This article was downloaded by: [Northeastern University]

On: 25 November 2014, At: 17:20

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

IMMUNE NETWORK SIMULATION WITH MULTIOBJECTIVE

GENETIC ALGORITHMS FOR MULTIDISCIPLINARY DESIGN
OPTIMIZATION
a a
A. KURAPATI & S. AZARM
a
Department of Mechanical Engineering , University of Maryland , College Park, Maryland,
20742, USA
Published online: 27 Apr 2007.

To cite this article: A. KURAPATI & S. AZARM (2000) IMMUNE NETWORK SIMULATION WITH MULTIOBJECTIVE GENETIC
ALGORITHMS FOR MULTIDISCIPLINARY DESIGN OPTIMIZATION, Engineering Optimization, 33:2, 245-260

To link to this article: http://dx.doi.org/10.1080/03052150008940919

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
 fig. Opr.. 2WO. Vol. 33, pp. 245-260 0 2000 OPA (Ovencar Publishen Assodation) N.V.
Rcprinu available directly from the publisher Published by limns under
Phot-pying psrmittcd by l i m w only thc Gordon and Brcach Scicna
Fublishcn imprint.
Printed in Malaysia.

IMMUNE NETWORK SIMULATION

WITH MULTIOBJECTIVE
GENETIC ALGORITHMS
Downloaded by [Northeastern University] at 17:20 25 November 2014

FOR MULTIDISCIPLINARY
DESIGN OPTIMIZATION
A. KURAPATI and S. AZARM*
Department of Mechanical Engineering, University of Maryland,
College Park, Maryland 20742, USA

(Received 7 May 1999; In final form 5 November 1999)

The paper presents a method called MOGA-INS for Multidisciplinary Design Opti-
mization (MDO) of systems that involve multiple competing objectives with a mix
of continuous and discrete variables. The method is based on the Immune Network
Simulation (INS) approach that has been extended by combining it with a Multi-
Objective Genetic Algorithm (MOGA). MOGA obtains Pareto solutions for multiple
objective optimization problems in an all-at-once manner. INS provides a coordination
strategy for subsystems in MDO to interact and is naturally suited for genetic algorithm-
based optimization methods. The MOGA-INS method is demonstrated with a speed-
reducer example, formulated as a two-level two-objective design optimization problem.

Keywordr: Multiobjective optimization; Genetic algorithms; Multidisciplinary

optimization

1. INTRODUCTION

Multidisciplinary Design Optimization (MDO) is a methodology for

optimization-based design of a system composed of multiple sub-
systems (or disciplines) that are coupled. Very often, the overall sys-
tem design is managed by a central authority (such as a team leader)
by coordinating the design activities of the disciplines involved in

*Corresponding author. e-mail: [email protected]

 246 A. KURAPATI AND S. AZARM

such a distributed design environment. With increasing complexity

in disciplinary analysis methods, the MDO problem has become more
and more challenging. It calls for more general techniques to solve a
broader class of MDO problems.
Many heuristic approaches to MDO have evolved over the years. A
general classification of MDO approaches would be 'hierarchic' [l] or
'non-hierarchic' 1121 depending on the type of systems to which they
apply. In hierarchic systems, children disciplines are coupled only to
Downloaded by [Northeastern University] at 17:20 25 November 2014

