Using Predictive Modeling Techniques

to Solve Multilevel Systems Design Problems

Richard J. Malak, Jr.1 and Edgar Galvan.2

Texas A&M University, College Station, TX, 77843

A challenging aspect of systems design is the search for a combination of concepts for

system components that together yield desirable system-level characteristics. These system-

level characteristics depend not only on the concepts designers choose, but also the details of

how they implement them. This yields a challenging search problem through a

heterogeneous and discontinuous space of system alternatives. Although designers can use

design optimization methods for this task, they can be slow because they entail solving a

different optimization problem for every valid combination of component-level concepts. In

this paper, we present a new approach to systems design based on the use of abstract

predictive models. Under this approach, designers abstract multiple physically

heterogeneous component-level concepts into a unified model that captures the salient

characteristics of the possible implementations of each concept. This enables them to search

the space of system alternatives quickly and effectively. We demonstrate the new approach

on a utility cart system design problem and compare it to optimization-based approaches

common in the literature. The new approach yields a system design as desirable as the one

we obtain from the best-performing optimization-based approach, but in less than one-tenth

the computational time.

1
Assistant Professor, Department of Mechanical Engineering, 3123 TAMU, College Station, TX 7743-3123.
2
Graduate Research Assistant, Department of Mechanical Engineering, 3123 TAMU, College Station, TX 7743-
3123.
1
American Institute of Aeronautics and Astronautics
 Nomenclature

b = damping coefficient

d = helical spring wire diameter

D = helical spring coil diameter

l = leaf spring length

a = leaf spring width

h = leaf spring thickness

Di = diameter of the gear

w = gear width

k = spring constant

Ng1 = first gear ratio

Ng2 = second gear ratio

Ng3 = third gear ratio

Rsusp = suspension reliability

Rtrans = transmission reliability

Csusp = suspension cost

Ctrans = transmission cost

dtow = distance towed

R = overall reliability

C = overall cost

tr = rise time

ts = settling time

ymax = maximum vehicle displacement

m = total vehicle mass

= max torque

= engine speed at max torque

= number of active helical spring coils

= number of leaves
2
American Institute of Aeronautics and Astronautics
 I. Introduction

URRENT trends are toward engineered systems with increased functionality, larger numbers of interacting
C components, and elevated expectations about cost, reliability, sustainability, and performance. To achieve these

high expectations, system designers must explore a broad space of alternatives to locate the few exceptional

solutions despite strict limitations on the time and design resources available. System designers often use

optimization methods to assist in this task. These methods are particularly useful for searching a well-defined

parameter space for the best set of parameter values. However, the problem of designing an engineered system—as

opposed to the problem of refining a system—is not so well defined. To be successful, designers require methods

that enable them to search the heterogeneous space of alternatives that occur in systems design problems.

The focus of this paper is on the point in a design process at which designers have determined the system

organization—i.e., how components interact to produce system-level functionality and behavior—but have not yet

determined how to implement the components physically. For example, designers may have settled on an aircraft

with two wings and two wing-mounted engines, but have not yet determined the specific type of engine, how to

actuate control surfaces, etc. Although designers have determined a system decomposition and hierarchy at this

point in a design process, considerable design freedom remains and it is not straightforward for them to apply

optimization methods to this problem. The challenge lies in the heterogeneous nature of the search space. Each

arrangement of component-level concepts represents a distinct search space, with no regularity between each search

space. Moreover, because each concept can involve different technologies and operate on different principles, there

exists no unified parametric description of the design alternatives. Consequently, a straightforward application of

optimization methods entails an optimization run for each combination of component-level concepts. For a complex

system, the number of valid combinations of component concepts can be unmanageably large. Introducing

additional concepts results in a combinatorial explosion in the number of valid system configurations. Technological

incompatibilities may mean the number of valid combinations is lower than the worst-case number. However, even

a dozen valid combinations can be intractable if it takes a large amount of design resources to evaluate each one.

System designers would benefit from an approach that allows for extensive search of the space of system

alternatives despite the challenge presented by heterogeneous search space.

In this paper, we demonstrate the use of predictive modeling in combination with optimization methods to

support systems design and compare this approach to those based strictly on optimization methods. Predictive

3
American Institute of Aeronautics and Astronautics
models capture associations between variables but do not have explanatory power or indicate causality 1. This makes

them promising as a mechanism for dealing with the mixed discrete-continuous optimization problem. As we

demonstrate, designers can abstract the key properties of many implementations of a component into a single

predictive model that they can use to perform system-level design exploration via optimization methods. The

system-level optimization problem yields the identity of the best concept and target specifications for designing it in

detail. Designers can follow this system-level problem by optimizing the components to these targets. Unlike classic

multilevel optimization procedures, the process does not require iteration between levels. This is because the

knowledge designers formalize into the predictive models ensures that the component target specifications are

technically feasible. Prior investigation into elements of this approach have proved promising 2-4, but the potential

benefits of abstracting physically heterogeneous concepts into a single model has not been studied. As we

demonstrate in this paper, this capability leads to gains in computational efficiency without sacrifices in solution

quality.

In the next section, we formulate the mixed discrete-continuous systems design problem that is the focus of the

paper. In section III, we describe how designers would solve this problem using two existing optimization-based
5, 6
approaches: an all-at-once optimization formulation and a formulation based on analytical target cascading .

Section IV is a description of the approach based on the use of predictive models to describe the capabilities of

several implementation concepts for a particular component. In section V we present a system design example on

which we compare the three approaches. Section VI contains an analysis and discussion of the comparison results.

II. Design Problem Formulation

Systems design is the process of transforming a problem statement into a detailed description of a system
7, 8
capable of solving the stated problem . The full systems design process involves many steps, both qualitative

(planning, problem clarification, alternative generation, etc.) and quantitative (engineering analysis, optimization,

etc.). In this paper we focus on the problem of designing the components of a fixed system hierarchy. Although the

fixed hierarchy imposes constraints on each component, system designers retain considerable design freedom at this

point in a project.

