Professional Documents
Culture Documents
HMK Dissertation
HMK Dissertation
net/publication/33984207
CITATIONS READS
135 926
1 author:
Harrison M Kim
University of Illinois, Urbana-Champaign
103 PUBLICATIONS 2,117 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Harrison M Kim on 28 May 2014.
SYSTEM DESIGN
by
Doctoral Committee:
by
large complex artifacts: How to propagate the desirable top level design specifications (or
consistent and efficient manner. Consistency means that all parts of the designed system
should end up working well together, while efficiency means that the product
development process itself should avoid iterations at later stages, which are costly in time
and resources.
In this dissertation target cascading is formalized in a process modeled as a
multilevel optimal design problem. Design targets are cascaded down to lower levels
using partitioning of the original problem into a hierarchical set of subproblems. For each
deviations from the propagated targets and thus achieve intersystem compatibility.
between the TC and the HOC formulations, i.e., both decomposed problems possessing
the same objective and constraint sets provides a theoretical foundation for asserting that
the TC algorithm will reach the target deviations.
vehicle chassis design problem. The target cascading process implementation in an actual
organization is also presented along with sample model forms for collecting and
To my parents
ii
I press on toward the goal to win the prize for which God has called me
Philippians 3:14
iii
ACKNOWLEDGEMENTS
The first meeting with my advisor, Panos Papalambros, on November 14, 1996 is
still vivid in my memory. He was visiting Korea and Japan and I was lucky to receive his
phone call early in the morning in my office. I went to see him in that afternoon and we
had a nice talk for about two hours and I was so excited about the potential chance to work
with him. Whenever I read my diary back in those days, I can still feel the excitement and
energy after the meeting with him. I was very fortunate to have Panos as my advisor,
teacher, and counselor and I want to sincerely thank Panos for his guidance throughout my
Ph.D. study.
I also like to thank Dr. Nestor Michelena for his idea and help from the beginning
of the target cascading project. Whenever I needed help, he was there to sit down with me
to figure out the problems. I wish him good luck and health. Many thanks go also to
Professor John Birge, Professor Huei Peng, and Dr. Tao Jiang for their advice on this
dissertation.
Colleagues at the ARC, Dr. Loucas Louca, Geoff Rideout, Dr. Zoran Filipi, Bo
Chen: I enjoyed working with them and their expertise was a great help. The ground
This dissertation was not possible without the support of my friends. First, my
dearest Eunji, for her love and support and also many meals she cooked for me. Last two
years were wonderful with her. All my precious ODE friends, Sig, Shinzi, Julie, Zhifang,
Mike, Ryan, Nnaemeka, Chris, Panayiotis, Matt, George, Jay, John, Hosam, Michael,
Olena: I thank all of them for the fun time and all the “research parties” every night.
Last but not least, I like to thank my parents for their love and prayers. I am glad
that now I can give them back something very small. My brother, Hyungham, and his
family, my beloved sister Minjoo, thank you all for the love and support.
iv
TABLE OF CONTENTS
DEDICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
CHAPTER 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
TARGET CASCADING IN PRODUCT DEVELOPMENT
CHAPTER 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
HIERARCHICAL OPTIMIZATION FOR LARGE-SCALE DESIGN
v
CHAPTER 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
TARGET CASCADING FORMULATION
CHAPTER 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
COORDINATION STRATEGY FOR TARGET CASCADING
CHAPTER 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
DEMONSTRATION STUDIES
CHAPTER 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
CONCLUSIONS
APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
vii
LIST OF FIGURES
FIGURE
1. Typical system life cycle (adapted from NASA Systems Engineering Handbook
[70]). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
14. Interaction of linking variables: a special case with a supersystem and two systems.
viii
16. Schematics of information exchange between system design and analysis model
17. Schematic of information exchange between subsystem design model and analysis
model: the subsystem analysis model takes subsystem level design variables and
18. Schematics of information exchange between level i design and analysis model and
20. Block D-decomposition of the functional dependence table (adapted from Michel-
ena et al [50]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
26. Cornering of a bicycle model (Bold faced symbols are design variables.) . . . . 88
31. Schematic of information exchange for vertical and cornering tire models . . . 98
ix
34. Comparison between design scenario A (baseline) and B . . . . . . . . . . . . . . . . 106
x
LIST OF TABLES
TABLE
3.1. Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.2. Optimal designs from All-At-Once (AAO) and Target Cascading (TC) models 82
E.1. List of hard point locations required for model input . . . . . . . . . . . . . . . . . . . 136
xi
E.3. Optional suspension parameters for future dynamic simulations . . . . . . . . . . 137
xii
CHAPTER 1
TARGET CASCADING IN PRODUCT
DEVELOPMENT
systems and it has become a central theme in the design of large complicated systems,
such as aircraft and automobiles. Systems engineering has been defined as (NASA
In simple terms, the approach consists of target identification, target cascading, creation of
multiple system design concepts, performance check, selection and implementation of the
degree of satisfaction. Targets are the goals to be achieved through the design process,
which are traditionally determined based on customer needs, corporate interests, and
government regulations.
is usually introduced at the early stages of the development process. A typical system life
cycle is given in Figure 1. The system definition stage includes target setting, locating
1
available analysis models, verification of design architecture, and formulation of design
problems at various levels of a modeling hierarchy. Once the overall system is defined, the
this stage and production and customer support then come along in a typical system life
cycle.
Subsystem definition
Production
System definition Fabrication,
Preliminary Detailed assembly,
design design integration, Customer
test support
Figure 1. Typical system life cycle (adapted from NASA Systems Engineering
Handbook [70]).
In the present chapter, the concept of a target cascading process will be introduced
within the overall product development effort. Starting with a general motivation for using
a systems approach, the main steps in product development will be outlined, where
mathematical models and computer simulations are indispensable tools for efficient, high-
quality design.
Much of the motivation for the work described in this dissertation comes from
recent efforts in the automotive, aerospace and other industries to formalize the product
development process and to take full advantage of computer aided engineering (CAE)
tools [16]. A complex product development process is most efficient when the various
required design tasks can be accomplished in a truly “concurrent” manner. In some sense,
concurrency means that tasks should be conducted in parallel without isolation from each
2
other, so that decisions are made taking all product requirements into account. Product
teams of diverse specialists are set up and communicate regularly. Model management and
group decision making in a computer-aided environment, and computer support for design
team collaboration have been suggested by many researchers (refer to [55][77][78][80] for
a few examples). Ironically, experience shows that such efforts also have serious adverse
effects because decisions are not made sufficiently quickly when trying to satisfy
everyone’s viewpoint. In another sense, then, concurrency means that each task should be
conducted in isolation or with the least amount of interaction with other tasks, so that each
specialist or team can concentrate on their own specialized task. Yet, clearly, interactions
do exist and task isolation could lead to costly downstream iterations. The effect of R&D
team co-location on communication patterns was studied by Bulte and Moenaert [15] in
terms of the volume of communication, which can effect the efficiency of the
collaboration.
The way out of this dilemma is to be able to identify the key links among design
tasks, reach the associated trade-off decisions early, and then let the individual tasks be
conducted separately. Subsequent interactions can take place when the specifications
given to each task turn out to be difficult or impossible to meet, and a refinement of joint
decisions is necessary. The target cascading process attempts to achieve just that. The
important specifications for the entire system as well as for each system element
(subsystems and components) are identified first, particularly those that will have
influence on other parts of the system. Values (targets) are assigned at the top level,
usually based on criteria supplied by management. These targets are “propagated” to the
rest of the system and appropriate values are assigned to the expected performance of each
element of the system. Design tasks are then executed for each individual element, and
any interaction with the rest of the system is revisited only when a target cannot be met.
One expects that the targets set are both consistent and achievable by the particular system
3
element. Target satisfaction is usually well understood by the specialists, so it is
x Supersystem x
Ca
s
ca
de
x System x
Do
Re
w
b
n
x Subsystem x
al
an
ce
x Component x
Up
Figure 2. Two approaches to target cascading.
Conceptually there can be two different ways for target cascading: an all-at-once
design models at multiple levels within one model. Top design targets are cascaded down
to the lowest level in one loop (see Figure 2). On the other hand, in a decomposition
approach targets are cascaded down level by level to lower levels and are rebalanced up
based on the lower level designs. In the all-at-once approach, complexity of the design
loop increases as the target cascading process is integrated in one loop. Detailed low level
models are required for the process, which might not be available at the early stages of the
design process. Also, some changes in one part of the design can affect the rest of the
design because of the integrated loop, which can hinder design autonomy in individual
design teams. In the decomposition approach, efficiency can become an issue because of
the added cost associated with coordinating analyses based on different disciplines. Also,
models for mid-level analysis might not be available, i.e., there can be a simulation and an
attendant design gap between the higher levels and the lower levels. Each approach has its
pros and cons; however, in the context of systems engineering, the decomposition
4
Target cascading in automotive vehicle design can be viewed as a four-step
process:
(iv) verify that the resulting design meets overall vehicle mission targets.
This process is efficient if appropriate CAE models are available to analyze design
decisions that can be made through a formal decision-making process. The results, of
course, will be as good as the available analysis models, and this is a practical challenge:
The models must capture the salient characteristics of the system interactions without
being burdened by cumbersome details. Typical CAE tools tend to be used for very
sophisticated analyses and are expensive to develop and to compute, while “back-of-the-
There are several research efforts that have explored system design methodologies
are examples of efforts in this direction (see, for example, Sobieski et al. [76], Cramer et
(Braun [11], Braun et al. [11][13][14], Tappeta and Renaud [82]). This approach attempts
to achieve overall design targets by coordinating the coupling effects among disciplines,
hierarchy (Braun et al. [14], Rawling and Balling [62]) and the use of a response surface
method combined with trust regions (Sobieski [71], Sobieski et al. [72]). In collaborative
5
between the interaction variables and the targets, and should become zero at the optimum.
system level. During iterations, subproblems may return different values for an
interdisciplinary variable, and this can cause convergence difficulties by not satisfying
equality constraints at the system level. These phenomena will be further investigated in
Chapter 2. The idea of minimizing discrepancies of the terms introduced among design
subproblems is very similar to the original Wismer and Chattergy dual coordination
method (Wismer and Chattergy [90]), where additional discrepancy terms are introduced
problems without fully addressing generalizations to the multilevel structure. The target
different from collaborative optimization, it shares ideas with Wismer and Chattergy’s
subproblems. Responses and linking variables are introduced to identify vertical and
given level are achieved by minimizing discrepancies of responses computed from levels
Target cascading is a system design approach enabling top level design targets to
be cascaded down to the lowest level of the modeling hierarchy. The steps involved are: (i)
development of appropriate models, (ii) partitioning the system, (ii) formulating the target
cascading problems for each element of the partition, and (iv) solving the partitioned
problem through a coordination strategy to compute all stated targets. In this section we
6
discuss the issues involved in these steps. Figure 3 gives a schematic of the target
cascading process in systems engineering. Starting from the top design targets, design
targets for each partitioned system are cascaded and the partitioned designs are evaluated.
fashion and decomposed systems can achieve the targets, then embodiment design follows
System No System
III Design III
Targets are set based on customer needs, government regulations, and company
interests. Goal setting theory goes back to early 20th Century studies in management
theory and psychology (Locke and Latham [44]). Translating customers’ needs into target
setting is performed by techniques such as quality function deployment (QFD) (Akao [1]).
With the help of QFD, the desired quality from a customer perspective is reflected in the
product. In engineering design, the appropriate “design authorities” set specific targets
using their expertise from past experience, market demand predictions, and regulations.
The effect of setting different targets in different design scenarios is investigated in the
expensive models are generally not appropriate for target cascading, inexpensive models
should replace expensive ones through surrogate modeling. Given the analysis models,
verification of the individual optimal design models for all system elements is necessary
for feasibility and boundedness checking (see, e.g., Papalambros and Wilde [58]). This
model development stage is the most time-consuming, but without good models
understanding of the relevant details of analysis models for target cascading practice in an
such as object, aspect (or discipline), and model based (Wagner [87]). Object and aspect
partitioning are “natural” partitions and typical large companies employ both partitions
partitions its organization into powertrain, body, chassis, or electronics divisions, but has
partitioning, design variables are categorized into linking variables, common to more than
In the target cascading formulation the easiest way is to start with an object
partition and recognize that each design problem at a given level is likely to be
8
multidisciplinary. The exact partitioning choice will also depend on the availability of
models, so this task should be done carefully and should be considered as subject to
After partitioning the original problem into subproblems at multiple levels, the
linking variables between subproblems at the same level and the responses linking
elucidated in Chapter 3, but the term is used here to indicate the output of an analysis
elements will be identified and cascaded in the objective functions of the lower level
hierarchical coordination strategy there is a master problem and one or more subproblems.
The master problem is solved for the linking variables that are then input as parameters to
the subproblems. The subproblems are then solved for local variables that are input as
However, unlike the usual hierarchical coordination, the linking variables are transferred
to the lower level problems as constant targets after solving the current level problem.
Furthermore, some of the current level design variables are also transferred to the lower
level as response targets. In the subsystem level problem, the objective function is to
minimize the discrepancy between these target values coming from the upper level system
and from the current level subsystem design, where the linking variables and responses are
now design variables not constants. These ideas will become more clear in the
9
1.2.5 Embodiment Design
Once targets are set for the individual design problems by a successful target
cascading process, all interactions among design problems, such as linking variables and
responses, are specified. Maintaining these values as fixed parameters, the simple models
used in the target cascading process are replaced with more detailed comprehensive
models for embodiment design, which will have many more variables and design degrees
of freedom. A new local optimal design problem can be formulated and solved. If current
design targets cannot be realized in the more detailed models (the new problem is
infeasible), the designer must either explore the local constraints for relaxation or return to
The steps described above are not simple, but they are systematic. Forcing the
design teams to create the correct models and to negotiate the selection of targets goes a
long way towards a successful process. The optimization formalism and attendant
numerical solutions help put any further needed negotiations on a rational basis.
This chapter has given an overview of the target cascading process in the context
coordination strategies for target cascading and its application to engineering design
frameworks are discussed, and in Section 2.2, coordination methods with proven
convergence are reviewed along with their pros and cons and range of applicability.
at multiple levels. Starting with the three-level case, general multilevel formulations are
10
given. Constraint qualifications for additional constraints introduced for the minimum
coordination (HOC) is reviewed and the two-level case of the TC formulation is proven to
A simple example to validate the target cascading approach and check whether
target cascading gives the same optimal solutions to optimization problem is given in
Chapter 5. A case study for a ground vehicle chassis design is then presented and the
organization is described. The process is explained step-by-step, from the analysis model
collection, to mapping the hierarchy, application of the target cascading, and embodiment
design. The chapter also reflects experiences gained during the application of the proposed
target cascading process at an industrial setting and at the Automotive Research Center at
and contributed to the placement of the methodology presented here on a theoretical and
practical foundation.
11
CHAPTER 2
HIERARCHICAL OPTIMIZATION
FOR LARGE-SCALE DESIGN
diverse but distinct design expertise towards achieving overall coordinated optimal
design. “Multidisciplinary design” essentially means that different design teams are
cooperating in the design process. Different design teams can be represented as: (1)
durability, and dynamics; (2) physically decomposed design entities, such as powertrain,
often set in advance at the higher level. However, an individual design entity can be
decomposed further at lower levels, and for this there are several partitioning methods
general overview of MDO formulations was presented by Cramer, et al. [18]. In this
chapter, several formulations and coordination strategies are reviewed relevant to target
Section 2.2 presents Wismer and Chattergy’s coordination methods for the constrained
12
optimization problem followed by some recent algorithms, Nelson’s sequentially
approach (Azarm and Li [4], Braun [11], Dantzig and Wolfe [19], Grana and Torrealdea
[25], Kim et al. [38], Kirsch [39][40], Michelena et al. [48][50], Nelson [53][54], Park et
al. [59], Shimizu and Aiyoshi [69], Sobieski [73][76]), design problems of multiple levels
making process is applied. In the nonhierarchical approach (Pan and Diaz [57], Renaud
and Gabriele [63][64], Sobieski [74]), decomposed design problems are organized in a
decision loop with no single “master” in the decision hierarchy.
Cramer et al. [18] for handling multiple design disciplines, and Sobieski’s hierarchical
(CO) is also discussed for solving multidisciplinary design problems with bilevel
following certain formal structures. Cramer et al. [18] formulated different ways to
13
integrate multiple analysis models under one design model following three model
structures: multidisciplinary feasible (MDF), individual discipline feasible (IDF), and all-
maintained during optimization. Here the “feasibility” for a single discipline means that
the equations the discipline code is intended to solve are satisfied. A multidisciplinary
analysis (MDA) is achieved when 1) we have single discipline feasibility in all disciplines,
and 2) the input to each discipline corresponds to the output of the other disciplines via
model feasibility is maintained at every optimization iteration. At the other extreme, in the
Between these extremes lies the IDF approach, where only individual discipline feasibility
The number of design problems in each of the formulations above is one, i.e., a
single design model sits on top of the MDF, IDF, or AAO analysis model integration. In
other words, Cramer’s formulations explore how to integrate multiple analysis models for
a single MDO design statement. Note the difference between the target cascading and the
analysis models for a single level are integrated under a design problem introduced to
achieve the overall design goal as a decision authority sitting on top of analysis models. In
the TC, the modeling hierarchy consists of multiple levels of analysis and design models
representing a multilevel decision making hierarchy. Thus the MDF, IDF, or AAO
approaches can be applied for each of the multiple design problems in the TC modeling
hierarchy.
problem requires an integrated set of analysis models and an optimizer (Figure 4). For a
14
given set of design variables x, the analysis returns the values of the constraints g, h and
within the integrated set of analyses as part of overall function evaluation. It is not
practical, and sometimes it is impossible, to have a single integrated analysis model for
MDO. For example, aeroelastic wing analysis includes structural analysis simulation and
fluid dynamics simulation, which are usually both stand-alone simulation programs and
Optimizer
Minimize x f x
subject to
gx d 0
h x = 0
x f g h
Integrated
Analysis
Among the MDF, IDF, and AAO approaches, the MDF approach is most similar to
the traditional integrated analysis approach in that feasibility among different analysis
models is tightly maintained during the optimization, although the analysis models
themselves are not integrated into a single simulation. Figure 5 illustrates the schematic of
the MDF method. Note that multidisciplinary analysis (MDA) is achieved in MDF. Given
a vector of design variables xD, MDA returns responses from discipline A and B, RA and
RB. The interdisciplinary design variable vector xAB denotes input variables to discipline B
from discipline A. Likewise, xBA denotes input variables to discipline A from discipline B.
