This is an electronic reprint of the original article.

This reprint may differ from the original in pagination and typographic detail.

Author(s): Miettinen, Kaisa; Hakanen, Jussi; Podkopaev, Dmitry

Title: Interactive Nonlinear Multiobjective Optimization Methods

Year: 2016

Version: Final Draft

Please cite the original version:

Miettinen, K., Hakanen, J., & Podkopaev, D. (2016). Interactive Nonlinear
Multiobjective Optimization Methods. In S. Greco, M. Ehrgott, & J. R. Figueira (Eds.),
Multiple Criteria Decision Analysis : State of the Art Surveys (pp. 931-980).
International Series in Operations Research and Management Science, 233. Springer
Science+Business Media. doi:10.1007/978-1-4939-3094-4_22

All material supplied via JYX is protected by copyright and other intellectual property rights, and
duplication or sale of all or part of any of the repository collections is not permitted, except that
material may be duplicated by you for your research use or educational purposes in electronic or
print form. You must obtain permission for any other use. Electronic or print copies may not be
offered, whether for sale or otherwise to anyone who is not an authorised user.
Chapter 1

INTERACTIVE NONLINEAR
MULTIOBJECTIVE OPTIMIZATION METHODS

Kaisa Miettinen, Jussi Hakanen

University of Jyväskylä
Department of Mathematical Information Technology
P.O. Box 35 (Agora), FI-40014 University of Jyväskylä, Finland
[email protected], http://www.mit.jyu.fi/miettine/engl.html
[email protected]

Dmitry Podkopaev
University of Jyväskylä
Department of Biological and Environmental Science
P.O. Box 35 (YA), FI-40014 University of Jyväskylä, Finland
[email protected]

Abstract An overview of interactive methods for solving nonlinear multiobjective

optimization problems is given. In interactive methods, the decision
maker progressively provides preference information so that the most
satisfactory Pareto optimal solution can be found for her or his. The
basic features of several methods are introduced and some theoretical
results are provided. In addition, references to modifications and ap-
plications as well as to other methods are indicated. As the role of
the decision maker is very important in interactive methods, methods
presented are classified according to the type of preference information
that the decision maker is assumed to provide.

Keywords: Multiple criteria decision making, Multiple objectives, Nonlinear opti-

mization, Interactive methods, Pareto optimality

1
2

1. Introduction
Nonlinear multiobjective optimization means multiple criteria decision
making involving nonlinear functions of (continuous) decision variables.
In these problems, the best possible compromise, that is, Pareto optimal
solution, is to be found from an (infinite) number of alternatives repre-
sented by decision variables restricted by constraint functions. Thus,
enumerating the solutions is impossible.
Solving multiobjective optimization problems usually requires the par-
ticipation of a human decision maker who is supposed to have insight
into the problem and who can express preference relations between alter-
native solutions or objective functions or some other type of preference
information. Multiobjective optimization methods can be divided into
four classes according to the role of the decision maker in the solution
process [79, 136]. If the decision maker is not involved, we use meth-
ods where no articulation of preference information is used, in other
words, no-preference methods. If the decision maker expresses prefer-
ence information after the solution process, we speak about a posteriori
methods whereas a priori methods require articulation of preference in-
formation before the solution process. The most extensive class is inter-
active methods, where the decision maker specifies preference informa-
tion progressively during the solution process. Here we concentrate on
this last-mentioned class and introduce several examples of interactive
methods.
In the literature, interactive methods have proven useful for various
reasons. They have been found efficient from both computational and
cognitive points of view. Because the decision maker directs the solution
process with one’s preferences, only those Pareto optimal solutions that
are interesting to her or him need to be calculated. This means savings
in computational cost when compared to a situation where a big set of
Pareto optimal solutions should be calculated. On the other hand, the
amount of new information generated per iteration is limited and, in this
way, the decision maker does not need to compare too many solutions
at a time. An important advantage of interactive methods is learning.
Once the decision maker has provided preferences, (s)he can see from
the Pareto optimal solutions generated, how attainable or feasible the
preferences were. In this way, the decision maker gains insight about the
problem. (S)he learns about the interdependencies between the objective
functions and also about one’s own preferences. The decision maker can
also change her or his mind after the learning, if so desired.
Many real-world phenomena behave in a nonlinear way. Besides, lin-
ear problems can always be solved using methods created for nonlinear
Interactive Nonlinear Methods 3

problems but not vice versa. For these reasons, we here devote ourselves
to nonlinear problems. We assume that all the information involved is
deterministic and that we have a single decision maker.
In this presentation, we concentrate on general-purpose interactive
methods and, thus, methods tailored for some particular problem type
are not included. In recent years, interactive approaches have been de-
veloped in the field of evolutionary multiobjective optimization (see, for
example, [15]), but we do not consider them here. The literature survey
of years since 2000 has been limited to journal articles (in English). We
describe in more detail methods with published applications.

2. Concepts
Let us begin by introducing several concepts and definitions. We study
multiobjective optimization problems of the form

minimize {f1 (x), f2 (x), . . . , fk (x)}

(1.1)
subject to x∈S

involving k (≥ 2) objective functions or objectives fi : S → R that

we want to minimize simultaneously. The decision (variable) vectors x
belong to the (nonempty) feasible region S ⊆ Rn . The feasible region is
formed by constraint functions but we do not fix them here.
We denote the image of the feasible region by Z ⊂ Rk and call it
a feasible objective region. Objective (function) values form objective
vectors z = f (x) = (f1 (x), f2 (x), . . . , fk (x))T . Note that if fi is to be
maximized, it is equivalent to minimize −fi .
We call a multiobjective optimization problem convex if all the objec-
tive functions and the feasible region are convex. On the other hand, the
problem is nondifferentiable if at least one of the objective or the con-
straint functions is nondifferentiable. (Here nondifferentiability means
that the function is not necessarily continuously differentiable but that
it is locally Lipschitz continuous.)
We assume that the objective functions are at least partly conflicting
and possibly incommensurable. This means that it is not possible to
find a single solution that would optimize all the objectives simultane-
ously. As the definition of optimality we employ Pareto optimality. An
objective vector is Pareto optimal (or noninferior or efficient or nondom-
inated) if none of its components can be improved without deterioration
to at least one of the other components. More formally, we have the
following definition.
4

Definition 1 A decision vector x∗ ∈ S is (globally) Pareto optimal if

there does not exist another decision vector x ∈ S such that fi (x) ≤
fi (x∗ ) for all i = 1, . . . , k and fj (x) < fj (x∗ ) for at least one index j.
An objective vector z∗ ∈ Z is Pareto optimal if there does not exist
another vector z ∈ Z such that zi ≤ zi∗ for all i = 1, . . . , k and zj < zj∗ for
at least one index j; or equivalently, z∗ is Pareto optimal if the decision
vector corresponding to it is Pareto optimal.

Local Pareto optimality is defined in a small neighborhood of x∗ ∈

S. Naturally, any globally Pareto optimal solution is locally Pareto
optimal. The converse is valid, for example, for convex multiobjective
optimization problems; see [22, 136], among others.
For the sake of brevity, we usually speak about Pareto optimality in
the sequel. In practice, however, we only have locally Pareto optimal
solutions computationally available, unless some additional requirement,
such as convexity, is fulfilled or unless we have global solvers available.
A Pareto optimal set consists of (an infinite number of) Pareto op-
timal solutions. In interactive methods, we usually move around the
Pareto optimal set and forget the other solutions. However, one should
remember that this limitation may have weaknesses. Namely, the real
Pareto optimal set may remain unknown. This may be the case if an
objective function is only an approximation of an unknown function or
if not all the objective functions involved are explicitly expressed.
Moving from one Pareto optimal solution to another necessitates trad-
ing off. To be more specific, a trade-off reflects the ratio of change in
the values of the objective functions concerning the increment of one
objective function that occurs when the value of some other objective
function decreases (see, for example, [24, 136]).
For any two solutions equally preferable to the decision maker there
is a trade-off involving a certain increment in the value of one objective
function that the decision maker is willing to tolerate in exchange for
a certain amount of decrement in some other objective function while
the preferences of the two solutions remain the same. This is called the
marginal rate of substitution (see, for example, [136] for further details
and properties).
Usually, one of the objective functions is selected as a reference func-
tion when trade-offs and marginal rates of substitution are treated. The
pairwise trade-offs and the marginal rates of substitution are generated
with respect to it.
Sometimes Pareto optimal sets are not enough but we need wider
or smaller sets: weakly and properly Pareto optimal sets, respectively.
An objective vector is weakly Pareto optimal if there does not exist
Interactive Nonlinear Methods 5

any other objective vector for which all the components are smaller.
Weakly Pareto optimal solutions are sometimes computationally easier
to generate than Pareto optimal solutions. Thus, they have relevance
from a technical point of view. On the other hand, a vector is properly
Pareto optimal if unbounded trade-offs are not allowed. For a collection
of different definitions of proper Pareto optimality, see, for example,
[136].
Multiobjective optimization problems are usually solved by scalar-
ization which means that the problem is converted into one or a fam-
ily of single (scalar) objective optimization problems. This produces a
new scalarized problem with a real-valued objective function, possibly
depending on some parameters. The resulting new problem must be
solved with a single objective optimization method which is appropriate
to the characteristics of the problem in question (taking into account,
for example, differentiability and convexity). When scalarization is done
properly, it can be guaranteed that the solution obtained is Pareto op-
timal to the original multiobjective optimization problems. For further
details see, for example, [136, 206].
Interactive methods differ from each other by the way the problem is
transformed into a single objective optimization problem, by the form
in which information is provided by the decision maker and by the form
in which information is given to the decision maker at each iteration of
the solution process.
One way of inquiring the decision maker’s opinions is to ask for satis-
factory or desirable objective function values. They are called aspiration
levels and denoted by z̄i , i = 1, . . . , k. They form a vector z̄ ∈ Rk to be
called a reference point.
The ranges of the objective functions in the set of Pareto optimal
solutions give valuable information to the decision maker about the pos-
sibilities and restrictions of the problem (assuming the objective func-
tions are bounded over S). The components of the ideal objective vector
z⋆ ∈ Rk are the individual optima of the objective functions. This vector
represents the lower bounds of the Pareto optimal set. (In nonconvex
problems, we need a global solver for minimizing the k functions.) Note
that we sometimes need a vector that its strictly better than the ideal
objective vector. This vector is called a utopian objective vector and
denoted by z⋆⋆ .
The upper bounds of the Pareto optimal set, that is, the components
of a nadir objective vector znad , are much more difficult to obtain. Actu-
ally, there is no constructive method for calculating the nadir objective
vector for nonlinear problems. However, a rough estimate can be ob-
tained by keeping in mind the solutions where each objective function
6

attains its lowest value and calculating the values of the other objectives.
The highest value obtained for each objective can be selected as the es-
timated component of znad . This approach was originally proposed in
[10] and later named as a pay-off table method. Some approaches for
estimating the nadir objective vector for nonlinear multiobjective opti-
mization are summarized in [136]. Examples of latest approaches include
[37, 38].
It is sometimes assumed that the decision maker makes decisions on
the basis of an underlying value function U : Rk → R representing
her or his preferences among the objective vectors [92]. Even though
value functions are seldom explicitly known, they have been important
in the development of multiobjective optimization methods and as a
theoretical background. Thus, the value function is sometimes presumed
to be known implicitly.
The value function is usually assumed to be strongly decreasing. In
other words, the preferences of the decision maker are assumed to in-
crease if the value of one objective function decreases while all the other
objective values remain unchanged. In brief, we can say that less is
preferred to more. In that case, the maximal solution of U is assured
to be Pareto optimal. Note that regardless of the existence of a value
function, in what follows, we shall assume that lower objective function
values are preferred to higher, that is, less is preferred to more by the
decision maker.
An alternative to the idea of maximizing some value function is sat-
isficing decision making [206]. In this approach, the decision maker
tries to achieve certain aspirations. If the aspirations are achieved, the
solution is called a satisficing solution.

3. Introduction to Interactive Methods

A large variety of methods has been developed for solving multiobjec-
tive optimization problems. We can say that none of them is generally
superior to all the others. As mentioned earlier, we apply here the clas-
sification of the methods into four classes according to the participation
of the decision maker in the solution process. This classification was
originally suggested in [79] and it was followed later, for example, in
[136].
While we discuss interactive methods, we divide them into ad hoc
and non ad hoc methods (based on value functions) as suggested in
[224]. Even if one knew the decision maker’s value function, one would
not exactly know how to respond to the questions posed by an ad hoc
method. On the other hand, in non ad hoc methods, the responses
Interactive Nonlinear Methods 7

can be determined or at least confidently simulated based on a value

function.
Before describing the methods, we mention several references for fur-
ther information. This presentation is mainly based on [136]. Con-
cepts and methods for multiobjective optimization are also treated in
[17, 24, 43, 44, 79, 129, 196, 206, 219, 223, 228, 242, 246, 271].
Interactive multiobjective optimization methods, in particular, are
collected in [153, 179, 207, 243, 255]. Furthermore, methods with ap-
plications to large-scale systems and industry are presented in [65, 216,
232].
We shall not discuss non-interactive methods here. However, we men-
tion some of such methods by name and give references for further infor-
mation. Examples of no-preference methods are the method of the global
criterion [270, 273] and the multiobjective proximal bundle method [145].
From among a posteriori methods we mention the weighting method
[56, 272], the ε-constraint method [64] and the hybrid method [32, 254]
as well as the method of weighted metrics [273] and the achievement
scalarizing function approach [257, 258, 259, 261]. Multiobjective evo-
lutionary algorithms are also a posteriori in nature, see, for example,
[15] and references therein. A priori methods include the value function
method [92], the lexicographic ordering [52] and the goal programming
[25, 26, 81, 194, 195].
In what follows, we concentrate on interactive methods. In interactive
methods, a solution pattern is formed and repeated several times. After
every iteration, some information is given to the decision maker and
(s)he is asked to answer some questions or to provide some other type
of information. In this way, only a part of the Pareto optimal solutions
has to be generated and evaluated, and the decision maker can specify
and correct her or his preferences and selections during the solution
process when (s)he gets to know the problem better. Thus, the decision
maker does not need to have any global preference structure. Further
information about the topics treated here can be found in [136, 153].
An interactive method typically contains the following main steps (as
outlined, for example, in [141]): (1) initialize (for example, calculate ideal
and nadir objective vectors and show them to the decision maker), (2)
generate a Pareto optimal starting point (some neutral compromise solu-
tion or solution given by the decision maker) and show it to the decision
maker, (3) ask for preference information from the decision maker (for
example, aspiration levels or number of new solutions to be generated,
depending on the method in question), (4) generate new Pareto optimal
solution(s) according to the preferences and show it/them and possi-
bly some other information about the problem to the decision maker,
8

(5) if several solutions were generated, ask the decision maker to select
the best solution so far, and (6) stop, if the decision maker wants to.
Otherwise, go to step (3).
Three main stopping criteria can be identified in interactive methods.
In the best situation, the decision maker finds a desirable solution and
wants to stop. Alternatively, the decision maker gets tired and stops or
some algorithmic stopping rule is fulfilled. In the last-mentioned case,
one must check that the decision maker agrees to stop.
As a matter of fact, as stated in [153], solving a multiobjective op-
timization problem with an interactive method can be regarded as a
constructive process where, while learning, the decision maker builds a
conviction of what is possible (that is, what kind of solutions are avail-
able and attainable) and confronting this knowledge with her or his
preferences that also evolve. Based on this understanding, in interactive
methods we should pay attention to psychological convergence, rather
than to mathematical convergence (like, for example, optimizing some
value function).
Sometimes, two different phases can be identified in interactive solu-
tion processes: learning phase and decision phase [153]. In the learning
phase, the decision maker learns about the problem and gains under-
standing of what kind of solutions are attainable whereas the most pre-
ferred solution is found in the decision phase in the region identified in
the first phase. Naturally, the two phases can also be used iteratively.
In what follows, we present several interactive methods. The idea is
to describe a collection of methods based on different approaches. In
addition, plenty of references are included. Note that although all the
calculations take place in the decision variable space, we mostly speak
about the corresponding objective vectors and refer to both as solutions
since the space is apparent from the context.
When presenting the methods we apply the classification given in
[125, 201] according to the type of preference information that the meth-
ods utilize. This is an important aspect because a reliable and an un-
derstandable way of extracting preference information from the decision
maker is essential for the success of applying interactive methods. The
decision maker must feel being in control and must understand the ques-
tions posed. Otherwise, the answers cannot be relied on in the solution
process. It is also important to pay attention to the cognitive load set on
the decision maker, as discussed in [112]. Applying the method should
not set too much cognitive load on the decision maker.
In the first class, the decision maker specifies aspiration levels (in
other words, a reference point) representing desirable objective function
values. In the second class, the decision maker provides a classification
Interactive Nonlinear Methods 9

indicating which of the objective function values should be improved,

maintained at the current value or allowed to impair. One should note
that providing aspiration levels and a classification are closely related
as justified in [148]. From classification information one can derive a
reference point but not vice versa. The third class is devoted to meth-
ods where the decision maker compares different solutions and chooses
a solution among several ones. The fourth class involves marginal rates
of substitution referring to the amount of decrement in the value of one
objective function that compensates to the decision maker an infinitesi-
mal increment in the value of another objective function while the values
of other objective functions remain unaltered. In addition to the four
classes given in [125, 201], we consider a fifth class devoted to naviga-
tion based methods where the decision maker moves around in the set
of Pareto optimal solutions in real time and controls the direction of
movement in different ways.

4. Methods Using Aspiration Levels

What is common to the methods in this section is a reference point
consisting of desirable aspiration levels. With a reference point, the de-
cision maker can conveniently express one’s desires without any cognitive
mapping as (s)he gives objective function values and obtains objective
function values generated by the method. Some of the methods in this
section utilize other types of preference information as well but the ref-
erence point is an integral element of each method.