Once designers have determined the system hierarchy, they must (1) determine the best concept for each

component in the hierarchy and (2) design each of the chosen component concepts in detail. These sub-problems are

4
American Institute of Aeronautics and Astronautics
dependent because the best implementation designers can Maximize:

achieve for a given concept will influence which With respect to:
 Component design concepts, for
arrangement of concepts is best overall. Figure 1 is a  Component design variables, for
Subject to:
mathematical formulation of the overall problem. It  Design variable bounds,
o for
consists of models at three levels of abstraction: a utility o for
 Engineering constraints,
model, system-level analysis models, and component- o for

level analysis models. o for

Where:
Utility Model. To be compatible with optimization  is the number of design variables for the
concept of the component
algorithms, designers must formalize their decision-  is the number of concepts for the component
 is a vector of system attributes
making preferences in a computer-interpretable form. A  for is a system attribute,
computed from component attributes
utility function, denoted , is a scalar function that  is a system-level analysis model
 is a vector of component
relates system-level attributes to a utility value that attributes
 for is a vector of
designers seek to maximize. System attributes are attributes for the component
 is a component-level analysis model
properties or measures of effectiveness that designers  is a vector of design variables for the
concept of the component
consider when determining relative preference among

different system implementations. For an automobile, Figure 1. System design formulation for a fixed
system hierarchy.
system-level attributes might include fuel economy, top speed, and production cost. Designers define a utility

function such that attribute vectors that are more preferred lead to larger utilities. There are multiple approaches by

which designers can do this. One approach is to focus on profit maximization as an objective, in which case a
5, 9
designer’s utility function computes profit and may include an adjustment to capture the designer’s risk attitude .

In this case, designers may consider also to be a function of enterprise-level design variables, such as

production quantities, product sales price, etc. Under multi-attribute utility theory (MAUT), one elicits a utility
10
function by answering a series of questions involving hypothetical choices involving lotteries . We use MAUT in

the example of Section V.

System-Level Analysis Models. Designers compute system attributes as a function of component attributes

using system-level analysis models, denoted for where is the number of attributes designers

use to describe the system in question. Each system-level analysis model takes one or more component-level

attributes as inputs. Each model may involve different formalisms as is appropriate for computing the attribute of

5
American Institute of Aeronautics and Astronautics
interest and in general is not a closed-form expression. For example, designers seeking to use fuel economy as a

system attribute may require a dynamic simulation of a system that incorporates energy losses and the impacts of

time-varying operating conditions.

In the interest of notational simplification, we state the computation of a system-level attribute as

for , where is a vector comprised of all component-level attribute values. However, we do not intend

this to imply that a particular system-level attribute is influenced by every component-level attribute.

In general, designers can require different system-level models depending on which components they consider.

For example, this occurs when the components induce system behaviors with dynamics of different orders or

involve incompatible physical phenomena. For example, designers might use different models to describe the

behavior of a turboprop aircraft and one with jet engines. However, in the interest of scope, we limit the current

paper to problems in which the same system-level models apply for all component concepts.

Component-Level Analysis Models. We denote component-level analysis models as for

, where is the number of components in the system and is an integer identifying which of

the concepts for the component designers are considering. The concept-specific indexing is necessary because

different implementation approaches can require different analysis models. Furthermore, different concepts

generally have different design variables. Thus, we denote the design vector for concept of component as .

Note however that we model all concepts for a given component as having the same attribute vector, . This is

because the focus of this research is on the point in a design project at which designers already have fixed the system

hierarchy. The hierarchy establishes an interface to the system for each component and therefore implies the

attributes that are important for system-level modeling.

III. Optimization-based Solution Approaches

In the following, we consider how designers might formulate the design problem of Figure 1 using solution

procedures based on optimization methods. In particular, we examine two approaches common in the literature: (1)

an integrated formulation, which we refer to as All-at-Once (AAO) optimization and (2) an approach for

hierarchically decomposed optimization, called Analytical Target Cascading (ATC).

6
American Institute of Aeronautics and Astronautics
A. All-at-Once Optimization
All-At-Once Approach
Under an AAO optimization approach, designers
Utility Maximization
Problem
formulate the design problem as a single optimization
z

problem that parallels the formulation of Figure 1. We System Analysis

depict this in Figure 2. The main practical consideration is y1 y2 ym

Component 1, Component 2, Component m,

that designers must repeat the AAO method for every Concept c1 Concept c2 Concept cm
Analysis Analysis … Analysis
combination of component concepts. Thus, this is a multi-
x1,c1 x2,c2 xm,cm
stage problem: (1) iterate through all combinations of

component concepts to find the best implementation of Figure 2. Schematic of All-at-Once (AAO)
approach for a specific selection of component
each concept relative to the system-level utility function concepts.

and (2) select the combination with the highest utility overall. Thus, the first stage involves classical nonlinear

optimization methods using the AAO framework and the second stage is a simple selection from a list of

alternatives.

B. Multilevel Optimization using Analytical Target Cascading

In an effort to manage complexity in a

Analytical Target Cascading Approach
systems design process, designers
Utility Maximization
sometimes decompose an optimization Problem

ztarget
problem in a way that reflects the z
System-Level
ATC System-Level Problem Analysis
y
organization of engineering expertise on a
y1,target y1,actual y2,target y2,actual ym,target ym,actual
design project. Designers can decompose a

problem among various collaborators ATC Component-

Level Problem
ATC Component-
Level Problem
… ATC Component-
Level Problem
11-13
along disciplinary boundaries , or
x1,c1 y1 x2,c2 y2 xm,cm ym
6, 14
according to system structure . We
Component 1, Component 2, Component m,
Concept c1 Concept c2 … Concept cm
focus in this paper on an approach called Analysis Analysis Analysis

Analytical Target Cascading (ATC)

because it follows a hierarchical system Figure 3. Schematic of Analytical Target Cascading (ATC)
based approach for a specified selection of component concepts.
structure 6, 14.