From the optimizer’s perspective, the coupled system of analyses disciplines is treated as a
single “black box” returning responses for the input design variables to evaluate design
15
objective and constraints. Whenever the design variable vector is provided to evaluate the
the optimum, the MDF approach can expend computational effort to enforce feasibility
among disciplines that is not necessary. However, this effort may not be wasted if the
Optimizer
xD RA, RB
Discipline Discipline
A B
xAB xBA
At the other extreme, the AAO approach does not guarantee feasibility for the
analysis models in any sense (individual discipline, multidisciplinary or even for a single
schematic of the AAO approach. Note that each discipline returns the residual of the
analysis model, WA and WB, rather than solving some sets of equations. The optimizer
provides analysis models with a vector of inputs and outputs of the analysis models, and
the analysis models calculate discrepancies (residuals) for the estimated inputs and
outputs. The computational burden to enforce feasibility for each iteration is relieved;
however, the number of constraints from the residual calculations and the additional
Discipline Discipline
A B
WA WB
position on a “feasibility spectrum” where the AAO and MDF approaches represent
extremes. The IDF approach maintains individual discipline feasibility, while the
given design variable xD and interdisciplinary variable xBA from the optimizer and returns
response RA and interdisciplinary variable xAB related to discipline B. Note that individual
discipline feasibility is maintained because disciplinary analyses take inputs from the
17
Optimizer
Discipline Discipline
A B
xAB, RA xBA, RB
allowed without affecting designs in other disciplines. Design variables are decomposed at
two different levels, system and subsystem, and system level variables are treated as
problem is to minimize constraint violations for its own set of design variables xj (j
denotes subsystem) while setting system-level design variables as constant. At the system
level, the original design objective is minimized with respect to system-level design
variables that are parameters in the subsystem. For example, a three-beam structure is
broken into individual beams either by the substructuring approach or by the finite
element method [34]. Each beam has its own type of cross section with detailed
dimensions under stress and buckling constraints. Given the cross sectional area of a beam
as a parameter, the subsystem beam design problem is to minimize stress and buckling
18
constraint violations by finding the detailed design of the cross section. At the system
level, the original design objective is minimized for the overall weight of the structure
under constraints for the overall assembled structure, such as the assembled structure
Minimize fx
x
subject to
(1)
g i x d 0 i = 1 } N sys
g j x d 0 j = 1 } N subsys
lower upper
x dxdx ,
where the constraints comprise the system level constraints gi and the subsystem level
constraints gj. The design variable x also comprises the system and subsystem level
Minimize C k x jk ;x i
x ij
subject to
(2)
x i = r x jk
lower upper
x jk d x jk d x jk ,
x i = r x jk (2a)
are added to the subsystem to match the system level design xi (treated as parameter). Here
r denotes a function evaluation and Ck is defined as the cumulative constraint violation
function (see Sobieski et al. [73]). As the subsystem level design variables are mutually
Minimize f xi
xi
subject to
g p x i d 0 p = 1, }, N sys
(3)
C k x i x jk d TOL
lower upper
xi d xi d xi
lower upper
x jk d x jk d x jk ,
where i denotes the system level and j denotes the subsystem level. TOL is a suitable
tolerance parameter, and Ck(xi,xjk) is a constraint violation function defined at the
subsystem level. In the subsystem optimization problem, the system-level design variables
xi are treated as parameters, while in the system optimization problem, the subsystem
design variables xjk can be changed within bounds. Bounds for the subsystem variables are
There have been observations about this formulation that some of the constraints
conflict resulting in poor convergence (see Thareja et al. [83]). The equality constraints
Eq. (2a) and the variable bound constraints in Eq. (2) might not have an overlapping
feasible region. Later Sobieski proposed new formulations that minimize deviations
between the system variable and the response from the subsystem instead of minimizing
the constraint violation at the subsystem level (see Sobieski [75]). The new k subsystem
optimization problem is
Minimize C k x k x i
xk
subject to
(4)
g k x k xi d 0
x klower d x k d x kupper .
20
In Eq. (4), xi are the variables at the system level which are treated as parameters at the
subsystem level. Local constraints gk are satisfied at the optimum with a minimum
deviation for the system-level design variables. The subsystem objective function Ck(xk,xi)
subsystem problem) and the subsystem response. Compared to the original formulation,
the equality constraint in the subsystem problem Eq. (2a) is removed and the objective is
At the system level, the design problem in Eq. (3) is changed as follows:
Minimize f xi
xi
subject to
g p x i d 0 p = 1, }, N sys (5)
C k x i d TOL
lower upper
xi d xi d xi
In the subsystem problem, the design objective Ck is changed so that it is now a constraint
at the system level within a tolerance TOL, while the variable bound constraint for the
constraints such as those in Eq. (2a) do not appear directly in the subsystem optimization
problem, design consistency between system and subsystem is satisfied within a tolerance
TOL.
optimization (see Braun [11]) and some other follow-up research (see Renaud and
Gabriele [63][64], Thareja and Haftka [84], Balling and Sobieski [5][6], and Vanderplaats
et al. [86]).
However, Sobieski’s framework does not have a convergence proof and does not
consider common design variables in more than one design discipline. Design autonomy
is allowed in that design change in one subsystem does not affect other subsystems, but
21
only because subsystem-level design variable sets are assumed mutually disjoint. Another
difficulty with the approach comes from the need to obtain optimum sensitivity
information from the subsystems. In case there is a subsystem design change or an update
in sensitivity, the subsystem design problem has to be solved again to get the optimum
sensitivity. Also, at the optimum, the TOL value will become zero; however, how to find
the appropriate tolerance value is not clearly suggested and the system level design
problem might fail to find a Karush Kuhn Tucker (KKT) point. This will be covered in
deviations between design targets and responses for system and subsystem (see Braun
[11], Braun et al. [12][13][14]). In this section, the collaborative optimization framework
2.1.3.1 Framework
The MDO problem presumes that multidisciplinary analysis models are coupled
either by common inputs or by taking other disciplines’ output as input. In the case of
aerodynamic structural design, the flow field near the wing and the deformed shape of the
wing are due to changes in other disciplinary analyses’ output. Schematically, the coupling
22
x 1 y p 1 R1
Disciplinary analysis 1
x 2 y p 2 R2
Disciplinary analysis 2
Local design variables are x1, x2, the linking variable is y and the parameters are p1,
p2. The quantities R1, R2 are outputs (responses) from each analysis model and they are
distinct (i.e., taken as sets that are mutually disjoint). The difference between Sobieski’s
presence of the linking variable and the coupling of analysis models in collaborative
levels following the fidelity of constraints, while in CO, all analysis models are assumed
In CO, for each analysis model the design problem is formulated to minimize the
deviation between the current analysis model response and the design “target” passed
from the master problem. In the master problem, subsystem objectives are posed as
equality constraints and the original design objective is minimized. The problem
23
System Level Optimizer
Goal: Original design objective
subject to
interdisciplinary compatibility
constraints
subsystem level and none of the original design constraints are included at the system
level. Instead, additional deviation constraints that are objective functions at the
subsystem levels are included at the system level design problem. In other words, the
original design objective is minimized at the system level while the original design
constraints are satisfied at the subsystem level. At the optimum, the subspace design
objectives will have zero values (i.e., a fully interdisciplinary compatible design) and the
Minimize fx
x
subject to
g i x d 0 i = 1, }, m (6)
lower upper
xj d xj d xj j = 1, }, n
24
This problem is decomposed into system and subsystem-level design problems. First, the j
h j' h j''
2 2
Minimize c j x j, x j = ¦ x ij – z i + ¦ y ij – z i
x j x j i=1 i=1
subject to (7)
g i x j x j d 0 i = 1, }, m j
lower upper
x kj d x kj d x kj k = 1 } h j'
lower upper
x lj d x lj d x lj l = 1 } n j – h j '
The subspace design objective is to minimize the deviation between targets from system
and subsystem level design. Targets are set for interdisciplinary design input and output.
The local design variable is x and the interdisciplinary input variable is x. The design
objective cj is defined as the least squared sum of deviations for the interdisciplinary input
vector xj of dimension hj’ and target vector z, and the interdisciplinary output vector yj of
dimension h j’’ and target z. The original design constraints g are partitioned into mj
function subject to additional constraints that are objective functions in the subsystem.
Minimize f z
z
subject to (8)
cj z = 0 j = 1, }, N
lower upper
z dzdz
Constraint cj is defined as the subsystem-level objective function in Eq. (7) and a total of N
constraints are treated as equality constraints. In the subsystem problem z is a parameter
vector to be set as target, but in the system problem z is a variable vector while the
25
At the optimum the subsystem objectives will have zero values and the system-
design. The concept of keeping the original design objective at the system level and
decomposing the analysis constraints into multiple subsystems has a great benefit in that
design autonomy is allowed to follow the analysis models. However, the additional
nonlinear constraints increase complexity in the design space at the system level and have
posing the optimality conditions. This issue will be presented in more detail in the
following section.
In this section the additional deviation constraints introduced at the system level
are shown not to satisfy the constraint qualifications for gradient-based optimality,
resulting in a failure to find KKT points for the system level optimization.
The form of deviation constraint differentiates two instances of CO. The first form,
CO1, is a linear constraint directly matching the subsystem-level response with the
c1 = z1 – x1 = 0
(9)
c2 = z2 – x2 = 0
The second form, CO2, is defined as the least squared sum of deviations in Eq. (7).
For example, for system variables z1 and z2, subsystems 1 and 2 are assumed to return
values for x1 and x2. Thus Eq. (10) is transformed as follows.
2
c1 = z1 – x1 = 0
2
(10)
c2 = z2 – x2 = 0
For both CO1 and CO2, if subsystem response and system variable match, constraints will
have zero values. However, these two formulations behave differently in KKT conditions.
26
For the system level optimization problem in Eq. (8), a KKT point should satisfy
the stationarity conditions and nonnegativity condition for the multipliers associated with
q
f z + ¦ O i c i z = 0
i=0
(11)
O z 0,
where
§ ·T
f z = ¨ wf , wf , }, wf ¸ (12)
© w z1 w z2 w z p¹
§ wc i wc i wc i · T
c i z = ¨ , , }, ¸ (13)
© w z1 w z2 w z p¹
Lagrange multiplier vector of dimension q. The condition in Eq. (11) can be posed only at
regular points, namely points where the gradients of the active constraints are linearly
independent. Otherwise the conditions will not be valid, hence the regularity is a form of
constraint qualification (see Bazaraa [9], Dennis and Schnabel [20], Papalambros and
Wilde [58], Saigal [66]). For CO2, Eq. (10) would require the existence of Lagrange
multipliers O , O such that
f z + O 1 c 1 z + O 2 c 2 z = 0, (14)
where
wc 1
w z1 2 z1 – x1
c 1 z = = (15)
wc 1 0
w z2
27
wc 2
w z1 0
c 2 z = = (16)
wc 2 2 z2 – x2
w z2
At the optimum the system variable is equal to the subsystem response, i.e., z 1
= x 1 ,
z 2
= x 2 . Then
f z + O 1 c 1 z + O 2 c 2 z = f z = 0 (17)
Thus the KKT conditions cannot be satisfied unless the gradient of the objective function
vanishes at the optimum, which can only happen by coincidence. The system-level
Jacobian vanishes at all feasible points, and the constraint qualification for regularity fails
at all feasible points. Alexandrov [3] points out this failure of the CO2 formulation to find
To avoid the constraint qualification failure, Braun [11] proposed the CO1
can be resolved by the CO1 formulation; however, the deviation constraint for the
interdisciplinary variable can lead to an empty feasible region. For example, for the
system-level variable z1, two interdisciplinary variables x11, x12 can have two different
problems are
2
Minimize c 1 x 11 = x 11 – z 1
x 11 x 21
subject to
g 1 x 11 x d0
21
lower upper
x 11 d x 11 d x 11
lower upper
x 21 d x 21 d x 21
(18)
28
2
Minimize c 2 x 12 = x 12 – z 1
x 12 x 22
subject to
g 2 x 12 x 22 d 0
lower upper
x 12 d x 12 d x 12
lower upper
x 22 d x 22 d x 22
(19)
where xij denotes variable i at subsystem j. Here x11 and x12 are the same interdisciplinary
inputs for subsystems 1 and 2. For a fixed target value z1, the optimal values for x11 and
x12 can be different when either of the subproblem objectives has nonzero optimum value.
In other words, the subsystem finds an optimum with nonzero minimum deviation from
the target assigned at the system level within the feasible region, and the deviation
2
c 1 = z 1 – x 11 = 0
2 (20)
c 2 = z 1 – x 12 = 0,
or
c 1 = z 1 – x 11 = 0
c 2 = z 1 – x 12 = 0
(21)
If x 11 z x 12 as explained above, for both CO1 and CO2, c1 and c2 cannot be satisfied at
the same time, i.e., the feasible region at the system level is empty. This empty feasible
region is not peculiar to this example. For any interdisciplinary input or output, if any of
the two subsystem problems returns different optima for the same variable, the system
level constraints will have empty feasible region resulting in a failure to find the optimum
The concept of minimizing the deviation between system and subsystem quantities
individual discipline, the task is to reach a design consistent with other disciplines, at the
same time designing the system element subject to its own constraints. This approach is
Another important idea is to decompose the design tasks so that not only the
design constraints but also the original design objective are decomposed into the different
Some of the coordination strategies with convergence proofs are presented in this
section. Wismer and Chattergy’s coordination [90] for decomposed problems has global
sequentially decomposed programming [53] has both global and local convergence
proofs. Some of the coordination ideas from these methods are utilized in a target
coordination [50][59] is also presented as it has convergence proofs for convex problems.
N
Maximize fi xi
¦
xi i=1
(22)
subject to
g i x 1 x 2 } x N t 0 i = 1 2 } N
where the objective function is decomposed into N components, and N sets of constraints
are posed (Figure 10). Here the objective function is dependent only on local variable of
30
the partitioned problem, while the constraints are dependent on local and linking
N N
T
L x P = ¦ fi xi + ¦ P i g i x 1 x 2 } x N (23)
i=1 i=1
By defining additional variable si, constraints gi can be transformed to be dependent on the
s i = x i i = 1 2 } N (24)
N
T T
L x O P s = ¦ f i x i + P i g i x i s 1 s 2 } s j } s N + Oi xi – si (25)
jzi
i=1
where O i = 0 if g i x 1 x 2 } x N is not a function of xi for all j z i .
System Optimize
Plant
Process
1
Process
2
... Process
N
31
As the objective and constraints are independent of the other subsystems, the
Lagrangian can be decomposed in two different ways: a feasible decomposition and a dual
treated as fixed parameters at the subsystem level; and in the dual feasible decomposition,
the Lagrange multipliers Oi are treated as parameters. The details follow in the next
sections.
FEASIBLE DECOMPOSITION
L i x i O i P i ;s i =
T
f i x i + P i g i x i ; s 1 s 2 } s j } s N
T
+ O i x i – s i i = 1 } N (26)
jzi
where si are the fixed parameters (separated by semi-colon) for the ith subsystem. The KKT
stationary point for each subproblem can be determined independently as a function of si.
Once all the KKT points are found at the subsystem level, the next step is to find optimal
values for the parameters si. Details of the algorithm are as follows:
1. Start with initial values for s1, ..., sN and set the iteration counter k = 0 .
k
2. Minimize L i x i Oi P i ;s i for i = 1 } N .
3. Update s ik + 1 .
k+1 k
4. If s i – s i d H , then stop. Else go to step 2 and increase k by 1.
The term “feasible” for this method comes from Eq. (24) that are always satisfied.
Interdisciplinary variables are duplicated at the subsystem level because the equality
constraints are imposed for them. Thus even if the optimization process is terminated
before complete convergence, the resulting solution is feasible throughout the disciplines.
constant at the subsystem level because of the additional equality constraints in Eq. (24)
32
and they are updated at the system level. Another way to achieve interdisciplinary
consistency is to fix the Lagrange multipliers associated with the additional constraints at
the subsystem level and keep updating them as the iteration proceeds at the system level.
As the system level problem is the dual of the subsystem level problem, minimization
problem at the subsystem level changes to dual maximization problem with respect to
s ij = x j i , j = 1 2 } N izj (27)
where sij indicates the presence of xj in the ith subsystem. Note that the individual
subsystem i has its own additional variables sij introduced by Eq. (27) for the original
L x O P s =
N N (28)
T T
¦ f i x i + P i g i x i s i1 s i2 } s ij } s iN + ¦ O ij x i – s ij
jzi j=1
i=1
where Oij = 0 if g i x 1 x 2 } x N is not a function of xj. The subsystem Lagrangians
become
L i x i P i s ij ;Oij =
N
T
f i x i + P i g i x i s i1 s i2 } s ij } s iN
T
O ij x i – s ij (29)
jzi
+ ¦
j=1
where the Lagrange multipliers Oij are treated as constants and stationary points with
respect to xi, sij, Pi can be determined as functions of Oij by the KKT conditions. Once
subsystem variables are determined, then the system-level optimizer determines optimal
33
k+1
3. Update O ij .
k+1 k
4. If O ij – O ij d H , then stop. Else go to step 2 and increase k by 1.
Wismer and Chattergy’s feasible and dual feasible methods maintain a descent
property for the constrained minimization problem that is critical to global convergence of
step algorithm to solve large scale problems by decomposition into subproblems and
follows:
1. If not near the solution, hold the linking variables constant and find a minimizer
the constraints and all of the variables in the original problem, find a satisfac-
Here linking variables are the variables common in more than one subproblems. Two
significant features of the SDP distinguish it from the other hierarchical coordination
methods: (1) In the master problem, sensitivities are not used. Instead, the next design
point is found by solving an approximate problem of the original problem including all of
the design variables and constraints. This approximate problem acts as the master
problem. (2) When near the solution, decomposition is not performed. Specific guidelines
for determining when one is “near the solution” are given in [53].
The SDP concept applied to the sequential quadratic programming (SQP) and
Yuan’s trust region (TR) algorithm [93] can be shown to have global and local
convergence proofs. As decomposition is not applied near the solution, the convergence
34
properties of SQP and TR methods can be utilized for SDP. Although it is advantageous to
be able to maintain the same convergence characteristics of SQP and TR in SDP, the size
overlapping coordination (HOC) (Macko and Haimes[46], Shima and Haimes[68], Park et
al. [59], Michelena et al. [50]). HOC maintains hierarchical coordination in a different
way in that two or more distinct decompositions serve to coordinate each other’s solutions
instead of using a master problem. Thus, there is no additional master problem defined in
subproblems.
2. Fix the linking variables yE to their values determined in step 1 and solve
Convergence is proven for convex problems with linear and nonlinear constraints. HOC
A merit function with penalty parameters is defined to have descent property during
iterations.
35
In the previous sections, a hierarchical design framework and associated
coordination methods for solving such hierarchical problems were reviewed. This design
not clear how each of these methods can be fitted in a systems engineering strategy.
