Multi-Objective Optimization

Carlos A. Coello Coello

[email protected]
CINVESTAV-IPN
Evolutionary Computation Group (EVOCINV)
Computer Science Department
Av. IPN No. 2508, Col. San Pedro Zacatenco

Mexico, D.F. 07360, MEXICO

Patna, India, December, 2016

LECTURE 1

Carlos A. Coello Coello Multi-Objective Optimization

Motivation

Most problems in nature have several (possibly conflicting)

objectives to be satisfied (e.g., design a bridge for which want
to minimize its weight and cost while maximizing its safety).
Many of these problems are frequently treated as
single-objective optimization problems by transforming all but
one objective into constraints.

Carlos A. Coello Coello Multi-Objective Optimization

Formal Definition

T
Find the vector ~x = x1 , x2 , . . . , xn which will satisfy the m


inequality constraints:

gi (~x ) 0 i = 1, 2, . . . , m (1)
the p equality constraints

hi (~x ) = 0 i = 1, 2, . . . , p (2)
and will optimize the vector function

~f (~x ) = [f1 (~x ), f2 (~x ), . . . , fk (~x )]T (3)

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

In order to know how good a certain solution is, it is

necessary to have some criteria to evaluate it. These criteria
are expressed as computable functions of the decision
variables, that are called objective functions.

In real-world problems, some of these objective functions are in

conflict with others, and some have to be minimized while
others are maximized. These objective functions may be
commensurable (measured in the same units) or
non-commensurable (measured in different units).

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

In Operations Research, it is a common practice to differentiate

among attributes, criteria, objectives and goals.

Attributes are often thought of as differentiating aspects,

properties or characteristics of alternatives or consequences.
Criteria generally denote evaluative measures, dimensions or
scales against which alternatives may be gauged in a value or
worth sense. Objectives are sometimes viewed in the same
way, but may also denote specific desired levels of attainment
or vague ideals. Goals usually indicate either of the latter
notions. A distinction commonly made in Operations Research
is to use the term goal to designate potentially attainable levels,
and objective to designate unattainable ideals.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Several researchers use the terms objective, criteria, and

attribute interchangeably to represent an MOPs goals or
objectives (i.e., distinct mathematical functions) to be achieved.
The terms objective space or objective function space are
also used to denote the coordinate space within which vectors
resulting from evaluating an MOP are plotted.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Ideal Objective Vector

It is an objective vector minimizing each of the objective
functions. The components zi of the ideal objective vector
z Rk are obtained by minimizing each of the objective
functions individually, subject to the constraints. That is, it is
obtained by solving:

minimize fi (~x ) (4)

subject to ~x F, for i = 1, . . . , k (5)

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Ideal Objective Vector

The ideal vector is unreachable in most cases (except when
there is no conflict among the objectives). However, the ideal
vector is adopted by some mathematical programming
techniques in which normally the idea is to minimize the
distance of a solution with respect to such ideal vector.

Generating the ideal vector is not particularly complicated

(except when some (or all) of the objective functions, when
considered in isolation, presents multimodality). However, its
generation has an additional computational cost that is not
always affordable. Some multi-objective metaheuristics adopt
an approximation of the ideal vector that is updated at each
iteration.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Utopian Objective Vector

Some authors (e.g., Miettinen [1999]) also consider the concept
of utopian objective vector.

The utopian objective vector is defined as: z Rk and it is an

infeasible objective vector whose components are formed by:

z = zi i (6)
for every i = 1, . . . , k, where zi is a component of the ideal
objective vector and i > 0 is a scalar which is relatively small,
but computationally significant. Clearly, the utopian objective
vector is strictly better (i.e., it strictly dominates) every Pareto
optimal solution.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Nadir Objective Vector

It refers to the upper bounds of the Pareto optimal set. It is
normally denoted as znad and its components are normally
quite difficult to obtain.

When computing the ideal vector, normally a payoff table is

created using the decision vectors obtained. Row i of the payoff
table displays the values of all the objective functions calculated
at the point where fi obtained its minimal value. Hence, zi is at
the main diagonal of the table. The maximal value of the
column i in the payoff table can be selected as an estimate of
the upper bound of the objective fi for i = 1, . . . , k over the
Pareto optimal set.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

In this figure, a black circle is used to indicate the Ideal

Objective Vectors and a gray circle is use to indicate the Nadir
Objective Vectors. It is worth noting that the Nadir Objective
Vector may be infeasible.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

This figure shows the Ideal Objective Vector (z ), the Utopian

Objective Vector (z ) and the Nadir Objective Vector (z nad ).