4.1 Reference Point Method

The reference point method [256, 257, 259] is based on vectors formed of
reasonable or desirable aspiration levels. These reference points are used
to derive scalarizing functions having minimal values at weakly, properly
or Pareto optimal solutions.
No specific assumptions are set in this method. The idea is to direct
the search by changing the reference point z̄h in the spirit of satisficing
decision making rather than optimizing any value function. It is impor-
tant that reference points are intuitive and easy for the decision maker
to specify and their consistency is not an essential requirement.
Note that specifying a reference point can be considered as a way of
classifying the objective functions. If the aspiration level is lower than
the current objective value, that objective function is currently unaccept-
able, and if the aspiration level is equal to or higher than the current
objective value, that function is acceptable. The difference here is that
the reference point can be infeasible in every component. Naturally,
10

trading off is unavoidable in moving from one Pareto optimal solution

to another and it is impossible to get a solution where all objective val-
ues are better than in the previous Pareto optimal solution but different
solutions can be obtained with different approaches.
Scalarizing functions used in the reference point method are so-called
achievement (scalarizing) functions and the method relies on their prop-
erties. We can define so-called order-representing and order-approximat-
ing achievement functions.
An example of a scalarized problem with an order-representing achieve-
ment function is
minimize max [ wi (fi (x) − z̄ih ) ]
i=1,...,k (1.2)
subject to x ∈ S,
where w is some fixed weighting vector with positive components. An
example of a scalarized problem with an order-approximating achieve-
ment function is
∑
k
minimize max [ wi (fi (x) − z̄ih ) ] + ρ wi (fi (x) − z̄ih )
i=1,...,k (1.3)
i=1
subject to x ∈ S,
where w is as above and ρ > 0 sets bounds for trade-offs.

Theorem 1 If the achievement function is order-representing, then its

solution is weakly Pareto optimal. If the function is order-approximating,
then its solution is Pareto optimal and the solution is properly Pareto
optimal if the function is also strongly increasing. Any (weakly) Pareto
optimal solution can be found if the achievement function is order-rep-
resenting. Finally, any properly Pareto optimal solution can be found if
the function is order-approximating.

The reference point method is very simple. Before the solution pro-
cess starts, some information is given to the decision maker about the
problem. If possible, the ideal objective vector and the (approximated)
nadir objective vector are presented. Another possibility is to minimize
and maximize the objective functions individually in the feasible region
(if it is bounded). Naturally, the maximized objective function values
do not typically represent components of the nadir objective vector but
they can give some information to the decision maker in any case.
The basic steps of the reference point algorithm are the following:
1 Select the achievement function. Present information about the
problem to the decision maker. Set h = 1.
Interactive Nonlinear Methods 11

2 Ask the decision maker to specify a reference point z̄h ∈ Rk .

3 Minimize the achievement function and obtain a (weakly, properly

or) Pareto optimal solution zh . Present it to the decision maker.

4 Calculate a number of k other (weakly, properly or) Pareto optimal

solutions with perturbed reference points z̄(i) = z̄h + dh ei , where
dh = ∥z̄h − zh ∥ and ei is the ith unit vector for i = 1, . . . , k.

5 Present the alternatives to the decision maker. If (s)he finds one of

the k + 1 solutions satisfactory, stop. Otherwise, ask the decision
maker to specify a new reference point z̄h+1 . Set h = h + 1 and go
to step 3.
The idea in perturbing the reference point in step 4 is that the deci-
sion maker gets a better conception of the possible solutions around the
current solution. If the reference point is far from the Pareto optimal
set, the decision maker gets a wider description of the Pareto optimal
set and if the reference point is near the Pareto optimal set, then a finer
description of the Pareto optimal set is given.
In this method, the decision maker has to specify aspiration levels
and compare objective vectors. The decision maker is free to change
her or his mind during the solution process and can direct the solution
process without being forced to understand complicated concepts and
their meaning. On the other hand, the method does not necessarily help
the decision maker to find more satisfactory solutions.
The reference point method is an ad hoc method because a reference
point cannot directly be defined based on a value function. On the
other hand, alternatives are easy to compare whenever a value function
is known.
Let us mention that a software family called DIDAS (Dynamic Inter-
active Decision Analysis and Support) has been developed on the basis
of the reference point ideas. It is described, for example, in [263].
Applications and modifications of the reference point method are pro-
vided in [12, 62, 155, 182, 183, 211, 213, 215, 225, 241, 244, 245, 260, 262].

4.2 GUESS Method

The GUESS method is also called a naı̈ve method [19]. The method is
related to the reference point method.
It is assumed that a global ideal objective vector z⋆ and a global nadir
objective vector znad are available. The structure of the method is very
simple: the decision maker specifies a reference point (or a guess) z̄h and
a Pareto optimal solution is generated which is somehow closest to the
12

reference point. Then the decision maker specifies a new reference point
and so on.
The general idea is to maximize the minimum weighted deviation from
the nadir objective vector. The scalarized problem to be solved is
[ ]
zinad − fi (x)
maximize min
i=1,...,k zinad − z̄ih (1.4)
subject to x ∈ S.
Notice that the aspiration levels have to be strictly lower than the com-
ponents of the nadir objective vector.

Theorem 2 The solution of (1.4) is weakly Pareto optimal and any

Pareto optimal solution can be found.

The GUESS algorithm has five basic steps.

1 Calculate z⋆ and znad and present them to the decision maker. Set
h = 1.
2 Let the decision maker specify upper or lower bounds to the objec-
tive functions if (s)he so desires. Update the problem, if necessary.
3 Ask the decision maker to specify a reference point z̄h between z⋆
and znad .
4 Solve (1.4) and present the solution to the decision maker.
5 If the decision maker is satisfied, stop. Otherwise, set h = h + 1
and go to step 2.
In step 2, upper or lower bounds mean adding constraints to problem
(1.4), but the ideal or the nadir objective vectors are not affected. The
only stopping rule is the satisfaction of the decision maker. No guidance
is given to the decision maker in setting new aspiration levels. This is
typical of many reference point based methods.
The GUESS method is simple to use and no consistency of the pref-
erence information provided is assumed. The only information required
from the decision maker is a reference point and possible upper and lower
bounds, which are optional. Note that inappropriate lower bounds may
lead into solutions that are not weakly Pareto optimal. Unfortunately,
the GUESS method relies heavily on the availability of the nadir objec-
tive vector, which is usually only an estimation.
The GUESS method is an ad hoc method. The existence of a value
function would not help in specifying reference points or bounds for the
objective functions. The method has been compared to several other
Interactive Nonlinear Methods 13

interactive methods in [18, 21, 34] and it has performed surprisingly

well. The reasons may be its simplicity and flexibility. One can say
that decision makers seem to prefer solution methods where they can
feel that they are in control.

4.3 Light Beam Search

The light beam search [82, 83] employs tools of multiattribute decision
analysis (see, for example, [246]) together with reference point ideas. The
basic setting is identical to the reference point method. The scalarized
problem to be solved is

∑
k
minimize max [ wi (fi (x) − z̄ih ) ] + ρ (fi (x) − z̄ih )
i=1,...,k (1.5)
i=1
subject to x ∈ S,

where w is a weighting vector, z̄h is the current reference point and

ρ > 0.

Theorem 3 The solution of (1.5) is properly Pareto optimal and any

properly Pareto optimal solution can be found.

The reference point is here assumed to be infeasible, that is, unattain-

able. It is also assumed that the objective and the constraint func-
tions are continuously differentiable and that the objective functions are
bounded over S. Furthermore, none of the objective functions is allowed
to be more important than all the others together.
In the light beam search, the decision maker directs the search by
specifying reference points. In addition, other solutions in the neigh-
bourhood of the current solution are displayed. Thus, the idea is iden-
tical to that of the reference point method. The main difference is in
the way the alternatives are generated. The motivation is to avoid com-
paring too similar alternatives or alternatives that are indifferent to the
decision maker. To achieve this goal, concepts of ELECTRE methods
(developed for handling with discrete problems in multiattribute decision
analysis) are utilized (see, for example, [200]).
It is not always possible for the decision maker to distinguish be-
tween different alternatives. This means that there is an interval where
indifference prevails. For this reason, the decision maker is asked to
provide indifference thresholds for each objective function. The line be-
tween indifference and preference does not have to be sharp, either. The
hesitation between indifference and preference can be expressed by pref-
erence thresholds. Finally, a veto threshold prevents a good performance
14

in some objectives from compensating for poor values on some other

objectives.
In the light beam search, outranking relations are established between
alternatives. An objective vector z1 is said to outrank z2 if z1 is at least
as good as z2 . The idea is to generate k new alternative objective vec-
tors such that they outrank the current solution. In particular, incom-
parable or indifferent alternatives are not shown to the decision maker.
The alternatives to be shown are called characteristic neighbours. The
neighbours are determined by projecting the gradient of one objective
function at a time onto the linear approximation of those constraints
that are active in the current solution.
We can now outline the light beam algorithm.
1 If the decision maker can specify the best and the worst values for
each objective function, denote them by z⋆ and znad , respectively.
Alternatively, calculate z⋆ and znad . Set h = 1 and z̄h = z⋆ .
Initialize the set of saved solutions as B = ∅. Ask the decision
maker to specify an indifference threshold for each objective. If
desired, (s)he can also specify preference and veto thresholds.

2 Calculate current Pareto optimal solution zh by solving (1.5).

3 Present zh to the decision maker. Calculate k Pareto optimal

characteristic neighbours of zh and present them as well to the
decision maker. If the decision maker wants to see alternatives
between any two of the k + 1 alternatives displayed, set their dif-
ference as a search direction, take different steps in this direction
and project them onto the Pareto optimal set before showing them
to the decision maker. If the decision maker wants to save zh , set
B = B ∪ {zh }.

4 If the decision maker wants to revise the thresholds, save them,

set zh = zh+1 , h = h + 1 and go then to step 3. If the decision
maker wants to give another reference point, denote it by z̄h+1 , set
h = h+1 and go to step 2. If the decision maker wants to select one
of the alternatives or one solution in B as a current solution, set it
as zh+1 , set h = h + 1 and go to step 3. If one of the alternatives
is satisfactory, stop.
The option of saving desirable solutions in the set B increases the
flexibility of the method. A similar option could be added to many
other methods as well.
The name of the method comes from the idea of projecting a focused
beam of light from the reference point onto the Pareto optimal set. The
Interactive Nonlinear Methods 15

lighted part of the Pareto optimal set changes if the location of the
spotlight, that is, the reference point or the point of interest in the
Pareto optimal set are changed.
In the light beam search, the decision maker specifies reference points,
compares alternatives and affects the set of alternatives in different ways.
Specifying different thresholds may be demanding for the decision maker.
Note, however, that the thresholds are not constant but can be altered at
any time. The developers of the method point out that it may be compu-
tationally rather demanding to find the exact characteristic neighbours
in a general case. It is, however, noteworthy that the neighbours can be
generated in parallel.
The light beam search is an ad hoc method because a value function
could not directly determine new reference points. It could, however,
be used in comparing alternatives. Remember that the thresholds are
important here and they must come from the decision maker.
A modification of the method is described in [260].

4.4 Other Methods Using Aspiration Levels

Many interactive methods of the class of methods using aspiration levels
originate from the goal programming approach because the interpreta-
tion of a goal and a reference point are closely related. Examples of
such methods include [130, 159, 181, 214, 233, 251]. Methods adopting
a fuzzy approach to setting aspiration levels have been proposed in [75,
77, 156, 157, 205]. Some other aspiration level based interactive methods
can be found in [14, 35, 61, 70, 96, 121, 180, 231, 234, 250, 252, 253].

5. Methods Using Classification

With a classification, the decision maker can express what kind of changes
should be made to the current Pareto optimal solution to get a more de-
sirable solution. Classification reminds the decision maker of the fact
that it is not possible to improve all objective values of a Pareto optimal
objective vector but impairment in some objective(s) must be allowed.
The methods presented in this section utilize different numbers of classes.
Some of the methods involve preference information other than classifi-
cation but classification is the core element in all of them.

5.1 Step Method

The step method (STEM) [10] is one of the first interactive methods
developed for multiobjective optimization problems. Here we describe
an extension for nonlinear problems according to [45] and [206], pp. 268–
269.
16

STEM is based on the classification of the objective functions at the

current iteration at zh = f (xh ). It is assumed that the decision maker
can indicate both functions that have acceptable values and those whose
values are too high, that is, functions that are unacceptable. Then
the decision maker is supposed to give up a little in the value(s) of
some acceptable objective function(s) fi (denoted by i ∈ I > ) in order to
improve the values of some unacceptable objective functions fi (denoted
by i ∈ I < ) (here I > ∪ I < = {1, . . . , k}). To be more specific, the decision
maker is asked to specify upper bounds εhi > fi (xh ) for the functions in
I >.
The only requirement in the method is that the objective functions
are bounded over S because distances are measured to the (global) ideal
objective vector. The first scalarized problem to be solved is
[ ]
ei
minimize max ∑k (fi (x) − zi⋆ )
i=1,...,k
j=1 ej
(1.6)
subject to x ∈ S,
1 zi −zi zinad −zi⋆
nad ⋆
where ei = zi⋆ zinad as suggested in [45], or ei = as
max [ |zinad |,|zi⋆ | ]
suggested in [243].
Theorem 4 The solution of (1.6) is weakly Pareto optimal. The prob-
lem has at least one Pareto optimal solution.
After the decision maker has classified the objective functions, the
feasible region is restricted according to the information of the decision
maker. The weights of the relaxed objective functions are set equal to
zero, that is ei = 0 for i ∈ I > . Then a new distance minimization
problem
[ ]
ei
minimize max ∑k (fi (x) − zi⋆ )
j=1 ej
i=1,...,k
subject to fi (x) ≤ εhi for all i ∈ I > , (1.7)
fi (x) ≤ fi (xh ) for all i ∈ I < ,
x∈S
is solved.
The basic phases of the STEM algorithm are the following:
1 Calculate z⋆ and znad and the weighting coefficients ei for i =
1, . . . , k. Set h = 1. Solve (1.6). Denote the solution by zh ∈ Z.
2 Ask the decision maker to classify the objective functions at zh
into I > and I < . If the latter class is empty, stop. Otherwise, ask
the decision maker to specify relaxed upper bounds εhi for i ∈ I > .
Interactive Nonlinear Methods 17

3 Solve (1.7) and denote the solution by zh+1 ∈ Z. Set h = h + 1

and go to step 2.
The solution process continues until the decision maker does not want
to change any component of the current objective vector. If the decision
maker is not satisfied with any of the components, then the procedure
must also be stopped.
In STEM, the decision maker is moving from one weakly Pareto opti-
mal solution to another. The idea of classification is quite simple for her
or him. However, it may be difficult to estimate appropriate amounts of
increment that would allow the desired amount of improvement in those
functions whose values should be decreased.
STEM is an ad hoc method because the existence of a value function
would not help in the classification process.
Applications and modifications of STEM are given in [7, 24, 36, 79,
85].

5.2 Satisficing Trade-Off Method

The satisficing trade-off method (STOM) [171, 175] utilizes classification
and reference points. As its name suggests, STOM is based on satisficing
decision making. The decision maker is asked to classify the objective
functions at the current solution zh = f (xh ) into three classes: the unac-
ceptable objective functions whose values should be improved (I < ), the
acceptable objective functions whose values may increase (I > ) and the
acceptable objective functions whose values are acceptable as they are
(denoted by I = ) (such that I < ∪ I > ∪ I = = {1, . . . , k}).
The decision maker only has to specify aspiration levels for the func-
tions in I < . The aspiration levels (that is, upper bounds) for the func-
tions in I > can be derived using so-called automatic trade-off. In addi-
tion, the aspiration levels for the functions in I = are set equal to fi (xh ).
All the three kinds of aspiration levels form a reference point z̄h .
Different scalarizing functions can be used in STOM. One alternative
is to solve the scalarized problem
[ ]
fi (x) − zi⋆⋆
minimize max
i=1,...,k z̄ih − zi⋆⋆ (1.8)
subject to x ∈ S,
where the reference point must be strictly worse in each component than
the utopian objective vector.
Theorem 5 The solution of (1.8) is weakly Pareto optimal and any
Pareto optimal solution can be found.
18

If weakly Pareto optimal solutions are to be avoided, the scalarized

problem to be solved is
[ ]
fi (x) − zi⋆⋆ ∑
k
fi (x)
minimize max +ρ
i=1,...,k z̄ih − zi⋆⋆ z̄ h
i=1 i
− zi⋆⋆ (1.9)
subject to x ∈ S,

where ρ > 0 is a sufficiently small scalar.

Theorem 6 The solution of (1.9) is properly Pareto optimal and any

properly Pareto optimal solution can be found.

Here the utopian objective vector must be known globally. However,

if some objective function fj is not bounded from below on S, then some
small scalar value can be used as zj⋆⋆ .
Assuming all the functions involved are differentiable the scalarizing
functions can be written in a differentiable form by introducing a scalar
variable α to be optimized and setting it as an upper bound for each
function in the max-term. Under certain assumptions, trade-off rate
information can be obtained from the Karush-Kuhn-Tucker multipliers
connected to the solution of this formulation. In automatic trade-off,
upper bounds for the functions in I > are derived with the help of this
trade-off information.
Let us now describe the STOM algorithm.
1 Select the scalarizing function. Calculate z⋆⋆ . Set h = 1.
2 Ask the decision maker to specify a reference point z̄h ∈ Rk such
that z̄ih > zi⋆⋆ for every i = 1, . . . , k.
3 Minimize the scalarizing function used. Denote the solution by zh .
Present it to the decision maker.
4 Ask the decision maker to classify the objective functions. If
I < = ∅, stop. Otherwise, ask the decision maker to specify new
aspiration levels z̄ih+1 for I ∈ I < . Set z̄ih+1 = zih for i ∈ I = .