7
American Institute of Aeronautics and Astronautics
 One challenge in applying ATC to the design problem defined in Figure 1 is that ATC is based on the notion of

target achievement rather than utility maximization. Researchers have demonstrated the use of ATC together with
5, 9
utility functions in the context of enterprise-driven optimization using on customer demand models . We are

unaware of prior application of ATC in concert with MAUT.

Like the AAO approach, ATC requires a two-phase strategy in which one first optimizes the system for every

combination of component concepts and then chooses the best overall combination. The distinction is that under

ATC, designers decompose the optimization of a particular combination of component concepts according to the

system hierarchy. Figure 3 is a schematic of the ATC-based approach for one selection of component concepts.

Designers must re-execute this for each combination of concepts. The utility maximization problem yields a target

system-level attribute vector, , that serves as an input to the ATC sub-problem. The final solution, the output

of the ATC procedure, is the one that minimizes deviation from this target. This procedure for coordinating the

utility-level and ATC problems is similar to one demonstrated previously in the literature5, 9, but is simplified due to

the fact that there are no enterprise-level variables in the current problem.

C. Limitations

These optimization-based approaches share one significant limitation: they search the heterogeneous space of

systems by enumerating all combinations of component concepts. This limitation is not severe if there are few

combinations or if each combination is very fast to evaluate. However, complex systems can involve dozens or

hundreds of components. Even with just a couple concepts for each component, the number of combinations can be

unreasonable. Add to this the likelihood of complex engineering analyses, such as computational fluid dynamics,

finite element modeling, and system dynamics simulations, and one can conclude that for most systems the

computational burden per combination is heavy. Parallelization of the optimization runs can help reduce the time

impact of the problem (each combination is an independent optimization problem), but this comes at an expense in

terms of computing resources and still may be insufficient to make a full search practical.

It is straightforward to compute a worst-case bound on the number of combinations designers need to evaluate. If

there are components in a system, denoted , and each component has associated with it physically

heterogeneous implementation concepts, then the upper bound on the number of optimization runs is

8
American Institute of Aeronautics and Astronautics
 . This upper bound can grow fairly quickly as the numbers of components and component concepts grows.

Introducing just one new component concept can increase the computational burden by a factor of .

In practice, this upper bound is likely to be fairly conservative given that technical incompatibilities may

preclude certain combinations of component concepts. For example, one cannot power a DC motor directly with

hydraulic fluid and therefore designers would not try to optimize a system involving a DC motor concept for an

actuator component and a fluid-power concept for power transmission. Nonetheless, technical incompatibilities

alone may not eliminate so many combinations that computation becomes trivial. System designers would benefit

from a better approach to this problem.

IV. Combining Optimization with Abstract Predictive Techniques

Predictive modeling techniques can be useful for improving upon existing optimization methods in systems

design. Predictive models allow a user to predict unknown values of one variable as a function of the others. They
1, 15
imply nothing about causality and have no explanatory power . Construed broadly, predictive modeling includes

response surface models based on computer experiments, a practice sometimes called meta-modeling. Several
16-19
authors report computational gains when applying this predictive approach to optimization problems . However,

although this approach can reduce the computational burden of a single optimization run, it does not solve the more

fundamental problem of potentially having an excessive number of combinations to optimize.

We consider a different approach for utilizing predictive modeling in which system designers abstract many

component concepts into a single predictive model. This enables designers to recast the heterogeneous systems

design problem into a homogeneous search space. This type of abstraction is not possible in every case (see the

discussion in Section VI), but it can yield considerable reductions in the numbers of independent optimization runs

that designers require. If designers can abstract as few as two concepts into a single model they can reduce the

required number of optimization runs for a system with components by a factor of .

The Abstract Predictive (AP) approach involves a three-step process of (1) abstraction and predictive modeling,

(2) predictive system-level optimization, and (3) component optimization. Figure is a schematic of this approach.

Unlike the illustrations of the AAO and ATC approaches (Figures 2 and 3, respectively), the procedure of Figure is

sufficient to explore all combinations of concepts without iteration.

9
American Institute of Aeronautics and Astronautics
 Abstract Predictive Approach
(1) Abstraction & Predictive Modeling Phase

Component 1
Concepts (2) Predictive System-Level Optimization Phase
A&PM
Concept
Concept
Concept
Analysis
Procedure
Analysis AP Component Model 1 y1
Analysis z
y2 Utility Maximization System-Level
A&PM AP Component Model 2 Analysis
Problem y
Component 2 Procedure … ym
Concepts
AP Component Model m y1,target y2,target yn,target
Concept
Concept
Concept
Analysis
Analysis
Analysis A&PM
… Procedure Component-Level
Optimization
Component-Level
Optimization … Component-Level
Optimization

Component m x1,c1 y1 x2,c2 y2 xn,cn yn

Concepts

Concept
Concept
Concept
Component 1,
Best Concept
Component 2,
Best Concept
… Component m,
Best Concept
Analysis
Analysis
Analysis Analysis Analysis Analysis

(3) Component Optimization Phase

Figure 5. Schematic of Abstract Predictive (AP) approach.

A. Abstraction and Predictive Model Generation

2, 3
The modeling approach is based on prior work involving predictive modeling in conceptual design . The prior

work involved only one concept per model and relied on data about prior implementations of a particular concept

(e.g., catalog and spec sheet data) rather than drawing samples from an engineering model. Here, we extend the

approach to the current problem.

