See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225485886

The weighted sum method for multi-objective optimization: New insights

Article  in  Structural and Multidisciplinary Optimization · June 2010

DOI: 10.1007/s00158-009-0460-7

CITATIONS READS

738 23,719

2 authors:

R. Timothy Marler Jasbir Singh Arora

RAND Corporation University of Iowa
89 PUBLICATIONS   5,131 CITATIONS    314 PUBLICATIONS   13,902 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Gruntsim View project

MALUM Terminus View project

All content following this page was uploaded by R. Timothy Marler on 22 May 2014.

The user has requested enhancement of the downloaded file.

Struct Multidisc Optim
DOI 10.1007/s00158-009-0460-7

RESEARCH PAPER

The weighted sum method for multi-objective

optimization: new insights
R. Timothy Marler · Jasbir S. Arora

Received: 17 April 2009 / Revised: 17 September 2009 / Accepted: 17 November 2009


c Springer-Verlag 2009

Abstract As a common concept in multi-objective opti- 1 Introduction

mization, minimizing a weighted sum constitutes an inde-
pendent method as well as a component of other methods. Despite deficiencies with respect to depicting the Pareto
Consequently, insight into characteristics of the weighted optimal set, the weighted sum method for multi-objective
sum method has far reaching implications. However, despite optimization (MOO) continues to be used extensively not
the many published applications for this method and the lit- only to provide multiple solution points by varying the
erature addressing its pitfalls with respect to depicting the weights consistently, but also to provide a single solu-
Pareto optimal set, there is little comprehensive discussion tion point that reflects preferences presumably incorporated
concerning the conceptual significance of the weights and in the selection of a single set of weights. However, the
techniques for maximizing the effectiveness of the method method’s effectiveness in this latter capacity has not been
with respect to a priori articulation of preferences. Thus, studied thoroughly. To be sure, there are many different
in this paper, we investigate the fundamental significance approaches for determining the weights (Marler and Arora
of the weights in terms of preferences, the Pareto optimal 2004), but ultimately, however, these are all just different
set, and objective-function values. We determine the factors processes for organizing one’s preferences and priorities.
that dictate which solution point results from a particular Instead of proposing yet another algorithm for translating
set of weights. Fundamental deficiencies are identified in preferences into weights, we focus on the mathematical
terms of a priori articulation of preferences, and guidelines characteristics of the solution and on the fundamental con-
are provided to help avoid blind use of the method. ceptual meaning of the weights. Furthermore, we determine
the factors that dictate which Pareto optimal solution point
Keywords Multi-objective optimization · Weighted sum · results from a particular set of weights. Through this anal-
Preferences ysis, we augment the currently available literature and pro-
vide new insight into the nature of the method and new ideas
for maximizing its effectiveness with respect to articulating
preferences.
New understanding of the fundamentals presented in this
paper, is useful in setting the weights directly (without the
use of an additional algorithm). In addition, understanding
inadequacies that are inherent in the method is instructive,
R. T. Marler (B)
and in this vein, we present deficiencies that cannot be
Center for Computer Aided Design, College of Engineering;
111 ERF, The University of Iowa, Iowa City, IA, 52242, USA avoided and that have not been discussed in the current
e-mail: [email protected] literature.
Setting weights is just one approach to articulating pref-
J. S. Arora
erences, and it is applicable to many different methods.
Center for Computer Aided Design, College of Engineering;
238 ERF, The University of Iowa, Iowa City, IA, 52242, USA Consequently, understanding how the weights affect the
e-mail: [email protected] solution to the weighted sum method has implications
 R.T. Marler, J.S. Arora

concerning other approaches that involve similar method- while its use for a posteriori articulation has been addressed
parameters. thoroughly in the literature.
Using the weighted sum method to solve the problem in
1.1 Overview of the formulation (1) entails selecting scalar weights wi and minimizing the
following composite objective function:
The general multi-objective optimization (MOO) problem
is posed as follows: 
k
U= wi Fi (x) (2)
i=1
Minimize : F (x) = [F1 (x) , F2 (x) , · · · , Fk (x)]T
x (1)
subject to : g j (x) ≤ 0; j = 1, 2, . . . , m If all of the weights are positive, as assumed in this
study, then minimizing (2) provides a sufficient condition
where k is the number of objective functions and m is the for Pareto optimality, which means the minimum of (2)
number of inequality constraints. x ∈ E n is a vector of is always Pareto optimal (Zadeh 1963; Goicoecheak et al.
design variables, and F (x) ∈ E k is a vector of objective 1982). Although some literature indicates that i=1 wi = 1
functions Fi (x) : En → 1
 E . The feasible design space and w ≥ 0, if any one of the weights is zero, there is a

