stop when the absolute values of the components of x are smaller than a
desired value ~ .
d. Extrapolation techniques. Fiacco and McCormick [25] have shown that the
104
2 Optimization Methods
H(r)
H(O) = a
~----~------------~----------~
aj and bj for the ith approximation are determined by fitting through the two
points of Hirj.]) and Hirj)
2
H(r:
" -I) = aI. + b.r}'1
I. r.-
=!.rmn. (r..,- I)
(2.119)
=fmin(r;)
(2.120)
g=E:
g= 0
1\
Feasible
region
g=E:
IILinear
extended
g=o
Feasible
Quadratic
extended
\If
f
Ci
(a)
(b)
105
4
Optimum for
1
r=
True constrained
optimum
(a)
(b)
= 10,
106
2 Optimization Methods
Figures 2.16a and 2.16b show contours of'll for r = 10 and r = 1. It can be
observed that the minimum for r = 1 is closer to the true constrained optimum of
the original problem. In addition, the closer to the constrained optimum the
minimum of 'I' is forced to lie, the more eccentric the function becomes. Variation
of'll,! and X; = X; = X with r are shown in Fig. 2.17.
The Augmented Lagrange Multiplier Method. Multiplier methods
combine the use of penalty functions with that of Lagrange multipliers. When
only Lagrange multipliers are employed, the optimum is a stationary point rather
than a minimum of the Lagrangian function. When only penalty functions are
used, ill-conditioning problems may be encountered during the solution process.
The object in combining both is to obtain an unconstrained minimization problem
that does not suffer from ill-conditioning. The conditions of optimality are included
in the optimization algorithm in order to improve its efficiency and reliability. As
a result, the dependency of the method on the choice of the penalty parameters is
reduced.
qt',
j', lOX'
50
40
30
20
10
100
200
107
z =f(X) ~ min
j
= 1,... , nit
(2.121)
Using an exterior penalty function, the following definition of 'II can be used [see
(2.112)]
L r[h (X)]2
"l
'(X.r) = /(X)+
(2.122)
j=l
L Aj
"l
cj>(X.A) = /(X) +
h/X)
(2.123)
j=l
The stationary conditions acj>/aX j together with the equality constraints are the
necessary Kuhn-Tucker conditions for optimality. It can be shown that the
minimum of the Lagrangian function subject to the equality constraints provides
the solution of the original problem (2.121). Thus, using the exterior penalty
function approach we defme the augmented Lagrangian as
A(X. A. r)=/(X)+
L {Ajhj(X)+r[hj(X)fl
"l
(2.124)
j=l
If all Aj = 0 we get the usual exterior penalty functions (2.122). On the other hand,
for the optimum values A"j the minimum of A (X, A, r) provides the true
minimum off{X) for any positive value of r. Then there is no need to use the large
value of r required in the case of an exterior penalty function.
In practice we do not know A" in advance, so we can modify A iteratively until
the optimum is reached. Assuming 1..=0 and an arbitrary small value of r, the
pseudo-objective function (2.124) is then minimized for the given A and r. To
obtain an estimate for the Lagrange multipliers we compare the stationary
conditions
(2.125)
with the exact conditions for the Lagrange multipliers
108
2 Optimization Methods
(2.126)
Comparing (2.125) and (2.126), Hestenes [53] suggested the estimation
A.(~+l)
= A.(~)
, + 2 r h.,(X(.t)
(2.127)
z =f(X) ~ min
j
= 1,..., n,
(2.128)
", (A.
)2
L
-L+ gj
j=l 2r
(2.129)
at
~ (A..
ago' =0
- ' +gj ) j=l 2r
aX
-+2r ..J
aX
(2.130)
(2.131)
where it is also required that
109
to large values.
h. Precise #X) 0 and hJ{X) 0 is possible.
c. Acceleration is achieved by updating the Lagrange multipliers.
d. The starting point may be either feasible or infeasible.
A(X,
A, r)=xr+xi+r
Assuming the initial values 1..(1) =0 and r(I) = 1, the solution of the unconstrained
minimization is X(I)T = {0.769, 0.769}, gi =gz = 0.155. Using (2.132) the
estimated Ais A(Z)T = {0.31 , 0.31}. Repeating the optimization for r(Z) = 10 and
1..(2) yields X(Z)T = {0.800 ,0.800], gi =gz =0, which is the true optimum. The
1..*
x*
r=1
0.82
r=1O
0.80
--====--n-=--r=100
0.78
A.*= 0.32
1~---l---'--'--J.......-L___"A.
0.2
0.4
0.6
110
2 Optimization Methods
a. The direction must be feasible. i.e., we can take at least a small step from Xq
along S that does not immediately leave the feasible domain. In problems
where the constraints at a point curve inward, this requirement is satisfied if
(2.134)
If a constraint is linear or outward curving, we may require
(2.135)
The interpretation of this condition is that the vector S must make an obtuse
angle with all constraint normals except that, for the linear or outward-curving
ones, the angle may go to 900. Any vector satisfying the strict inequality lies
at least partly in the feasible region (see Fig. 2.19).
b. The direction must be usable, i.e., the value off is improved. This requirement
is satisfied if
~-----------------------------------------Xl
Fig. 2.19.
Feasible directions S.
L -_ _ _ _ _ _-'---'---'---'--_ _ _ _ _ _
111
XI
STVf < 0
(2.136)
X2
S(O-) 00)
Xl
I - -_ _ _------------~-
112
2 Optimization Methods
(2.137)
~~max
STVgj + OJ ~
S;
j= 1, ... , J
(2.138)
sTvr + ~ s; 0
(2.139)
-1 s; S s; 1
(2.140)
113
~-------/~~--------------------Xl
Optimum
f along S.
(2.141)
or
(2.142)
If the result is negative. we proceed and find a ne,w direction at Xq+ I. otherwise the
minimum of/is along the line segment Xq+l - Xq in the feasible region.
114
2 Optimization Methods
6
Objective function contours
~----~------L2------L~------4L------L------6L----Xl
(a)
gl ==
g2 ==
xl /20 - X2 + 1 ~ 0
(b)
xi 120 - Xl + 1 ~ 0
xi = {6.
11 S
Minimizing f(a) along the line defined by (d), we find the boundary point
X~ ={2.764, 1. 382}. To compute the direction S2' the following LP problem is
formulated at X2
13-+ max
(e)
5.528}
{Si' S2} { 2.764 +13 ~ 0
The solution is S~
={Si'
X3
2.764}
={ 1.382
+a
{-I.
O}
1.0
(j)
The value of a* is selected by minimizing the objective function along this line.
The result is Xr
new direction by
= {2.073.
S3
=-vr=-{ 4.146}
4.146
X4
={2.073}_a{4.146}
(g)
and ~ is computed by
2.073
4.146
(h)
X4
which is the optimal solution.
_{1.056}
1.056
(i)
116
2 Optimization Methods
c!>(A) ~ max
j = 1,... , nil
(2.145)
",
c!>(X, A) = f(X)+
L Ajgj(X)
j=l
(2.146)
(2.147)
It may be advantageous in some cases to solve the dual problem and then retrieve
the optimum primal variables X*. The motivation for such an approach is that in
many design problems only a few constraints are critical at the optimum. Therefore
only a few 'Aj. are nonzero.
The point (X*, A*) defines a saddle point of the Lagrangian function given by
(2.146). This will correspond to a maximum with respect to A and a minimum
with respect to X. Thus, we can define the Lagrangian in terms of A alone as the
dual function (2.147).
Since the dual function c!>(A) is a maximum with respect to A at the optimum,
the object is to fmd
maxc!>(A) = max minc!>(X, A)
)..
)..
(2.148)
117
Dual methods have been used extensively in linear programming to improve the
optimization efficiency. In nonlinear programming the dual formulation is
particularly attractive in cases where the primal problem is convex and
mathematically separable. If the primal problem and its dual are both convex, their
respective solutions satisfy the same optimality conditions. Both problems are
equivalent and their optimal values are equal, that is
cjI(A*) = JtX*)
(2.150)
Convex approximate problems will be discussed in Sect. 3.1.3. The objective and
constraint functions are said to be separable if each can be expressed as the sum of
functions of the individual design variables
f(X)
= L h(X
"
j )
j=l
gj(X) =
(2.151)
L" gjj(Xj )
=1... n,
j=l
The primal formulation does not benefit much from the separability. However, the
dual formulation does, because the Lagrangian function in this case
",,,
"
cjI(X. A)=
L h(X )+ L Aj L gjj(Xj )
j
j=l
j=l
(2.152)
j=l
"
L
j=l
Aj~
n:n[h(Xj )+
j=l
j= 1,...
",
L Ajgjj(Xj)]--+ffiaX
n,
(2.153)
The Lagrangian cjl(A) is therefore easy to calculate. Furthermore, the single variable
optimization problem has often a simple algebraic structure and it can be solved in
closed form, yielding thus an explicit dual function. It should be noted, however,
that because of the nonnegativity conditions that the dual variables must fulfil, a
direct solution of the dual problem from the stationary conditions is usually
difficult if not impossible.
118
2 Optimization Methods
An interesting property of the dual function is that its ftrst partial derivatives are
given by the primal constraint values. that is
acjl
aA.
= g .[X(A)]
J
(2.154)
z = xl' + xi + x; ~ min
1
1
gl =X2 + X2 -1 ~ 0
1
2
g2 =
xi + x; -1 ~ 0
where
Each one of the functions iloh and h can be minimized separately to get the
minimum of cjl(X. A) with respect to X
X1 --
119
'1114
Jl.l
X3 --
'11/4
Jl.2
max
A~O
~(A) = 5.828
Since the constraints A ~ 0 are not active, we have from (2.154)
= X; = 1.307
X; = 1.554
f(X)
= 5.828
That is
max ~(A)
=minJtX)
120
2 Optimization Methods
projecting the steepest descent direction for the objective function to the tangent
plane. Consequently, the new design will usually be infeasible and a series of
correction steps need to be executed to reach the feasible region. The step size
specification is arbitrary and the constraint correction process might be tedious.
Despite these drawbacks the method has been applied to several structural design
problems.
The reduced gradient method is based on a simple variable elimination technique
for equality constrained problems. Dependent and independent variables are
identified in the linearized subproblem and the dependent variables are eliminated
from it. The generalized reduced gradient (GRG) method is an extension of the
reduced gradient method to accommodate nonlinear inequality constraints. A search
direction is selected such that for any small move the current active constraints
remain active. If some active constraints are not satisfied due to nonlinearity of the
constraint functions, the Newton-Raphson procedure is used to return to the
constraint boundary. Thus, the GRG method is similar, in this respect, to the
gradient projection method. Although the method appears complicated, relative to
SUMT or other methods, the efficiency is often improved. However, the method
does have some drawbacks:
during the one-dimensional line search. If the problem is highly nonlinear, the
use of these iterations may become ill-conditioned and may not converge.
b. If there are many inequality constraints and inequalities are converted to
equalities by adding slack variables, the problem size might become large. This
problem can be overcome if only potential active constraints are considered.
c. The method tends to move from one constraint vertex to the next. If the
number of independent variables is large the convergence might be slow.
d. A feasible starting point must be selected. In addition, the design process
produces a sequence of infeasible designs. This feature is undesirable in many
applications.
Exercises
2.1 Find the minimum of the function
Exercises
121
vr
xi
xf
a. Calculate
sf = {Xl'
OJ. s~ = {O. X2 }
are conjugate with respect to a at the points Xl and X3 (Note: a = 1/2 H).
2.3 Given the constrained optimization problem
Xl
XloX2~0
xf = {4.
2}.
xI = (2.
6).
Xr = {4. OJ. Show graphically the gradient vectors of the constraints and the
objective function at the three points.
2.4 Given the general quadratic function
q = XT a X + XT b + c
Substitute X = X* + a S* and derive a formula for a* (corresponding to the
minimum value of q) in terms of a, b, X* and S*.
a. The formula derived in Exercise 2.4; plot/versus a for the two directions;
b. the golden section method;
c. the quadratic fitting; and
d. the cubic fitting.
122
2 Optimization Methods
2.6 Show that successive directions of steepest descent are orthogonal to one
another in the algorithm of (2.56).
2.7 Given the function
xi
a.
h.
/ = 25X~
+ 20xi - 2Xl
X~ = {I. I}.
X~ = {3, I}.
X2 ;
a.
Z= 3X l + 2X2 ~ max
2Xl+X2~IO
Xl+X2~8
Xl~
XltX2~0
h.
Z=X l +X2
min
2Xl+X2~4
Xl + 3X 2 ~ 6
XltX2~0
C.
Z=2Xl+X2~max
5X l + X 2 ~ 30
3Xl+X2~20
3X l +
2X2~
28
XltX2~0
Exercises
a.
123
X h X2
b.
c.
Z=2Xl +X2
max
-2Xl+X2~4
Xl +2X2~2
Xl>X2~0
Q.
Solve by the exterior penalty function for '1 = 1.0, '2 = 10.0. Show graphically
contours of'll = 1, 2 and 4 for each case.
b. Solve by the interior penalty function for '1 = 1.0,'2 = 0.1. Show graphically
contours of 'II = 5 ,10,15 for each case.
Q.
X I -X2 - 2
=0
124
2 Optimization Methods
2.14 Solve the problem of Exercise 2.11 by the method of feasible directions.
Perform two iterations, starting at the initial point
the directions of move in the space of Xl and X2
xf = {7.
Slate the dual problem in terms of A alone and solve for the optimum values of A,
Xl, and X 2 What are the optimum values of the primal and the dual objective
functions?
2.16 Given the problem
5
2
.
Z =-+---+mm
Xl
X2
Xl + X2 - 3 ~ 0
g2=4X l +X2-6~0
g3 Xl + 3X2 - 20 ~ 0
gl
Solve this problem using dual methods. Find the optimum values of the primal
and dual variables and the primal and dual objective functions. Draw the design
space for the primal problem and verify the solution.
3. Approximation Concepts
One of the main obstacles in the solution of optimal design problems is the high
computational cost required for solving large scale problems. Applications of
approximation concepts in structural optimization have been motivated by the
following characteristics of the design problem:
-
The problem size (number of variables and constraints) is usually large. Each
element involves at least one variable, and various failure modes under each of
several load conditions must be considered.
The constraints are usually implicit functions of the design variables. That is,
evaluation of the constraints for any given design involves solution of a set of
simultaneous equations. In addition, it is often required to calculate constraint
derivatives with respect to design variables.
In general, the solution of optimal design problems is iterative and consists of
repeated analyses followed by redesign steps. The number of redesigns (or
repeated analyses) is usually a function of the problem dimensionality.
For practical design problems each redesign involves extensive calculations and the
number of redesigns is large. Consequently, the total computational effort might
become prohibitive. Introduction of approximate models of the structural behavior
in terms of the design variables is intended to reduce the computational cost and
allows the solution of practical design problems. It is recognized that only methods
which do not involve many implicit analyses are suitable for practical design
applications. In structural optimization, the analysis task will require most of the
computational effort. As a result, approximation techniques used to solve a
structural optimization problem might affect the overall computational cost more
than the choice of the optimization method.
In this chapter, approximation concepts for structural optimization are discussed.
In Sect. 3.1 methods for calculating derivatives of displacements with respect to
design variables are first introduced. These derivatives are needed for effective
approximations of the constraints and efficient solution of the optimization
problem. Intermediate variables, often used in structural optimization, and methods
based on sequential approximations are then discussed. In Sect. 3.2 various
approximate behavior models for evaluation of displacements, stresses and forces in
terms of the design variables are presented.
126
3 Approximation Concepts
(3.1)
(3.2)
This can be done by a simple recursion algorithm. The displacements are then
computed by forward and backward substitutions.
For a problem with n design variables X j (i = 1..... n). finite difference
derivative calculations of the displacements with respect to design variables require
to repeat the analysis for (n+I) different stiffness matrices. However. the
derivatives can be calculated analytically in more efficient ways. and the large
number of analyses associated with fmite difference calculations can be avoided. In
this section three alternative methods for such analytical calculations of derivatives
are discussed. Further developments in this area are reviewed by Haftka and
Adelman [47].
Direct Method. In this approach the displacements r are expressed in terms of
the independent design variables X by (3.1). Implicit differentiation of (3.1) with
respect to Xj yields
(3.4)
in which
oR! IOXj}
.
{
ax =
:
, oR",loXj
or! IOXj}
.
ax. ={ :
, or", loXj
oR
or
127
(3.5)
oK!",
oXj
oK
oK"""
(3.6)
oXj
ar ={ar
ax
aXl
'
ar}
ax"
..
(3.7)
(3.8)
arrax.
arJ _
T _
ar
ax
-=Vr-I-
ax
(3.10)
where Ij is a vector having unit value at the jth location and zeros elsewhere.
In many problems where the load vector R is independent of the design variables
j =0, and (3.4) is reduced to
aR/aX
ar
aX
aK
aX
K-=--r
j
(3.11)
128
3 Approximation Concepts
Ir
K - 1 and
(3.12)
~j is defmed as
K~j=Ij
or!
- ' =Vr!=~~ V
oX
'
(3.15)
where matrix V is defined by (3.8). Once the system (3.13) is solved, the adjointvariable vector ~j is repeatedly used in (3.14) for all variables. Since (3.13) and
(3.1) have the same coefficient matrix K, again only forward and backward
substitutions are needed to solve for ~j
Virtual-Load Method. This method is also based on the assumption that the
dependent displacements r are expressed in terms of the independent design
variables X. To calculate derivatives of displacements with respect to design
variables, any desired displacement rj is expressed as
(3.16)
where Qj is a virtual-load vector, having unit value at the jth location and zeros
elsewhere. Differentiation of (3.16) with respect to X gives
or!
oX
'
Or
'ax
(3.17)
K r?
=Qj
129
(3.18)
Vr!
J
=(r(2)TK
2!.
J
al(
(3.19)
Substimting (3.9) into (3.19), the following expression for the derivatives of rj is
obtained
VrJ
=(r?)T V
(3.20)
i=I, ... ,n
(3.21)
In the calculations of aK/aXi only the stiffness matrix of the ith member, K i , can
be considered and (3.21) becomes
_J
i= 1, ... ,n
(3.22)
The vectors r? and r can accordingly be reduced to contain only those degrees of
freedom associated with the ith element.
In many strucbJral design problems Ki is a linear function of Xi. Thus
i =l, ... ,n
(3.23)
i=I, ... ,n
(3.24)
Comparison or Methods. Arora and Haug [3] analyzed the different methods
to design sensitivity analysis. It has been shown that the virtual-load method can
be derived from either the direct method or the adjoint-variable method.
130
3 Approximation Concepts
;j is calculated by
I j =Qj
(3.25)
K ;j= Qj
(3.26)
Comparing (3.26) and (3.18). it can be seen that the adjoint-variable vector
identical to the virtual-displacement vector
rf
;j =rf
;j is
(3.27)
ar/ax
~=K-IV
ax
(3.29)
Vr! =Q~K-IV
J
(3.30)
Based on (3.18) and on the symmetry of K. this equation can be written in the
fonn of (3.20). obtained by the virtual-load method.
A summary of the various methods for calculating the derivatives of the
displacements is given in Table 3.1. It has been shown that the three design
sensitivity analysis methods give the same results. However. as noted by Arora and
Haug [3]. there are some differences in generality and efficiency of the individual
methods. The adjoint-variable and direct methods are more general than the virtualload method and can be extended to include other behavior functions. As to
efficiency considerations. both the adjoint-variable and the virtual-load methods are
superior to the direct method in cases where derivatives of a limited number of
displacements must be calculated. Let J be the number of displacements to be
considered. The adjoint-variable method then requires calculation of J adjoint
vectors;j by (3.13). and the virtual-load method requires calculation of J virtualdisplacement vectors
ar/ax;
that must
131
Unknowns
Direct
ar/ax
Adjoint-variable
Virtual-load
Number of
unknown vectors
Equations
n nL
K~=V
ax
T ar
VrJT =1J ax
K ~j= I j
VrJ =~]V
K rf =Q j
~j
r~
J
VrJ =(rflV
J
(3.9)
(3.10)
(3.13)
(3.15)
(3.18)
(3.20)
Example 3.1. Consider the three bar truss shown in Fig. l.ll. The equilibrium
equations are
]=
E [0.707X1
0.707X1 + Xz r
100
{14.14}
14.14
(a)
132
3 Approximation Concepts
X -T
={1.0,
r -T
1.0}
={0.0666,
0.0276}
art
art
axl aX2
orax
(b)
-=
or;
or;
axl ax2
Direct method. The matrix V is first calculated by (3.8)
300
0.707
0]
1. 707
orax =-
or- [-0.0666
ax = -0.0114
[14.14
0]
5.86 8.28
0]
-0.0162
(d)
(e)
g}
~07 1. ~07]~l =
300 [0.
= {0.00471,
O}
~~
={O,
0.00195}
(g)
133
(h)
and
ar/ax is given by
ar
[-0.0666
ax
=(Vrl Vr2) =
T
-0.0114]
-0.0162
(i)
Q[ = (1.
Q~
o)
= (0.
(j)
I)
300
[ 0.707
0] Q
1. 707 r l
{I}
300 [
0.707
0
0]
1. 707 r
{O}1
(k)
=(rIQ)TV ={-o.0666.
O}
(m)
Vr;T
=(rlV ={-o.0114.
-0.0162}
Y;
=Y,{X;)
(3.31)
Y;=Xt'
(3.32)
134
3 Approximation Concepts
y=I
(3.33)
The reason for this is that displacement and stress constraints for detenninate
structures are often linear functions of the reciprocal variables. For statically
indetenninate structures, the use of these variables still proves to be a useful device
to obtain better approximations. Intennediate variables are usually most effective
for some homogeneous functions as will be shown later in Sect. 3.2.3.
One disadvantage of the reciprocal approximation is that it becomes infinite if
any Xj is zero. This difficulty can be overcome by the simple transfonnation [48]
1
Xj +OXj
y=--I
(3.34)
where the values of ~Xj are typically small compared to representative values of
the corresponding X/so
Consider the first order Taylor series expansion of a constraint function g in
tenns of the design variables Xj , denoted as a direct approximation gD
gD == g + L,"
-og (X. -
j=! oXj
x.)
I
(3.35)
gR == g +
Og. )]
[(Xj
"o
Lj=! -(
.) (r; - r; ) = g + L
or;
j=!
oXj
. . "
(3.36)
135
The sign of each term in the sum is determined by the sign of the ratio
(ag"' axi ) , Xi which is also the sign of the product Xi (ag"' aXi ). Since the
constraint is expressed as g ~ 0, a more conservative approximation is the one
which is more positive. It is possible, therefore, to create a conservative
approximation gc which includes the more positive term for each design variable
L
II
gc =g"
(Xi
i=1
agO
(X-X)
'
,
aXi
(3.38)
where
a~ll
, X:
-'
Xi
~o
if
Xi
agO
aXi
if
Xi
agO <0
aXi
(3.39)
It should be noted that this approach does not ensure that the approximate
constraint is conservative with respect to the true constraint. The conservative
approximation is only more conservative than both the linear and the reciprocal
approximations. The conservative approximation has the advantage of being
convex approximation [30]. However, it has been found that this approximation
tends to be less accurate than either the direct or the reciprocal approximation.
The method of moving asymptotes (MMA), proposed by Svanberg [134], is
intended to introduce more accurate approximations. Using this method the
intermediate variables are defined such that the degree of convexity, and hence
conservativeness, of the approximation can be adjusted. Instead of using direct and
reciprocal variables, the method employs the intermediate variables
1
y=-,
or
Xi-L i
(3.40)
y=-,
Ui-Xi
where Li and U i are specified parameters that may be changed. Based on this
transformation, the moving asymptotes approximation gM is formulated as
-. + " -L
a" (
gM =g
~
,=1 ax.'
A.)
'+-P-'X-L., ,
(X.
U, -X,
(3.41)
136
3 Approximation Concepts
where
={ (Ui -
(x.
&
Pi
={O
_.
Xi )2
-(Xi -Li)
8 =8 -
a8 laxi >0
if
a8 laxi ~O
if
a8laxi ~O
if
a8 laxi <0
~
a8
~ -
. ax
&=1
if
&
(Xi
U-X~
&
&
Pi)
+---i.........X-.
&
&
(3.42)
(3.43)
Since all the coefficients in (3.41) are non-negative, the approximations for 8 are
convex functions. The approximated functions are driven by the selected values for
the parameters Li and Ui which act as asymptotes.
The moving asymptotes approximation is general; the direct approximation and
the conservative approximation can be viewed as the following special cases:
- For Li -+_00, Ui -+00, no intermediate variables are considered, and the direct
formulation (3.35) is obtained.