5 Use automatic trade-off to obtain new levels (upper bounds) z̄ih+1

for the functions in I > . Set h = h + 1 and go to step 3.
The decision maker can modify the levels calculated based on trade-
off rate information if they are not agreeable. On the other hand, the
decision maker can specify those upper bounds herself or himself, if so
desired. If trade-off rate information is not available, for example, in a
case when the functions are nondifferentiable, STOM is almost the same
Interactive Nonlinear Methods 19

as the GUESS method. The only difference is the scalarizing function

used.
There is no need to repeat comments mentioned in connection with
STEM and the GUESS method. In all of them, the role of the decision
maker is easy to understand. STOM requires even less input from the
decision maker if automatic trade-off is used.
As said before, in practice, classifying the objective functions into
three classes and specifying the amounts of increment and decrement
for their values is a subset of specifying a new reference point. A new
reference point is implicitly formed.
STOM is an ad hoc method like all the other classification based
methods. However, one must remember that the aim of the method is
particularly in satisficing rather than optimizing some value function.
Modifications and applications of STOM are described in [95, 154,
165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 185, 249].

5.3 Reference Direction Method

In the classification based reference direction (RD) method [177, 178], a
current objective vector zh is presented to the decision maker and (s)he
is asked to specify a reference point z̄h consisting of desired levels for the
objective functions. However, as the idea is to move around the weakly
Pareto optimal set, some objective functions must be allowed to increase
in order to attain lower values for some other objectives.
As mentioned earlier, specifying a reference point is equivalent to an
implicit classification indicating those objective functions whose values
should be decreased till they reach some acceptable aspiration level,
those whose values are satisfactory at the moment, and those whose
values are allowed to increase to some upper bound. We denote again
these three classes by I < , I = and I > , respectively. Furthermore, we
denote the components of the reference point corresponding to I > by εhi
(at iteration h) because they represent upper bounds.
Here, steps are taken in the reference direction z̄h −zh and the decision
maker specifies a priori the number of steps to be taken, that is, the
number of solutions to be generated. The idea is to move step by step
as long as the decision maker wants to. In this way, extra computation
is avoided when only those alternatives are calculated that the decision
maker wants to see.
20

Alternatives are generated along the reference direction by solving the

scalarized problem
[ ]
fi (x) − zih
minimize max
i∈I < zih − z̄ih
subject to fi (x) ≤ εhi + α(zih − εhi ) for all i ∈ I > , (1.10)
fi (x) ≤ zih for all i ∈ I = ,
x ∈ S,

where 0 ≤ α < 1 is the step-size in the reference direction, z̄ih < zih for
i ∈ I < and εhi > zih for i ∈ I > .

Theorem 7 The solution of (1.10) is weakly Pareto optimal for every

0 ≤ α < 1 and any Pareto optimal solution can be found.

The steps of the RD algorithm are the following:

1 Find a starting solution z1 and show it to the decision maker. Set
h = 1.

2 If the decision maker does not want to decrease any component of

zh , stop. Otherwise, ask the decision maker to specify z̄h , where
some of the components are lower and some higher or equal when
compared to those of zh . If there are no higher values, set P = r =
1 and go to step 3. Otherwise, ask the decision maker to specify
the maximum number of alternatives P (s)he wants to see. Set
r = 1.

3 Set α = 1 − r/P . Solve (1.10) and get zh (r). Set r = r + 1.

4 Show zh (r) to the decision maker. If (s)he is satisfied, stop. If

r ≤ P and the decision maker wants to see another solution, go to
step 3. Otherwise, if r > P or the decision maker wants to change
the reference point, set zh+1 = zh (r), h = h + 1 and go to step 2.
The RD method does not require artificial or complicated informa-
tion from the decision maker; only reference points and the number of
intermediate solutions are used. Some decision makers may appreciate
the fact that they are not asked to compare several alternatives but only
to decide whether another alternative is to be generated or not.
The decision maker must a priori determine the number of steps to
be taken, and then intermediate solutions are calculated one by one
as long as the decision maker wants to. This has both positive and
negative sides. On one hand, it is computationally efficient since it may
Interactive Nonlinear Methods 21

be unnecessary to calculate all the intermediate solutions. On the other

hand, the number of steps to be taken cannot be changed.
The RD method is an ad hoc method because a value function would
not help in specifying reference points or the numbers of steps to be
taken. It could not even help in selecting the most preferred alternative.
Here one must decide for one solution at a time whether to calculate
new alternative solutions or not. If the new alternative happens to be
less preferred than its predecessor, one cannot return to the previous
solution.
Applications and modifications of the RD method are described in
[60, 146].

5.4 NIMBUS Method

The NIMBUS method was originally presented in [136, 145, 146] but
here we describe the so-called synchronous version introduced in [149].
Originally, NIMBUS was particularly directed for nondifferentiable prob-
lems but nowadays it is a general interactive multiobjective optimization
method for nonlinear problems.
NIMBUS offers flexible ways of performing interactive consideration
of the problem and determining the preferences of the decision maker
during the solution process. Classification is used as the means of inter-
action between the decision maker and the algorithm. In addition, the
decision maker can ask for intermediate Pareto optimal solutions to be
generated between any two Pareto optimal solutions.
In the classification, the decision maker can easily indicate what kind
of improvements are desirable and what kind of impairments are tol-
erable. Opposed to the classification based methods introduced so far,
NIMBUS has five classes available. The decision maker examines at ev-
ery iteration h the current objective vector zh and divides the objective
functions into up to five classes according to how the current solution
should be changed to get a more desirable solution. The classes are
functions fi whose values
should be decreased (i ∈ I < ),
should be decreased till an aspiration level z̄ih < zih (i ∈ I ≤ ),
are satisfactory at the moment (i ∈ I = ),
are allowed to increase till an upper bound εhi > zih (i ∈ I > ), and
are allowed to change freely (i ∈ I ⋄ ),
where I < ∪ I ≤ ̸= ∅ and I > ∪ I ⋄ ̸= ∅.
22

In addition to the classification, the decision maker is asked to specify

the aspiration levels and the upper bounds if the second and the fourth
class, respectively, are used. The difference between the classes I < and
I ≤ is that the functions in I < are to be minimized as far as possible but
the functions in I ≤ only as far as the aspiration level.
As mentioned, NIMBUS has more classes than STEM, STOM or the
RD method. This means that the decision maker has more freedom and
flexibility in specifying the desired changes in the objective values. Note
that not all of the classes have to be used. The availability of the class
I ⋄ means that some functions can be left unclassified for a while to be
able to follow how their values change while the others are classified.
After the classification information has been obtained, a scalarized
problem is solved and the Pareto optimal solution obtained reflects the
desires of the decision maker as well as possible. In this way, the decision
maker can learn about the attainability of her or his preferences. In the
synchronous version of NIMBUS [149], the idea is to provide to the de-
cision maker up to four slightly different Pareto optimal solutions based
on the same preference information. The decision maker can decide how
many solutions (s)he wants to see and compare. In this way, the decision
maker can learn more about what kind of solutions are available in the
area of the Pareto optimal set that (s)he is interested in.
After the classification, up to four scalarized problems are solved. The
one that follows the classification information closest is
[ ]
fi (x) − zi⋆ fj (x) − ẑj ∑k
fi (x)
minimize max , + ρ
i∈I <
≤
zi − zi zj − zj
nad ⋆⋆ nad ⋆⋆ z
i=1 i
nad − zi⋆⋆
j∈I
subject to fi (x) ≤ fi (xh ) for all i ∈ I < ∪ I ≤ ∪ I = ,
fi (x) ≤ εi for all i ∈ I ≥ ,
x ∈ S,
(1.11)
where a so-called augmentation coefficient ρ > 0 is a relatively small
scalar and zi⋆ for i ∈ I < are components of the ideal objective vector.
The weighting coefficients 1/(zjnad − zj⋆⋆ ) involving components of the
nadir and the utopian objective vectors, respectively, have proven to
facilitate capturing the preferences of the decision maker well. They
also increase computational efficiency [150].
The other three problems are based on a reference point. As men-
tioned in Section 3, one can derive a reference point from classification
information. If the decision maker has provided aspiration levels and
upper bounds, they are directly used as components of the reference
point. Similarly it is straightforward to use the current objective func-
tion value of the class I = . In the class I < , the component of the ideal
Interactive Nonlinear Methods 23

objective vector is used in the reference point and in the class I ⋄ , the
component of the nadir objective vector is used. In this way, we can
get a k-dimensional reference point and can solve reference point based
scalarized problems. In the synchronous NIMBUS, the problems (1.4)
of GUESS, (1.3) of the reference point method and (1.8) of the STOM
method are used.

Theorem 8 The solution of (1.11) is Pareto optimal.

The decision maker can also ask for intermediate solutions between
any two Pareto optimal solutions xh and x̂h to be generated. This
means that we calculate a search direction dh = x̂h − xh and provide
more solutions by taking steps of different sizes in this direction. In other
words, we generate P − 1 new vectors f (xh + tj dh ), j = 2, . . . , P − 1,
where tj = Pj−1
−1 . Their Pareto optimal counterparts (by setting each of
the new vectors at a time as a reference point for (1.3)) are presented to
the decision maker, who then selects the most satisfying solution among
the alternatives.
The NIMBUS algorithm is given below. The solution process stops
if the decision maker does not want to improve any objective function
value or is not willing to impair any objective function value.
We denote the set of saved solutions by A. At the beginning, we
set A = ∅. The starting point of the solution process can come from
the decision maker or it can be some neutral compromise [261] between
the objectives. The nadir and and utopian objective vectors must be
calculated or estimated before starting the solution process.
The main steps of the synchronous NIMBUS algorithm are the fol-
lowing.

1 Generate a Pareto optimal starting point.

2 Ask the decision maker to classify the objective functions at the

current Pareto optimal solution and to specify the aspiration levels
and upper bounds if they are needed.

3 Ask the decision maker to select the maximum number of different

solutions to be generated (between one and four) and solve as many
problems (listed above).

4 Present the different new solutions obtained to the decision maker.

5 If the decision maker wants to save one or more of the new solutions
to A, include it/them to A.
24

6 If the decision maker does not want to see intermediate solutions

between any two solutions, go to step 8. Otherwise, ask the deci-
sion maker to select the two solutions from among the new solu-
tions or the solutions in A. Ask the number of the intermediate
solutions from the decision maker.

7 Generate the desired number of intermediate solutions and project

them to the Pareto optimal set. Go to step 4.

8 Ask the decision maker to choose the most preferred one among
the new and/or the intermediate solutions or the solutions in A.
Denote it as the current Pareto optimal solution. If the decision
maker wants to continue, go to step 2. Otherwise, stop.

In NIMBUS, the decision maker is free to explore the Pareto optimal

set, to learn and also to change her or his mind if necessary. Further-
more, the selection of the most preferred alternative from a given set is
possible, but not necessary. The decision maker can also extract unde-
sirable solutions from further consideration. Unlike some other classifi-
cation based methods, NIMBUS does not depend entirely on how well
the decision maker manages in the classification. It is important that
the classification is not irreversible. If the solution obtained is not sat-
isfactory, the decision maker can go back to the previous solution or
explore intermediate solutions. The method aims at being flexible and
the decision maker can select to what extent (s)he exploits the versa-
tile possibilities available. The method does not introduce too massive
calculations, either.
Being a classification based method, NIMBUS is ad hoc in nature. A
value function could only be used to compare different alternatives.
An implementation of NIMBUS is available on the Internet. This
WWW-NIMBUS system is at the disposal of every academic Internet
user at http://nimbus.mit.jyu.fi/. Positive sides of a WWW implemen-
tation are that the latest version of the system is always available and the
user saves the trouble of installing the software. The operating system
used or compilers available set no restrictions because all that is needed
is a WWW browser. Furthermore, WWW provides a graphical user
interface with possibilities for visualizing the classification phase, alter-
native solutions etc. The system contains both a nondifferentiable local
solver and a global solver (genetic algorithm). For details, see [147, 149].
The first version of WWW-NIMBUS was implemented in 1995. Then,
it was a pioneering interactive optimization system on the Internet.
There is also an implementation of NIMBUS in the Windows/Linux
operating systems called IND-NIMBUS [1, 137]. It can be connected to
Interactive Nonlinear Methods 25

different simulation and modelling tools like Matlab and GAMS. Sev-
eral local and global single objective optimization methods and their
hybrids are available. It is also possible to utilize, for example, the opti-
mization methods of GAMS. IND-NIMBUS has different tools for sup-
porting graphical comparison of selected solutions and it also contains
implementations of the Pareto Navigator method and the NAUTILUS
method (see Subsections 8.2 and 6.2, respectively).
Applications and modifications of the NIMBUS method can be found
in [47, 66, 67, 68, 71, 72, 114, 115, 138, 142, 145, 146, 148, 151, 152, 202,
209, 218].

5.5 Other Methods using Classification

Interactive physical programming is an interactive method developed for
trade-off analysis and decision making in multidisciplinary optimization
[236]. It is based on a physical programming approach to produce Pareto
optimal solutions [133]. A second order approximation of the Pareto op-
timal set at the current Pareto optimal solution is produced and the
decision maker is able to generate solutions in the approximation obey-
ing her or his classification. However, this necessitates differentiability
assumptions. A modification can be found in [76].
Some other classification based methods can be found in [8, 90, 135].

6. Methods where Solutions are Compared

In this section we present some methods where the decision maker is
assumed to compare Pareto optimal solutions and select one of them.
Thus, the decision maker is not assumed to provide much information
but the cognitive load related to the comparison naturally depends on
the number of alternatives to be considered.

6.1 Chebyshev Method

The Chebyshev method was originally called the Tchebycheff method.
It was proposed in [219], pp. 419–450 and [222], refined in [220] and it
is also known by the name interactive weighted Tchebycheff procedure.
The idea in this weighting vector set reduction method is to develop a
sequence of progressively smaller subsets of the Pareto optimal set until
a most preferred solution is located.
This method does not have too many assumptions. All that is as-
sumed is that the objective functions are bounded (from below) over
S. To start with, a (global) utopian objective vector z⋆⋆ is established.
Then the distance from the utopian objective vector to the feasible ob-
26

jective region is minimized by solving the scalarized problem

∑
k
lex minimize max [ wih (fi (x) − zi⋆⋆ ) ], (fi (x) − zi⋆⋆ )
i=1,...,k (1.12)
i=1
subject to x ∈ S.
The notation above means that if the min-max problem does not have
a unique solution, the sum term is minimized subject to the obtained
solutions.
Theorem 9 The solution of (1.12) is Pareto optimal and any Pareto
optimal solution can be found.
In the Chebyshev method, different Pareto optimal solutions are gen-
erated by altering the weighting vector wh . At each ∑ iteration h, the
weighting vector set W h = {wh ∈ Rk | lih < wih < uhi , k h
i=1 wi = 1} is
reduced to W h+1 , where W h+1 ⊂ W h . At the first iteration, a sample
of the whole Pareto optimal set is generated by solving (1.12) with well
dispersed weighting vectors from W = W 1 (with li1 = 0 and u1i = 1).
The space W h is reduced by tightening the upper and the lower bounds
for the weights.
Let zh be the objective vector that the decision maker chooses from
the sample at the iteration h and let wh be the corresponding weighting
vector in the problem. Now a concentrated group of weighting vectors
centred around wh is formed. In this way, a sample of Pareto optimal
solutions centred about zh is obtained.
Before the solution process starts, the decision maker must set the
number of alternative solutions P to be compared at each iteration and
the number of iterations to be taken itn. We can now present the main
features of the Chebyshev algorithm.
1 Set the set size P and a tentative number of iterations itn. Set
li1 = 0 and u1i = 1 for all i = 1, . . . , k. Construct z⋆⋆ . Set h = 1.
2 Form the weighting vector set W h and generate 2P dispersed
weighting vectors wh ∈ W h .
3 Solve (1.12) for each of the 2P weighting vectors.
4 Present the P most different of the resulting objective vectors to
the decision maker and let her or him choose the most preferred
among them.
5 If h = itn, stop.
6 Reduce W h to get for W h+1 , set h = h + 1 and go to step 2.
Interactive Nonlinear Methods 27

Problem (1.12) is solved more that P times so that solutions very close
to each other do not have to be presented to the decision maker. On the
other hand, the predetermined number of iterations is not necessarily
conclusive. The decision maker can stop iterating when (s)he obtains a
satisfactory solution or continue the solution process longer if necessary.
In this method, the decision maker is only asked to compare Pareto
optimal objective vectors. The number of these alternatives and the
number of objective functions affect the easiness of the comparison. The
personal capabilities of the decision maker are also important. Note
that some consistency is required from the decision maker because the
discarded parts of the weighting vector space cannot be restored.
It must be mentioned that a great deal of calculation is needed in the
method. That is why it may not be applicable for large and complex
problems. However, parallel computing can be utilized when generating
the alternatives.
The Chebyshev method is a non ad hoc method. It is easy to compare
the alternative solutions with the help of a value function.
Applications and modifications of the Chebyshev method are given in
[2, 87, 93, 126, 185, 192, 208, 221, 230, 265].

