Struct Multidisc Optim 22, 208–218  Springer-Verlag 2001

Pareto analysis in multiobjective optimization using

the collinearity theorem and scaling method
E.M. Kasprzak and K.E. Lewis

Abstract This paper presents a method to predict the recent evolution. Indeed, interest in decision theory can
relative objective weighting scheme necessary to cause be traced back to the 1950’s (e.g. Luce and Raiffa 1957),
arbitrary members of a Pareto solution set to become op- but the application of decision theory to multidisciplinary
timal. First, a polynomial description of the Pareto set design is relatively recent and reviewed by Lewis and
is constructed utilizing simulation and high performance Mistree (1998).
computing. Then, using geometric relationships between In decision theory, and more specifically in decision-
the member of the Pareto set in question, the location based design, there are two primary steps: generate the
of the utopia point and the polynomial coefficients, the option space, and select the best option (Hazelrigg 1996).
weighting of the performance metrics which causes a par- The option space is the accumulation of all potential
ticular member of the Pareto set to become optimal is design solutions. Choosing from among this space is
determined. The use of this technique, termed the scal- certainly not trivial but rather a function of tradeoffs
ing method, is examined via using a sample problem from and compromises. In this work, we are concerned with
the field of vehicle dynamics optimization. The scaling both of these steps. The decision making environment
method is based on the collinearity theorem which is also for this work is multiattribute design problems where
presented in the paper. there is more than one attribute under question. Also, we
are dealing with mathematical representations of design
Key words multiobjective optimization, Pareto set, problems and therefore are working with multiobjec-
objective weighting tive optimization problems. This does not exclude the
results of this paper from nonmathematical representa-
tions, such as look-up tables or logical rule-based models,
but we focus on mathematical models specifically. We in-
tend on providing decision support for determining the
1 optimal solution, given a set of operating conditions,
Introduction assumptions, risk profiles, and preferences. We do not in-
tend to construct risk profiles or model preferences, but
It is widely recognized that design is a series of compro- simply acknowledge that they must play a role in this
mises. Compromises are made using tradeoffs between kind of decision making problem.
performance, cost, risk, and quality attributes. In a mul- In multiobjective optimization problems, there are
tiattribute design problem there are typically an infinite two primary approaches to finding the preferred, if not
number of “optimal” solutions, based on the preferences optimal, design. The first involves determining the rela-
and risk assessments of the designer(s). The final design tive importance of the attributes and aggregating the
will be a result of numerous tradeoffs, many times ad-hoc, attributes into some kind of overall objective. Then, solv-
each aimed at making a compromise decision between ing the optimization problem presumably would generate
conflicting attributes. Viewing design formally as a col- the optimal solution for a given set of attribute impor-
lection of compromise decisions, however, is only a very tances. The second approach involves populating a num-
ber of optimal solutions along the Pareto frontier and
July 25, 2000
then selecting one based on the values of the attributes for
E.M. Kasprzak1 and K.E. Lewis2 a given solution. In both cases, there are complications.
1 First, coming up with exact relative attribute weights is
Milliken Research Associates, Inc., 245 Brompton Road,
a daunting task with complicated ramifications (Messac
Williamsville, NY 14221-5942, USA
e-mail: [email protected] 2000). Only in the rarest of circumstances can this ap-
2
Department of Mechanical and Aerospace Engineering, Uni- proach be taken and relied upon. Second, while selecting
versity at Buffalo, 1010 Furnas Hall, Buffalo, NY 14260, USA from a set of Pareto solutions may seem straightforward
e-mail: [email protected] once a set of preferences are established, it may result
 209

in a nonoptimal design for certain design operating con- and Papalabmros (1996) also looked at using nonlinear
ditions. It is difficult to relate the choice of Pareto so- weights to better capture the non-convex Pareto set.
lution to any kind of optimal criteria. In this work, we An alternative to these methods, compromise pro-
develop a technique that can provide decision support gramming, was developed in the 1970’s and it, too, is
for both of the approaches to multiobjective optimization a subject of continuing research. This approach has the
problems. In the next section, we provide the necessary advantage of being able to generate many points on the
background on Pareto sets for the development of the efficient frontier. It works from a stationary utopia point
technique. and, through variation of a vector of weights, intersects
points in the Pareto set. Chen et al. (1999) used this
method in an approach to robust design with reason-
2 able success, although implementation of compromise
Pareto set background programming is notably more difficult than the weighted
sum method. Tappeta and Renaud (1999) used compro-
Since multiobjective optimization problems are in ques- mise programming to find an initial Pareto solution, and
tion, we make use of the concept of Pareto sets, an effi- then constructed local approximations to the Pareto sur-
cient frontier of solutions in the performance space. There face and iteratively presented the decision maker with
continue to be two primary challenges in Pareto sets: pop- a set of solution options to choose from.
ulating the Pareto set or finding Pareto solutions, and Yoo and Hajela (1999) effectively use an immune-
selecting from among the Pareto solutions. These chal- based genetic algorithm (GA) to generate the Pareto-
lenges parallel the challenges of decision-based design. Edgeworth solutions in one run of the GA. Narayanan
They are analogous to determining potential solutions and Azarm (1999) improve multiobjective GA’s using fil-
and selecting from among the solutions. In Pareto analy- tering, mating restrictions, and the idea of objective con-
sis, there are added restrictions on the criteria for decid- straints in order to detect Pareto solutions in the noncon-
ing upon a solution’s inclusion in the set. A design vector vex region of the Pareto set.
x∗ is a Pareto optimum if and only if, there is no other This work will use yet another method to determine
feasible vector x such that the Pareto set: a grid search of the design space. Effi-
cient points from the grid search are fit by a polynomial
fi (x) ≤ fi (x∗ ) , j = 1, . . . , m to approximate the Pareto set, resulting in several advan-
tages over other methods. This is pursued in Sect. 3. Once
and the Pareto set has been generated, the next challenge is
choosing the best solution from among the set. This is the
fi (x) < fi (x∗ ) for at least one i , 1 ≤ i ≤ m. (1)
subject of the next section.
We address both of these challenges in this work, al-
though most of the paper is focused on the second chal-
lenge. In the next sections, we address both challenges 2.2
and the current research in each area. Selection from among the Pareto solutions