- For Li 0, Ui -+00 , the conservative formulation (3.38) is obtained.
Other values of Li and Ui are acceptable, and these values may even be modified
during the solution process. However, it is not at all straight forward to find
suitable values for the asymptotes. To avoid the possibility of any unexpected
division by zero, move limits
xl'
xf and XiU
4 < xf and
< Ui . The closer Li and Ui are chosen to X; , the more curvature is given to
the approximate function, and the more conservative becomes the approximation of
the original function. Choosing Li and Ui far away from Xi , then the approximate
function becomes close to linear.
Both formulations (3.38) and (3.41) are based on first order convex
approximations that attempt to simulate curvature of the functions. The latter
method (MMA) offers more flexibility through the moving asymptotes Li and Ui.
Example 3.2. To illustrate the various approximations, consider the constraints
xl 120 - X2 + 1 ~ 0
82 =.xi 120-X1 +1~0
81 =.
obtained.
137
(a)
gl
(b)
(c)
The various approximations are demonstrated in Fig. 3.1. It can be observed that
xt,
10
8
6
10
138
3 Approximation Concepts
L 1)J:fik + L YZdZk
1ft
1ft
k=O
k=O
(3.45)
In this equation. fu andflk are values offl andlz at the preselected points Xu and
Xlk (k =O... m). respectively. where
L 1)kXU
(3.46)
L YZkX
(3.47)
1ft
Xl =
k=O
1ft
Xz =
2k
k=O
L 1)k = L Y
1ft
1ft
k=O
k=O
2k
= 1.2
fll
Xl
10
Fig. 3.2.
X II
X 12
X13
Linearized segments.
X 14
=1
(3.48)
= 0.1.2... m
139
We require that for every i (i = 1,2) at most two adjacent YiA: be positive. If, for
example, Yll and YIZare nonzero with all other Yu; zero then the value of Xl is
given by
(3.49)
with
Yll + Yl2 = 1
(3.50)
The two adjacent nonzero Y's identify the segment where the final solution lies.
The number (m) of segments determines the degree of approximation; the larger the
m the closer will be the approximation to the original function.
" dj*
L
-(Xj-X;)~min
j=l aXj
" agj*
*
g*.+ L -(X-X-)<O
J
j=l
ax.
1-
(3.51)
j = 1... ng
The LP problem can be solved repeatedly, redefming X* each time as the optimal
solution of the preceding problem. This procedure, however, will not always
converge to the optimum. Problems which may arise during the solution process
include:
140
3 Approximation Concepts
(a)
(h)
141
First LP solution
with move limits
I
;1
True
/
OPtimu";......-//
First LP solution
/
,//
without move limits I / / /
//
)././
/,
Fig. 3.4.
Move limits.
(3.52)
where AXL and AXu, called the move limits, are suitably chosen vectors of
positive constants (see Fig. 3.4). To solve a problem we first choose a starting
point X and linearize the objective function and the constraints in the
neighborhood of X. We solve the linearized problem (3.51) and (3.52) and redefine
X as the optimum solution to the preceding problem. The nonlinear functions are
relinearized about the new X and the process is repeated until either no significant
improvement occurs in the solution or successive solutions start to oscillate
between the vertices of the feasible region. In the latter event we may reduce the
values of the bounds AXL and ax U and continue.
For computational efficiency it is desirable to choose large values for the move
limits, so that the imposed limits will not slow convergence. However, these
bounds should be gradually shrunk as the design approaches the optimum. One
reason for the need to shrink the move limits is that the accuracy of the
approximations is required to be higher when we get close to the optimum. The
move limits are typically reduced by ten to fifty percent of their previous values
until convergence. Convergence is assumed if the change in the optimal Z for two
successive LP subproblems is smaller than a desired value and the nonlinear
constraints are satisfied within a desired tolerance. The method of approximate
programming is applicable to nonconvex problems and produces feasible or nearly
feasible intermediate solutions with good accuracy. The method differs from the
cutting-plane method in that there is no link between subsequent LP problems.
142
3 Approximation Concepts
Objective function
contours
1 iO'!!/~/~/;;:/~=,---_
X .,.'
X;
~ optimum
-\
xi ~ min
gl == Xf /20 - X2 + 1 ~ 0
g2 ==
xi 120 - Xl + 1~ 0
Z=-(Xl 2
+X2
.
+X2X2)~mm
143
for the fllSt three iterations are shown in Table 3.2 and in Fig. 3.5. Convergence is
fast although the starting point is far from the optimum and no move limits have
been assumed.
q
1
3
4
(6.3)
{O.378 -0.573}
{O.924 1.028}
{1.055 1.055}
(3.53)
where II~II is the length of the search direction and !; is a specified small positive
number. Substituting
1I2
II~XII= (~ Mf J
(3.54)
It can be noted (Fig. 3.6) that the new design is required to be in a hypersphere of
radius!; with origin at the current poinL Thus. the approximate subproblem to be
solved at each iteration is: find ~X such that
=l ... n,
(3.56)
144
3 Approximation Concepts
Feasible region
vr
where
and Vg; are computed at X, and the last equation is a quadratic step
size constraint. A solution of this problem may not exist if I; is too small and the
current design X is infeasible.
The problem (3.56) can be transformed into the following QP [2]: find aX such
that
Z =f + VrTaX + aXT aX ~ min
j =
t .... n,
(3.57)
'
(3.58)
j = 1..... n,
where (Xi and (Xij are defined by (3.39). Fleury and Braibant [30] employed these
approximations in the convex linearization method (CONLIN). It has been found
that in some cases the convex approximations scheme used in CONLIN might not
145
!=!-* +
Li=1"
R.)
al'* ((X.-.:
_'J_
aXi
" ag~
_ -* ~)
gj =gj + .J -
i=1 aXi
Ui
Xi
+-P-~- ~min
((X"
I)
Ui -Xi
Xi
Li
R..) - 0
+--- <
PI)
Xi -Li
(3.59)
j
= l. .... n
,l
g;
146
3 Approximation Concepts
(3.60)
(3.61)
where K is the stiffness matrix for any assumed design, and r a is a vector of
approximate displacements computed for this design. If r a are the exact
displacements, then Ra =R (the actual given loading). That is, the approximate
displacements r a can be viewed as the exact displacements for Ra. The difference
between the fictitious loading and the real loading
(3.62)
indicates the discrepancy in satisfying the original equilibrium conditions due to
the approximate displacements.
The problem considered in this section can be stated as follows:
147
Given an initial design variables vector X*, the corresponding stiffness matrix
K*, and the displacements r*, computed by the equilibrium equations (1.26)
K* r* = R
(3.63)
The elements of the load vector R are often assumed to be independent of the
design variables and the stiffness matrix K* is usually given from the initial
analysis in the decomposed form
(3.64)
x =X* +.1.X
(3.65)
K = K* +.1.K
(3.66)
where.1K is the change in the stiffness matrix due to the change .1X.
c. The object is to find efficient and high quality approximations of the modified
displacements r due to various changes in the design variables .1.X, without
solving the modified analysis equations
K r = (K* + .1.K)r = R
(3.67)
In this formulation, the elements of the stiffness matrix are not restricted to certain
forms and can be general functions of the design variables. That is, the design
variables X may represent coordinates of joints, the structural shape, geometry,
members' cross sections, etc. Once the displacements are evaluated, the stresses can
readily be determined explicitly by (1.29). Thus the presented approximations of r
are intended only to replace the set of implicit analysis equations (3.67).
Local Approximations: Series Expansion. A common approach is to
consider the first terms of a series expansion, to obtain the approximate
displacements r a
(3.68)
The Taylor series expansion is one of the most commonly used approximations in
structural optimization. The first three terms, obtained by expanding r about X*,
are given by
rl =r*
r2 = Vr;.1.X
r3j
(3.69)
*
= lJ2.1.X H j .1.X
148
3 Approximation Concepts
where the displacements r*, the matrix of flfSt derivatives Vr; ,and the matrix of
second derivatives H; ,are computed at X*. The scalar r3j is the jth component of
vector r3. To reduce the computational effort, linear approximations are often used.
These require evaluation of the flfSt derivatives which can readily be calculated by
the methods discussed in Sect. 3.1.1. It should be noted, however, that the fll'storder approximations may be insufficient in many cases and second-order models
might be needed. The latter methods can be divided into two groups:
K* r = R - AK r
(3.70)
K* r(k+l) = R - AK r(k)
(3.71)
where r(k+l) is the value of r after the kth cycle, and assuming the initial value r(l)=
r*, the following binomial series expansion is obtained
ra = (1- B + B2 - ... ) r*
(3.72)
B == K*l AK
(3.73)
That is, the flfSt three terms of the series are given by
rl =r*
r2=-B r*
r3= B2r*
(3.74)
K*r2
=-L\K r*
149
(3.75)
(3.80)
where IIBII is the norm of B. It will be shown later is Sect. 3.2.3 that the terms of
the Binomial series (3.74) are equivalent to those of the Taylor series (3.69) for
homogeneous displacement functions.
Series expansions are local approximations, based on information of a single
design. As a result, the quality of the approximations might be sufftcient only for
a limited region. Several methods have been proposed to improve the series
convergence. These include the Jacobi iteration, block Gauss-Seidel iteration,
dynamic acceleration methods and scaling of the initial design [62, 67]. These
means may considerably improve the results with a moderate computational effort.
Improved displacement approximations, based on combining local and global
approximations, will be discussed in Sect. 3.2.2.
Global Approximations: The Reduced Basis Method. In this approach
[34] it is assumed that the displacement vector r of a new design can be
approximated by a linear combination of s linearly independent basis vectors rh
r2, ... ,r.. of previously analyzed basis designs (where s is assumed to be much
smaller than the number of degrees of freedom m), that is
(3.81)
or in matrix form
(3.82)
150
3 Approximation Concepts
where
(3.83)
rB
Y = rj R
axm mx!
(3.84)
KR Y = RR
axa ax!
ax,
(3.86)
151
series might even diverge. Scaling of the initial design can greatly improve the
quality of the approximations. Scaling of the initial stiffness matrix K is defined
by [67]
K =)JK.
(3.87)
where J.1 is a positive scalar multiplier. From (3.63), (3.67) and (3.87) it is clear
that the exact displacements after scaling can be calculated directly by
(3.88)
Note that (3.87) does not require linear dependence of K on X. Furthermore, in
many cases, where the elements of K are nonlinear functions of X, the matrix )JK.
does not correspond to an actual design. That is, the matrix K computed by (3.87)
does not have the usual physical meaning. Scaling of the initial stiffness matrix
will improve the quality of the approximations, if the known displacements J.1- 1 r
[see (3.88)] provide better initial data than the original displacements r.
To obtain the displacement approximations, the modified stiffness matrix K [see
(3.66)] is expressed in terms of J.1 by (Fig. 3.7)
(3.89)
That is, if an initial design )JK* is assumed instead of K*, the modified stiffness
matrix K is expressed in terms of the corresponding changes in the stiffness matrix
AK" instead of AK . From (3.89), AK" is given by
(3.90)
Consider the recurrence relation (3.71), with J.l.K, AK~ and WI r assumed
instead of K*, .11K and r, respectively. The resulting series is [see (3.72)]
(3.91)
where B~ is given by
I-J.1
J.1
J.1
B =-- 1+- B
"
(3.92)
(3.93)
152
3 Approximation Concepts
in which m denotes the order of matrix B. A major drawback of using the criterion
(3.93) is that the elements of matrix B must be calculated. Since this opemtion
involves much computational effort, we may use an alternative criterion that
minimizes the Euclidean norm of the second term in (3.91) , that is [85],
UB" r*U ~ min
(3.94)
Substituting B" from (3.92) into (3.94), differentiating and setting the resUlt equal
to zero, yields
a
J.l=-
(3.95)
where
L ('i;-ru)Z
III
a=
;=1
L (rJ-rurz;)
(3.96)
III
b=
;=1
ru are the given elements of rl = r*, and ruare the computed elements of rz [see
(3.74)]. The effect of J.l on the quality of the approximations will be demonstmted
later in this section.
Combined First-Order Approximations and Scaling. Because of
efficiency considemtions, frrst-order series approximations are often used. If only
two terms of the binomial series (3.72) are considered, the following first-order
approximations are obtained
r .. ={I-B)r*
(3.97)
153
(3.98)
Neglecting the second order tenn 11K I1r, premultiplying by K*1 and substituting
(3.63) and (3.73) gives (3.97).
It will be shown now that by combining two types of scaling:
-
(3.100)
where Q is a scalar. By evaluating ra for any given J.1 by means of (3.99), the
latter displacements can then be scaled by (3.100) such that the fInal displacements
are improved. Substituting (3.99) into (3.100) yields
(3.101)
Thus, each evaluation of the displacements involves the following two steps:
Y2 =Q 1.1"2
and substituting (3.102) into (3.101) yields
(3.102)
154
3 Approximation Concepts
(3.103)
where
(3.104)
(3.105)
That is, (3.101) which is based on combining the two types of scaling, is
equivalent to the two-terms expression of the reduced basis method (3.103). It
should be noted that ~ and Q can be determined uniquely for any y by (3.102).
The above combination of a first-order series expansion and the reduced basis
method can be generalized to any number of terms in the series, as will be shown
subsequently.
Combined Series Expansion and Reduced Basis. The drawbacks of
series expansion and the reduced basis method motivated combination of the two
approaches to achieve an improved solution procedure [80]. In this procedure, the
computed terms of a series expansion are used as high quality basis vectors in a
reduced basis expression. The advantage is that the efficiency of local (series
expansion) approximations and the improved quality of global (reduced basis)
approximations are combined to obtain an effective solution procedure. The
solution process involves the following steps:
b. The basis vectors rj of a series expansion [i.e. (3.69) or (3.74)] are calculated,
(3.106)
It has been noted earlier that if r a is the vector of the exact displacements, then
AR =O. Thus, AR can be used to evaluate the quality of the approximations. Let
us derme the common measure of smallness of AR(y) by the quadratic form
(3.107)
Substituting (3.106) into (3.107) , differentiating with respect to y and setting the
result equal to zero, we obtain the following linear equations in the form of (3.86)
a y =b
155
(3.108)
a =(KrBl(KrB)
(3.109)
This alternative method for determining y can be used instead of (3.85) in step c of
the solution process. The two criteria have been compared using several numerical
examples [80]. It has been found that although the method (3.109) provides smaller
q(y) values, better results might be obtained by the method (3.85).
Another method, that combines the reduced basis approach and the first-order
Taylor series approximations of the displacements, has been proposed by Noor and
Lowder [101]. The assumed basis vectors are rl
= r*
and rj+l
=or" fiJX j
(j =I ... ,n) . These vectors are normalized to overcome numerical roundoff errors.
An advantage of this choice is that it contains the sensitivity analysis vectors.
However, many of the computed derivatives may correspond to nonactive
constraints and are not needed for the optimization process. In addition, this
method is not efficient in problems with a large number of design variables, where
n derivative vectors are to be computed and an (n+l)x(n+l) system of equations
must be solved. Recently, Noor and Whitworth [102] proposed to express the
modified analysis equations in terms of a single parameter, and to choose the basis
vectors as the various-order derivatives of the displacements with respect to the
parameter, evaluated at the original design. It has been found that a small number
of basis vectors is often sufficient to obtain good results.
Computational Considerations. The quality of the results and the efficiency
of the calculations are two conflicting factors that should be considered in selecting
an approximate reanalysis model. That is, better approximations are often achieved
at the expense of more computational effort. In this subsection, some
computational considerations associated with the presented approximations are
discussed. Consider first the two methods of calculating the basis vectors, namely
the Taylor series and the binomial series.
Assuming the common fmt-order Taylor series expansion, once the matrix Vr;
is available each redesign involves only calculation of the product Vr;AX . This
is probably the most efficient reanalysis method. However, it has been noted that
the quality of the results might be insufficient for large changes in the design
variables. In addition, the second-order Taylor series expansion is usually not
practicable owing to the large computational effort involved in calculation of the
second-order derivative matrices H j An exception is the common case of
homogeneous displacement functions, discussed in Sect. 3.2.3, where the Taylor
series and the binomial series become equivalent.
The advantage of using the binomial series is that, unlike the Taylor series,
calculation of derivatives is not required. This makes the method more attractive in
general applications where derivatives are not available. Calculation of each term of
the binomial series involves only forward and backward substitutions, if K* is
156
3 Approximation Concepts
given in the decomposed form of (3.64). Thus. the second order terms can readily
be calcuJated. In the case of fll'St-order approximations. determination of only Br
must be repeated for each trial design. This requires calculation of a single vector
by forward and backward substitutions.
As to the selection of the number of basis vectors to be considered in the
combined approximations. it has been noted that. in general. second-order
approximations (three basis vectors) provide better results than first order
approximations (two basis vectors).
To evaluate the computational effort involved in the combined approximations.
compared with conventional local methods. assume the fll'St-order approximations.
The series approximations require only calculation of the second basis vector
[(3.69) or (3.74)]. The combined approximations require. in addition. introduction
of the modified stiffness matrix K. calcu1ation of the products given by (3.85) or
(3.109). determination of y by solving the set of (2 x 2) equations (3.86) or
(3.108) and multiplication of the two basis vectors by y. Certainly. these
operations increase the computational cost. However. the result is often
considerably better approximations. particularly in cases of large changes in the
design variables. Tttat is. high quality approximations can be obtained in cases
where the local series approximations provide meaningless results. Consequently.
exact analyses which involve more computational effort are not required in cases
where the local approximations provide insufficient results.
Finally. the local approximations may be viewed as a special case of the
combined approximations where y 1.0 is selected. An additional advantage of
the combined approximations is that the errors involved in the approximations can
be evaluated by AR and q.
Example 3.4. Consider the truss shown in Fig. 3.8 with ten cross-sectional
area variables Xj (i = 1... 10). subjected to two loading conditions. Assuming
E=L =1.0 and the initial design X=1.0. the corresponding displacements are
Loading A:
LoadingB:
'1
360
(I)
(2)
'4
t 50
(b)
(a)
8.
t 50
36{
h50
(2)
Loading A. b. Loading B.
L50
157
Method
Exact
(3.109)
(3.85)
DisElacements
19.52 53.17 23.49 115.5 -26.51 120.52 -20.48 57.54
19.51 53.17 23.48 115.5 -26.52 120.51 -20.48 57.54
19.51 53.17 23.48 115.5 -26.52 120.52 -20.48 57.54
Exact
(3.109)
~3.85~
xT=
The resulting displacements obtained by the methods (3.85) and (3.109) are
summarized in Table 3.3. It can be observed that excellent results have been
obtained for these large changes (up to 900%) in the design variables. The high
quality of the results can be explained by the scaling procedure. That is, the
modified design is relatively close to the scaling line J.1K*.
Considering the optimal designs [62]
Loading A:
LoadingB:
xT =
results obtained for these very large changes in the cross-sections (up to +905%
and -90% simultaneously) by various methods are shown in Table 3.4. It can be
observed that solution by the Taylor series (3.69) is meaningless. SpecifIcally,
since the series diverges the results obtained by the second-order Taylor series
approximations (three terms of the series, or three basis vectors) are worse than
those obtained by the first-order approximations (two terms of the series).
Relatively good results have been obtained by both combined methods (3.85) and
(3.109). Considering three terms of the series (second-order approximations), the
quality of the results is further improved. It has been noted [80] that the results
obtained by the method (3.85) are better than those obtained by (3.109) even in
cases of larger q values.
Example 3.5. Consider the thirteen-bar truss with the initial geometry and
loading shown in Fig. 3.9a. The modulus of elasticity is 10,000, the initial cross
sections are X* = 1.0, and the unknowns are the horizontal (to the right) and the
vertical (upward) displacements in joints B, C, D, F, G and H, respectively.
Assume first the following modifIed cross-sectional areas
158
3 Approximation Concepts
Table 3.4. First- and second-order approximations, ten-bar truss, optimal solutions.
Exact
(3.69)
2
3
2
3
2
3
(3.109)
(3.85)
25.0
-1078
7095
24.5
21.8
25.7
22.6
75.0
-2201
13620
68.8
67.4
75.0
70.1
Displacements
40.5 184.4 -50.0
-1223 -4831 1218
8615 31588 -8649
31.6 160.7 -40.1
35.6 172.5 -47.6
34.4 175.3 -43.8
37.7 181.4 -50.6
75.0
-2189
13729
66.0
63.5
69.6
66.1
38.1
-1111
7189
31.9
31.0
33.6
32.3
175.0
-5031
33615
153.3
156.8
161.7
164.0
200.0
-4873
31736
171.1
185.4
186.7
195.0
-50.0 200.0
1406 -5125
-1106133887
-40.1 172.6
-47.2 180.5
-42.3 182.4
-50.0 189.0
-25.0
1219
-8739
-22.7
-24.6
-24.7
-26.0
75.0
-2272
13814
77.0
75.9
84.0
79.1
-25.0 75.0
1360 -2469
-10949 15209
-22.4 78.2
-24.5 75.5
-23.4 82.5
-26.0 78.6
Results obtained for these large changes in cross sections (up to 900%) by various
approximate methods are shown in Table 3.5. It can be observed that:
- Results obtained by the local flrst- and second-order series expansion [(3.69) or
(3.74), which are equivalent in this case] are meaningless. Once again, the series
diverges due to the large changes in the design.
5.0
15.0
13
'''''
60
80
80
'''''
60
(a)
~,
y
(b)
""
159
Table 3.S. First- and second-order approximations, thirteen-bar truss, modified cross
sections.
Displacement
Nwnber
1
2
3
4
5
6
7
8
9
10
11
12
(3.69) or (3.74)
2 Terms 3 Terms
-3.90
35.7
-2.12
19.3
-11.85 103.8
26.3
-3.10
-20.70
177.2
-3.30
27.4
-4.00
36.5
2.10
-19.6
-11.70 102.5
3.10
-26.4
-21.00 178.4
3.30
-27.5
(3.109)
2 Terms 3 Terms
0.051
0.048
0.028 0.026
0.178 0.167
0.050 0.047
0.338 0.319
0.058 0.055
0.054 0.051
-.029
-.027
0.173 0.162
-.050
-.046
0.348 0.329
-.059
-.056
(3.85)
2 Terms 3 Terms
0.048 0.048
0.026 0.026
0.168 0.167
0.047 0.047
0.320 0.319
0.055 0.055
0.050 0.051
-.027
-.027
0.162 0.162
-.047
-.047
0.329 0.329
-.056
-.056
Exact
Method
0.048
0.026
0.167
0.047
0.319
0.055
0.051
-.027
0.162
-.047
0.329
-.056
- Very good results have been obtained by the first- and the second-order
approximations for both methods (3.85) and (3.109).
Consider the modified cross-sectional areas X as in the previous case and a
single geometric variable Y representing the span (Fig. 3.9b), with the optimal
value Y = 153 [85]. Results obtained for these large changes in both the geometry
(155% in Y) and the cross-sections (up to 900%) by various approximations are
shown in Table 3.6. Improved results have been obtained for the second-order
approximations (three terms) by both methods (3.85) and (3.109). Specifically,
assuming three terms for the method (3.85), the errors in most displacements do
not exceed 5%.
Table 3.6. First- and second-order approximations, thirteen-bar truss, modified
geometry and cross sections.
Displacement Exact
Nwnber
Method
1
0.0098
2
0.0101
3
0.0400
4
0.0162
5
0.0811
6
0.0134
7
0.0114
8
-.0101
0.0360
9
10
-.0166
11
0.0902
12
-.0133
(3.109)
Error
2 Terms 3 Terms
~%l
0.0043 0.0105
7
0.0059 0.0091
10
0.0230 0.0364
9
0.0105 0.0142
12
10
0.0518 0.0726
0.0095 0.0114
15
0.0061 0.0108
5
-.0063
-.0089
12
0.0185 0.0339
6
-.0104
-.0141
15
0.0603 0.0817
9
-.0102
-.0120
10
(3.85)
Error
2 Terms 3 Terms
~%l
0.0089
14
0.0112
0.0084
0.0099
2
0.0378
0.0396
1
0.0146
4
0.0155
0.0788
0.0795
2
0.0140
0.0127
5
0.0108
0.0117
3
-.0088
-.0096
5
0.0330
0.0367
2
-.0154
7
-.0145
0.0895
1
0.0879
0
-.0133
-.0147
160
3 Approximation Concepts
(3.110)
(3.113)
and substituting (3.111) into (3.113), the latter approximations for homogeneous
displacement functions become
(3.114)
Intermediate Variables. Assume intermediate variables of the form Y; = X;'"
[see (3.32)]. The resulting displacements are homogeneous functions of degree nlm
in V. Therefore, the rust-order Taylor series expansion (3.114) is
r"
(3.115)
For any given n, we can choose the value of m such that the approximations are
improved. Considering the common reciprocal cross-sectional variables Y; = l/X;
where n=m=-l, then (3.115) becomes
(3.116)
Since the displacements in this case are homogeneous functions of degree 1 in V,
we have from (3.112)
161
(3.117)
That is. the approximations (3.116) are exact along the scaling line Y = J.1Y*.