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

In multiobjective optimization problems, there are three

possible situations:
Minimize all the objective functions
Maximize all the objective functions
Minimize some and maximize others

For simplicity reasons, normally all the functions are converted

to a maximization or minimization form. For example, the
following identity may be used to convert all the functions which
are to be maximized into a form which allows their minimization:

max fi (~x ) = min(fi (~x )) (7)

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Having several objective functions, the notion of optimum

changes, because in MOPs, the aim is to find good
compromises (or trade-offs) rather than a single solution as in
global optimization.

The notion of optimum that is most commonly adopted is that

originally proposed by Francis Ysidro Edgeworth (in 1881) in
his book entitled Mathematical Psychics.
Carlos A. Coello Coello Multi-Objective Optimization
Basic Concepts

This notion was generalized by the italian economist Vilfredo

Pareto (in 1896) in his book Cours dEconomie Politique.
Although some authors call Edgeworth-Pareto optimum to this
notion (originally called ophelimity) it is normally preferred to
use the most commonly accepted term: Pareto optimum.
Carlos A. Coello Coello Multi-Objective Optimization
Basic Concepts

Pareto Optimality
We say that a vector of decision variables ~x F is Pareto
optimal if there does not exist another ~x F such that
fi (~x ) fi (~x ) for all i = 1, . . . , k and fj (~x ) < fj (~x ) for at least
one j (assuming that all the objectives are being minimized).

Other important definitions

In words, this definition says that ~x is Pareto optimal if there
exists no feasible vector of decision variables ~x F which
would decrease some criterion without causing a simultaneous
increase in at least one other criterion. This concept normally
produces a set of solutions called the Pareto optimal set. The
vectors ~x corresponding to the solutions included in the Pareto
optimal set are called nondominated. The image of the Pareto
optimal set is called the Pareto front.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts
Pareto Dominance
A vector ~u = (u1 , . . . , uk ) is said to dominate ~v = (v1 , . . . , vk )
(denoted by ~u  ~v ) if and only if u is partially less than v , i.e.,
i {1, . . . , k }, ui vi i {1, . . . , k} : ui < vi .

Pareto Optimal Set

For a given MOP ~f (x), the Pareto optimal set (P ) is defined as:

P := {x F | x 0 F ~f (x 0 )  ~f (x)}. (8)

Pareto Front
For a given MOP ~f (x) and Pareto optimal set P , the Pareto
front (PF ) is defined as:

PF := {~u = ~f = (f1 (x), . . . , fk (x)) | x P }. (9)

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

The concept of Pareto dominance implies that, for a solution

to dominate another one, it should not be worse in any
objective and must be strictly better in at least one of them.

Consequently, when comparing two solutions A and B, using

Pareto dominance, there are three possible outcomes:
1 A dominates B.
2 A is dominated by B.
3 A and B are not dominated by each other (i.e., they are
both non-dominated).

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Properties of the dominance relation

Cormen et al. [1990] provide the properties of the dominance
relation:
Reflection: The dominance relation is not reflexive,
because any relation p does not dominate itself.
Symmetry: The dominance relation is not symmetric
because p  q does not imply q  p. In fact, the opposite
is true. In other words, if p dominates q, then q does not
dominate p. Therefore, the dominance relation is
asymmetric.
Antisymmetry: Since the dominance relation is not
symmetric, it cant be antisymmetric.
Transitivity: The dominance relation is transitive. This is
because if p  q and q  r , then p  r .

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Properties of the dominance relation

Another interesting property of the dominance relation is that if
a solution p does not dominate another solution q, this does not
imply that q dominates p.

For a binary relation to qualify as an order relation, it must be at

least transitive [Chankong & Haimes, 1983]. Thus, the
dominance relation is an order relation. However, since the
dominance relation is not reflexive, it is a strict partial order.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Properties of the dominance relation

In general, if a relation is reflexive, antisymmetric and transitive,
it is called (in a general sense) a partial order. A set in which a
partial order is defined is called partially ordered set.