Figure 4 is a summary of the abstraction and

1. Identify component attributes and each as a dominator or a
parameter
predictive modeling procedure. In Step 1, 2. For each concept
i. Generate samples of design variables
designers identify the principal abstraction of the ii. Run component-level analysis model on samples to
generate attribute data
component by identifying attributes of the iii. Filter attribute data using parameterized Pareto dominance
iv. Fit a model to the non-dominated attribute data
component—i.e., measurable properties of the 3. Abstraction procedure
i. Generate samples of each concept model at the same
locations
component that are inputs to higher-level models
ii. Perform parameterized Pareto dominance on the data
iii. Fit a model to the non-dominated data
or for which decision makers have preferences
Figure 4. Procedure for creating an abstract predictive
directly. In addition, designers must classify each model for a particular component

attribute as a dominator or a parameter attribute. This classification relates to how changes in the attribute value

affect system-level preferences. An attribute is a dominator if it is something that designers generally prefer to
10
American Institute of Aeronautics and Astronautics
maximize or minimize, assuming all other factors are equal. Definition (Parameterized Pareto Dominance): An
alternative having attributes is parametrically Pareto
Cost, reliability, lifetime, environmental impact, and fuel
dominated by one with attributes if
, and , where is
efficiency are examples of dominator attributes. All other
the set of parameter attribute indexes and is the (non-
attributes are classified as parameter attributes. These are empty) set of dominator attribute indexes.

attributes with no problem-independent

preference Figure 6. Definition of parameterized Pareto
dominance.
association. For example, maximizing the bore of a

hydraulic cylinder is desirable in some applications (those for which power output is paramount), but the opposite is

preferred in other situations (when speed is paramount).

In Step 2, designers generate individual predictive models for each concept. The third sub-step—data filtering

using parameterized Pareto dominance—is particularly important when designers use samples from analysis models

as their data source. Dominance-based filtering eliminates data corresponding to provably bad solutions—ones that a

designer never would implement. The result is a more accurate predictive relationship among the component-level

attributes. For example, suppose designers seek to predict production cost as a function of efficiency and maximum

power output for an electric motor and they have the data points ($75, 0.9, 750 W) and ($300, 0.9, 750 W). To

include the second point during fitting would yield overly pessimistic predictions about what production costs are

achievable at particular levels of efficiency and power output.

Parameterized Pareto dominance is an extension of classical Pareto dominance to the situation in which some

attributes are not dominators2. This is necessary to support dominance analysis in a multilevel systems context.

Figure 6 contains the definition of parameterized Pareto dominance. It is mathematically sound in that it will not

eliminate any solution that could be part of the optimal system4.

Step 3 is the extension to support abstraction across component concepts. The procedure involves a second

application of parameterized Pareto dominance because different concepts can dominate in different regions of the

attribute space. This procedure yields the overall non-dominated set that can be the basis for an accurate predictive

model.

11
American Institute of Aeronautics and Astronautics
B. Predictive System-Level Optimization

Designers use the predictive models of components to

Find:
replace the component-level analysis models during

optimization. Let denote the abstract predictive model Subject to:



associated with the component. Also let denote


the component-level attribute to be predicted, denote
(a)
the vector of attributes used as independent variables in the
For each component, :
predictive model and such that , and Find:

denote the valid domain of values. Given this notation,

Subject to:

one can formulate the predictive system-level optimization

problem as in Figure 7(a). Designers need to execute this
(b)
optimization only once.
Figure 7. Mathematical definition of (a)
One can interpret the output of this as a prediction,
, predictive system-level optimization problem to
find component-level attribute targets and (b)
about the best component-level attribute values designers component-level optimization.

can achieve when designing to the objectives formalized in the utility function, . If the predictive models are

perfectly accurate, then this prediction corresponds to the optimal attribute values.

It is straightforward to recover the best concept for a given component based upon the attribute predictions. Let

denote the predictive model for the concept for the component and denote the

attribute values predicted to be optimal for this component. Then the best concept is the one that minimizes

. If multiple concepts yield very similar results, designers can choose from among them

arbitrarily.

C. Component Optimization

Given target attribute values for a component and the identity of the best concept for implementing that

component, designers can proceed to develop each component independently by optimizing to the specified targets.

This is a straightforward application of engineering optimization techniques. One optimization run is required for

each component. Figure 7(b) is a summary of this problem

12
American Institute of Aeronautics and Astronautics
 V. Engineering Example: Design of a Utility Cart

A. Design Objectives and Problem Formulation

1. Design Problem Scenario and System-Level Modeling

We use a utility cart (UC) design problem to demonstrate the AP approach to systems design and to compare it

with the AAO and ATC approaches. A utility cart is a small, four-wheeled vehicle commonly used by grounds

keepers and maintenance staff. Our focus is on the design methods and their comparisons rather than the design

results per se. In the design scenario, we assume designers already have determined the UC system architecture, but

have not yet determined how to implement its suspension and its transmission. The UC seats two people and has

room to carry equipment and supplies as well as the capability to tow loads. It is powered by an internal-combustion

engine that is connected to the rear drive wheels through a three-speed transmission whose implementation has not

yet been determined. Table 1 is a listing of the environmental variables and fixed design parameters associated with

the utility cart design problem.

There are four principal system-level design attributes of interest in this problem: cost, reliability, towing

performance, and ride quality.

System Cost. We model system cost as the sum of Table 1. Environmental variables and static design
variables relevant to the design problem.
component costs plus a fixed assembly cost. We Parameter Value

assume the assembly cost is independent of the System and Environment

Vehicle Curb Weight, 650 (kg)
component implementations.
Hill incline
System Reliability. We model system reliability Towed Mass, 500 (kg)
Max Engine Torque, 125 (Nm)
as the product of component reliabilities. No Engine Speed at Max Torque, 4000 (RPM)

redundancies exist in the system, so a failure of any Transmission

Application Factor, 1.7
component implies a failure of the system. We Gear Quality Factor, 8
Bending Strength Geometry Factor, 0.24
evaluate component reliabilities using static failure
Bending Fatigue Strength, 200 (MPa)
models under operating conditions representative of Differential ratio, 4

Suspension
anticipated operational scenarios.
Active Helical Spring Coils, 8
Towing Performance. We evaluate towing Number of Leaves, 5

performance using a test scenario of the utility cart towing a 500 kg mass up a 20° incline. The instantaneous

13
American Institute of Aeronautics and Astronautics
acceleration is found at each time step using a model of vehicle dynamics and the engine power curve given the

current wheel speed and gear ratio. When the engine reaches a predetermined speed the model shifts to a higher gear

if it is available. If no higher gear is available the utility cart stops accelerating. The simulation ends after sixty

seconds and outputs the final distance traveled.