parent disciplines and not to each other. However, in non-hierarchic

systems, there is no restriction on how the disciplines are coupled.
Solving a non-hierarchic MDO problem directly can be a challenging
task [3]. Some researchers [2, 41 have attempted to classify the various
MDO approaches. Approaches that have evolved more recently are
Collaborative Optimization (CO) [5,6], Multi-Objective Collaborative
Optimization (MOCO) 171, and single-objective Immune Network
Simulation (INS) [8], among others. Nevertheless, there are other
features in MDO which have not yet been exploited or accounted for.
Some of these features are discussed next.
Most of the MDO approaches handle only continuous design vari-
ables wherein they require gradient information to coordinate discipli-
nary solutions. In reality, however, MDO problems may have a mix
of continuous and discrete design variables, and for such problems
the gradient information may not be available. Some work in MDO
has been done to handle mixed design variables [9, 101 wherein discrete
variables are only handled at the disciplinary level. Others, as in INS
[8], have used Genetic Algorithms (hereafter referred to as GAS) [I I]
to solve single-objective MDO problems with mixed variables at both
coordination (system-level) and disciplinary levels. However, it is quite
possible that an MDO problem may have more than one design
objective. There is more than one solution for such a multiobjective
MDO problem, and these solutions are referred to as Pareto (or non-
inferior) solutions. Some of the traditional multiobjective optimiza-
tion methods [12], like the weighting method, have been used in
multiobjective MDO with continuous variables (see, for example,
MOCO [7]) to obtain Pareto solutions in a point-by-point manner.
The shortcomings of the weighting method, however, have well been
documented [13]. On the other hand, Multi-Objective Genetic Algo-
rithms (MOGAs) [14, 151 have been developed to generate Pareto
 IMMUNE NETWORKS AND MDO 247

solutions in an alliat-once manner for single-level design optimization

problems but not for an MDO problem.
The overall goal of this paper is to develop an approach, which can
(i) solve an MDO problem with multiple conflicting system objectives
and obtain Pareto solutions in an all-at-once manner, and (ii) handle a
mix of discrete and continuous design variables and support subsystem
autonomy in a multidisciplinary environment. The proposed MDO
approach, MOGA-INS, extends and integrates MOGA [I 51 and INS
Downloaded by [Northeastern University] at 17:20 25 November 2014

[8] techniques.

2. THE MOGA-INS APPROACH

2.1. Background
MOGA is based on GAS that have been modified to handle multi-
objective optimization so that the problem need not be converted
into a series of single-objective optimization problems [15]. The fitness
assignment strategy in MOGA is based on the concept of non-infer-
iority and not on the objective function value as in a single objective
GA. This is the fundamental difference between MOGAs and single
objective GAS. Since any GA-based method works with a population
of solutions, MOGA exploits this feature to detect the Pareto solu-
tions in a single stage (or in an all-at-once manner). The idea is to iden-
tify non-inferior individuals in a population and have them evolve
towards the final Pareto solution set as the population is evolved.
The INS strategy can be used only in conjunction with a GA-based
optimization approach and so far has been applied only to single
objective MDO problems [8]. The INS strategy is inspired by the
way the biological immune system functions. The immune system in
the body is a defense mechanism, capable of identifying foreign mole-
cules (termed as antigens) that enter the body and manufacturing anti-
bodies to destroy the antigens. The antibodies are assembled from
components manufactured in different subsystems of the immune
system using a combinatorial mechanism. The antibody has to match
the structure of the antigen in order to eliminate the antigen. So, to
apply the INS strategy, it is necessary to define the antigens for which
the antibody population evolves. By simplifying the representation of
antibodies and antigens to binary strings, INS can be adapted along
 248 A. KURAPATl AND S. AZARM

with a G A for the solution of MDO problems. In INS, it is important

to know the definition for a generalist antibody. For instance, as-
suming that matching the binary digits in an antigen is tantamount
to matching the complex three dimensional structure of the antigen,
the antibody represented by 101 is a generalist antibody with respect
to the antigens 11 1, 100 and 001, since it matches in exactly two of
the three binary digits in each antigen.
Downloaded by [Northeastern University] at 17:20 25 November 2014

2.2. Assumptions
The assumptions made for the execution of MOGA-INS are listed in
the following:
(a) The MDO problem is either in a hierarchical form, or has been
converted to that form.
(b) Each subsystem solves a multiobjective optimization problem,
with its multiple design objectives being the same in each sub-
system.
(c) The ith subsystem handles a portion of the overall design vector
x designated by xi and the design variable vectors are mutually
exclusive among the subsystems.
(d) All subsystems have information about all the variables (both
independent and state variables), either as variables in that sub-
system or as fixed parameters.