This illustrates the advantage of using the reciprocal variables for approximations
near the scaling line in structures with cross-sectional variables.
Combined Taylor Series and Scaling. For a point X along the scaling line
(3.118)
X = J.1X*
the displacements and displacement derivatives are [see (3.110) and (3.112)]
r = J.1" r*
Vrx = J.1II-IVrx
(3.119)
(3.121)
1
1
a=ST-=P-=PY
(3.122)
(3.123)
a=P Y
(3.124)
162
3 Approximation Concepts
This equation is based on the relations (aT"' alj)Y" = 0 (i = 1, ... , n), which are
true for a general statically indeterminate structure. Based on (3.125), it can be
seen that the approximations (3.116) and (3.123) are equivalent In addition, it can
be shown that (3.124) is equivalent to the first-order Taylor series expansion in the
reciprocal variables. That is, the first-order Taylor series approximations in the
nvrx"_
- - K"-lnK"
v x r"
(3.127)
(3.131)
order Taylor series and the first-order binomial series are equivalent.
163
Example 3.6. Consider the three-bar truss shown in Fig. 1.11 with the
following single stress constraint (Fig. 3.10)
(a)
=14.14
0";
*T
VO"x
=(-11.714. -2.426)
(b)
Assuming the fIrst-order Taylor series expansion about X, the approximate stress
constraint is
(c)
The intersection of the scaling line X = IlX* through X with the constraint
surface 0"1 = 20 is at the point X**T = (0.707, 0.707). At this point
b ').\)
.,)
0\
Xz
2.0
~)
\
\
()"\
d-
01
v--"""
\
\~ \ \
b ').\) \
1.6
I
\
I
I
\
\
\
1.2
< 20
\
\
0.8
r-
(Y") <20
\
\
\ \
\ \
\
(Y') =
0.4
0.2
0.4
1.2
0"1 ~
20 .
164
3 Approximation Concepts
a~* = 20.0
(d)
(e)
It can be observed from Fig. 3.10 that both approximate constraints (c) and (e)
represent parallel lines in the design space. Also, the line (e) is tangent to the
constraint surface al =20 at the intersection point X**.
Repeating the above procedure for the reciprocal design variables Y, we obtain
for both y*T = {1.0, 1.0} and y**T = {1.414. 1.414}
al (y*) = a 1(y**) = 11.714l} + 2.426Y2:5; 20
if)
(3.133)
in which X* is a given initial design, dX* is a given direction vector in the design
space, and the variable ex determines the step size. Since only a single variable is
involved, approximations along a line require much less computations. In this
section, local (series expansion) and global (polynomial fitting) approximations
along a line are presented. Other methods are discussed elsewhere [65].
Dependence of K on ex. The elements of the stiffness matrix are some
functions of ex. A common form of the modified stiffness matrix is [see (3.66)]
K = K* + f(ex)dK*
(3.134)
165
(3.135)
(3.136)
ax!
Y; = aX!
(3.137)
and obtain the linear relationship (3.135) in terms of L\Yj In cases where such
transformations are not possible, linear approximations of the nonlinear terms of
K are often sufficiently accurate.
Local Approximations: Series Expansion. Assuming approximations
along the line defined by (3.133), the Taylor series expansion about X"' (a =0) is
given by
(3.138)
The displacement derivatives can readily be calculated by the methods discussed in
Sect. 3.1.1. Assuming that R is independent of a and considering the direct
method, differentiation of (3.1) with respect to a gives
(3.139)
166
3 Approximation Concepts
aK
aa =M(
(constant)
(3.140)
Assuming that the relationship (3.134) holds, then the binomial series (3.72)
along the line (3.133) is reduced to the explicit expression
r = [1- !(a) B* + f(a) B2 - ... ]r
(3.141)
(3.143)
(3.145)
Comparing (3.144) and (3.145) it can be seen, once again, that the Taylor series
and the binomial series approximations are equivalent if the linear relationship
(3.136) holds.
Global Approximations: Polynomial Fitting. Polynomial fitting
techniques, discussed in Sect. 2.2.1, can be used to obtain explicit approximations
of the displacements along a line. While series expansions are based on a single
exact analysis, polynomial fitting techniques usually require analyses or calculation
of the displacement derivatives for several designs. Since more information is used
in the latter techniques, the quality of the approximations is higher at the expense
of more computational effort.
Assuming, for example, the quadratic fitting
r(a) = a + ba + ca2
167
(3.146)
r" = a + b + c
for a = a" = 1
.)
'(
r(a)=r ,Or
+-a+ r '" -r -Or- a 2
aa
aa
(3.148)
This equation is based on two exact analyses and calculation of the displacement
derivatives at a single point. Another possibility, that does not involve evaluation
of derivatives, is to use results of three exact analyses. Substituting the computed
values of r' (for a' = 0), r" (for a" =0.5), r'" (for a" = 1) into (3.146)
and solving for a, b, c, we find
r = r' + (-3r + 4r" - r*) a + (2r' - 4r" + 2r"')a2
(3.149)
=a + ba + ca2 + da3
(3.150)
can be determined from the values of r, ar'/aa (for a' = 0) and r", ar"/aa
(for a" = 1). The result, based on exact analyses and derivative calculations for
two designs, is
+ ( 2r , -2r "ar
+ - +ar")
-- a3
aa
(3.151)
aa
Example 3.7. Consider the three-bar truss shown in Fig. 3.11. The modulus
of elasticity is 30, 000 and the cross-sectional areas are 1.0. The geometric
variables flo f2 represent the location of the free node. Two cases of changes in
the geometry along a line in the design space have been considered
168
3 Approximation Concepts
20
Fig. 3.11.
Three-bar truss.
Case a
Y2
100
(l
Caseb
Results obtained for the horizontal displacement by the following methods [85] are
shown in Fig. 3.12:
(a)
0.25
(b)
0.25
l1....f
0.20
0.20
0.15
0.15
0.10
0.10
0.50
a
~
1.00
0.50
1.00
,..a
169
- A =exact solution;
- B =the flrst-order Taylor series (3.138) ;
- C the second-order Taylor series (3.138) ;
-D =the quadratic fitting (3.148) ;
-E =the quadratic fitting (3.149);
- F = the cubic fitting (3.151) ;
- G =the fU'St-order binomial series (3.72).
It can be seen that in both cases methods B and G provide poor results. Better
results are obtained by method C, and the best results are achieved by the
polynomial fitting techniques (methods D. E. F).
3.2.5 Approximate Force Models
Forces as Intermediate Response Quantities. In statically determinate
structures the element forces are independent of the cross-sectional variables while
the stresses are not. This implies that in statically indeterminate structures the
stresses are more nonlinear functions of the variables than element forces.
Consequently, improved approximations of stresses could be developed by using
element forces as intermediate response quantities. That is, instead of using flrstorder approximations of the stresses
II
~_.
0'.=0'.+
"
-ClUj
j=l
ax
I
(X--X-)
I
(3.152)
+ I" -aA;
aX
A.
0'.
A.
J
= _,
= __
Xj
:=j-~l
(x. I
x)
I
_ _ _ _ __
Xj
(3.153)
(3.154)
It should be noted that the approximation (3.153) is exact for statically determinate
structures.
170
3 Approximation Concepts
NiX) == N(X)
can be used for simple evaluation of the constraint functions. The displacements
and the forces corresponding to any approximate redundant forces N a can be
calculated directly by (1.23) and (1.24). It should be noted that member forces
corresponding to approximate redundant forces will satisfy equilibrium but not
necessarily compatibility conditions.
For any given design represented by a flexibility matrix F with approximate
redundant forces N a, we may define a vector of fictitious displacements Ba by
(3.156)
If Na are the exact redundant forces then Ba = B. That is, the approximate redundant
forces can be viewed as the exact forces for a structure with displacements Ba in the
direction of the redundants. The difference
(3.157)
=B* + AB
(3.158)
171
(3.159)
=N + aN =NA - CN
(3.160)
=~. + a~
(3.161)
=NA
(3.162)
(3.163)
(3.164)
(3.165)
172
3 Approximation Concepts
(3.167)
in which k denotes the term in the series and
(3.168)
It can be noted that if only the rrrst terms of the BSA (3.164) are considered, then
these approximations become equivalent to the FOA (3.160).
The First-Order Taylor series expansion of N about X* is given by
(3.169)
It has been shown [63] that for cross-sectional variables the frrst-order Taylor series
expansion in the reciprocal variables
(3.170)
is equivalent to the FOA (3.160).
Similar to displacement approximations, first-order approximations of forces
may provide poor results for large changes in the design variables. More accurate
results can be obtained by the BSA when the series converges and several tenns are
considered. However, the series might diverge for large changes in the design.
Combined Series Expansion and Scaling. A procedure that greatly
improves the quality of the BSA (3.164) is presented in this subsection. Similar to
scaling of the stiffness matrix [see (3.87)] the method [83] is based on scaling of
the initial flexibility matrix F*
F
=JlF*
(3.171)
= F* + AF =JlF* + AFIL
(3.172)
1-1l
1
C =--I+-C
...
Il
Il
173
(3.175)
Using this last expression in (3.174) and expanding (I + C...>-t, the following
Improved Binomial Series Approximations (IBSA) of N in terms of Il are obtained
1
2
N=-(I-C ... +C ... - ... )N A
Il
(3.176)
For Il = 1, C ... becomes equivalent to C and the IBSA is reduced to the BSA
(3.164).
The object now is to select Il such that the convergence properties and
approximation qualities of the series (3.176) are improved over those of the series
(3.164). It has been shown [67, 85] that various selections of Il may lead to
improved approximations of displacements. One possibility is to assume a
criterion similar to (3.94), that minimizes the Euclidean norm of the second term
in the series (3.176)
(3.177)
This does not involve a prior determination of elements of matrix C and the
calculation requires only forward and backward substitutions. By this criterion, the
scaling factor Il is determined from
(3.178)
in which NJj and N2i are the element of the vectors Nt and N 2, respectively, as
defmed by (3.165). The advantage of this criterion is that all terms can readily be
computed. The effect of Il on the quality of the approximations is demonstrated by
the following example.
Example 3.8. Consider the twenty-five-bar truss shown in Fig. 3.13 with the
initial and modified designs
XT= 10
L\XT = {95, 100,82,83,80,99, 88,95, 96, 88, 97, 82,
81,94,88, 88, 96, 93, 91,86,95, 87,92, 84, 85}
where X is the vector of cross-sectional areas. The applied loads are
Node
x
100
0
50
50
1000 -500
1000 -500
o
o
0
0
174
3 Approximation Concepts
Fig. 3.13.
Twenty-five-bar truss.
(3.160)
77.5
168.8
-136.6
999.5
-1242.1
1400.1
875.8
(3.164)
Two #
10.5
25.7
-39.2
186.4
-236.5
-270.8
157.8
Five #
27.1
59.1
-80.7
403.5
-510.8
-580.8
340.0
(3.176) ##
Two #
Five #
126.7
125.9
193.1
193.3
-147.2
-146.1
1007.9
1008.4
-1261.6
-1262.0
-1381.8
-1380.5
829.3
830.6
Exact
126.7
193.2
-146.0
1008.4
-1261.9
-1382.0
830.7
#
Tenns
## Il = 0.10
The chosen redundants are the forces in members 19-25, and the changes in the
design variables are about 900%. Redundant forces obtained by the FOA (3.160),
the BSA (3.164) and the IBSA (3.176) are given in Table 3.7 [83]. It can be
observed that poor results have been obtained by the FOA (3.160). The series
(3.164) and (3.176) converge to the exact solution. However, it has been noted
that only two terms are sufficient to obtain good results by the IBSA (3.176)
Exercises
175
while more than 50 terms are required by the BSA (3.164) to reach a similar
accuracy level. This example illustrates the effectiveness of scaling on the quality
of the results.
Exercises
3.1 Consider the three-bar truss shown in Fig. 1.11. with the given analysis
equations (c), (d) in example 1.7
a. Assuming the initial design
xf
3.2 The symmetric truss shown in Fig. 3.14 is subjected to a single load P = 10.
The members' cross-sectional areas are Xl> X z and 1/2 Xl as shown and the
modulus of elasticity is E =30.000.
a. Show that the fIrst-order Taylor series expansions of the displacements rl rz in
Perform three iterations of the sequential linear programming method. Assume the
xf
initial point
={7.0, 4.0}. and the move limits L1Xu = L1XL [see (3.52)] of
2.0. 1.0. and 0.5 for the fIrst. the second and the third iteration. respectively. Show
graphically the linearized feasible region for each of the iterations.
3 Approximation Concepts
176
C.L.
p= 10
4Xl
(5)
rl
Xl
iXI
(4)
100
(I)
(3)
B
I r2
4Xl
I.
Fig.
100
.1
100
3.14.
3.4 Assume the continuous beam shown in Fig. 3.15. The five design variables
are Xi = (EII!)i (i = 1... 5). Assuming the initial design X* = 1.0 the
following results are given from the initial analysis
28
-7.5
30
-8
-8
2
30
-7.5
-7.5
28
~5l
r.
0.10526}
={ 0.07895
0.07895
0.10526
Assume a change in the initial design variables such that the modified design is
given by
(X)T = {5.0. 5.2.4.8.5.2. 5.0}
Evaluate the modified displacements by the following methods. assuming only two
terms (rl and r~ for each of the series:
the Taylor series (3.69) ;
b. the binomial series (3.74) ;
c. the scaled binomial series [(3.91). (3.92). (3.95). and (3.96)] ;
d. the combined series expansion (3.74) and reduced basis [(3.82) through (3.86)] ;
Q.
Exercises
177
Evaluate the errors (3.106) in the modified analysis equations obtained by each of
the four methods.
3.5 Assume that X is a point on the constraint surface rj = rX. Show that for
any point X = J,1X on the scaling line through X*, where r* = (lIJ,1)r**, the
linearized consttaint is parallel to the consttaint tangent hyperplane at X** and can
be expressed as
"
':lo **
"'"
ar
~_J_X;
;=1
ax;
=J,1(J,1-2)rjU
3.6 Evaluate the modified displacements of exercise 3.4 along the line
X =X*+a AX*
for a = 0.2S, O.SO, 0.7S, 1.0. Draw the displacements versus a as obtained by
each of the four approximate methods. Compare the results with the exact
solution.
3.7 Consider the problem formulated in exercise I.S. Assuming the initial design
(X*)T = {SO, SO}, introduce the stress approximations in sections B. C under the
loadP1 by:
={30,40}.
a
Fig.
I
(>...
2i
3.15.
3
,...,...
"
2i
2i
4
,...,...
4
2i
4 Design Procedures
The design problem is formulated. The design variables are chosen, the
constraints and the objective function are defined and an analysis model is
introduced. This step is of crucial importance for the solution process. A poor
problem formulation might lead to incorrect results and/or prohibitive
computational cost. Various formulations have been discussed in Chap. 1.
The optimization method is selected. One of the methods presented in Chap. 2
might be suitable for the solution process. In general, the reliability and ease of
use of the method are more important than its computational efficiency. Since
most of the cost of optimization is associated with the exact analysis and
derivative calculations, efficiency of the method used to solve the problem is
not a major consideration in choosing the method.
Approximations are introduced. It has been noted that approximations are
essential in most practical design problems. Using linking and basis reduction
methods, it is possible to reduce the number of independent design variables.
Scaling of variables, constraint normalization and constraint deletion techniques
(Sect. 1.3.4) are all intended to improve the solution efficiency. Approximate
behavior models, discussed in Chap. 3, are often necessary in order to reduce the
number of exact analyses during the solution process.
A design procedure is established. The problem formulation, the chosen
optimization method and the approximation concepts are integrated to introduce
an effective solution strategy. In this chapter, various design procedures
demonstrate the solution methodology.
180
4 Design Procedures
Z=11X
min
(J~X ~ A ~ (J~X
(yield conditions)
C A=Ru
(equilibrium)
(4.1)
18 1
z = t1){ ~ min
(4.2)
The number of variables in this LP formulation [n+(nL nR)] is smaller than the
number of variables in formulation (4.1). In addition, the equality constraints have
been eliminated.
A major complicating factor in the plastic design of trusses by LP is the
variability of the compressive limiting stresses. Several approaches have been
proposed to treat buckling in compression members. The ultimate stress in
member i which buckles is
(4.3)
in which (JEi = Euler buckling stress, E = modulus of elasticity, li = unbraced
length of member, and rGi =critical radius of gyration. In general, the stress in each
member must satisfy
(4.4)
However, if the Euler buckling stresses (JEi were entered into the yield conditions, a
nonlinear programming problem would result. Russell and Reinschmidt [116]
found that when evaluating the results of the LP problem, it is useful to ignore the
computed member areas and design for the computed member forces, which
constitute a force system in equilibrium. The compressive limiting stresses are
then modified and the LP problem is solved repeatedly until convergence.
Transformation of Variables. It has been noted in Sect. 2.3.2 that all
variables in a standard LP formulation are assumed to be nonnegative. Since the
member forces are not restricted to nonnegative values, transformation of variables
may be used to account for the unrestricted variables. This can be done in several
ways, briefly described herein.
Considering the formulation (4.1), the member forces A can be represented by
the difference of two vectors of nonnegative variables A' and A by
II
A=A'_A"
(4.5)
182
4 Design Procedures
Substituting (4.5) into (4.1) the number of variables in the LP problem becomes
(2nL+1)n. However, the advantage of this approach is that only nL n yield stress
conditions are required instead of 2nL n in the original problem. To see this
possibility, we may rewrite the yield conditions in terms of A' and An and obtain
the following n inequalities for each loading condition
(4.6)
A procedure can be established to guarantee that for any member i either the
constraint Xj ~ A'j Icrf or Xj ~ _Anj IcrT will be considered, depending on the sign
of the force.
An alternative approach is to use an axis transfer of the form
A = A' - Ao 1.0
(4.7)
(4.8)
The variables A' and the inequalities (4.10) can be used instead of A and the
original yield stress conditions. Since all variables in a standard LP formulation are
assumed to be nonnegative, only the right-hand side of the inequalities (4.10) must
be considered.
Finally, we may introduce a vector AU of positive constants or linear functions
of X which have numerical values larger than the expected values of A. Defming
the new variables A' and replacing the original variables A by
A'=A+AU
(4.11)
then all variables in the LP problem will be nonnegative. Choosing the upper
bounds
(4.12)
183
(4.13)
It can be noted that in cases where cr~ = -cr~, the constraints (4.10) and (4.13)
become identical.
Considering the formulation (4.2), similar transformations can be applied. The
transformation (4.7) becomes
(4.14)
and the number of variables is increased by one. Alternatively, the transformations
(4.8) and (4.11), respectively, become
N' =N _NL
(4.15)
N'=N +Nu
(4.16)
PI =20.0
P2 = 30.0
P3 = 40.0
Assuming the formulation (4.1), the problem is to find the cross-sectional areas
XT = {Xl> X2 X3 } and the members force matrix A=[Al> A2 , A 3] (corresponding
to the three loading conditions) such that
Z = {141.4. 100.0. 141.4} X -+ min
-20]
20
(b)
(c)
[1 0
1 .J2
-1]
[40
0
1 A = 40 30.J2
(a)
184
4 Design Procedures
100
~ I
Xl
PI
100
= 20
P2 = 30
P3 = 40
(d)
2l.21
(J~ [X, X, X] ~ [ 0
20 21.21
(e)
*T
z* = 1206
I'
360
360
(I)
(2)
~I
(6)
100
360
100
(j)
185
Example 4.2. The ten-bar truss shown in Fig. 4.2 is designed to resist a single
ultimate loading. Limiting stresses of a L = -25, aU = 25 are assumed for all
members. Choosing the forces in members 7 and 10 as redundants N1 and N 2
respectively, the LP plastic design problem can be stated in the form of (4.2) as
follows: fmd the cross-sectional areas XT={X 1 X2 XlO } and the redundants
NT={N1 N 2 } such that
Xj
+ 1.414 x 360
j=1
-~{~: }~
XlO
L
10
Z = 360
Xj
~ min
(a)
j=1
-0.707
300
0
0
0
-0.707
-100
-0.707
0
-100
0
-0.707
100
-0.707 -0.707
+
0
0
-0.707
0
1.0
0
-282.8
1.0
0
141.4
0
1.0
0
0
1.0
{::}s~f: }
(b)
XlO
Z' = 15,840
(c)
That is, the unnecessary members 2,5,6 and 10 have been eliminated by the LP.
In order to compare the elastic and plastic optimal designs, limiting stresses and
loadings are assumed to be the same in both cases. Since the optimal solution
represents a statically determinate structure, the plastic and elastic optimal designs
are identical. If lower bound constraints on X are considered so that 0.1 ~ X, the
resulting optimal plastic design is
X*T
Z* = 15,910
(d)
Z*
=15,934
(e)
Frame Structures. The frame optimal design problem can be stated as the
following LP, similar to (4.1) : find Mpl and M such that
186
4 Design Procedures
Z=lMpl~min
(4.17)
C M=R,.
where M is a vector of moments Mj,j = 1,... , I, in statically admissible moment
field corresponding to collapse, and I is the number of critical sections where
plastic hinges may form. L is a linking matrix of 0, 1 elements. If Lji = 0, the ith
plastic moment does not govern section j. The inequality constraints require that
the admissible moments nowhere exceed the plastic moment capacities of the
members. Strictly, these constraints must apply at any point in the structure, but
in practice it is necessary to confine their application to I possible hinge
positions. For prismatic members this can be achieved by considering moments
only at the ends of members and at the position of maximum moment in loaded
elements. The number of variables in the LP problem (4.17) is I+I (I being the
number of plastic moments). The number of independent equations of equilibrium
is nE = I - nR
The frame design problem (4.17) can be stated as the following equivalent LP,
similar to (4.2) : find Mpl and N such that
Z = iT Mpl ~ min
-L
Mpl ~ Mp
+ MNN
Mpl
(4.18)
in which Mp is the vector of moments due to the applied loading, and MN is the
matrix of moments due to unit value of redundants, both computed in the primary
structure. The number of variables in this LP formulation is l+nR and the number
of inequality constraints is 21.
Though most practical designs for regular frames will tend not to have reverse
taper in the column members, we may consider constraints of the form
(4.19)
in which M pl. is plastic moment of columns at the ith story. We may consider
also linking constraints to ensure a desired ratio between Mpl. and Mpl'+1
(4.20)
where ~ is a given constant. The above linear constraints can be included in the
LP formulation.
Example 4.3. To illustrate the LP formulation (4.17) consider the frame
shown in Fig. 4.3 with 14 critical cross sections and 4 groups of plastic moments.
Bending moments with tension at inner fibers of columns or at lower fibers of
187
beams are assumed to be positive. Results for this example have been presented by
Cohn et al. [18].
The optimal design problem is to find M~l = (M pll' M p l 2' M pl 3' M pl.) and
T
M = (MI' M 2
M 14 ) such that
Z=
lTMpl
~ min.
(a)
6
0 0 0 0 0 0 0 0 0 0 0 -1 2 -1
12
0 0 0 0 -1 2 -1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 -1 1 1 -1 0 0 0
3
-1 1 1 -1 0 0 0 0 0 0 0 0 0 0
9
M=
0 0 0 0 0 0 0 0 -1 0 0 1 0 0
0
0 0 0 0 0 0 0 0 0 0 -1 0 0 1
0
0 -1 0 0 1 0 0 1 0 0 0 0 0 0
0
0
0 0 0 -1 0 0 1 0 0 1 0 0 0 0
C
Pul
(b)
Ru
-L
Mpl
:5:M:5:L Mpl
(c)
in which
IT
1
1
1
1
0
0
0
L=
0
0
0
0
0
0
0
0
0
0
0
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
(d)
188
4 Design Procedures
3Pu
Pu -
CD
M,'J
M,/4
@
M,ll
6Pu
CD CD
CD
2Pu -----+-
MplJ
MplJ
CD m
7?
I
I
I'
Fig. 4.3.
CD
2l
I,
I
13l
,I
0) @
M,a
8)
-t
2l
'77 ~
I
.. !
-!
I
13l
I
-'
Frame example.