Just as several methods exist to determine what solu-

2.1 tions compose the Pareto set of efficient solutions, there
Population of the Pareto set are several techniques to determine which member of the
Pareto set is the optimal solution. Similar to decision-
The notion of an efficient frontier of solutions is not based design, the simplest method is to choose a so-
new (e.g. Luce and Raiffa 1957), but methods of gen- lution based on the values of the objectives and how
erating Pareto sets continue to be a topic of research. well they match the preferred values. This method is
The weighted sum method of generating Pareto sets was used by Nelson et al. (1999) when manipulating multi-
shown to work well with convex problems decades ago ple Pareto solution sets in product platform design. Das
by Geoffrion (1968), and while it is still a very popular (1999) introduces the concept of “order of efficiency” as
method its deficiencies have been noted. Messac (2000) an attempt to create a meta-metric, stronger than the
has effectively illustrated the problems associated with Pareto conditions, in order to rank order Pareto solu-
choosing weights for an aggregate objective function and tions. Hazelrigg (1996) argues that the selection from
has derived conditions that predict which Pareto solu- a set of solutions should be guided by a “meta-objective”
tions can be found using weights (Messac et al. 2000). of maximizing profit. Designer preferences can be ac-
Dennis and Das (1997) notes that, even with convex prob- commodated by using utility theory to establish a rank-
lems, taking an even spread of weights will not result ordering of solutions based on uncertain and changing
in an even spread of points in the Pareto set, making individual and group preferences (Callaghan and Lewis
some sections of the Pareto set difficult to populate. Koski 2000). Horn et al. (1994) use multi-attribute utility an-
(1985) emphasizes the weighted sum method’s inade- alysis to select the preferred solution from the Pareto set
quacies when dealing with nonconvex problems. Athan of solutions. Eschenauer et al. (1990) describes the most
210

widely used method: Lp norms. This technique minimizes

the distance from the Pareto set to an ideal solution (i.e.
utopia point) to find the optimal solution according to the
following formula:

m 1/p

(fi (x) − fi∗ )
p
minimize . (2)
i=1

Typical applications of the Lp norm are the L1 , L2 and

L∞ norms (where p = 1, 2 and ∞, respectively). A fine
summary of solution techniques is provided by Palli et al.
(1998) should the reader desire a more in-depth explo-
ration of the subject.
In this work, we utilize the L2 norm (p = 2) to de-
rive a technique for finding the optimal attribute weights Fig. 1 A generic performance space
necessary to make any member of the Pareto set opti-
mal. In other words, we determine what design conditions
are required to make a certain Pareto solution the pre- simultaneously. As such, the Pareto set is sometimes re-
ferred solution. In Sect. 3, we present a new theorem, the ferred to as the “efficient frontier” of the performance
collinearity theorem, in a general sense and then apply space. Whatever compromise design is chosen, it must
it in a new method called the scaling method in Sect. 4. be a member of the Pareto set, and the remainder of
We then illustrate the use of these techniques in a vehicle possible outcomes can be ignored. They represent design
dynamics design example. First, however, it is instructive points that give suboptimal performance on both axes
to step back and examine the performance space in more simultaneously.
detail. This is done in the next section.