2.3. The MOGA-INS Steps

The steps in the MOGA-INS procedure are given as follows.
Step I - Identification of a hierarchically decomposed system
Hierarchical decomposition may be done by an ad-hoc technique
across disciplinary boundaries, or by using a formal decomposition
approach. For converting a non-hierarchic system into a hierarchic
one, a coupled three-discipline (-subsystem) problem is considered, as
shown in Figure 1, wherein the three subsystems are nonhierarchically
coupled through the various input and output vectors. The subsystems
are basically coupled through the shared design and state variables.
In Figure I and Figure 2, x,hi contains design variables shared by
subsystem i with other subsystems. Denote the outcome of the ope-
+
ration: ( ~ s hUl ~ s h 2 U ~ s h-3Xshl n X s h 2 - X s h 2 n X j h 3 - Xshl n X s h 3 f
 IMMUNE NETWORKS A N D MDO

Subsystem 2 Subsystem 3
Downloaded by [Northeastern University] at 17:20 25 November 2014

XlhZ . x1 gl. ./i L?J,./~ Xrhl .XI

FIGURE 1 A threediscipline coupled system.

@ SSI Q SS2 @ SS3

FIGURE 2 Venn diagram for the design vector.

x , h l + l Xsh2 n xSh3)by x S h . Then, x S h r X I , x2 and x3 (the shared design

variables, and the variables handled exclusively in subsystems 1,2 and 3,
respectively) together form the vector of system variables, x, as shown in
Figure 2. The subsystems are also coupled through the state vectors yu
(output vector of the ith subsystem required as an input to the jth
subsystem; i,j= 1,2,3). The variables in the state vector are known as
state variables. The quantity giis the constraint set handled in subsystem
i and fi is the objective function vector computed in subsystem i
(i= 1,2,3), which is to be minimized. MOGA-INS cannot be applied
directly to a system in the form shown in Figure 1 because it does not
satisfy the assumptions required for applying MOGA-INS.
Figure 3 shows the system in Figure 1 after it has been converted to
a hierarchically decomposed form in order to apply MOGA-INS. In
 A. KURAPATI A N D S. AZARM
Downloaded by [Northeastern University] at 17:20 25 November 2014

FIGURE 3 Decomposed structure for MOGA-INS.

Figure 3, XI, X2 and X3 (which represent the design variables handled

by subsystems 1, 2 and 3, respectively) are three mutually exclusive
vectors to satisfy assumption (c). For instance, XI could contain
variables from region 1 and a few variables from regions 4, 5, 6 and
7 (Fig. 2), where x = ( X I , X2, X3) Because of assumption (d), each
subsystem also has information about the state vector yi. In Figure 3,
Xi' is the optimum design vector population in subsystem i based on
a disciplinary optimization and x + is a set of updated design vectors
generated by INS and fed to each subsystem by the coordinator.
Step 2 - Coding of variables for MOGA This step is illustrated
graphically in Figure 4. A random initial population of size P cor-
responding to the design vector x of string length L is created. This
is nothing but the antibody population and is used by the coordina-
tor to perform INS.
Also, a random initial population in each subsystem with the same
string length L is created, the difference being that only certain
portions of the string are active in each subsystem. For instance, the
portion of the string corresponding to XI having a string length L1
would be active in subsystem 1. The remaining portion of the string
in subsystem 1 would correspond to X2 and X3, which are treated as
fixed parameters in this subsystem. As the design variable vectors
 IMMUNE NETWORKS AND MDO
Downloaded by [Northeastern University] at 17:20 25 November 2014

Suhsgaem I

FIGURE 4 MOGA-INS structure for a three-discipline coupled system.

are mutually exclusive, if a certain portion of the string is active (or

varied) in one subsystem, then it is treated as fixed in every other
subsystem. If Liis the active string length in subsystem i (i= 1,2,3),
+ +
then L = L, L2 L,.The population size in each subsystem need
not be equal. The active string length in each subsystem is small
Downloaded by [Northeastern University] at 17:20 25 November 2014
Downloaded by [Northeastern University] at 17:20 25 November 2014
 254 A. KURAPATl AND S. AZARM

