Você está na página 1de 17

This article was downloaded by: [University of Cambridge]

On: 09 October 2014, At: 18:00


Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Engineering Optimization
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/geno20

INTERACTIVE DESIGN OPTIMIZATION


a a
J. S. ARORA & C. H. TSENG
a
Optimal Design Laboratory, Department of Civil and Environmental Engineering , The
University of Iowa , Iowa City, IA, 52242, USA
Published online: 27 Apr 2007.

To cite this article: J. S. ARORA & C. H. TSENG (1988) INTERACTIVE DESIGN OPTIMIZATION, Engineering Optimization, 13:3,
173-188, DOI: 10.1080/03052158808940953

To link to this article: http://dx.doi.org/10.1080/03052158808940953

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
Eng. Opr., 1988, Vol. 13, pp. 173-188 0 1988 Gordon and Breach Science Publishers, Inc.
Reprints available directly from the publisher Printed in Great Britain
Photocopying permitted by license only

INTERACTIVE DESIGN OPTIMIZATION

J. S. ARORA and C. H. TSENG


Optimal Design Laboratory,
Department of Civil and Environmental Engitreering,
The University of Iowa,
Iowa City, IA 52242 USA
Downloaded by [University of Cambridge] at 18:00 09 October 2014

(Receiued 27 May 1987)

Most optimization algorithms have uncertainty in their computational steps. Therefore, it is prudent to
interactively monitor their progress and change the course of computations if problems are encountered in
the expected performance. Proper interactive algorithms need to be developed and implemented into user-
friendly software. The present paper describes general algorithms suitable for interactive design optimiza-
tion. Several queries that can aid the design optimization process are described and algorithms to
implement them are given. Using these queries the designer can actually guide the design process toward
improved designs and, finally, the best design. Several other capabilities, such as graphics that can aid the
interactive design decision making process are also discussed and demonstrated with example problems.

KEYWORDS: CAD, design optimization, structures, interactive computing,

1 INTRODUCTION

During the last twenty years considerable numerical work has been done to show that
nonlinear programming methods can be used to optimize engineering systems.
However, many optimization techniques are inefficient because they need many
function and gradient evaluations. Each call for function evaluation requires solution
of a large complex analysis model. T o overcome the computational expense, approxi-
mations are introduced in optimization algorithms when they are applied to engineer-
ing design problems. With these approximations, the algorithms lose their robustness.
In addition, most optimization algorithms have uncertainty in their computational
steps. If they are not properly safe-guarded, the algorithm can give erratic behavior.
This can lead to frustrations with optimization methods for applications in engineer-
ing design.
In this regard, one area that has received very little attention is the development
of general algorithms suitable for use in interactive environments. It is possible
that some algorithms are more suitable for interactive use than others. Also the extent
of user interaction needs to be investigated. Use of interactive graphics can play a
major role. With proper display of information (graphic and numeric) the designer
can control the design process by changing the design data and other parameters at
each iteration. The purpose of this paper is to study the role of the designer in the
interactive process. Various capabilities needed to properly guide the optimum design
process will be discussed. Algorithms suitable for interactive processing will be
described and demonstrated.
A considerable body of literature on mathematical programming and structural
optimization methods is available. A comprehensive review of this literature is beyond
the scope of this paper. However, recent articles5.' and the literature cited in them
would reasonably represent the state-of-art in engineering design optimization.
174 J. S. ARORA AND C. H. TSENG

2 DESIRED INTERACTIVE CAPABILITY

Design Optimization Problem


Many engineering design optimization problems can be described by the following
general nonlinear programming, Problem PS.9:
Find design variable vector x E Rn that solves the problem
minimize f (x) for x E S (1)
where S is a constraint subset of R" given as

Here m is the total number of constraints, p is the number of equalities, ( m - p) is


Downloaded by [University of Cambridge] at 18:00 09 October 2014

the number of inequalities, x, and xu are the lower and upper bounds o n the
variables.
It is important to realize that many constraints in Eq. ( 2 ) may be implicit functions of
the design variables. These include stress, displacement, natural frequency, buckling
load and other such constraints.

Whur is Inreracriue Design Oprimizarion?


Interactive design optimization is a computer-aided environment where the designer
has the opportunity to interact with the process and make design decisions. A key

General Purpose versus Domain Specific


Interaction

L-i Data Entry

I Optimization
Aloorithms
Design I Design Problem
Definition

Interactive
Graphics

General Purpose
Design Optimization
Software
, ,

Figure I Conceptual layout of interactive design optimization envirpnment.