Using (4.8) with A, A' and AL replaced by M, M' and -L M pl' respectively, the
resulting number of variables is 18, the number of equality constraints is 8 and the
number of inequalities is only 14. The optimal solution is
M~,
=P"t
189
z= 51P. t2
"
MT =P ul (-3,0,3, -3, 1.5,4.5, -4.5, -1.5, -1.5, 1.5, -1.5, -1.5, 1.5, -1.5)
(e)
(j)
MN
the resulting nonnegative variables are Mpl and N'. The optimal solution is
190
4 Design Procedures
3p',
N6
~r""
N4
l.5Pu l
p.,--
4.5Pu i
N)
"""''\r''''
Nl~
)NS
)N2
(b)
(a)
05
0.5
. h--::::::OO<~~-T""1
0.5
rT--~O-::::::-""""
1.0
0.5
1.0 r--r---<>--~-'11.0
1.0
Ns
= 1.0
(c)
191
M~l
11.5)
Z=51Put 2
Z =(l'X -+ min
(4.21)
This formulation is similar to that of (4.2), but the loadings and the limiting
stresses are different; while service loads and allowable stresses are considered in
(4.21), ultimate loads and yield stresses are assumed in (4.2).
The LP formulation can be preserved, under certain circumstances, even if
displacement constraints are considered. This is the case, for example, in
continuous beams when the displacements are given in the form of (l.33). We note
first that the internal forces A are linear functions of N. Also, a statically
equivalent internal force system corresponding to the virtual loads may be selected
so that the forces
A;9
expressed as
Iij = Iijo +
Iijk Nk
Tij
can be
(4.22)
where T ijO and T ijk are constant coefficients. From (l.33) and (4.22), any
displacementDi can be expressed as
(4.23)
Assuming a continuous beam and choosing the bending moments over the
supports of the member under consideration as redundants Nil N 2 , then the
displacementDi in the hth member becomes
192
4 Design Procedures
D. = TiIIO
I
+ Till! N! + TiII2 N2
(4.24)
Xit
Df
and
(4.25)
which can be added to the LP problem (4.21).
Effect or Compatibility. The LP formulation (4.21) is most suitable for
optimal design of trusses. If no lower bounds on cross-sectional areas X are
considered, the LP method has the ability to make unnecessary members to vanish
from the structure. This topic of optimizing the topology will be discussed later in
Sect. 4.7. If the resulting optimal solution represents a statically determinate
structure, the compatibility conditions are always satisfied and the LP solution is
the final optimum. In cases where the optimal solution represents a statically
indeterminate structure the compatibility conditions may not be satisfied, and the
LP solution is not the final optimum.
To illustrate the effect of compatibility conditions on the optimum, solutions of
two formulations have been compared [70]: the LP, where compatibility conditions
are neglected and the NLP, where the latter conditions are considered. In general,
the optimal LP solution, Z;;, will be at least as good as the optimal NLP
solution Z~, that is
(4.26)
Therefore, Z;; can be viewed as a lower bound on ZNLP' It has been shown that
under certain circumstances the optimal solutions of the two problems are identical
[70]. In general, three different situations might be encountered :
Q.
Z;'<Z~
(4.27)
b. The two optimal objective functions are identical, but the LP problem
possesses multiple optimal force distributions, including the one corresponding
to the optimal NLP force distribution, that is
(4.28)
193
This situation might occur in cases where the LP objective function contours
are parallel to the boundary of the feasible region.
c. The two optimal solutions are identical
z:.r, =z~
(4.29)
N~=N~LP
That is, the compatibility conditions do not affect the optimal solution. This
situation might occur for certain geometries and loading conditions.
Several procedures may be used in case a to modify the LP solution such that
compatibility requirements are satisfied. Denoting the strain in the redundant
member i under the load condition q as Eiq' the corresponding redundant forces as
Niq' and the implied areas as Xiq' then
N
X. =--.!!L
"I
EE.
q= 1. 2... nL
(4.30)
"I
where nL is the number of loading conditions. Obviously the value of Xiq would be
unique for all q if the design had been fully compatible. It has been shown [24]
that the conditions of compatibility are equivalent to the conditions that lead
toward the uniqueness of X iq for all q. That is, in a compatible design the
following relations must be satisfied
=2, 3,..., nL
(4.31)
Equation (4.31) produces (nL - 1) equality constraints for the uniqueness of the area
of each redundant member i. Based on (4.30), these constraints can be expressed as
q = 2, 3,... , nL
(4.32)
Assume that the variables of (4.32) are N iq and N i .q.l while Eiq and Ei,q.l are taken
as known constants whose approximate magnitudes are supplied by elastic analysis
of the optimal LP design X~. Thus (4.32) are linear and suitable for use together
with the original LP. After the initial solution of the LP, subsequent iterations
involve both the LP and (4.32) simultaneously, resulting in a series of modified
designs X~ whose compatibility violations successively diminish until the final
optimal design is reached. This procedure is quite efficient [24] but it may not
converge to the true optimum in some cases [61].
An alternative iterative procedure to consider compatibility, buckling, and
discrete rolled steel sections is as follows [111]:
1. Specify all given parameters and the initial values for the allowable
compression stresses, to start the iteration.
194
4 Design Procedures
2. Solve the LP problem (4.21) for the member areas and forces, satisfying the
equilibrium conditions, the stress constraints, and the area bounds.
3. For each member, using the computed member forces from step 2, select the
minimum section (from the table of available rolled steel sections) which
satisfies the stress constraints. The member areas computed by the LP are
ignored; the use of the computed member forces gives improved convergence
because the correct allowable compression stress can be computed for each
candidate section, using code formulas and the tabulated radius of gyration. The
LP is used to derive a distribution of forces in the optimal structure; the crosssectional areas determined by the LP are only an intermediate step.
4. If the process has converged, go to step 5. If it has not converged, replace the
lower stress bound for each member in the LP by the allowable compression
stress computed for the section selected in step 3, and return to step 2. In the
case of discrete sections and code-allowable compression stresses, convergence
is obtained when for each member the same section is selected in two
successive cycles.
5. If the resulting structure is statically determinate, the design is complete. If the
design is statically indeterminate, add the compatibility conditions, and obtain
an exact elastic analysis.
6. Select new sections as necessary based on the code requirements and the
member forces computed in step 5. If the convergence criterion is met, the
design is complete; if not, return to step 5.
As there is no explicit objective function in steps 5 and 6, the resulting design will
be fully stressed but not necessarily minimum weight.
Example 4.4. Consider again the three-bar truss shown in Fig. 4.1 and
subjected to three loading conditions. In order to compare elastic and plastic
optimal designs, limiting stresses and loadings are assumed to be the same in both
cases. Choosing the forces in member 2 as redundants NT = {N I' N z, N 3 }
(corresponding to the three loading conditions), the resulting LP problem is given
by (d), (e) in example 4.1. The plastic and the elastic design formulations are
similar, however, in the latter formulation compatibility conditions must be
considered. The optimal LP solution is
X
"T
z" = 1206
(a)
Analysis of this structure shows that the elastic forces NE do not match the
optimal forces N" determined by LP. The strains 2q of the redundant member 2,
are computed for all loading conditions (q = 1,2,3) and the LP is now solved with
the added constraints (4.32). These requirements are modified after each cycle, and
the global optimum, reached after six cycles, is [24]
"T
X = {8.0, 1.5, O}
z = 1281
(b)
It can be seen that the method has the ability to eliminate unnecessary members of
the structure. Structural analyses are performed only to update the terms in (4.32).
Solving the NLP problem by a numerical search technique, the solution is [127]
-1
-=1597
195
(c)
which is not the optimum. That is. the true optimal solution could not be reached
by the general NLP formulation. This phenomenon of singular solutions in
optimization of structural topologies will be discussed later in Sect. 4.7.3.
N~ =N~-N~
(4.33)
= 1... nL
(4.35)
This is the case also in SIS if the conditions (4.28) or (4.29) hold for all loading
conditions. In summary. the various possible cases of LP solutions are
characterized as follows:
196
4 Design Procedures
100
(a)
(b)
Example 4.5. The eleven-bar truss shown in Fig. 4.6a is subjected to two
loading conditions (P l and P z, respectively). The allowable stresses for all
members are aU -oL 20.0 and the assumed loads are P 1 P z 10.0. Conditions
of symmetry are considered to reduce the number of variables and constraints. No
lower bounds are imposed on cross sections, that is XL = O. Two selections of
redundants have been assumed: the forces in elements 1 and i. and the force in
element 4. The results are summarized in Tables 4.1 and 4.2, and the optimal
design is shown in Fig. 4.6b [74]. The optimal objective function value is Z' =
125. It can be observed that the optimal structure is statically indeterminate, and a
single set of prestressing forces can be chosen to maintain compatibility at the
optimum for the two loading conditions [see (4.33)].
= =
= =
Load 1
3.54
-3.54
0
5.00
0
0
A~I
X~I
Load 2
-3.54
3.54
0
5.00
0
0
0.177
0.177
0
0.250
0
0
N*LP
3.54
5.00
Loading 1
N*E
N*P
N*LP
4.25
4.00
-0.71
1.00
-3.54
5.00
Loading 2
N*E
N;
-2.83
4.00
-0.71
1.00
197
a. Exterior (or infeasible) methods. based on search for the optimum outside the
feasible region. where all intermediate solutions lie in the infeasible region and
converge to the optimum from the outside. A major shortcoming of these
methods is that the search cannot be stopped with a feasible solution before the
optimum is reached. Examples for this class of optimization methods include
exterior penalty-function techniques and sequential linear programming. Such
methods are usually not suitable for evaluating improved feasible designs.
Intermediate solutions obtained by exterior methods may often be viewed as a
lower bound on the optimum.
b. Interior (or feasible) methods. where all intermediate solutions lie in the
feasible region and converge to the optimum from the interior side of the
acceptable domain. The advantage of these methods is that we may stop the
search at any time and end up with an improved feasible design. Moreover. for
some methods the constraints become critical only near the end of the solution
process; thus instead of taking the optimal design we can choose a suboptimal
but less critical design. Another advantage is that a near optimal solution can
be achieved with a reduced computational effort. Examples for this class of
optimization methods include interior penalty-function techniques and the
method of feasible directions. Intermediate solutions obtained by feasible
methods may be viewed as an upper bound on the optimum.
Introducing Feasible Designs. One problem in using interior optimization
methods is that it might be difficult to find an initial feasible design. particularly
in problems with a narrow feasible region. Several methods can be used for this
purpose. some of which are discussed in this section. Such methods are useful also
in identifying situations where no feasible solutions exist.
Assume the inequality constraints (2.111)
j = 1.
n,
(4.36)
198
4 Design Procedures
=1..... p-l
(4.38)
KS =
.!.In[t
p
exp(pgj)]
(4.39)
j=!
199
(4.42)
XL SXSX U
(4.43)
in which J1 is a variable scaling multiplier (J1 > 0). Equation (4.43) defines the
scaling of X*. Under certain circumstances, analysis of a scaled design is trivial,
provided the initial design X* has already been analyzed. In the presentation that
follows, the displacement method of analysis is considered. A similar approach can
be applied for the force method, or other analysis methods. The scaling procedure
will be used in several design procedures throughout this chapter.
Assume that the elements of R in (1.26) are independent of X and the elements
of the stiffness matrix are linear functions of the design variables, that is,
L XiK ?
II
K=
i=1
(4.44)
in which K? are matrices of constant elements. The displacements for any design
on the line defmed by (4.43) are then given by
1
*)
r(jlX * ) = -r(X
(4.45)
J1
The relationship (4.44) is typical for various structures such as trusses, where Xi
are the cross-sectional areas, or beams, where Xi are the moments of inertia.
To find the resulting stresses, the scaled displacements are substituted into the
stress-displacement relations. Assuming that the elements of Sin (1.29)
(4.46)
cr=Sr
(4.47)
J1
In cases where the above conditions for stresses do not hold, simplified expressions
can be obtained. For frame elements, cr can be expressed as an explicit function of
J1 under certain assumptions [64].
200
4 Design Procedures
The significance of the relations (4.45) and (4.47) is that the exact displacements
and stresses at any point along the design line (4.43) can readily be determined
without solving the set of implicit equations (1.26). The scaling procedure can be
used to achieve the best feasible design along a given design line, as will be
demonsttate<l subsequently.
Consider, for illustrative purposes, the two-dimensional design space shown in
Fig. 4.7 where r and o are given displacements and stresses at X. It is assumed
that the objective function is increased with ~. From (4.45) and (4.47), the
constraints (4.42) can be expressed for any point on the design line (4.43) as
rL~!r~rU
~
OL
~ !oo ~ OU
(4.48)
XL
~~X~
XU
or, alternatively
(4.49)
where
rj0 /rj
II
r-
L}
rj /rj
LO
U
L
= max{ 0/0
.0,/0,
. .
I.J
"
"
x!- / X~
I
II U = min(X~
,....
.
, / X~)
,
I
(4.50)
(4.51)
The elements of rL and oLin (4.50) are assumed to be negative. That is, all
expressions in this equation are positive. From (4.49), the best feasible design
along the design line (4.43) can readily be determined by
201
It can be noted that if JlL> JlU, no feasible designs exist along the design line. That
is, a necessary and sufficient condition for a feasible design along this line is
(4.53)
For some standard beam cross sections the stiffness matrix elements are
functions ofaXt , in which Xi are the cross-sectional areas and a and b are given
constants. Choosing
(4.54)
as design variables, then the displacements can be scaled by (4.45).
A typical relationship for a general frame structure is given by
L [XiKFi+d(Xi)KNd
II
K=
(4.55)
i=l
in which KFi and KNi are matrices of constant elements, representing the
contribution of flexural deformations and axial deformations, respectively. Analysis
of the scaled design by (4.45) is possible if
(4.56)
where c is a constant. In cases where the relationship (4.56) does not hold, linear
approximations may be assumed.
Another approach, which might prove useful in cases where the conditions
(4.45) cannot be used and the displacement derivatives are available, is to assume
the fIrst-order approximations of rj in terms of Jl to obtain
(4.57)
Considering the constraint
(4.58)
then from (4.57) and (4.58) we have
) L
II
r +
i=l
~.
urj
)<r.U
_(IIV.
-X
ax....... - }
(4.59)
202
4 Design Procedures
J.I.>I+
rU -r"
1
ar; "
~
L.J - X
aXj
j=1
(4.60)
L aXar~ X."
II
_l
II
>
j=1
t""-
a"
+ L -X
lax.'
II
r"
-rU
rj
j=1
(4.62)
"
'
Equations (4.60) or (4.62) may be used instead of (4.49) and (4.50) to find the best
feasible design.
'
Conditions of Feasibility. For some problems it is difficult to determine a
priori whether feasible solutions exist or not. However, it might be possible to
show for a specific problem that there is no feasible region. This is the case, for
example, if the constraints (4.36) can be expressed in the form
(4.63)
where C is a vector of preassigned parameters. Based on the transitivity property,
necessary conditions for a feasible solution are
(4.64)
It should be noted that the conditions (4.64) are independent of the design variables
value and can be checked before solving the complete problem. Furthermore, in
cases where the conditions (4.64) are not met it might be possible to evaluate
modified values of C such that these conditions will be satisfied, as illustrated in
the next subsection.
In cases where the problem can be stated in a linear programming form it is
possible to identify situations where no feasible solutions exist. It can be shown
that in such cases the dual optimal solution is unbounded.
203
Assuming the displacement method formulation [see (1.26) and (1.29)], the
constraints (4.42) for the given design X can be expressed as
(4.65)
where X, K and S are constant. It can be observed that all the constraints are
linear functions of the preassigned parameters. Define a linear objective function
(4.66)
where b is a vector of constants and C is the vector of preassigned parameters to be
modified. It is then possible to find optimal values of the preassigned parameters
by solving the linear programming problem (4.65) and (4.66). A direct solution
can often be achieved, as will be illustrated in the following example.
Example 4.6. Consider the typical prestressed concrete member of a uniform
cross section continuous beam shown in Fig. 4.8 (all dimensions are in tons and
meters). A parabolic tendon is assumed with Yl' Y2 being the tendon's coordinate
variables, and P is the variable prestressing force. Two loading conditions have
been considered:
J
Fig. 4.8. Prestressed concrete beam.
204
4 Design Procedures
(a)
(stress)
(displacement)
(concrete coverage)
(b)
(c)
(d)
The design space for three different values of B 1 is shown in Fig. 4.9 and the
results are summarized in Table 4.3.
In case A (Fig. 4.9a), a wide feasible region is obtained. However, if the
constraint (d) is modified such that O.SX 1 :::; Xo no feasible solutions can be
achieved. In case B (Fig. 4.9b), the feasible region is reduced to a line segment,
where a transitivity condition becomes equality. Specifically, the first equation of
the stress constraints (b) becomes
-3214:S; -10.9Xo- 47.6Xl :s; -3214
(e)
if)
(g)
1.000
0.916
0.800
Feasible Space
Wide
Line Segment
58.7
64.5
0.900
0.816
205
70
60
40
/5'0
70
90
Xo
4'
'50
70
90
Xo
/~O
= 1.000,
b. Bl
~ ~o
(e)
(b)
(a)
70
= 0.916,
C.
Bl
= 0.800.
That is, the transitivity conditions are not satisfied for both if) and (g), and no
feasible solutions can be achieved for these concrete dimension~
Assuming the infeasible design Bl = 0.800, P =64.5, Yl - Y2=0.70, the stress
constraints can be expressed in terms of the preassigned parameters as follows
crf -1125DL:S; -3628.125:S; 1.25cr~ -1406. 25TL
(h)
(i)
Q.
crf:s; -1378
crf ~-234
cri:s; -1762.5
cr~ ~472.5
(j)
That is, the allowable stresses for the second loading condition must be
modified.
b. For the given allowable stresses, the modified loadings needed to convert the
design into a feasible one are
DL
1.89
TL:S; 2.50
That is, the live load must be reduced to 0.50 to achieve a feasible design.
(k)
206
4 Design Procedures
W=TX
(4.68)
in which T is a (2 x n) rotation matrix and n is the order of X. The elements of T
are the direction cosines
T = [COS(Wt.xl)... COS(Wt.x,,)]
cos(W2XI ).. COS(W2 X,,)
(4.69)
In Fig. 4.10 the direction of WI is selected along X*, and the direction of W 2 is
normal to X* in the given design plane.
207
..
Fig. 4.10. A design plane determined by X and l1X.
Although various methods can be used to select .1X*, only the SLP method
(Sect. 3.1.3) will be considered in this section. That is, the direction vector
.1X =X-X
(4.70)
or, alternatively
X = J.1(X + a.1X*)
(4.71)
w =J.1(W + a.1W*)
(4.72)
208
4 Design Procedures
(4.74)
then it can be shown [84] that the same ratio is obtained also for the approximate
scaled design Wa
P= ra(Wa) I r(Wa)
(4.75)
That is, the errors in the displacements at Wa depend only on the errors in W, and
not on the distance between W and Wa.
w,
________
~-L
__
~_~
209
Example 4.7. Consider the ten-bar truss shown in Fig. 4.2 with a lower bound
on cross sections XL =0.1. Two cases of stress constraints will be considered:
case A:
CJU =
CJU
The assumed initial design is XO = 10, the objective function represents the
volume and the assumed convergence criterion is
(Z<Ic-l) _ Z<Ic / Z<Ic.l):s; 0.5%
where k denotes the iteration cycle. The iteration history for the presented
procedure is shown in Fig. 4.12 [84]. The optimal solution in case A is
XT = (7.94, 0.10, 8.06, 3.94, 0.10, 0.10, 5.74, 5.57, 5.57,0.1)
Z* = 15,932
This solution is achieved after four iteration cycles (Fig. 4.12a). The results for
caseB are
XT = (7.90,0.10,8.10,3.90,0.10,0.10,5.80,5.51,3.68,0.14)
Z* = 14,976
It instructive to note that the true optimum, achieved after four iteration cycles, is
not a fully-stressed design.
Z/IOOO
35
35
25
25
15
15
-IL--1--11--+--1-1-.. k
2
TL--t--tl-t---l-I__
.. k
2
4
(a)
(b)
a. Case A b. Case B.
21 0
4 Design Procedures
211
by relative variations in the size of the other members and the iterative process
converges rapidly. However, for structures with the so-called hybrid action, the
convergence might be slow. Reasons for the significance of FSD include:
a. Engineering experience indicates that a good design is often one in which each
X(l+l)
I
=X(l) ~
(4.76)
esC!
esf
X(l)
Af l ),
(4.77)
X[l+l)
A(l)
~_._
X (l+l) __
j
esC!
I
can be computed by
(4.78)
212
4 Design Procedures
That is, the new design variable is the computed critical member force divided by
the allowable member stress. Geometrically, (4.76) means that the exact constraint
surface is approximated by a plane normal to the ith axis in the design space. Such
an approximation is called of zero order.
Using a simple stress-ratio redesign method, convergence will occur in one step
for a statically determinate structure. The number of iterations required for
convergence in statically indeterminate structures is not, in general, linked closely
to the number of variables in the problem. It is this fact which makes the use of
FSD so attractive.
Increasing the stiffness of a member will result in more force in that member.
Thus, it might be reasonable to increase (or reduce) Xj by a factor which is larger
(or smaller) than the stress ratio. To improve the convergence, an "over-relaxation"
factor, v ,greater than unity can be used [40] in (4.76)
v>1
(4.79)
A critical stress can be determined, for which the required X[k+l) is the largest.
The zero-order stress-ratio approximations (4.76) may not be sufficient in some
cases and convergence might be slow. Using the first-order approximations (4.81)
may lead to a divergent calculation when the initial design represents a poor
(oa load
k ) may require
estimate of the final design. In addition, calculation of
ll
much computational effort. Several approaches have been proposed to overcome
these difficulties. In the mixed mode approach [38], the stress-ratio rule is applied
for a few cycles, after which the first-order rule of (4.81) is used. Another
possibility is to use hybrid methods [32] based on zero and first-order
approximations.
Minimum
size
size
213
"
100
100
100
(b)
(a)
Fig. 4.13.
100
8.
It should be noted that in certain cases convergence of the iterative fully stressed
design procedure cannot be achieved since there is no FSD. To illustrate this
situation, consider a truss with n members (or design variables), nR redundants, and
nL load conditions. There are (n-nR) independent element equilibrium equations for
each load condition. The requirement that the number of independent equations
exceeds the n unknowns can be written as
(4.82)
If this requirement is violated, the member sizes cannot be chosen independently
and a fully stressed design might not exist
-25
~ (Ji ~
CaseB:
-25
-50
~ (J9 ~
-25
-70
~ (J9 ~
Case C:
~ (Ji ~
~ (Ji ~
25
for i = 1, ... , 10
25
50
for i = 1, ... , 8, 10
25
70
for i = 1,... , 8, 10
Results reported by Berke and Khot [11] are given in Table 4.4. For cases B and C
the solution is not improved, it rather converged to a fully stressed but nonoptimal
design. Figure 4.13 shows that different load paths have been obtained for case A
and for cases B and C. In the latter cases application of the stress-ratio rule may
214
4 Design Procedures
lead to elimination of member 9 by dividing with its greater allowable stress in the
stress-ratio algorithm. At the true optimum for cases B and C member 9 is not
fully stressed (cr9 = 37.5). An identical optimum would be obtained for all similar
problems with
Table 4.4.
Member
1
2
3
4
5
6
7
8
9
10
Volume
crlf ~ 37.5.
(4.83)
(4.84)
The objective function value is then calculated for X(k), which is the best
feasible design along the scaling line through X(k). If some convergence
criteria are satisfied (that is, no further improvements of the objective function
value can be achieved), then the solution process is terminated and the current
scaled design is the final solution.