is defined as X = x g j (x) ≤ 0, j = 1, 2, ..., m . The potential for the solution to be only weakly Pareto optimal.
feasible criterion space is defined as Z = {F (x) |x ∈ X }. With regards to the design space, minimizing the
The idea of a solution for (1) can be unclear, because weighted sum provides a necessary condition if the multi-
a single point that minimizes all objectives simultaneously objective problem is convex, which means the feasible
usually does not exist. Consequently, the idea of Pareto opti- design space is convex (each constraint is convex) and all of
mality is used to describe solutions for MOO problems. A the objective functions are convex (Geoffrion 1968; Koski
solution point is Pareto optimal if it is not possible to move 1985; Miettinen 1999). A function defined on a convex set
from that point and improve at least one objective function is convex if and only if the Hessian matrix of the function
without detriment to any other objective function. Alterna- is positive semidefinite or positive definite at all points in
tively, a point is weakly Pareto optimal if it is not possible the set. Thus, if the Hessian matrix for each constraint and
to move from that point and improve all objective functions for each objective function is positive semidefinite or posi-
simultaneously. tive definite, then the weighted sum method can provide all
Solving a particular MOO formulation can serve as a Pareto optimal points (the complete Pareto optimal curve).
necessary and/or sufficient condition for Pareto optimality, With regards to the criterion space, because the Pareto
and a method’s performance in this capacity is important. optimal set is always on the boundary of the feasible crite-
However, these characterizations veer slightly from their rion space Z, some studies suggest that Z must be convex
meaning in terms of single-objective optimization. If a for- for the weighted sum method to yield all of the Pareto opti-
mulation provides a necessary condition, then for a point mal points (Stadler 1995; Athan and Papalambros 1996).
to be Pareto optimal, it must be a solution to that formu- However, Lin (1975) demonstrates that, in fact, if Z is
lation. Consequently, the method can yield every Pareto only p-directionally convex (Holtzman and Halkin 1966)
optimal point as a solution. However, some solutions to a for any definitely negative vector p, then the weighted sum
formulation that provides a necessary condition may not be approach provides a necessary condition. The term defi-
Pareto optimal. Alternatively, if a formulation provides a nitely negative implies p < 0. p-directionally convex is
sufficient condition, then its solution is always Pareto opti- defined as follows:
mal, although such a formulation may not be able to capture
Definition 1 p-Directionally Convex: Given a nonzero vec-
all of the Pareto optimal points.
tor p ∈ E k , Z ⊂ E k is said to be p-directionally convex if
Typically, there are infinitely many Pareto optimal solu-
given any two different points in Z, F1 and F2 , and any two
tions for a multi-objective problem. Thus, it is often neces-
positive scalars, w1 and w2 , with w1 + w2 = 1, there is a
sary to incorporate user preferences for various objectives in
positive number β such that w1 F1 + w2 F2 + βp∈Z.
order to determine a single suitable solution. With methods
that incorporate a priori articulation of preferences, the user This means that only the portion of the feasible criterion
indicates preferences before running the optimization algo- space that is visible when looking in the direction of p from
rithm and subsequently allows the algorithm to determine within the feasible space needs to be convex. Therefore,
a single solution that presumably reflects such preferences. only the Pareto optimal hypersurface needs to be convex for
Alternatively, with a posteriori articulation of preferences, the weighted sum method to provide a necessary condition
one manually selects a single solution from a representa- for Pareto optimality. Currently, the authors do not know of
tion of the Pareto optimal set. This paper focuses on the any method for predicting whether or not the Pareto optimal
use of the weighted sum method for a priori articulation, hypersurface is convex.
The weighted sum method for multi-objective optimization