INTERACTIVE DESIGN OPTIMIZATION 175

component of such a capability is software that has several interactive optimization


algorithms implemented in a robust manner. It has commands available for design
queries and decision making. Two levels of interaction are possible as shown in the
conceptual layout ofthe interactive design optimization environment of Figure I. The
first level is the general purpose interaction shown on the left side. This type of
interaction is not tied to any application and uses only general optimization
terminology. The domain specific interaction is shown on the right side of the figure.
This type of interaction uses the domain-specific terminology and commands, and
must be developed for each application area.'.12 The present paper concentrates on
the general purpose interactive design optimization capability.

Case jiir lnreractive Design Oprimizarion


Downloaded by [University of Cambridge] at 18:00 09 October 2014

Design of practical engineering systems is a fairly complex process. Several software


components, such as model generation, analysis, constraint definition, design sensitiv-
ity analysis, optimization and post-processor, must be integrated to create a design
capability. Most of the components are quite large and it is not a trivial matter to
integrate them. Errors can occur in many places. In addition there are many
uncertainties in the problem formulation-loads for the system must be estimated,
material properties can vary within certain ranges, modeling of the system for analysis
purpose can introduce errors, and the analysis of the system is usually only
approximate. Thus, the problem cannot be stated in a precise form for complete
analysis and there are uncertainties in the design data. A solution to the problem need
not necessarily exist. Many times, the formulation of the problem must be developed
as a part of the design process. Therefore, it is neither desirable nor useful to optimize
an inexact problem to its conclusion in a batch environment. It would be a complete
wastage of valuable human, as well as computer, resources to finally discover that
wrong data were used or a constraint was inadvertantly omitted. In addition, most
optimization algorithms have many uncertainties in their computational steps and
their implementation is an art. All these difficulties indicate that it is prudent to
monitor design optimization processes interactively and change the course of the
calculations or abort them if the progess is not as expected.
Most nonlinear programming methods require the evaluation of problem functions
several times before an optimum design is reached. For practical problems such
evaluation usually needs enormous computational effort due to the implicit nature of
the functions. It may be useful to interactively guide the optimization process,
especially in the initial phase. This way the design can be improved for a few iterations
and then the process can be turned over to a more robust algorithm to obtain the
precise optimum.
There are several other advantages in having algorithms and software capable of
interaction and design decision making:
I ) Designer experience and intuition can be interjected into the design process.
2) Practical designs can be assured.
3) Insights for a particular design problem can be gained leading to better design
concepts.

Interactive software can be menu-driven, command-driven or both, and should have


help facilities. It should be able to treat general nonlinear programming as well as
176 J. S. ARORA AND C. H. TSENG

unconstrained problems. It should be able to treat equality, inequality and design


variable bound constraints. I t should have a choice of a few good algorithms that are
robustly implemented. It should trap user's mistakes and not abort abnormally.

Ir~rerucrivednru prepururion The software should have a module for interactive data
preparation and editing. The commands for data entry should be explicit and simple
so that even a beginner can easily follow it. The user should be able to change any data
that has been previously entered. In addition, minimum input data should be
required.

It~rercrctioe decision mukiny All general purpose design optimization software re-
quires the following information about the problem to be solved: i) input data, such as
number of design variables, number of constraints, etc., ii) cost and constraint
Downloaded by [University of Cambridge] at 18:00 09 October 2014

functions, and iii) gradients of cost and constraint functions. It should be possible to
provide this information easily for various applications. I t should be also possible to
utilize the interactive dcsign optimization software in a batch environment. The
system should have default values for parameters which are the best ones determined
through expertise and numerical experimentations.
The designer should be able to guide the problem solving process. For example, the
program can be run for a certain number of iterations and interrupted to see if the
process is progressing satisfactorily. Then a decision to change o r continue the course
of the calculations can be made. For example, the following capabilities are desirable:
I) If there are constraint violations, the designer may want to know whether they
can be corrected without any penalty on the cost function. If this cannot be done, the
penalty on the cost function to correct the constraints should be made available.
2) When the design is in the feasible region, the system should have the capability
to perform calculations and determine if the cost function can be reduced by a certain
percentage and still remain feasible.
3) If the iterative process does not progress well, then the designer should be able to
restart the program from any previous iteration o r any other design.
4) At the optimum point, the penalty to tighten a constraint o r the gain to relax it
should be displayed. This information is available from the Lagrangian multipliers for
the constraints.
5) It should be possible to change the design data (such as design variables and
their limits, convergence criteria, etc.) for a design problem during the iterative
process. After monitoring the process for a few iterations it may be necessary to
change the problem o r program parameters. This should be possible without
terminating the program.
6) I t should be possible to assign fixed values to design variables; i.e., it should be
possible to change the status of a design variable interactively.
7) The designer may want to re-examine the problem formulation or design data.
Therefore, it should be possible to exit from the program at any point.
8) I t should be possible to display the status of the design at any iteration, such as
current values of the design variables, cost function, maximum constraint violation
and other such data.
9) The designer should be able to run the algorithm one iteration at a time o r for
several iterations.
INTERACTIVE DESIGN OPTIMIZATION