4. Each design variable is modified independently by
- ( l+1) _
-max(
Xi
A(l)
A(l)
) _
.:.:i.-.-a;
L
U '-L-,Xi
,-l, ... ,n
(Ji
215
(4.85)
(Ji
1.0
0.8
0.6
0.4 -
/ /
/'
-(2)
/' /'
/'
/'./ /
/'./
0.2
..--
/'
"--X(3)
/~(3)
./ /
0.4
X(4) 4
( 4 ) . _0-_ X_
_
_ _ 5
- - - - -
-.:::. -
,:::-::;"'.::;-::::=---
/'
:~~::~\\~~\
./
I
0
/' / '
Objective
function
contours
0.2 -
0.6
-Q..(
i
X
-(5)
0.8
x(S)
1.0
1.2
216
4 Design Procedures
Solving by the combined stress-ratio and scaling procedure, the assumed initial
design variables are given by X(l)T
4.14). Computing J,1(1)
=(l.O,
J,1(1) =0.7CY7
X(l)T = (0.707,0.707)
i 1) =270.7
From (4.85), the new relative design variables X(2)T = (0.707, 0.414) are
computed, forming the line 0-2. Scaling of the design yields
J,1(2) = 1.094
X(2)T = (0.773,0.453)
= (0.773,
X(3)T = (0.815,0.337)
Z(2)
=264.0
0.320) and
z(3)
=264.3
Thus, X(2) represents the best design, which is close to the true optimum
XT = (0.788,0.410)
-r =263.9
It has been noted that, depending on the initial design, the solution may not
converge to the true optimum. Choosing, for example, X(4)T
initial design, we obtain
J,1(4) =0.890
From (4.85),
X(S)T
J,1(S)
X(4)T = (0.890,0.178)
=(0.890,
= 1.012
= (1,0.2)
z<4)
=269.5
Z<S)
= 270.3
as the
0.156) and
X(S)T
= (0.900,0.158)
In this case the solution process does not converge to the true optimum. The
initial design X(4) is the best, but nonoptimal, solution.
Table 4.5. Three-bar truss example, solution by stress-ratio (4.78).
Cyelek
0
1
2
3
X(l)
X(l)
1
1.0
0.707
0.173
0.815
2
1.0
0.414
0.320
0.261
a{l)
~l)
a~l)
14.14
21.87
21.09
20.71
8.28
15.47
16.32
16.89
-5.86
-6.40
-4.17
-3.82
1.0
20.0
217
Z=
L ljXj ~min
(4.86)
j=l
~ T:
LJ -' =r
(4.87)
j=l Xj
(4.88)
At the optimum the following conditions must be satisfied [see (2.16)]
h= t
.... n
(4.89)
'\ T"
l"-"'2=0
X"
h=l, ... n
(4.90)
218
4 Design Procedures
Equation (4.90) states that at the optimum. the quantity l"X~ I T" is the same for
every design variable. This form of optimality criteria is useful only if every
member is subject to change. In general. only part of the design variables are
determined by the constraint (4.87). To consider this possibility. the variables in
the structure are divided into two groups - active and passive variables. A variable
becomes passive if its value is determined by considerations other than those
associated with the displacement constraint (4.87). This topic of active and passive
groups of variables will be discussed later in this section. At present it is assumed
that there are I (~n) active variables. Denoting the contribution of the passive
members to the displacement expression as ro and to the objective functions as Zo.
then the problem (4.86) and (4.87) becomes
I
Z=Zo+
I
L (Xj ~min
(4.91)
j=l
T:
..J -' = r - ro == r *
(4.92)
j=l Xj
h = 1. ... .1
Substituting (4.93) into (4.92) for all variables (h
gives
1 I
.,ff. = . .
(4.93)
L ~T;lj
(4.94)
r j=l
h= 1. ... .1
(4.95)
Equation (4.95) represents the criteria which must be satisfied at the optimum of
the problem (4.91) and (4.92).
Multiple-Displacement Constraints. Multiple-displacement constraints
may arise in the case of multiple-loading conditions. or multiple active constraints
under a single loading. or both. A major difficulty in deriving the optimality
criteria for such problems is associated with prediction of the set of active
constraints. Consider the following optimal design problem with J equality
displacement constraints and I active members: find XT = (Xl ... XI) such that
Z = Zo +
219
tjXj
-+ min
j=1
L
j=1
(4.96)
T
-L- r =0
X.
J
= 1, ... .1
'
(4.97)
(4.98)
xf=
t A.j (Tjh)
j=1
(4.99)
h=I, ... ,!
th
Equations (4.99) represent the optimality criteria which must be satisfied at the
optimum.
Recurrence Relations for Redesign. Different numerical procedures for
redesign can be established for similar optimality criteria. In this subsection
redesign rules to satisfy the displacement optimality criteria, based on physical
considerations, will be presented.
Consider frrst the optimality criteria for a single-displacement constraint as given
by (4.95). The latter equation can be used as a recurrence relation for an iterative
process of determining Xh. Similar to the stress-ratio rule for achieving a fully
stressed design, (4.95) can be written in the form [9]
h= 1, ... ,/
(4.100)
where the superscript k denotes the iteration cycle. The iterative process consists of
applying successively analysis of the structure and the rule of (4.100). This rule is
based on the assumption that the elements of matrix T are constant at a given
iteration, i.e., the force redistribution is not sensitive to changes in member sizes.
This is an intuitive recurrence relation for modification of the design variables to
220
4 Design Procedures
satisfy the optimality criteria (4.95) with no guarantees on convergence. The latter
depends on the force redistribution during the solution process.
The optimality criteria for multiple-displacement constraints, given by (4.99).
can be expressed as the following recurrence relations
h=I ... , /
(4.101)
These relations are similar to those of (4.100) and could be used to obtain X it .
However, the multipliers Aj cannot be eliminated in a simple manner and explicit
expressions for Xit cannot be obtained. The optimal solution has to satisfy the /
conditions (4.99) and the J constraints (4.96). These are (/+J) nonlinear equations
with / unknown design variables Xit and J unknown Lagrange multipliers Aj.
Several methods have been proposed to solve this problem. One major difficulty
common to all approaches is that the constraints which are active at the optimum
are not known in advance. Thus any procedure has to be capable of deleting
constraints which are inactive at the optimum. This can be done by considering the
Kuhn-Tucker necessary conditions for optimality. The above considerations
indicate that OC methods are most suitable for problems with a singledisplacement constraint.
Passive Variables. Different considerations may affect the selection of the
passive variables and several approaches of dividing the design variables into the
two groups of active and passive variables are possible. If the variables are
separated into active and passive ones then the optimality criteria are necessary
conditions. On the other hand. if all variables are assumed to be active and can be
modified, the final design will not be required to satisfy the optimality criterion.
The latter is neither a necessary nor a sufficient condition. but only an expedient
way to modify the design variables giving a practical solution. If for a member the
size required to satisfy another constraint is greater than the size required to satisfy
a certain displacement constraint. that member should be considered passive.
Another criterion for the selection of passive members can be observed from an
examination of (4.100). If Ti for a member is negative. its inclusion as an active
member would require calculation of ~1il; introducing unacceptable imaginary
numbers. The corresponding criterion for a passive member is T; < O. indicating
that reduction in displacement occurs when the member size is decreased rather than
increased. This means that the member size will be determined by other constraints
(such as minimum size. stress. or other displacement). Such a member may be
treated as passive in the redesign to satisfy the appropriate displacement constraint.
4.3.3 Design Procedures
Combined Optimality Criteria and Scaling. It has been noted that the
OC method is most suitable for problems with a single displacement constraint. In
the case of multiple-displacement constraints, (4.100) can be applied to each
221
constraint separately and the largest value of each design variable is selected for the
final design. This extension of the recurrence relation to multiple loading
conditions and displacement constraints is accomplished in a manner similar to that
used in the stress-ratio procedure. That is, the largest design variable computed by
(4.100) is taken, considering all loading conditions and displacement constraints. If
we have nL loading conditions and J displacement constraints, X" is computed by
(4.100) for the J nL combinations. This procedure, proposed by Gellatly and Berke
[40], is called the envelope method. Since the method is based on the optimality
criteria for a single constraint (4.95), the solution usually does not satisfy the
correct optimality criteria for multiple constraints (4.99). However, solutions
reached by this approximate method are often very close to those obtained by the
correct criteria for multiple constraints.
The design procedure described in Sect. 4.3.1 for stress and member-size
constraints can be extended to include also displacement constraints. The procedure
presented here is based on combining the stress-ratio rule, the displacement criteria
and the scaling procedure. The solution process involves the following steps:
1. An initial design
x(/c)
is assumed (k=I).
100
100
-I
"
(2)
(I)
Xl
T
(3)
X2
Xl
100
222
4 Design Procedures
x,
Feasible region
1
2
3
4
5
Xl
0.943
0.943
0.943
0.943
0.943
X2
0.105
0.191
0.320
0.485
0.648
Xl
1.997
1.726
1.345
0.919
0.390
X2
0.333
0.575
0.897
1.225
1.014
Max. Values
Xl
1.997
1.726
1.345
0.943
0.943
X2
0.333
0.575
0.897
1.225
1.014
After Scaling
Xl
1.907
1.602
1.213
0.943
0.943
X2
0.318
0.534
0.809
1.225
1.014
Z
571.3
506.5
424.0
389.2
368.0
223
Various relations can be used to modify iteratively the design variables X and the
Lagrange multipliers A. in the redesign step. The optimality criteria are used to
derive recurrence relations to modify the design variables. and the constraint
equations are used to establish relations to update the Lagrange multipliers.
In general. the constraints that are active at the optimum are not known in
advance. At each iteration the set of active constraints can be checked and updated
as necessary. Since the Lagrange multipliers must be nonnegative at the optimum.
the constraints corresponding to negative multipliers at a given iteration may be
deleted from the set of active constraints. Convergence difficulties may arise in
cases when the initial design is far from the optimum or if constraint switching
occurs from iteration to iteration. However. convergence near the optimum is
usually very rapid. Various recurrence relations for redesign are discussed and
compared elsewhere [12.58.59].
4.3.4 The Relationship Between OC and MP
The MP methods. presented in Chap. 2. are general and most suitable for problems
with multiple types of constraints. These methods exhibit usually good
convergence properties. The computational process is stable but the convergence
near the optimum might be slow. In addition. the number of iterations increases
rapidly with the number of design variables .
Physical OC methods are not as general as MP methods. but their efficiency is
usually good. Mathematical OC methods are more general. but convergence
difficulties can arise in cases where the initial design is far from the optimum.
The efficiency of MP methods has been greatly increased by using
approximation concepts (Chap. 3). In addition. OC methods have been extended to
include more complex design conditions and rigorous criteria. The different
approaches have many common ideas and the positive features can be used to
establish better solution methods. The relationship between approximate
formulations and solution methods used in both approaches is briefly discussed in
this section.
Approximate Problem Formulations. The original design problem is
usually transformed into a sequence of approximate problems. solved successively.
The definition of the solution method used depends on the formulation and on the
algorithm for solving the approximate problems. Classification of different
approximate problem formulations and possible methods of solution have been
proposed elsewhere [32]. Assuming homogeneous displacement and stress
functions. it has been noted in Sect. 3.2.3 [(3.121) through (3.135)] that the
224
4 Design Procedures
Therefore, the solution process requires a large number of analyses and much
computational effort However, the convergence is usually guaranteed.
b. K ~ 00 , the approximate problem is solved exactly after each sttuctural
analysis. This can be done by generalization of the optimality criteria, as
proposed by Fleury and Sander [32]. The explicit GOC are derived from the
Kuhn-Tucker conditions of the approximate problem. Since the approximate
problem formulation is equivalent to using a first-order Taylor series expansion
in the reciprocal variables, the problem can be solved exactly also by MP
methods applied to the linearized constraints. This means that solution by GOC
is equivalent to solving completely the approximate problem with the same
linearized constraints by MP. The properties of solution by this approach are
similar to those of OC methods, i.e. the solution is more efficient but the
convergence is uncertain.
c. K. limited number, the approximate problem is solved partially. Updating the
approximate constraints after a limited number of steps, the approach becomes
mixed with properties lying between those ot OC approaches and those of MP.
Actually, both MP and OC methods often do not solve the approximate
problem exactly.
225
!&Yd.3.
Level 2
l.&Yill
(b)
(a)
8.
226
4 Design Procedures
min
(4.102)
h(X) = 0
It has been shown [4] that the relationship between the problem variables,
objective function and constraint functions can be described by the problem matrix.
An entry V in position i, j of this matrix indicates that function i depends on the
jth subvector of variables. Depending on the variables selected to formulate it, a
problem may have many different structures. In this section only some of the
many possible problem formulations will be considered and classified according to
the structure of the problem matrix.
From the stand point of decomposition, a problem having an additively separable
objective function and a block-diagonal problem matrix (Fig. 4.18) is ideal, since
227
L h(X;)
s
Z=
i=l
(4.103)
i = 1. ... s
resulting in s independent subproblems each having a certain!.{Xi) as an objective
function. It can be observed that any design satisfying the optimality conditions for
each of the subproblems will satisfy those conditions for the original problem.
A nested structure may be often obtained by partitioning the variable vector into
two subvectors Y and X. The resulting problem is to find Y and X such that
Z =f(Y, X)
min
g(Y. X) $; 0
h(Y. X)
(4.104)
=0
This problem can be stated in a two-level form where the lower level variables X
are optimized for fixed values of the upper level variables Y = yo. Satisfying the
optimality conditions for the first- and second-level subproblems in this case
guarantees that the conditions for the original problem are also satisfied.
Complex design problems usually cannot be formulated with a block-diagonal
structure. A more typical structure is block-angular with a number of coupling
variables and/or constraints (Fig. 4.19) . This problem can often be formulated as a
two-level problem. At the upper level. the coordinating second-level variables Y
affect directly the upper level constraints. At the lower level the first-level local
subsystem variables Xi affect directly the lower level constraints.
Assuming additively separable objective function. a problem with only coupling
variables (Fig. 4.19a) can be formulated as: find Y and Xi (i = 1... s) such that
Z = fo(Y) +
i=l
(global constraints)
gi(Y' X)$;O}
hi(Y. X)=O
(4.105)
228
4 Design Procedures
Variables
Xl
X2
X3
X4
Objective function
Constraints
V
V
Fig. 4.18.
It can be shown that a design satisfying the optimality conditions for the first- and
second-level problems will satisfy those conditions for the original problem.
F(X) N = o(X)
(e)
(d)
j=1
i= I, ... ,n
(e)
It should be noted that the optimal solution in this case can be determined directly.
Alternatively, denoting the redundant forces N = Y, the problem can be
formulated in the form of (4.104) as follows: find X and Y such that
if)
(g)
F(X)Y = o(Y)
(h)
229
Variables
Objective function
V V V V
V V V V
Global constraints
V V V V
V V V V
V
V V
V
V
V
V
Local constraints
V V
(a)
V
V
V
(b)
V V
V
V
V
V
(c)
This fonnulation is similar to the integrated problem fonnulation (a), (b) and (c),
except that a two-level optimization is considered here, where the design variables
X are selected at the frrst level for the assumed redundant forces Y =yo. The latter
forces are then updated at the second level and the frrst-Ievel problem is solved for
the modified Y.
Example 4.12. Consider an optimal plastic design of the three bar truss shown
in Fig. 4.15, subjected to two distinct loadings PI and P2 Choosing the force in
member 2 as a redundant (denoted Ni for the ith loading condition), the integrated
problem is to find XT = {Xl' X 2 } and NT = {NI N 2 } such that
Z = 282.8 Xl + 100 X2 ~ min
(a)
(b)
230
4 Design Procedures
design. The goal coordination method, on the other hand, will provide infeasible
intermediate designs. That is, full convergence must be achieved in order to obtain
a feasible design. Another drawback of the goal coordination method is that it is
guaranteed to work, loosely speaking, only for convex programs. Due to these
limitations only the model coordination method will be considered in the remainder
of this section.
Consider the typical formulation (4.105). It has been noted that a natural
approach to solve the problem is to use a two-level optimization procedure where
the optimization of the local (subsystem) variables Xi is nested inside an upperlevel optimization of the global variables Y. The resulting first- and second-level
problems are formulated as follows:
First-level problem. For given values Y =yo the problem is decomposed into the
following s independent subproblems: find Xi such that
gi(YO' Xi) ~ 0
hi (yo, Xi)
(4.106)
=0
=min Zi
(4.107)
H(Y) = fo(Y) +
L Hi(Y)~ min
i=1
(4.108)
ho(Y) = 0
An additional constraint on Y is that the first-level problem has a feasible
solution, i.e., that H(Y) exists. The two-level problem is solved iteratively as
follows:
1.
2.
3.
4.
231
1. The second-level problem must be easy to solve. That is, either the space of Y
must be small enough to search, or else gradients must exist so as to permit
efficient optimization techniques. Calculation of derivatives of the subsystems
optima with respect to Y is demonstrated elsewhere [128].
2. The first-level problem must be formulated in such a way that it can be
decomposed into simple independent subproblems.
One problem associated with the two-level approach is that for some values of Y
there may be no feasible solution to the first-level problem. To overcome this
difficulty we may add constraints to the upper-level problem that help prevent
infeasible solution of the lower level problem, as will be illustrated in Sects. 4.4.2
and 4.4.3. Other means such as an extended penalty function [46] and an envelope
function which replaces a vector of constraints [6, 129] have been proposed.
Example 4.13. The object of this example is to illustrate system oriented
decomposition of a coupled statically indeterminate structure. Using this procedure,
decomposition is associated with dividing the structure into a number of
substructures along boundary lines. The coordinating variables represent the
behavior of the structure (displacements and forces) along these lines, and each
substructure corresponds to a proper first-level subproblem. Consider the
continuous beam shown in Fig. 4.20a (all dimensions are in kilograms and
meters) with the following functions of the four cross-sectional variables Xi
cross - sectional areas
=6 x 10-2 Xi
Moments of inertia
= 1. 8 x
Allowable moments
A;u
.}
10-5 Xi
= 1. ... ,4
= _A;L =9 X 103 Xi
Boundary line
(a)
_
.....
(b)
= 10.0
~c------,~
.....
- _.
-100
=10.0
-~- -----~-I
l=10.0
_I
...------
First substructure
Second substructure
232
4 Design Procedures
-9 X 103
54 Xl
12X2
+300
-400
Xl
X2
X2
X3
36X2
+400
+
-400
0
+400
-300
X3
X4
[54Xl + 12X,
36X2
0
0
(a)
36X2
0
12X2
0
12X2 36X3
36X3 12X3
0
54X4
m~9XlO'
36X,
12X2 + 12X3
36X
o3
36X3
12X3 + 54X4
Xl
X2
X2
(b)
X3
X3
X4
1{"} rOO}
re =
0
rD
-100
(c)
The boundary line through C separates the structure into two substructures and the
chosen coordinating variables are
re= Yl
A3 = -A4 = Y2
(rotation at C)
(bending moment at C)
(d)
(e)
Equation (e) is based on the relation A3 +A4= O. The second equation of (c).
which includes variables of both substructures. can now be decomposed by
separating the contributions of the individual substructures
~
"
-i;
(()
The first and the third equations of (c) include only variables of a single
substructure. For assumed fixed values
11 = Yjo.
Y2
= y20
(a) through (c) can be transformed into the following two independent first-level
subproblems. each with its own variables. constraints. and objective function:
233
The second-level problem is to find Y1 and Y2 such that the first-level solution
exists and the overall objective function (a) is minimized. A similar approach has
been demonstrated also for the force method of analysis [81].
4.4.2 Two-level Design of Prestressed Concrete Systems
Problem Formulation. The object in this section is to illustrate a two-level
optimal design of prestressed concrete systems. Considering a typical indeterminate
structure (Fig. 4.21) , the problem under consideration is to find the vector of
concrete dimensions B, the initial prestressing force P, and the vector of tendon's
coordinates Y such that
(objective function)
aL ~ O'(B, P, Y) ~ O'u
(stress)
DL~
D(B, P, Y)
DU
(displacement)
BL~B ~Bu
(concrete dimensions)
pL~p ~pu
(prestressing force)
yL
(tendon's coordinates)
yu
(4.109)
234
4 Design Procedures
and D =vector of displacements. The fact that the three types of design variables in
this formulation are of fundamentally different nature may produce numerical
problems in the solution process. In addition, the number of variables and
constraints might be large even in simple structures. In the presented two-level
procedure the variables P and Y are optimized at the first level and the variables B
are treated at the second level. That is, the two-level formulation is based on
process oriented decomposition. Considering the formulation (4.104), it will be
shown that the first-level problem can be stated in an LP form, thereby allowing
an efficient solution. For purpose of illustration, uniform concrete dimensions
have been assumed. That is, the relative stiffnesses of the members are constant,
and the bending moments, stresses and displacements (due to both external loadings
and prestressing) are explicit functions of the design variables. The number of
stress and displacement constraints can be reduced by considering only potentially
critical constraints. The latter can be identified before solving the problem, as will
be shown subsequently.
The normal stress constraints for a given cross section are
L
YqP
a <---q -
.4.: (B)
Mq +YqM(P.Y)
W,(B)
<a
-
u
q
(top fiber)
(4.110)
apr =
max
q
(L
a q +Mq)f
- - Yq
W,(B)
Mq)f
aL
aq - Yq
pb = max (L
q
Wb(B)
u= mm. (u +W,(B)
Mq)f
- - Yq
apr
aq
u . (u
(4.111)
Mq)f
a pb = mm a - - Y
q
q Wb(B)
q
and substituting into (4.110) , the following reduced set of stress constraints is
obtained
aL < __P__ M(P. Y) < aU
Pr - Ac(B)
W,(B) - Pr
P
M(P.Y)
u
<---+
<a
Pb -
Ac(B)
Wb(8) -
Pb
(4.112)
235
That is, for any assumed B there are only two stress constraints for each" point,
instead of the original 2 nL constraints (nL =number of loading conditions).
Similarly, the displacement constraints are
(4.113)
in which Dq(B) displacement due to the qth external loading, and D(B, P, Y)
displacement due to the initial prestressing force, both being explicit functions of
the variables. Defming
D;
=max[D;
- Dq (B)] I 'Yq
q
(4.114)
D; ~ D(B,P, Y) ~ D}t
(4.115)
Based on (4.112) and (4.115), the problem (4.109) can be slated as the following
explicit design problem for uniform concrete dimensions [68]: find B,P, Y, such
that
Z=
CcV(B)+CpP~
min
CJ
Ac(B)
W,(B) -
PI
L <
u
- - P- + M(P,Y)<
CJ
Ph - Ac (B)
Wh(8)
Ph
(4.116)
D; ~D(B,P,Y)~D}t
yL~y ~Yu
In this formulation the bounds CJ~, ' CJ~, ' CJ~h' CJ~h ' D;, D}t are functions of
only B. The stress and displacement constraints are written for all potentially
critical points in the structure. The elimination process (4.111) and (4.114) may
considerably reduce the number of constraints in structures subjected to several
loading conditions.
236
4 Design Procedures
Assuming the equivalent load method, the problem (4.109) can be stated for any
assumed 8 8 0 as follows [60] : fmd P and Y such that:
Z=P
--+
min
(stress)
(displacement)
(4.117)
pL<5:P <5:pu
(prestressing force)
yL<5: Y <5:Yu
(tendon's coordinates)
237
(4.118)
To convert the NLP problem (4.117) into an LP. assume the transfonnation of
variables
j= 1... n
(4.119)
Substituting (4.118) and (4.119) into (4.117). the problem can be stated as the
following LP: find Xj (j = 0.1 ... n) such that
z=XO ~ min
L ajXj
$;
O"~
L bjXj
$;
D~
II
O"~
$;
j=O
II
Dft $;
j=O
(4.120)
pL $;Xo $; pU
Similar to the problem (4.116) only two behavior constraints are considered in
this fonnulation at each point instead of 2 nL constraints in the original
fonnulation (4.117).
Second-Level Problem. The object at the second level is to optimize the
concrete dimension B. We first introduce constraints on B that help prevent
infeasible solution of the first-level problem. Considering the stress constraints
(4.112). necessary conditions for a feasible solution (on the basis of the transitivity
property) are
O"PI $; 0" PI
(4.121)
Substituting (4.111) into (4.121). denoting q = 1.2.3.4 for the four tenns in the
latter equation and rearranging yields
238
4 Design Procedures
(4.122)
"(3 M 4
= W;L
L-b
"(30'4 -"(40'3
The right hand side tenns in these equations ( w,L and WbL ) are constant lower
bounds on the section moduli. Similarly, from the displacement constraints
(4.115), necessary conditions for a feasible solution are
U
D pL<D
p
(4.123)
Substituting (4.114) into (4.123) , denoting q = 5,6 for the two tenns in the latter
equation and rearranging yields
(4.124)
in which / (8) = moment of inertia of the concrete cross section; d s, d 6 = the
displacements Ds(8), D 6(8), respectively, assuming / (8) = 1.0 ; and /L = constant
lower bound on / (8). Equations (4.122) and (4.124) are necessary conditions for a
feasible solution of the first-level problem. Thus, the second-level constraints are
W,(8) ~ W,L}
Wb(8) ~ WbL
/ (8) ~/L
( stress)
(displacement)
(4.125)
(side constraints)
Assuming the objective function
z= V (8) ~ min
(4.126)
239
= 0.5,
BIU
= to,
Bf
= 0.1,
= 0.2; the bounds on tendon's coordinates are fL=O.05, yU =Bl - 0.05; and the
bounds on the prestressing force are pL = 10, pU = 100.