However, it is important to emphasize that the dominance

relation is not reflexive and is not antisymmetric. Therefore, the
dominance relation is not a partial order, but only a strict partial
order.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Weak Pareto Optimality

A point ~x F is a weakly Pareto optimal solution if there is
no ~x F such that fi (~x ) < fi (~x ), for i = 1, . . . , k .

Strong Pareto Optimality

A point ~x F is a strongly Pareto optimal solution if there is
~x F such that fi (~x ) fi (~x ), for i = 1, . . . , k.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Pareto Front
For a given multi-objective optimization problem ~f (x) and a
Pareto optimal set P , the Pareto Front (PF ) is defined as:

PF := {~u = ~f = (f1 (x), . . . , fk (x)) | x P }. (10)

Carlos A. Coello Coello Multi-Objective Optimization
Basic Concepts

Pareto Front
In general, it is impossible to find an analytical expression that
represents the line or hyper-surface corresponding to the
Pareto Optimal Front. This is possible only in very simple
(textbook) cases.

The normal procedure for generating (an approximation of) the

Pareto optimal front of a problem is to compute all the (or as
many as possible) feasible points and to obtain their
corresponding objective function values. When we had
obtained a sufficient number of such points, it is possible to
determine which are the nondominated solutions from them
(the Pareto optimal set). As indicated before, the image of the
Pareto optimal set is the Pareto front.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Algorithm 1 to Obtain Nondominated Solutions (Simple Method)

Algorithm
Step 1: Set a counter i = 1 and create a (empty) set of
nondominated solutions P 0
Step 2: For a solution j P, where P is the population
(j 6= i), check if the solution j dominates solution i.
If it dominates it, go to Step 4.
Step 3: If there are more solutions left in P, increase j by
one and go to Step 2; otherewise, P 0 = P 0 {i}
Step 4: Increase i by one. If i N, go to Step 2; otherwise,
stop and declare P as the nondominated set.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts
Algorithm 2 to Obtain Nondominated Solutions (Continuous
Update)
Algorithm
Step 1: Initialize P 0 = {1}. Set a counter i = 2.
Step 2: Set j = 1.
Step 3: Compare solutions i and j from P 0 .
Step 4: If i dominates j, then delete the j th member of P 0 .
If j < |P 0 |, then increase j by one and go to Step 3.
Otherwise, go to Step 5.
Alternatively, if the j th member of P 0 dominates i,
increase i by one and go to Step 2.
Step 5: Insert i in P 0 . If i < N, increase i by one and
go to Step 2.
Otherwise, stop and declare P 0 as the nondominated
set.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Computational Efficiency
Both Algorithm 1 and Algorithm 2 shown before have an
algoritmic complexity O(MN 2 ) in the worst case. In this case M
is the number of objectives and N is the number of solutions.
However, in practice, Algorithm 2 requires about half of the
computational effort required by Algorithm 1.

So the obvious question is: can we obtain nondominated

solutions in a more efficient way (computational
speaking)?

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Algorithm 3 to Obtain Nondominated Solutions

Theoretically, the most efficient algorithm that we can have for
obtaining nondominated solutions is the one proposed by Kung
et al. [1975]. This algorithm requires the set (or population) to
be sorted based on the first objective. Then, the population is
recursively divided in two halves: (S) superior and (I) inferior.
Knowing that the first half is better than the second in terms of
the first objective function, the inferior half is checked (in terms
of Pareto dominance) with respect to the superior half.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Algorithm 3 to Obtain Nondominated Solutions

The solutions of (I) which are not dominated by any member of
(S) are combined with the members of (S) to form a mixed
population M. This union and dominance checking starts with
the most inner case (in which there is only one member either
in S or in I, after performing several recursive divisions of the
population) and then the algorithm continues in a bottom up
manner.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Algorithm 3 to Obtain Nondominated Solutions

This algorithm has a complexity O(N(log M)M2 ) for M 4 and
O(N log N) for M = 2 y M = 3.

So, as we increase the number of objectives, Algorithm 3 also

approximates the quadratic complexity of Algorithm 1 and
Algorithm 2. However, for 2 or 3 objectives, this algorithm is
clearly more efficient.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

For more on Kungs Algorithm

H.T. Kung, F. Luccio, and F.P. Preparata, On finding the
maxima of a set of vectors, Journal of the Association for
Computing Machinery, 22(4):469476, 1975.