5700. -
5 4 0 0 . ..
5 1 0 0 . .~
4800. ..
4 5 0 0 . ..
4 2 0 0 . ..
3 9 0 0 . ..
3 6 0 0 . ..
3 3 0 0 . ..
3 0 0 0 . ..
2 7 0 0 . ..
2400. ..
2100. ..
1800. ..
1 5 0 0 . ..
Downloaded by [University of Cambridge] at 18:00 09 October 2014

1200. ..
300. ..
600. 0. //
2,
1. 3. 4. 5. 6. '.
ITERATION8. NO.9. 10. 12. 13. 14.

I
Figure 2 Cost function history lor ten bar structure.

10) It should be possible to change the algorithm during the iterative process.
I I) The designer should be able to request constraint correction with an x % limit
on the increase in cost.
12) The designer should be able to request various graphical displays.

Interactive graphics Graphical display of data is a convenien~way to interpret


results and draw conclusions. Interactive graphics can play a major role in design
decision making during the iterative optimization process. Possible graphical displays
at each iteration are:
1) Cost function, convergence parameter and maximum constraint violation
histories. As an example, Figure 2 shows the cost function history for a ten member
structure optimization problem discussed later in the paper. It can be seen that after
the 8th iteration the cost function does not change very much. Large changes occur
during the initial iterations because the starting point is highly inreasible.
2) Histories of design variables. These can be used to observe trends in design
variables and possibly extrapolate their values. These plots are quite similar to the one
for the cost function shown in Figure 2.
3) Histories of constraint functions. This can show constraints that are not playing
any role in the design process. It can also show dominant constraints. These plots are
also quite similar to the one for the cost function shown in Figure 2.
4) Sensitivity coefficients for the cost and constraint functions. These can be
displayed in the form of bar charts. These are nothing but normalized gradients of cost
and constraint functions. They show sensitive or insensitive variables and functions.
Figure 3 shows the relative sensitivity of constraint number 3 (stress in member 2)
to all the design variables for the ten member structure. The sensitivity chart shows
that the stress in member 2 can be reduced by decreasing the size of members 2,5 and
10 or by increasing the size of the remaining members.
J. S. A R O R A A N D C. H. TSENG

1.0 - -
R
E 0.8 ..
L
A 0.6 ..
T
0.4 ..
0.2 ..
1 3 4 6 7 8 9
8 . 8 . n O.V. YO.
Y 2 5 10
5 -0.2..

;
I
-0.4..

-8.6..
Downloaded by [University of Cambridge] at 18:00 09 October 2014

-8.8 ..

Figure 3 Constraint sensitivity bar chart for ten member structure (stress in member 2).

Figure 4 shows the relative sensitivity of the cost function to the design variable for
the ten member structure. The figure shows that reduction in any variable reduces the
cost function. More reduction is achieved with members 7-10 as compared with
members 1-6. Since the cost [unction is the weight of the structure, Figure 4 gives
correct sensitivity information; i.e., reduction in the cross-sectional area reduces the
cost function.
Figure 5 shows the sensitivity of all the constraints to design variable number 4 for
the ten member structure. It can be seen that constraint numbers 5 and 8 (stress in
members 4 and 7) are affected the most with any change in this design variable.

I COST S E N S I T I V I T Y CHRRT

Figure 4 Cost function sensitivity bar chart Tor ten member structure
INTERACTIVE DESIGN OPTIMIZATION

I DESIGN V R R I A l L E S E N S I T I V I T Y C H A R T D.V. NO. 4


Downloaded by [University of Cambridge] at 18:00 09 October 2014

Figure 5 Sensitivity of all constraints to a design variable for ten member structure.

Note that the preceding sensitivity charts are plotted at a particular design, so they
will change from one iteration to the next. It can be seen that using the interactive
capabilities, designers can observe the progress of the optimization process. They can
learn more about the behavior of the design problem and perhaps refine its
formulation.

3 INTERACTIVE ALGORITHMS

It is clear from the preceding discussion that for a useful interactive capability proper
algorithms must be implemented into a well designed software. Some optimization
algorithms are not suitable for designer interaction. Their steps are in some sense
closed-ended allowing little opportunity for the designer to change the course of the
iterative design process. However, it turns out that the quadratic programming (QP)
subproblem^^.'^ used in them can be utilized also to devise algorithms suitable for an
interactive environment. The algorithms fall into the class of search methods where
the basic iterative equation is x ' ~ + "= xIk'+ dx; 6x = ad; k = 0, 1, 2, . ..; where k is
the iteration counter. 6x is the total change in design, d is a direction vector and a is a
step size. Usually d is found by solving a subproblem and the step size a is found by
minimizing a descent function in the direction d.$.' One dimensional search may
require several cost and constraint function evaluations, so it can be quite expensive
and should be avoided as far as possible. A slightly different concept is used in
determining the step size. It requires the solution of one of four subproblems at each
step. These subproblems can also answer some of the interactive queries posed earlier.