Sobieski’s framework and collaborative optimization do not have convergence proofs and
are currently limited to bi-level systems, although multilevel extensions are possible.
different subproblems, such as common design inputs and outputs, are not clearly
engineering process. Motivation comes from the question: How can the design targets be
subsystem design targets so that a concurrent detailed design process for subsystems can
follow and lead to the final design? Target cascading will function at the early stages of
design problems at different levels, and design and analysis models are distinctly
for finding the optimal solution of problems posed within this framework and convergence
36
CHAPTER 3
TARGET CASCADING
FORMULATION
multilevel modeling hierarchy composed of analysis models and design models is defined
identifying the required hierarchical information flow among the various models. The
Target cascading can handle multi-level modeling hierarchies. In this section the
distinction between design and analysis models and their interactions are presented.
Supersystem
37
Table 3.1: Nomenclature
Symbol Description
Po original design problem
Psuper supersystem level target cascading optimization problem
Psys system level target cascading optimization problem
Psub subsystem level target cascading optimization problem
RL target values of R from a lower level
RU target values of R from an upper level
R responses computed by analysis models
T design targets
f objective for the original design problem
g inequality constraints for the original design problem
h equality constraints for the original design problem
r response function
x vector of all design variables ( x̃ , y)
x̃ local design variables
xmin lower bound of x
xmax upper bound of x
y design linking variables
yL target values of y from a lower level
yU target values of y from an upper level
HR target deviation tolerance for responses
Hy target deviation tolerance for linking variables
c component level
mi number of inequality constraints
me number of equality constraints
n number of design variables
sub subsystem level
super supersystem level
sys system level
values of a large number of design variables for the vehicle (supersystem), systems,
subsystems, and components that result in a final product that meets customer, corporate
and regulatory requirements. In this dissertation, top level model in the modeling
38
hierarchy is defined as the supersystem (or vehicle in the vehicle design). Under the
supersystem, system, subsystem, and component models are also defined in the decision
requirements can be translated into quantifiable vehicle design targets, such as fuel
The original design problem P0 can be stated as follows: find a design that
minimizes the deviations between the overall design targets and responses while satisfying
all constraints. Alternatively, determine the values of supersystem, system, subsystem and
P 0 : Minimize T–R
x
where
R = rx
subject to
(30)
gi x d 0 i = 1 } m i
hj x = 0 j = 1 } m e
min max
xk d xk d xk k = 1 } n .
The objective is defined as the discrepancy between the target T and the response
R obtained from the analysis models r(x); g and h are inequality and equality design
constraint vectors, and the design variable x is defined within lower and upper bounds, xmin
and xmax.
In theory, given the models for supersystem, systems, subsystems and components,
formulating and solving the above design problem as a single optimization problem is
possible using classical optimization techniques. However, this single problem approach
39
is often impractical and even computationally impossible. An alternative is to solve the
cascading process. The target cascading problem can be stated then as follows: given a set
of supersystem level targets, and models for all design elements (namely, supersystem, and
overall design problem, satisfying feasibility of element designs, and achieving top
supersystem targets.
will have a multilevel hierarchical structure. Two types of models exist in a modeling
hierarchy of the target cascading process: the optimal design model P and the analysis
model r (Figure 12). Figure 13 shows interactions between analysis models and design
problems at the system level. Optimal design models call analysis models to evaluate
supersystem, system and subsystem responses. Thus, analysis models take design
variables and parameters, as well as lower level responses, and return responses to upper
level design problems. A response is defined as an output from an analysis model, and a
linking variable is defined as a common design variable among the same level design
cascading, the design model is not limited to calling only one analysis model. For
analysis model called from an individual subsystem design model with no analysis models
In the early stages of the design process, targets must be set before detailed
embodiment design takes place. Hence, the analysis models in Figures 12 and 13 should
be simple yet descriptive of the relations between upper level responses and lower level
40
responses and design variables. These models can be low-degree of freedom analysis
models, response surface models from design of experiments, approximate models based
Analysis Model
P
super
Supersystem
Level (super)
r
super ... r
super
P sys
... P sys
System
Level (sys)
r sys r r sys r sys
sys
r sub rsub r
sub
r
sub ... r sub r sub r sub r sub
Supersystem
Level (super) U U L L
R sys y sys R sys y sys
System
P
Level (sys) sys
U L U L
y sub 1 y sub 1 y sub 2 y sub 2
Subsystem
Level (sub) P sub1 P sub 2
41
3.1.3 Interactions between Analysis Model and Design Model
As mentioned, analysis models are called by design models as shown in Figure 12.
The system design problem Psys calls an analysis model r sys by providing system design
variables (local and linking), system design parameters, and subsystem responses as input,
design problems in the previous section. For example, in Figure 14, systems sys1 and sys2
can share the same design variables ysys that are input to analysis models r sys1 , and r sys2 ,
i.e., they share system linking variables. System design problems are most likely to find
different values for the system linking variable, and so the supersystem design problem
Psuper must coordinate these different system linking variables until they match.
P
super
Supersystem
Level (super)
r U U L L U U L L
super R sys y
1 sys R sys y R sys 2 y R sys y
1 sys sys 2 sys
P sys1 P sys2
System
Level (sys) x̃ sys1 y R sys1 x̃ sys 2 y sys
sys R sys
2
r sys1 r sys2
Figure 14. Interaction of linking variables: a special case with a supersystem and two
systems. For each design problem, one analysis model is assigned.
In case more than one analysis models are called by the design model,
Section 2.1.1. Analysis models can be integrated in one of three ways: MDF
42
Once). In the context of linking variables, when more than one analysis models exist,
common input or output from one discipline that is input to other discipline can be
variable and a design linking variable is that a design linking variable is common to the
same-level design problems, while an analysis linking variable is common to the analysis
models called by the same design problem. In the TC framework, design linking variables
are coordinated through formulation and coordination, and analysis linking variables are
handled by one of the three ways (MDF, IDF, AAO) defined by Cramer et al [18].
A target is defined as a design goal to be achieved that is either cascaded from the
upper level design problem or passed up from the lower level design problem. A target
cascaded from the upper level is easily understood in the context of top-down systems
engineering. On the other hand, a target cascaded from the lower level design problem is
the lower level. Thus, design targets are included not only in the objective function but
In Figure 13, targets for system responses and system linking variables R Usys and
U
y sys are cascaded down from the supersystem level. After solving the system-level design
L
problem Psys, target values for system responses and system linking variables R sys and
L
y sys are passed up to the supersystem level. Likewise, for subsystem 1, R Usub1 and ysub1
U
are cascaded down as targets from the system-level design problem, whereas R Lsub1 and
L
y sub1 are returned to the system level. Responses R sub1 from subsystem 1, local design
variables, x̃ sys1 , and linking variables, y sys1 , are input to the analysis model r sys1 ,
departments and makes final design decisions for optimal design of the overall system.
43
This top level is denoted as the supersystem and a schematic diagram of the information
exchange between the top supersystem and several systems is shown in Figure 15.
Interactions between supersystem and system design problems occur only through system
responses and system linking variables. The supersystem design model collects this
system-level information and calls its own analysis model to generate the supersystem
responses. Then, based on target values for system responses and system linking variables,
the supersystem optimization problem finds the optimal design, one with the minimum
deviation from the lower-level design and with the minimum deviation from the overall
Overall Target
Supersystem Level
Supersystem Level
Target Cascading
Design Model
Supersystem Supersystem
Response Response
System Response
System Linking Variable
At the top level of the design hierarchy the problem is stated as follows: minimize
P super : Minimize
R super – T super + H R + H y
x̃ super y sys R sys H R H y
where
R super = r super R sys x̃ super
(31)
subject to
L L
R sys – R sys d H R y sys – y sys d H y
The original objective that minimizes deviations between design targets Tsuper and
coordinate values of the responses from the system Rsys, and the system linking variables
ysys. If there are p system level problems, then R sys = R sys1 } R sysp and
R sysi R sysj = for i z j . System linking variables are defined as common design
where < is essentially some type of norm, and y Lsysi is a system linking variable
calculated at the system optimal design problem i. One instance of < can be
L 1 L L L
y sys – y sys { --- y sys – y sys1 + } + y sys – y sysi + } + y sys – y sysp . (33)
p
At convergence, the deviation tolerances become zero as the system linking variables
converge to the same values for the different systems. The values of the system responses
L L
are matched with R sys , where R sys is the target response calculated at the system
optimal design problem. While system linking variables are common to system-level
design problems, system analysis linking variables are interdisciplinary variables for
45
different analysis models, i.e., common to analysis models. Finally, g super and h super
are inequality and equality design constraints at the supersystem level and are subsets of
the original constraints g and h. The norm used for the deviation constraints will be
problems, the system-level design problem is in the middle of the modeling hierarchy
capturing all of the general features in the TC formulation. In the case of the supersystem
design, there exist only one design problem at the top level. At the system level each
design department can be regarded as one system design model. Individual design
decisions are made within each design model and some of these design decisions can
affect other design models, i.e., other design departments, via the linking variables. Thus,
models: single design model at the supersystem level, multiple design models sharing
linking variables at the system level. The interactions between system level design and
analysis models and subsystem level design models are shown schematically in Figure 16.
U U
P sys : Minimize R sys – R sys + y sys – y sys + H + H y
R
subject to
(34)
L L
R sub – R sub d H R y sub – y sub d H y
Similarly at the system level, the optimal design problem is stated as in Eq. (34):
minimize the deviations for system responses and system linking variables, subject to
46
system design constraints and deviation constraints that coordinate subsystem responses
System System
Response Response
Subsystem Response
Subsystem Linking Variable
Figure 16. Schematics of information exchange between system design and analysis
model and subsystem design model
The objective function minimizes the discrepancy between the current system-
U
level responses Rsys and the targets set at the upper (supersystem) level R sys , as well as
U
between system linking variables ysys and the targets set at the supersystem level y sys .
U
Here R sys and y Usys are determined by solving the supersystem design problem in Eq.
(31). Target deviation tolerances are minimized to achieve consistent design with
minimum discrepancies between the subsystem level responses R sub and the target
L
responses R sub from the subsystem design problem, as well as between the subsystem
level linking variables y sub and the target values y Lsub from the subsystem design
47
problem. Since the system level is located in the middle of the overall hierarchy, this
response targets from the lower level (superscript L), and response targets from the upper
Figure 12. As the design problem calls for analysis models to retrieve responses, the
subsystem design model calls for subsystem analysis models that take subsystem level
design variables and parameters as input and return subsystem responses as output.
Constraints to coordinate linking variables from below are not necessary, since the
subsystem design problem is at the bottom of the modeling hierarchy. Instead subsystem
linking variables common to subsystem problems are coordinated at the system level
The subsystem design problem is stated in Eq. (35): minimize the deviations for
subsystem responses and subsystem level linking variables subject to subsystem design
constraints. Formally,
P sub : Minimize
U U
R sub – R sub + y sub – y sub
x̃ sub y sub
48
Subsystem Response Target
Subsystem Linking Variable Target Subsystem Design
Linking Variables
Subsystem Subsystem
Response Response
Subsystem Subsystem
Analysis Model Analysis Model
Figure 17. Schematic of information exchange between subsystem design model and
analysis model: the subsystem analysis model takes subsystem level design variables
and parameters and returns subsystem responses.
At the bottom of the modeling hierarchy, the subsystem design variables are input
to analysis models r sub returning responses to the subsystem level as output. In Eq. (35),
the objective is to minimize the deviations between the subsystem responses R sub and the
U
targets set from the system level R sub , as well as between the subsystem linking variables
U
y sub and the targets from the system level y sub . Target deviation tolerance constraints
are not introduced in Eq. (35) because there are no lower level design models that need to
be coordinated.
The modeling structure presented for the three-level hierarchy in Figure 12 can be
extended to a general multilevel hierarchical structure. For the j partitioned element at the
i level, the design problem is given targets for the i level response and the i level linking
variable for the objective function. Based on the (i+1) level design in the form of
constraint and analysis model evaluation from the i level analysis model, the TC design
49
problem finds the design with the minimum deviation from the targets from above and
from below. Also, the design linking variables from different (i+1) level design problems
Problem Pij for the j partition element at the i level is defined as follows: minimize
the deviations between current level responses R ij and linking variables y ij from one
U U
level below and the targets R ij and y ij , subject to design constraints, and tolerance
constraints that coordinate responses and linking variables from one level below.
Level i Level i
Level i
Target Cascading Target Cascading
Design Model Design Model
Level i Level i
Response Response
Figure 18. Schematics of information exchange between level i design and analysis
model and level (i+1) design model
50
Formally,
U U
P ij : Minimize R ij – R ij + y ij – y ij + H R + H
y
where R = r R
ij ij i + 1 j x̃ ij y ij
subject to (36)
L L
R i + 1 – R i + 1 d H R y i + 1 – y i + 1 d H y
g ij R ij x̃ ij y ij d 0
h ij R ij x̃ ij y ij = 0
This concludes our discussion of the formal statement of the problem. The next
level design problem to coordinate deviations between system and subsystem do not
system level problem in the collaborative optimization framework fail to satisfy the
regularity constraint qualification. It is important to check this issue for the additional
51
constraints introduced in the TC framework. For constraints gij, hij regularity of all
feasible points is assumed and so the gradient vectors for all active constraints are
The general TC design problem formulation in Eq. (36) includes the deviation
L
R i + 1 – R i + 1 d H R
L
y i + 1 – y i + 1 d H y (37)
Deviations can be defined as the least square norm. From the definitions of response and
2
R i + 1 – R i + 1 = §© R i + 1 – R i + 1 ·¹
L L
2 2 (38)
y i + 1 – y i + 1 = --- § §© y i + 1 – y i + 1 1·¹ + } + §© y i + 1 – y i + 1 p·¹ ·
L 1 L L
p © ¹
Here p is the number of lower level design problems that return linking variable values to
2
§R L ·
c1 : © i + 1 – R i + 1 ¹ d H R
1§§ (39)
--- © y i + 1 – y i + 1 1·¹ + } + §© y i + 1 – y i + 1 p·¹ · d H y
L 2 L 2
c2 :
p © ¹
The response and linking variables are assumed to be scalars for simplicity in presenting
the next argument. The constraints in Eq. (39) will be active and the bounds HR, Hy are
included in the scalar objective to be minimized.
52
wc 1
w R i + 1
wc 1
2 §© R i + 1 – R i + 1 ·¹
L
T
w y i + 1
c 1 = = 0 (40)
wc 1 –1
w HR 0
wc 1
wHy
wc 2
w R i + 1
wc 2 0
T
w y i + 1 y
1
– ------ y
L L
+ } + y i + 1 p
c 2 = = i + 1 2p i + 1 1 (41)
wc 2 0
w HR –1
wc 2
w Hy
It is immediately evident that c 1 and c 2 are linearly independent even when complete
consistency occurs for the lower level responses and linking variables, resulting in zero
exceedingly simple, it captures the distinct characteristics of CO and TC. For the CO
53
2
Minimize x
1
x1
subject to (42)
0 d x1 d 1
Each variable bound constraint can be regarded as a disciplinary analysis model and
partitioning within the CO framework generates three design problems: one system level
U 2
P sub1 : Minimize x 1 – x 1
x1
(43)
subject to
0 d x1
U 2
P sub 2 : Minimize x 1 – x 1
x1
(44)
subject to
x1 d 1
The CO formulation is then stated with two subsystem objectives as equality constraints in
P sys : Minimize x 2
1
x1
subject to (45)
2
c 1 = §© x 1 – x 1 1·¹ = 0
L
2
c 2 = §© x 1 – x 1 2·¹ = 0
L
As an aside, note that, compared to the original problem in Eq. (42), two additional
constraints are added in the system level problem. Indeed, an MDO approach is not
54
increased problem size by additional constraints and design variables. However, in most
cases, an MDO decomposition approach may be the only alternative for solving multi-
L L L L
CO2 Case I: x 1 1 z x 1 2 , 0 x 1 1 1 , 0 x 1 2 1
In this case, the subsystem level responses are different and are set as constants in
the system level deviation constraints. This case accounts for most common situations in
design practice, in that different design teams working independently return different
values for the common parts within the feasible domain. As deviation constraints are
simultaneously.
2
§x – x 1 1·¹ = 0
L
© 1
§x
2 (46)
– x 1 2·¹ = 0
L
© 1
In other words, by introducing additional deviation constraints, the system level design
The subsystem problems return exactly the same value for the common design
within the feasible region, i.e., c 1 = c 2 . Subsystem-level optima are found inside the
feasible domain, i.e., there are no active variable bound constraints. The KKT stationarity
f + O 1 c 1 + O2 c 2 = 0, (47)
L L
multiplier. A feasible point x 1 = x 1 1 = x 1 2 is not a regular point because the
constraints c 1 c 2 are co-linear. At the feasible point, the left hand side in Eq. (47)
becomes
55
T T T T
f + O 1 c 1 + O 2 c 2 = f z 0 . (48)
TC Case: 0 d x 1L 1 d 1 , 0 d x 1L 2 d 1
The optimization problem in Eq. (42) can be decomposed into system and
subsystems following the TC formulation. Design problems for system and subsystem in
TC are similar to CO design problems. The subsystem design problems are the same as
those in CO, Eq. (43) and Eq. (44). Original variable bound constraints are decomposed
The system level TC design problem is stated introducing the additional deviation
P sys : Minimize x 2 + H
1
x 1 H
subject to
2 2 (49)
c 1 = §© x 1 – x 1 1·¹ + §© x 1 – x 1 2·¹ – H d 0
L L
c2 = –x 1 d 0
c3 = x1 – 1 d 0
and returning subsystem responses. The system design problem must satisfy the KKT
stationarity condition
T T T T
f + O1 c 1 + O 2 c 2 + O3 c 3 = 0, (50)
where
56
wf
T w x1 2x
f = = 1 (51)
wf 1
wH
wc 1
T w x1 L L
2 x 1 – x 1 1 + 2 x 1 – x 1 2
c 1 = = (52)
wc 1 –1
wH
wc 2
w x1
= –1
T
c 2 = (53)
wc 2 0
wH
wc 3
w x1
= 1
T
c 3 = (54)
0
wc 3
wH
L L
2x
1 + O 2 x 1 – x 1 1 + 2 x 1 – x 1 2 + O – 1 + O 1 = 0 . (55)
1 2 0 3 0 0
1 –1
Depending on the constraint activity, three cases should be considered: (1) interior
optimum, i.e., neither constraint c 2 nor constraint c 3 is active, (2) constraint c 2 is active,
multipliers O 2 , O 3 for constraints c 2 and c 3 become zero. From Eq. (55) O 1 = 1 and
--- x 1 1 + x 1 2 satisfy the KKT condition. The optimum x 1
is cascaded to the
x 1
= 1
L L
3
subsystem problems in Eq. (43) and Eq. (44). As the current optimum is an interior
optimum, variable bounds in the subproblems are satisfied and the target is perfectly
57
achieved at the subsystem level, i.e., subsystem solutions are x 1 1
= --1- x 1 1 + x 1 2
L L
3
1 L L
and x 1 2 = --- x 1 1 + x 1 2 . Now the system-level solution is updated with new
3
subsystem-level optima:
2 L
x 1
= 1
--- x 1 1
+ x 1 2
= --- x 1 1 + x 1 2 d 1--- x 1 1 + x 1 2 ,
L L L
(56)
3 9 3
which is less than the system-level solution obtained in the first iteration, i.e., as iteration
subsystem level and the subproblems in Eq. (43) and Eq. (44) will match the target value,
L
x 1 2 = 0 , updated system-level problem is stated as follows:
P sys : Minimize x 2 + H
1
x 1 H
subject to
2 2
(57)
c1 = x1 + x1 – H d 0
c2 = –x 1 d 0
c3 = x1 – 1 d 0
resulting in x 1 = 0 and H = 0 .
based on the target x 1 = 1 from the system level are x 1 1
= 1 and x 1 2
= 1 . Then
U
L
updated system-level problem based on the subsystem-level design x 1 1 = 1 and
L
x 1 2 = 1 is stated
58
P sys : Minimize x 2 + H
1
x 1 H
subject to
2 2
(58)
c1 = x1 – 1 + x1 – 1 – H d 0
c2 = –x 1 d 0
c3 = x1 – 1 d 0
2 2
with the optimal solution x 1
= --- and H = --- . Next iteration will follow the first case with
3 9
both constraints c 2 and c 3 inactive.