2.3
A conceptual look at the performance space and axis
scaling

Pareto set analysis typically takes place in the perform-

ance space. This is the space created by considering the
design objectives as coordinate axes. On these axes the
performance of each possible design is plotted, one point
per design. Each possible design has associated with it
a specific performance or outcome for each objective. In
general, the plot of possible outcomes appears as a region
in the performance space, as depicted in a general two ob-
jective problem in Fig. 1. Fig. 2 Optimal compromise solution using the L2 norm
Figure 1 notes that the intention is to minimize both
objectives. The individual optimum for each objective is
also noted in Fig. 1. These correspond to the best pos- The concept of utopia point is important when using
sible performance for each given objective. Unfortunately, distance functions, such as the L2 norm. It is the theoret-
two distinct points appear, one for each objective. Since ical best performance point that can be achieved. While
both objectives cannot be simultaneously optimized, any more than one method for locating the utopia point exists
attempt to choose a single design to perform well across (e.g. Miettinen 1999), this discussion presents the defin-
both objectives will necessarily be a compromise design. ition which agrees with the vehicle dynamics problems
Being optimal for one objective implies being suboptimal presented later. In this case, the utopia point is taken to
for the other. reside at the origin of the coordinate system. While we
Figure 2 shows how a certain distance measure, the L2 present the results of the example using this utopia point
norm, determines the optimal compromise design. First, location, the derivation in Sect. 4 is general. The develop-
the range of possible designs is narrowed significantly ments can be used with other utopia point definitions.
through the concept of a Pareto set, a subset of the set Now, according to the L2 norm method (Eschenauer
of possible outcomes shown in Fig 1. The Pareto set has et al. 1990), the optimal compromise design is the mem-
the property that, for any point in the Pareto set, there ber of the Pareto set which lies geometrically closest to
does not exist another point in the set of possible out- the utopia point, calculated in terms of vector distance in
comes with a better performance on both objective axes the performance space. Figure 2 locates the optimal point
 211

using the L2 norm method. A vector is drawn from the sented. In the next section the collinearity theorem is
utopia point to the optimal point on the Pareto set. To introduced and a proof is given.
emphasize that this is, indeed, the closest point, a circle
centered at the utopia point with the radius of the vector
is included. Since no other points on the Pareto set appear 3
within this circle the point shown must be the closest to Mathematical basis: the collinearity theorem
the utopia point.
This sample problem is investigated with an arbitrary When analysing problems involving compromise deci-
baseline weighting of the performance objectives. That is, sions, the construction of a Pareto set is useful in identi-
a certain relative importance (ratio of objective weights) fying all possible “good” solutions from the set of possible
is assumed to produce Figs. 1 and 2. At this point the designs. Each member of the Pareto set is potentially
precise ratio used is not important. What is important, the optimal solution to the problem at hand, depending
however, is how Fig. 2 changes if the relative importance on the relative weights of the objectives. Deciding which
of the weights changes. Suppose the scenario in Figs. 1 Pareto set member is optimal requires these weights to be
and 2 is generated by considering the two objectives to known so that existing methods, such as the Lp -norm and
have equal importance. If the importance placed on ob- utopia point concept, can be applied, thereby determin-
jective I doubles (such that objective I is twice as import- ing the optimal solution. Definitions of these decision the-
ant as objective II) a different compromise design may be ory terms and techniques are summarized by Palli et al.
the optimal solution. (1998), many of which follow from the early game the-
Figure 3 shows how this change in relative importance ory and compromise decision work described by Luce and
of the objectives is handled. Compared with Fig. 2, the Raiffa (1957).
objective I axis in Fig. 3 is stretched to twice its original A substantial number of analysis techniques exist, as
length, reflecting the doubling of the importance on that reflected by Chen et al. (1999). Every technique study-
axis. This reshapes or “rescales” the set of possible out- ing compromise decisions which strives to identify an
comes and the Pareto set. As a result, the member of the optimum point is constrained by a given set of perform-
Pareto set which is closest to the utopia point is differ- ance metrics. That is, had the relative objective weighting
ent than that in Fig. 2. A new optimal compromise design been different, the point chosen from the Pareto set as
is located using the L2 norm in Fig. 3 for the case when optimal would likely also have been different. While the
objective I is twice as important as objective II. designs composing the Pareto set are, by definition, inde-
pendent of any specific objective weighting, the optimal
Pareto set solution is necessarily a function of the particu-
lar objective weighting under discussion.
This section states and proves a new theorem termed
the collinearity theorem. It has been developed as an ex-
tension of the L2 norm in the performance space. The
L2 norm locates the best point in a Pareto set as the
one which lies geometrically closest to the utopia point.
This is visualized as a circle, centered at the utopia point,
which determines the optimum point as the first design
point encountered as the radius of the circle is increased.
Instead, the collinearity theorem relates the best point
in a Pareto set to the utopia point in terms of another
geometric construct, the shape of the Pareto set itself.
It is the inclusion of information regarding the shape of
the Pareto set which gives the collinearity theorem its
Fig. 3 Rescaled performance space doubling the importance value. While the L2 norm and collinearity theorem are in
of objective I
agreement in that they always determine the same Pareto
point to be the optimal solution, the collinearity theorem
It is important to note that the design points compos- lends itself more readily to expansion and generalization.
ing the Pareto Set are no different in Fig. 3 than they were Indeed, the collinearity theorem is the underlying mathe-
with the baseline weighting of the objectives. By defin- matical premise for the scaling method, developed later in
ition Pareto set points are independent of the relative this paper in Sect. 5.
importance of the design objectives. The only difference
is the way these points are plotted in the performance 3.1
space. In Sect. 4 this geometric interpretation of objective Development of the collinearity theorem
weighting is applied in the scaling method.
Through the use of a generic example, a conceptual For the time being assume that the Pareto set is described
summary of performance space analysis has been pre- by a line  in the performance space whose equation or
212