Bf
C'l
Bl
....
II)
ci
II
ci
II
~~
8C'l
ci
II
~~
1.00
0.90
llmin
0.80
0.70
~~'--~1------~1----~172~--~~~
0.10
0.15
O. 0
B2
240
4 Design Procedures
The optimal design problem (4.116) is to fmd BIoB1' P, Y Io Y1, such that
(a)
0.10 S B1 S 0.20
lOSP S 100
0.05 S YI S BI - 0.05
0.05 S Y1 S Bl - 0.05
B1!if ~ 0.079
(b)
0.50 S Bl S 1.00
0.10 S B1 S 0.20
z=p
241
min
-554.2
10 ~P
0.05
Y1 ~ 0.95
0.05
Yz ~ 0.95
P(Y1 - Y:z}
1000
569.3
100
(c)
The design space of the problem (b) is shown in Fig. 4.22. The optimal solution
is
B!"un
= {O.92,
Y1 = 0.87, Yz = 0.05. Solving the frrst-Ievel problem for B[max = {l. 00, 0.1}
(maximum depth and minimum width) yields P =58.7. It has been shown [68]
that, depending on the ratio CclCp , the optimal solution is, either at B min (for
high CclCp ratios) or at B1max (for low CclCp ratios).
4.4.3 Multilevel Design or Indeterminate Systems
In most structural optimization formulations, design variables are treated as the
independent variables, while behavior variables are the dependent ones. This
appears to be a rational choice, since for any assumed design the behavior can be
determined uniquely by the analysis equations. However, it has been noted in Sect
1.3.4 that this approach requires multiple repeated analyses for successive
modifications in the design variables. An alternative two-level formulation,
presented in this section, is to choose some behavior quantities as second-level
independent variables. The resulting frrst-Ievel problem is simple and can readily be
solved. The presented approach is intended to reduce the number of repeated
analyses in structures subjected to multiple loading conditions.
Problem Formulation. For simplicity of presentation, continuous beams
will be assumed in the presentation that follows. The problem under consideration
is to find the independent behavior variables Y (bending moments) and the
dependent design variables X such that
Z =f(Y, X) -+ min
M (Y)
MpiX)
ML~M (Y)~MU
M~ ~ Mpt(X) ~ M~
XL ~X~XU
(4.127)
242
4 Design Procedures
max[M~i' Mi(yO)]SMp'i(Xi)SM~'i
xf SXi sxf
in which subscript i denotes quantities associated with the ith element. Since the
number of variables in each subproblem is small and the constraints are expressed
explicitly in terms of the variables, solution of the flfSt-level problem is simple.
To introduce second-level constraints, we have from (4.127)
(4.129)
These constraints are necessary conditions for a feasible solution of the flfSt-level
problem. Thus, the task in the second-level problem is to find Y such that
i=l
i=l
(4.130)
243
Assuming an initial uniform cross section, the elastic envelope moments M are
shown in Fig. 4.23b. Choosing the moments over the interior supports as the
independent variables, Y1 and Y2 , the moments M(Y) are shown in Fig. 4.23c.
For any assumed Y, the moments M I (Y) , M 3(Y), and M s(Y) can readily be
determined by the equilibrium conditions. Two cross-sectional variables, Xii and
X i2 , are assumed for each of the five elements (Fig. 4.23d).
(i
0.02
10.01
(a)
0.01
0.02
0.02
I O.QI
0.01
1
~
tl
i
1
Q""'"
1000
'<''
l3
t4
1000
0.02
q~
"""
t
t
1000
(b)
M2=Y I
(c)
(d)
M4 =Y2
-<f7A\<17!f\~
1.0
STEEL
CROSS - SECTION
Xn
1.0
xf ={Xii
X i2 }
244
4 Design Procedures
5
Z=
(4.131)
in which Ai(X i ) is the cross-sectional area. and li(Y) is the length of the ith
element. selected as the distance between zero moment sections (see Figs. 4.23a
and 4.23c). Formulating the problem in the design space. the number of
independent design variables is 10. Assuming the three-level formulation. the
number of independent variables is two. The fIrst-level problem can be decomposed
into fIve subproblems. each of which having only two design variables. The threelevel formulation is as follows [69].
(4.132)
H=
i=!
i=!
(4.133)
i = 1... 5
Third-level problem. Elastic analysis is carried out and the bounds MiL and MY
are modifIed. if necessary .
245
Example 4.15. To illustrate numerical results, the following values have been
assumed (all dimensions are in tons and centimeters; due to symmetry, Y1 =Y2 =Y)
M~l. = 2200
M;l. = 800.
Xj~ =
xf2 = 10.
xg = xg = 40
Mf =Mf =1810
Mf=M~=0.6x2170.
M =MY =2170
Mf = 0.6 x 1000.
Mf = 1000
The expressions for Aj(Xj) and M pl. (Xj) are based on the cross section shown in
Fig. 4.23d. The lower and upper bounds on Mj are Mf = 0.6Mj (40% maximum
moment redistribution) and MY = Mj , where Mj are the envelope moments.
From the second-level constraints, we obtain the feasible region for Yas
1500
1900
At the ftrst level we ftnd the optimal cross-sectional dimensions for any assumed
Y. Evidently, the feasible region of the ftrst-Ievel subproblems depends on the
assumed values of Y. Solving the first- and the second-level problems, the
resulting optimal moments are
M; =0.79M =785
Elastic analysis of the optimal structure and modiftcation of the bounds ML and
MU (third-level) do not affect the optimal solution. That is, two elastic analyses
are sufftcient to obtain the ftnal optimum [69].
246
4 Design Procedures
1. Passive control systems, where there is no external energy supply that may
generate control forces. Passive viscoelastic dampers and tuned mass dampers,
used in some tall structures to reduce their dynamic response, are examples for
this class of control systems. Similarly, base isolation devices can be used to
reduce buildings seismic response.
2. Active control systems that rely on the availability of an external energy
supply and generally consist of a feedback control system that is designed to
generate corrective forces. Several such systems have been studied analytically
and experimentally. These include active tuned mass dampers, active tendons
and active pulse systems.
Passive control is most appealing from the reliability and maintainability points of
view. Active control systems, on the other hand, are usually expensive and difficult
to maintain, but they can be designed to control the dynamic response more
precisely. Two design problems will be discussed in this section:
1. The problem of optimizing a passive structural control system where the
structure itself is given (Sect. 4.5.1). Some considerations for selecting optimal
control systems, intended to improve the structural behavior under static loads,
are discussed. It will be shown that the problem can be stated in a linear
programming form where the control forces or control gains for displacements
are the independent variables.
2. The problem of improving the optimal design by control variables (Sect.
4.5.2). Potential savings in weight achieved by introducing passive control
devices are demonstrated, and a solution procedure for the optimal design of
controlled structures subjected to static loads is presented.
4.5.1 Optimal Control of Structures
Problem Statement. Consider a given structure subjected to external static
loads, with given displacements ro and stresses (Jo, determined by analysis of the
structure. It is assumed that some constraints are violated due to excessive loads or
other reasons, and it is therefore necessary to introduce control forces or gains for
displacements to improve the behavior of the structure. Ideally, a system of control
forces could be applied in directions opposite to those of the external loads.
However, this solution is usually not practical due to the large number of forces. A
reasonable criterion in selecting a system of control variables is to consider both
the number and the magnitude of control forces.
Mathematically, either control forces, Pc, or control gains for displacements, Y e,
may be chosen as independent variables. Assuming control gains for displacements
247
(4.134)
where the elements of K and S are constant. Considering truss structures for
illusttative purposes, the elements of the load vector R(Yc) are determined by (Fig.
4.24a)
(4.135)
where E, aj and ti are given constants. The resulting load vector can be expressed as
a linear function of Yc as
(4.136)
where the elements of matrix c are independent of Y c. Substituting (4.136) into
(4.134), the following linear expressions are obtained
(4.137)
where the elements of matrices ryand ayare independent of Yc.
In the absence of specific requirements, the objective function Z is assumed as a
linear function of Yc' that is
(4.138)
where the elements of vector Zyare constant. To account for negative values of
Yc' it is possible to express Z in terms of new variables X as
Z=Z~X
(4.139)
248
(a)
(b)
(c)
4 Design Procedures
FWi'/Yffff4l
rU ____ 0
Ycj
~
rU ----
Pci
IE
y--
Yei
R / Ycj) ----
R/Ycj1
l,
~i
R/Ycj)
o _ _ rj2
o _ _ rj2
~I
~i
Fig. 4.14. Truss element: a. Fixed ends, effect of control gains for displacements,
b. Effect of joint displacements, c. Final position.
Z=Z~X~min
(4.141)
Since there is a linear relationship between control gains for displacements Yc and
the corresponding control forces Pc (Fig. 4.24c), the problem (4.141) can be stated
as the following equivalent LP in terms of Pc: fmd Pc and X such that
Z=Z~ X~min
(4.142)
where the elements of Zp, rp and CJp are constant. The two LP formulations can
readily be extended to include structures subjected to multiple loading conditions.
Design Considerations. For statically determinate structures, Y c will not
cause any stresses and forces, and the resulting displacements can be determined
directly by kinematic relations. For statically indeterminate structures, any gains
for displacements Y c will result in self-equilibrating forces and corresponding
stresses in the structure. These forces and stresses can be introduced by a reduced set
of gains for displacements corresponding to a selected set of redtmdant forces N.
Assuming Pc = N, the displacements and the stresses due to Pc are given by
249
(4.143)
where rN and aN are matrices of stresses and displacements, respectively, due to
unit value of N in the primary structure. It can be observed that equivalent stress
distributions can be obtained by various selections of the redundant forces N. The
number of possible selections of N might be very large in structures with high
degree of statical indetenninacy. Unlike the stresses, the displacements depend on
both the magnitude and the location of the control variables. That is, various
selections of N may result in equivalent stress distributions but different
displacement fields in the structure. Thus, one consideration for selecting the set of
control variables is their effect on the displacement constraints.
To find the optimal location and magnitude of the control variables, it is
possible to assume fIrst control forces or control gains for displacements in all
potential locations (members). This initial set of variables is optimized by solving
the LP problem (4.141) or (4.142). It is expected that due to the nature of the
problem, only a small number of the variables will be non-zero at the optimum.
Indeed, experience has shown [79] that usually there are only a small number of
non-zero control variables at the LP optimum. The LP procedure can also identify
situations where the problem has no feasible solution. For cases where the LP
solution does not provide a sufficiently small number of non-zero control
variables, an alternative solution procedure has been proposed [79].
Example 4.16. The seven-bar truss shown in Fig. 4.25 is subjected to a single
loading condition of two loads. Assume E = 30,000, aU = -aL = 25 and an upper
bound on the vertical displacement at joint 0, r U = 2.0. The cross-sectional areas
and stresses due to the given loads are shown in Table 4.7.
The given vertical displacement at joint 0 is ro = 2.21. That is, the constraints
aL $; a6 and ro $; r U are violated. Assuming fIve cases of a single control force (at
members 1, 2, 4, 5 and 6), the resulting constraints on the control force are
summarized in Table 4.8. Identical force distributions can be obtained by
assuming each of the five control forces if only stress constraints are considered.
The optimal solutions in this case are Pel = -5.80, P e2 = -5.80, or P,;04= -5.80. For
both stress and displacement constraints the only optimal solution is P e2 = -7.55.
Table 4.7. Cross sectional areas
Member i
1
2
3
4
5
6
7
a;
10.0
10.0
4.0
4.0
8.0
4.0
6.0
ai
17.66
-22.23
25
-5.86
21.82
-27.05
23.57
(J0i.
seven-bar truss.
250
4 Design Procedures
o
100
360
Fig. 4.25.
100
360
Seven-bar truss.
Stress
constraints
-18.00 S Pel S -5.80
-18.00 S Pe2 S -5.80
-18.00 S P c4 S -5.80
8.20 S PeS S 25.44
8.20 S Pc6 S 25.44
Displacement
constraints
Pel S 7.90
P e2 S -7.55
Pc4 S -21.68
PeS S 30.66
P c6 S -54.56
Stress and
displacement constraints
7.90SPel S -5.80
-18.00 S Pe2 S -7.55
-18.00 S Pc4 S-21.68
30.66 SPeS S 25.44
8.20 S P c6 S -54.56
F(X, Y)N
=~(X, Y)
(4.144)
251
Considering the second approach, it has been shown [76] that the final optimal
solution is often identical or very close to the LB solution. Therefore, a small
number of iteration cycles is needed to arrive at the optimum and the solution
efficiency is greatly improved. Starting with the LB solution X~. N~. Z~, we
first check the compatibility conditions. In cases where the latter conditions are
violated it is still possible to find Y and N such that the LP solution is the final
optimum, as will be shown subsequently.
Assuming the formulation (4.144) and the given LP solution X~, the problem
of optimal control (OPC) is formulated as follows: find Y such that
(4.145)
In this formulation the implicit expression for N, obtained from the compatibility
conditions, is substituted into the stress constraints. One approach to obtain a
feasible solution of the OPC problem is to use the KS function (4.39), where all
the constraints (4.145) are expressed in the form #Y) :5: O. If the optimal solution,
252
4 Design Procedures
, Y~pc, does not satisfy all the constraints, no feasible control forces can be
achieved and it is necessary to modify the LB problem constraints. Several methods
can be used for this purpose [76].
Based on these considerations the following two-level design procedure can be
used:
1. The LB problem (1.68) is solved to obtain X~ (fIrst-level solution).
2. The ope problem (4.145) is solved to obtain Y~pc satisfying all the
constraints (second-level solution).
3. If all the constraints are satisfIed, the fInal optimum is reached. Otherwise the
LB constraints are modifIed.
Steps 1 through 3 are repeated as necessary until the fInal solution is reached. It
has been observed that in many cases the optimum is reached in one iteration cycle
[76].
Example 4.17. The grillage shown in Fig. 4.26a is subjected to a single load
P, acting at point C. The structure is simply supported at points A. E. F and J. A
uniform cross section with
EI
Cl=-=2.5
GJ
has been assumed, where EI is the flexural rigidity and GJ is the torsional rigidity.
A single design variable, X, representing the bending moment capacity is assumed,
and the objective function is Z = 36 X .
Two types of control devices are demonstrated in this example:
(a)
:
l
3.75
0
3.75
rcJS:J
Y=F
3.75
(b)
:}
3.75
(c)
[ltD 7
--.JfY=5 0
Fig. 4.26.
Grillage example.
253
::J:::::~
(a)
lB75P
J.B75P
1111111111111111111111111111"'111111111111111111111111111
O.7P
(b)
I.B75P
3.05P
Fig. 4.27. Optimal moment diagrams, grillage example: a. LB and NLP solution
b. UB solution.
Xi
1.875P
1.875P
3.050P
Z*
67.5P
67.5P
l09.8P
254
4 Design Procedures
min
(displacement constraints)
(stress constraints)
(4.146)
255
In general, upper and lower bounds on design and behavior variables are assumed to
be constant. If stability of members is considered, the elements of oL are functions
of both Y and X. This situation will be discussed later in this section.
The difficulties associated with the common problem of cross-sections
optimization (Le., the implicit nature of the behavior constraints and the large
numbers of variables and constraints) are magnified in problems of geometrical
optimization. That is, the problem size is increased and the need for multiple
repeated analyses is a major obstacle in applying optimization methods to large
scale structures. In addition, changes in the geometrical variables may be of
different magnitude than changes in cross sections. Experience has shown that
combining these two types of variables, which are of fundamentally different
nature, may produce a different rate of convergence and ill-conditioning problems.
Thus, two-level optimization is most suitable.
In this section two approaches for solving the geometrical optimization problem
will be demonstrated:
- Simultaneous optimization of all design variables (Sect. 4.6.1).
_ Multilevel optimization, where the two types of variables are optimized in
different levels (Sect. 4.6.2).
*):5:0" U
0"L :5:0" * +VO"x(X-X
(4.147)
XL:5: X :5:Xu
in which all derivatives are computed at X*, Y*, the elements of the bounds on Y
and X are given by
.L
{y
= max.
L
Y - 6Y
.L =max.
{XL
X
- 6X
.U
{yu
= mm.
Y
.U = mm.
. {XU
+ 6Y
X + 6X
(4.148)
U
256
4 Design Procedures
and _!!lyL. !!lYu _IlXL !!lXu. are vectors of given move limits [see (3.5.2)].
Using the SLP procedure (Sect. 3.1.3). the LP problem is solved iteratively. X
and Y redefined each time as the optimum solution to the preceding problem. In
addition. it might be necessary to modify the move limits in each iteration. As
noted in Sect. 3.1.3. we may start with relatively large move limits and end up
with limits corresponding to the required accuracy. One possible procedure to deal
with the geometrical variables is to increase or decrease the corresponding move
limit for each step of iteration. If a coordinate change goes in the same direction
twice we increase the move limit. If. instead. the change is opposite to the one
before. then we decrease the move limit.
Example 4.18. Consider the symmetrical bridge truss shown in Fig. 4.28 (all
dimensions are in newtons and meters). The bridge is subjected to five alternative
loading conditions Pq =3 X 106 (q = 1.. 5). The geometrical variables are the
coordinates of the five arch joints. which lie initially on a parabolic arch.
To consider stability constraints. assume a factor of safety against buckling. 'YE.
in the elastic range. In the plastic range. a parabolic design formula is suggested
that guarantees a continuous transition from the elastic range [104]. In order to
know whether we are in the elastic or the plastic range. defme the limiting forces
AP
= - (~r
E~E
(a)
where aP is the stress at the limit of proportionality between stress and strain and
a is a cross-sectional constant given by
(b)
rG
257
(d)
For the suggested design formula in the plastic range, the necessary area
given by
Xmin
is
(e)
equal to zero. From (e) it follows that for the allowable compressive forces in the
plastic range
for
Determining AL by (d) or by if) we may find crL and derivatives of the allowable
compressive stresses can now be computed. The corresponding terms of
cr*L + Vcr;L(X - X*) + Vcr;L(y - y*) are substituted in (4.147) in place of crL.
Four cases of constraints have been solved, as summarized in Table 4.10 [104].
The assumed fixed bounds on stresses, for the cases in which stability
considerations are neglected, are
considered are
'YE
= 2.5
a.
= 0.8
= 2.1xlO11
= 7850 kg/m 3
Stability constraints
Not considered
Not considered
Considered
Considered
Maximum displacement
Not considered
0.01
Not considered
0.01
258
4 Design Procedures
._
. . 6.29
1~13.62
~-
(c)
1. 5.78
Cd)
.
cases a , b, c, d [104].
Fig. 4.29. Optimal geo metrles,
259
a
1868
1656
11.3
b
3526
2911
17.4
c
3141
2905
7.5
d
3668
3315
9.6
The optimal geometries are shown in Fig. 4.29. It can be seen that the horizontal
changes of the positions of the joints are toward the supports for the four cases. To
illustrate the effect of changes in the geometry, the above four cases were
optimized also for the fixed initial geometry shown in Fig. 4.28. Comparison of
results, given in Table 4.11, shows that savings in weight due to changes in
geometry are 7.5 to 17.4 percent. Some of the optimal solutions are not fully
stressed designs.
4.6.2 Approximations and Multilevel Optimization
It has been noted in Chap. 3 that approximation concepts can greatly reduce the
computational effort involved in cross-sections optimal design. Similar
approximations may be applied also in geometrical optimization. Multilevel
optimization, discussed in Sect. 4.4 is most suitable for the problem under
consideration, where different types of variables are involved. It combines efficient
suboptimization for member sizes, reduction in the number of design variables
optimized simultaneously and improved convergence. In this section,
approximation concepts and multilevel optimization are combined to achieve
effective solution procedures.
Consider again the geometrical optimization problem (4.146). Assuming the
force method of analysis, the problem can be formulated as follows [see (1.64)] :
find Y and X such that
Z=j(Y,
X)~min
(4.149)
260
4 Design Procedures
261
If displacement constraints are not considered, Steps 2 and 5 (the fIrst level) are
reduced to a direct determination of the cross-sectional variables. The choice of Y
in Step 3 (the third level) is subject to the side constraints on Y.
It should be noted that all intermediate values of the variables Y (two-level
formulation) or the variables Y and N (three-level formulation) are feasible. That
is, the iteration can be terminated always with a feasible - even though nonoptimal
- solution, whatever the number of cycles. While the three-level formulation may
prove useful for some specifIc applications, the two-level procedure is usually
most effective [66].
Example 4.19. The symmetric four-bay frame shown in Fig. 4.30 is subjected
to a single loading of six loads (all dimensions are in tons and meters). The
allowable stresses are aU = -aL = 15,000. For simplicity of presentation, the
following relationships have been assumed for the three cross-sectional design
variables
i = 1,2,3
in which ai = the i th cross-sectional area ; Wi = the modulus of section ; and Ii =
the moment of inertia. Two geometric variables (Y1 and Y:0 have been considered
with the following side constraints
Assuming a grid of points with intervals of 2.0 in the geometric variables space,
the lower bound solution of the AEP is [66]
Y1 = 12.0
Y2 = 4.0
Z = 602
The exact solution of the cross-sections optimization problem for this geometry is
Z =682, and the true optimum for the assumed grid is
I 5.0 I 10.0 I
lIE
lIE
)I
lIE
10.0
2.0--o~_-I---r---l_..,----L_.-----,-_..-_2.0
x,
x,
XI
X,
I 20.0-Y I
....
;Jr .'1(
I';
lIE
I';
I 20.0-Y I
)1'.
lIE
..
262
4 Design Procedures
Y1 = 14.0
Y2 = 4.0
Z = 668
That is, the true optimum is only 2% lighter than the final result obtained from
the AEP solution (Z = 682).
263
Most studies on optimal topologies deal with truss structures. This may be
attributed to the fact that the truss by its very nature is most suitable for
optimization of the topology. It possesses usually many nodes and elements that
can be deleted or retained without affecting the functional requirements. In addition.
the truss is a relatively simple. yet nontrivial. structure. It is therefore an ideal
system for the demonstration of some properties and characteristics associated with
optimal topologies. Although most of the discussion in this section is related to
trusses. the extension to other classes of structure is often straightforward.
While the displacement method formulation is the prevalent structural analysis
tool in current computational practice. the force method formulation is adopted in
many topological optimization problems due to several reasons:
a. The objective function represents the weight and can be expressed in linear
Assume a grid of points that may be connected by many potential truss members
to form a ground structure with a finite number of members. Considering the
above mentioned simplifications. the problem can be stated in the LP form (1.69):
fmd the cross-sectional areas X and the members' forces A such that
Z=lTX~min
(objective function)
(stress constraints)
(4.150)
(side constraints)
CA
=R
(equilibrium conditions)
264
4 Design Procedures
where the stress constraints and the equilibrium conditions are formulated for all
loading conditions. It has been noted that the LP problem (4.150) can be expressed
in terms of the redundant forces N as [see (1.68)]: find X and N such that
Z=(fX~min
(4.151)
In this formulation the equality constraints have been eliminated and the number of
variables is reduced (only the redundant forces are considered as independent
variables instead of all member forces). The stress constraints are formulated for all
loading conditions.
One advantage of the LP formulation is that the global optimum is reached in a
finite number of steps. Thus, large structures with many members and joints can
efficiently be solved. In addition, problems of singular optimal topologies are
eliminated and some effects of the optimal solutions can be evaluated directly.
In the case of zero lower bounds on cross sections
(4.152)
the LP method has the ability to make unnecessary members vanish from the
structure to obtain the minimum weight design. The optimal structure in such
cases satisfy equilibrium and stress constraints, but it might not satisfy
compatibility conditions or may represent unstable configuration under a general
loading. That is, the LP solution is not the final optimum and some modifications
might be required. However, this solution may be viewed as a lower bound on the
optimum. The various types of the optimal topologies are discussed subsequently
in Sect. 4.7.2. Some properties of these topologies are introduced in Sect. 4.7.3
and two-stage design procedures are presented in Sect. 4.7.4.
265
These results illustrate how mobile topologies might be obtained for particular
loading cases.
100
Fig. 4.31.
Seven-bar truss.
266
4 Design Procedures
;
ffi
P2 (= PI/Z)
I
I
(a)
Fig. 4.32.
(b)
~-P2
(c)
(d)
267
(4.157b)
It will be shown in example 4.21 that elimination of members might change the
design space, if the accurate stress constraints (4.156a) are considered. Problems of
discontinuity due to deletion of members cannot occur if the linear approximate
constraints (4. I 57a) are assumed.
Assuming the accurate displacement constraints and decreasing the cross section
of a specific member, its relative stiffness and internal forces are reduced.