1.2 Review of the literature and motivation lation of preferences. First, many authors demonstrate the
method’s inability to capture Pareto optimal points that lie
Following the introduction of the weighted sum method on non-convex portions of the Pareto optimal curve (Koski
by Zadeh (1963), the method has been mentioned promi- 1985; Stadler and Dauer 1992; Stadler 1995; Chen et al.
nently in the literature. However, most general treatments of 1999; Athan and Papalambros 1996; Huang et al. 2007).
MOO simply outline the weighted sum approach and indi- Secondly, it is well known that the method does not pro-
cate that it provides a Pareto optimal solution. Intricacies of vide an even distribution of points in the Pareto optimal set,
the method and of the solutions that it yields are typically with a consistent change in weights, where even implies a
not discussed. In particular, the significance of the weights consistent Euclidean distance between consecutive solution
is not thoroughly explored, and thus, despite the presence of points. Das and Dennis (1997) evaluate why this deficiency
many algorithms for determining weight values, no funda- occurs, and for a two-objective problem, they derive the
mental guidelines have been presented for selecting weights explicit form of the unique Pareto optimal curve for which
for accurate a priori articulation of preferences. an even distribution of weights yields an even distribution of
The weighted sum method is often presented strictly as Pareto solutions. In general, however, the points that result
a tool, especially over the past few years, and literature from a systematic selection of weights may cluster together.
regarding examples of applications is extensive. However, These two deficiencies are often answered with alternative
the focus is on the application, and the problems tend to methods for representing a Pareto optimal set (Das and
be limited to those with just two objective functions. For Dennis 1998; Tappeta et al. 2000; Messac and Mattson
instance, as development for a new approach, Koski and 2002; Messac et al. 2003; Huang et al. 2007; Zhang et al.
Silvennoinen (1987) provide an early application and use 2008). Such alternatives can be effective and valuable, but
the weighted sum method to obtain multiple Pareto opti- they are usually independent of the weighted sum method.
mal solutions with a systematic change in the weights, Although the difficulties mentioned above are well doc-
while minimizing the volume and the nodal displacement umented, they concern the method’s use for yielding a
of a four-bar space truss. Kassaimah et al. (1995) use complete Pareto optimal set rather than a single solution (a
the method for the two-objective optimization of laminated priori articulation of preferences). Criteria for maximizing
plates, where critical buckling shear stress is maximized and the effectiveness of the method in the latter respect are not
deflection is minimized. A few different somewhat arbitrary available. In fact, there are few, if any numerical studies of
sets of weights are considered, and the corresponding solu- the weighted sum method that focus on a priori articulation
tions are compared, but the method itself is not studied and of preferences. Essentially, there is no fundamental unifying
the articulation of preferences is not considered. Proos et al. analysis of the method.
(2001) apply the method to topology optimization, mini- Many researchers have, however, developed systematic
mizing compliance and maximizing the first mode of the approaches to selecting weights, surveys of which are pro-
natural frequency, for two-dimensional plane stress prob- vided by Eckenrode (1965), Hobbs (1980), Hwang and
lems. Again, the weights are altered to yield a representation Yoon (1981), and Voogd (1983). We identify two broad
of the Pareto optimal set. Saramago and Steffen (1998) use classes of approaches, but with these approaches, weights
a weighted sum to combine two objective functions in an may still be set ineffectively in terms of accurately articulat-
optimization-based approach to predicting robotic motion, ing preferences. With rating methods, the decision-maker
but the weights have the same value and are thus irrelevant. assigns independent values of relative importance to each
Marler and Arora (2005) study a three-objective prob- objective function. With ranking methods (Yoon and Hwang
lem, but they use the weighted sum method as a platform for 1995), which can be viewed as a subset of rating methods,
studying various function-transformation methods and their the different objective functions are ordered by importance.
affect on the depiction of the Pareto optimal set. It is found Then, the least important objective receives a weight of one,
that a convex combination of functions is advantageous for and integer weights with consistent increments are assigned
depicting the Pareto optimal set, as apposed to determin- to objectives that are more important. The same approach
ing a single solution. In addition, it is shown that some is used with categorization methods, in which different
transformation methods can actually be detrimental to the objectives are grouped in broad categories such as highly
process of obtaining a diverse and uniform spread of Pareto important and moderately important. Ratio questioning or
optimal points, and criteria are presented for determining paired comparison methods are also common and provide
when certain transformation methods (in conjunction with systematic means to rate objective functions by comparing
the weighted sum method) may not perform well. two objectives at a time. In this vein, Saaty (1977, 2003),
Some literature provides a more substantial look at the and Saaty and Hu (1998) provide an eigenvalue method
method, focusing on its two primary deficiencies when for determining weights, which involves k(k − 1)/2 pair-
representing the Pareto optimal set for a posteriori articu- wise comparisons between objective functions. This yields
 R.T. Marler, J.S. Arora

a comparison matrix, and the eigenvalues of the matrix are A preference function is an abstract function (of points
used as the weights. Rao and Roy (1989) discuss a method in the criterion space) in the mind of the decision-maker,
for determining weights based on fuzzy set theory. While which perfectly incorporates his/her preferences. Most
disregarding preferences, Gennert and Yuille (1988) suggest MOO methods that involve minimizing a single aggregated
that the weights should be selected to maximize the final objective function, attempt to approximate the preference
minimum value of the weighted sum function. For the most function with some mathematical representation, called a
part, algorithms for determining weights that reflect user utility function. The gradients of the preference function
preferences either fall into the category of rating methods P[F(x)] and the utility function in (2) are given respectively
or paired comparison methods. as follows:
Although the above-mentioned algorithms provide valid
processes for determining the weights, there is little rigor- 
k
∂P
∇x P [F (x)] = ∇x Fi (x) (3)
ous investigation as to what the weights actually represent. ∂ Fi
i=1
Thus, this paper studies the significance of the weights
themselves. By considering the fundamental significance of
the weights from different viewpoints, we uncover intrinsic