mathematical representation is known. A technique for By the definition of optimal point using the L2 norm,
determining this representation of the Pareto set is given point B is known to be the closest member of  to the
in Sect. 3.2. With this information the collinearity theo- utopia point, A. Thus, a circle drawn about A with radius
rem can be defined as follows. AB will touch line  only at point B. Since the circle and
the line at point B have continuous slopes the circle cen-
The collinearity theorem. An internal point B in a Pareto tered at A must be tangent to  at B. The tangent to  at
set  is an optimal point if and only if the utopia point,
B is T . A circle’s radius is always perpendicular to a tan-
point B and the instantaneous centre of curvature of gent on the circumference, so line segment AB must be
the Pareto set  at point B are collinear, provided the
perpendicular to T .
instantaneous centre of curvature of the Pareto set at
Both AB and BC are perpendicular to T and there-
point B does not lie between point B and the utopia fore parallel to each other. Parallel line segments, by
point.
definition, have the same slope. Since point B is com-
Exception. If the Pareto set does not have a continu- mon to both line segments we know that A, B and C
ous slope at B or if B is an endpoint (i.e. “noninternal” must be members of the same line. This completes the
point) of the Pareto set then this condition need not
proof of the collinearity theorem for internal Pareto set
be met for point B to be optimal. points.
As an aid to the proof to this theorem, consider Fig. 4 At points where the slope of the Pareto set is dis-
where continuous or where the optimal point is a Pareto set
endpoint the collinearity theorem does not need to be
 is a curve representing a Pareto set, satisfied for optimality to exist, thus the reference to “in-
A is the utopia point, ternal points” of the Pareto set. Noninternal points can
B is the point on  closest to the utopia point, be corners, jumps or endpoints of the Pareto set. These
C is the instantaneous centre of curvature of  at B, and points do not have a continuous derivative. As such, these
T is the tangent to  at B. points do not have an instantaneous centre of curvature
and cannot satisfy the collinearity theorem for optimal-
ity. Noninternal points are typically few and can easily
be checked individually. Through the use of the scaling
method they can often be readily examined without addi-
tional analysis.
This proof has assumed that the objectives are being
minimized. This is not a requirement, but presented for
illustrative purposes.

3.2
Polynomial description of the Pareto set

The approach of the collinearity theorem is based upon

knowledge of the shape of the Pareto set. It is the local
curvature of the Pareto set which determines what rela-
tive objective weighting will cause a local Pareto set mem-
ber to become optimal.
Section 2.1 presented various different techniques to
Fig. 4 Geometry associated with the collinearity theorem
populate the Pareto set. Our approach centres on a grid
search of the design space followed by a polynomial
A proof of the collinearity theorem is now presented. fit in the performance space. First, a discretization of
Three points are said to be collinear if they are points on the design space is conducted. Each combination of de-
the same line, according to elementary geometry. Addi- sign variables is evaluated to determine the correspond-
tionally, a line drawn between any two points on a line will ing performance with respect to the design objectives.
have the same slope as the original line. If A, B and C in The results are then plotted in the performance space
Fig. 4 are collinear then line segment AB and line segment and, in a manner similar to that shown in Fig. 2, the
BC must have the same slope as line segment AC. By ver- points which satisfy the criteria for the Pareto set are
ifying this is the case the collinearity of A, B and C can be identified.
proven. Despite the fact that most of the design points cho-
Proof. Because C is the instantaneous centre of curva- sen will not be members of the efficient frontier there are
ture of  at point B the line segment BC is, by definition, distinct advantages to the grid search approach. Perhaps
perpendicular to the tangent of  at B. The tangent to  at key among these are the fact that the entire Pareto Set,
point B is denoted T in Fig. 4. both convex and nonconvex regions, is guaranteed to be
 213

located given a fine enough discretization of the design

space. This overcomes a principal shortcoming of any of
the techniques of Sect. 2.1. Furthermore, advanced grid
search techniques may be employed to enhance the effi-
ciency of the search, although this has not been addressed
by the authors.
With a representative number of Pareto set points de-
termined by the grid search, the process of polynomial fit-
ting can begin. In a two-objective problem, one objective
is represented as a function of the other and a polynomial
fit applied in the performance space. Since a polynomial
has two fewer inflection points than the order of the poly-
nomial, the use of a high order is important to capture
both the convex and nonconvex regions in the approxima-
Fig. 5 Relating the collinearity theorem to Pareto set termi-
tion. The result is a simple, continuous description of the nology
Pareto set.
The fitting of a polynomial allows information about
the shape of the Pareto set to be easily accessed as re-
quired by the collinearity theorem. While it has not been
pursued, a series of splines, done properly, would also
satisfactorily represent the Pareto set for use with the
collinearity theorem. Also, when dealing with three or
more objectives the Pareto set is no longer a line but
rather a hyperplane. In such problems a response surface
would be fit to the Pareto set. This, too, is the subject of
future research.