6.2 NAUTILUS Method

The NAUTILUS method, introduced in [140], has a different philosophy
from many other interactive methods. It is based on the assumptions
that past experiences affect the hopes of decision makers and that people
do not react symmetrically to gains and losses. This is derived from the
prospect theory of [86]. Typically, interactive multiobjective optimiza-
tion methods move around the set of Pareto optimal solutions according
to the preference of the decision maker and (s)he must trade-off, that
is, give up in some objective functions in order to enable improvement
in some others to get from one Pareto optimal solution to another. But
according to the prospect theory, decision makers may have difficulties
in allowing impairment, the decision maker may get anchored in the
vicinity of the starting point and the solution process may even be pre-
maturely terminated.
The NAUTILUS method is different from most interactive methods
because it does not generate Pareto optimal solutions at every iteration.
Instead, the solution process starts from the nadir objective vector rep-
resenting bad values for all objective functions. In this way, the decision
maker can attain improvement in each objective function without any
trading-off and can simply indicate how much each of the objectives
should be improved. It has also been observed that the decision maker
28

may be more satisfied with a given solution if the previous one was very
undesirable and this lays the foundation of the NAUTILUS method.
The method utilizes the scalarized problem (1.3) of the reference point
method but unlike other methods utilizing this problem where weights
are kept unaltered during the whole solution process while their purpose
is mainly to normalize different ranges of objectives, in NAUTILUS the
weights have a different role as proposed in [124]. In NAUTILUS, the
weights are varied to get different Pareto optimal solutions and some
preference information is included in the weights. As mentioned earlier,
the optimal solution of problem (1.3) is assured to be Pareto optimal for
any reference point (see, for example, [136]).
As said, the NAUTILUS method starts from the nadir objective vec-
tor and at every iteration the decision maker gets a solution where all
objective function values improve from the previous iteration. Thus,
only the solution of the last iteration is Pareto optimal. To get started,
the decision maker is asked to give the number of iterations (s)he plans
to carry out, denoted by itn. This is an initial estimate and can be
changed at any time.
As before, we denote by zh the objective vector corresponding to
the iteration h. We set z0 = znad . Therefore, z0 (except in trivial
problems) is not Pareto optimal. Furthermore, we denote by ith the
number of iterations left (including iteration h). Thus, it1 = itn. At
each iteration, the range of reachable values that each objective function
can have without impairment in any other objective function (in this
and further iterations) will shrink. Lower and upper bounds on these
reachable values will be calculated when possible. For iteration h, we
denote by zh,lo = (z1h,lo , . . . , zkh,lo )T and zh,up = (z1h,up , . . . , zkh,up )T these
lower and upper bounds, respectively. Initially, z1,lo = z⋆ and z1,up =
znad . This information can be regarded as an actualization of the pay-
off table (see, for example, [136]) indicating new ideal and nadir values
at each iteration, thus informing the decision maker of what values are
achievable for each objective function.
For iteration h − 1, the objective vector zh−1 = (z1h−1 , . . . , zkh−1 )T is
shown to the decision maker, who has two possibilities to provide her or
his preference information:
1 Ranking the objective functions according to the relative impor-
tance of improving current objective function values. Here the
decision maker is not asked to give any global preference ranking
of the objectives, but the local importance of improving each of
the current objective function values. (S)he is asked to assign ob-
jective functions to classes in an increasing order of importance
for improving the corresponding objective value zih−1 . With this
Interactive Nonlinear Methods 29

information the k objective functions can be allocated into index

sets Jr which represent the importance levels r = 1, . . . , s, where
1 ≤ s ≤ k. If r < t, then improving objective function values in
Jr is less important than improving objective function values in
Jt . Each objective function can belong to only one index set, but
several objectives can be in the same index set Jr . We then set
1
wih = for all i ∈ Jr , r = 1, . . . , s. (1.13)
r(zinad− zi⋆⋆ )

2 Answering the question: Assuming you have one hundred points

available, how would you distribute them among the current objec-
tive values so that the more points you allocate, the more improve-
ment on the corresponding current objective value is desired? If
the decision maker gives phi points to the objective function fi , we
set ∆qih = phi /100 and
1
wih = for all i = 1, . . . , k. (1.14)
∆qih (zinad − zi⋆⋆ )

We set z̄h = zh−1 , and wi = wih (i = 1, . . . , k), as defined in (1.13) or

(1.14), depending on the way the decision maker specifies the preference
information and solve the scalarized problem (1.3). Let us denote by xh
the Pareto optimal decision vector obtained and set f h = f (xh ). Then,
at the next iteration we take a step from the current solution towards
f h and show to the decision maker
ith − 1 h−1 1
zh = h
z + h f h. (1.15)
it it

As mentioned, if h is the last iteration, then ith = 1 and zh = f h is the

most preferred Pareto optimal objective vector and xh is the correspond-
ing solution in the decision space. But if h is not the last iteration, then
zh can even be an infeasible vector in the objective space. Nevertheless,
it has the following properties:

Theorem 10 At any iteration h, components of zh are all better than

the corresponding components of zh−1 .

It is important to point out that although zh is not a Pareto optimal

objective vector of problem (1.1) (if h is not the last iteration), and it
may even be infeasible for this problem, it is assured to either be in the
30

feasible objective set Z for problem (1.1) or there is some Pareto optimal
objective vector where each objective function has a better value. On
the other hand, each objective vector zh produced has better objective
function values than the corresponding values in all previous iterations.
In addition, at each iteration, a part of the Pareto optimal set is elim-
inated from consideration in the sense that it is not reachable unless a
step backwards is taken.
Vectors zh,lo providing bounds for the objective values that can be
attained at the next iteration can be calculated by solving k problems of
the ε-constraint method so that each objective function is optimized in
turn and the upper bounds for the other objective functions are taken
from the corresponding components of zh−1 .
Thus, the attainable values of zh are bound in the following way:
zih ∈ [zih,lo , zih−1 ] (i = 1, . . . , k).
By denoting zh,up = zh−1 , we have
zih ∈ [zih,lo , zih,up ] (i = 1, . . . , k). (1.16)

Depending on the computational cost of solving the k problems of

the ε-constraint method, it must be evaluated whether these bounds are
worth to be calculated at each iteration. If this is regarded to be too
time-consuming, calculating the bounds can be skipped.
In addition, a measure of the closeness of the current vector to the
Pareto optimal set can be shown to the decision maker. This allows the
decision maker to determine whether the approach rhythm to the Pareto
optimal set is appropriate or whether it is too fast or too slow. The
decision maker can affect this by adjusting the number of iterations still
to be taken. Given the information available, the decision maker may
take a step backwards if (s)he does not like the new solution generated
or the bounds and/or change the number of remaining iterations. In the
latter case, we assign a new value to ith . In the former case, the decision
maker can either:
continue with old preference information. A new solution is ob-
tained by considering a smaller step-size starting from the previous
solution (for example, a half of the former step-size), or
provide new preference information. Then a new iteration step is
taken, starting from zh−1 .
To get started, the ideal and the nadir objective vectors must be
calculated or estimated. Then, an overview of the NAUTILUS algorithm
can be summarized as follows.
Interactive Nonlinear Methods 31

1 Ask the decision maker to give the number of iterations, itn. Set
h = 1, z0 = f 1,up = znad , f 1,lo = z⋆ and it1 = itn.
2 Ask the decision maker to provide preference information in either
of the two ways and calculate weights wih (i = 1, . . . , k).
3 Set the reference point and the weights and solve problem (1.3) to
get xh and the corresponding f h .
4 Calculate zh according to (1.15).
5 Given zh , find f h+1,lo by solving k ε-constraint problems. Fur-
thermore, set f h+1,up = zh . Calculate the distance to the Pareto
optimal set.
6 Show the current objective values zih (i = 1, . . . , k), together with
the additional information [fih+1,lo , fih+1,up ] (i = 1, . . . , k) and the
distance to the decision maker.
7 Set a new value for ith if the decision maker wants to change the
number of remaining iterations.
8 Ask the decision maker whether (s)he wants to take a step back-
wards. If so, go to step 10. Otherwise, continue.
9 If ith = 1, stop with the last solution xh and f h . Otherwise, set
ith+1 = ith − 1 and h = h + 1. If the decision maker wants to
give new preference information, go to step 1. Alternatively, the
decision maker can take a new step in the same direction (using
the preference information of the previous iteration). Then, set
f h = f h−1 , and go to step 4.
10 Ask the decision maker whether (s)he would like to provide new
preference information starting from zh−1 . If so, go to step 2.
Alternatively, the decision maker can take a shorter step with the
same preference information given in step 2. Then, set zh = 12 zh +
1 h−1
2z and go to step 5.

The algorithm looks more complicated than it actually is. There are
many steps to provide to the decision maker different options of how to
continue the solution process. A good user interface plays an important
role in making the options available intuitive.
The NAUTILUS method has been located in this class of methods
because the decision maker must compare at each iteration the solution
generated to the solution of the previous iteration and decide whether to
proceed or to go backwards. Naturally, preference information indicating
32

how important it is to improve each of the objective functions from their

current levels is also needed.
NAUTILUS is ad hoc in nature because all preference information
needed cannot be obtained from a value function.
A modification of the NAUTILUS method is presented in [210].

6.3 Other Methods where Solutions are

Compared
Methods where the decision maker is asked to compare different solutions
have been developed rather recently. Such methods targeted at nonlinear
problems can be found in [27, 88, 91, 102, 118, 127, 128].

7. Methods Using Marginal Rates of

Substitution
In this section, we present methods that utilize preference information in
the form of marginal rates of substitution or desirability of trade-off in-
formation provided. These methods are included here because they have
played a role in the history of developing interactive methods. They aim
at some sort of mathematical convergence in optimizing an estimated
value function rather than psychological convergence. It is important
that the decision maker understands well the concepts used in these
methods to be able to apply them.

7.1 Interactive Surrogate Worth Trade-Off

Method
The interactive surrogate worth trade-off (ISWT) method is introduced
in [23] and [24], pp. 371–379. The ISWT method utilizes the scalarized
ε-constraint problem where one of the objective functions is minimized
subject to upper bounds on all the other objectives:
minimize fℓ (x)
subject to fj (x) ≤ εj for all j = 1, . . . , k, j ̸= ℓ, (1.17)
x ∈ S,
where ℓ ∈ {1, . . . , k} and εj are upper bounds for the other objectives.
Theorem 11 The solution of (1.17) is weakly Pareto optimal. The de-
cision vector x∗ ∈ S is Pareto optimal if and only if it solves (1.17) for
every ℓ = 1, . . . , k, where εj = fj (x∗ ) for j = 1, . . . , k, j ̸= ℓ. A unique
solution is Pareto optimal for any upper bounds.
The idea of the ISWT method is to maximize an approximation of an
underlying value function. A search direction is determined based on the
Interactive Nonlinear Methods 33

opinions of the decision maker concerning trade-off rates at the current

solution. The step-size to be taken in the search direction is determined
by solving several ε-constraint problems and asking the decision maker
to select the most satisfactory solution.
It is assumed that the underlying value function exists and is implicitly
known to the decision maker. In addition, it must be continuously dif-
ferentiable and strongly decreasing. Furthermore, the objective and the
constraint functions must be twice continuously differentiable and the
feasible region has to be compact. Finally, it is assumed that the Pareto
optimality of the solutions of the ε-constraint problem is guaranteed and
that trade-off rate information is available in the Karush-Kuhn-Tucker
(KKT) multipliers related to the ε-constraint problem.
Changes in objective function values between a reference function fℓ
and all the other objectives are compared. For each i = 1, . . . , k, i ̸= ℓ,
the decision maker must answer the following question: Let an objective
vector zh be given. If the value of fℓ is decreased by λhi units, then the
value of fi is increased by one unit (or vice versa) and the other objective
values remain unaltered. How desirable do you find this trade-off?
The response of the decision maker indicating the degree of preference
is called a surrogate worth value. According to [23, 24] the response must
be an integer between 10 and −10 whereas it is suggested in [238] to use
integers from 2 to −2.
The gradient of the underlying value function is then estimated with
the help of the surrogate worth values. This gives a search direction
with a steepest ascent for the value function. Several different steps
are taken in the search direction and the decision maker must select
the most satisfactory of them. In practice, the upper bounds of the
ε-constraint problem are revised based on surrogate worth values with
different step-sizes.
The main features of the ISWT algorithm can be presented with four
steps.

1 Select fℓ to be minimized and give upper bounds to the other

objective functions. Set h = 1.

2 Solve (1.17) to get a solution zh . Trade-off rate information is

obtained from the KKT multipliers.

3 Ask the decision maker for the surrogate worth values at zh .

4 If some stopping criterion is satisfied, stop. Otherwise, update the

upper bounds with the help of the answers obtained in step 3 and
solve several ε-constraint problems. Let the decision maker choose
34

the most preferred alternative zh+1 and set h = h + 1. Go to step

3.

As far as stopping criteria are concerned, one can always stop when
the decision maker wants to do so. A common stopping criterion is the
situation where all the surrogate worth values equal zero. One more
criterion is the case when the decision maker wants to proceed only in
an infeasible direction.
In the ISWT method, the decision maker is asked to specify surrogate
worth values and compare Pareto optimal alternatives. It may be diffi-
cult for the decision maker to provide consistent surrogate worth values
throughout the solution process. In addition, if there is a large number
of objective functions, the decision maker has to specify a lot of surro-
gate worth values at each iteration. On the other hand, the easiness of
the comparison of alternatives depends on the number of objectives and
on the personal abilities of the decision maker.
The ISWT method can be regarded as a non ad hoc method. The sign
of the surrogate worth values can be judged by comparing trade-off rates
with marginal rates of substitution (obtainable from the value function).
Furthermore, when comparing alternatives, it is easy to select the one
with the highest value function value.
Modification of the ISWT method are presented in [24, 28, 49, 63, 69].

7.2 Geoffrion-Dyer-Feinberg Method

In the Geoffrion-Dyer-Feinberg (GDF) method proposed in [57], the ba-
sic idea is related to that of the ISWT method. In both the methods, the
underlying (implicitly known) value function is approximated and max-
imized. In the GDF method, the approximation is based on marginal
rates of substitution.
It is assumed that an underlying value function exists, is implicitly
known to the decision maker and is strongly decreasing with respect to
the reference function fℓ . In addition, the corresponding value function
with decision variables as variables (i.e., arguments) must be continu-
ously differentiable and concave on S. Furthermore, the objective func-
tions have to be continuously differentiable and the feasible region S
must be compact and convex.
Let xh be the current solution. We can obtain a local linear approxi-
mation for the gradient of the value function with the help of marginal
rates of substitution mhi involving a reference function fℓ and the other
Interactive Nonlinear Methods 35

functions fi . Based on this information we solve the problem

(∑
k )T
maximize −mhi ∇x fi (xh ) y
(1.18)
i=1
subject to y ∈ S,
where y ∈ Rn is the variable. Let us denote the solution by yh . Then,
the search direction is dh = yh − xh .
The following problem is to find a step-size. The decision maker can
be offered objective vectors where steps of different sizes are taken in
the search direction starting from the current solution. Unfortunately,
these alternatives are not necessarily Pareto optimal.
Now we can present the GDF algorithm.
1 Ask the decision maker to select fℓ . Set h = 1.
2 Ask the decision maker to specify marginal rates of substitution
between fℓ and the other objectives at the current solution zh .
3 Solve (1.18). Set the search direction dh . If dh = 0, stop.
4 Determine with the help of the decision maker the appropriate
step-size th to be taken in dh . Denote the corresponding solution
by zh+1 = f (xh + th dh ).
5 Set h = h + 1. If the decision maker wants to continue, go to step
2. Otherwise, stop.
In the GDF method, the decision maker has to specify marginal rates
of substitution and select the most preferred solution from a set of alter-
natives. The theoretical foundation of the method is convincing but the
practical side is not as promising. At each iteration the decision maker
has to determine k − 1 marginal rates of substitution in a consistent and
correct way. On the other hand, it is obvious that in practice the task
of selection becomes more difficult for the decision maker as the number
of objective functions increases. Another drawback is that not all the
solutions presented to the decision maker are necessarily Pareto optimal.
They can naturally be projected onto the Pareto optimal set but this
necessitates extra effort.
The GDF method is a non ad hoc method. The marginal rates of
substitution and selections can be done with the help of value function
information. Note that if the underlying value function is linear, the
marginal rates of substitution are constant and only one iteration is
needed.
Applications and modifications of the GDF method are described in
[4, 40, 42, 51, 53, 73, 79, 84, 143, 144, 164, 186, 197, 203, 219, 264].
36

7.3 Other Methods Using Marginal Rates of

Substitution
Although preference information about relative importance of different
objectives in one form or another is utilized in many interactive methods,
there are very few methods where the desirable marginal rates of substi-
tute are the main preference information. Such methods are presented
in [120, 131, 268].

8. Navigation Methods
By navigation we refer to methods where new Pareto optimal solution al-
ternatives are generated in a real-time imitating fashion along directions
that are derived from the information the decision maker has specified.
In this way, the decision maker can learn about the interdependencies
among the objective functions. The decision maker can either continue
the movement along the current direction or change the direction, that
is, one’s preferences. Increased interest has been devoted to navigation
based methods in the literature in recent years. In these methods, the
user interface plays a very important role in enabling the navigation.