shaft design practices. The problem has seven design variables and on
each variable an upper and lower limit is imposed. There are 11 other
inequality constraints, of which one is an objective constraint, i.e., the
objective function ( f2) appearing also as a constraint. The mathema-
tical formulation, Eq. (I), of the single level problem is now described.
The first design objective, fi, is to minimize the volume. The second
design objective, f2, is to minimize the stress in one of the two gear
shafts. The seven variables are, gear face width ( X I )teeth
, module (x2),
Downloaded by [Northeastern University] at 17:20 25 November 2014

number of teeth of pinion (x3 - integer variable), distance between

bearings I ( 4 , distance between bearings 2 ( x ~ )diameter
, of shaft 1
(x6 - discrete variable), and diameter of shaft 2 ( x , - discrete
variable). The variables xs and x7 can take 1 I discrete values uniformly
distributed within their respective allowable range. The upper and
lower limits on the design variables, Eq. (2), are implicit for a GA
based optimization, and hence are not counted as separate constraints.

Minimize f~ = 0 . 7 8 5 4 ~ 1 4 ( 1 0 ~ $+31 4 . 9 3 3 -43.0934)

~~
- I .508x1(4 +G)+7.477($+x$)
+0.7854(x&+x&)

\/(%)'+I.@ x 10'
Minimize f2 =
0.14
Subject to:
 IMMUNE NETWORKS AND MDO 255

The lower and upper limits on the seven variables are as follows:
Downloaded by [Northeastern University] at 17:20 25 November 2014

3.1. Application of MOGA-INS

Step I The decomposition has been done in an ad-hoc fashion. The
problem has been posed as a two subsystem problem. It was
decided to group the constraints g,, g, and g , , (with the
corresponding active design variables being x5 and x7) into
subsystem 2, and the remaining constraints into subsystem 1
(with the corresponding active design variables being x,, xz,
x3, xq and x6).
Step 2 The string length corresponding to the design variables has 36
binary bits. A random population with this string length is
initialized. This is the antibody population. The string length
corresponding to the active variables in subsystem 1 has 26
bits while that in subsystem 2 has 10. A random population
is initialized in both subsystems, using the design variables
information mentioned above.
Step 3 MOGA is performed for five cycles. One cycle consists of
performing MOGA on subsystem 1 and subsystem 2 for a
few generations using the information updated from the pre-
vious cycle.
Step 4 INS is performed on this decomposed problem using an
antibody population generated in Step 2. Some copies of
generalist antibodies are introduced into the two subsystems
to update the variables in the subsystems.
Step 5 Steps 3 and 4 are repeated until convergence.
The Pareto set results obtained using MOGA-INS is compared
against the Pareto set generated using a single-level MOGA (or
MOGA-SL), as shown in Figure 6. The single level MOGA required
5109 function calls to generate 12 Pareto points using a population size
of 50 individuals (426 function calls per Pareto point). The multi-level
 A. KURAPATI A N D S. AZARM

1800 -
1600 -

- 1400 -
5 1200
-e
U)

U)

8o I
roo0

600
400
-

,
0
0

::,ooo;: -; , ,
Downloaded by [Northeastern University] at 17:20 25 November 2014

2000 2500 3000 3500 4000 4500 5000

volume (1,)

FIGURE 6 Comparison of Pareto sets for the speed reducer example.

MOGA-INS required 21931 function calls to generate 7 Pareto points,

again using a population size of 50 individuals in each subsystem
(3133 function calls per Pareto point). The two solutions are in good
agreement as they lie on approximately the same Pareto frontier.

3.2. Discussion: improvements and Limitations

For the speed reducer example, the computation effort (in terms of
number of function calls) required by MOGA-INS is nearly four times
that required by a single level MOGA. Let us refer to the MOGA-INS
procedure described so far as MOGA-INS-Old. T o improve the
efficiency of MOGA-INS-Old, a numerical experiment was carried out
to find the effects of various factors. Some of the factors considered
are the constraint set partitioning between the subsystems, popula-
tion size, the mutation probability and the crossover type. Let us now
refer to the procedure with all these factors considered as MOGA-
INS-New. The number of function calls and CPU time (a measure
of the computation effort) depend on the population size as shown
in Figures 7(a) and 7(b), respectively.
For the example, it was found that the computation effort is
significantly increased with duplication of constraints in the sub-
systems. The computational effort for MOGA-INS in the range of
population size 50 to 500 is always more than for the single level
MOGA (MOGA-SL in Fig. 7). The mutation rate is kept constant
throughout in MOGA-INS-Old. In MOGA-INS-New, the mutation
 IMMUNE NETWORKS AND MDO 257