Ride Quality. We evaluate ride quality using the time response of the UC system, modeled using a quarter-car

model (Figure 8), to a typical speed bump at 15 mph (approx. 24 kph). We consider ride quality to be an aggregate

of three system behavioral attributes: settling time, vertical displacement, and rise time.
x
Objectives are to minimize settling time and vertical displacement, and to maximize rise m

time. Tradeoffs among these attributes are determined using a utility function (described k b

below). The quarter-car model enables us to relate suspension component attributes and
u(t)
system-level attributes. The vehicle is modeled as spring mass damper system that
Figure 8. Quarter
ignores the effects of the tire’s mass and springiness. The resulting transfer function is Car Model.

where is the damping coefficient is the spring constant, and is the vehicle mass. The input is a half-

cycle sinusoid that represents the speed bump. The rise time, settling time and vertical displacement are determined

from the model output. The displacement information is also used to determine the spring reliability.

2. Utility Model

We use techniques from the MAUT literature to determine a utility function for the UC system. The overall

utility function is:

where is the utility function for the settling time attribute at , is for the rise time

attribute, is for the vertical displacement attribute, is for the

tow distance attribute, is for the reliability attribute, is for the cost attribute, and is the

system level attribute vector. The design problem is to find the component-level attribute values that maximize

system utility, .

14
American Institute of Aeronautics and Astronautics
3. Component Concepts

In the interest of scope, we limit the example to the design of the transmission and suspension components. We

consider three physically heterogeneous suspension concepts and two for the transmission (forward speeds only),

yielding six unique combinations of concepts.

Transmission. We consider two three-speed transmission concepts:

TO
DIFFERENTIAL

FROM
3-Speed 8-Gear (T3-8): Figure 9(a). Basic three speed transmission ENGINE

system with 8 gears total. The engine shaft turns the four layshaft LAY SHAFT

(a)
gears. The three remaining gears on the layshaft turn the gears

corresponding to each speed. Design variables consist of the widths

TO
DIFFERENTIAL
FROM
and diameters of the eight gears. ENGINE

(b)
 3-Speed 6-Gear (T3-6): Figure 9(b). Simplified three speed
Figure 9. Layout of both
transmission system with six gears total. The engine shaft has three transmission concepts: (a) 3-
speed 8-gear transmission; (b)
gears which turn the gears corresponding to each speed. In this 3-speed 6-gear transmission
concept there is no layshaft, reducing cost at the expense of reducing the compactness of the transmission.

Design variables consist of the widths and diameters of the six gears.

Both transmission concepts are subject to geometric constraints to ensure proper meshing of the gears. Given the

design variable values and operating conditions, we compute five component-level attributes:

 Cost: The sum of material costs and parts. Classified as a dominator attribute (less is better).

 Reliability: The probability that the transmission will function without failure under anticipated operating

conditions. Classified as dominator attribute (more is better).

 Three Gear Ratios: The three available ratios of transformation from transmission input shaft to output shaft.

Classified as a parameter attribute.

Suspension. We consider three concepts for the suspension design. We assume identical suspension designs are

used on all four wheels. All suspension concepts have the same shock absorber configuration.

15
American Institute of Aeronautics and Astronautics
 Helical Spring (HS): Figure 10(a). Helical
To Chassis To Chassis
compression spring of uniform coil diameter. We To Chassis

consider the wire and coil diameters as variables

during the design process.

To Wheel
To Wheel To Wheel
 Leaf Spring (LS): Figure 10(b). Leaf spring with
(a) (b) (c)
rectangular cross-section. We consider the overall leaf
Figure 10. Layout of the three suspension concepts:
(a) helical spring (b) leaf spring (c) combination helical
spring length, and the thickness and width of each
and leaf spring
leaf as design variables.

 Combination Helical and Leaf Spring (HLS): Figure 10(c). Helical and leaf spring systems described above
aaa
combined in parallel.

The suspension concepts are subject to geometric constraints to ensure the design specifications are physically

feasible. There are four attributes for the suspension component:

 Cost: The sum of material costs and parts. Classified as a dominator attribute (less is better).

 Reliability: The probability that suspension function without failure under anticipated operating conditions.

Classified as dominator attribute (more is better).

 Spring Constant: The ratio of the force affecting the spring to the displacement. Classified as a parameter

attribute.

 Damping Coefficient: The ratio of the force affecting the damper to the velocity. Classified as a parameter

attribute.

Component reliabilities, helical

Table 2. Utility cart design problem component-level variables
spring constant, and gear ratios are Component Concept Design Variables Attributes
Suspension Helical Spring
calculated from the design variables Leaf Spring
Helical + Leaf
using standard engineering models20.
Transmission Six gear
The leaf spring constant and reliability Eight gear

are determined using beam theory 21. Transmission and spring costs are computed based on the amount of material

needed to produce their components. Shock absorber cost is an empirical relationship that depends on damping

coefficient. Table 2 is a summary of the design variables relevant to this example.

16
American Institute of Aeronautics and Astronautics
B. Design Approaches for Comparison

1. All-at-Once Formulation

In the AAO optimization approach, an optimizer at the system level searches the component-level design

variable space directly as illustrated in Figure 2. Each combination of component-level concepts has a different set

of design variables and requires different component-level models. Consequently, the entire optimization process is

repeated for each of the six combinations of design concepts. The execution time for each optimization run is

recorded along with the optimization results.

2. Analytical Target Cascading Formulation

The ATC-based optimization approach entails multiple coordinated optimizers as illustrated in Figure 3. The

utility-level problem is formulated as

subject to:

where is the vector of system-level attributes, is the system-level utility function, and and are

system-level constraint functions. This problem is solved once for all concepts and yields a target for the ATC

procedure, .