8.1 Reference Direction Approach

The reference direction approach [103, 108] is also known by the name
visual interactive approach. It contains ideas from, for example, the GDF
method and the reference point method. However, more information is
provided to the decision maker.
In reference point based methods, a reference point is projected onto
the Pareto optimal set by optimizing an achievement function. Here,
instead, a so-called reference direction as a whole is projected onto the
Pareto optimal set. It is a vector from the current solution zh to the
reference point z̄h . In practice, steps of different sizes are taken along
the reference direction and projected. The idea is to plot the objective
function values on a computer screen as value paths. The decision maker
can move the cursor back and forth and see the corresponding numerical
values at each solution.
Solutions along the reference direction are generated by solving the
scalarized problem

[ ]
fi (x) − z̄ih
minimize max
i∈I wi (1.19)
subject to z̄h = zh + tdh+1 ,
x ∈ S,
Interactive Nonlinear Methods 37

where I = {i | wi > 0} ⊂ {1, . . . , k} and t has different discrete nonneg-

ative values. The weighting vector can be, for example, the reference
point specified by the decision maker.
Theorem 12 The solution of (1.19) is weakly Pareto optimal.
The algorithm of the reference direction approach is as follows.
1 Find an arbitrary objective vector z1 . Set h = 1.
2 Ask the decision maker to specify a reference point z̄h ∈ Rk and
set dh+1 = z̄h − zh .
3 Find the set Z h+1 of weakly Pareto optimal solutions with different
values of t in (1.19).
4 Ask the decision maker to select the most preferred solution zh+1
in Z h+1 .
5 If zh ̸= zh+1 , set h = h + 1 and go to step 2. Otherwise, check
the optimality conditions. If the conditions are satisfied, stop.
Otherwise, set h = h + 1 and set dh+1 to be a search direction
identified by the optimality checking procedure. Go to step 3.
Checking the optimality conditions in step 5 is the most complicated
part of the algorithm. Thus far, no specific assumptions have been set
on the value function. However, we can check the optimality of zh+1
if the cone containing all the feasible directions has a finite number of
generators. We must then assume that an underlying value function
exists and is pseudoconcave on Z. In addition, S must be convex and
compact and the constraint functions must be differentiable.
The role of the decision maker is similar in the reference point method
and in the reference direction approach: specifying reference points and
selecting the most preferred alternative. But by providing similar refer-
ence point information, in the reference direction approach, the decision
maker can explore a wider part of the weakly Pareto optimal set. This
possibility brings the task of comparing the alternatives.
The performance of the method depends greatly on how well the deci-
sion maker manages to specify the reference directions that lead to more
satisfactory solutions. The consistency of the decision maker’s answers
is not important and it is not checked in the algorithm.
The reference direction approach can be characterized as an ad hoc
method like the other reference point based methods. The aim is to
support the decision maker in getting to know the problem better.
A dynamic user interface to the reference direction approach and its
adaptation to generalized goal programming is introduced in [110]. This
38

method for linear multiobjective optimization problems is called the

Pareto race.
Applications and modifications of the reference direction approach are
described in [11, 101, 103, 104, 105, 106, 107, 109].

8.2 Pareto Navigator Method

Pareto Navigator is an interactive method utilizing a polyhedral ap-
proximation of the Pareto optimal set for convex problems [48]. Pareto
Navigator consists of two phases, namely an initialization phase, where
the decision maker is not involved, and a navigation phase. In the initial-
ization phase, a relatively small set of Pareto optimal objective vectors
is assumed to be available to form a polyhedral approximation of the
Pareto optimal set in the objective space. These objective vectors can
be computed, for example, by using some a posteriori approach.
Pareto Navigator has been developed especially for the learning phase
of interactive solution processes introduced in Section 3 and for computa-
tionally expensive problems where objective function and/or constraint
function value evaluations may be time-consuming because the problem
is, for example, simulation-based. In these problems, computing Pareto
optimal solutions can take a lot of time. For this reason, besides the
original (computationally expensive) problem, an approximation is used
to enable fast computations so that the decision maker does not need to
wait for new solutions being generated based on her or his preferences.
In Pareto Navigator, the decision maker is not involved in the part
of the solution process where the set of objective vectors representing
the Pareto optimal set is generated. Once the approximation has been
created based on the objective vectors available, the original problem is
not solved (in the navigation phase). When the navigation phase starts,
the decision maker can navigate dynamically in the approximated Pareto
optimal set in real time since approximated Pareto optimal solutions can
be produced by solving linear programming problems that are compu-
tationally inexpensive.
Whenever the decision maker has found an interesting approximated
Pareto optimal solution, the corresponding solution to the original prob-
lem can be generated by solving problem (1.3) with the approximated
solution as a reference point. This can be seen as projecting the ap-
proximated solution to the Pareto optimal set of the original problem.
However, this step may take time if the original problem is computa-
tionally expensive.
Interactive Nonlinear Methods 39

As mentioned, the multiobjective optimization problem is assumed to

be convex, that is, the objective functions and the feasible region must
be convex. The algorithm of Pareto Navigator is as follows.

1 Compute first a polyhedral approximation of the Pareto optimal

set in the objective space based on a small set of Pareto optimal
objective vectors. Use the extreme values present in this set to ap-
proximate the ideal and nadir objective vectors. Ask the decision
maker to select a starting point for navigation (for example, one
of the Pareto optimal objective vectors available).

2 Show the current objective vector to the decision maker and ask
her or him whether a preferred solution has been found. If yes, go
to step 6. Otherwise, continue.

3 Ask the decision maker whether (s)he would like to proceed to some
other direction. If the decision maker does not want to change the
direction, go to step 5.

4 Ask the decision maker to specify how the current objective vector
should be improved by giving aspiration levels for the objectives.
To aid her or him, show the ideal and the nadir objective vectors.
Based on the resulting reference point z̄ and the current objective
vector zc , set a search direction.

5 Ask the decision maker to indicate a speed of movement, that is, a

step-size α > 0 to the direction specified. Generate approximated
Pareto optimal solutions in the direction specified by using a ref-
erence point based approach for each step in the direction starting
from the current objective vector zc . Once an approximated so-
lution is produced, it is instantly shown to the decision maker.
New approximated solutions are produced to the direction speci-
fied until the decision maker stops the movement. Then go to step
2.

6 Once the decision maker has found a satisfactory solution, stop.

Project the approximated Pareto optimal solution to the actual
Pareto optimal set and show the resulting solution to the decision
maker.

The search direction is based on decision maker’s preferences and

there are different ways of defining a direction where to move on the
approximation. In Pareto Navigator, the direction is specified by d =
z̄ − zc . The approximated Pareto optimal solutions are then computed
40

by solving scalarized problems of the form

minimize max wi (zi − z̄i (α))
i=1,...,k (1.20)
subject to Az ≤ b,
where z̄(α) = zc + αd is the reference point depending on the step pa-
rameter α > 0 (being varied) to the direction d and wi , i = 1, . . . , k,
are the scaling coefficients. The scaling coefficient can be set as one di-
vided by the difference of the estimated nadir and ideal objective values.
The linear constraints of problem (1.20) form a convex hull for a set of
Pareto optimal solutions used to form the polyhedral approximation and,
in practice, the reference point z̄(α) is projected to the nondominated
facets of the convex hull.
The objective function of problem (1.20) is nonlinear with respect to
z but can be linearized by adding a new real variable ξ ∈ R replacing
the max term. The resulting problem is then linear with respect to a
new variable z′ = (ξ, z)T . Due to linearity, approximated Pareto optimal
solutions can be produced and shown to the decision maker in real time
by moving the reference point along the direction d. This is done by
increasing the value of α. At any point, the decision maker is able to find
the closest actual Pareto optimal solution for any approximated Pareto
optimal solution. However, as said, this can be time consuming.
Because the decision maker must specify desirable objective function
values, this method is ad hoc by nature.
During the navigation, the approximated solutions are shown to the
decision maker by presenting the approximated values as a continuous
path (value path) for each objective function separately (bar charts can
be used as well). Pareto navigator is implemented in the IND-NIMBUS
system [1, 137] and the graphical user interface development is described
in [237].

8.3 Pareto Navigation Method

The Pareto Navigation method developed in [160] assumes the convexity
of all objective functions and a convex feasible region. Similar to the
Pareto Navigator method, the idea is to enable a fast generation of
new solutions in the navigation phase. Thus, the method starts with
formulating a surrogate problem based on a set of pre-computed Pareto
optimal decision vectors {x(1) , . . . , x(m) }. The most preferred solution
is sought among their convex hull
 
∑m ∑
m 
X = vj x(j) : vj = 1, vj ≥ 0 for all j = 1, . . . , m .
 
j=1 j=1
Interactive Nonlinear Methods 41

This allows replacing the feasible region of the original problem with the
set of convex combination coefficients v1 , . . . , vm in the definition of X .
The current state of the navigation process is represented by the cur-
rent Pareto optimal solution xh and the vector of current upper bounds
b ∈ Rk on objective function values. Using the surrogate problem with
these bounds as additional constraints, the ideal objective vector is cal-
culated and the nadir objective vector is estimated via a pay-off table.
They define ranges of objective function values for Pareto optimal solu-
tions. These ranges together with the current solution are displayed in
a radar chart also known as a spider-web chart.
By moving sliders on the radar chart with the mouse, the decision
maker can provide two types of preference information: upper bounds
on objective values and a desired value (aspiration level) of any objective
function. Changes made by the decision maker are immediately reflected
in the current state of the navigation process and shown in the radar
chart. Setting the upper bounds influences the objective function ranges
as described above. Setting the value of any objective function fi∗ to a
desired value τ yields updating the current solution with the solution of
the following problem
minimize max yi − fi (xh )
i=1,...,k,
i̸=i∗ ( )
∑
m
subject to y=f vj x(j) + s,
j=1
yi ≤ bi , i = 1, . . . , k,
yi∗ = τ,
∑
m
vj = 1,
j=1
v and s are non-negative.
By using the two above-described mechanisms of expressing prefer-
ences the decision maker explores the set of Pareto optimal solutions
of the surrogate problem until a most preferred or satisfactory solution
is found. Because the decision maker must provide upper bounds and
aspiration levels, the method is ad hoc by nature.
The method has been developed and implemented for intensity mod-
ulated radiation therapy treatment planning. Therefore, in addition to
the radar chart, some application-specific information about the cur-
rent solution (treatment plan) is displayed. Nevertheless, there are no
obstacles of adapting the method elsewhere when the multiobjective op-
timization problem is convex and the convex hull of some finite set of
pre-calculated Pareto optimal solutions may serve as a good enough ap-
proximation of the Pareto optimal set.
42

8.4 Other Navigation Methods

Other navigation based methods developed for nonlinear multiobjective
optimization problems, implemented as software tools include [122, 123].
A collection of methods and software for solving linear multiobjective
optimization problems [5, 6] can also be mentioned as they can partly
be extended to nonlinear problems.

9. Other Interactive Methods

The number of interactive methods developed for multiobjective opti-
mization is large. So far, we have given several examples of them. Let us
next mention references to some more methods based on miscellaneous
ideas: [9, 29, 30, 31, 39, 46, 50, 54, 55, 78, 89, 94, 97, 100, 116, 117, 119,
132, 158, 162, 163, 176, 187, 188, 189, 190, 198, 199, 204, 212, 216, 217,
223, 226, 227, 229, 235, 248, 266, 267, 269].

10. Comparing the Methods

None of the many multiobjective optimization methods can be claimed
to be superior to the others in every aspect. One can say that selecting
a multiobjective optimization method is a problem with multiple ob-
jectives itself. The properties of the problem and the capabilities and
the desires of the decision maker have to be charted before a solution
method can be chosen. Some methods may suit some problems and some
decision makers better than some others.
A decision tree is provided in [136] for easing the method selection.
The tree is based on theoretical facts concerning the assumptions on the
problem to be solved and the preferences of the decision maker. Further
aspects to be taken into account when evaluating and selecting methods
are collected, for example, in [13, 58, 74, 80, 136, 228, 239, 240].
In addition to theoretical properties, practical applicability, in partic-
ular, plays an important role in the selection of an appropriate method.
The difficulty is that practical applicability is hard to determine without
experience.
Some comparisons of the methods have been reported in the literature.
They have been carried out with respect to a variety of criteria and
under varied circumstances. Instead of a human decision maker, one can
sometimes employ value functions in the comparisons. Unfortunately,
replacing the decision maker with a value function does not fully reflect
the real usefulness of the methods. One of the problems is that value
functions cannot really help in testing ad hoc methods.
REFERENCES 43

Tests with human decision makers are described in [16, 18, 20, 21, 33,
34, 41, 111, 134, 184, 247] while tests with value functions are reported
in [3, 59, 161, 191]. Finally, comparisons based on intuition are provided
in [45, 98, 99, 113, 131, 135, 185, 193, 207, 243, 246].

11. Conclusions
We have outlined several interactive methods for solving nonlinear multi-
objective optimization problems and indicated references to many more.
One of the challenges in this area is spreading the word about the exist-
ing methods to those who solve real-world problems. Another challenge
is to develop methods that support the decision maker even better. User-
friendliness cannot be overestimated because interactive methods must
be able to correspond to the characteristics of the decision maker. Spe-
cific methods for different areas of application that take into account the
characteristics of the problems are also important.
An alternative to creating new methods is to use different methods in
different phases of the solution process. This hybridization means that
the positive features of various methods can be exploited to their best
advantage in appropriate phases. In this way, it may also be possible to
overcome some of the weaknesses of the methods like proposed. Ways
to enable changing the type of preference information specified, that is,
the method used during the solution process are presented in [125, 201].
The decision maker can be supported by using visual illustrations
and further development of such tools is essential. For instance, one
may visualize (parts of) the Pareto optimal set and, for example, use 3D
slices of the feasible objective region (see [122, 123], among others) and
other tools. On the other hand, one can illustrate sets of alternatives by
means of bar charts, value paths, spider-web charts and petal diagrams
etc. For more details see, for example, [136] and references therein as
well as [139] for a more detailed survey.

References
[1] IND-NIMBUS website. http://ind-nimbus.it.jyu.fi/.
[2] P. J. Agrell, B. J. Lence, and A. Stam. An interactive multicri-
teria decision model for multipurpose reservoir management: The
Shellmouth reservoir. Journal of Multi-Criteria Decision Analysis,
7:61–86, 1998.
[3] Y. Aksoy, T. W. Butler, and E. D. Minor III. Comparative stud-
ies in interactive multiple objective mathematical programming.
European Journal of Operational Research, 89:408–422, 1996.
44

[4] J. E. Al-alvani, B. F. Hobbs, and B. Malakooti. An interac-

tive integrated multiobjective optimization approach for quasicon-
cave/quasiconvex utility functions. In A. Goicoechea, L. Duck-
stein, and S. Zionts, editors, Multiple Criteria Decision Making:
Proceedings of the Ninth International Conference: Theory and
Applications in Business, Industry, and Government, pages 45–
60, New York, 1992. Springer-Verlag.
[5] C. H. Antunes, M. J. Alves, A. L. Silva, and J. N. Clı́maco. An
integrated MOLP method base package – a guided tour of TOM-
MIX. Computers & Operations Research, 19:609–625, 1992.
[6] C. H. Antunes, M. P. Melo, and J. N. Clı́maco. On the integra-
tion of an interactive MOLP procedure base and expert system
technique. European Journal of Operational Research, 61:135–144,
1992.
[7] J.-P. Aubin and B. Näslund. An exterior branching algorithm.
Working Paper 72-42, European Institute for Advanced Studies in
Management, Brussels, 1972.
[8] N. Baba, H. Takeda, and T. Miyake. Interactive multi-objective
programming technique using random optimization method. In-
ternational Journal of Systems Science, 19:151–159, 1988.
[9] J. F. Bard. A multiobjective methodology for selecting subsystem
automation options. Management Science, 32:1628–1641, 1986.
[10] R. Benayoun, J. de Montgolfier, J. Tergny, and O. Laritchev. Lin-
ear programming with multiple objective functions: Step method
(STEM). Mathematical Programming, 1:366–375, 1971.
[11] H. P. Benson and Y. Aksoy. Using efficient feasible directions
in interactive multiple objective linear programming. Operations
Research Letters, 10:203–209, 1991.
[12] E. Bischoff. Two empirical tests with approaches to multiple-
criteria decision making. In M. Grauer, M. Thompson, and A. P.
Wierzbicki, editors, Plural Rationality and Interactive Decision
Processes, pages 344–347, Berlin, Heidelberg, 1985. Springer-Ver-
lag.
[13] E. Bischoff. Multi-objective decision analysis – the right ob-
jectives? In G. Fandel, M. Grauer, A. Kurzhanski, and A. P.
Wierzbicki, editors, Large-Scale Modelling and Interactive Deci-
sion Analysis, pages 155–160. Springer-Verlag, 1986.
[14] P. Bogetoft, Å. Hallefjord, and M. Kok. On the convergence of
reference point methods in multiobjective programming. European
Journal of Operational Research, 34:56–68, 1988.
REFERENCES 45

[15] J. Branke, K. Deb, K. Miettinen, and R. Slowinski, editors. Multi-

objective Optimization: Interactive and Evolutionary Approaches.
Springer-Verlag, Berlin, Heidelberg, 2008.
[16] K. Brockhoff. Experimental test of MCDM algorithms in a modu-
lar approach. European Journal of Operational Research, 22:159–
166, 1985.
[17] J. T. Buchanan. Multiple objective mathematical programming:
A review. New Zealand Operational Research, 14:1–27, 1986.
[18] J. T. Buchanan. An experimental evaluation of interactive MCDM
methods and the decision making process. Journal of the Opera-
tional Research Society, 45:1050–1059, 1994.
[19] J. T. Buchanan. A naı̈ve approach for solving MCDM problems:
The GUESS method. Journal of the Operational Research Society,
48:202–206, 1997.
[20] J. T. Buchanan and J. L. Corner. The effects of anchoring in
interactive MCDM solution methods. Computers & Operations
Research, 24:907–918, 1997.
[21] J. T. Buchanan and H. G. Daellenbach. A comparative evalua-
tion of interactive solution methods for multiple objective decision
models. European Journal of Operational Research, 29:353–359,
1987.
[22] Y. Censor. Pareto optimality in multiobjective problems. Applied
Mathematics and Optimization, 4:41–59, 1977.
[23] V. Chankong and Y. Y. Haimes. The interactive surrogate worth
trade-off (ISWT) method for multiobjective decision-making. In
S. Zionts, editor, Multiple Criteria Problem Solving, pages 42–67,
Berlin, Heidelberg, 1978. Springer-Verlag.
[24] V. Chankong and Y. Y. Haimes. Multiobjective Decision Making
Theory and Methodology. Elsevier, New York, 1983.
[25] A. Charnes and W. W. Cooper. Management Models and Indus-
trial Applications of Linear Programming, volume 1. John Wiley
& Sons, New York, 1961.
[26] A. Charnes and W. W. Cooper. Goal programming and multiple
objective optimization; part 1. European Journal of Operational
Research, 1:39–54, 1977.
[27] J. Chen and S. Lin. An interactive neural network-based approach
for solving multiple criteria decision-making problems. Decision
Support Systems, 36:137–146, 2003.
[28] T. Chen and B.-I. Wang. An interactive method for multiobjec-
tive decision making. In A. Straszak, editor, Large Scale Systems:
46