For more information on this topic

J.L. Bentley, H.T. Kung, M. Schkolnick, and C.D.
Thompson, On the Average Number of Maxima in a Set
of Vectors and Applications, Journal of the Association
for Computing Machinery, 25(4):536543, October 1978.
Jon Louis Bentley, Multidimensional
Divide-and-Conquer, Communications of the ACM,
23(4):214229, April 1980.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

For more information on this topic

Jon L. Bentley, Kenneth L. Clarkson, and David B. Levine,
Fast Linear Expected-Time Algorithms for Computing
Maxima and Convex Hulls, Algorithmica, 9:168183,
1993.
Lixin Ding, Sanyou Zheng, and Lishan Kang, A Fast
Algorithm on Finding the Non-dominated Set in
Multi-objective Optimization, in Proceedings of the 2003
Congress on Evolutionary Computation (CEC2003),
Vol. 4, pp 25652571, IEEE Press, Canberra, Australia,
December 2003.
Michael A. Yukish, Algorithms to Identify Pareto Points
in Multi-Dimensional Data Sets, PhD thesis, College of
Engineering, Pennsylvania State University, USA, August
2004.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts
For more information on this topic
Parke Godfrey, Ryan Shipley, and Jarek Gryz, Maximal
Vector Computation in Large Data Sets, Technical
Report CS-2004-06, Department of Computer Science and
Engineering, York University, Canada, December 2004.
K.K. Mishra and Sandeep Harit, A Fast Algorithm for
Finding the Non Dominated Set in Multiobjective
Optimization, International Journal of Computer
Applications, 1(25):3539, 2010.
Maxim Buzdalov and Anatoly Shalyto, A Provably
Asymptotically Fast Version of the Generalized Jensen
Algorithm for Non-dominated Sorting, in Thomas
Bartz-Beielstein et al. (Eds), Parallel Problem Solving from
Nature - PPSN XIII, 13th International Conference, pp.
528537. Springer. Lecture Notes in Computer Science
Vol. 8672, Ljubljana, Slovenia, September 13-17 2014.
Carlos A. Coello Coello Multi-Objective Optimization
Basic Concepts

Optimality Conditions
Fritz-Johns Necessary Condition. A necessary condition for
x to be Pareto optimal is that there exist vectors 0 and
u 0 (where RM , u RJ and , u 6= 0) such that the
following conditions hold:
PM
Pf
m=1 m fm (x ) j=1 uj gj (x ) = 0
1

2 uj gj (x ) = 0 for every j = 1, 2, . . . , J.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Optimality Conditions
These conditions are very similar to the Kuhn-Tucker conditions
of optimality for single-objective problems. The difference lies
on the addition (in this case) of the vector of the gradients of the
objectives.

For an unconstrained multi-objective optimization problem, the

previous theorem requires the following condition:
M
X
m fm (x ) = 0 (11)
m=1

for a solution to be Pareto optimal.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Optimality Conditions
For nonlinear objective functions, it is expected that the partial
derivatives are nonlinear. For a given vector , it is possible to
check the non-existence of a Pareto optimal solution using the
previously defined conditions.

If the necessary conditions are not satisfied, then there does

not exist a Pareto optimal solution corresponding to the given
vector . It is worth noting, however, that since this is a
necessary condition, the existence of a solution that is Pareto
optimal is not guaranteeed. In other words, a solution that
satisfies these conditions is not necessarily Pareto optimal.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

Optimality Conditions
Kuhn-Tuckers Sufficiency Conditions for Pareto
Optimality: Lets assume that the objective functions are
convex and the constraints are non-convex. Lets assume that
the objective functions and the constraints are continuously
differentiable in a feasible solution x . A sufficient condition for
x to be Pareto optimal is that there exist vectors > 0 and
u 0 (where RM and u RJ ) such that the following
equations hold:
PM
PJ
m=1 m fi (x ) j=1 uj gj (x ) = 0
1

2 uj gj (x ) = 0 for every j = 1, 2, . . . , J.