Although this is not a general case, it clearly shows the difference between CO and TC in
terms of the regularity condition of the constraints. The deviation limit H is always
positive if the Euclidean norm is used to define the deviation. Thus the nonnegativity
constraints for H are redundant and if included in the formulation, they will be dominated
by the deviation constraints, which are always active by the first Monotonicity Principle
The deviation constraints in the TC formulation are bounded by upper limits H that
are treated as additional design variables. The upper limits are included in the objective
have inherently the form of multiobjective functions. In this dissertation equal weights are
assigned for each term in the objective function. Rather than exploring the Pareto space
for multiple designs by assigning different weights on the objective functions, finding out
a design that meets targets serves the TC process. In this section, the H-inequality
in the TC formulation.
59
In a multiobjective optimization problem, one of the multiple objective functions
is selected as the primary objective function and optimized, while the other objective
functions are converted into constraints. Often engineering design problems are
Minimize ^ f x f x } f x `
1 2 m
x
n
fi R o R (59)
n
xR
where fi (i = 1, ..., m) is a scalar objective function and x is n-dimensional design variable.
one by converting the objectives, except the selected primary one, into constraints
(henceforth some referred to as “objective constraints”) with an Hbound, and solving the
resulting problem for multiple H values, updated with some heuristic rule.
Minimize f k x
x
subject to (60)
fi x d Hi i = 1 } m i z k.
Choo [79], Miettinen [51], Palli et al. [56]). For each constraint, an Hinterval is selected,
namely,
In the TC design problem additional constraints are included in the constraint set
as inequality constraints bounded by limits that are themselves design variables. These
deviation tolerance limits (HR, Hy) are also appended in the objective function effectively
forming a multiobjective model. The additional deviation constraints can be regarded as
60
transformed objectives (i.e., “objective constraints”). Thus the TC design objective is
stated as minimizing two deviations: a deviation for a target given from above (primary
objective) and a deviation for a target given from below (objective constraint). In the H-
inequality constraint method an interactive filtering is applied to the bounds to guide the
search towards a desirable Pareto optimum, i.e., the bounds are treated as parameters
variables, so the deviation constraints are always active by the first Monotonicity Principle
(Papalambros and Wilde [58]) and the transformed objective constraints are equivalently
objective and equal weights (i.e., no weights at all), this is done for simplicity of
weights. For example, the deviation terms can be assigned with higher weights, i.e.,
deviation limits are multiplied by higher weights to put more emphasis on consistent
while avoiding the limitations of Sobieskie’s framework for its not having linking
variables defined in the formulation and the failure to meet constraint qualifications in the
collaborative optimization. The simple example in the previous section illustrated and
confirmed these arguments. Starting from the supersystem design problem, the system,
different design problems at different levels in the form of responses and linking variables.
The major difference between the TC formulation and the CO formulation is the
way design problems are formulated in multiple levels. In CO, design problems are
formulated for individual disciplines and on top of the disciplinary design problems one
61
system level design problem coordinates deviations from disciplines. Besides CO’s failure
design problems added under the two-level hierarchy. These additional design problems
will be a burden in the coordination process if some design changes occur in the hierarchy.
In TC, design problems are formulated for individual design tasks at different levels and
deviation terms in the objectives and constraints. Thus additional coordination problems
part of the overall framework. The second major issue is how to solve the problem posed
in the TC hierarchy, i.e., how to define an effective coordination strategy. This is the topic
62
CHAPTER 4
COORDINATION STRATEGY FOR
TARGET CASCADING
The TC modeling hierarchy of design models for two levels (supersystem and
system) is given in Figure 19. The number of system design problems is p; Psuper denotes
the supersystem design problem; and Psys1, Psys2,..., Psysp denote system design problems.
Each of the design problems calls for analysis models to obtain responses for the current
design. Design problems for each level are stated in Eq. (62) and Eq. (63). System level
design problem in Eq. (63) is for the i decomposed problem. Note that the index i indicates
i partitioning.
In the system design problem, constraints for subsystem linking variables (see Eq.
(34)) are omitted with the deviation limits H y because there are no more design problems
below the system level. TC formulations use several terms defined in a norm. Unless
¦ xj
2
otherwise noted, the symbol x = designates the least square norm.
63
Targets
Supersystem
P super
Level (super)
System
Level (sys) P sys1 P sys2 ... P sysp
where
R super = r super R sys x̃ super
subject to (62)
L L
R sys – R sys d H R y sys – y sys d H y
P sysi : U U
Minimize R sysi – R sysi + y sysi – y sysi + H y
64
The limit vectors for deviations, H R , H y , are treated as additional optimization variables,
so the deviation constraints become always active. The deviation constraints in Eq. (62)
and Eq. (63) can be transferred to the objective functions and the variable bounds can be
L
P super : Minimize R
super
–T
super
+ R sys – R sys + y sys – y Lsys
x̃ super y sys R sys
subject to
g super R super x̃ super d 0
h super R super x̃super = 0
U U
P sysi : Minimize R sysi – R sysi + y sysi – y sysi
x̃ sysi y sysi
subject to (65)
g sysi §© R sysi x̃ y sysi· d 0
¹
sysi
h sysi §© R sysi x̃ y · d0
sysi sysi¹
Note that system linking variables y sys are included in both supersystem and
system problems as variables, and the system responses R sysi appear in both problems,
65
4.2 Hierarchical Overlapping Coordination
Minimize f x
x
subject to (66)
gx d 0
hx = 0
n n n
where x is the vector of optimization variables, f: o and g i : o are
n
convex and differentiable, and h i : o are affine functions.
Boolean matrix, the function dependence table (FDT). The (i, j) entry of the FDT is one if
the i function depend on the j variable, and zero otherwise. By reordering the variables and
constraints in the original problem Eq. (66), the FDT of D-decomposition problem is
yD xD xD ... xD
2 pD
1
gD
1 B
D1 A
hD D1
1
gD
2 B A
hD D2 D2
2
...
...
...
gD
pD B A
hD Dp Dp
pD D D
Figure 20. Block D-decomposition of the functional dependence table (adapted from
Michelena et al [50])
66
In Figure 20, x D is the vector of local variables associated with block A D for
i i
subproblem D i ; y D is the vector of n D linking variables for the D-decomposition; and p D
denotes the number of subproblems in the D-decomposition. Assuming that the objective
function f is D-additive separable (in general, HOC can be used if the objective function
can be written as monotonic functions of local objective functions derived from the D and
E decompositions), the original optimization problem in Eq. (66) takes the following
form:
pD
Minimize f D y D +
0 ¦ f D y D x D
i i
x i=1
subject to (67)
g D y D x D d 0 i = 1 } p D
i i
h D y D x D = 0 i = 1 } p D
i i
n
For a given vector d , fixing the D-linking variables y D = d D in Eq. (67) gives the
following Problem D, P D .
P D : For each i = 1 } p
D
Minimize f D d D x D
i i
xD
i
subject to (68)
g D d D x D d 0
i i
h D d D x D = 0.
i i
Similarly, Problem E, P E , can be defined and solved for Edecomposition after fixing the
E-linking variables. In other words, for the same problem, two decomposition methods are
applied and the problems are solved fixing its own linking variables.
composed of a master problem and subproblems in bilevel fashion. Solving the master
67
The HOC maintains hierarchical coordination in a different way in that two or
more distinct decompositions for the same design problem serve to coordinate each
other’s solutions instead of using a master problem. Thus, there is no additional master
problem defined in HOC, but Problem D passes its optimal solution to Problem E and vice
1. Fix the linking variables y and solve problem Dby solving p independent sub-
D D
problems.
2. Fix the linking variables y to their values determined in step 1 and solve problem
E
ProblemD
Subproblem Subproblem
... Subproblem ...
1 pD
y D
linking variables
y E
Subproblem Subproblem
... Subproblem ...
1 pE
ProblemE
68
Solve Problem Dwith fixed
linking variables y . D
Yes
Check termination criteria. Stop
No
fD yD x – fD yD xD H (69)
i+1 i + 1 Di + 1 i i i
fE yE x – fE yE xE H (70)
i+1 i + 1 Ei + 1 i i i
(1) If the HOC algorithm started with a feasible point, then at each stage of the
(a) f x D t f x E t f x D
i i i + 1
min
(b) lim f x D = lim f x E = f
t f .
iof i iof i
69
½f ½f
(3) Any accumulation point x
of either ® x D ¾ or ® x E ¾ solves both
¯ i¿
i=1
¯ i¿
i=1
Problems D and Problems E.
the existence of feasible domain is assumed, thus the property (1) above is maintained.
The second property is a key element to proving the global convergence through a non-
ascent property of the algorithm (see Macko and Haimes [46], Shima and Haimes[68],
In the TC formulations given in Eq. (62) and Eq. (63) for the two-level case, the
system responses R sys and the system linking variables y sys obtained in the supersystem-
level problem are cascaded as fixed targets to the system level. At the system level, for
minimum deviation from the targets, optimal system responses and system linking
variables are obtained. These optimal values are fed back to the supersystem level in the
TC process is similar to the HOC in that variables in one design problem are fixed
in the other design problem and the process is repeated until convergence. In this section,
the analogy between the TC and the HOC is presented and some properties of the HOC
The HOC in essence solves the original design problem in Eq. (66) by using at
least two different decompositions, i.e., two different arrangements of the original design
vector x . In order to apply HOC properties to TC, the design problem to be solved should
be the same at each level, i.e., when solving the supersystem-level problem and the
First, to show the model equivalence between the supersystem and the system,
consider the following AAO formulation for the bilevel TC formulation by putting Eq.
70
super sys super sys
P AAO : Minimize f AAO = R super – Tsuper + R sys – R sys + y sys – y sys
subject to
g super R super x̃ super d 0, h super R super x̃ super = 0
where R sys = R sys1 } R sysi } R sysp . Although the system responses and system
linking variables are common to both problems, the terms in the objective including
R sys y sys are differentiated. The superscript U and L in the objective have been
substituted by super and sys, respectively. Although R sys y sys refer to the same entity in
the design problem, when the system-level and the subsystem-level problems are solved
sequentially, they can have different values during the iteration process and they are noted
as super super
Rsys y sys and Rsys sys
sys y sys , respectively.
System F F F V V V
Table 4.1 gives the information flow between the supersystem and the system. V
denotes variables and F denotes fixed parameters. Variables (V) at the supersystem level
become fixed (F) at the system level and vice versa. Note that the system responses and
the system linking variables R sys y sys become targets when they are fixed either at the
71
Eq. (71) the supersystem-level design problem and system-level design problem are
generated when fixing different variables as stated in Table 4.1. At the system level, the
objective should be identical with the supersystem objective to apply HOC. From the
AAO problem statement in Eq. (71), the supersystem objective and the system objective
should be
f AAO
= f super
= f sys + R super – T super (72)
super sys super sys
= R super – Tsuper + R sys – R sys + y sys – ysys .
The terms in the objective become either parameters or variables at different levels
following Table 4.1. The objective functions at the supersystem and system level are
constant and becomes irrelevant in the system objective function even if it is included in
the objective. Thus those terms and constraints irrelevant to each of the problems are
omitted resulting in the TC formulations for the supersystem and the system as follows.
subject to
g super R super x̃ super d 0, h super R super x̃ super = 0
sys sys
with respect to x̃ sys y sys R sys
sys § sys· (74)
sys © x̃ sys y sys ¹
where R sys = r
subject to
g sys § R sys x̃ sys y sys · d 0 , h sys §© R sys x̃ sys y sys ·¹ = 0
sys sys sys sys
© ¹
72
The above problems indicate that, by differentiating the supersystem-level
variables and system-level variables, the original AAO problem can be decomposed into
the system-level problem. Thus the same problem structure of the HOC can be applied to
L L
1. For fixed response and linking variable targets R sys y sys , solve supersystem prob-
lem.
U U
2. For fixed response and linking variable targets R sys y sys , solve p system problems.
L L
3. Go to step 1 with the fixed response and linking variable targets R sys y sys deter-
mined in step 2.
4. Repeat until convergence.
properties described in the previous section. The non-ascent property is critical to proving
global convergence and it is part of the original HOC algorithm, which can be also
applicable to the TC. This non-ascent property will be examined in the following section.
Here we assume that the feasible space for the subproblems is not empty. When
solving the individual design problems in the TC, any general convergent algorithm can
In the HOC algorithm, the decomposed design problem contains local and linking
variables. In the TC, there exist no local variables in decomposed problem as in Table 4.1.
73
This coupling between the supersystem-level problem and the system-level problem
might cause some convergence difficulty (Michelena et al [50], Park et al [59]), which
might need to be further investigated in the TC case. Utilizing the non-ascent proof of the
The supersystem problem and the system problem in the TC formulation can be
sys sys
where x super = x̃ super Rsuper
sys
super
y sys and x sys = x̃ sys, y sys , Rsys , and all the
Then the combined objective Eq. (75) maintains non-ascent property, i.e.,
f § x super x sys · t f§x x sys · t f § x super x sys · , where i is the
© i i – 1¹ © super i i¹ © i + 1 i¹
iteration index.
Proof
the objectives at the supersystem and the system can be obtained from the All-At-
Once objective f AAO as shown in Eq. (72). The supersystem objective is identical
with the AAO objective and the system objective is also identical with the AAO
74
4.5 Concluding Comments
The HOC has been proven to have a non-ascent property and this idea was applied
decomposition and the system-level problem was regarded as E-decomposition. Fixing the
design from one decomposition, finding out the optimal design of the current
decomposition and feeding back the current design as fixed parameters to the other
decomposition exactly follow the idea of the TC formulations. The optimal design from
the supersystem level is cascaded down to the system level as targets and the system-level
Comparison of the HOC and the TC shows that, in case of the TC, the two design
problems (the supersystem and system problems) are strongly coupled. It is because of
meeting the requirements of the HOC that the objective and the constraints of two
decomposed problems should be identical. Thus in the case of TC, all the variables in both
supersystem and system are coupled, i.e., variables in one level are fixed parameters in the
other. This strong coupling might affect the performance of the HOC, since the number of
the design variables is not significantly reduced.
Another important point is that the convergent solutions from the TC will likely
match the optimal solutions of the AAO formulations. The TC iterations will likely create
The study presented in this chapter is mainly focused on the two-level case of the
TC formulations. The two-level case can be extended to the multiple level cases still
coordination steps when the multiple levels are involved. This also needs to be further
investigated.
75
CHAPTER 5
DEMONSTRATION STUDIES
Case studies applying the target cascading (TC) formulation demonstrate the
and demonstrates that the TC decomposition gives the same solution as the original
problem. The second study is a TC application to vehicle chassis design and provides
sufficient breadth and depth, in that the problem is multilevel and contains all elements of
the TC formulation; namely, responses, local variables, and linking variables at multiple
levels.
globally optimal solution (see Wilde and Beightler [88] or Beightler and Phillips [10]).
The problem here is such a geometric programming problem and it serves to compare the
target cascading solution with the All-At-Once (AAO) solution. The problem is selected
having a known unique global solution relieves the solution strategy from the added
The original problem has a quadratic objective function, fourteen design variables,
four equality and six inequality constraints, and nonnegativity constraints for all design
variables as shown in Eq. (76). The equality constraints h1,..., h4 can be directed as active
76
inequality constraints in the negative unity form (see Papalambros and Wilde [58]). The
unique global minimum is shown in the AAO column of Table 5.2. The original problem
2 2
Minimize f = x1 + x2
x 1 x 2 } x 14
subject to
–2 2 2 –2 2 2
x3 + x4 x5 + x6 x8 + x9
2
-d1
g 1 : ------------------- -d1
g 2 : -------------------
2
g 3 : -----------------
2
-d1
x5 x7 x 11
–2 2 2 –2 2 2
x 8 + x 10 x 11 + x 12 x 11 + x 12
-d1
g 4 : ---------------------- -d1
g 5 : ----------------------- -d1
g 6 : ----------------------- (76)
2 2 2
x 11 x 13 x 14
2 2 –2 2 2 2 2 2
h1 : x = x3 + x4 + x5 h2 : x2 = x5 + x6 + x7
1
2 2 –2 –2 2 2 2 2 2 2
h 3 : x 3 = x 8 + x 9 + x 10 + x 11 h 4 : x 6 = x 11 + x 12 + x 13 + x 14
x 1 x 2 } x 14 t 0
2 2
Minimize f = x1 + x2
x 1 x 2 } x 14
where
2 –2 2 1e2
R 1 = x 1 = r 1 x 3 x 4 x 5 = x 3 + x 4 + x 5
2 2 2 1e2
R 2 = x 2 = r 2 x 5 x 6 x 7 = x 5 + x 6 + x 7
2 –2 –2 2 1e2
R 3 = x 3 = r 3 x 8 x 9 x 10 x 11 = x 8 + x 9 + x 10 + x 11
2 2 2 2 1e2
R 4 = x 6 = r 4 x 11 x 12 x 13 x 14 = x 11 + x 12 + x 13 + x 14 (77)
subject to
–2 2 2 –2 2 2
x3 + x4 x5 + x6 x8 + x9
2
-d1
g 1 : ------------------- -d1
g 2 : -------------------
2
g 3 : -----------------
2
-d1
x5 x7 x 11
–2 2 2 –2 2 2
x 8 + x 10 x 11 + x 12 x 11 + x 12
2
-d1
g 4 : ---------------------- -d1
g 5 : -----------------------
2
-d1
g 6 : -----------------------
2
x 11 x 13 x 14
x 1 x 2 } x 14 t 0
77
If we assume that x1, x2, x3, x6 are responses from analysis models h1, h2, h3, h4,
the equality constraints can be regarded as analysis models and the overall problem can be
The problem in Eq. (77) is partitioned as shown in Figure 23. Here the top level is
the system level (denoted as “s”) and the lower level is the subsystem level (denoted as
“ss”). The system-level design problem evaluates responses from the analysis models r1,
r2 to achieve overall targets for x1, x2. At the subsystem level, individual design problems
are formulated for each analysis model r3 and r4, and they share linking variable. Inputs to
the system analysis models are x3, x4, x5, x6, x7, while the analysis models return system
responses x1, x2. Inputs to subsystem analysis models are x8, x9, x10, x11, x12, x13, x14,
returning the subsystem responses x3, x6. The linking variable between the subsystem-
responses and variables is given in Table 5.1. Only the values of x3, x6 and x11 are passed
up and cascaded down between the system and subsystem optimal design problems. Top
level target values are set as zero, i.e. under a perfect design scenario, x1, x2 will have zero
values.