The ATC problem is partitioned into system-level and component-level problems based on an object partitioning

of the UC system into a system level and two component-level problems (the transmission and the suspension). At

the ATC system level, an optimizer minimizes deviation between the target identified at the utility level, ,

and what is feasible, . This search is conducted over the space of component-level attributes. Component

attribute targets are cascaded down to the component optimizers, which search the space of component design

variables for a particular design concept. Since component-level variables and models are different across the

concepts, designers must repeat the entire process for each combination of components.

3. Abstract Predictive Formulation

The AP approach allows a designer to search the design space through high-level properties of the system

common to all component concepts. We formulate the AP approach according to the steps defined in Figure 4. We

17
American Institute of Aeronautics and Astronautics
 Predictive model of
Analysis attribute relationships
Generate Model for a particular concept
Design
PDOM Kriging e.g.,
Variable e.g., Costsusp  Tsusp ,c1  k , R, b 
Samples d 4G
k
8D3 N

Repeat for each concept of a given component type

Concept 1
Predictive
Model
Predictive model of
attribute relationships
Generate Concept 2 for all concept
Input Predictive
Model e.g.,
Costsusp  Tsusp  k , R, b 
Attribute PDOM Kriging
Samples
…

Concept K
Predictive
Model

Figure 11. Procedure for generating an abstract predictive model for a particular component type.

reported the results of the first step—attribute identification and classification—when describing the component

concepts in the preceding section. The second step—component attribute mapping—involves sampling the

component-level analysis model for each component concept, post-processing the results to eliminate undesirable

data, and fitting a model to the remaining data. Figure 11 is an illustration of this process. Input samples can be

random or generated systematically. A feasibility check is necessary because the sampling procedure may produce

inputs that are physically unachievable. Feasible samples are analyzed to determine the corresponding component-

level attribute data. The attribute data are post-processed using parameterized Pareto dominance (PDOM) to

eliminate points that correspond to designs that are provably inferior from a decision-making perspective. A model

is fit to the remaining non-dominated data. We use kriging techniques and the DACE kriging toolbox for this 22.

The final step is to abstract all concepts for a given component into a single model. This follows a procedure

similar to Step 2, except that one samples the previously fitted models rather than detailed engineering models. Also,

to ensure a proper comparison of the models, it is helpful to sample all concept models at the same input

combinations. One then applies PDOM to the attribute data and fits a model to the resulting non-dominated set. We

also use kriging for this model.

18
American Institute of Aeronautics and Astronautics
 With the abstract predictive models in place, designers can use them in place of the component level analysis.

Following Figure 7, the design problem becomes

subject to:

where

Constraints and are imposed on the attributes by the physical constraints of the

components. For example, it is not possible to have a spring constant less than zero. They also serve to ensure the

validity of inputs to the predictive models (e.g., to ensure that inputs are not from parts of the attribute space for

which no data was collected and for which the model may be a poor fit). The output of this optimization

problem, , are the target attribute values for the components during detailed design optimization (c.f., process flow

in Figure ). However, before designing the components in detail, one must determine which concept is best for each

component.

Designers can identify the best concept through a straightforward procedure of comparing predictions to the

component-specific predictive models to the target attribute vector. The best concept for the component is the

one that minimizes where is the element of that corresponds to the attribute

predicted by the component attribute prediction models. For example, for the suspension, one would evaluate

for .

Once designers have identified the best concept for a component, they can use standard engineering optimization

techniques to identify design variable settings (c.f., Figure ). One can formulate this as a target-achievement type

problem, where the design objective is to minimize deviation for the target attribute vector for that component. For

19
American Institute of Aeronautics and Astronautics
the component, designers can formulate this as , where is

the analysis model and is the vector of component-level design variables, and is the best concept.

C. Optimization Results

Table 3 contains the results from the UC design Table 3. Results for utility cart design problem using three
approaches. The most preferred design in each case is the
problem. The data includes the final system-level T3-6 transmission combined with the leaf spring
suspension concept.
and component-level attributes, computation time
AAO ATC AP
and utility value of the most preferred design for Utility 0.841 0.664 0.808
System-Level Attributes
each of the three solution approaches. In each case, Cost ($US) 901 365 1080
Reliability 0.996 0.949 0.997
Distance Towed (m) 256 56 337
the system using the 3-speed 6-gear transmission Settling Time (s) 1.86 1.03 2.03
Rise Time (s) 0.55 0.58 0.54
and leaf spring suspension is the most preferred. Displacement (m) 0.072 0.069 0.060
Suspension (LS)
We measure computational time as the elapsed Spring Constant, (N/m) 6752 1194 4113
Damping, (Ns/m) 2249 1670 2045
time for all computational activities related to each Reliability, 0..996 0.998 0.997
Cost, ($US) 178 56 119
approach. These times are measured on a mid- Transmission (T3-6)
Gear Ratio 1, 1.66 1.66 2.02
range PC workstation with a 2.93 GHz clock speed. Gear Ratio 2, 1.32 1.03 1.89
Gear Ratio 3, 0.99 1.05 0.82
Although run times will vary depending on Reliability, 0.999 0.995 0.999
Cost, ($US) 723 310 961
hardware and software configurations, one can Total Computation Time (s) 13659 5789 1331
Speedup Factor (rel. to AAO) 1 2.35 10.3
expect the relative results to hold across platforms. Pct Utility Difference w/ AA0 0 21% 4%

For the AAO approach, the computational time we report is the time required to optimize each of the six

combinations of concepts. For the ATC approach, this is the time to identify a target attribute vector at the utility

level plus the time to apply ATC to each of the six

Table 4. Computational time by step for AAO
concept combinations (the same target applies to all and ATC approaches.