4
Application of the collinearity theorem:
the scaling method
Fig. 6 Rescaled abscissa by an amount k
Now that the collinearity theorem is stated and proven,
it is developed into a method to predict the scaling of the
objective weights needed to cause a member of the Pareto The collinearity theorem allows the appropriate value
set to become the optimal solution. This new procedure, of k to be determined. Three points, the utopia point, the
termed the scaling method, has been developed to realize point we wish to make optimal (f1 , f2 ) and the centre of
the information the collinearity theorem makes available. curvature of the Pareto set at (f1 , f2 ) from Fig. 5, must be
The goal is to predict the relative objective weighting re- collinear for optimality to occur. Equating slopes between
quired to cause any member of the Pareto set to become the first pair of points and the second pair of points (refer
the optimal solution on the basis of the information con- to Fig. 6) and considering objective f2 to be a function of
tained in the shape of the Pareto set. objective f1 , gives the relationship
Figure 5 presents the performance space for a general
compromise decision problem. The point (f1∗ , f2∗ ) is the f2 − f2v 1
=−  . (3)
optimal solution for a baseline weighting of the objectives kf1 − kf1u f2
on the abscissa and ordinate. This baseline weighting of
See the Appendix for a detailed explanation of how rela-
objectives f1 and f2 is arbitrary, although for simplicity it
tionships change when the abscissa is rescaled. Solving for
is convenient to weight the objectives equally. Other im-
k, the only unknown, gives
portant points on this figure are the utopia point, denoted
(f1u , f2u ), the instantaneous centre of curvature at the (f2 − f2v )f2
optimal point, denoted (f1c ∗ ∗
, f2c ), and a nonoptimal point k=− . (4)
f1 − f1u
on the Pareto set, (f1 , f2 ).
Stretching or “rescaling” the f1 -axis by an appropri- Through the use of (4) the change in relative objective
ate amount k will, in general terms, cause Fig. 5 to look weights, k, needed to change a suboptimal member of
like Fig. 6. The point (f1 , f2 ) is now (kf1 , f2 ). Because the Pareto set (f1 , f2 ) into the optimal point can be de-
of the value of k selected, the formerly suboptimal point termined. Implementation of (4) to predict under what
(f1 , f2 ) is now the optimal point (kf1 , f2 ). The previously preferences a member of the efficient frontier becomes the
optimal point is now suboptimal. optimal solution is termed the scaling method. It is an
214

extension of the collinearity theorem and it predicts the trol over the performance of the car. Furthermore, dur-
change in relative weighting of the objectives, k, from the ing a lap at a typical race, a driver may sees a number
baseline condition as a function of of different types of corners and straights. In Fig. 7, the
configuration for the Indianapolis Motor Speedway Road-
1. the original (unscaled) location of the point of interest course, site of the Formula One United States Grand Prix
on the Pareto Set (f1 , f2 ); in September 2000, is shown. There are many turns and
2. the original slope, f2 , of the Pareto set at the point of straight-aways, each with its own optimal vehicle charac-
interest (f1 , f2 ); and teristics. The optimal car for the sweeping final corner, an
3. the original location of the utopia point (f1u , f2u ). 840-foot radius turn on the lower right, is different than
the tight 114-foot radius turn in the upper centre of the
This approach is illustrated in the vehicle design prob- track.
lem that follows.

5
Case study: vehicle dynamics design

Whether it is NASCAR, CART, Formula One, or even

local racetracks, the difference between winning a race
and not winning comes down to the ability of a driver
to get the most out of his or her racecar. While having
a talented driver is always desirable, even the most tal-
ented driver can do nothing more than realize the full Fig. 7 Indianapolis motor speedway roadcourse configura-
potential of the vehicle. The core vehicle design, how it tion
is “set-up” and the “tuning” done by a race team are
aimed at an optimal compromise that allows the driver to
repeatedly turn fast lap times at a particular racetrack. Designing a car to perform well across turns of all radii
Any advantage gained though vehicle design and tuning, on a single track involves a set of complicating tradeoffs.
however small, increases the vehicle’s potential and, with Simulating these tradeoffs is a difficult task, as a complete
a talented driver, will translate into an increase in on- vehicle simulation is difficult to model, very complex and
track performance. Vehicle simulations are now used not computationally challenging to run efficiently. In this pa-
only prior to and during a race weekend to guide tun- per, we use a simplified vehicle model to illustrate the use
ing of the race car, but also in the design phase where of the techniques introduced in Sects. 3 and 4 to model
parameters which are not adjustable must be set and this type of problem and to predict under what race con-
optimized. ditions (combination of radii) a particular vehicle design
The scope of vehicle simulation continues to grow. Ad- will be optimal.
vances in the memory and speed of computers have led The vehicle model is based on the classic bicycle model
to the use of increasingly complex vehicle models with of Milliken and Milliken (1995), which has been expanded
an ever expanding scope. The modelling of performance to include four individual wheels. Equations of motion
around a single corner has developed into full lap analy- are written for lateral acceleration, longitudinal accel-
sis. Even further, it is now possible to analyse all the eration and yaw acceleration. The tyres, which may be
tracks on the season schedule when designing a new car different front and rear, are modelled using tabular tyre
and attempt to optimize core design parameters, those data measured on a state-of-the-art tyre testing machine
which are not easily changed once the car is built, be- and represent numerous tyre nonlinearities such as load
fore the car even exists. An example of one such vari- sensitivity and slip angle saturation. Wheel loads are cal-
able is the longitudinal (fore/aft) centre of gravity loca- culated based on static load, aerodynamic downforce and
tion. The range of possible adjustment on this variable lateral load transfer.
throughout a season is very small. It must be optimized Through computer simulation, the vehicle is entered
in the design stage before the car is constructed. Further into a special race for which the compromises involved
importance is placed on the use of vehicle simulations can be easily studied. This race is defined as the time
as racing sanctioning bodies impose tighter restrictions to complete “x laps on a small radius (100 foot ra-
on the amount of on-track testing teams are allowed to dius) circle and y laps on a large (400 foot radius) ra-
do. dius circle”, each at best steady-state speeds. This ab-
Vehicle design is considered an excellent example of straction of a true racetrack is especially well-suited to
a multidisciplinary design optimization problem. There highlighting how the objectives have competing opti-
are vehicle dynamicists, aerodynamicists, tyre designers, mal designs. Since the vehicle cannot be optimized for
engine builders, shock absorber specialists, mechanics, both radii simultaneously due to vehicle aerodynamics,
a driver, all of whom have different outlooks and con- tyre performance and other speed dependent vehicle
 215