Theory and Applications 1983, pages 277–282. Pergamon Press,

1984.
[29] J. N. Clı́maco and C. H. Antunes. Flexible method bases and
man-machine interfaces as key features in interactive MOLP ap-
proaches. In P. Korhonen, A. Lewandowski, and J. Wallenius, edi-
tors, Multiple Criteria Decision Support, pages 207–216. Springer-
Verlag, 1991.
[30] J. N. Clı́maco and C. H. Antunes. Man-machine interfacing in
MCDA. In G. H. Tzeng, H. F. Wand, U. P. Wen, and P. L.
Yu, editors, Multiple Criteria Decision Making – Proceedings of
the Tenth International Conference: Expand and Enrich the Do-
mains of Thinking and Application, pages 239–253, New York,
1994. Springer-Verlag.
[31] J. N. Clı́maco, C. H. Antunes, and M. J. Alves. From TRIMAP
to SOMMIX – building effective interactive MOLP computational
tools. In G. Fandel and T. Gal, editors, Multiple Criteria Deci-
sion Making: Proceedings of the Twelfth International Conference,
pages 285–296, Berlin, Heidelberg, 1997. Springer-Verlag.
[32] H. W. Corley. A new scalar equivalence for Pareto optimization.
IEEE Transactions on Automatic Control, 25:829–830, 1980.
[33] J. L. Corner and J. T. Buchanan. Experimental consideration of
preference in decision making under certainty. Journal of Multi-
Criteria Decision Analysis, 4:107–121, 1995.
[34] J. L. Corner and J. T. Buchanan. Capturing decision maker pref-
erence: Experimental comparison of decision analysis and MCDM
techniques. European Journal of Operational Research, 98:85–97,
1997.
[35] J. P. Costa and J. N. Clı́maco. A multiple reference point par-
allel approach in MCDM. In G. H. Tzeng, H. F. Wand, U. P.
Wen, and P. L. Yu, editors, Multiple Criteria Decision Making –
Proceedings of the Tenth International Conference: Expand and
Enrich the Domains of Thinking and Application, pages 255–263,
New York, 1994. Springer-Verlag.
[36] Y. Crama. Analysis of STEM-like solutions to multi-objective
programming problems. In S. French, R. Hartley, L. C. Thomas,
and D. J. White, editors, Multi-Objective Decision Making, pages
208–213. Academic Press, 1983.
[37] K. Deb and K. Miettinen. Nadir point estimation using evolution-
ary approaches: Better accuracy and computational speed through
focused search. In M. Ehrgott, B. Naujoks, T. J. Stewart, and
REFERENCES 47

J. Wallenius, editors, Multiple Criteria Decision Making for Sus-

tainable Energy and Transportation Systems, Proceedings, pages
339–354, Berlin, Heidelberg, 2010. Springer-Verlag.
[38] K. Deb, K. Miettinen, and Chaudhuri S. Towards an estimation
of nadir objective vector using a hybrid of evolutionary and local
search approaches. IEEE Transactions on Evolutionary Computa-
tion, 14:821–841, 2010.
[39] A. Diaz. Interactive solution to multiobjective optimization prob-
lems. International Journal for Numerical Methods in Engineering,
24:1865–1877, 1987.
[40] J. S. Dyer. Interactive goal programming. Management Science,
19:62–70, 1972.
[41] J. S. Dyer. An empirical investigation of a man-machine inter-
active approach to the solution of the multiple criteria problem.
In J. L. Cochrane and M. Zeleny, editors, Multiple Criteria De-
cision Making, pages 202–216, Columbia, South Carolina, 1973.
University of South Carolina Press.
[42] J. S. Dyer. A time-sharing computer program for the solution of
the multiple criteria problem. Management Science, 19:1379–1383,
1973.
[43] J. S. Dyer and R. K. Sarin. Multicriteria decision making. In A. G.
Holzman, editor, Mathematical Programming for Operations Re-
searchers and Computer Scientists, pages 123–148. Marcel Dekker,
1981.
[44] M. Ehrgott. Multicriteria Optimization. Springer-Verlag, Berlin,
Heidelberg, 2000.
[45] H. A. Eschenauer, A. Osyczka, and E. Schäfer. Interactive multi-
criteria optimization in design process. In H. Eschenauer, J. Koski,
and A. Osyczka, editors, Multicriteria Design Optimization Pro-
cedures and Applications, pages 71–114, Berlin, Heidelberg, 1990.
Springer-Verlag.
[46] H. A. Eschenauer, E. Schäfer, and H. Bernau. Application of in-
teractive vector optimization methods with regard to problems in
structural mechanics. In H. A. Eschenauer and G. Thierauf, ed-
itors, Discretization Methods and Structural Optimization – Pro-
cedures and Applications, pages 95–101, Berlin, Heidelberg, 1989.
Springer-Verlag.
[47] P. Eskelinen and K. Miettinen. Trade-off analysis approach for
interactive nonlinear multiobjective optimization. OR Spectrum,
34:803–816, 2012.
48

[48] P. Eskelinen, K. Miettinen, K. Klamroth, and J. Hakanen. Pareto

Navigator for interactive nonlinear multiobjective optimization.
OR Spectrum, 32:211–227, 2010.
[49] J. Feng. Dual worth trade-off method and its application for solv-
ing multiple criteria decision making problems. Journal of Systems
Engineering and Electronics, 17:554–558, 2006.
[50] P. A. V. Ferreira and J. C. Geromel. An interactive projection
method for multicriteria optimization problems. IEEE Transac-
tions on Systems, Man, and Cybernetics, 20:596–605, 1990.
[51] P. A. V. Ferreira and M. E. S. Machado. Solving multiple-objective
problems in the objective space. Journal of Optimization Theory
and Applications, 89:659–680, 1996.
[52] P. C. Fishburn. Lexicographic orders, utilities and decision rules:
A survey. Management Science, 20:1442–1471, 1974.
[53] T. L. Friesz. Multiobjective optimization in transportation: The
case of equilibrium network design. In N. N. Morse, editor, Orga-
nizations: Multiple Agents with Multiple Criteria, pages 116–127,
Berlin, Heidelberg, 1981. Springer-Verlag.
[54] L. R. Gardiner and R. E. Steuer. Unified interactive multiple ob-
jective programming. European Journal of Operational Research,
74:391–406, 1994.
[55] L. R. Gardiner and R. E. Steuer. Unified interactive multiple objec-
tive programming: An open architecture for accommodating new
procedures. Journal of the Operational Research Society, 45:1456–
1466, 1994.
[56] S. Gass and T. Saaty. The computational algorithm for the para-
metric objective function. Naval Research Logistics Quarterly,
2:39–45, 1955.
[57] A. M. Geoffrion, J. S. Dyer, and A. Feinberg. An interactive
approach for multi-criterion optimization, with an application to
the operation of an academic department. Management Science,
19:357–368, 1972.
[58] M. Gershon and L. Duckstein. An algorithm for choosing of a
multiobjective technique. In P. Hansen, editor, Essays and Sur-
veys on Multiple Criteria Decision Making, pages 53–62, Berlin,
Heidelberg, 1983. Springer-Verlag.
[59] M. Gibson, J. J. Bernardo, C. Chung, and R. Badinelli. A compar-
ison of interactive multiple-objective decision making procedures.
Computers & Operations Research, 14:97–105, 1987.
REFERENCES 49

[60] V. G. Gouljashki, L. M. Kirilov, S. C. Narula, and V. S. Vassilev. A

reference direction interactive algorithm of the multiple objective
nonlinear integer programming. In G. Fandel and T. Gal, editors,
Multiple Criteria Decision Making: Proceedings of the Twelfth In-
ternational Conference, pages 308–317, Berlin, Heidelberg, 1997.
Springer-Verlag.
[61] J. Granat and M. Makowski. Interactive specification and analysis
of aspiration-based preferences. European Journal of Operational
Research, 122:469–485, 2000.
[62] M. Grauer, A. Lewandowski, and A. Wierzbicki. DIDASS – the-
ory, implementation and experiences. In M. Grauer and A. P.
Wierzbicki, editors, Interactive Decision Analysis, pages 22–30.
Springer-Verlag, 1984.
[63] Y. Y. Haimes. The surrogate worth trade-off (SWT) method and
its extensions. In G. Fandel and T. Gal, editors, Multiple Criteria
Decision Making Theory and Application, pages 85–108, Berlin,
Heidelberg, 1980. Springer-Verlag.
[64] Y. Y. Haimes, L. S. Lasdon, and D. A. Wismer. On a bicriterion
formulation of the problems of integrated system identification and
system optimization. IEEE Transactions on Systems, Man, and
Cybernetics, 1:296–297, 1971.
[65] Y. Y. Haimes, K. Tarvainen, T. Shima, and J. Thadathil. Hierar-
chical Multiobjective Analysis of Large-Scale Systems. Hemisphere
Publishing Corporation, New York, 1990.
[66] J. Hakanen, Y. Kawajiri, K. Miettinen, and L. T. Biegler. In-
teractive multi-objective optimization for simulated moving bed
processes. Control and Cybernetics, 36:282–320, 2007.
[67] J. Hakanen, K. Miettinen, and K. Sahlstedt. Wastewater treat-
ment: New insight provided by interactive multiobjective optimiza-
tion. Decision Support Systems, 51:328–337, 2011.
[68] J. Hakanen, K. Sahlstedt, and K. Miettinen. Wastewater treatment
plant design and operation under multiple conflicting objective
functions. Environmental Modelling & Software, 46:240–249, 2013.
[69] W. A. Hall and Y. Y. Haimes. The surrogate worth trade-off
method with multiple decision-makers. In M. Zeleny, editor, Mul-
tiple Criteria Decision Making Kyoto 1975, pages 207–233, Berlin,
Heidelberg, 1976. Springer-Verlag.
[70] Å. Hallefjord and K. Jörnsten. An entropy target-point approach
to multiobjective programming. International Journal of Systems
Science, 17:639–653, 1986.
50

[71] J. P. Hämäläinen, K. Miettinen, P. Tarvainen, and J. Toivanen. In-

teractive solution approach to a multiobjective optimization prob-
lem in a paper machine headbox design. Journal of Optimization
Theory and Applications, 116:265–281, 2003.
[72] E. Heikkola, K. Miettinen, and P. Nieminen. Multiobjective opti-
mization of an ultrasonic transducer using NIMBUS. Ultrasonics,
44:368–380, 2006.
[73] T. Hemming. Some modifications of a large step gradient method
for interactive multicriterion optimization. In N. N. Morse, editor,
Organizations: Multiple Agents with Multiple Criteria, pages 128–
139, Berlin, Heidelberg, 1981. Springer-Verlag.
[74] B. F. Hobbs. What can we learn from experiments in multiobjec-
tive decision analysis? IEEE Transactions on Systems, Man, and
Cybernetics, 16:384–394, 1986.
[75] C. Hu, Y. Shen, and S. Li. An interactive satisficing method based
on alternative tolerance for fuzzy multiple objective optimization.
Applied Mathematical Modelling, 33:1886–1893, 2009.
[76] H. Huang and Z. Tian. Application of neural network to inter-
active physical programming. In J. Wang, X. Liao, and Z. Yi,
editors, Advances in Neural Networks, pages 725–730. Springer-
Verlag, Berlin, Heidelberg, 2005.
[77] H.-Z. Huang, Y.-K. Gu, and X. Du. An interactive fuzzy multi-
objective optimization method for engineering design. Engineering
Applications of Artificial Intelligence, 19:451–460, 2006.
[78] M. L. Hussein and F. S. A. El-Ghaffar. An interactive approach
for vector optimization problems. European Journal of Operational
Research, 89:185–192, 1996.
[79] C.-L. Hwang and A. S. M. Masud. Multiple Objective Decision
Making – Methods and Applications: A State-of-the-Art Survey.
Springer-Verlag, Berlin, Heidelberg, 1979.
[80] C.-L. Hwang and K. Yoon. Multiple Attribute Decision Making
Methods and Applications: A State-of-the-Art Survey. Springer-
Verlag, Berlin, Heidelberg, 1981.
[81] J. P. Ignizio. Goal Programming and Extensions. Lexington Books.
D.C. Heath and Company, 1976.
[82] A. Jaszkiewicz and R. Slowiński. The light beam search – outrank-
ing based interactive procedure for multiple-objective mathemati-
cal programming. In P. M. Pardalos, Y. Siskos, and C. Zopounidis,
editors, Advances in Multicriteria Analysis, pages 129–146, Dor-
drecht, 1995. Kluwer Academic Publishers.
REFERENCES 51

[83] A. Jaszkiewicz and R. Slowiński. The ‘light beam search’ approach

– an overview of methodology and applications. European Journal
of Operational Research, 113:300–314, 1999.
[84] P. Jedrzejowicz and L. Rosicka. Multicriterial reliability optimiza-
tion problem. Foundations of Control Engineering, 8:165–173,
1983.
[85] I.-J. Jeong and K.-J. Kim. D-STEM: A modified Step method with
desirability function concept. Computers & Operations Research,
32:3175–3190, 2005.
[86] D. Kahneman and A. Tversky. Prospect theory: An analysis of
decision under risk. Econometrica, 47:263–291, 1979.
[87] I. Kaliszewski. A modified weighted Tchebycheff metric for mul-
tiple objective programming. Computers & Operations Research,
14:315–323, 1987.
[88] I. Kaliszewski. Using trade-off information in decision-making al-
gorithms. Computers & Operations Research, 27:161–182, 2000.
[89] I. Kaliszewski and W. Michalowski. Searching for psychologically
stable solutions of multiple criteria decision problems. European
Journal of Operational Research, 118:549–562, 1999.
[90] I. Kaliszewski, W. Michalowski, and G. Kersten. A hybrid in-
teractive technique for the MCDM problems. In M. H. Karwan,
J. Spronk, and J. Wallenius, editors, Essays in Decision Making:
A Volume in Honour of Stanley Zionts, pages 48–59. Springer-
Verlag, Berlin, Heidelberg, 1997.
[91] I. Kaliszewski and S. Zionts. Generalization of the Zionts-
Wallenius algorithm. Control and Cybernetics, 33:477–500, 2004.
[92] R. L. Keeney and H. Raiffa. Decisions with Multiple Objectives:
Preferences and Value Tradeoffs. John Wiley & Sons, 1976.
[93] J. Kim and S.-K. Kim. A CHIM-based interactive Tchebycheff
procedure for multiple objective decision making. Computers &
Operations Research, 33:1557–1574, 2006.
[94] S. H. Kim and T. Gal. A new interactive algorithm for multi-
objective linear programming using maximally changeable domi-
nance cone. European Journal of Operational Research, 64:126–
137, 1993.
[95] S. Kitayama and K. Yamazaki. Compromise point incorporating
trade-off ratio in multi-objective optimization. Applied Soft Com-
puting, 12:1959–1964, 2012.
52

[96] K. Klamroth and K. Miettinen. Integrating approximation and in-

teractive decision making in multicriteria optimization. Operations
Research, 56:222–234, 2008.
[97] L. Y. Klepper. An interactive method for optimization of irradi-
ation plans in radiation therapy of malignant tumors. Biomedical
Engineering, 40:291–297, 2006.
[98] M. Kok. Scalarization and the interface with decision makers in
interactive multi objective linear programming. In P. Serafini, edi-
tor, Mathematics of Multi Objective Optimization, pages 433–438,
Wien, New York, 1985. Springer-Verlag.
[99] M. Kok. The interface with decision makers and some experimen-
tal results in interactive multiple objective programming method.
European Journal of Operational Research, 26:96–107, 1986.
[100] M. Kok and F. A. Lootsma. Pairwise-comparison methods in
multiple objective programming, with applications in a long-term
energy-planning model. European Journal of Operational Re-
search, 22:44–55, 1985.
[101] M. Köksalan and R. D. Plante. Interactive multicriteria optimiza-
tion for multiple-response product and process design. Manufac-
turing & Service Operations Management, 5:334–347, 2003.
[102] M. M. Köksalan and H. Moskowitz. Solving the multiobjective de-
cision making problem using a distance function. In G. H. Tzeng,
H. F. Wand, U. P. Wen, and P. L. Yu, editors, Multiple Criteria
Decision Making – Proceedings of the Tenth International Confer-
ence: Expand and Enrich the Domains of Thinking and Applica-
tion, pages 101–107, New York, 1994. Springer-Verlag.
[103] P. Korhonen. Reference direction approach to multiple objec-
tive linear programming: Historical overview. In M. H. Karwan,
J. Spronk, and J. Wallenius, editors, Essays in Decision Making:
A Volume in Honour of Stanley Zionts, pages 74–92. Springer-
Verlag, Berlin, Heidelberg, 1997.
[104] P. Korhonen and M. Halme. Using lexicographic parametric pro-
gramming for searching a nondominated set in multiple objective
linear programming. Journal of Multi-Criteria Decision Analysis,
5:291–300, 1996.
[105] P. Korhonen and J. Laakso. A visual interactive method for solving
the multiple-criteria problem. In M. Grauer and A. P. Wierzbicki,
editors, Interactive Decision Analysis, pages 146–153. Springer-
Verlag, 1984.
REFERENCES 53