-e- MOGA-INS-Old
MOGA-SL
. MOGA-INS-New
Downloaded by [Northeastern University] at 17:20 25 November 2014

Pop size

250
+MOGA-INS-Old
-4- MOGA-SL
o x MOGA-INS-New
150
E
'1

a
m
50 /
/-

0
0 50 100 150 200 250 300 350 400 450 500
(b) pop size

FIGURE 7 Function calls (a) and elapsed times (b) v5. population size

rate was allowed to change from a high initial value of 0.2 to a low
value of 0.002, as MOGA-INS proceeded. The contention is that the
higher the mutation rate, the more is the diversity in the population
Downloaded by [Northeastern University] at 17:20 25 November 2014
Downloaded by [Northeastern University] at 17:20 25 November 2014
 260 A. KURAPATI AND S. AZARM

NASAIUSAFIISSMO Symposium on Multidisciplinary Analysis and Optimization,

Panama City, Florida, September 7-9, 1994, Paper No. AIAA-944325-CP.
[6] Braun, R. D. and Kroo, I. M. (1995). Development and application of the
collaborative optimization architecture in a multidisciplinary design environment.
Multidisciplinary Design Optimization: State of the Art, Alexandrov, N. and
Hussaini. M. Y. (Eds.), Proceedinas - of
. the ICASEINASA Lamley - . Workshop on
M D O , SIAM, ~ h i l a d e l ~ h i a .
[7] Tappeta, R. V. and Renaud, J. E. (1997). Multiobjective collaborative optimiza-
tion. Trans. ASME. Journal o f Mechanical Desipn. - . 119.. 403-41 1.
[8] Hajela, P. and ~ e e J. , (199j). GAS in decomposition based design-subsystem
interactions through immune network simulation. Structural Optimization, 14,
248 - 255.
Downloaded by [Northeastern University] at 17:20 25 November 2014

[9] Balling, R. J. and Gale, D. L. (1998). Collaborative optimization of systems

involving discrete design at the discipline level. Trans. ASME, Journal o j
Mechanical Design, 120, 32- 39.
[lo] Praharaj, S. and Azarm, S. (1992). Two-level nonlinear mixed discrete continuous
optimization based design: an application to printed circuit board assemblies.
Trans. ASME, Journal of Electronic Packaging, 114(4), 425-435.
[I I] Levine, D. (1996). PGA Pack ParallelGenetic Algorirhm Library. Mathematics and
Computer Science, Argonne National Lab, Argonne, IL.
[I21 Chankong, V. and Haimes, Y. Y. (1983). Mulriobjective Decision Making: Theory
and Merhodology, Elsevier Science Publishing Co., Inc., New York.
[I31 Das, I. and Dennis, J. E. (1997). A closer look at drawbacks of minimizing
weighted sums of objectives for pareto set generation in multicriteria optimization
problems. S~ructuralOprimizarion, 14(1), 63-69.
[I41 Fonseca, C. M. and Fleming, P. J. (1993). Genetic algorithms for multiobjective
optimization: formulation, discussion, and generalization. Proceedings ofrhe FiJih
Internarional Conference on Genetic Algorithms, Forrest, S. (Ed.), pp. 416-423.
[IS] Narayanan, S. and Azarm, S. (1999). On improving multiobjective genetic algo-
rithms for design optimization. Slrucrural Oplimization, 18, 146- 155.
[I61 Goldberg, D. E. (1989). Generic Algorithms in Search, Optimizarion and Machine
Learning. Addison-Wesley, New York.
[I71 Golinski, J. (1970). Optimal synthesis problems solved by means of nonlinear
programming and random methods. Journal of Mechanisms, 5, 287-309.