k
flaws in the rating methods and paired comparison meth- ∇x U = wi ∇x Fi (x) (4)
ods for a priori articulation of preferences. In addition, we i=1
identify which factors govern which solution is determined
from a specific set of weights, and we explain why even Each component of the gradient ∇x P qualitatively repre-
with an effective algorithm that helps one decide on values sents how the decision-maker’s satisfaction changes with
for the weights, a weighted sum is not necessarily capable a change in the design point and a consequent change in
of incorporating preferences completely. function values. Comparing (3) and (4) suggests that if the
weights are selected properly, then the utility function can
have a gradient that is parallel to the gradient of the prefer-
ence function. This is significant because the purpose of any
2 Interpretation of the weights
utility function is to approximate the preference function,
so imposing similarities between these two functions such
One often views the weights as general gauges of relative
as parallel gradients, can result in a more accurate repre-
importance for each objective function. However, selecting
sentation of one’s preferences. If the utility function should
a set of weights that reflects preference towards one objec-
ideally be the same as the preference function, then certainly
tive or another can be difficult, because preferences tend
the gradients of the two functions should be the same as
to be indistinct. In addition, even with full knowledge of
well.
the objectives and satisfactory selection of weights, the final
The above relationships indicate that wi represents
solution may not necessarily reflect the intended preferences
∂ P/∂ Fi . Conceptually, ∂ P/∂ Fi is the approximate change
that are supposedly incorporated in the weights. Nonethe-
in the preference-function value (change in a decision-
less, the specific Pareto optimal point that is provided as the
maker’s satisfaction) that results from a change in the
solution depends on which weights are used, so it is impor-
objective-function value for Fi . In this way, ∂ P/∂ Fi pro-
tant to determine how the weights relate to preferences,
vides a mathematical definition or metric for the importance
to the Pareto optimal set, and to the individual objective
of Fi . However, it only makes sense to consider the impor-
functions. These relationships are explored in this section.
tance of an objective or change in preference-function value,
in relative terms. Thus, the value of a weight is signifi-
2.1 Preferences cant relative to the values of other weights; the independent
absolute magnitude of a weight is irrelevant in terms of
First, the relationship between preferences and weights is preferences.
examined. Concurrently, the theoretical foundation of rat-
ing methods is studied. We discover that such methods 2.1.1 Deficiencies in the method
allow one to inadvertently set the weights according to the
relative magnitude of the objective functions rather than Although (3) and (4) provide a general, conceptual idea of
the relative importance of the objectives. In this way, we what the weights represent in terms of preferences, using
demonstrate how function-transformation can be advanta- them to set precise values for weights can be difficult. In
geous with a priori articulation of preferences. In addition, fact, a fundamental deficiency in the weighted sum method
we reveal inherent deficiencies in the weighted sum method is that it can be difficult to discern between setting weights to
with respect to its capacity for incorporating preferences. compensate for differences in objective-function magnitudes
The weighted sum method for multi-objective optimization

and setting weights to indicate the relative importance of an inate the aggregate objective function can help one set the
objective as is done with the rating methods. For practical weights to reflect preferences more accurately.
purposes, ∂ P/∂ Fi can be approximated as P/Fi , and As an example demonstrating the significance of the rel-
when objective functions have different ranges and orders ative values of the weights, consider the following MOO
of magnitude, an appropriate value for Fi may not be problem:
apparent. For example, consider two objective functions to
be minimized, where function-one is stress, which for the Find : x = [x1 , x2 ]T
sake of argument ranges between 10,000 and 40,000 psi, F = 20 (x1 − 0.75)2 + (2x2 − 2)2 (5)
to minimize : 1
and function-two is displacement, which ranges between 1 x F2 = (x1 − 2.5)2 + (x2 − 1.5)2
and 2 in. If displacement were to decrease by one unit, the
consequence would be a significant improvement; P/F2 The design space for this problem is shown in Fig. 1, and
would be relatively high. Concurrently, if stress were to the criterion space is shown in Fig. 2.
decrease by a unit of one, the consequence would be negligi- When F1 is minimized independently, it has a minimum
ble; P/F1 would be insignificant. The result would be a value of zero, and F2 has a value of 3.3125. When F2 is min-
relatively large value for w2 , compared to w1 . However, this imized independently, it has a minimum value of zero, and
result does not necessarily reflect the relative importance of F1 has a value of 62.25. F1 has a range between zero and
the objectives; the values for the weights have been dictated 62.25, and F2 has a range between zero and 3.3125. Thus,
purely by the relative magnitudes of the objective functions. the range of values for F1 is much larger than that of F2 . If
Alternatively, one could consider percentage changes in this problem is solved using the weighted sum method, and
the objective functions rather than an absolute change of weights are selected with relative values to suggest equal
unity as discussed above. However, then the question arises importance of eth two objectives (i.e. w1 = w2 = 1), then
as to what one should consider when evaluating percent- the solution is x = (0.833, 1.100) and F = (0.179, 2.938).
ages. Should Fi be a percentage of the average value for Yet, even when considered to be equally important, F1 has
Fi , the maximum value for Fi , the minimum value for Fi , a value close to its absolute minimum, and F2 has a value
or some factor of the range of values for Fi ? The choice is close to its maximum. If the weights are set to suggest that
arbitrary. F1 is twice as important as F1 (i.e. w1 = 1, w2 = 2), then
Consequently, one may set the weights to reflect the solution is x = (0.909, 1.167) and F = (0.179, 2.938).
objective-importance directly, but this approach can also Even when one weight is twice as large as the other, the
lead to difficulties. With the previous example, suppose dis- solution does not change a significant amount. However,
placement is twice as important as stress. Considering the when the objective functions are divided by their respective
relative value of the weights, if w1 = 1, then w2 = 2. maxima and thus normalized, then it becomes easy to set the
However, given the relative magnitudes of the objective weights such that they are significant relative to each other
functions, stress would still dominate the weighted sum, and and relative to the objective functions values. Then, with the
the use of w2 = 2 would be irrelevant. Again, the magni- objective functions normalized such that they both a have a
tudes of the objective functions affect how the weights are range between zero and one, if the weights are again set
set, and the articulation of preferences becomes blurred at
best.
This difficulty arises because what dictate the solution to x2