[106] P. Korhonen and J. Laakso. On developing a visual interactive

multiple criteria method – an outline. In Y. Y. Haimes and
V. Chankong, editors, Decision Making with Multiple Objectives,
pages 272–281, Berlin, Heidelberg, 1985. Springer-Verlag.
[107] P. Korhonen and J. Laakso. Solving generalized goal programming
problems using a visual interactive approach. European Journal of
Operational Research, 26:355–363, 1986.
[108] P. Korhonen and J. Laakso. A visual interactive method for solving
the multiple criteria problem. European Journal of Operational
Research, 24:277–287, 1986.
[109] P. Korhonen and S. C. Narula. An evolutionary approach to sup-
port decision making with linear decision models. Journal of Multi-
Criteria Decision Analysis, 2:111–119, 1993.
[110] P. Korhonen and J. Wallenius. A Pareto race. Naval Research
Logistics, 35:615–623, 1988.
[111] P. Korhonen and J. Wallenius. Observations regarding choice be-
haviour in interactive multiple criteria decision-making environ-
ments: An experimental investigation. In A. Lewandowski and
I. Stanchev, editors, Methodology and Software for Interactive De-
cision Support, pages 163–170. Springer-Verlag, 1989.
[112] O. Larichev. Cognitive validity in design of decision-aiding tech-
niques. Journal of Multi-Criteria Decision Analysis, 1:127–138,
1992.
[113] O. I. Larichev, O. A. Polyakov, and A. O. Nikiforov. Multicriterion
linear programming problems. Journal of Economic Psychology,
8:389–407, 1987.
[114] T. Laukkanen, T.-M. Tveit, V. Ojalehto, K. Miettinen, and C.-J.
Fogelholm. An interactive multi-objective approach to heat ex-
changer network synthesis. Computers and Chemical Engineering,
34:943–952, 2010.
[115] T. Laukkanen, T.-M. Tveit, V. Ojalehto, K. Miettinen, and C.-
J. Fogelholm. Bilevel heat exchanger network synthesis with an
interactive multi-objective optimization method. Applied Thermal
Engineering, 48:301–316, 2012.
[116] R. Lazimy. Interactive relaxation method for a broad class of inte-
ger and continuous nonlinear multiple criteria problems. Journal
of Mathematical Analysis and Applications, 116:553–573, 1986.
[117] R. Lazimy. Solving multiple criteria problems by interactive de-
composition. Mathematical Programming, 35:334–361, 1986.
54

[118] R. Lazimy. Interactive polyhedral outer approximation (IPOA)

strategy for general multiobjective optimization problems. Annals
of Operations Research, 210:73–99, 2012.
[119] E. R. Lieberman. Soviet multi-objective mathematical program-
ming methods: An overview. Management Science, 37:1147–1165,
1991.
[120] G. V. Loganathan and H. D. Sherali. A convergent interactive
cutting-plane algorithm for multiobjective optimization. Operation
Research, 35:365–377, 1987.
[121] A. V. Lotov. Computer-based support for planning and negotia-
tion on environmental rehabilitation of water resource systems. In
D. P. Loucks, editor, Restoration of Degraded Rivers: Challenges,
Issues and Experiences, pages 417–445. Kluwer Academic Publish-
ers, 1998.
[122] A. V. Lotov, V. A. Bushenkov, and O. L. Chernykh. Multicriteria
DSS for river water-quality planning. Microcomputers in Civil
Engineering, 12:57–67, 1997.
[123] A. V. Lotov, V. A. Bushenkov, and G. K. Kamenev. Interactive
Decision Maps: Approximation and Visualization of Pareto Fron-
tier. Kluwer Academic Publishers, Norwell, 2004.
[124] M. Luque, K. Miettinen, P. Eskelinen, and F. Ruiz. Incorporating
preference information in interactive reference point methods for
multiobjective optimization. Omega, 37:450–462, 2009.
[125] M. Luque, F. Ruiz, and K. Miettinen. Global formulation for
interactive multiobjective optimization. OR Spectrum, 33:27–48,
2011.
[126] M. Luque, F. Ruiz, and R. E. Steuer. Modified interactive Cheby-
shev algorithm (MICA) for convex multiobjective programming.
European Journal of Operational Research, 204:557–564, 2010.
[127] P. Mackin, A. Roy, and J. Wallenius. An interactive weight space
reduction procedure for nonlinear multiple objective mathematical
programming. Mathematical Programming, 127:425–444, 2011.
[128] B. Malakooti and V. Raman. An interactive multi-objective ar-
tificial neural network approach for machine setup optimization.
Journal of Intelligent Manufacturing, 11:41–50, 2000.
[129] R.T. Marler and J. S. Arora. Survey of multi-objective optimiza-
tion methods for engineering. Structural and Multidisciplinary Op-
timization, 26:369–395, 2004.
REFERENCES 55

[130] A. S. M. Masud and C. L. Hwang. Interactive sequential goal pro-

gramming. Journal of the Operational Research Society, 32:391–
400, 1981.
[131] A. S. M. Masud and X. Zheng. An algorithm for multiple-objective
non-linear programming. Journal of the Operational Research So-
ciety, 40:895–906, 1989.
[132] Z. Meng, R. Shen, and M. Jiang. An objective penalty functions al-
gorithm for multiobjective optimization problem. American Jour-
nal of Operations Research, 1:229–235, 2011.
[133] A. Messac. Physical programming: Effective optimization for com-
putational design. AIAA Journal, 34:149–158, 1996.
[134] W. Michalowski. Evaluation of a multiple criteria interactive pro-
gramming approach: An experiment. INFOR: Information Sys-
tems & Operational Research, 25:165–173, 1987.
[135] W. Michalowski and T. Szapiro. A bi-reference procedure for
interactive multiple criteria programming. Operations Research,
40:247–258, 1992.
[136] K. Miettinen. Nonlinear Multiobjective Optimization. Kluwer Aca-
demic Publishers, Boston, 1999.
[137] K. Miettinen. IND-NIMBUS for demanding interactive multiob-
jective optimization. In T. Trzaskalik, editor, Multiple Criteria
Decision Making ’05, pages 137–150, Katowice, 2006. The Karol
Adamiecki University of Economics in Katowice.
[138] K. Miettinen. Using interactive multiobjective optimization in con-
tinuous casting of steel. Materials and Manufacturing Processes,
22:585–593, 2007.
[139] K. Miettinen. Survey of methods to visualize alternatives in mul-
tiple criteria decision making problems. OR Spectrum, 36:3–37,
2014.
[140] K. Miettinen, P. Eskelinen, F. Ruiz, and M. Luque. NAUTILUS
method: An interactive technique in multiobjective optimization
based on the nadir point. European Journal of Operational Re-
search, 206:426–434, 2010.
[141] K. Miettinen and J. Hakanen. Why use interactive multi-objective
optimization in chemical process design? In G. P. Rangaiah, edi-
tor, Multi-Objective Optimization Techniques and Applications in
Chemical Engineering, pages 148–183. World Scientific Publishers,
Singapore, 2009.
[142] K. Miettinen, A. V. Lotov, G. K. Kamenev, and V. E. Berezkin.
Integration of two multiobjective optimization methods for non-
56

linear problems. Optimization Methods and Software, 18:63–80,

2003.
[143] K. Miettinen and M. M. Mäkelä. An interactive method for non-
smooth multiobjective optimization with an application to optimal
control. Optimization Methods and Software, 2:31–44, 1993.
[144] K. Miettinen and M. M. Mäkelä. A nondifferentiable multiple cri-
teria optimization method applied to continuous casting process.
In A. Fasano and M. Primicerio, editors, Proceedings of the Seventh
European Conference on Mathematics in Industry, pages 255–262,
Stuttgart, 1994. B. G. Teubner.
[145] K. Miettinen and M. M. Mäkelä. Interactive bundle-based method
for nondifferentiable multiobjective optimization: NIMBUS. Opti-
mization, 34:231–246, 1995.
[146] K. Miettinen and M. M. Mäkelä. Comparative evaluation of some
interactive reference point-based methods for multi-objective opti-
misation. Journal of the Operational Research Society, 50:949–959,
1999.
[147] K. Miettinen and M. M. Mäkelä. Interactive multiobjective opti-
mization system WWW-NIMBUS on the Internet. Computers &
Operations Research, 27:709–723, 2000.
[148] K. Miettinen and M. M. Mäkelä. On scalarizing functions in mul-
tiobjective optimization. OR Spectrum, 24:193–213, 2002.
[149] K. Miettinen and M. M. Mäkelä. Synchronous approach in interac-
tive multiobjective optimization. European Journal of Operational
Research, 170:909–922, 2006.
[150] K. Miettinen, M. M. Mäkelä, and K. Kaario. Experiments with
classification-based scalarizing functions in interactive multiob-
jective optimization. European Journal of Operational Research,
175:931–947, 2006.
[151] K. Miettinen, M. M. Mäkelä, and T. Männikkö. Optimal control
of continuous casting by nondifferentiable multiobjective optimiza-
tion. Computational Optimization and Applications, 11:177–194,
1998.
[152] K. Miettinen, J. Mustajoki, and T. J. Stewart. Interactive mul-
tiobjective optimization with NIMBUS for decision making under
uncertainty. OR Spectrum, 36:39–56, 2014.
[153] K. Miettinen, F. Ruiz, and A. P. Wierzbicki. Introduction to mul-
tiobjective optimization: Interactive approaches. In J. Branke,
K. Deb, K. Miettinen, and R. Slowinski, editors, Multiobjective Op-
REFERENCES 57

timization: Interactive and Evolutionary Approaches, pages 27–57.

Springer-Verlag, Berlin, Heidelberg, 2008.
[154] K. Mitani and H. Nakayama. A multiobjective diet planning sup-
port system using the satisficing trade-off method. Journal of
Multi-Criteria Decision Analysis, 6:131–139, 1997.
[155] U. Mocci and L. Primicerio. Ring network design: An MCDM ap-
proach. In G. Fandel and T. Gal, editors, Multiple Criteria Deci-
sion Making: Proceedings of the Twelfth International Conference,
pages 491–500, Berlin, Heidelberg, 1997. Springer-Verlag.
[156] C. Mohan and H. T. Nguyen. Reference direction interactive
method for solving multiobjective fuzzy programming problems.
European Journal of Operational Research, 107:599–613, 1998.
[157] C. Mohan and H. T. Nguyen. An interactive satisficing method for
solving multiobjective mixed fuzzy-stochastic programming prob-
lems. Fuzzy Sets and Systems, 117:61–79, 2001.
[158] M. A. Moldavskiy. Singling out a set of undominated solutions in
continuous vector optimization problems. Soviet Automatic Con-
trol, 14:47–53, 1981.
[159] D. E. Monarchi, C. C. Kisiel, and L. Duckstein. Interactive mul-
tiobjective programming in water resources: A case study. Water
Resources Research, 9:837–850, 1973.
[160] M. Monz, K. H. Kufer, T. R. Bortfeld, and C. Thieke. Pareto
navigation – algorithmic foundation of interactive multi-criteria
IMRT planning. Physics in Medicine and Biology, 53:985–998,
2008.
[161] J. Mote, D. L. Olson, and M. A. Venkataramanan. A comparative
multiobjective programming study. Mathematical and Computer
Modelling, 10:719–729, 1988.
[162] A. M’silti and P. Tolla. An interactive multiobjective nonlinear
programming procedure. European Journal of Operational Re-
search, 64:115–125, 1993.
[163] H. Mukai. Algorithms for multicriterion optimization. IEEE
Transactions on Automatic Control, 25:177–186, 1980.
[164] K. Musselman and J. Talavage. A tradeoff cut approach to multiple
objective optimization. Operations Research, 28:1424–1435, 1980.
[165] H. Nakayama. On the components in interactive multiobjective
programming methods. In M. Grauer, M. Thompson, and A. P.
Wierzbicki, editors, Plural Rationality and Interactive Decision
Processes, pages 234–247, Berlin, Heidelberg, 1985. Springer-Ver-
lag.
58

[166] H. Nakayama. Sensitivity and trade-off analysis in multiobjec-

tive programming. In A. Lewandowski and I. Stanchev, editors,
Methodology and Software for Interactive Decision Support, pages
86–93. Springer-Verlag, 1989.
[167] H. Nakayama. Satisficing trade-off method for problems with
multiple linear fractional objectives and its applications. In
A. Lewandowski and V. Volkovich, editors, Multiobjective Prob-
lems of Mathematical Programming, pages 42–50. Springer-Verlag,
1991.
[168] H. Nakayama. Trade-off analysis based upon parametric optimiza-
tion. In P. Korhonen, A. Lewandowski, and J. Wallenius, editors,
Multiple Criteria Decision Support, pages 42–52. Springer-Verlag,
1991.
[169] H. Nakayama. Trade-off analysis using parametric optimization
techniques. European Journal of Operational Research, 60:87–98,
1992.
[170] H. Nakayama. Engineering applications of multi-objective pro-
gramming: Recent results. In G. H. Tzeng, H. F. Wand, U. P.
Wen, and P. L. Yu, editors, Multiple Criteria Decision Making –
Proceedings of the Tenth International Conference: Expand and
Enrich the Domains of Thinking and Application, pages 369–378,
New York, 1994. Springer-Verlag.
[171] H. Nakayama. Aspiration level approach to interactive multi-
objective programming and its applications. In P. M. Pardalos,
Y. Siskos, and C. Zopounidis, editors, Advances in Multicriteria
Analysis, pages 147–174, Dordrecht, 1995. Kluwer Academic Pub-
lishers.
[172] H. Nakayama and K. Furukawa. Satisficing trade-off method with
an application to multiobjective structural design. Large Scale
Systems, 8:47–57, 1985.
[173] H. Nakayama, K. Kaneshige, S. Takemoto, and Y. Watada. An ap-
plication of a multi-objective programming technique to construc-
tion accuracy control of cable-stayed bridges. European Journal of
Operational Research, 87:731–738, 1995.
[174] H. Nakayama, J. Nomura, K. Sawada, and R. Nakajima. An ap-
plication of satisficing trade-off method to a blending problem of
industrial materials. In G. Fandel, M. Grauer, A. Kurzhanski, and
A. P. Wierzbicki, editors, Large-Scale Modelling and Interactive
Decision Analysis, pages 303–313. Springer-Verlag, 1986.
REFERENCES 59

[175] H. Nakayama and Y. Sawaragi. Satisficing trade-off method for

multiobjective programming. In M. Grauer and A. P. Wierzbicki,
editors, Interactive Decision Analysis, pages 113–122. Springer-
Verlag, 1984.
[176] H. Nakayama, T. Tanino, and Y. Sawaragi. An interactive opti-
mization method in multicriteria decisionmaking. IEEE Transac-
tions on Systems, Man, and Cybernetics, 10:163–169, 1980.
[177] S. C. Narula, L. Kirilov, and V. Vassilev. An interactive algorithm
for solving multiple objective nonlinear programming problems. In
G. H. Tzeng, H. F. Wand, U. P. Wen, and P. L. Yu, editors, Multi-
ple Criteria Decision Making – Proceedings of the Tenth Interna-
tional Conference: Expand and Enrich the Domains of Thinking
and Application, pages 119–127, New York, 1994. Springer-Verlag.
[178] S. C. Narula, L. Kirilov, and V. Vassilev. Reference direction
approach for solving multiple objective nonlinear programming
problems. IEEE Transactions on Systems, Man, and Cybernet-
ics, 24:804–806, 1994.
[179] S. C. Narula and H. R. Weistroffer. Algorithms for multi-objective
nonlinear programming problems: An overview. In A. G. Lockett
and G. Islei, editors, Improving Decision Making in Organisations,
pages 434–443, Berlin, Heidelberg, 1989. Springer-Verlag.
[180] S. C. Narula and H. R. Weistroffer. A flexible method for nonlin-
ear multicriteria decisionmaking problems. IEEE Transactions on
Systems, Man, and Cybernetics, 19:883–887, 1989.
[181] P. Nijkamp and J. Spronk. Interactive multiple goal programming:
An evaluation and some results. In G. Fandel and T. Gal, editors,
Multiple Criteria Decision Making Theory and Applications, pages
278–293, Berlin, Heidelberg, 1980. Springer-Verlag.
[182] W. Ogryczak. Preemptive reference point method. In J. Clı́maco,
editor, Multicriteria Analysis, pages 156–167, Berlin, Heidelberg,
1997. Springer-Verlag.
[183] M. Olbrisch. The interactive reference point approach as a solu-
tion concept for econometric decision models. In M. J. Beckmann,
K.-W. Gaede, K. Ritter, and H. Schneeweiss, editors, X. Sympo-
sium on Operations Research, Part 1, Sections 1-5, pages 611–619,
Knigstein, 1986. Verlag Anton Hain Meisenheim GmbH.
[184] D. L. Olson. Review of empirical studies in multiobjective math-
ematical programming: Subject reflection of nonlinear utility and
learning. Decision Sciences, 23:1–20, 1992.
60

[185] D. L. Olson. Tchebycheff norms in multi-objective linear program-

ming. Mathematical and Computer Modelling, 17:113–124, 1993.
[186] K. R. Oppenheimer. A proxy approach to multi-attribute decision
making. Management Science, 24:675–689, 1978.
[187] A. Osyczka and S. Kundu. A new method to solve generalized
multicriteria optimization problems using the simple genetic algo-
rithm. Structural Optimization, 10:94–99, 1995.
[188] R. Ramesh, M. H. Karwan, and S. Zionts. Theory of convex cones
in multicriteria decision making. Annals of Operations Research,
16:131–147, 1988.
[189] R. Ramesh, M. H. Karwan, and S. Zionts. Interactive multicriteria
linear programming: An extension of the method of Zionts and
Wallenius. Naval Research Logistics, 36:321–335, 1989.
[190] R. Ramesh, M. H. Karwan, and S. Zionts. Preference structure
representation using convex cones in multicriteria integer program-
ming. Management Science, 35:1092–1105, 1989.
[191] G. R. Reeves and J. J. Gonzalez. A comparison of two interactive
MCDM procedures. European Journal of Operational Research,
41:203–209, 1989.
[192] G. R. Reeves and K. R. MacLeod. Robustness of the interactive
weighted Tchebycheff procedure to inaccurate preference informa-
tion. Journal of Multi-Criteria Decision Analysis, 8:128–132, 1999.
[193] P. Rietveld. Multiple Objective Decision Methods and Regional
Planning. North-Holland Publishing Company, 1980.
[194] C. Romero. Handbook of Critical Issues in Goal Programming.
Pergamon Press, 1991.
[195] C. Romero. Extended lexicographic goal programming: A unifying
approach. Omega, 29:63–71, 2001.
[196] R. E. Rosenthal. Principles of multiobjective optimization. Deci-
sion Sciences, 16:133–152, 1985.
[197] E. E. Rosinger. Interactive algorithm for multiobjective optimiza-
tion. Journal of Optimization Theory and Applications, 35:339–
365, 1981. Errata Corrige in Journal of Optimization Theory and
Applications, 38:147–148, 1982.
[198] A. Roy and P. Mackin. Multicriteria optimization (linear
and nonlinear) using proxy value functions. In P. Korhonen,
A. Lewandowski, and J. Wallenius, editors, Multiple Criteria De-
cision Support, pages 128–134. Springer-Verlag, 1991.
REFERENCES 61