Computational Time (s)

concepts, so this portion of the time is counted only AAO ATC
Utility Level Optimization n/a <1
once). Table 4 is a breakdown of the computational time Concept Combinations:
Transmission Suspension
for each operation in the AAO and ATC approaches. T3-8 HS 2485 1196
T3-8 LS 1754 1006
Optimization times for individual concepts ranged from T3-8 HLS 1590 785
T3-6 HS 2969 885
785 s (about 13 min) to 2969 s (about 50 min). T3-6 LS 2235 810
T3-6 HLS 2626 1107
The AP approach involves a greater number of steps, Total 13659 5789

20
American Institute of Aeronautics and Astronautics
but the accounting procedure is similar. Computational time for this approach includes the time to:

1. Sample the design variable space of each of the five component concept models and analyze the samples (i.e.,

compute component-level attributes for each design vector for each of the five component concepts)

2. Perform parameterized Pareto dominance on the component-level attribute data for each for the five component

concepts.

3. Fit models to each of the five parameterized Pareto frontiers

4. Sample each of the five component concept models

5. Perform parameterized Pareto dominance on the new sample data for each component type (i.e., this happens

twice, once for the suspension and once for the transmission)

6. Fit a new model to the parameterized Pareto frontier data for each of the two component types

7. Conduct a system-level optimization using the abstract predictive models to identify the most preferred

component-level attributes

8. Determine which concept was the most preferred for each of the two components

9. Optimize the design of the most preferred concept for each of the two components

Table 5. Computational time by step for the AP approach.

Computational Time (s)
AP Approach
1. Component Concept Sampling and Analysis (all 6 concepts) 18
2. Parameterized Pareto dominance (all 6 concepts) <1
3. Component Concept Model Fitting (all 6 concepts) 1
4. Component Concept Model Sampling (all 6 concepts) <1
5. Parameterized Pareto dominance (both components) 10
6. Abstract Predictive Model fitting (both components) 51
7. Predictive System-level Optimization 524
8. Component Concept Identification <1
9. Component-level Design Optimization (both components) 727
Total 1331

Although there are many steps, most are extremely fast to execute. Table 5 is a breakdown of the execution time for

each step of the AP approach. Only the system-level optimization (step 7) and component-level design optimization

(step 9) took more than a minute. No step took longer than 13 minutes.

Table 3 also contains two important comparison results: (1) the relative speedup of ATC and AP over AAO and

(2) the amount by which the solutions of ATC and AP underperform AAO, measured as the percent difference

between the utilities achieved. The ATC-based approach yields a moderate speedup relative to AAO of a factor of

21
American Institute of Aeronautics and Astronautics
2.35. This means that the ATC approach is a little more than twice as fast as AAO. However, the ATC approach

yields a utility that is 21% off from the AAO approach. In contrast the AP approach yields a utility very close to that

of the AAO approach (4% difference) with a speedup factor of 10.3. This means the AP approach yields an answer

of equal quality to AAO but at a speed that is over ten time faster.

VI. Discussion

For the UC system design problem, the AP approach produced similar results as the AAO approach in less than

one-tenth of the computational time. In contrast, the ATC-based approach yield a solution faster than AAO (in about

half the time), but the solution was significantly lower in quality. When considering computational expense and

solution quality, the AP approach is superior to the others.

As we noted in previously, ATC is not inherently a utility-based method. Although our implementation of ATC

with a utility function parallels those found in prior work, none of them is a mathematically sound procedure that is

guaranteed to converge to an optimal solution (globally or locally). Consequently, the poor solution quality of the

ATC-based result was not unexpected. The problem is that the ATC approach minimizes a weighted distance

between the target system-level attributes identified at the utility level and the actual system attributes that are

physically feasible. However, minimizing this distance is not the same as maximizing overall utility.

Our principal motivation for including ATC in the current analysis is that it possesses a strong structural parallel

with the AP approach—i.e., both follow a problem decomposition based on a system hierarchy. The main

motivation for using ATC is that it supports engineering processes organized according to this hierarchy. Our results

show that the AP approach offers both speed and solution quality advantages over ATC while following this same

organizational structure. This suggests that an AP-based approach may be a superior design strategy.

Although one cannot draw strong conclusions from an individual study, there is reason to believe that these

computational advantages will occur in other problems and possibly will be greater. This study involves a system

with two components having two and three concepts each for a total of six system concept combinations. We expect

that as the total number of component-level concepts increases, so too will the advantages of the AP approach. In the

UC system, adding an additional transmission concept would add three more system concept combinations for a

total of nine. A combinatorial growth pattern also exists when increasing the number of components in the problem.

In the UC system, adding a third component with only two concepts would double the number of system

22
American Institute of Aeronautics and Astronautics
combinations to twelve. Computational time for the AAO and ATC approaches grows invariably as more concepts

and components are added. One can think of this as there being an average computational cost associated with

evaluating any concept combination for a given problem and the overall computational cost is this average time

multiplied by the number of combinations. Thus, as the number of combinations grows, so does the overall

computation time. In contrast with this growth pattern, the AP approach as a whole takes only a small amount of

time more than it does to complete the ATC optimization for one concept combination. Adding concepts while

keeping the number of components fixed may affect the times associated with steps 1-5 in the AP procedure (c.f.,

Table 5), but it would have little or no effect on the two steps that dominate the computation time (steps 7 and 9).

Adding components with multiple concepts likely would impact all steps of the AP procedure, but there is no reason

for one to believe the degree of impact would scale directly with the number of combinations as it does for the other

approaches. Most likely, the AP approach is most sensitive to the number of samples one takes in step 1 (which in

turn can affect the dominance and model fitting times as well as model accuracy and final solution quality) and the

number of attributes required in the predictive model. The precise computational characteristics of the AP approach

are a subject for future study.

The practical advantages of the AP approach can be limited somewhat by engineering details that make certain

combinations of component-level concepts incompatible with each other. Moreover, just because designers have

four concepts for one component and three for another does not mean they have twelve valid system combinations

in total. This does not detract from the preceding results, but does mean the actual advantages of the AP approach

will be less than one would obtain from a worst-case combinatorial analysis.