Subproblem definition The subproblem for calculating the search direction d can be
derived in several ways. One way is by writing linear Taylor expansions of cost and
180 J. S. ARORA AND C. H. TSENG

constraint functions. A quadratic step size constraint is then imposed to obtain a


bounded subproblem as
minimize (c, d) (3)
subject to linearized constraints

and a quadratic step size constraint 0.5(d, d) I r2 (6)


Here I, is a collection of indices of potential &-activeconstraints:
Downloaded by [University of Cambridge] at 18:00 09 October 2014

I, = { i : i = 1, p; gi(x) + E > 0, i > p, E > 0)


and r > 0 is a constant, c is the gradient of the cost function, Vgi is the gradient of a
constraint function, and the notation (A, B) = A'B has been used. For various
implicit functions, these gradients are evaluated using efficient procedures given in
Refs [2], 131, [9] and [lo].

Cosr Reduction ( C R ) Algorirhm


Since the cost reduction problem is solved from a feasible or almost feasible point, all
the constraints are satisfied. The cost reduction subproblem can be defined by adding
the quadratic step size constraint of Eq. (6) to the cost function as
minimize (c, d) + 0.5(d,d) (7)
subject to (N, d) = 0; (A, d) I b; x,, < xi + di I xi, (8)
Columns of matrices N and A contain gradients of equality and potential inequality
constraints respectively. Equations (8) show that the dot product of d with all the
columns of N is zero. Therefore, the search direction d lies in a plane tangent to the
equality constraints. The right hand side vector b for the inequality constraints in
Eq. (8) contains zero elements corresponding to the active constraints and positive
elements corresponding to the inactive constraints (i.e., bi = -9,). If an active
constraint remains satisfied at equality; i.e., (au),d) = 0 where a"' is the gradient of the
ith inequality, the direction d is in a plane tangent to that constraint. Otherwise, it
must point into the feasible region for the constraint.
The QP subproblem defined in Eqs (7) and (8) can be solved for the cost reduction
direction by any of the available subroutines. After the direction has been determined,
the step size can be calculated by a line search on the proper descent function.
Alternatively, we can require a certain reduction in the cost function and determine
the step size that way. For example, we can require a fractional reduction y in the cost
function (for five percent reduction, y = 0.05). If a is a step size along d, then a first
order change in cost is given as al(c, d)l. Equating this to the required reduction in
cost Iyf 1, the step size is calculated as a = Iyf l/l(c, d)l, with (c, d) # 0.
INTERACTIVE DESIGN OPTIMIZATION 181

Constraint Correction ( C C ) Algorithm


If at a design point, constraint violations are very large, it may be useful to find out if a
feasible design can be obtained. Several algorithms can be used to correct constraints.
We shall describe a procedure that is a minor variation of the CR algorithm. A QP
subproblem that gives constraint correction can be obtained from Eqs (7) and (8) by
neglecting the term related to the cost function. In other words we do not put any
restriction on the changes in the cost function, and define the Q P subproblem as
minimize 0.5(d, d) (9)
subject to (N, d) = e; (A, d) I b; Xi, I xi + di Ixi" (10)
with ei = - g i and bj = - g j for potential inequalities. This subproblem finds the
Downloaded by [University of Cambridge] at 18:00 09 October 2014

direction of the smallest distance to the constraint boundary (linear approximation)


from an infeasible point. Equation (9) essentially says: find a direction d having the
shortest path to the linearized feasible region from the current point. Equations (lo)
impose the requirement of constraint corrections. After the direction has been found, a
step size can be determined to make sure that the constraint violations are improved.
Note that usually constraint correction results in increase in the cost. However, there
can be some unusual cases where constraint correction can be accompanied by a
reduction in the cost function.