behaviours, a compromise design is needed. This com- Figure 9 shows the grid search points which are mem-
promise design is one which allows the vehicle to per- bers of the Pareto set along with the results of fitting
form optimally for the race, not just an individual circle. a seventh-order polynomial via least-squares criteria.
The ratio of x to y, denoted k can be viewed as the The polynomial gives a representative description of the
weighting of two design objectives, the elapsed times Pareto set throughout the range of the Pareto set, except
on each individual radius. The shortest time to com- at the far right end where the fit acquires a positive slope.
plete the total distance around both radii will win the As a result, analysis of this Pareto set for cases where the
race. 400’ radius design dominates the compromise design can
For this study, the vehicle has two design variables. be expected to produce slightly inaccurate results. While
These variables, roll stiffness distribution (K’) and weight the use of an eighth order fit resolves this inaccuracy, the
distribution (a’), are two of the three “magic numbers” seventh-order fit is presented here to highlight potential
Wright (1998) identifies as being fundamental to race errors caused by a fit which does not fully represent the
car design and performance. They are used by the vehi- Pareto set.
cle designer and vehicle dynamicist to “tune” the car’s
handling in an attempt to optimize performance. The an-
alysis begins with a grid search of the design space from
which the performance space is plotted and the Pareto set
identified. The design space has been discretized into 546
(a’, K’) pairs. Each design pair results in a certain level of
performance on each of the two radii. Figure 8 plots the
performance space for the vehicle for a baseline race de-
fined as one lap on each radius (k = 1). This gives each
radius equal importance. Thus, the axes are the time re-
quired to complete one lap on each of the 100’ radius and
400’ radius circles.
From Fig. 8 the optimum lap times for each radius are
easily located. For the 100’ radius the leftmost point on
this figure gives the optimum where the vehicle can do no
better than approximately an 8.88 second lap. Likewise,
for the 400’ radius the bottom most point in the plot gives
the optimum. This point indicates approximately a 9.1
second lap as the best possible. These two points are con-
sistent with the detailed vehicle simulations of Kasprzak Fig. 9 The Pareto set with seventh-order polynomial
(1998). The two optima do not occur at the same point
so the race definition, combining laps on each radius, will
need a compromise solution to arrive at the optimal de- Figure 10 presents the outcome of applying the scaling
sign and performance. method (4) to the seventh-order fit Pareto set. This figure
associates the optimal race definition k with the vehicle
performance on the 100 foot radius. Note that since the
origin (0,0) is being used as the utopia point, (4) reduces
to a simple form. Since Fig. 9 shows a one-to-one corres-
pondence of performance on the 100 foot radius to that
on the 400 foot radius, Fig. 10 also specifies the optimal
400 foot radius performance for any race definition k. Fur-
thermore, since each point on the Pareto set is the result
of a unique combination of design variables, a relationship
between race definition and vehicle design is also acquired
for the entire range of k values.
Consider the case of a race definition k = 15. That is,
the race will contain 15 laps on the 100 foot radius and
one lap on the 400 foot radius, each at best steady-state
performance. Figure 10 predicts an optimal performance
on the 100 foot radius circle of 8.894 seconds per lap, or
133.4117 seconds total time on the 100 foot radius. If the
abscissa of Fig. 9 is rescaled by an amount equal to k (15)
then Fig. 11 is generated. It shows the shape of the Pareto
set in terms of the elapsed time on each radius for an
Fig. 8 Grid search results in the performance space k = 15 race. A circle drawn about the utopia point (0,0),
216

optimum point. For the central, nonoptimal point, the

centre of curvature of the Pareto set lies between the
Pareto set and the utopia point. This point does not sat-
isfy the collinearity theorem and is therefore not a can-
didate to be an optimal point, even though (4) pro-
duces this point as a solution. Thus, the scaling method
and the collinearity theorem are best used in tandem
to arrive at optimal solutions in the Pareto set. Calcu-
lation of the centre of curvature in addition to the use
of (4) gives a more complete picture of the Pareto set’s
behaviour.