Another limitation on the practical advantages of the AP approach is that designers may be unable to abstract all

concepts for a given component into a unified model. For example, we use the three gear ratios in the predictive

transmission model in our example, which means we are unable to represent a four-speed transmission. We still

would be able to abstract concepts for four-speed transmissions into a single model, but would have to deal with

three-speed and four-speed transmissions independently during optimization. Thus, the AP approach would be

beneficial, but to a lesser extent than if we could abstract all transmission concepts into a single model.

23
American Institute of Aeronautics and Astronautics
 VII. Summary

In this paper, we have presented and demonstrated a new approach for solving system design problems that

involve concept selection at the component level. This approach involves the use of abstract predictive models that

capture the capabilities of multiple component-level concepts into a single model. This has the effect of converting a

heterogeneous and combinatorial problem into a single search over a homogeneous space. In comparison with other

approaches common in the literature, the new approach offers significant computational advantages (a more than

ten-fold speedup) without major sacrifices in solution quality. Although our comparison is limited to one system

design problem, there is reason for one to believe the relative results will generalize to other problems. Future work

will involve additional assessment of the computational characteristics of the approach and demonstration on

problems of greater complexity.

Acknowledgments

This work is supported by the Department of Mechanical Engineering at Texas A&M University.

References

1
Geisser, S., Predictive Inference: An Introduction, Chapman & Hall, New York, 1993.
2
Malak, R. J. and Paredis, C. J. J., "Using Parameterized Pareto Sets to Model Design Concepts," Journal of Mechanical

Design, Vol. 132, No. 4, 2010

3
Malak, R. J., Tucker, L. and Paredis, C. J. J., "Compositional Modeling of Fluid Power Systems using Predictive Tradeoff

Models," International Journal of Fluid Power, Vol. 10, No. 2, 2009, pp. 45-55.
4
Malak, R. J., Using Parameterized Efficient Sets to Model Alternatives for Systems Design Decisions, Ph.D. Thesis, Georgia

Institute of Technology, Mechanical Engineering, Atlanta, GA, 2008.

5
Kim, H. M., Kumar, D. K. D., Chen, W. and Papalambros, P. Y., "Target Exploration for Disconnected Feasible Regions in

Enterprise-Driven Multilevel Product Design," AIAA Journal, Vol. 44, No. 1, 2006, pp. 67-77.
6
Kim, H. M., Michelena, N. F., Papalambros, P. Y. and Jiang, T., "Target Cascading in Optimal System Design," Journal of

Mechanical Design, Vol. 125, No. 3, 2003, pp. 474-480.

7
Buede, D. M., The Engineering Design of Systems, John Wiley & Sons, New York, 2000.
8
Sage, A. P. and Armstrong Jr., J. E., Introduction to Systems Engineering, Wiley and Sons, 2000.

24
American Institute of Aeronautics and Astronautics
 9
Kumar, D. K. D., Chen, W. and Kim, H. M.,"Multilevel Optimization for Enterprise-Driven Decision-Based Product

Design," Decision Making in Engineering Design, K. E. Lewis, W. Chen and L. C. Schmidt Eds., American Society of

Mechanical Engineers, 2006.

10
Keeney, R. L. and Raiffa, H., Decisions with Multiple Objectives, Cambridge University Press, Cambridge, UK, 1993.
11
Sobieszczanski-Sobieski, J. and Haftka, R. T., "Multidisciplinary Aerospace Design Optimization: Survey of Recent

Developments," Structural and Multidisciplinary Optimization, Vol. 14, No. 1, 1997, pp. 1-23.
12
Kroo, I. and Manning, V., "Collaborative Optimization: Status and Directions," 8th AIAA/NASA/ISSMO Symposium on

Multidisciplinary Analysis and Optimization, 2000, paper No. AIAA-2000-4721.

13
Gu, X., Renaud, J. E., Ashe, L. M., Batill, S. M., Budhiraja, A. S. and Krajewski, L. J., "Decision-based Collaborative

Optimization," Journal of Mechanical Design, Vol. 124, No. 1, 2002, pp. 1-13.
14
Michelena, N. F., Park, H. and Papalambros, P. Y., "Convergence Properties of Analytical Target Cascading," AIAA

Journal, Vol. 41, No. 5, 2003, pp. 897-905.

15
Rygielski, C., Wang, J.-C. and Yen, D., "Data Mining Techniques for Customer Relationship Management," Technology in

Society, Vol. 24, No. 4, 2002, pp. 483-502.

16
Simpson, T. W., Mauery, T. M., Korte, J. J. and Mistree, F., "Kriging Models for Global Approximation in Simulation-

Based Multidisciplinary Design Optimization," AIAA Journal, Vol. 39, No. 12, 2001, pp. 2233-2241.
17
Sobieski, I. P. and Kroo, I. M., "Collaborative Optimization using Response Surface Estimation," AIAA Journal, Vol. 38,

No. 10, 2000, pp. 1931-1938.

18
Wang, G. G. and Shan, S., "Review of Metamodeling Techniques in Support of Engineering Design Optimization," Journal

of Mechanical Design, Vol. 129, No. 4, 2007, pp. 370-380.

19
Jin, R., Du, X. and Chen, W., "The Use of Metamodeling Techniques for Optimization under Uncertainty," Structural and

Multidisciplinary Optimization, Vol. 25, No. 2, 2003, pp. 99-116.

20
Shigley, J. E. and Mischke, C. R., Mechanical Engineering Design, McGraw-Hill, New York, NY, 2001.
21
Tambača, J., "Derivation of a Model of Leaf Springs," Proceedings of the Conference on Applied Mathematics and

Scientific Computing, 2005, pp. 305-315.

22
Lophaven, S. N., Nielsen, H. B. and Sondergaard, J., "DACE: A Matlab Kriging Toolbox," Technical Report, IMM-TR-

2002-12, Technical University of Denmark, 2002.

25
American Institute of Aeronautics and Astronautics