Algorithm for Constraint Correction at Constant ( C C C ) Cost


In some instances, the constraint violations are not very large. It is useful to know
whether a feasible design can be obtained without any increase in the cost. This shall
be called a constant cost subproblem. It turns out that by adding a very simple
additional constraint to the QP subproblem given in Eqs (9) and (10) of the previous
subsection, the constant cost subproblem can be defined. The additional constraint
simply requires the current linearized cost to either remain constant or reduce; i.e.,
(c,d) I 0. The constraint imposes the condition that the direction d be either
orthogonal to the gradient of the cost function or make an angle between 90 and 270
degrees. If the constraint is active, then there is no change in the linearized cost
function value. However, when the problem functions are nonlinear, there may be
some change in the original cost function. If the constraint is inactive, then there is
actually some reduction in the linearized cost function along with correction of the
constraints. This is a desirable situation. Thus, we observe that the constant cost
problem seeks a shortest path to the feasible region that either reduces the linearized
cost function or keeps it unchanged.
Note that the constant cost QP subproblem can be infeasible if the current cost
function contour does not intersect the feasible region. This can happen in practice, so
the QP subproblem should be properly solved. If it turns out to be infeasible, then the
constraint (c d) 5 0 must be relaxed, and the cost function must be allowed to
increase to obtain a feasible point.

Constraint Correction at Specified ( C C S ) Cost Algorithm


As observed in the previous subsection, the constant cost subproblem can be
infeasible. In that case, the current value of the cost function must be allowed to
increase. This can be done quite easily by slightly modifying the constant cost
182 J. S. ARORA AND C. H. TSENG

constraint as (e, d) 5 A, where A is a specified maximum increase in the cost. The


increase in cost can be specified based on the condition that the new cost does not go
beyond the previous cost at a feasible point, if known. Note again that the Q P
subproblem in this case can be infeasible if the increase in the cost specified in A is not
enough. Therefore A may have to be adjusted.

Ohservutions on Algorithms

The algorithms presented in the preceding sections are useful for interactive design
optimization. There are, however, certain limitations that should be clearly under-
stood:
I ) All the algorithms use linear approximations for the cost and constraint
Downloaded by [University of Cambridge] at 18:00 09 October 2014

functions. Thus all the conclusions stated previously are for only the linearized
subproblem.
2) The step size a is calculated using the desired reduction y and a linear
approximation for the cost function. The actual reduction in the cost function may be
smaller or larger than y depending on the nonlinearity of the cost function.
3) When the constraints are nonlinear, several constraint correction steps are
usually needed to reach the feasible region.
4) There are several cost reduction directions at a given feasible point. They depend
on the definition of the QP subproblem. It is difficult to determine the best possible
direction.
5) The Lagrange multipliers evaluated during the solutions of the Q P subproblems
can be quite different from their optimum values.

4 INTERACTIVE DESIGN OPTIMIZATION SOFTWARE IDESIGN

Most of the general purpose computer programs for design optimization do not have
an interactive capability or graphics. A program called IDESIGN (Interactive
DESIGN Optimization of Engineering system^)^.^ has most of the facilities discussed
earlier and will be briefly described. These capabilities permit the engineer to interact
with and control the optimization process. One can backtrack to any previous design
or manually input a new trial design. Design information can be displayed in a variety
of ways. The package has been designed to accommodate both experienced users and
beginners. The beginner can respond to one menu at a time as guided by the on-line
instruction. The expert can prepare an input data file and thus bypass immediate
menus. The software identifies and helps the user to correct improper responses.
The program can solve any general nonlinear programming problem given in Eqs
(I) and (2), linear programming problems and unconstrained problems. The program
has the option of following algorithms that have been robustly implemented by
extensive testing: (1) Cost function bounding (CFB) a l g o r i ~ h m ~(2)
, ~ ;Pshenichny's
linearization method (LINRM)'; (3) Sequential quadratic programming (SQP)
algorithm that generates and uses approximate second order information for the
Lagrange functionL1;(4) A hybrid method that combines the cost function bounding
and the sequential quadratic programming algorithms1'; and (5) Conjugate gradient
method for unconstrained problems.
INTERACTIVE DESIGN OPTIMIZATION

User Interface for IDESIGN


IDESIGN consists of a main program and several standard subroutines. In order to
solve a problem, the user must prepare additional subroutines for the program. The
input data such as the initial design, lower and upper limits on design variables,
problem parameters, and the program parameters to invoke options must also be
provided. The input data and options available in the program are described in the
User's manuaL6
The user must describe the design problem by coding the following four FOR-
TRAN subroutines:
USERMF: Minimization (cost) Function evaluation subroutine
USERCF: Constraint Functions evaluation subroutine
Downloaded by [University of Cambridge] at 18:00 09 October 2014

USERMG: Minimization (cost) function Gradient evaluation subroutines


USERCG: Constraint functions Gradient evaluation subroutine
A fifth subroutine USEROU can be provided to perform post optimality analyses for
the optimum solution and obtain more output.
Capabilities of IDESIGN
IDESIGN has most of the interactive graphics and decision making capabilities
outlined earlier. The program has been used to solve several classes of optimal design
problems:
1) Several small scale engineering design problems having explicit cost and
constraint functions.
2) Structural design problems modeled using finite elements, such as trusses,
frames, mixed finite elements, bridges, industrial buildings, high rise buildings, plate
girders, machine elements and many others.
3) Dynamic response optimization applications, such as vibration isolation, steady
state response, design for earthquake resistance, worst case design and transient
response problems.
4) Biomechanics applications such as muscle force distribution and contact force
determination problems.
5) Optimal control of systems; structural, mechanical and aerospace applications.
6) System identification problems, such as environmental and material modelling
problems.