78
P : Minimize 2 2
s x +x +H +H +H
1 2 1 2 3
x 1 x 2 } x 7 x 11 H 1 H 2 H 3
subject to
L 2 L 2
§
x – x 11ss1· + § x 11 – x 11ss2· d H 1
© 11 ¹ © ¹
§ L 2 L 2
x – x · d H 2 § x 6 – x 6 · d H 3
© 3 3¹ © ¹
–2 2 2 –2
x +x x5 + x 6
3 4
g : --------------------- d 1 g 2 : --------------------- d 1 x 1 x 2 } x 7 x 11 t 0
1 2 2
x5 x7
2 –2 2 1e2 2 2 2 1e2
x 1 = r 1 x 3 x 4 x 5 = § x 3 + x 4 + x 5· x 2 = r 2 x 5 x 6 x 7 = § x 5 + x 6 + x 7·
© ¹ © ¹
P s s1 : Minimize
§ U 2 U 2 P s s2 : Minimize U 2 U 2
x
© 3
– x 3 ·¹ + §© x 11 – x 11·¹ §
x – x 6 ·¹ + §© x 11 – x 11·¹
x 3 x 8 x 9 x 10 x 11 x 6 x 11 x 12 x 13 x 14 © 6
x 11
subject to subject to
2 2 –2 2 2 –2 2 2
x8 + x9 x 8 + x 10 x 11 + x 12 x +x
g : ------------------ d 1 g 4 : ------------------------ d 1 11 12
3 2 2 g 5 : ------------------------ d 1 g 6 : ------------------------ d 1
x 11 x 11 2 2
x 13 x 14
x x 6 x 11 x 12 x 13 x 14 t 0
3 8 9 10 11 t 0
x x x x
2 –2 –2 2 1e2 2 2 2 2 1e2
x 3 = r 3 x 8 x 9 x 10 x 11 = § x 8 + x 9 + x 10 + x 11· x 6 = r 4 x 11 x 12 x 13 x 14 = § x 11 + x 12 + x 13 + x 14·
© ¹ © ¹
79
With two analysis models r1, r2, the system level design problem is stated as
follows.
Ps: Minimize 2 2
x1 + x2 + H + H2 + H
1 3
x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 11 H 1 H 2 H 3
where R = x = r x x x = x 2 + x – 2 + x 2
1e2
1 1 1 3 4 5 3 4 5
2 2 2 1e2
R 2 = x 2 = r 2 x 5 x 6 x 7 = x 5 + x 6 + x 7
subject to
–2 2 2 –2
x3 + x4 x5 + x6
- d 1
g 1 : ------------------- -d1
g 2 : -------------------
x5
2
x7
2 (78)
2 2
§x – x 11ss1·¹ + §© x 11 – x 11ss2·¹ d H 1
L L
© 11
§x L 2 L 2
© 3
– x 3·¹ d H 2 §x
© 6
– x 6·¹ d H 3
x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 11 t 0
The analysis models r1, r2 evaluate current design to return system responses x1, x2 for
which targets are set to be zero. Besides the original constraints g1, g2 and nonnegativity
constraints for the variables, three additional constraints are introduced: one constraint for
the subsystem linking variable x11, and two constraints for the subsystem responses x3, x6.
These new constraints impose deviation limits on the lower level designs. Three additional
design variables H, H, H are introduced for the deviation limits and are added to the
objective function. The parameters x L11ss1 x 11ss2
L
are linking variable values from
L L
subsystems 1 and 2, and x 3 x 6 are the subsystem response values.
The two subsystem design problems are formulated as follows. For each problem,
the design objective is to minimize the deviations between targets and responses or linking
variables. The subsystem 1 design problem obtains a response x3 from the analysis model
r3, and the subsystem 2 design problem obtains a response x6 from the analysis model r4.
Note that x11 is common to both design problems, i.e., it is a linking variable. The
80
deviation design objectives are formulated for these targets. Constraints g3, g4 and g5, g6
2 U ·2
P ss1 : Minimize § x – x U· + § x
© 3 3¹ © 11 – x 11¹
x 3 x 8 x 9 x 10 x 11
2 –2 –2 2 1e2
where R 3 = x 3 = r 3 x 8 x 9 x 10 x 11 = x 8 + x 9 + x 10 + x 11
subject to (79)
2 2 –2 2
x8 + x9 x 8 + x 10
2
- d 1
g 3 : ----------------- -d1
g 4 : ----------------------
2
x 11 x 11
x 3 x 8 x 9 x 10 x 11 t 0
2 U ·2
P ss2 : Minimize § x – x U· + § x
© 6 6¹ © 11 – x 11¹
x 6 x 11 x 12 x 13 x 14
2 2 2 2 1e2
where R 4 = x 6 = r 4 x 11 x 12 x 13 x 14 = x 11 + x 12 + x 13 + x 14
subject to (80)
2 –2 2 2
x 11 + x 12 x 11 + x 12
- d 1 g 6 : -----------------------
g 5 : -----------------------
2 2
-d1
x 13 x 14
Table 5.2 shows the optimization results obtained from the AAO and the target
cascading models. Objective, variables and tolerance values are given in the Table. The
solution from the AAO model is the unique global optimum for this posynomial
programming problem. The TC solution is the same within a tolerance. The TC solution
was obtained following top-down implementation fashion. After solving the system-level
design problem, targets for the subsystem problems were cascaded based on the optimal
design at the system level, and the subsystem optimal designs were passed back to the
system level after minimizing deviations between the responses and targets. This top-
81
down and bottom-up process completed one iteration loop in the TC process and it was
Ideally, the tolerances should be equal to zero. In current case study, if the
tolerance is tightened, it will converge to zero. This case study can represent a hierarchical
MDO case composed of two levels and two disciplines. Analysis models at the system and
subsystem can represent different degree of fidelity, i.e. r1, r2 represent low-fidelity
simple models and r3, r4 represent high-fidelity detailed models for two different
disciplines. Also, there is defined a linking variable common to subsystem level design
problems. Thus this example can serve as a validation case of hierarchical TC design
Table 5.2: Optimal designs from All-At-Once (AAO) and Target Cascading (TC)
models
AAO TC
x1 2.84 2.80
x2 3.09 3.03
x3 2.36 2.35
x4 0.76 0.76
x5 0.87 0.87
x6 2.81 2.79
x7 0.94 0.95
x8 0.97 0.97
x9 0.87 0.87
x10 0.80 0.80
x11 1.30 1.30
x12 0.84 0.84
x13 1.76 1.75
x14 1.55 1.54
f 17.61 17.02
H1 N/A 0.0001
H2 N/A 0.0017
H3 N/A 0.0029
82
5.2 Chassis Design for Ride and Handling Targets
A target cascading case study for chassis design of a typical sport-utility vehicle
(SUV) is performed in this section. The targets and models are selected for the purpose of
(MDO) strategy and as a rigorous methodology for product development. The analysis
models and targets are not unrealistic and are appropriate for our present purpose.
Although the actual vehicle design process is simplified here, the primary goal in
constructing the study is to represent a TC modeling with sufficient breadth and depth
83
5.2.1 Problem Description
Figure 24 gives a schematic of SUV design problem structure showing the flow of
design information between and across levels. Supersystem-level targets are prescribed.
For a given vector of design variables, “half-car” and “bicycle” analysis models (Section
5.2.2) generate responses for the target vector. The computed variable values are then
cascaded to the system-level design problem as targets. For example, the front suspension
stiffness is changed to achieve the desired front suspension first natural frequency. Once
an optimal value of the front suspension stiffness is found at the supersystem design
problem, that value becomes a target value at the system-level design problem, in which
the suspension design variables (coil spring stiffness and free length) are altered to achieve
value. The computed values of the variables such as the coil spring stiffness that gives the
optimal suspension stiffness, are then cascaded to the component level as targets. The
spring component variables are optimized to achieve minimal deviation from the targets
supersystem (vehicle) level become targets at the system level, where system level
variables (tire inflation pressure) are changed to meet the stiffness targets. In tire design
models for vertical and cornering stiffnesses, the inflation pressure is common, i.e., the
In Figure 24, each block indicates an optimal design model where design decisions
are made to achieve minimum deviation from the targets. Each design model calls one or
more analysis models to evaluate the current design. The supersystem-level design
problem contains two analysis models, the half-car model and bicycle model mentioned
above. System level analysis models for the front and rear suspensions utilize the
AUTOSIM modeling software (Hogland [32][33]). The tire models call the equations of
motions described in Appendix C and the front and rear spring models call the relevant
84
design equation described in Appendix D. These analysis models will be further described
Once the supersystem design targets are cascaded down to the lowest level, the
resulting design information must then be passed back to higher levels up to the top level.
In general, it will not be possible to achieve the target values exactly in each design
problem, due to constraints and variable bounds or to lower level responses. For example,
the front suspension stiffness obtained from the system-level optimization problem might
not match the target values from the supersystem level due to constraints on coil spring
free length and stiffness. Similarly, upon cascading the desired coil spring stiffness to the
coil spring component design problem, packaging or fatigue constraints might result in
spring stiffnesses deviating from the specified target value. Deviation in spring coil
stiffness will subsequently result in a deviation of the overall suspension stiffness, which
85
in turn will affect the first ride frequency of the vehicle. Thus an iterative process in top-
down and bottom-up fashion will lead to a consistent design considering limitations of the
design space and unearth potential discrepancies between overall system responses and
targets.
At the supersystem level, the ride quality and handling targets are as follows:
the driver’s perception of ride is affected by many factors (Ferris [21]). Use of natural
frequency as targets is valid, given that there exist sets of frequency values that are
accepted as necessary conditions for desirable ride characteristics for different classes of
vehicles.
The vehicle can be represented by a simple half-car model as in Figure 25. In this
model, the frequencies can be calculated in closed form as functions of sprung mass (Ms),
front and rear unsprung masses (Musf, Musr), and suspension stiffnesses. Equations of
motion are included in Appendix A. In this study the sprung and unsprung masses are
assumed to be prescribed a priori, and fixed are design parameters as opposed to design
variables in the optimization problem. The vehicle body is assumed as a single rigid body
mass. If more detail is needed, body can be represented as a finite element model (Kikuchi
and Malen [37]). The main design variables at the supersystem level are the stiffnesses of
the front and rear suspensions (Ksf, Ksr) and of the tires (Ktf, Ktr). Also, the distances a and
b from the vehicle center of mass to the front and rear axles are chosen as design variables.
Omission of the damping values (Csf, Csr) of the shock absorbers will not appreciably
86
affect the natural frequencies of the suspension, as the overall suspension damping ratio in
The first natural frequencies of the suspensions are primarily affected by changing
the front and rear suspension stiffnesses. Modifying the distances a and b also affects the
natural frequencies by changing the portion of the sprung mass that is carried by the front
and rear suspensions. In other words, the first natural frequencies are mostly affected by
x b a
Ms , J T
z
K sr C sr K sf C sf
M usr M usf
K tr K tf
V r(t) V f(t)
Figure 25. Half-Car model (Bold faced symbols are design variables.)
Next to the sprung mass of a vehicle, the axles and wheels (which make up the
unsprung mass) are the second largest masses capable of separate resonances as rigid
bodies. The wheel hop frequencies are functions of the prescribed unsprung masses and of
the tire stiffnesses that vary as a function of inflation pressure (Pif, Pir front and rear,
respectively) and normal vertical load (Gillespie [23]).
Pitch is defined as the angular component of ride vibrations of the sprung mass
about the vehicle y-axis, which is the lateral direction out the right side of the vehicle. The
pitch natural frequency is a function of the pitch moment of inertia (J), the distances a and
87
The handling target is represented by the understeer gradient, which describes how
the steer angle of the vehicle must be changed with the radius of turn or lateral
acceleration. The steering gradient is a measure of the magnitude and direction of the
steering input for a vehicle to track a curve of constant radius R with forward velocity u.
For the purpose of understeer analysis, it is convenient to represent the vehicle by the
bicycle model shown in Figure 26. The understeer gradient is a function of a and b and of
the tire lateral cornering stiffnesses CDf, CDr for front and rear. Refer to Appendix B for
Targets are set for the ride and handling characteristics described above,
comprising a multiobjective function for the top supersystem level that will minimize
exchange in the supersystem design problem is shown in Figure 27. The models needed at
Gf
Forward Velocity u
C Df
a
R
C Dr
Figure 26. Cornering of a bicycle model (Bold faced symbols are design variables.)
88
min||T super-Rsuper||+HR+H y
T super = target vector (given)
Subject to constraints on:
Rsuper = target values from supersystem model
amin d a d amax , bmin d bd bmax RS = responses from system models
yS = linking variables (tire pressures)
~ a ½ Convergence tolerances
xs ® ¾ ||RS-RSL|| d HR
¯b ¿ 1/2(||yS-yS3L||+||yS-yS4L||) d Hy
K SF ½
°K °
° SR ° RL
° K TF ° SF ½
s4
L
RS ® ¾ ° ° U L U y s4
° K TR ° ° SR ° R U R R R
°C ) ° RL R U
RL s3 s3 s4
° ° °° TF °° s1 s1 s2 s2 y Us
R super ® ¾ y Us
¯°C ) ¿° ° TR ° y Ls3
° P° To / from tire
HALF CAR VERTICAL ° ° cornering model
AND CORNERING °¯ kUS °¿ To / from To / from
MODEL
To / from tire
front rear vertical model
suspension suspension
The supersystem design problem is stated in the same manner as Eq. (31).
P super : Minimize
R super – T super + H R + H y
x̃ super y sys R sys H R H y
where
R super = r super R sys x̃ super
subject to (81)
L L
R sys – R sys d H R y sys – y sub d H y
For current exercise, target, design variables (local and linking), and responses are
89
U U U U
P super : Minimize Z –Z + Z –Z + Z –Z + Z –Z +
sf sf sr sr tf tf tr tr
U U
Z p – Z p + k us – k us + H R + H y
with respect to
Z sf Z sr Z tf Z tr Z p k us a b K sf K sr K tf K tr C Df C Dr P if P ir
where
K sf K sr K tf K tr
Z sf = --------- Z sr = ---------- Z tf = ------------ Z tr = -------------
M sf M sr M usf M usr
Kp
Zp = ------- 2a + b
K p = ---------------------------------------------------------------
J K sf + K tf K sr + K tr
--------------------------- + -----------------------------
aK sf K tf bK sr K tr
mb ma
k us = ------------- – -------------
LC Df LC Df (82)
subject to
L L
K sf – K sf d H R1 K sr – K sr d H R2
L L
K tf – K tf d H R3 K tr – K tr d H R4
L L
C Df – C Df d H R5 C Dr – C Dr d H R6
1§§ L ·2 § L · 2·
--- ¨ ¨ P if – P if ¸ + ¨ P if – P if ¸ ¸ d H y1
2©© v¹ © c¹ ¹
1§§ L ·2 § L · 2·
--- ¨ ¨ P ir – P ir ¸ + ¨ P ir – P ir ¸ ¸ d H y2
2©© v¹ © c¹ ¹
min max
a dada
min max
b dbdb
90
As noted previously, the masses are parameters at the supersystem level. The input
vector to the analysis models (half-car and bicycle models) is partitioned into local
variables a, b and system responses Ksf, Ksr, Ktf, Ktr, CDf, CDr.
Hy to the system-level linking variables. The norms of the deviation constraints are also
minimized, and so the deviation constraints are always active. Besides the deviation
constraints, constraints for supersystem local variables gsuper, hsuper are enforced. In the
study, a linking variable vector consisting of front and rear tire inflation pressures is
cascaded to both the vertical and the cornering tire stiffness models as targets. These target
values are determined upon solution of the supersystem problem, where the deviation
with respect to linking variables. Also, at an ideal convergence, the system response R sys
L
will match R sys , the target response calculated at the system level design problem.
In the current study, there exist four design models at the system level, two
suspension models for the front and rear, and two tire models for the vertical and
cornering stiffness. Each design model contains one analysis model. Suspension models
contain more detailed design models for coil spring design at the subsystem level, but the
tire models have no more design models below, but they share linking variables that are
coordinated at the supersystem level. Front/Rear suspensions that were represented with a
91
single spring damper system in the half-car model are now extended into more detailed
models at the system level. Similarly for tires, the single spring model is extended with
with a single spring and damper. At the system level, this single spring and damper is
replaced with the kinematic suspension analysis model by Hogland [32] (Figure 28) and
the suspension system in the vehicle is decomposed into front and rear suspensions.