the weighted sum are the relative magnitudes of the terms 4

wi Fi , in (2), not just the magnitudes of the functions or

weights alone. The value of a weight is significant not only 3 Pareto Optimal
Set (Segment BC)
relative to other weights but also relative to its correspond-
ing objective function. This is a critical idea, although it is
2
often overlooked. With many weighted methods, including C D
the weighted sum method, preferences are articulated a pri- B F2
ori by using weights to induce the dominance of a particular 1
term in a formulation (and the objective function associated
F1
with the term). However, the process of selecting weights 0
and thus indicating preferences is complicated if this dom-
inance is instead, induced by relative function-magnitude. A
-1 x1
Thus, when using weights to represent the relative impor-
0 1 2 3 4
tance of the objectives, transforming the functions so that
they all have similar magnitudes and do not naturally dom- Fig. 1 Example problem design space
 R.T. Marler, J.S. Arora

F2
6 It is interesting to note that from a computational per-
D
spective, nonlinear composite objective functions often help
5
to find Pareto optimal solutions on non-convex Pareto opti-
4 Pareto Optimal mal sets more effectively (Messac et al. 2000a, b; Marler
C Set (Segment BC)
3 2009). Such functions typically involve various types of
2
parameters, not just weighting factors. Thus, they provide
a more complex and versatile utility function for altering
1
B
A preferences.
F1
20 40 60 80 100
2.2 Pareto optimal set
Fig. 2 Example problem criteria space
In this section and the next, factors that affect the solution
to the weighted sum method are determined. In addition,
such that w1 = w2 = 1, the solution is x = (1.598, 1.412) fundamental characteristics of the weights are presented
and F = (15.054, 0.822). with regard to the Pareto optimal set, whereas the pre-
Further study of (2), (3), and (4) reveals another funda- vious section concerned fundamental characteristics with
mental deficiency in the weighted sum method: the weighted respect to preferences. The characteristics discussed here
sum utility function is only a linear approximation (in the provide a theoretical foundation for the paired comparison
criterion space) of the preference function. Equation (2) methods. We point out a fundamental deficiency with such
is a linear function of the objective functions; it is the methods for articulating preferences, and we discover that
most basic approximation of one’s preference function. The function transformations can actually be detrimental with
consequences of this characteristic are explained as fol- these methods.
lows. ∂ P/∂ Fi may actually be a nonlinear function of Figure 3 depicts a general illustrative model of a Pareto
the objective functions, changing from point to point in optimal set in the criterion space, for two objective func-
the criterion space. However, using a weighted sum as tions.
a utility function to approximate the preference function, The weights represent the gradient of U in (2) with
inherently assumes that this term is constant, represented respect to the vector function F(x), shown as follows:
by a scalar weight. Thus, values that are selected for the ⎧ ⎫
⎪ ∂U ⎪
weights are only locally relevant. That is, they are not ⎪
⎨ ⎪
⎬  
necessarily appropriate for the complete criterion space, ∂ F1 w1
∇F U = = (6)
only for the point at which they are determined, presum- ⎪
⎪ ∂U ⎪ ⎪ w2
⎩ ⎭
ably the starting point (initial guess). As the design point ∂ F2
changes with an iterative optimization process, the values
of the objective functions change, and consequently, prefer- The Pareto optimal solution (for a given set of weights) is
ences (between various points in the criterion space) may found by determining where the U -contour with the lowest
change in a nonlinear fashion. For instance, in the con-
text of the stress-displacement example described above,
minimizing displacement may be more important than min-
F2
imizing stress only until stress reaches a relatively high
Pareto Z
value, suggesting potential failure. Then, minimizing stress
Optimal
becomes more critical, and its associated weighting value Set
should be increased. Anticipating such a change and altering Direction of
weighting values while an optimization problem is running Decreasing
is impractical. Thus, decision-makers implicitly assume that U-values
Δ−
∂ P/∂ Fi is constant, when they determine a set of weights to ( FU )
reflect the relative importance of the objectives. This means
that one assumes that the gradient of the preference function Utility Function Contours
U = w1 F1 ( x ) + w2 F2 ( x )
(in the criterion space) is constant, which is not necessarily
the case. This is why even with a process that enables one to
determine acceptable values for the weights, the final solu-
tion to the weighted sum problem may not accurately reflect F1
initial preferences that were supposedly incorporated in the
weights. Fig. 3 Pareto optimal set in the criterion space
The weighted sum method for multi-objective optimization