Consequently, its contribution to the accurate displacement expression (4.156b)
might approach zero. That is, deletion of a member will not result in discontinuity
or singular solutions similar to those obtained in the case of stress constraints.
Considering the approximate displacement constraints (4.157b), the result might
be meaningless in cases where members are eliminated from the structure [77].
Another difficulty that might be encountered in topological optimization is that
several local optima representing different topologies might exist. Local optima
representing three different topologies are shown in Fig. 1.17 for the grillage of
Fig. 1.16a. Evidently, local optima do not exist in the LP problem.
268
4 Design Procedures
Example 4.21. Consider the three-bar truss shown in Fig. 4.33 and subjected
to a single vertical load. The members length is [=l.0, and the allowable stresses
are crU =-crL= 20.0. The design variables are X I =cross-sectional area of member
1, X2 =cross-sectional area of members 2 and 3, and the accurate stress constraints
(4.1500) are
- 20 XI
30 Xl
3XI + X2
- 20 X2
14.14 -
- 20 X2
-10 +
20XI
42.42 XI
3XI + X2
30 XI
3XI + X2
20X2
(a)
20X2
- 20 XI
- 20 X2
14.14 - 1.414 N
- 20 X2
-10 + N
20 X2
(b)
20 X2
The design spaces for the accurate stress constraints and for the approximate
stress constraints are shown in Fig. 4.34. To illustrate the singularity
phenomenon, consider a parametric objective function
Z = aXl + X 2
(c)
10.0
269
X,
(a)
~
?
'"-:"
(I)
0.5
1.0
1.5
X,
2(q. ?: 3)
X,
!i
0.5
X,
B.
Point A:
Xl
=0.5,
X2 =0
(for a S 3)
Point B:
Xl
=0,
X2 = 1.5
(for a
(d)
~
3)
It should be noted that at point B member 1 and the corresponding active constraint
are eliminated, and the line segment BC is included in the feasible region.
Therefore, the solution (d) is incorrect for certain a values and the true optimum
points are given by
(for a S 1.414)
Point A:
Point C:
Xl
=0,
X2 =0.707 (for a
(e)
~
1.414)
That is, for 1.414 < as 3, the solution process will converge to point A and for
a ~ 3 the solution process will converge to point B while the true optimum in
both cases is at point C.
For the approximate constraints (b) a two-dimensional feasible region can be
introduced for any assumed N. Eliminating the latter variable, the resulting design
space is shown in Fig. 4.34b and the solution process converges to the true
270
4 Design Procedures
optimum points (e). Since both points A and C represent statically determinate
structures, compatibility conditions can always be satisfied and the approximate
solution is the final optimum.
Multiple Optimal Topologies. Solving the LP problem it is possible that,
for certain geometries, identical optimal objective function values are obtained for
multiple force distributions in the structure. The various optimal force
distributions usually correspond to several different optimal topologies. Some
properties of such particular geometries, where the optimal objective function
value becomes independent of the force distribution, are demonsttated subsequently.
For purpose of illustration, assume a structure subjected to a single loading
condition such that each element is fully stressed. From (4.157a) we obtain for the
force in the ith member
Ap., + A I,.
N =a~" X.
nl
(4.158)
af
-+ (Api
~ t
Z = .J
T
ANi
N)
i=l ai
(4.159)
(4.160)
~ t
Z = Zo + .J
-T ANiT N
i=l a i
(4.161)
The objective function will become independent of the force distribution in the
structure if
(4.162)
The conditions (4.162) form a system of nR equations (nR = the number of
redundant forces) expressed in terms of the geometric parameters of the structure. In
some cases it is possible to find a geometry satisfying these conditions, where
multiple optimal topologies (MOT) are obtained. MOT occur in cases where the
271
objective function contours are parallel to the boundary of the feasible region, and
the LP problem possesses an infinite number of optimal solutions and several
corresponding optimal topologies.
Situations where MOT occur can be identified from the LP solution. Consider
the LP problem where all inequalities are transformed into equalities by adding
surplus or slack variables S. Assume a basic feasible solution, given in the
canonical form
(4.163)
where the elements of a, b, c and Zo are constant, and subscripts B and N denote
basic and nonbasic variables, respectively. The LP solution is optimal if Cj ~ 0
(j=I, 2, 3). If the latter conditions hold and at least one C; equals zero, then the
problem might have multiple basic optimal solutions representing MOT. If C; < 0
at least for one nonbasic variable, the solution is nonoptimal and an improved
basic feasible solution with a corresponding topology can be introduced.
Assume an MOT geometry where all basic optimal solutions with a
corresponding set of T optimal topologies have been determined. The basic
optimal topologies can be used now to introduce new optimal topologies, not
corresponding to the basic solutions. This can be done by assuming linear
combinations of the basic feasible solutions X t , Nt [78]
L atX t
T
X =
tal
=L
(4.164)
atNt
t=l
o~
T
at
L at = 1
t= 1, ... ,T
(4.165)
t=l
Equations (4.164) and (4.165) provide multiple optimal cross-sectional areas and
force distributions corresponding to all optimal topologies. New combined optimal
topologies are obtained for some at values other than 0 and 1. Based on these
definitions, any specific basic optimal topology is given by a certain at equals
272
4 Design Procedures
unity and all the remaining at equal to zero. The basic optimal topologies are
usually statically detenninate structures, therefore the LP solution is the final
elastic optimum.The combined topologies, on the other hand, represent SIS but
compatibility conditions might be satisfied.
An important property of the MOT geometry is that new optimal topologies are
introduced from existing basic optimal topologies by combination rather than the
common ground structure approach of elimination.
Example 4.22 [78]. The fifteen-bar symmetric truss shown in Fig. 4.35 is
subjected to three loads P acting simultaneously. Assume nonuniform depth with a
single geometric variable Y, allowable stresses (Ju=-aL=20.0 and P=10.0 (all
dimensions are in kips and inches). Two SDS basic optimal topologies are
obtained for Y = 100 (Figs. 4.300, 4.36b), and a combined SIS optimal topology
(Fig. 4.36c) is introduced by
Variation of Z" with Y and the corresponding optimal topologies are shown in
Fig. 4.37. A mobile topology (Fig. 4.36d) is obtained for Y 75, where active
constraints in the internal members change from tension to compression.
Assuming unifonn depth Yand two fixed supports (Fig. 4.38), nineteen optimal
topologies can be introduced (Fig. 4.39). The five basic SDS optimal topologies
are shown in Fig. 4.39a, four of which being mobile and two are asymmetric
structures. However, these two types of structure are needed to introduce the
complete set of fourteen combined SIS optimal topologies (Fig. 4.39b), five of
which also being mobile structures. Variation of Z' with Y is shown in Fig. 4.40.
It can be seen that the optimum for fixed supports (Z' = 632.4), is significantly
better then the optimum for nonfixed horizontal support (Z' = 836.7). In both cases
Z is not sensitive to changes in Y near the optimum. It has been found that
compatibility conditions are satisfied for the combined SIS optimal topologies.
4x 100
(a)
(b)
(e)
(d)
z
1400
1300
1200
1100
1000
900
~~~~~~~~~~---y
Fig. 4.37. Variation of Z with Y, fifteen-bar truss, nonuniform depth.
273
274
4 Design Procedures
YI~
4 x 100
.It
(a)
I'
~
_.:.11'-
~
-
, I "
/IVTVr'\.
Lj,~~~
(b)
, I "
L_~:l_~
II
10
~
/--~~-~
12
13
14
15
16
17
18
I "
19
275
z
Non fixed
horizonlal support
900
800
1'"
167.3
700
Z' 632.4
600
L,
1'" .126.5
I
50
75
a, XL = ( > 0). No members can be eliminated from the ground structure and
the accurate NLP formulation (4.156) will give the optimum of this problem.
It has been noted that the approximate formulation (4.157) usually will give a
better solution, since compatibility conditions are not considered. Assuming
mixed formulation, with approximate stress constraints (4.157a) and accurate
displacement constraints (4.156b), the resulting frrst-stage solution will be in
between the former two cases.
b. XL = O. It has been shown that the accurate formulation (4.156) might lead
to incorrect results in the case of a singular optimum. To overcome this
difficulty, it is possible to assume the mixed formulation, or the approximate
LP formulation (4.151), where compatibility conditions and displacement
constraints are neglected. The approximate displacement constraints cannot be
used if members are eliminated from the structure.
The various possibilities for the first-stage formulations are given in Table 4.12
[77]. In summary, approximations assumed at the first stage may include:
-
276
4 Design Procedures
Size
constraints
XL =
XL=O
Stress
constraints
Accurate
Approximate
Approximate
Accurate
Approximate
Approximate
Displacement
constraints
Accurate
Accurate
Approximate
Accurate
Accurate
Neglected
The criteria for selecting the type of approximation could reflect both the solution
efficiency and the chance to arrive at the true optimum.
Several procedures might be assumed at the second stage for modifying the flfStstage approximate solution. In case a (XL> 0), it is possible to eliminate
members with X =XL and modify the design accordingly. Accurate formulation of
both stress and displacement constraints is needed at this stage. In case b ( XL =0),
modification of the first-stage solution is required to account for displacement
constraints and compatibility conditions. The accurate NLP problem can be solved
by available methods (such as those presented in Chap. 2) for the given topology
obtained by the LP solution.
Example 4.23. The ten-bar truss ground structure shown in Fig. 3.8a is
subjected to two loads acting simultaneously. The allowable stresses are aU =-oL =
25, the modulus of elasticity is 30,000 and the upper bound on the vertical
displacement at joint A is DU =2.0. To illustrate the effect of various constraints
on the optimum, the following cases have been solved [77]:
A. Only stress constraints.
(a)
(b)
(e)
277
Member
1
2
3
4
5
6
7
8
9
10
Z
B
7.99
0.1
8.06
3.94
0.1
0.1
5.74
5.57
5.57
0.1
15,932
A
8.0
0
8.0
4.0
0
0
5.66
5.66
5.66
0
15,840
C
9.9
0
8.0
4.94
0
0
5.66
7.0
7.0
0
18,210
D
9.9
0.1
8.0
4.94
0.1
0.1
5.6
7.0
7.0
0.1
18,343
The displacement constraint significantly affects the optimum but does not
change the optimal topology;
the lower bound constraints do not appreciably affect the optimum.
To illustrate the effect of the topology on the optimum, optimal solutions for
various topologies (Fig. 4.41) are compared in Table 4.14. It can be seen that
topology (4.41a) is much better than topologies (4.41b) and (4.41c).
Optimal Topologies or Controlled Structures. An alternative approach
for the second-stage solution is to apply a set of control forces such that the
accurate constraints will be satisfied at the fl1'St-stage optimum [86]. The rationale
of this approach is that the approximate optimum is usually better than the
accurate one, since some constraints (compatibility conditions and displacement
constraints) are neglected. The required control forces can readily be determined by
solving an LP problem, as shown in Sect. 4.5.1. The optimal solution will give
the minimum magnitude and number of control forces needed to convert the
infeasible frrst-stage solution into a feasible one.
Table 4.14. Optimal cross-sections, effect of topology, ten-bar truss.
Member
1
2
3
4
5
6
7
8
9
10
Z
8.0
4.0
0
4.0
4.0
4.0
12.0
4.0
4.0
8.0
4.94
5.53
5.53
4.0
15.1
6.16
6.16
5.66
5.66
5.66
0
11.31
5.66
11.31
0
5.66
5.66
7.0
7.0
0
11.31
7.81
12.33
0
8.72
15,840
17,276
17,276
18,210
20,028
22,803
278
4 Design Procedures
Graphics
display
scrcen
279
The algorithm is stopped. The output includes the current values of the
geometrical variables and also a file containing all the information needed to
restart the algorithm.
The optimality criteria method is used to find the optimum weight of the
structure obtained when the algorithm is stopped, taking into consideration the
minimum size, stress, and displacement constraints.
The unconstrained minimization algorithm is restarted with its input being the
calculated optimum weight and the restart file saved when it was stopped.
280
4 Design Procedures
When the process converges the value of the geometrical variables and the
optimum weight are output. and the optimum geometry is displayed on the screen.
4.8.2 Graphical Interaction Programs
Two display programs are used for the graphical interaction (Fig. 4.42):
Q.
a. The weight against the number of cycles for the optimality criteria method.
h. The weight against the value of a geometrical variable and the weight against
the number of uni-dimensional minimizations for the UM method.
4.8.3 Design Procedure
Using interaction, the display programs, and the three optimization methods,
several alternative procedures can be used to optimize a structure whose topology,
281
geometry and member sizes are treated as variables. These alternative procedures
could differ by the order in which they use the optimization methods and by the
amount of interaction involved. The more the engineering experience of the user,
the more he could depend on his own judgement to guide the design. The four steps
of the procedure described here are: establishing starting structure, establishing
preliminary topology and geometry, establishing final geometry and establishing
fmal topology and member sizes.
Establishing Starting Structure. A grid set covering the design space is
first introduced. Obviously, this grid set must include the load and support points.
A starting topology is established by connecting the grid points with as many
members as desired. The designer may connect only those members which seem to
be reasonable, based on engineering and practical considerations. Equal initial
starting sizes are assigned and the applied loads are introduced. All these actions are
done interactively on the screen which provides an instant graphical check of the
data. The consuaints applied to the structure are also defined.
Establishing Preliminary Topology and Geometry. The purpose of
this step is to establish a preliminary topology by deleting members from the
starting structure built in the first step. Grid points are also deleted if all the
members connected to them are deleted. This indirectly establishes a preliminary
geometry.
Taking into account all the constraints, the OC program is first applied to
evaluate the allowable compressive stress. If buckling and displacement consuaints
are not considered, this substep is not needed. The LP program is then applied. The
structure is displayed, after deleting members that reach zero size and grid points for
which all connected members have been deleted. If the resulting topology is
mobile, members that are needed to produce stability are interactively added.
Finally, the OC program is applied to consider the compatibility conditions and
the displacement constraints.
Establishing Final Geometry. The geometrical variables are chosen, based
on practical considerations. We may impose symmetry on the structure and decide
which grid coordinates or geometrical dimensions will be treated as independent
variables. The remaining coordinates or dimensions could either by linked to the
ones taken as variables or could be fIXed. The unconstrained minimization method
is applied, with the optimality criteria method as its function calculator. The input
is the topology obtained in the previous step and equal initial starting sizes. The
resulting geometry can be rounded off to practical dimensions; the optimality
criteria method is used to obtain the weight of the rounded-off structure.
Establishing Final Topology and Member Sizes. The purpose of this
step is to apply fmal interactive changes to the structure. Starting with the
structure obtained in the previous step, the designer can interactively delete or add
members to produce a better structure based on his experience and taking into
consideration the construction limitations. For each of the alternative structures
resulting from the deletion or addition of members, the optimality criteria method
is applied. The optimum weight of each of the alternative structures is recorded.
282
4 Design Procedures
The final optimal structure chosen is not necessarily the one with the least weight;
for there could be a structure, with slightly higher weight than the least weight
one, but still considered better because of other engineering or construction
considerations. In this final step of the procedure the designer uses his experience
to the utmost advantage in moulding the structure to its fmal design.
Example 4.24. Consider the transmission tower with the initial grid set,
including the load points and the supports, shown in Fig. 4.43a. The following
design data have been assumed:
-
a;E
7tEX (d=~
8l;
t
t) = 7.87 x 10
+ d
X
-t
l;
Allowable compressive stresses for the kth cycle, c:#/c), are computed by
This average is taken to prevent oscillations in the values of oL from one cycle
to the other.
Displacement limits: DU = 1.0 in the direction of P 2
Minimum size XL = 0.1.
Loading conditions (Fig.4.43a): loading a: PI + P 2 + P 3 + P4 ;
loading b: PI + P 2
The design procedure for this example involved the following steps [131]:
-
Starting with the 8 point grid set, 19 members were interactively built (Fig
4.43b).
Three cycles of the optimality criteria method were applied to obtain the
starting allowable compressive stresses for the LP run.
The LP was applied and members reaching zero size were deleted (Fig 4.43c).
The optimality criteria method was applied to introduce a compatible structure,
taking into account the minimum size, stress and displacement constraints. The
final weight obtained is 4.487.
Considering the structure as symmetrical around a vertical axis, the width at the
base is taken as a geometrical variable. The width at the top is fixed and the
width at intermediate stories is taken as varying linearly between them. Using
as input the topology shown in Fig. 4.43c and equal starting sizes, the UM
program is applied. The final weight obtained is 1.349 and the width at the base
is 152.60. Rounding off the width at the base to a practical dimension of 153,
the resulting structure is shown in Fig. 4.43d. Applying the optimality criteria
283
method to the rounded-off structure, the weight practically did not change. This
step represents a 70 percent saving in weight, as compared to the starting
geometry whose optimal weight was 4.487.
=40 t
P3
P =5
4~
t.<;=40
P=15
~
80
0
.2
.2
80
BC
(a)
-.!L+
(b)
(c)
(e)
(j)
284
4 Design Procedures
Structure
Figure 4.43c
Figure 4.43d
Figure 4.43e
Number of members
13
13
8
Weight
4.487
1.349
1.348
Exercises
4.1 Consider the three-bar truss shown in Fig. 1.11. Choose the force in member
2 as a redundant, N, and consider only the load Pl. Assume the bounds on stresses
aU =- oL =40.
b. Assuming the moments over the interior supports B and C as redundants N 1 and
N 2, respectively, formulate the LP problem (4.18).
c. Assuming N 1 =N 2 =N, and M pn =Mpo. =M pl3' illustrate graphically the
feasible region and find the optimal solution.
Exercises
5.0
10.0
5.0
00
0)
0@
>L
100
285
100
200
600
Fig.
4.44.
4.3 The frame of example 4.3, shown in Fig. 4.3, is subjected to two loading
conditions: loading 1 - the vertical loads, and loading 2 - the horizontal loads.
Choose redundant forces as shown in Fig. 4.5.
h. Solve the problem by an LP computer program and find the optimal solution.
1.0
I
yo.s
:1:
20.0
Fig.
4.45.
.. Y l
10.0
.1
20.0
10.0
286
4 Design Procedures
Boundary
line
l20
llO
M"lJ
200
ZS
Mpl/
M,,12
200
200
ZS
D
! 40
M,,12
6.
E
200
Second substructure
First substructure
Fig.
po
4.46.
4.6 The continuous prestressed concrete beam shown in Fig. 4.45 (all dimensions
are in tons and meters) is subjected to the following loading conditions:
l. P + (dead load = 2.0 tIm); bounds on stresses aU = ISO,
2. 0.8P + (total load 3.0 tIm); bounds on stresses aU 0,
crL =- 1500.
crL = - 1200.
The design variables are P, YI> Y2 and the tendon is assumed to be parabolic. The
constraints are related to the stresses in the cross sections of Yl and Y2 and to the
design requirements
j = 1,2
b. Show graphically the feasible region for Y2 = 0.9 and find the optimal P and Yl.
Z=400L M pli
i=l
where M pli (i= 1, 2) are the plastic moments for the two substructures. Choose the
moment Me as the coordinating variable. Illustrate graphically the variation of Zuun
with the coordinating variable.
Exercises
287
"I
~I
l",,,,l,,,,,~,~:,J;:lI
I
I
~~ 1
V".
I'. -
~1
I.
Fig.
i.
8.0
_I.
A...J
I.
8.0
Y '1'
I 15.0 -2Y . 1
. . .Y. .
15.0
i
Se..:tion A-A
8.0
4.47.
4.9 Given the reinforced concrete system of a slab, two beams, and eight identical
columns shown in Fig. 4.47 (all dimensions are in tons and meters). The slab is
subjected to a uniformly distributed load of 2.0 tlm Z and the design variables are
Xl' Xz, X3 and Y. Consider the following constraints:
Design constraints,
3.5 ~ Y ~ 6.5
Slab constraints,
Ms(Y) ~ 190X1Z
Beam constraints,
Mb(Y) ~ 38Xi
Column constraints,
~(Y) ~650xi
in which Ms(Y), Mb(y), Ac(Y), = maximum bending moment at the slab support,
maximum bending moment at the beam support, and maximum load at the
columns, respectively. The objective function represents the concrete cost
Z = 400X1 + 20Xz + 130xi ~ min
Formulate the ftrst- and second-level problems to be solved by model coordination.
Compute Zmin for Y = 3.5, 5.0, 6.5, and ftnd the optimal solution by quadratic
interpolation. Use the following data:
y
3.5
5.0
6.5
9.0
5.0
7.5
90.0
70.0
62.0
123.7
96.8
84.7
288
4 Design Procedures
Z =282.8XI + l00Xl
where Xl and Xl are design variables representing the cross-sectional areas. The
modulus of elasticity is 30,000, and the allowable stresses are aU=20.0, aL= -15.0.
b. Assume the initial design (X;)T = {I, I} and the initial move limits given by
- 0.8 < IiX < 0.8. Apply one iteration of the sequential LP and scale the
resulting design to obtain a feasible solution.
4.11 Consider the grillage of exercise 1.8, shown in Fig. 1.22. The bounds on
stresses are aU
aL 20 in beam 1, and aU aL= 10 in beam 2. The lower
bound on X is XL 10 and the modulus of elasticity is E 30,000.
=- =
=
=-
a. lllustrate graphically the feasible region and the objective function contours in
the space of Xl and Xl .
b. Solve the problem by the solution process for optimization in design planes.
Assume the initial design (XjT = {20, ISO} and choose (l1X*)T = {I, -I} [see
(4.71)]. Show the intermediate solutions in the design plane.
c. Solve part b for XL= O. Show that the optimal solution is a singular point
4.12 Solve the problem of exercise 1.3 by the procedure of (4.83) through (4.85).
Assume the lower bounds on the design variables
design (X;)T
={10, 10}.
100
Fig.
4.48
Exercises
100
100
289
~I
10.0
Fig.
4.49.
4.13 The truss shown in Fig. 4.48 is subjected to two distinct loading conditions,
PI = 20.0 and P2 = 20.0. The objective function is Z =282.8X1 + lOOX2 ' the
allowable stresses are aU= 20.0, oL=-15.0, and the modulus of elasticity is 30,000.
the two loading conditions). Find the optimal solution and compare the result
with that of part a.
c. Find the control forces in member 2, Nel and NeZ' required to maintain
compatibility at the optimum of part h.
5.0
5.0
b
Fig.
4.50.
200
JOI
!
JE
600
5.0
200
:::h....
-1. ~
100
loE
290
4 Design Procedures
4.14 Consider the truss shown in Fig. 4.49 with two design variables: the crosssectional area X and the angle Y. The upper bound on the stresses is aU =20 and
the bounds on Yare Yu = 6()0, yL = 300. The objective function is Z =X. Solve
the nonlinear programming problem by sequential LP [see (4.147)] without move
limits. Choose the initial design X(l) = 0.4. y(1) = 30 0
4.15 Assuming a two-level formulation, solve the problem of exercise 4.14,
where X is the frrst-level variable and Y is the second-level variable. Choose the
initial design yel) = 300.
4.16 The symmetric continuous beam shown in Fig. 4.50 is subjected to a single
loading condition of three concentrated loads. The two design variables are the
cross-sectional area X and the distance between the supports Y. Assume a uniform
cross section with the modulus of section W = )(2/6. The bounds on stresses are
aU = -aL = 20.0 and the bounds on Y are yL = 120, Yu = 280. The objective
function represents the volume of material. Formulate the two-level problem,
where X is the first-level variable and Y is the second-level variable, and find the
optimal solution.
4.17 The symmetric truss shown in Fig. 4.51 is subjected to two concentrated
loads, PI = 10.0 and P 2 = 10.0. The bounds on stresses are aU = -aL = 20.0, and
the lower bound on cross-sectional areas is XL = O. The objective function
represents the volume of material. Formulate the LP problem (4.151) and find the
optimal topologies for Y = 50, Y = 100 and Y = 150 and for each of the following
loading cases:
l",
.
J~I20a
li
Fig.
4.51.
Exercises
4 x 100
Fig.
291
4.52.
4.18 The symmetric twenty-one-bar truss shown in Fig. 4.52 is subjected to three
concentrated loads P =10, acting simultaneously. The bounds on stresses are aU =
_aL = 20.0, and the objective function represents the volume of the truss.
Formulate and solve the LP problem (4.151) for Y = 100, Y = 150 and Y =200.
Show all optimal topologies and find the optimal value of Y by quadratic
interpolation.
4.19 The five-bar truss shown in Fig. 4.53 is subjected to two loads, PI and P 2 ,
acting simultaneously. The bounds on stresses are aU = _aL = 20.0, and the
objective function represents the volume of material. Formulate and solve the LP
problem (4.151) for:
a. PI = 10.0, P 2 = 5.0;
b. PI = 5.0, P2 = 10.0;
c. PI =P2 = 10.0.