Hogland’s suspension model was designed to enable design decision making for the ride
quality in the early stage of design process by utilizing a high-level vehicle ride
performance model and a low-level suspension design model. A high-level model, such as
RideSim [52], captures the overall vehicle performance while a low-level model, such as
shows a schematic of the simulation-based design process for vehicle ride using RideSim
and AutoSim. In the present study, the high-level RideSim model is replaced with a half-
Hogland’s main focus was on a low-level suspension model that has following
distinct features:
• The model has been designed to feed results into a large vehicle model specifi-
cally for vehicle ride. This gives an engineer the ability to study how suspen-
• The model has been written in AutoSim and the kinematic and dynamic equa-
• The model is totally parametric and can be easily modified to study simple
92
The present model is based on the actual suspension design of an existing sport-
utility vehicle. The suspension model takes as inputs basic suspension design points and
some component properties. The required hard point locations for the model are given in
Table E.1 of Appendix E. All units are entered in mm, with x, y, and z coordinate
directions consistent with vehicle design standards. Table E.2 in the same Appendix shows
other mandatory properties of the suspension that must be known and entered as
parametric inputs along with associated units. The suspension model is very flexible and
executable PC file generated using AutoSim by Hogland [32] and Rideout [65]. Since
AutoSim derives the equations of motion (Hann et al. [30]), a program is given that is
completely parametric and dependent on user inputs for key suspension properties and
geometry. In other words, by changing some values in the input file, the suspension model
can be used to evaluate different design specifications. The input file is described in
Appendix F.
z
Lo
Kspring
F
Among the input parameters from Appendix E, the suspension travel zsmax is
chosen as local variable; the linear coil spring stiffness KL, bending stiffness of the coil
spring KB, and the free length L0, are chosen as subsystem responses from the coil spring
design problem. There is no linking variable between the front and rear design problem.
93
Changing the free length of the spring changes the trim position of the nonlinear
RideSim
Target Suspension
Suspension Performance
Performance Response
AutoSim
L
R Us1 R s1
~
x SS 1 ^zsmax ` f
K Lf ½ K Lf ½
K Lf ½ U ° ° ° °
° ° R ss1 ® K Bf ¾ R L
ss1 ® K Bf ¾
R ss1 ® K Bf ¾
°L ° R s1 ^K `sf
° L0 f °
¯ ¿ ° L0 f °
¯ ¿
¯ 0f ¿
m
rS1 SS1
HOGL AND MU L T I -
B ODY SU SPE NSI ON
is shown in Figure 30. The TC system-level design problem for the front suspension
U
P s ys1 : Minimize K sf – K sf + H R1 + H R2 + H R3
subject to
L
K Lf – K Lf d H R1
L
K Bf – K Bf d H R2
(84)
L
L 0f – L 0f d H R3
min max
zsmaxf d zsmax f d zsmax f
min max
K Lf d K Lf d K Lf
min max
K Bf d K Bf d K Bf
min min
L 0f d L 0f d L 0f
For a given target value for suspension stiffness from the supersystem TC problem in Eq.
(82), the objective is to minimize the discrepancy between target and response. As there is
no linking variable at the subsystem level, the term for minimizing the linking variable
deviation is not included in the objective function. Besides the original variable bound
constraints for suspension design, additional deviation constraints from component level
L L L
are included in the constraint set. Deviations for component level responses K Lf K Bf L 0f
95
Vertical/Cornering Tire Models
The tire was represented as a single spring in the half-car model in the
supersystem. At the system level, two different aspects of the same tire analysis model,
vertical and cornering, are considered, and for each aspect a TC design problem is
formulated. Detailed equations for tire stiffness by Wong [92] are given in Appendix C.
A schematic of the information exchange for the vertical and cornering tire models
are shown in Figure 31. The design models for the vertical and cornering tire stiffness are
U U
P sys3 : Minimize K Tf – K Tf + K Tr – K Tr
K Tr = RVTIRE P if P ir
(85)
subject to
min max
P if d P if d P if
min max
P ir d P ir d P ir
U U
P sys3 : Minimize C Df – C Df + C Dr – C Dr
min max
P ir d P ir d P ir
96
Table 5.5: Summary of responses and variables at the system level
Design problem P P P P
s1 s2 s3 s4
Responses ( R s ) K
sf
K
sr K K
tf tr
C Df C Dr
Responses from subsystem level ( R ss ) K Lf K Bf L 0f K Lr K Br L 0r N/A N/A
In the tire models, the objective function is to minimize deviations for front and
rear tire stiffnesses (vertical K Tf K Tr or cornering C Df C Dr ) subject to variable bound
constraints for the tire inflation pressures for the front and rear P if P ir . The vertical and
cornering stiffness of the front and rear tires are calculated by the tire models FVTIRE,
FCTIRE, RVTIRE and RVTIRE (Appendix C). The inflation pressures for the front and
rear tires are linking variables that are coordinated at the supersystem level as in Eq. (84).
The values of the inflation pressure returned to the supersystem level by the vertical
stiffness model in general will be different from the values generated by the cornering
stiffness design model. Design autonomy in the two different system-level design
problems will give each of the design models freedom to have different linking variable
values from each other. Discrepancies between the linking variables are minimized at the
97
R Us3 R Ls3
y US yL
S3
Pif ½
yS3 ® ¾
¯ Pir ¿ K Tf ½
R s3 ® ¾
¯ K Tr ¿
rS3
R ADIAL T IR E
VE R T ICAL S T IF F NE S S
RUs4 L
R s4
yUS yL
S4
Pif ½
yS4 ® ¾
¯Pir ¿ C ½ I
Rs 4 ® ¾
¯C ¿ U
rS4
RADIAL T IRE CORNERING
S T IFFNES S
Figure 31. Schematic of information exchange for vertical and cornering tire models
98
5.2.4 Subsystem Level: Coil Spring Model
At the subsystem level below the suspension model, the front and rear coil spring
design models minimize the difference between target coil spring stiffness and the
response generated by the spring design analysis model. The coil spring design model
between target and actual linear spring stiffness, bending stiffness, and free length, while
The detailed equations for coil spring design including the above constraints are given in
Appendix D. A schematic of the information exchange for subsystem coil spring design is
shown in Figure 32. Target values for linear spring stiffness, bending stiffness, and free
length are cascaded down from the system level. Subsystem design variables are the wire
diameter, coil diameter, and pitch. Once optimal design is found, the updated target values
99
R USS1 L
R SS1
d f ½
~ ° ° K Lf ½
x SS 1 ®D f ¾ ° °
°p ° R ss1 ® ¾
¯ f¿ I
°L °
¯ 0f ¿
rSS1
S PR ING DE S IGN
E QUAT IONS
RUSS 2 R LSS 2
dr ½
~ ° ° K Lr ½
xSS 2 ®Dr ¾ ° °
°p ° R ss 2 ® U¾
¯ r¿
°L °
¯ 0r ¿
rSS2
S PR ING DE S IGN
E QUAT IONS
100
5.2.5 Design Scenario A: Equally Weighted Ride and Handling Targets
The models in the study were implemented in MATLAB and sequential quadratic
programming (SQP) was used to solve the individual TC optimization problems. The half-
car and bicycle models were written in MATLAB at the supersystem level, and the
suspension model was a DOS-executable file at the system level. Tire models were written
in MATLAB also, while the coil spring analysis models were written as Microsoft Excel
environment.
The computational process used to solve the TC problem in this study was as
follows (Figure 24): First, the top level vehicle design problem was solved and system
level targets were cascaded. Second, four system-level problems were solved
independently based on the targets assigned from the top level. Third, subsystem-level
problems for the front/rear coil spring design were solved. Based on the subsystem-level
responses, system-level design problems for front/rear suspension design were solved
again and all the system-level responses and linking variables from the four system design
problems were fed back to the top level, completing one iteration. This process was not
used as a formal coordination algorithm. Rather, iterations were terminated when the
deviation terms became smaller than tolerance H . Typically this was achieved within ten
iterations.
The study begins with baseline case with top level design targets that could be
conceivably specified at a high level from a number of departments. The targets are set
based on accepted design practice and parameters for typical sport-utility vehicles. The
objectionable vertical vibration in the 4-8 Hz range. Natural frequency for the rear
suspension is higher than the front aiming at mitigating vehicle’s tendency to be set into
pitching motion when the front wheels encounter a bump. The desired pitch frequency is
101
set below 1-2 Hz pitch plane vibration frequency range in correlation with occupant’s
discomfort.
The baseline study attempts to satisfy all departments involved by assigning equal
weight for each target after scaling. Deviation quantities were scaled to the same order of
The target values and responses from the baseline study are given in Table 5.6.
Figure 33 shows normalized comparisons of targets and responses, where “1” denotes an
exact match, greater than 1 denotes exceeding the target, and less than 1 denotes not
reaching the target. Exceeding the target does not necessarily mean better design, i.e.,
better than expected, because responses are normalized and the closer the response value
is to 1, the better the target is achieved. The optimal design from design scenario A is
given in Table 5.7. It is shown that TC yields a consistent design such that for a given
design quantity, such as front suspension stiffness, that is cascaded down from the level i
to the level (i+1) as a target, the response from the analysis model at the level (i+1) for
that design quantity matches the target closely, within a tolerance. Similarly, linking
variables converge to a single value within tolerance for each system they affect. For
example, the front suspension stiffness is 36.93 N/mm at the supersystem level and
matches the overall front suspension stiffness of 37.09 N/mm from the system-level
design problem. For the linking variable, the front tire inflation pressure in the vertical tire
model is 124.8 kPa, and is 123.35 kPa for the cornering stiffness model. If the tolerance is
tightened closely, then the responses and linking variables will match more closely. Note
that in Table 5.7 some of the variables are hitting lower or upper bounds, for example, the
lower bound for the front suspension stiffness is active (boldface). This indicates that if
the variable bounds are relaxed, then overall response will be changed for better
achievement of targets. Different design scenarios are discussed in the following sections.
102
Table 5.6: Supersystem targets and vehicle responses: Design scenario A
Target Desired value Response
Front suspension first natural frequency Zsf [Hz] 1.20 1.11
Rear suspension first natural frequency Zsr [Hz] 1.44 1.55
Front suspension wheel hop frequency Ztf [Hz] 12.00 11.55
Rear suspension wheel hop frequency Ztr [Hz] 12.00 11.55
Pitch natural frequency Zp [Hz] 0.50 0.87
Understeer gradient kus [rad/m/s2] 0.00719 0.00610
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
Front Ride Rear Ride Front Wheel Rear Wheel Pitch Under-steer
Freq. Freq. Hop Freq. Hop Freq. Frequency Gradient
103
Table 5.7: Baseline design scenario A
Supersystem Design Initial Values Optimal Values Lower Bounds Upper Bounds
CG distance to front [m] 1.32 1.25 1.25 1.39
CG distance to rear [m] 2.38 2.39 2.31 2.45
Front suspension stiffness [N/mm] 40 40.83 13.13 56.25
Rear suspension stiffness [N/mm] 40 40 25.7 40
Front tire stiffness [N/mm] 20 30 12.31 30
Rear tire stiffness [N/mm] 20 30 11.95 30
Front cornering stiffness [N/rad/10e-4] 10 10.36 4.81 12.88
Rear cornering stiffness [N/rad/10e-4] 10 8.96 4.81 12.88
104
The final response values match the targets closely, with the exception of the pitch
natural frequency and the understeer gradient. These quantities are both functions of the
distances a and b, which are hitting variable bounds. In the following section, two
different design scenarios B and C will be presented. Design scenario B uses different
target values, and design scenario C uses a modified design space. Indeed when the
there are two options to exercise: (1) changing targets, or (2) changing the design space.
From the baseline study in design scenario A, the design authority must assess the
deemed too high, the target cascading can be reapplied either with a different target value
(i.e. different objective function) or different design space. In design scenario B, the target
value for pitch frequency was decreased to 0.3 Hz in an attempt to increase the deviation
between the target and the response value, possibly causing the TC process to reduce the
final pitch frequency. No changes were made to the feasible design space.
Responses after changing the target value are given in Table 5.8. Compared to the
baseline responses, changing the target value alone has negligible effect on the response
and on the final design, especially when the baseline optimum is bound by active
constraints, (Figure 34). Design scenario B leads to a negligible change from the baseline
design scenario A. In Figure 34 the pitch frequency is compared to the same target value
from the baseline study. In other words, ratio “1” for pitch frequency means a perfect
match with the target value for design scenario A, which is 0.5 not 0.3. The facts that
suspension stiffness is affecting pitch frequency and the lower bound for the suspension
stiffness is active suggest that relaxing the feasible domain for the suspension stiffness
will lead to better achievement of the target. This will be investigated in design scenario C.
105
Table 5.8: Supersystem targets and vehicle responses: Design scenario B
Target Desired value Response
Front suspension first natural frequency Zsf [Hz] 1.20 1.11
Rear suspension first natural frequency Zsr [Hz] 1.44 1.54
Front suspension wheel hop frequency Ztf [Hz] 12.00 11.54
Rear suspension wheel hop frequency Ztr [Hz] 12.00 11.54
Pitch natural frequency Zp [Hz] 0.30 0.87
Understeer gradient kus [rad/m/s2] 0.00719 0.00590
1.8
1.6
1.4
1.2
Baseline
1
Scenario B
0.8
0.6
0.4
0.2
0
Front Ride Rear Ride Front Rear Wheel Pitch Under-steer
Freq. Freq. Wheel Hop Hop Freq. Frequency Gradient
Freq.
Assuming that relaxing some of the variable bounds is allowed, for example that
suspension stiffness can be lowered to 120 N/mm, the TC process is repeated for the SUV
design. As noted in the previous section, changing the target value hardly affects overall
responses when the optimal design is found at the boundaries of feasible space.
106
For design scenario C, target values are kept the same as in the baseline design in
scenario A. Instead, the variable bounds for coil spring stiffnesses in the front and rear
suspensions are relaxed. The design authority receives feedback from the baseline design
and realizes that the targets assigned for each department are not achievable within the
current design specification. Also, in case of boundary optima, changing target values
does not help to produce better design in terms of achieving targets closely. Thus the
design departments are allowed to change design specifications by changing the feasible
The current case study adopts a change in the feasible space by relaxing variable
bounds for certain design variables. As a result, the pitch frequency that had the most
significant discrepancy from the target value, has now a better response, i.e., it is closer to
the target value compared to design scenario B. Table 5.9 summarizes top level responses
and Table 5.10 gives detailed designs for low levels. Table 5.10 suggests that the TC
process in design scenario C achieves consistency in the optimal design, matching the
107
1.8
1.6
1.4
1.2 Run 1
1 OPTION A
0.8 OPTION B
0.6
0.4
0.2
0
Front Ride Front Pitch
Freq. Wheel Hop Frequency
Freq.
108
Table 5.10: Baseline design scenario C
Supersystem Design Initial Values Optimal Values Lower Bounds Upper Bounds
CG distance to front [m] 1.32 1.25 1.25 1.39
CG distance to rear [m] 2.38 2.45 2.45 2.45
Front suspension stiffness [N/mm] 40 36.93 13.13 56.25
Rear suspension stiffness [N/mm] 40 36.41 25.7 40
Front tire stiffness [N/mm] 20 30 12.31 30
Rear tire stiffness [N/mm] 20 30 11.95 30
Front cornering stiffness [N/rad/10e- 10 10.55 4.81 12.88
Rear cornering stiffness [N/rad/10e-4 10 8.79 4.81 12.88
109
5.2.8 Numerical Characteristics and Discussion
Both scenarios B and C lead to a significant discrepancy for the fifth target, the
pitch frequency with slightly better results under the scenario C, changing design space.
Changing targets (scenario B) does not have much effect on the response after the TC
process because some of the constraints were active and some variables were bounded by
variable bounds. Although the target was not closely matched, changing design space
(scenario C) has more effect on the response in terms of achieving the targets closer.
by component level. Once the process reached the bottom component level, responses are
fed back to the system level and finally to the top vehicle level completing one iteration
loop. Four design problems at the system level can be solved in a parallel fashion, but in
this case they were solve sequentially, still maintaining independent solution process for
each problem. Comparing computational efficiency for each scenario was not considered
Design authority and teams can learn from these scenarios that when there is
discrepancy for certain targets, it is critical to have alternative design options rather than
assigning new target values. Alternative design options include changing design space,
design with different material, or different design configurations.
consistent design using decomposition and multilevel framework. Two successful cases
First, geometric programming case demonstrates that TC gives the same optimal
optimization approach such as AAO. With more robust convergence proofs in CHAPTER
110
4, the TC approach guarantees that it gives the optimal solution to the original large-scale
design problem. Second, SUV chassis design case demonstrates the multilevel nature of
the target cascading framework. The case study is composed of seven design problems in
three levels and composed of eight analysis models written in heterogeneous computing
collaboration for large-scale design. Systems engineering practice using the TC follows in
the next chapter demonstrating that TC approach can be applied in a real design
organization.
111
CHAPTER 6
TARGET CASCADING IN SYSTEMS
ENGINEERING PRACTICE
described. Starting from collecting analysis models using model forms, analysis models of
interest are mapped onto a design hierarchy and design models are posed at different
levels of the hierarchy and linked to the analysis models required for computing the
objective and constraint functions. The TC process is imposed on the design modeling
hierarchy and lower level targets are identified taking into account interactions among
different design models in the form of linking variables and responses. Based on the multi-
level system targets obtained from the TC process, embodiment design with more detailed
models follows.
As defined in the Section 3.1.2, an analysis model serves to evaluate the current
design by taking design variables and parameters as input and returning responses as
output. For large-scale system design, design teams can have numerous analysis models
“simple” models are preferred for a quick TC process in the early stages of product
development. High fidelity complex models will substitute for the simple models in a
subsequent embodiment design process that follows the TC. These simple models can be
112
relations. Expensive models for embodiment design are high-degree of freedom
multibody dynamics models, finite element models with large number of elements,
Design teams will first collect analysis models to build an analysis model database
with individual model specifications stored along with the models. For the collection of
analysis models, models forms A and B (see Table 6.1 and Table 6.3) will serve to identify
input/output relationships and components for a specific analysis model and to map those
First, the Model Form A summarizes input and output relationships including
information on the assumptions, fidelity, and computational cost of the model. Table 6.1
gives an example of Model Form A. A model for durability analysis takes three types of
inputs: parameters, variables, and responses from the lower level. Parameters are fixed
during the analysis, for example, five road profiles for durability analysis are fixed
parameters that were obtained a priori. The distinction between variables and responses
from the lower level might not be as clear as for the parameters. Considering the modeling
hierarchy, if there is another layer of models below the current analysis and the output
from these lower-level models affects the current analysis model output, then the inputs to
the current analysis model are categorized as the responses from the lower level.