possible value intersects the boundary of the feasible crite- The left side of (9) can be approximated as F2 /F1 . Then,
rion space. The U -contours are linear and are thus tangent assuming w1 = 1 is again a reference weight and F1 = 1
to the Pareto optimal curve at the solution point. Changes (comparable to using 1 = 1 as above), (9) is reduced to
in the relative values of the weights result in changes in the
1
orientation of the U -contours. The subsequent change in the w2 ≈ − (10)
F2
solution depends on the shape of the Pareto optimal curve in
the criterion space. This has significant repercussions when There is no negative sign in (8), because the numerator is
trying to provide an even spread of Pareto optimal points by assumed to represent a decrease of one.
solving a series of weighted sum problems, for a posteriori When using the concepts described above, the objective
articulation of preferences. In fact, some more complicated functions should not be transformed. This is because using
methods have been developed so that the shape of the Pareto (7) involves comparing the different objective functions and
optimal hypersurface has a minimal affect the accuracy with evaluating trade-offs (how much of a loss in an objective
which it is depicted (Marler and Arora 2004). is one willing to sacrifice for a gain in another objective).
In terms of higher dimensions (more than two objec- In contrast to the idea of evaluating relative importance,
tive functions), the minimum value of (2) for a fixed set which was discussed with regard to the rating methods,
of weights determines a supporting hyperplane to Z (i.e. assigning values to the weights based on trade-offs can be
U (F) = constant) (Zadeh 1963; Gembicki 1974), and the hampered by function transformations. Trade-off compar-
normal of the hyperplane is the vector of weights. The isons between functions are easier when the functions retain
Pareto optimal hypersurface is always on the boundary of Z, their original physical significance and original units. How-
so the supporting hyperplane is tangent to the hypersurface ever, most transformation methods yield unitless functions,
at the point that minimizes the weighted sum. the values of which have little or no physical significance.
Steuer (1989) uses such a hyperplane to provide an When using the stress-displacement example, one may
interpretation of the weights, which we suggest forms the be willing to accept an increase in displacement of 0.002 in.
basis for paired comparison methods. This interpretation if stress were to reduce by 5,000 psi. If stress is the reference
is explained as follows. w1 = 1 is considered a reference function, then a decrease (improvement) of 1 psi must cor-
weight, and F1 is the reference function (theoretically, any respond to a displacement increase of 5,000 × 0.002 = 10.
function can be used as a reference). Then, i represents the Therefore, if the reference weight is w1 = 1, then w2 = 10.
amount by which Fi must increase in order to compensate However, if the objectives are transformed or normalized
for a decrease (improvement) of 1 in F1 , while remaining and thus become unitless, it becomes difficult to conduct
on the hyperplane. Each weight is then defined as any type of trade-off analysis; it is difficult to conceptualize
what level of increase in normalized displacement one might
1 accept in exchange for a reduction in normalized stress of
wi = lim (7) one.
i →0 i
Although (7) is derived based on the mathematical char-
acteristics of (2), it provides another general method for
If one assumes 1 = 1, then each weight is approximated as articulating preferences, and here we point out an inher-
ent deficiency in this method. Note that (7) is a relationship
1 that necessarily exists at the solution point for the weighted
wi ≈ (8)
i sum. However, a user’s decisions concerning trade-offs
can change depending on the point at which the differ-
ent objective functions are evaluated. One may be forced
Thus, wi represents the trade-off between Fi and the ref-
to determine trade-offs based on values for the different
erence function, at the solution point to the weighted sum
functions when the functions are evaluated at a point other
problem. This is the idea behind paired comparison methods
than the solution point to which (7) applies. Therefore, as
for setting weights.
with (3) and (4), (7) only provides an approximate means
To clarify this idea, note that one can draw the same con-
for incorporating preferences in the weighted sum utility
clusion using Fig. 3 for the case of two objective functions.
function a priori.
Considering that the solution to the weighted sum problem
Figure 3 and (6) provide reinforcement of the conclu-
is always Pareto optimal, the slope of the Pareto optimal
sions drawn in Section 2.1 concerning the significance of
curve in Fig. 3 is determined as follows:
the weights. Note that the direction of the gradient of U
in the criterion space, and thus the orientation of the U-
d F2 w1 contours and corresponding solution point, depends on the
=− (9)
d F1 w2 relative values (not the absolute values) of the weights. In
 R.T. Marler, J.S. Arora