Fig. 10 Corresponding optimal scaling predicted by (4)

representative of the L2 norm technique for locating an

optimum point, identifies 113.4117 seconds as the opti-
mum performance on the 100 foot radius, identical to that
predicted by the scaling method. The L2 norm and the
scaling method each predict the same optimal perform-
ance, as desired.

Fig. 12 Theoretical optimal scaling when k = 8.4

The other two points identified in this k = 8.4 example

are indeed optimal design points and produce identical
race times to one another. The left most optimum results
in a race time of (8.4 × 8.9031) + 9.1688 = 83.9545 sec-
onds while the right most optimum gives (8.4 × 8.9198)
+ 9.0282 = 83.9545 seconds. These are identical, optimal
race times. In contrast, the centre point gives a race time
of (8.4 × 8.9114) + 9.2459 = 84.102 seconds, nearly 0.15
seconds slower than the other two design points. As pre-
dicted by the collinearity theorem, because the centre of
curvature of the Pareto set is between the utopia point
and the Pareto set, this point is not optimal.
These two race definitions serve to show the valid-
Fig. 11 Verification of theoretical optimal scaling when
k = 15 ity and usefulness of the scaling method and collinearity
theorem. A more detailed discussion of this example is
provided by Kasprzak et al. (1999). Additional comments
Close examination of Fig. 10 reveals a region where about these techniques are provided in the next section.
the scaling k is associated with more than one perform-
ance criteria on the 100 foot radius. One such value is
k = 8.4. For this race definition three optimal 100 foot 6
radius performance points are predicted: 8.9198, 8.9114 Comments on the scaling method
and 8.9031 seconds per lap on the 100 foot radius. These
are denoted by the dashed lines in Fig. 12. Rescaling When using the scaling method, arriving at a polynomial
the original Fig. 9 by k (8.4) produces Fig. 12. In this which describes the Pareto set accurately is critical. Use
figure the circle about the utopia point touches only of a high-order polynomial is desirable so that both
two of the three points. The middle point is not an the concave and convex regions of the Pareto set can
 217

be represented. A polynomial curve has two less inflec- of the Pareto set at the point of interest are collinear.
tion points than the order of the polynomial and these From this the scaling method is derived, which allows
polynomial inflection points may not all fall within the the relative objective weighting to be predicted for any
range of the Pareto set. It is very important that the point on the Pareto set, based on the polynomial fit.
polynomial chosen meets the criteria used to specify the Alternatively, the optimal performance for all ratios of
Pareto set. As noted earlier, the seventh-order fit pre- objective weights can be readily determined. Thus, from
sented violates a Pareto set criterion by reversing the a grid search of the design space the shape of the en-
sign of its slope near the right end of the fit. This re- tire Pareto set can be determined, as well as the entire
sults in negative scaling values being predicted for this mapping of the optimum Pareto set solution to the rela-
region, seen at the right end of the trace in Fig. 10. tive objective weighting. This provides a vast amount of
These negative values have no physical meaning and lead information and practical insight to the problem under
to solution inaccuracies. The use of an eighth-order fit consideration. Verification of this method against re-
with this example, while not presented here, with this sults obtained using an L2 norm shows the new methods
example provides a better representation of the Pareto produce results complementary to the established tech-
set. nique. While the method presented here addresses only
The polynomial fit can only be as good as the data two-objective compromise decision making, the merits
to which it is applied. A sufficiently fine grid search of of extending the collinearity theorem and the scaling
the design space is required to populate the Pareto set. method to multiple dimensions are clear. Describing the
While this method has the distinct advantage of find- Pareto set as a hyperplane in n-dimensions and apply-
ing the entire Pareto set, both concave and convex re- ing the collinearity theorem in n-dimensional space ap-
gions, the search increment may be very small and re- pear to be possible. While the concepts involved remain
quire a large number of points to be calculated. Since unchanged, the mathematical expression of these ideas
most points calculated will not be Pareto set points this in n-dimensions will be more complicated than in two
method may, at first glance, be deemed inefficient. Still, dimensions.
the grid search technique is guaranteed to find the en-
tire Pareto set, given a fine enough discretization. This Acknowledgements The authors would like to thank Milliken
alone is a key point. The combination of a grid search and Research Associates, Inc. for their support of this research.
polynomial to describe the Pareto set has the ability to The authors would also like to thank the National Science
provide an accurate, smooth and continuous description Foundation, Grant DMI-9875706, for their support.
of the whole Pareto set, regardless of the existence of both
convex and nonconvex regions.
The collinearity theorem and scaling method exploit References
the use of this grid search and polynomial description of
Athan, T.W.; Papalambros, P.Y. 1996: A note on weighted
the Pareto set. These techniques pull a vast amount of criteria methods for compromise solutions in multi-objective
information out of the polynomial fit. The optimal ob- optimization. Eng. Opt. 27, 155–176
jective performances can be calculated for any and all
ratios of objective weights with ease via (4). To be of Callaghan, A.; Lewis, K. 2000: A 2-phase aspiration-level and
real use to the designer, relating these performance re- utility theory approach to large scale design. 2000 ASME
sults to their corresponding design variables is desirable. Design Engineering Technical Conf., DETC2000/DTM-14569
While these results appear in the performance space, an- Chen, W.; Wiecek, M.M.; Zhang, J. 1999: Quality util-
other method has been developed, termed the contour ity: a compromise programming approach to robust design.
method, which relates these results to the design space. ASME J. Mech. Des. 121, 179–187
This method, presented by Kasprzak and Lewis (2000),
Das, I. 1999: A preference ordering among various Pareto op-
further enhances the utility of the scaling method and
timal alternatives. Struct. Optim. 18, 30–35
collinearity theorem.
Dennis, J.E.; Das, I. 1997: A closer look at drawbacks of min-
imizing weighted sums of objective for Pareto set generation
7 in multicriteria optimization problems. Struct. Optim. 14,
Conclusions 63–64