Otherwise, the rest of the inputs are categorized as variables that do not have direct
correlation with the lower-level models. The outputs from an analysis model are responses
into the higher level; for example, the durability model takes system-level design variables
and responses at the system level, and returns supersystem-level output responses.
assumptions, including the computer platform on which the model was developed,
software necessary to run the model, and components or file structure of the model. The
model fidelity column describes the fidelity of the model which can be subjective
113
depending on the assumed viewpoint, but it is needed to provide some judgement on the
Next the Model Form B must be created. This form indicates where a given model
within the system. In Table 6.3, a vehicle system is decomposed into body, chassis and
powertrain, and the attributes of interest are handling, NVH (Noise, Vibration, and
drivability, ride, and acoustic signature. Any existing correlations between the
decomposed subsystems and attributes are ranked into four levels: primary, strongest,
attribute, the cell in the table is empty. For a model, it is preferred to have one or two (not
many) primary relations to obtain a more meaningful mapping of the characteristics of the
model. Using these two Model Forms A and B a database is built to store all the necessary
information regarding analysis models so that they may be pulled out of the database
when designers want to utilize them to evaluate some aspect of the current design.
dispersed, who contributed a large variety of simulation modules related to some aspect of
ground vehicle performance. The model collection process took place over a period of
twelve months and upon completion of the extensive model collection efforts, researchers
in the ARC have built up broad knowledge of ARC’s current research and development in
the context of the systems engineering. Also a mapping of the analysis models in the TC
framework has lead each area to evaluate current research effort and direction, for
example, each area could diagnose where the research effort is being duplicated, where
they lack of appropriate analysis models, or where they have strong expertise of the model
development.
114
Table 6.1: Model Form A: Summary description of model
Computati
Model formats, Fidelity
onal Cost
contents and (High,
Inputs Outputs (Seconds,
assumptions Medium,
Hours, or
(actual software) or Low)
Days)
• 5 Road • Body Front • Jounce Travel • Peak Force to Matlab: 3 .m files Medium Hours
Profiles Mass and Rate Body and 2 .mat files
• ... • Body Rear Mass • Shock Jounce • ...
• Wheel Base and Rebound
Length • ...
• Body CG Height
• ...
In the ARC, the Model Forms A and B were utilized to collect all the available
analysis models from five research thrust areas: (1) intelligent vehicle dynamics and
control, (2) synthetic and virtual environment, (3) high performance structures and
materials, (4) advanced and hybrid powertrain, and (5) integrated system design and
simulation. The model collection activity was designed to allow understanding of each
analysis model and to get the model ready to be mapped in a target cascading framework
for vehicle design. Table 6.4 summarizes the collection of the ARC models in Model
Form B. Looking at this table, one can also tell which is the main focus of current research
and development effort in the ARC. A design authority can combine multiple models for
multidisciplinary design from the table. Table 6.2 identifies the models represented by
116
Table 6.3: Model Form B: Model subsystem/attribute interaction
Place your analysis model subsystems in the boxes. Indicate the relation of your model to attributes listed using the symbols
below.
✪ Primary relation
★ Strongest relation
✩ Stronger relation
❍ Weak relation
Attributes
Subsystem Fuel
Acceleration
Handling NVH Economy & Durability Weight ...
Performance
117
Emissions
Chassis ❍ ✩ ...
( e.g. Suspension,
Tires, ...)
Body ❍ ✩
(e.g. Structure,
Joints, ...)
Powertrain ✩ ✩ ✪ ★
(e.g. Engine,
Transmission,
Final Drive, ...)
Table 6.4: ARC model collection
✪ Primary relation ★ Strongest relation ✩ Stronger relation ❍ Weak relation
Subsystem Attributes
Chassis ✪(11) ✪(12) ❍(21) ✩(21) ❍(26) ★(11) ★(12) ✩(11) ✩(12) ✩(11) ✩(12) ★(11)
( e.g. Suspension, ✪(13) ❍(15) ✪(35) ❍(27) ★(15) ★(16) ★(26) ★(27) ✩(15) ✩(16) ★(12)
Tires, ...) ✩(16) ✪(17) ★(38) ❍(28) ★(21) ✪(23) ❍(28) ❍(26) ❍(13)
✪(18) ✪(21) ★(39) ❍(35) ✪(24) ❍(26) ✩(32) ✪(33) ❍(27) ✩(34) ✪(15)
★(25) ✪(26) ★(51) ❍(38) ❍(27) ❍(28) ★(34) ★(35) ✩(38) ✩(39) ✪(16)
✪(27) ✪(28) ❍(39) ✩(35) ✩(38) ✪(38) ✪(39) ❍(17)
★(51) ✩(32) ✪(51) ❍(18)
118
❍(33) ❍(34)
✩(41)
Body ❍(26) ❍(27) ✪(31) ✩(11) ✩(12) ★(27) ✪(11) ✪(12) ❍(41) ★(31)
(e.g. Structure, ❍(28) ✩(32) ✪(32) ✪(23) ✪(24) ❍(28) ✪(13) ✪(15)
Joints, …) ❍(33) ❍(34) ✩(41) ✩(31) ✪(16) ✪(17)
✩(43) ✩(32) ✪(33) ✪(18) ❍(31)
★(34) ✪(32) ✩(34)
✩(41)
Powertrain ❍(14) ✩(22) ❍(22) ★(14) ★(22) ✪(14) ✪(22) ❍(22) ❍(22) ✩(36) ✪(14)
(e.g. Engine, ✩(25) ❍(26) ✪(36) ✩(36) ★(37) ✪(23) ✪(24) ★(36)
Transmission, ❍(27) ★(37) ★(41) ★(42) ✪(25) ✩(36) ✩(37)
Final Drive, …) ✩(41) ✩(42) ✩(41) ★(43) ✪(37) ✪(41)
✩(42) ✪(42) ✪(43)
✩(43) ✪(44)
One research question that needs to be further investigated in the future is how to
pull out appropriate models from the analysis model database, in other words, how the
analysis model library is queried. The designer should interact with the database system
by submitting requests for data selection (queries). In the systems engineering process,
designer sends queries to the database to select appropriate models and this process can be
Once the specifications (requirements) are given, they are sent to the database in the query
language, and the database returns a model or a set of models that satisfy the requests. In
the target cascading to date, this analysis model selection process is executed manually in
that the designer picks the right model based on the information provided by Model Forms
A and B.
systems are the focus of intense interest. An agent is a computer system situated in some
environment, capable of autonomous action within this environment in order to meet its
design objectives (see Jennings et al. [36]). An agent-based system in combination with
database queries can be implemented for the target cascading system engineering design
framework in terms of model selection. An intelligent design agent can select an analysis
model based on the queries to the analysis model database to meet certain requirements.
meet the target, there can be several design alternatives. An agent-based system can
automate the model selection process for TC and a subsequent embodiment design
process by returning a set of design alternatives (options) for the same design objectives.
are usually heterogeneous and it is critical to collaborate between different design teams
across heterogeneous design environment. Software such as Windchill from PTC [61]
119
provides a web-based portal over the Internet to facilitate collaboration between design
teams for an effective product data management. In combination with the intelligent
design agent system, TC process can be automated over the Internet and implemented in
After collecting models from the design teams, the next question is: At which level
of the target cascading modeling hierarchy should the analysis model be located? In order
words, which model has the proper degree of resolution or fidelity for a specific job in the
TC design process? In the TC hierarchy, simple models are usually located at higher levels
to provide quick evaluations of the proposed designs while more complex models are
teams. Collaboration among design engineers to determine the relative positions of the
own analysis models and conduct a discussion with a worksheet, if necessary, to map
those models in the overall hierarchy. This is an ideal application of the multifunctional
In the study presented in Section 5.2, as one goes down from the supersystem to
system, subsystem, and component levels, the complexity of analysis models increases.
First, there are eight different models available to apply the target cascading process for
ride and handling quality achievement. All of them are different in terms of complexity,
models are mapped as follows. At the top level, the half-car and bicycle models are
relatively simple and quick to evaluate capturing the overall behavior of the vehicle. The
120
suspension stiffness that was represented with a single spring at the top level is computed
with a more detailed geometric suspension model at the system level. Likewise, the tire
model was represented as a single spring at the supersystem level, and it is replaced with
horizontal and vertical tire models at the system level. Furthermore, the spring
characteristics in the suspension model were calculated with a more detailed coil spring
In a large organization, mapping the right model in the right position in the TC
hierarchy is critical to achieving consistent design. Different teams can have different
terminology to describe the same physical entity and a complex model in one team can be
regarded as a very simple model for the design objectives of another team. The next
question then is how to build design models based on the analysis model hierarchy for
Design models are differentiated from analysis models in the TC process. Design
decisions are made in the design models and analysis models evaluate design decisions in
conjunction with design models. A design model can call multiple analysis models.
It is not clear whether mapping the analysis model precedes building the design
models, but these are two distinct steps in the TC process. Top design targets are set by a
design authority, and a design problem is formulated at the top level to achieve the targets
In this stage of the TC process, design objectives and constraints for design models
below the top level are formulated based on the interactions between multiple design
straightforward because the targets are assigned a priori. However, design models below
the top level should consider the interactions between other design models via responses
and linking variables, which comprise the design objectives and deviation constraints. In
121
other words, based on the analysis modeling hierarchy defined in the previous section,
interactions among different design models at different levels should be identified in this
step.
Once the interactions between design models are identified, the next task is to
build design models until the comprehensiveness of the modeling hierarchy in the design
modeling hierarchy means that details of designs are realized up to design authority’s
satisfaction. In other words, if design models are posed at high levels only, overall design
from the TC process will be too abstract. If design models are posed in too many levels
with too much complexity of analysis models, the advantage of the TC process to cascade
design targets to low-level design teams will be lost. Design models follow formulations
presented in CHAPTER 3.
After design models are formulated in conjunction with analysis models to achieve
top-level targets, the individual problems are solved to find their optima. Consistency is
achieved throughout the TC process by maintaining the same values for the linking
variables, and matching responses to the targets from higher levels. Starting from the top
level targets, the optimal design is cascaded down to a lower level, and the lower-level
design problems are solved to achieve minimum deviations from the higher level designs.
Once the top-down solution process is completed, the lower-level designs are fed back to
the higher-level problems. If the lower-level designs deviate from the targets within their
specified design space, the higher-level design problems will adjust their design and this
design models at multiple levels and the optimal solutions for each of the design problems.
With the modeling hierarchy in the TC process, the design authority can assess overall
122
system performance in terms of how well the targets are achieved. As the TC process
should take place in the early stages of product development, further detailed design, i.e.,
embodiment design, should follow while fixing targets, responses, and linking variables to
Analysis models used in the TC process should be quick to run and accurate to
evaluate design alternatives. These simple models are replaced with more detailed high-
fidelity models in the embodiment design step. Interactions among design problems, i.e.,
responses, and linking variables, are identified in the TC process. Holding these values as
constant, analysis models are substituted with more detailed models. For example, the
body of a vehicle in the half-car model was represented with a single mass in the TC
process. Now the body is replaced with a body-in-white (BIW) finite element model
composed of beams or an even more detailed model such as a full-scale finite element
model with thousands of nodes and elements. If the mass of the entire body was found in
the target cascading process, it is fixed as the optimal value and geometric design follows
with detailed models resulting in geometric descriptions of different parts of the body.
The embodiment design process leads to the final product for the large-scale
system design. If the target cascading process were to be done with the detailed models for
embodiment design, the process might take a lot of computational effort. Hence detailed
models are preferred in the embodiment design step after the TC process. Holding the
interactions between other design models constant, if a detailed design cannot be found
within design space in the embodiment design step, then design teams must go back to the
TC process to adjust the interactions with other design problems. This feedback process is
similar to the different design scenarios in the SUV chassis design in Section 5.2 in that, if
the targets cannot be achieved, the designer can either go back to the higher level to adjust
target values or to change current design space. Also in case of the embodiment design,
123
the design can try to change design space, for example, by relaxing some constraints,
using different material, or by changing configurations. Or the design can go back to the
Forms devised for collecting analysis models are beneficial to design organizations in that
they facilitate understanding current research and development efforts and cooperating
with different design teams for further understanding of analysis models. Analysis models
are required to comply the syntax of the Model Forms to be further understood in a
multilevel modeling hierarchy. By mapping analysis models at the appropriate level of the
modeling hierarchy and by formulating proper design models in multiple levels, top-down
and bottom-up target cascading process can be fully integrated at the early stages of
product development. Researchers in the ARC participated in this model collection and
mapping effort and successful outcome was drawn resulting in overall modeling hierarchy
at current state.
model database, pulling out right models for a specific design purpose (model query),
generating multiple design alternatives. These issues will be further addressed in the
following chapter.
124
CHAPTER 7
CONCLUSIONS
qualifications was given to show difficulties associated with MDO formulations such as
that have convergence proofs were reviewed as well. The algorithms reviewed have a
single level structure of analysis models, i.e., the large-scale problem is decomposed in
multiple pieces and on top of them there is a master (coordination) problem. On the other
hand, target cascading is posed for a multiple-level hierarchy of analysis models, for
which the previous coordination methods are not readily applicable, although they may
models and analysis models. The interactions between different design models are clearly
125
qualifications for the additional deviation constraints introduced in the TC framework
In Chapter 4 a coordination algorithm was presented for the two-level case of the
TC formulation. The equivalence between the TC and HOC formulations for the purpose
of coordination was shown and the possibility of proving a non-ascent property for TC via
the HOC formulation was explored. It was shown that at least a non-ascent property can
be easily established. However, true descent and convergence to an optimal point requires
further investigation. This is because the HOC proof is based on the assumption that the
number of linking variables is small relative to the total, while in the TC-HOC
equivalence transformation this assumption is not satisfied. Empirical evidence shows that
problem was presented to show that the TC solution matches the AAO optimal solution.
An engineering design example for an automotive chassis design composed of three levels
of design models was presented and extensively studied. Different design scenarios were
process includes analysis model collection, mapping of the models, and target cascading
implementation. Target cascading was shown to serve as a decision support tool in a real
7.2 Contributions
From a design viewpoint, the main benefit of the proposed approach for target
cascading is reduction in large-scale design cycle time, avoidance of design iterations late
in the development process, and increased likelihood that physical prototypes will be
closer to production quality. Design iterations are reduced by integrating the target
126
propagation and target matching processes into a single procedure. Using a partitioning
that comprises systems, subsystems, and components reduces the complexity of the
overall design problem and allows more systematic concurrent design of the system’s
elements.
multilevel formulation for the target cascading process in systems engineering and a
coordination algorithm with a non-ascent property. The proposed framework is suited for
large-scale design problems and can handle the associated multilevel design process in a
consistent manner.
points:
different design teams, while at the same time allowing design autonomy
problem.
127
one may expect loss of efficiency when compared to an AAO strategy, the
design problems are usually discrete or mixed. To avoid being stuck at a local
process is underway using the extendible markup language (XML) and the
TC model selection and mapping process, which is being done manually for
now.
with target cascading should be applied to allow design agents with appropriate
intelligence for follow-up design actions with different options. This idea is an
extension of item 3. Based on the query, the database will return possibly
multiple analysis model candidates. Each candidate will have trade-offs. If the
TC process can give a brief overview of the design outcome for various design
128
5. Target cascading should be expanded to include financial modeling.
129
APPENDICES
130
APPENDIX A
Half-Car Model
Based on the input variables a b K sf Ksr K tf K tr , output frequencies are calculated
The first natural frequencies are calculated considering front and rear of the
vehicle individually. Front and rear frequencies are calculated by the following formula.
K
sf
Z = --------- (A.2)
sf M
sf
K
sr
Z = ---------- (A.3)
sr M
sr
where Msf and Msr are the portions of sprung mass by the front and rear
suspensions respectively.
K
tf
Z = ------------ (A.4)
tf M
usf
K
tr
Z = ------------- (A.5)
tr M
usr
where Musf and Musr are the front and rear unsprung masses. The equations are
based on the assumptions that an order-of-magnitude difference exists between sprung and
The pitch natural frequency is a function of pitch stiffness Kp, pitch moment of
inertia J and the pitch stiffness is a function of front/rear suspension and tire stiffnesses,
and the distances a and b from the center of mass to the front and rear axles.
131
K
p
Z = ------- (A.6)
p J
where the pitch stiffness is
2a + b
K = ---------------------------------------------------------------
p K + K K + K . (A.7)
sf tf sr tr
--------------------------- + -----------------------------
aK K bK K
sf tf sr tr
132
APPENDIX B
Bicycle Model
The steady-state cornering equations are derived from the application of Newton’s
Second Law along with the equation describing the geometry in turns. For the purpose of
The angle between the direction of heading and the direction of travel of a tire G f
is calculated by
2
G f = --- + § ------------- – -------------· -----
L mb ma u
(B.1)
R © LC Df LC Df¹ R
where L is the wheelbase, R is the radius of the turn, m is the mass of the vehicle,
a b are the distances from the center of gravity to the front and the rear, and u is the
describes how the steer angle of the vehicle must be changed with the radius of turn or the
mb ma
k us = ------------- – ------------- . (B.2)
LC Df LC Df
It consists of two terms, each of which is the ratio of the load on the axle (front or
rear) to the cornering stiffness of the tires on the axle. The BICYCLE model in Eq. (82)
takes distances a b and the cornering stiffnesses C Df C Dr as input and return the
133
APPENDIX C
Tire Model
The tire stiffnesses are calculated along the direction of the vehicle: vertical or
cornering direction. The stiffness of the tire in the vertical direction is a function of the
inflation pressure P i [kPa] and the datum vertical load on the tire F m [N]. The load due to
vehicle mass is distributed between the front and rear tires as a function of the tire
stiffness to rolling stiffness. The remaining coefficients are from a curve fit to sample data
where
9.81Mb
F = ------------------- . (C.2)
m a+b
Similarly, the cornering stiffnesses depend on pressure and load:
–6 2 –3 – 2 180
CD = F m – 2.668 u10 P i + 1.605 u10 P i – 3.86 u10 --------- [N/rad]. (C.3)
S
134
APPENDIX D
Coil Spring Equations
The spring design module calculated linear stiffness and bending stiffness as a
function of coil diameter D, wire diameter d, and pitch p. At the optimum, the difference
between the calculated values and the optimal spring parameters from the suspension
4
Gd
K L = ---------------------------------- (D.1)
3 L 0 – 3d
8D § ------------------·
© p ¹
where L 0 is spring free length in mm, G is modulus of rigidity of spring material.
4
EGd
K B = --------------------------------- (D.2)
16D 2G + E
The following constraints incorporate the formulae for shear stress and factor of safety for
fatigue failure. The first constraint requires that the maximum shear stress be less than the
maximum allowable shear stress divided by the factor of safety in shear n s . The second
constraint ensures that the actual fatigue failure factor of safety is greater than the
S su
F a + F m u § --------3 + --------2· – ------
8D 4
-d0 (D.3)
© Sd Sd ¹ n s
3
§ S su S se Sd ·
n f – ¨ ---------------------------¸ e § § ------- + 2· e § ------- – 3· F a S su + § ------- + 1· e § -------· F m S se· d 0
4D 4D 2D 2D
(D.4)
© 8D ¹ © © d ¹ © d ¹ © d ¹ © d¹ ¹
where S su is the maximum allowable shear stress, S se is fatigue endurance limit, and
135
APPENDIX E
Hogland’s Suspension Model Inputs
Table E.1: List of hard point locations required for model input
Geometric Location
136
Table E.2: List of vehicle/component properties required for input
137
Table E.3: Optional suspension parameters for future dynamic simulations
138
APPENDIX F
Input File for Suspension Model
PARSFILE
* Multibody system dynamic simulation.