Gradient evaluation 'The following capabilities to evaluate gradients and check


gradient expressions are available:
1) If the user does not program gradient expressions in USERMG and USERCG
subroutines, the program has an option to automatically calculate them.
2) An option is available to determine the optimum value of the change in the
variables for finite difference gradient evaluation.
3) If the user has programmed gradient expressions in USERMG and USERCG
subroutines, an option is available to verify them. If the gradient expressions are in
error, an option is available to either stop the program or continue its execution.
These options have proven to be extremely useful in practical applications.
184 J. S. ARORA AND C. H. TSENG

5 EXAMPLE PROBLEMS

Ten Member Structure


T o demonstrate the use of the interactive capabilities, a ten member structure that has
been used in numerous investigations is selected. The problem is described on page
242 in Ref. [9] and design data are given there. Loading Case IT is considered and
stress, displacement, member buckling, fundamental vibration frequency and design
variable bound constraints are imposed. There are ten design variables for the
problem-the cross sectional area of each member.

Continuous variable optimization The problem is solved in Ref. [9] with the lower
limit on the area as 0.10 in2. We shall, however, use 1.62 in%s the lower limit because
Downloaded by [University of Cambridge] at 18:00 09 October 2014

this is the smallest available section, and we want to compare optimum solutions with
discrete and continuous variables. The optimum weight reported in Ref. [I31 is
4783 Ibs with active constraints as natural frequency, vertical deflection at node 2,
stress in members 2,5, and 6, and minimum sizefor members 6 and 10. With the lower
bound and the starting design as 1.62, the optimum solution is: design variables =
+ + +
(2.82808 01, 1.62, 2.7262E 01, 3.3737E 01, 1.62, 4.0026, 1.3595E -t 0.1,
1.75448 + 01, 1.9130E + 01, 1.62); cost function = 5,396.5 Ibs, number of itera-
tions = 19, number of analyses = 21, maximum constraint violation = 8.244E - 10,
and convergence parameter = 2.660E - 05. The active constraints at the optimum
include frequency, stress in member 2, displacement at node 2 in the y direction, and
lower bound on variables 2, 5 and 10. Stress in member 8 is also nearly active. The
Lagrange multipliers for the six normalized colistraints are 3.924E + 02,33068 + 01,
+ + +
4.9678 03, 7.829, 2.051E 02, 1.405E 02. Note that very severe convergence
criteria are used in obtaining the precise optimum. With less strict criteria, a practical
optimum can be obtained with fewer analyses.
T o see the advantage of the interactive design optimization environment, we shall
solve the same problem using the interactive commands. Table 1 gives a snapshot of

Table I Interactive solution lor ten member structure with continuous


variables.

It. Maximum Cost Algorithm Data used


no. violation function used
in percent

CC constraint violation
CC constraint violation
CC constraint violation
CC constraint violation'
CC constraint violation
CCC violation, cost history
CC constraint violation
CCC violation, cost history
CCC violation, cost history
CC violation history
INTERACTIVE DESIGN OPTIMIZATION

Table 2 Interactive solution for ten member structure with discrete


variables.

It. Max. viol. Cost Algorithm Variables


no. in percent function used

6 2.02OE + 01 4722.73 CC 5(1.62), lO(1.62)


7 2.418 5389.28 CCC 2(1.80)
11 1.223E -01 5402.62 SQP I(30.0), 6(3.84), 7(13.5)
13 5.204E -04 5411.13 SQP 3(26.5), 9(19.9)
14 1.388 5424.69 - q13.5). S(16.9)
Downloaded by [University of Cambridge] at 18:00 09 October 2014

the design conditions at various iterations, decisions made and the data used in
making the decisions. It can be seen that for the first five iterations the constraint
violations are very large, so the constraint correction algorithm is used to correct the
constraints. Constraint correction a t constant cost is sought at iteration 6, and
constraint correction is desired a t iteration 7. At iteration 8 the cost function has a
value of 5382.61 and the maximum violation is only 2.4 percent. This may be
considered as a n engineering optimum. Another reasonable design is obtained at the
11th iteration with a maximum constraint violation of 1.45 percent and design
+ +
variables of (2.6836958 01, 1.734398, 2.6185E 01, 1.407173E + 01, 1.62, 3.77835,
+ + +
1.414866E 01, 1.846209E 01, 1.94570E 01, 1.62). At the 11th iteration a
command is given to find the mathematical optimum using the SQP algorithm. The
program finds the same optimum as before using 12 more analyses.
It can be seen from this example that designer interaction is extremely useful in
speeding up the optimization process. In the preceeding demonstration only the
constraint violation and cost function histories are used in interactive decision
making. Other capabilities, such as graphics, can be also profitably used to further
speed up the process or gain insights into the problem behavior.