addition, the orientation of the U-contours relative to the 3 Discussion and conclusions
Pareto optimal set also depends on the range of objective-
function values for points in the set. That is, for a given set This paper has provided insight into how the weighted sum
of weights (a given gradient vector for the weighted sum), method works and has explored the significance of the
scaling an objective function changes the Pareto optimal weights with respect to preferences, the Pareto optimal set,
solution point that the weighted sum method provides. Thus, and the objective-function values. We have revealed fun-
as suggested earlier, the value of a weight is significant rel- damental deficiencies with respect to preferences, and we
ative to the values of other weights and relative to the value have shown that although the weighted sum method is easy
of its corresponding objective function. to use, it provides only a linear approximation of the pref-
erence function. Thus, the solution may not preserve one’s
2.3 Objective functions initial preferences no matter how the weights are set, a cru-
cial idea that is often overlooked. The solution depends on
Whereas the discussion above considers the significance multiple factors, one of which is the relative magnitude of
of weights in terms of the Pareto optimal set, this section the objective functions. However, when setting the weights,
focuses on the relationship between a set of weights and only the relative importance of the objectives should be
the objective function values. First, consider a problem with considered, not the relative magnitudes of the function val-
two unconstrained objective functions: ues. In this way, we argue that function-transformation can
be helpful with a priori articulation of preferences, when
U = w1 F1 (x) + w2 F2 (x) (11) the weights are used to represent the relative importance
of the objectives (as with rating methods). This finding
For U to have a minimum, it is necessary that the gradient
provides an important guideline that should always be con-
of U be equal to zero, as follows:
sidered when setting weights and/or transforming objective
∇x U = w1 ∇x F1 + w2 ∇x F2 = 0 (12) functions. Alternatively, we argue that if the weights are
viewed as representing trade-offs between objective func-
Assuming the weights are positive and noting that mini- tions (as with paired comparison methods), then retaining
mizing a weighted sum always provides a Pareto optimal the original units of the objectives, without transformation,
solution, (12) indicates that at a Pareto optimal point, the is advantageous.
gradients of the two objective functions are co-linear and Much of the discussion in this paper pertains to a pri-
point in opposite directions. Essentially, the linear com- ori articulation of preferences, and although
ksome literature
bination of the gradients equals zero. In compliance with suggests that the weights be set such that i=1 wi = 1 and
the definition of Pareto optimality, this suggests that mov- w ≥ 0, which results in a convex combination of objective
ing from a solution that satisfies (12), in order to improve a functions, when using the weighted sum for a priori articula-
function, is detrimental to at least one other function. tion of preferences, there is no need to place any restriction
General constrained problems are considered using on the values of the weights other than w ≥ 0, which
Karush–Kuhn–Tucker necessary conditions, which are sim- ensures Pareto optimality. In fact, using unrestricted weights
plified as follows: can be beneficial, making the determination of appropriate
weight values easier. However, when systematically alter-

k 
ma
ing the weights for a posteriori articulation of preferences,
wi ∇x Fi = − μi ∇x gi (13)
using a convex combination of objective function can help
i=1 i=1
avoid repeating weighting vectors in terms of relative values
where m a is the number of active constraints and μi is (Marler and Arora 2005). Thus, weights should be treated
the Lagrange multiplier for gi(x). Equation (13) indicates differently when used for a priori articulation than they are
that a linear combination of the objective function gradi- when used for a posteriori articulation.
ents is equal and opposite to a linear combination of the Conclusions based on this work and guidelines for setting
constraint gradients (for active constraints), at a Pareto opti- weights are presented as follows:
mal point. If a constraint is active at the solution point for a
weighted sum, which necessarily provides a Pareto optimal 1) When setting weights directly for a priori articulation
point, then that constraint forms part of the Pareto optimal of preferences, the value of a weight must be sig-
set. Equations (12) and (13) suggest that the Pareto opti- nificant relative to other weights and relative to its
mal point resulting from the use of a specific set of weights corresponding objective function.
depends on active constraints that form part of the Pareto 2) When using a ranking method to set weights, all objec-
optimal set and on the relationship between the gradients of tive functions should be transformed such that they have
the different objective functions. similar ranges.
The weighted sum method for multi-objective optimization