* Version created by AutoSim 2.8 on April 4, 2000.
* Copyright 1999. Mechanical Simulation Corporation. All rights reserved.
* PARAMETER VALUES
139
MDD 1.54 , mass of DD (kg)
NEUANG 15.54 , term in coefficient in Moment due to Spring (deg)
RBNDBMPLNGTH 16.8 , argument to REBSTOP in rebound stop force
(mm)
RTIME 0 , CALC -- Computational efficiency (sec/sim. sec) (-)
SPINXADJ -0.887 , NIL (mm)
SPINYADJ -1.377 , NIL (mm)
SPINZADJ -12.18 , NIL (mm)
SPRFREEL 420 , negative term in coefficient in Spring Force (mm)
TOEADJ -0.151 , NIL (deg)
XP(1) 877.5 , X coordinate (mm)
XP(10) 855.44 , X coordinate (mm)
XP(14) 1013 , X coordinate (mm)
XP(15) 1026 , X coordinate (mm)
XP(16) 1026.9 , X coordinate (mm)
XP(17) 1016.8 , X coordinate (mm)
XP(18) 1 , X coordinate (mm)
XP(2) 1197.5 , X coordinate (mm)
XP(21) 1029.2 , X coordinate (mm)
XP(22) 1028.6 , X coordinate (mm)
XP(23) 1150.83 , X coordinate (mm)
XP(3) 972 , X coordinate (mm)
XP(34) 1124.5 , X coordinate (mm)
XP(4) 1285.5 , X coordinate (mm)
XP(5) 991.3 , X coordinate (mm)
XP(6) 1017.7 , X coordinate (mm)
XP(7) 1000 , X coordinate (mm)
XP(9) 845.1 , X coordinate (mm)
YP(1) 402.2 , Y coordinate (mm)
YP(10) 284.18 , Y coordinate (mm)
YP(14) 437.5 , Y coordinate (mm)
YP(15) 481.5 , Y coordinate (mm)
YP(16) 427.5 , Y coordinate (mm)
YP(17) 470.6 , Y coordinate (mm)
YP(18) 1 , Y coordinate (mm)
YP(2) 379.6 , Y coordinate (mm)
YP(21) 696.2 , Y coordinate (mm)
YP(22) 528.2 , Y coordinate (mm)
YP(23) 505.3 , Y coordinate (mm)
YP(3) 251.6 , Y coordinate (mm)
YP(34) 515.9 , Y coordinate (mm)
YP(4) 325 , Y coordinate (mm)
YP(5) 686.5 , Y coordinate (mm)
YP(6) 636.3 , Y coordinate (mm)
YP(7) 795 , Y coordinate (mm)
YP(9) 690.7 , Y coordinate (mm)
140
ZP(1) 763.6 , Z coordinate (mm)
ZP(10) 532.45 , Z coordinate (mm)
ZP(14) 709.2 , Z coordinate (mm)
ZP(15) 377.2 , Z coordinate (mm)
ZP(16) 684.5 , Z coordinate (mm)
ZP(17) 374 , Z coordinate (mm)
ZP(18) 1 , Z coordinate (mm)
ZP(2) 698 , Z coordinate (mm)
ZP(21) 472.51 , Z coordinate (mm)
ZP(22) 698 , Z coordinate (mm)
ZP(23) 467.1 , Z coordinate (mm)
ZP(3) 462 , Z coordinate (mm)
ZP(33) 547.8 , Z coordinate (mm)
ZP(34) 497.7 , Z coordinate (mm)
ZP(4) 462 , Z coordinate (mm)
ZP(5) 407.7 , Z coordinate (mm)
ZP(6) 700.4 , Z coordinate (mm)
ZP(7) 501.71 , Z coordinate (mm)
ZP(9) 475 , Z coordinate (mm)
141
* Forces applied by road
ROADF-TABLE
10, 0
50, 10000
110, -4000
ENDTABLE
* INITIAL CONDITIONS
END
142
Table G.1: Model Form A: Summary description of model
Computati
Model formats, Fidelity
onal Cost
contents and (High,
Inputs Outputs (Seconds,
assumptions Medium,
Hours, or
(actual software) or Low)
Days)
APPENDIX G
Model Forms
Design)
143
Table G.2: Model Form B: Model subsystem/attribute interaction
Place your analysis model subsystems in the boxes. Indicate the relation of your model to attributes listed using the symbols
below.
✪ Primary relation
★ Strongest relation
✩ Stronger relation
❍ Weak relation
Attributes
Subsystem Fuel
Acceleration
Handling NVH Economy & Durability Weight ...
144
Performance
Emissions
Chassis ❍ ✩ ...
( e.g. Suspension,
Tires, ...)
Body ❍ ✩
(e.g. Structure,
Joints, ...)
Powertrain ✩ ✩ ✪ ★
(e.g. Engine,
Transmission,
Final Drive, ...)
BIBLIOGRAPHY
145
BIBLIOGRAPHY
1. Akao, Y., Quality function deployment: integrating customer requirements into product
design, 1990, Productivity Press.
4. Azarm, S. and Li, W., “A two level decomposition method for design optimization,”
1988, Engineering Optimization, Vol. 13, pp. 211-224.
6. Balling, R. J. and Sobieszczanski-Sobieski, J., “An algorithm for solving the system-
level problem in multilevel optimization,” 1994, AIAA Paper No. 94-4333, 5th
AIAA/USAF/NASA/ISSMO Symposium on Multidisciplinary Analysis and
Optimization, Panama City, Florida.
8. Barthelemy, J. and Riley, M., “Improved multilevel optimization approach for the
design of complex engineering systems,” 1988, AIAA Journal, Vol. 26, No. 3, pp.
353 ~ 360.
9. Bazaraa, M.S., Sherali, H. D., and Shetty, C. M., Nonlinear Programming, 2nd edition,
John Wiley & Sons, 1993.
10. Beightler, C. S. and Phillips, D. T., Applied Geometric Programming, 1976, John
Wiley and Sons, Inc., New York.
146
11. Braun, R., Collaborative Optimization: An Architecture For Large-Scale Distributed
Design, 1996, Doctoral Dissertation, Stanford University.
12. Braun, R., Gage, P., Kroo, I., and Sobieski, I., “Implementation and Performance
Issues in Collaborative Optimization,” 1996, AIAA Paper 96-4017.
13. Braun, R. and Kroo, I., “Development and Application of the Collaborative
Optimization Architecture in a Multidisciplinary Design Environment,”
Multidisciplinary Design Optimization: State-of-the-Art, N. Alexandrov and M.
Hussaini (eds.), SIAM, 1996.
14. Braun, R., Moore, A., and Kroo, I., “Use of the Collaborative Optimization
Architecture for Launch Vehicle Design,” 1996, Proceedings of the 6th AIAA/
NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization, AIAA-
96-4018-CP.
15. Bulte, C. and Moenaert, R., “The Effects of R&D Team Co-location on
Communication Patterns among R&D, Marketing, and Manufacturing,”
Management Science, Vol. 44, No. 11, pp. S1 ~ S18, 1998.
16. Burgess, M. A., “The future of aircraft design,” September 1991, Aerospace America,
pp. 22 ~ 25.
17. Chankong, V. and Haimes, Y., Multiobjective Decision Making: Theory and
Methodology, 1983, Elsevier Science Publishing Co. Inc., New York.
18. Cramer, E., Dennis, J., Frank, P., Lewis, R, and Shubin, G., “Problem Formulation for
Multidisciplinary Optimization,” 1994, SIAM Journal of Optimization, 4 (4):
754~776.
19. Dantzig, G. B. and Wolfe, P., “Decomposition principles for linear programs,” 1960,
Operations Research, Vol. 8, pp. 101-111.
20. Dennis, J.E. and Schnabel, R.B., Numerical Methods for Unconstrained Optimization
and Nonlinear Equations, 1983, Prentice-Hall Series in Computational
147
Mathematics.
21. Ferris, J., “Factors Affecting Perceptions of Ride Quality in Automobiles,” 1999,
ASME International Conference and Exhibition.
22. Gage, P., New approaches to Optimization in Aerospace Conceptual Design, 1995,
Doctoral Dissertation, Stanford University.
24. Gittins, D., Query Language Systems, 1986, Edward Arnold Publishers Ltd.
25. Grana, M. and Torrealdea, F. J., “Hierarchically structure systems,” 1986, European
Journal of Operational Research, Vol. 25, pp. 20 ~ 26.
26. Haftka, R., “An Improved Computational Approach for Multilevel Optimum Design,”
1984, J. Struct. Mech., Vol. 12, No. 2, pp. 245 ~ 261.
27. Haimes, Y., Wismer, D., and Lasdon, L., “On a Bicriterion Formulation of the
Problems of Integrated System Identification and System Optimization,” 1971,
IEEE SMC-1, pp. 296 ~ 297.
28. Haimes, Y. and Li, D., “Hierarchical Multiobjective Analysis for Large-scale Systems:
Review and Current Status,” 1988, Automatica, Vol. 24, No. 1, pp. 53 ~ 69.
29. Han, S.-P., Superlinearly Convergent Variable Metric Algorithms for General
Nonlinear Programming Problems, 1974, Doctoral Dissertation, University of
Wisconsin.
30. Hann, S. A., Nakamura, S. and Sayers, M. W., “Painless Derivation and Programming
of Equations of Motion for Vehicle Dynamics,” 1992, Proceedings from AVEC
1992, SAE Paper No. 923010.
31. Hansen, M., “The Search-Transfer Problem: The Role of Weak Ties in Sharing
Knowledge across Organization Subunits,” Administrative Science Quarterly, Vol.
148
44, i1, p. 82, 1999.
33. Hogland, D., A Parametric Model to Generate Subsystem Constitutive Laws for a
Vehicle Ride Model, 2000, M.S. Thesis, The University of Michigan.
34. Huebner, K. H. and Thornron, E. A., The finite element method for engineers, 1982,
Wiley, New York.
35. IEEE, “IEEE Standard for Application and Management of the Systems Engineering
Process,” IEEE Std 1220-1998.
37. Kikuchi, N. and Malen, D., Automotive Body Structure Course Note, 1998, The
University of Michigan.
38. Kim, H. M., Michelena, N., Papalambros, P., and Jiang, T., “Target Cascading in
Optimal System Design,” Proceedings of DETC 2000: 26th Design Automation
Conference, DETC2000/DAC-14265, 2000.
39. Kirsch, U., “Multilevel Approach to Optimum Structural Design,” 1975, Journal of
the Structural Division, Proceedings of the American Society of Civil Engineers,
Vol. 101, No. ST4, pp. 956 ~ 974.
40. Kirsch, U., “An Improved Multilevel Structural Synthesis Method,” 1985, J. Struct.
Mech., Vol. 13, No. 2, pp. 123 ~ 144.
41. Kroo, I., Altus, S., and Braun, R., “Multidisciplinary Optimization Methods for
Aircraft Preliminary Design,” 1994, AIAA Paper 94-4325-CP.
42. Kroo, I., “MDO for Large-Scale Design,” Proceedings of the ICASE/NASA Langley
Workshop on Multidisciplinary Design Optimization, N. Alexandrov and M.
Hussaini (eds.), SIAM, Hampton, VA, March 13-16, 1995.
149
43. Krishnamachari, R. S. and Papalambros, P., “Optimal Hierarchical Decomposition
Synthesis Using Integer Programming,” 1997, Trans. ASME Journal of
Mechanical Design, Vol. 119, No. 4, pp. 440-447.
44. Locke, E. A. and Latham, G. P, A theory of goal setting and task performance, 1990,
Prentice Hall.
47. Malmborg, C. J. and Simons, G. R., “Integrating logistical and processing functions
through mathematical modelling: A case study,” 1989, Appl. Math. Modelling,
Vol. 13, pp. 357 ~ 364.
48. Michenela, N., Kim, H. M., and Papalambros, P. Y., “A System Partitioning and
Optimization Approach to Target Cascading,” 1999, Proceedings of 12th
International Conference on Engineering Design (ICED), pp. 1109 ~ 1112.
50. Michelena, N., Papalambros, P., Park, H. A., Kulkarni, D., “Hierarchical Overlapping
Coordination for Large-Scale Optimization by Decomposition,” 1999, AIAA
Journal, Vol. 37, No. 7, pp. 890-896.
51. Miettinen, K., “Review of Nonlinear MCDM Methods,” 1997, 6th International
school of MCDA, Turku, Finland.
52. Mousseau, C. W., Karamihas, S. M, Gillespie, T. D., Sayer, M., Stevenson, L. A.,
Johnson, T. A. and Brown, A. M., “Computer Synthesis of Light Truck Ride Using
150
a PC Based Simulation Program,” 1999, SAE Paper No. 1999-01-1976.
53. Nelson, S., Optimal Hierarchical System Design via Sequentially Decomposed
Programming, 1997, Ph.D. Dissertation, The University of Michigan.
55. Nunamaker, J., Applegate, L., and Konsynski, B., “Computer-Aided Deliberation:
Model Management and Group Decision Support,” 1988, Operations Research,
Vol. 36, No. 6, pp. 826 ~ 848.
56. Palli, A., Azarm, S., McCluskey, P., and Sundararajan, R, “An Interactive Multistage
H-inequality constraint method for multiple objective decision making,” 1998,
Trans. ASME Journal of Mechanical Design, Vol. 120, No. 4, pp. 678 ~ 686.
57. Pan, J. and Diaz, A. R., “Some Results in Optimization of Nonhierarchic Systems,”
1990, ASME Journal of Mechanical Design, vol. 112 pp. 399 ~ 405.
58. Papalambros, P. and Wilde, D., Principles of Optimal Design: Modeling and
Computation (1st Ed.), 1988, Cambridge University Press, New York.
59. Park, A., Michelena, N., Kulkarni, D., Papalambros, P., “Convergence Criteria for
Hierarchical Overlapping Coordination of Linearly Constrained Convex
Problems,” Computational Optimization and Applications (accepted for
publication).
60. Powell, M. J. D., “The Convergence of Variable Metric Methods for Nonlinear
Constrained Optimization Calculations,” 1978, Nonlinear Programming 3,
Academic Press, pp. 27-64.
151
Conceptual Design,” 1998, Proceedings of the 7th AIAA/NASA/ISSMO Symposium
on Multidisciplinary Analysis and Optimization, AIAA-98-4919.
66. Saigal, R., Continuous Optimization, 1997, Course pack, University of Michigan.
67. Shigley, J. E. and Mischke, C. R., Mechanical Engineering Design (5th Ed.), 1989,
McGraw-Hill Book Co.
69. Shimizu, K. and Aiyoshi, E., “Hierarchical multi-objective decision systems and
power-decentralized systems for general resource allocation problem,” 1980, Keio
Engineering Reports, Vol. 33, No. 2, pp. 13 ~ 29.
70. Shishko, R., NASA Systems Engineering Handbook, 1995, NASA SP-6105.
72. Sobieski, I., Manning, V., and Kroo, I., “Response Surface Estimation and Refinement
in Collaborative Optimization,” 1998, Proceedings of the 7th AIAA/NASA/ISSMO
Symposium on Multidisciplinary Analysis and Optimization, AIAA-98-4753.
73. Sobieszczanski-Sobieski, J., James, B. B., and Dovi, A. R., “Structural Optimization
by multilevel decomposition,” 1985, AIAA Journal, vol. 21, pp. 1291 ~ 1299.
152
74. Sobieszczanski-Sobieski, J., “Optimization by Decomposition: A step from hierarchic
to nonhierarchic systems,” 1988, Recent Advances in Multidisciplinary Analysis
and Optimization, NASA CP-3031, Hampton, VA.
75. Sobieszczanski-Sobieski, J., “Two alternative ways for solving the coordination
problem in multilevel optimization,” 1993, Structural Optimization, vol. 6, pp. 205
~ 215.
76. Sobieski, J., James, B, and Riley, M., “Structural Sizing by Generalized, Multilevel
Optimization,” 1987, AIAA Journal, Vol. 25, No. 1, pp. 139~145.
77. Sriram, D., Logcher, R., Wong, A., and Ahmed, S., “A Case Study in Computer-Aided
Cooperative Product Development,” 1990, Technical Report No: IESL-90-01,
Department of Civil Engineering, Massachusetts Institute of Technology.
78. Stefik, M., Foster, G., Bobrow, D., Kahn, K., Lanning, S., and Suchman, L., “Beyond
the Chalkboard: Computer Support for Collaboration and Problem Solving in
Meetings,” 1987, Communications of the ACM, Vol. 30, No. 1, pp. 32 ~ 47.
79. Steuer, R. E. and Choo, E.-U., “An interactive weighted Tchebycheff procedure for
multiple objective programming,” 1983, Mathematical Programming, Vol. 26, No.
3, pp. 326 ~ 344.
80. Stressing, J., “System-level design tools,” April 1989, Computer-Aided Engineering
Journal, pp. 44 ~ 48.
81. Tang, C. S. and Yoo, S.-C., “System planning and configuration problems for optimal
system design,” 1991, European Journal of Operational Research, Vol. 54, pp.
163 ~ 175.
82. Tappeta, R. and Renaud, J., “Multiobjective Collaborative Optimization,” 1997, Trans.
ASME Journal of Mechanical Design, Vol. 119, No. 3, pp. 403 ~ 411.
83. Thareja, R. and Haftka, R. T., “Numerical difficulties associated with using equality
constraints to achieve multi-level decomposition in structural optimization,” 1986,
153
AIAA paper No. 86-0854.
84. Thareja, R. and Haftka, R. T., “Efficient single-level solution of hierarchical problems
in structural optimization,” 1990, AIAA Journal, Vol. 28, No. 3, pp. 506 ~ 514.
85. Ulrich, K. and Eppinger, S., Product Design and Development, McGraw-Hill, 1995.
87. Wagner, T.C., A General Decomposition Methodology for Optimal System Design,
1993, Doctoral Dissertation, The University of Michigan.
89. Wilson, R. B., “A simplicial algorithm for concave programming,” 1963, Doctoral
Dissertation, Harvard University.
90. Wismer, D. and Chattergy, R., Introduction to Nonlinear Programming, 1978, North-
Holland, New York.
91. Wong, A., Sriram, D., and Logcher, R., “User Interfaces for Cooperative Product
Development,” 1990, Research in Engineering Design.
92. Wong, J. Y., Theory of Ground Vehicles (2nd Ed.), 1993, John Wiley & Sons Inc.
93. Yuan, Y., “Conditions for Convergence of Trust Region Algorithms for Nonsmooth
Optimization,” 1985, Mathematical Programming, Vol. 31, pp. 220 ~ 228.
154