[199] A. Roy and J. Wallenius. Nonlinear multiobjective optimiza-

tion: An algorithm and some theory. Mathematical Programming,
55:235–249, 1992.
[200] B. Roy. The outranking approach and the foundations of ELEC-
TRE methods. In C. A. Bana e Costa, editor, Readings in Multi-
ple Criteria Decision Aid, pages 155–183, Berlin, Heidelberg, 1990.
Springer-Verlag.
[201] F. Ruiz, M. Luque, and K. Miettinen. Improving the compu-
tational efficiency of a global formulation (GLIDE) for interac-
tive multiobjective optimization. Annals of Operations Research,
197:47–70, 2012.
[202] H. Ruotsalainen, K. Miettinen, J.-E. Palmgren, and T. Lahtinen.
Interactive multiobjective optimization for anatomy based three-
dimensional HDR brachytherapy. Physics in Medicine and Biol-
ogy, 55:4703–4719, 2010.
[203] S. Sadagopan and A. Ravindran. Interactive algorithms for mul-
tiple criteria nonlinear programming problems. European Journal
of Operational Research, 25:247–257, 1986.
[204] M. Sakawa. Interactive multiobjective decision making by the se-
quential proxy optimization technique: SPOT. European Journal
of Operational Research, 9:386–396, 1982.
[205] M. Sakawa and K. Yauchi. An interactive fuzzy satisficing method
for multiobjective nonconvex programming problems through
floating point genetic algorithms. European Journal of Operational
Research, 117:113–124, 1999.
[206] Y. Sawaragi, H. Nakayama, and T. Tanino. Theory of Multiobjec-
tive Optimization. Academic Press, Orlando, Florida, 1985.
[207] W. S. Shin and A. Ravindran. Interactive multiple objective op-
timization: Survey I – continuous case. Computers & Operations
Research, 18:97–114, 1991.
[208] J. Silverman, R. E. Steuer, and A. W. Whisman. A multi-period,
multiple criteria optimization system for manpower planning. Eu-
ropean Journal of Operational Research, 34:160–170, 1988.
[209] K. Sindhya, V. Ojalehto, J. Savolainen, H. Niemistö, J. Hakanen,
and K. Miettinen. Coupling dynamic simulation and interactive
multiobjective optimization for complex problems: An APROS-
NIMBUS case study. Expert Systems with Applications, 41:2546–
2558, 2014.
[210] K. Sindhya, A. B. Ruiz, and K. Miettinen. A preference based in-
teractive evolutionary algorithm for multiobjective optimization:
62

PIE. In R.H.C. Takahashi, K. Deb, E.F. Wanner, and S. Greco,

editors, Evolutionary Multi-Criterion Optimization: 6th Interna-
tional Conference, Proceedings, pages 212–225, Berlin, Heidelberg,
2011. Springer-Verlag.
[211] A. M. J. Skulimowski. Decision Support Systems Based on Refer-
ence Sets. Wydawnictwa AGH, Kraków, 1996.
[212] R. Slowinski. Interactive multiobjective optimization based on
ordinal regression. In A. Lewandowski and V. Volkovich, editors,
Multiobjective Problems of Mathematical Programming, pages 93–
100. Springer-Verlag, 1991.
[213] A. Song and W.-M. Cheng. A method for multihuman and multi-
criteria decision making. In A. Sydow, S. G. Tzafestas, and
R. Vichnevetsky, editors, Systems Analysis and Simulation 1988 I:
Theory and Foundations, pages 213–216, Berlin, 1988. Akademie-
Verlag.
[214] J. Spronk. Interactive multifactorial planning: State of the art. In
C. A. Bana e Costa, editor, Readings in Multiple Criteria Decision
Aid, pages 512–534, Berlin, Heidelberg, 1990. Springer-Verlag.
[215] A. Stam, M. Kuula, and H. Cesar. Transboundary air pollution in
Europe: An interactive multicriteria tradeoff analysis. European
Journal of Operational Research, 56:263–277, 1992.
[216] R. B. Statnikov. Multicriteria Design: Optimization and Identifi-
cation. Kluwer Academic Publishers, Dordrecht, 1999.
[217] R. B. Statnikov and J. Matusov. Use of pτ -nets for the approx-
imation of the Edgeworth-Pareto set in multicriteria optimiza-
tion. Journal of Optimization Theory and Applications, 91:543–
560, 1996.
[218] I. Steponavice, S. Ruuska, and K. Miettinen. A solution process
for simulation-based multiobjective design optimization with an
application in paper industry. Computer-Aided Design, 47:45–58,
2014.
[219] R. E. Steuer. Multiple Criteria Optimization: Theory, Computa-
tion, and Applications. John Wiley & Sons, 1986.
[220] R. E. Steuer. The Tchebycheff procedure of interactive multi-
ple objective programming. In B. Karpak and S. Zionts, editors,
Multiple Criteria Decision Making and Risk Analysis Using Mi-
crocomputers, pages 235–249, Berlin, Heidelberg, 1989. Springer-
Verlag.
[221] R. E. Steuer. Implementing the Tchebycheff method in a spread-
sheet. In M. H. Karwan, J. Spronk, and J. Wallenius, editors, Es-
REFERENCES 63

says in Decision Making: A Volume in Honour of Stanley Zionts,

pages 93–103. Springer-Verlag, Berlin, Heidelberg, 1997.
[222] R. E. Steuer and E.-U. Choo. An interactive weighted Tchebycheff
procedure for multiple objective programming. Mathematical Pro-
gramming, 26:326–344, 1983.
[223] R. E. Steuer and L. R. Gardiner. Interactive multiple objective
programming: Concepts, current status, and future directions. In
C. A. Bana e Costa, editor, Readings in Multiple Criteria Decision
Aid, pages 413–444, Berlin, Heidelberg, 1990. Springer-Verlag.
[224] R. E. Steuer and L. R. Gardiner. On the computational testing of
procedures for interactive multiple objective linear programming.
In G. Fandel and H. Gehring, editors, Operations Research, pages
121–131, Berlin, Heidelberg, 1991. Springer-Verlag.
[225] R. E. Steuer, J. Silverman, and A. W. Whisman. A combined
Tchebycheff/aspiration criterion vector interactive multiobjective
programming procedure. Management Science, 39:1255–1260,
1993.
[226] R. E. Steuer and M. Sun. The parameter space investigation
method of multiple objective nonlinear programming: A compu-
tational investigation. Operations Research, 43:641–648, 1995.
[227] R. E. Steuer and A. W. Whisman. Toward the consolidation of
interactive multiple objective programming procedures. In G. Fan-
del, M. Grauer, A. Kurzhanski, and A. P. Wierzbicki, editors,
Large-Scale Modelling and Interactive Decision Analysis, pages
232–241. Springer-Verlag, 1986.
[228] T. J. Stewart. A critical survey on the status of multiple criteria
decision making theory and practice. Omega, 20:569–586, 1992.
[229] M. Sun, A. Stam, and R. E. Steuer. Solving multiple objective pro-
gramming problems using feed-forward artificial neural networks:
The interactive FFANN procedure. Management Science, 42:835–
849, 1996.
[230] M. Sun, A. Stam, and R. E. Steuer. Interactive multiple objec-
tive programming using Tchebycheff programs and artificial neural
networks. Computers & Operations Research, 27:601–620, 2000.
[231] T. Sunaga, M. A. Mazeed, and E. Kondo. A penalty function for-
mulation for interactive multiobjective programming problems. In
M. Iri and K. Yajima, editors, System Modelling and Optimization,
pages 221–230. Springer-Verlag, 1988.
[232] M. T. Tabucanon. Multiple Criteria Decision Making in Industry.
Elsevier, Amsterdam, 1988.
64

[233] M. Tamiz and D. F. Jones. A general interactive goal program-

ming algorithm. In G. Fandel and T. Gal, editors, Multiple Criteria
Decision Making: Proceedings of the Twelfth International Confer-
ence, pages 433–444, Berlin, Heidelberg, 1997. Springer-Verlag.
[234] C. G. Tapia and B. A. Murtagh. The use of preference criteria
in interactive multiobjective mathematical programming. Asia-
Pacific Journal of Operational Research, 6:131–147, 1989.
[235] C. G. Tapia and B. A. Murtagh. A Markovian process in interac-
tive multiobjective decision-making. European Journal of Opera-
tional Research, 57:421–428, 1992.
[236] R. V. Tappeta, J. E. Renaud, A. Messac, and G. J. Sundararaj.
Interactive physical programming: Tradeoff analysis and decision
making in multidisciplinary optimization. AIAA Journal, 38:917–
926, 2000.
[237] S. Tarkkanen, K. Miettinen, J. Hakanen, and H. Isomäki. Incre-
mental user-interface development for interactive multiobjective
optimization. Expert Systems with Applications, 40:3220–3232,
2013.
[238] K. Tarvainen. On the implementation of the interactive surro-
gate worth trade-off (ISWT) method. In M. Grauer and A. P.
Wierzbicki, editors, Interactive Decision Analysis, pages 154–161,
Berlin, Heidelberg, 1984. Springer-Verlag.
[239] A. Tecle and L. Duckstein. A procedure for selecting MCDM
techniques for forest resources management. In A. Goicoechea,
L. Duckstein, and S. Zionts, editors, Multiple Criteria Decision
Making: Proceedings of the Ninth International Conference: The-
ory and Applications in Business, Industry, and Government,
pages 19–32, New York, 1992. Springer-Verlag.
[240] J. Teghem Jr., C. Delhaye, and P. L. Kunsch. An interactive de-
cision support system (IDSS) for multicriteria decision aid. Math-
ematical and Computer Modelling, 12:1311–1320, 1989.
[241] A. Udink ten Cate. On the determination of the optimal temper-
ature for the growth of an early cucumber crop in a greenhouse.
In M. Grauer, M. Thompson, and A. P. Wierzbicki, editors, Plu-
ral Rationality and Interactive Decision Processes, pages 311–318,
Berlin, Heidelberg, 1985. Springer-Verlag.
[242] D. Vanderpooten. Multiobjective programming: Basic concepts
and approaches. In R. Slowinski and J. Teghem, editors, Stochas-
tic versus Fuzzy Approaches to Multiobjective Mathematical Pro-
REFERENCES 65

gramming under Uncertainty, pages 7–22, Dordrecht, 1990. Kluwer

Academic Publishers.
[243] D. Vanderpooten and P. Vincke. Description and analysis of some
representative interactive multicriteria procedures. Mathematical
and Computer Modelling, 12:1221–1238, 1989.
[244] R. Vetschera. Feedback-oriented group decision support in a refer-
ence point framework. In A. Lewandowski and V. Volkovich, edi-
tors, Multiobjective Problems of Mathematical Programming, pages
309–314. Springer-Verlag, 1991.
[245] R. Vetschera. A note on scalarizing functions under changing sets
of criteria. European Journal of Operational Research, 52:113–118,
1991.
[246] P. Vincke. Multicriteria Decision-Aid. John Wiley & Sons, Chich-
ester, 1992.
[247] J. Wallenius. Comparative evaluation of some interactive ap-
proaches to multicriterion optimization. Management Science,
21:1387–1396, 1975.
[248] S. Wang. Algorithms for multiobjective and nonsmooth optimiza-
tion. In P. Kleinschmidt, F. J. Radermacher, W. Schweitzer, and
H. Wildermann, editors, Methods of Operations Research 58, pages
131–142, Frankfurt am Main, 1989. Athenum Verlag.
[249] S. Wang. An interactive method for multicriteria decision mak-
ing. In K. H. Phua, C. M. Wang, W. Y. Yeong, T. Y. Leong, H. T.
Loh, K. C. Tan, and F. S. Chou, editors, Optimization: Techniques
and Applications, volume 1, pages 307–316. Proceedings of the In-
ternational Conference (ICOTA), World Scientific Publishing Co.,
1992.
[250] H. R. Weistroffer. Multiple criteria decision making with interac-
tive over-achievement programming. Operations Research Letters,
1:241–245, 1982.
[251] H. R. Weistroffer. An interactive goal-programming method for
non-linear multiple-criteria decision-making problems. Computers
& Operations Research, 10:311–320, 1983.
[252] H. R. Weistroffer. A combined over- and under-achievement pro-
gramming approach to multiple objective decision-making. Large
Scale Systems, 7:47–58, 1984.
[253] H. R. Weistroffer. A flexible model for multi-objective optimiza-
tion. In J. Jahn and W. Krabs, editors, Recent Advances and
Historical Developments of Vector Optimization, pages 311–316,
Berlin, Heidelberg, 1987. Springer-Verlag.
66

[254] R. E. Wendell and D. N. Lee. Efficiency in multiple objective

optimization problems. Mathematical Programming, 12:406–414,
1977.
[255] D. J. White. A selection of multi-objective interactive program-
ming methods. In S. French, R. Hartley, L. C. Thomas, and D. J.
White, editors, Multi-Objective Decision Making, pages 99–126.
Academic Press, 1983.
[256] A. P. Wierzbicki. The use of reference objectives in multiobjective
optimization. In G. Fandel and T. Gal, editors, Multiple Criteria
Decision Making Theory and Applications, pages 468–486, Berlin,
Heidelberg, 1980. Springer-Verlag.
[257] A. P. Wierzbicki. A mathematical basis for satisficing decision
making. Mathematical Modelling, 3:391–405, 1982.
[258] A. P. Wierzbicki. A methodological approach to comparing para-
metric characterizations of efficient solutions. In G. Fandel,
M. Grauer, A. Kurzhanski, and A. P. Wierzbicki, editors, Large-
Scale Modelling and Interactive Decision Analysis, pages 27–45.
Springer-Verlag, 1986.
[259] A. P. Wierzbicki. On the completeness and constructiveness of
parametric characterizations to vector optimization problems. OR
Spectrum, 8:73–87, 1986.
[260] A. P. Wierzbicki. Convergence of interactive procedures of mul-
tiobjective optimization and decision support. In M. H. Karwan,
J. Spronk, and J. Wallenius, editors, Essays in Decision Making:
A Volume in Honour of Stanley Zionts, pages 19–47. Springer-
Verlag, Berlin, Heidelberg, 1997.
[261] A. P. Wierzbicki. Reference point approaches. In T. Gal,
T. J. Stewart, and T. Hanne, editors, Multicriteria Decision Mak-
ing: Advances in MCDM Models, Algorithms, Theory, and Ap-
plications, pages 9–1–9–39. Kluwer Academic Publishers, Boston,
1999.
[262] A. P. Wierzbicki and J. Granat. Multi-objective modeling for
engineering application in decision support. In G. Fandel and
T. Gal, editors, Multiple Criteria Decision Making: Proceedings
of the Twelfth International Conference, pages 529–540, Berlin,
Heidelberg, 1997. Springer-Verlag.
[263] A. P. Wierzbicki and J. Granat. Multi-objective modeling for en-
gineering applications: DIDASN++ system. European Journal of
Operational Research, 113:372–389, 1999.
REFERENCES 67

[264] H.-M. Winkels and M. Meika. An integration of efficiency pro-

jections into the Geoffrion approach for multiobjective linear pro-
gramming. European Journal of Operational Research, 16:113–127,
1984.
[265] E. Wood, N. P. Greis, and R. E. Steuer. Linear and nonlinear
applications of the Tchebycheff metric to the multi criteria water
allocation problem. In S. Rinaldi, editor, Environmental Systems
Analysis and Management, pages 363–376. North-Holland Publish-
ing Company, 1982.
[266] J.-B. Yang. Gradient projection and local region search for multi-
objective optimization. European Journal of Operational Research,
112:432–459, 1999.
[267] J.-B. Yang, C. Chen, and Z.-J. Zhang. The interactive step trade-
off method (ISTM) for multiobjective optimization. IEEE Trans-
actions on Systems, Man, and Cybernetics, 20:688–695, 1990.
[268] J.-B. Yang and D. Li. Normal vector identification and interactive
tradeoff analysis using minimax formulation in multiobjective op-
timization. IEEE Transactions on Systems, Man, and Cybernetics,
32:305–319, 2002.
[269] J.-B. Yang and P. Sen. Preference modelling by estimating local
utility functions for multiobjective optimization. European Journal
of Operational Research, 95:115–138, 1996.
[270] P. L. Yu. A class of solutions for group decision problems. Man-
agement Science, 19:936–946, 1973.
[271] P. L. Yu. Multiple-Criteria Decision Making Concepts, Techniques,
and Extensions. Plenum Press, New York, 1985.
[272] L. Zadeh. Optimality and non-scalar-valued performance criteria.
IEEE Transactions on Automatic Control, 8:59–60, 1963.
[273] M. Zeleny. Compromise programming. In J. L. Cochrane and
M. Zeleny, editors, Multiple Criteria Decision Making, pages 262–
301, Columbia, South Carolina, 1973. University of South Carolina
Press.