Eschenauer, H.; Koski, J.; Osyczka, A.E. 1990: Multicriteria

Starting with a grid search of the design space, the design optimization: procedures and applications. Berlin, Hei-
Pareto set is populated and then approximated using delberg, New York: Springer
a least-squares fit polynomial. This Pareto set is based
on a baseline goal definition with equally weighted ob- Geoffrion, A.M. 1968: Proper efficiency and the theory of vec-
tor maximization. J. Math. Analysis Appl. 22, 618–630
jectives, referred to as k = 1. The collinearity theorem
states that any internal member of the Pareto set is op- Hazelrigg, G. 1996: Systems engineering: an approach to
timal only if the utopia point, the point of interest in information-based design. Upper Saddle River, NJ: Prentice-
the Pareto set and the instantaneous centre of curvature Hall
218

Horn, J.; Nafpliotis, N.; Goldberg, D.E. 1994: A niched Pareto Miettinen, K.M. 1999: Nonlinear multiobjective optimization.
genetic algorithm for multiobjective optimization. Proc. 1st Boston: Kluwer
IEEE Conf. on Evolutionary Computation, IEEE World
Congr. of Computational Intelligence (held in Piscataway, Milliken, W.F.; Milliken, D.L. 1995: Race car vehicle dynam-
NJ), Vol. 1, pp. 82–87 ics. Warrendale, PA: SAE Int. No R-146

Kasprzak, E.M. 1998: Multivariate optimization and game Narayanan, S.; Azarm, S. 1999: On improving multiobjective
theory applications in vehicle dynamics simulations. M.S. genetic algorithms for design optimization. Struct. Optim. 18,
Thesis, Department of Mechanical and Aerospace Engineer- 146–155
ing, University at Buffalo, NY
Nelson, S.; Parkinson, M.; Papalambros, P. 1999: Multicri-
Kasprzak, E.M.; Lewis, K.E. 2000: An approach to facilitate teria optimization in product platform design. Proc. ASME
decision tradeoffs in Pareto solution sets. J. Eng. Value Cost Design Engineering Technical Conf. (held in Las Vegas),
Analysis 3, 173–187 DETC99/DAC-8676

Kasprzak, E.M.; Lewis, K.E.; Milliken, D.L. 1999: Steady- Palli, N.; McCluskey, P.; Azarm, S.; Sundararajan, R. 1998:
state vehicle optimization using Pareto-minimum analysis. An interactive multistage ε-inequality constraint method for
SAE Trans., J. Passenger Cars 107, 2624–2631 multiple objectives decision making. ASME J. Mech. Des.
120, 678–686
Koski, J. 1985: Defectiveness of weighting method in multicri-
terion optimization of structures. Comm. Appl. Numer. Meth. Tappeta, R.; Renaud, J. 1999: Interactive multiobjective op-
1, 333–337 timization design strategy for decision based design. Proc.
ASME Design Engineering Technical Conf. (held in Las
Lewis, K.; Mistree, F. 1998: The other side of multidisci- Vegas, NV), DETC99/DAC-8581
plinary design optimization: accommodating a multiobjec-
tive, uncertain and non-deterministic world. Eng. Opt. 31, Wright, P. 1998: What is McLaren’s secret? Racecar Eng. 8,
161–189 6–7

Luce, R.D.; Raiffa, H. 1957: Games and decisions. New York: Yoo, J.; Hajela, P. 1999: Immune network simulations in mul-
John Wiley ticriterion design. Struct. Optim. 1, 85–94

Messac, A. 2000: From the dubious construction of objective

functions to the application of physical programming. AIAA Appendix
J. 38, 155–163

Messac, A.; Sundararaj, J.G.; Tappeta, R.V.; Renaud, J.E. Table 1 supports the development in Sect. 4. It shows how
2000: The ability of ojective functions to generate non-convex various Pareto Set parameters change with rescaling of
Pareto frontiers. AIAA J. 38, 1084–1091 the abscissa in the performance space by an amount k.

Table 1 Effects of abscissa scaling on Pareto set attributes

Before rescaling After rescaling by k

suboptimal point of interest on Pareto set (x1 , y1 ) (kx1 , y1 )

instantaneous slope of Pareto m1 m1 /k
set at point of interest
m1
equation of line tangent to Pareto y1 = m 1 x 1 + b y1 = k (kx1 ) + b1
set at point of interest y 1 = m 1 x 1 + b1
slope of line normal to Pareto −1/m1 −1/m1
set at point of interest
Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.