3) When using a paired comparison method to set weights Geoffrion AM (1968) Proper efficiency and the theory of vector
or when viewing the weights as trade-offs, objective maximization. J Math Anal Appl 22:618–630
Goicoechea A, Hansen DR, Duckstein L (1982) Multiobjective deci-
functions should not be transformed/normalized. sion analysis with engineering and business applications. Wiley,
4) When articulating well-understood preferences with New York
the paired comparison methods, unrestricted positive Hobbs BF (1980) A comparison of weighting methods in power plant
weights should be used. siting. Decis Sci 11:725–737
Holtzman JM, Halkin H (1966) Directional convexity and the maxi-
5) The weighted sum provides only a basic approximation mum principle for discrete systems. SIAM J Control 4:263–275
of one’s preference function. It is fundamentally inca- Huang C-H, Galuski J, Bloebaum CL (2007) Multi-objective Pareto
pable of incorporating complex preference information. concurrent subspace optimization for multidisciplinary design.
Thus, even if one determines acceptable values for the AIAA J 45:1894–1906
Hwang C-L, Yoon K (1981) Multiple attribute decision making, meth-
weights a priori, the final solution may not accurately ods and applications: a state-of-the-art survey. In: Beckmann M,
reflect initial preferences. Kunzi HP (eds) Lecture notes in economics and mathematical
6) The Pareto optimal solution that results from a specific systems, no 186. Springer, Berlin
set of weights depends on constraints that form part of Kassaimah SA, Mohamed AM, Kolkailah FA (1995) Bi-criteria opti-
mum design of laminated plates under uniform load and shear.
the Pareto optimal set, the relationships between the In: Proceedings of the 27th international SAMPLE technical
gradients of the different objective functions, the rel- conference (held in Albuquerque, NM), 27, pp 731–737
ative magnitudes of the objective functions, and the Koski J (1985) Defectiveness of weighting method in multicrite-
shape of the Pareto optimal hypersurface. rion optimization of structures. Commun Appl Numer Methods
1:333–337
7) The weighted sum method provides a basic and easy- Koski J, Silvennoinen R (1987) Norm methods and partial weighting
to-use approach for multi-objective optimization and in multicriterion optimization of structures. Int J Numer Methods
is useful as such, but alternate methods should be Eng 24:1101–1121
considered for the following cases: Lin JG (1975) Three methods for determining Pareto-optimal solu-
tions of multiple-objective problems. In: Ho YC, Mitter SK (eds)
a. Non-convex or potentially non-convex problems Directions in large-scale systems. Plenum, New York
where it is critical to depict the complete Pareto Marler RT (2009) A study of multi-objective optimization methods for
engineering applications. VDM, Saarbrucken
optimal set thoroughly Marler RT, Arora JS (2004) Survey of multi-objective optimization
b. When one wishes to depict the Pareto optimal set methods for engineering. Struct Multidiscipl Optim 26:369–395
for even a convex problem thoroughly, but needs Marler RT, Arora JS (2005) Transformation methods for multi-
to minimize computational expenses (minimize the objective optimization. Eng Optim 37:551–569
Messac A, Mattson CA (2002) Generating well-distributed sets of
number of times (1) must be solved) Pareto points for engineering design using physical programming.
c. When it is critical to articulate complex preferences Eng Optim 3:431–450
accurately Messac A, Sukam CP, Melachrinoudis E (2000a) Aggregate objective
functions and Pareto frontiers: required relationships and practical
implications. Optim Eng 1:171–188
Messac A, Sundararaj GJ, Tappeta RV, Renaud JE (2000b) Ability
References of objective functions to generate points on nonconvex Pareto
frontiers. AIAA J 38(6):1084–1091
Athan TW, Papalambros PY (1996) A note on weighted criteria meth- Messac A, Ismail-Yahaya A, Mattson CA (2003) The normalized nor-
ods for compromise solutions in multi-objective optimization. mal constraint method for generating the Pareto frontier. Struct
Eng Optim 27:155–176 Multidiscipl Optim 25:86–98
Chen W, Wiecek MM, Zhang J (1999) Quality utility—a compromise Miettinen K (1999) Nonlinear multiobjective optimization. Kluwer
programming approach to robust design. J Mech Des 121:179– Academic, Boston
187 Proos KA, Steven GP, Querin OM, Xie YM (2001) Multicriterion
Das I, Dennis JE (1997) A closer look at drawbacks of minimiz- evolutionary structural optimization using the weighted and the
ing weighted sums of objectives for Pareto set generation in global criterion methods. AIAA J 39:2006–2012
multicriteria optimization problems. Struct Optim 14:63–69 Rao JR, Roy N (1989) Fuzzy set theoretic approach of assigning
Das I, Dennis JE (1998) Normal-boundary intersection: a new method weights to objectives in multicriteria decision making. Int J Syst
for generating the Pareto surface in nonlinear multicriteria opti- Sci 20:1381–1386
mization problems. SIAM J Optim 8:631–657 Saaty TL (1977) A scaling method for priorities in hierarchies, multi-
Eckenrode RT (1965) Weighting multiple criteria. Manage Sci 12:180– ple objectives and fuzzy sets. J Math Psychol 15:234–281
192 Saaty TL (2003) Decision-making with the AHP: why is the principal
Gembicki FW (1974) Performance and sensitivity optimization: a eigenvalue necessary. Eur J Oper Res 145:85–91
vector index approach. PhD dissertation, Case Western Reserve Saaty TL, Hu G (1998) Ranking by eigenvector versus other methods
University, Cleveland, OH in the analytic hierarchy process. Appl Math Lett 11:121–125
Gennert MA, Yuille AL (1988) Determining the optimal weights in Saramago SFP, Steffen V Jr (1998) Optimization of the trajectory plan-
multiple objective function optimization. In: Second international ning of robot manipulators taking into account the dynamics of
conference on computer vision (held in Los Alamos, CA), Insti- the system. Mech Mach Theory 33:883–894
tute of Electrical and Electronics Engineers, Piscataway, NJ, pp Stadler W (1995) Caveats and boons of multicriteria optimization.
87–89 Microcomput Civ Eng 10:291–299
 R.T. Marler, J.S. Arora

Stadler W, Dauer JP (1992) Multicriteria optimization in engineering: Voogd H (1983) Multicriteria evaluation for urban and regional plan-
a tutorial and survey. In: Kamat MP (ed) Structural optimiza- ning. Pion, London
tion: status and promise. American Institute of Aeronautics and Yoon KP, Hwang C-L (1995) Multiple attribute decision making, an
Astronautics, Washington, DC introduction. Sage, London
Steuer RE (1989) Multiple criteria optimization: theory, computation, Zadeh LA (1963) Optimality and non-scalar-valued performance crite-
and application. Krieger, Malabar ria. IEEE Trans Automat Contr AC-8:59–60
Tappeta RV, Renaud JE, Messac E, Sundararaj GJ (2000) Interactive Zhang K-S, Han Z-H, Li W-J, Song W-P (2008) Bilevel adaptive
physical programming: tradeoff analysis and decision making in weighted sum method for multidisciplinary multi-objective opti-
multicriteria optimization. AIAA J 38:917–926 mization. AIAA J 46:2611–2622

View publication stats