Show the optimal topologies and check whether the resulting structures are mobile
(mechanisms).
Fig.
4.53.
References
1. Abu Kassim, A. M., and Topping, B. H. V.: Static reanalysis: a review. J. Struct.
Engrg., ASCE 113, 1029-1045 (1987)
2. Arora, J.S.: Introduction to optimum design. Singapore: McGraw-Hill Book
Company 1989
3. Arora, J.S., and Haug, E.J.: Methods of design sensitivity analysis in structural
optimization. J. AIAA 17, 970-974 (1979)
4. Barthelemy, J-F.M.: Engineering design applications of multilevel optimization
methods. In: Brebbia C.A., and Hernandez. S. (eds.): Computer aided optimum
design of structures. Berlin: Springer-Verlag 1989
5. Barthelemy, J-P.M., and Haftka, R. T.: Recent advances in approximation concepts
for optimum structural design. In: Proceedings of NATO/DFG ASI on Optimization
of large structural systems. Berchtesgaden, Germany, September 1991
6. Barthelemy, J-F.M., and Riley, M.F.: An improved multilevel optimization
approach for the design of complex engineering systems. J. AIAA 26, 353-360
(1988)
7. Bendsoe, M. P., and Mota Soares, C. A. (eds.): Proceedings of NATO ARW on
Topology design of structures. Sesimbra, Portugal June 1992
8. Bendsoe, M. P., and Kikuchi, N.: Generating optimal topologies in structural
design using the homogenization method. Computer Methods in Appl. Mech. and
Engrg. 71, 197-224 (1988)
9. Berke, L.: An efficient approach to the minimum weight design of deflection
limited structures. USAF-AFFDL-TR-70-4-FDTR, May 1970
10. Berke, L.: Convergence behavior of optimality criteria based iterative procedures.
USAF AFFDL-TM-72-1-FBR, January 1972
11. Berke, L., and Khot, N.S.: Use of optimality criteria methods for large scale
systems, AGARD-SL-70, 1974
12. Berke, L., and Khot, N.S.: Performance characteristics of optimality criteria
methods. In: Rozvany, G.I.N. and Karihaloo, B.L. (eds.): Structural optimization.
Dordrecht: Kluwer, 1988
13. Bertsekas, D.P.: Multiplier methods: a survey. Automatica 12, 133-145 (1976)
14. Carmichael, D.G.: Structural modelling and optimization. Chichester, UK: Ellis
Horwood 1981
15. Carroll, C.W.: The created response surface technique for optimizing nonlinear
restrained systems. Operations Res. 9, 169-184 (1961)
16. Cassis, I.H., and Schmit, L.A.: On implementation of the extended interior penalty
function. Int. J. Num. Meth. Engrg. 10, 3-23 (1976)
17. Cheney, E.W., and Goldstein, A.A.: Newton's method for convex programming on
Tchebycheff approximation. Num. Math I, 253-268 (1959)
18. Cohn. M.Z., Ghosh, S.K., and Parimi, S.R.: Unified approach to the theory of
plastic structures. J. Eng. Mech. Div., ASCE 98, 1133-1158 (1972)
294
References
19. Courant, R.: Variational methods for the solution of problems of equilibrium and
vibrations. Bull. Am. Math. Soc. 49, 1-23 (1943)
20. Dantzig, G.: Linear programming and extensions. Princeton, N.J.: Princeton
University Press 1963
21. Davidon, W.C.: Variable metric method for minimization. Argonne Nat. Lab. ANL5990, Rev., 1959
22. Dorn, W.S., Gomory, R.E., and Greenberg, H.J.: Automatic design of optimal
structures. Journal de Mecanique 3, 25-52 (1964)
23. Eschenauer, H., Koski, J., and Osyczka, A.: Multicriteria design optimization procedures and applications. Berlin: Springer-Verlag 1990
24. Farshi, B. and Schmit, L.A.: Minimum weight design of stress limited trusses. J.
Struct. Div., ASCE 100, 97-107 (1974)
25. Fiacco, A.V., and McCormick, G.P.: Programming under nonlinear constraints by
unconstrained minimization: a primal-dual method. Research Analysis
Corporation, Bethesda, Md., Tech. Paper RAC-TP-96, 1963
26. Fiacco, A.V., and McCormick, G.P.: Nonlinear programming, sequential
unconstrained minimization techniques. New York: John Wiley & Sons 1968
27. Fletcher, R.: An ideal penalty function for constrained optimization. J. Inst.
Maths. Appl. 15, 319-342 (1975)
28. Fletcher, R., and Powell, M.J.D.: A rapidly convergent descent method for
minimization. Comput. J. 6, 163 (1963)
29. Fletcher, R., and Reeves, C.M.: Function minimization by conjugate gradients.
Comput. J. 7, 149 (1964)
30. Fleury, C. and Braibant, V.: Structural optimization: a new dual method using mixed
variables. Int. J. Num. Meth. Engrg. 23, 409-428 (1986)
31. Fleury, C. and Geradin, M.: Optimality criteria and mathematical programming in
structural weight design. Computers and Structures 8, 7-18 (1978)
32. Fleury, C. and Sander, G.: Structural optimization by finite element. University of
Liege, Belgium, LTAS Report NBR-SA-58, January 1978
33. Fox, R.L.: Optimization methods for engineering design. Reading, Mass.:
Addison-Wesley 1971
34. Fox, R.L. and Miura, H.: An approximate analysis technique for design
calculations. J. AIAA 9, 177-179 (1971)
35. Fox R.L. and Schmit L.A.: Advances in the integrated approach to structural
synthesis. 1. Spacecraft and Rockets 3, (1966)
36. Fuchs, M.B: Linearized homogeneous constraints in structural design. Int. J. Mech.
Sci. 22, 33-40 (1980)
37. Fuchs, M.B.: Explicit optimum design. Int. 1. Solids Structures 18, 13-22 (1982)
38. Gallagher, R.H.: Fully stressed design. In: R.H. Gallagher and O.C. Zienkiewicz
(eds.): Optimum structural design. New York: John Wiley & Sons 1973
39. Gass, S.I.: Linear programming. New York: McGraw-Hill Book Company 1958
40. Gellatly, R.A. and Berke, L.: Optimal structural design. USAF AFFDL-TR-70-165,
April 1971
41. Gerard, G.: Minimum weight analysis of compressive structures. New York: New
York Univ. Press 1956
42. Gill, P.E., Murray, W., and Wright, M.H.: Practical optimization. New York:
Academic Press 1981
43. Griffith, R.E., and Stewart, R.A.: A nonlinear programming technique for the
optimization of continuous processing systems. Manage. Sci. 7, 379-392 (1961)
44. Gurdal, Z., and Haftka, R. T.: Optimization of composite laminates. In:
Proceedings of NATO/DFG ASI on Optimization of large structural systems.
Berchtesgaden, Germany, September 1991
45. Hadley, G.: Linear programming. Reading, Mass.: Addison-Wesley 1962
References
295
46. Haftka, R.T.: An improved computational approach for multilevel optimum design.
1. Struct. Mech. 12, 245-261 (1984)
47. HaftIca, R.T., and Adelman, H.M.: Recent developments in structural sensitivity
analysis. Structural Optimization I, 137-151 (1989)
48. HaftIca, R.T., Gurdal, Z., and Kamat, M.P.: Elements of structural optimization.
Dordrecht: Second Edition, Kluwer Academic Publishers 1990
49. HaftIca, R.T., and Starnes, 1.H.: Application of a quadratic extended interior penalty
function for structural optimization. 1. AIM 14, 718-724 (1976)
50. Hajali, R.M., and Fuchs, M.B.: Generalized approximations of homogeneous
constraints in optimal structural design. In: Brebbia C.A., and Hernandez, S. (eds.):
Computer aided optimum design of structures. Berlin: Springer-Verlag 1989
51. Haug, E.1.: A review of distributed parameter structural optimization literature. In:
Haug, E.1., and Cea, 1. (eds.): Optimization of distributed parameter structures.
Proceedings of NATO ASI. Iowa City: Sijthoff and Noordhoff, Alphen ann der Rijn
1981
52. Hemp, W.S.: Optimum structures. Oxford, UK.: Clarendon Press 1973
53. Hestenes, M.R.: Multiplier and gradient methods. 1. Optimization Theory and
Applications 4, 303-320 (1969)
54. Heyman, 1.: Plastic design of beams and frames for minimum material
consumption. Quarterly of Appl. Math. 8, 373-381 (1951)
55. Hodge, P.G. Ir.: Plastic analysis of structures. New York: McGraw-Hill Book
Company 1959
56. Kavlie, D.: Optimum design of statically indeterminate structures. University of
California, Berkeley, Ph.D. Thesis, 1971
57. Kelley, H.I.: The cutting plane method for solving complex programs. SIAM 1. 8,
703-712 (1960)
58. Khot, N.S.: Algorithms based on optimality criteria to design minimum weight
structures. Engineering Optimization 5, 73-90 (1981)
59. Khot, N.S., Berke, L., and Venkayya, V.B.: Comparison of optimality criteria
algorithms for minimum weight design of structures. 1. AIM 17, 182-190 (1979)
60. Kirsch, U.: Optimized prestressing by linear programming. Int. 1. Num. Meth.
Engrg. 7, 125-136 (1973)
61. Kirsch, U.: Optimal design of trusses by approximate compatibility. Computers
and Structures 12, 93-98 (1980)
62. Kirsch, U.: Optimum structural design. New York: McGraw-Hill Book Company
1981
63. Kirsch, U.: Approximate structural reanalysis based on series expansion. Compo
Meth. Appl. Mech. Eng. 26, 205-223 (1981)
64. Kirsch, U.: Optimal design based on approximate scaling. 1. Struct. Div., ASCE
108, 888-910 (1982)
65. Kirsch, U.: Approximate structural reanalysis for optimization along a line. Int. 1.
Num. Meth. Engrg. 18, 635-651 (1982)
66. Kirsch, U.: Synthesis of structural geometry using approximation concepts.
Computers and Structures 15, 305-314 (1982)
67. Kirsch, U.: Approximate behavior models for optimum structural design. In: Atrek,
E., et al (eds.): New directions in optimum structural design. New York: Iohn Wiley
& Sons 1984
68. Kirsch, U.: A bounding procedure for synthesis of prestressed systems. Computers
and Structures 20, 885-895 (1985)
69. Kirsch, U.: An improved multilevel structural synthesis. lour. of Structural
Mechanics 13, 123-144 (1985)
70. Kirsch, U.: The effect of compatibility conditions on optimal design of flexural
structures. Int. 1. Num. Meth. Engrg. 24, 1173-1185 (1987)
296
References
71. Kirsch, U.: Optimal topologies of flexural systems. Engineering Optimization 11,
141-149 (1987)
72. Kirsch, U.: Optimal topologies of truss structures. Computer Methods in Appl.
Mech. and Engrg. 72, 15-28 (1989)
73. Kirsch U.: Optimal topologies of structures. Appl. Mech. Rev. 42,223-239 (1989)
74. Kirsch, U.: Effect of compatibility and prestressing on optimized structures. J.
Struct. Eng., ASCE 115, 724-737 (1989)
75. Kirsch, U.: Feasibility and optimality in structural design. Computers and
Structures 41, 1349-1356 (1991)
76. Kirsch, U.: Improved optimum structural design by passive control. Engineering
with Computers 5, 13-22 (1989)
77. Kirsch, U.: On singUlar topologies in optimum structural design. Structural
Optimization 2, 133-142 (1990)
78. Kirsch, U.: On the relationship between optimum structural geometries and
topologies. Structural Optimization 2, 39-45 (1990)
79. Kirsch, U.: Optimal design of structural control systems. Engineering Optimization
17, 141-155 (1991)
80. Kirsch, U.: Reduced basis approximations of structural displacements for optimal
design. J. AIAA 29, 1751-1758 (1991)
81. Kirsch, U., and Moses, F.: Decomposition in optimum structural design. J. Struct.
Div., ASCE 105, 85-100 (1979)
82. Kirsch, U., and Taye, S.: On optimal topology of grillage structures. Engineering
with Computers 1, 229-243 (1986)
83. Kirsch, U., and Taye, S.: High quality approximations of forces for optimum
structural design. Computers and Structures 30, 519-527 (1988)
84. Kirsch, U., and Taye, S.: Structural optimization in design planes. Computers and
Structures 31, 913-920 (1989)
85. Kirsch, U., and Toledano, G.: Approximate reanalysis for modifications of
structural geometry. Computers and Structures 16, 269-279 (1983)
86. Kirsch, U., and Topping, B.H.V.: Minimum weight design of structural topologies.
J. of Computing in Civil Engineering, ASCE 118, 1770-1785 (1992)
87. Kresselmeir, G., and Steinhauser, R.: Systematic control design by optimizing a
vector performance index. In: Proceedings of IFAC Symposium on Computer Aided
Design of Control Systems. Zurich, Switzerland: 1979, pp. 113-117
88. Kuhn, H.W., and Tucker, A.W.: Nonlinear programming. In: Proc. 2d Berkeley
Symp. Math. Stat. Probab., Berkeley, Calif., 1950, pp. 481-492
89. Lasdon, L.S.: Optimization theory for large systems. London: The Macmillan
Company 1970
90. Lellep, J., and Lepik, U.: Analytical methods in plastic structural design.
Engineering Optimization 7, 209-239 (1984)
91. Leunberger, D.G.: Introduction to linear and nonlinear programming. Reading,
Mass: Addison-Wesley 1984
92. Lev, O.E. (ed.): Structural optimization: recent developments and applications.
New-York: ASCE 1981
93. Levy, R., and Lev, 0.: Recent developments in structural optimization. J. Struct.
Eng. ASCE 113, 1939-1962 (1987)
94. Massonnet, C.E., and Save, M.A.: Plastic analysis and design, vol. 1. Blaisdell
Publishing Company 1965
95. Maxwell, C.: Scientific papers ll. Cambridge University Press, 175 (1890)
96. Mesarovic, M.D., Macko, D., and Takahara, Y.: Two coordination principles and
their application in large scale systems control. In: Proc. IFAC Congress, Warsaw,
Poland, 1969
References
297
97. Mitchell, A.G.M.: The limits of economy of material in framed structures. Phil.
Mag., series 6, 8, 589-597 (1904)
98. Miura, H., and Schmit, L.A.: Second order approximations of natural frequency
constraints in structural synthesis. Int. J. Num. Meth. Engrg. 13, 337-351
(1978)
99. Moe J., and Gisvold K.M. (eds.): Optimization and automated design of
structures. Div. of Ship Structures, Norwegian Institute of Technology,
Trondheim, Rep. SK/M21, December 1971
100. Niordson, FJ., and Pedersen, P.: A review of optimal structural design. In:
Proceedings, 13th Int. Congress on Theoretical and Applied Mechanics. Berlin:
Springer-Verlag 1973
101. Noor, A. K., and Lowder, H. E.: Approximate techniques for structural reanalysis.
Computers and Structures 4, 801-812 (1974)
102. Noor, A. K., and Whitworth, S.A.: Reanalysis procedure for large structural
systems. Int. J. Num. Meth. Engrg. 26, 1729-1748 (1988)
103. Olhoff, N., and Taylor, J.E.: On structural optimization. J. Appl. Mech. 50,
1139-1151 (1983)
104. Pedersen P.: On the optimal layout of multi-purpose trusses. Computers and
Structures 2, 695-712 (1972)
105. Pickett, R.M., Rubinstein, M.F., and Nelson, R.B.: Automated structural
synthesis using a reduced number of design coordinates. J. AIAA II, 489-494
(1973)
106. Pope, G.G., and Schmit L.A. (eds.): Structural design applications of
mathematical programming. AGARDograph No. 149, AGARD-NATO
Publication, February 1972
107. Powell, M.J.D.: An efficient method of finding the minimum of a function of
several variables without calculating derivatives. Compo J. 7, 155 (1964)
108. Prager, W., and Shield, R.T.: Optimal design of multi-purpose structures. Int. J.
Solids Structures 4, 469-475 (1968)
109. Prager, W., and Taylor, J.E.: Problems of optimal structural design. Journal of
Applied Mechanics 35, 102-106 (1968)
110. Reinschmidt, K.F.: Review of structural optimization techniques. In: Proceedings
of The ASCE Annual Convention. Philadelphia: 1975
111. Reinschmidt, K.F., and Russell, A.D.: Applications of linear programming in
structural layout and optimization. Computers and Structures 4, 855-869 (1974)
112. Rosen, J.B.: The gradient projection method for nonlinear programming. J. Soc.
Industrial and Appl. Math. 9, 514-532 (1961)
113. Rozvany, GJ.N.: Structural design via optimality criteria. Dordrecht: Kluwer
1989
114. Rozvany, GJ.N., and Mroz, Z.: Analytical methods in structural optimization.
Appl. Mech. Rev. 30, 1461-1470 (1977)
115. Rozvany, GJ.N., and Ong, T.C.: Update to analytical methods in structural
optimization. App. Mech. Rev. Update. 289-302 (1986)
116. Russell, A.D., and Reinschmidt, K.F.: Discussion of optimum design of trusses
for ultimate loads. J. Struct. Div., ASCE, 97, 2437-2442 (1971)
117. Schmit, L.A.: Structural design by systematic synthesis. In: Proceedings of the
2nd Conference on Electronic Computation. New York: American Society of
Civil Engineering 1960, pp. 105-122
118. Schmit, L.A.: Structural synthesis 1959 - 69: A Decade of progress. In: Japan U.S. seminar on matrix methods of structural analysis and design. Tokyo: 1969
119. Schmit, L.A.: Structural synthesis - its genesis and development. J. AIAA 19,
1249-1263 (1981)
298
References
120. Schmit, L.A.: Structural optimization - some key ideas and insights. In: Atrek,
E., et al (eds.): New directions in optimum structural design. New York: Iohn
Wiley & Sons 1984
121. Schmit, L.A., and Farshi, B.: Some approximation concepts for structural
synthesis. I. AIAA 12, 692-699 (1974)
122. Schmit, L.A., and Fleury, C.: Discrete-continuous variable structural synthesis
using dual methods. I. AIAA 18, 1515-1524 (1980)
123. Schmit L.A., and Mallett R.H.: Structural synthesis and design parameter
hierarchy. I.StrucL Div., ASCE, 89, August 1963
124. Schoeffler, I.D.: Static multilevel systems. In: D.A. Wismer (ed.): Optimization
methods for large scale systems. New York: McGraw-Hill Book Company 1971
125. Shanley, F.R.: Weight strength analysis of aircraft structures. New York:
McGraw-Hill Book Company 1952
126. Sheu, C.Y., and Prager, W.: Recent developments in optimal structural design.
Appl. Mech. Rev. 21, 985-922 (1968)
127. Sheu, C.Y., and Schmit, L.A.: Minimum weight design of elastic redundant
trusses under multiple static loading conditions. 1. AIAA 10, 155-162 (1972)
128. Sobiszczanski-Sobieski I., lames, B.B., and Dovi, A.R.: Structural optimization
by multilevel decomposition. I. AIAA 23, 1775-1780 (1985)
129. Sobioszczanski-Sobieski, I., lames, B.B., and Riley, M.F.: Structural sizing by
generalized multilevel optimization. 1. AIAA 25, 139-145 (1987)
130. Sobiszczanski-Sobieski: Optimization decomposition - a step from hierarchic to
non-hierarchic systems. 2nd NASA Air Force symposium on recent advances in
multidisciplinary analysis and optimization. NASA CP-3031, part 1 (1988)
131. Somekh, E., and Kirsch, U.: Interactive optimal design of truss structures. CAD
lour. 13, 253-259 (1981)
132. Starnes, I.H. Ir., and Haftka, R.T.: Preliminary design of composite wings for
buckling stress and displacement constraints. 1. Aircraft 16, 564-570 (1979)
133. Stewart m, G.W.: A modification of Davidon's minimization method to accept
difference approximations to derivatives. I. ACM 14, 72 (1967)
134. Svanberg, K.: The method of moving asymptotes - a new method for structural
optimization. Int. I. Num. Meth. Engrg. 24, 359-373 (1987)
135. Sved, G., and Ginos, L.: Structural optimization under mUltiple loading. Int. I.
Mech. Sci. 10, 803-805 (1968)
136. Topping, B.H.V.: Shape optimization of skeletal structures: A review. I. Struct.
Eng., ASCE 109, 1933-1951 (1983)
137. Vanderplaats, G.N.: Structural optimization - past, present and future. I. AIAA
20, 992-1000 (1982)
138. Vanderplaats, G.N.: Numerical optimization techniques for engineering design.
New York: McGraw-Hill Book Company 1984
139. Vanderplaats, G. N., and Salagegheh, E.: A new approximation method for stress
constraints in structural synthesis. I. AIAA 27, 352-358 (1989)
140. Venkayya, V.B.: Structural optimization: A review and some recommendations.
Int. I. Num. Meth. Engrg. 13, 205-228 (1978)
141. Venkayya, V.B., Khot, N.S., and Reddy, V.S.: Optimization of structures based
on the study of strain energy distribution, AFFDL-TR-68-150, 1968
142. Wasiutynsky, A., and Brandt, A.: The present state of knowledge in the field of
optimum design of structures. Appl. Mech. Rev. 16, 341-350 (1963)
143. Weaver, W. Ir., and Gere I.M.: Analysis of framed structures, second edition. New
York: Van Nostrand Reinhold 1980
144. Zangwill, W.I.: Nonlinear programming via penalty functions. Manage. Sci. 13,
344-358 (1967)
145. Zoutendijk, G.: Methods of feasible directions. Amsterdam: Elsevier 1960
Subject Index
300
Index
Continuous problems, 4, 8
Control of structures, 246-253, 211
active, 246
forces, 246
gains for displacements, 246
optimal, 246-249
passive, 246
Convergence criteria, 103
Convex cone, 62
Convex function, 65-66
Convex linearization, 144-145
Convex programming, 66, 111
Convex set, 65-66
Coordination, 225-226
Cost function (see Objective function)
Cubic fitting, 10-12, 161
Cubic function, 10
Cutting-plane method, 140
Decomposed matrix, 141, 111
Decomposition, 225
process oriented, 225, 234
system oriented, 225, 231
Derivatives of constraints, (see
Sensitivity analysis)
Design:
constrained, 29
deterministic, 3
probabilistic, 3
unconstrained, 29
Design approaches, 3
Design line, 199, 201
Design plane, 206-208
Design procedures, 119-289
Design process, 1
Design space, 28, 31
Design variables, 25-21
configurational (see geometrical)
continuous, 25
cross-sectional, 25, 21
discrete, 25
geometrical, 25, 26
integer, 26
material, 25, 26
pseudodiscrete, 26
topological, 25, 26
Detailing stage, 1
Dimensions, 12n
Direct search methods, 13
Direct update methods, 19-80
Direction vector, 110-111
Discrete parameter optimization, 4, 8, 9
Distributed parameter optimization, 4, 8
Index
Integrated problem formulation (see
Simultaneous analysis and design)
Interactive optimization, 278
Interior methods, 197
Intermediate response, 169
Intermediate variables, 133-139, 160161
Interpolation functions, 138-139
Intersection, 29
Interval of uncertainty, 67
Jacobi iteration, 149
Kresselmier-Steinhouser function, 198
Kuhn-Tucker conditions, 62-63, 210
Lagrange multipliers, 59-60, 107, 116,
217-218, 223
Lagrangian function, 60, 217
Limit analysis, 180
Limit design, 20
Linear programming, 4, 22, 48, 80-97,
278
directions fmding by, 111-112
elastic design by, 191-198
frame design by, 185-186
geometrical interpretation of, 184
optimal topology by, 263-267
optimality condition of, 86
phase I of, 89
phase II of, 89
plastic design by, 180-191
prestressed concrete optimization by,
236-237
sensitivity analysis of, 97
sequence of, 98, 139-142, 255
standard form of, 82
truss design by, 180-181
Linking of design variables, 36
Load:
overload, 3, 8, 9
service load, 3, 8, 9
factor, 21
Local minimum (see Relative
minimum)
Lower bound of the optimum, 197,251
Lower bound theorem of limit analysis,
180
Marginal price, 63
Mathematical programming, 4, 7, 57,
223
Mechanism of collapse, 20
Minimization along a line, 66-72
301
302
Index
Spri nger-Verlag
and the Environment
We
We
ness partners - paper mills, printers, packaging manufacturers, etc. - to commit themselves
to using environmentally friendly materials and
production processes.
The