Table 3 Interactive solution for 200 member structure.


- -

It. Maximum violation Cost Algorithm


no. in percent function used

CC
CC
CCC
CR
CCC
CR
CCC
CR
CCC
CR
CCC
CR
CR
CCC
186 J. S. ARORA A N D C. H. TSENG

Table 4 Comparison of interactive and automated solutions for two hundred member
structure.

Description Interactive Automated (SQP)

Maximum constraint violation (%) 7.705E - 03 -


8.8878 03
Optimum weight (Ib) 29526 29409
Number of iterations 77 105
Number of cost funclion evaluations 1 19 159
Number of total gradient evaluations 5203 9760
CPU time for analysis (sec) 1352(10.3%) 2663(9.l %)
CPU time for design sensitivity analysis (sec) 7221(55.0%) 13328(45.7%)
Total CPU time (sec) 13110 29156

Note: CPU time is for APOLLO DN560


Downloaded by [University of Cambridge] at 18:00 09 October 2014

Discrete variable optimization The capability of IDESlGN to fix variables can be


used to calculate a discrete optimum. It is desired that the final design variable values
must correspond to the following discrete values'(in2) (double angle section available
in the American Institute of Steel Construction code): 1.62, 1.80, 1.99, 2.13,2.38,2.62,
2.63, 2.88, 2.93, 3.09, 3.13, 3.38, 3.47, 3.55, 3.63, 3.84, 3.87, 3.88, 4.18, 4.22, 4.49, 4.59,
4.80,4.97, 5.12, 5.74, 7.22, 7.97, 11.5, 13.5, 13.9, 14.2, 15.5, 16.0, 16.9, 18.8, 19.9,22.0,
22.9,26.5, 30.0, 33.5. This is only a partial list of the available sections. It is chosen just
to demonstrate the procedure for discrete variable optimization.
Table 2 shows various iterations used to reach the discrete optimum point. The first
five iterations are the same as shown in Table 1. At the 6th iteration, it is determined
that design variables 5 and 10 are not changing, so they are fixed to their current
value. Similarly at other iterations, variables are assigned valucs from the available
set. At the 14th iteration, all variables have discrete values, the constraint violation is
about 1.4 percent and the optimum weight is 5424.69 which is an increase of less than
I percent from the previously obtained interactive continuous optimum.
It should be noted that with discrete variables several solutions near the true
optimum point are possible. A different sequence of fixing variables can give a
different solution. For example, starting from the optimum solution with continuous
variables, the following acceptable discrete solutions are obtained interactively:
a) 30.0, 1.62, 26.5, 13.9, 1.62, 4.18, 13.5, 18.8, 18.8, 1.62; cost = 5485.6, maximum
violation = 4.167% for stress in member 2.
b) Same as a) except that the 8th design variable is 16.9; cost = 5388.9 and
maximum violation = 0.58 %.
c) Same as a) except design variables 2 and 6 are 2.38 and 2.62; cost = 5456.8,
maximum violation = 3.74% for stress in member 2.
d) Same as c) except design variable 2 is 2.62; cost = 5465.4; all constraints are
satisfied.

Two Hundred Member Structure


A structure having 200 members is described on page 250 in Ref. [9] and was selected
for demonstrating the use of the interactive capabilities for large scale problems. The
structure has 77 joints, 150 degrees of freedom and three independent loading
conditions. Design variable linking is used to give
. . 96 design variables (cross-sectional
INTERACTIVE DESIGN OPTIMIZATION 187

areas) for 200 members. The problem has 1051 constraints (member stress, nodal
displacement, buckling, and fundamental vibration frequency). Explicit design vari-
able bound constraints are also imposed. To demonstrate the use of the interactive
capabilities, a bad starting design consisting of 1 in2 for all members is selected.
However, in actual practice a "better" starting point can be selected by experienced
engineers. The initial weight is 9963 1b and the maximum constraint violation is
6303 %. For obtaining optimum solutions, the constraint violations should be less
than 1 % and the length of the search step should be less than 0.05.
The optimum weight reported in Ref. [9] is 29725 Ib and a slightly better solution of
29409.2 Ib is reported in Ref. [13]. The iteration history with interactive commands is
given in Table 3 for some iterations. It can be seen that for the first five iterations the
maximum violation is very large. The constraint correction algorithm is used which
corrects constraints so that the maximum violation is under 20%. The constraint
Downloaded by [University of Cambridge] at 18:00 09 October 2014

correction at constant cost is used at iteration 6. At iteration 7, the violation is only


0.035 % and the cost function has a value of 37478. The cost reduction algorithm is
used which reduces the cost function by 475 Ib but constraint violation is increased.
The remaining iterations use the same procedure for correcting constraint violation,
using either constraint correction or constant cost algorithm. A t feasible points the
cost reduction algorithm is used to reduce the cost function. At iteration 77, both the
violation (0.008 %) and convergence parameter (0.046) satisfy the specified tolerance.
The optimum solution is 29526 Ib.
A comparison of interactive and automated solutions is given in Table 4. As the
results show, the total CPU time with the interactive procedure is less than half of that
with the automated procedure, and the number of iterations is also reduced. However,
the optimum cost function with the interactive procedure is sightly higher than that
with the automated procedure. The most important factor affecting efficiency in this
problem is the total number of constraint gradient evaluations. Note also that a
substantial amount of CPU time is used in design sensitivity analysis; 55% for the
interactive and 45.7% for the automated procedures. These results show that
interactive design optimization can be a powerful tool even for large scale problems.

6 CONCLUDING REMARKS

Interactive capabilities for practical applications of optimization techniques have


been described. Algorithms suitable for designer interaction have been given and
explained. A strong case was made for designer interaction in the optimum design
process. An interactive design optimization software having a wide range of capabili-
ties was described. The advantages of designer interaction were demonstrated by
solving a structural design problem. It was shown that with designer interaction, the
numerical optimum design process can be speeded up. In addition practical designs
can be assured by assigning discrete values to the design variables.
The interactive process is fairly efficient. For the present examples, it took about
one hour of connect time on an Apollo Workstation to solve all the problems after the
program was set up and input data had been entered. In conclusion, designer
interaction during the optimization process is highly desirable and recommended for
practical applications.
J. S. ARORA AND C. H. TSENG

REFERENCES

1. Adeli. H. and Phan, K. (1986) Interactive Computer-Aided Design o l Nonhybrid and Hybrid Plate
Girders. Computers & Structures, 22, (3). 267-290.
2. Arora. J. S. and Govil, A. K. (1977) Design Sensitivity Analysis with Substructuring. J. Engineering
Mechanics Division. ASCE. 103, (EM 4), 537-548.
3. Arora, J. S. and Haug. E. J. (1979) and (1980) Methods of Design Sensitivity analysis in Structural
Optimization. AlAA J., 17, (9), 970-74; and 18, (11). 1406-1408.
4. Arora. J. S. (1984) An Algorithm for Optimum Structural Design Without Line Search. Chapter 20 in
New Directions in Optimum Strurlural Design, (E. Atrek; R. H. Gallagher. K. M. Ragsdell and 0. C.
Zienkiewicz (Eds.)), John Wiley & Sons, N.Y., 429-441.
5. Arora. J. S. and Thanedar, P. B. (1986) Computational Methods for Optimal Design of Large Complex
Systems. Computational Mechanics. 1, 221-242.
6. Arora, J. S. and Tseng. C. H. (1987) /DESIGN User's Manual: Version 3.5. Optimal Design
Laboratory. Civil and Environmental Engineering, The University o l lowa, lowa City, IA 52242.
Downloaded by [University of Cambridge] at 18:00 09 October 2014

7. Belegundu, A. D. and Arora. J. S. (1984) A Study of Nonlinear Programming Methods lor Structural
Optimization. Int. J . Jor Numerical Methods in Engrg., 21, 1583-1599.
8. Belsare, S. V. and Arora, J. S. (1983) An Algorithm lor Engineering Design Optimization. Int. J.jor
Numerical Methods in Engineering, 19, 841-858.
9. Haug, E. 1. and Arora. J. S. (1979) Applied Optimal Design: Mechanical and Structural Systems, Wiley
Interscience, John Wiley & Sons, N.Y.
10. Hsieh. C. C. and Arora, J. S. (1984) Design Sensitivity Analysis and Optimization of Dynamic
Response. Computer Merhods in Applied Mechanics and Engineering, 43, 195-215.
I I. Lim. 0 . K. and Arora, J. S. (1986) An Active Set RQP Algorithm for Engineering Design Optimization.
Computer Methods in Applied Mechanics and Engineering, 57, 51-65.
12. Saouma. V. E. and Sikiotis, E. S. (1985) Interactive Graphics Nonlinear Optimization. Computers and
Structures. 21. (4). 759-770.
13. Thanedar, P. B.. Arora. J. S. and Tseng, C. H. (1986) A Hybrid Optimization Method and Its Role in
Computer-Aided Design. Computers and Structures. 23. (3). 305-314.

Você também pode gostar