Carlos A. Coello Coello Multi-Objective Optimization

Basic Concepts

For more information

Kaisa M. Miettinen, Nonlinear Multiobjective Optimization,
Kluwer Academic Publishers, Boston, Massachusetts, 1999.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Multiobjective optimization is an intrinsic part of economic

equilibium theory and, as such, it can be said to have been
founded by Adam Smith in his famous treatise entitled An
Inquiry into the Nature and Causes of the Wealth of
Nations, in 1776.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

The concept of general economic equilibrium is normally

attributed to Leon Walras (1834-1910). Within economic
equilibrium theory, the most relevant works (besides those of
Walras) are those from Jevons and Menger on utility theory,
and the work on welfare theory by Edgeworth and Pareto,
spanning the period from 1874 to 1906.
Carlos A. Coello Coello Multi-Objective Optimization
Historical Highlights of Multi-Objective Optimization

A second area that is considered to be one of the main

precursors of multi-objective optimization is the inception of the
psycological theory of games and the notion of (game) strategy.

Games of chance have a very ancient history. However, Felix

Edouard Justin Emile Borel (1871-1956) is normally considered
as the one who started the psicological theory of games and
the one who introduced the formal definition of strategies that
are based on analyzing the psychology of the opponent.
Carlos A. Coello Coello Multi-Objective Optimization
Historical Highlights of Multi-Objective Optimization

Game Theory
The so-called game theory can be traced back to a work by
Borel from 1921. However, many historians normally attribute
the origins of game theory to a paper from the famous
hungarian mathematician John von Neumann which was orally
presented in 1926 and published in 1928.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Game Theory
In 1944, John von Neumann and Oskar Morgenstern
mentioned (in their famous book on Game Theory) that they
had found a problem in economics that was a peculiar and
disconcerting mixture of several problems in conflict with each
other which could not be solved with the classical optimization
methods known at that time. It remains a mystery why is that
von Neummann did not get interested in this peculiar problem.
Carlos A. Coello Coello Multi-Objective Optimization
Historical Highlights of Multi-Objective Optimization

In 1951, Tjalling C. Koopmans edited a book entitled Activity

Analysis of Production and Allocation, in which the concept
of efficient vector (which is the same as a nondominated
vector) was used in a meaningful way for the very first time.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Mathematical Foundations
The origins of the mathematical foundations of multi-objective
optimization can be traced back to the period from 1895 to
1906 in which Georg Cantor and Felix Hausdorff established
the foundations of ordered spaces of infinite dimensions.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Mathematical Foundations
Cantor also introduced equivalent classes and established the
first set of sufficiency conditions for the existence of a utility
function.

Hausdorff provided the first example of a complete ordering.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Mathematical Foundations
However, it was the concept of the maximum vector problem
introduced by Harold W. Kuhn and Albert W. Tucker (1951)
which allowed multi-objective optimization to become a
mathematical discipline on its own.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Mathematical Foundations
It is well-known that the now famous conditions of optimality
commonly attributed to Kuhn and Tucker had been previously
stated and proved by W. Karush in an unpublished Masters
thesis in 1939.

Kuhn and Tucker gave credit to Karush, which is the reason

why many books call them Karush-Kuhn-Tucker (KKT)
Conditions.
Carlos A. Coello Coello Multi-Objective Optimization
Historical Highlights of Multi-Objective Optimization

Mathematical Foundations
Nevertheless, the theory of multi-objective optimization
remained practically unexplored during the 1950s. It was until
the 1960s, in which the mathematical foundations of the area
were consolidated when Leonid Hurwicz generalized Kuhn and
Tuckers results to topological vector spaces.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Kenneth J. Arrow did some very important pioneering work in

the 1950s using the concept of admissible points and stating
his famous impossibility theorem which relates to multi-criteria
decision making.
Carlos A. Coello Coello Multi-Objective Optimization
Historical Highlights of Multi-Objective Optimization

Goal Programming
Perhaps the most important outcome from the 1950s was the
development of Goal Programming, which was originally
introduced by Abraham Charnes and William Wager Cooper in
1957. However, Goal Programming became popular in the
1960s.
Carlos A. Coello Coello Multi-Objective Optimization
Historical Highlights of Multi-Objective Optimization

Applications
The first application of multi-objective optimization outside
economics was done by Koopmans (1951) in production theory.
Later on, Marglin (1967) developed the first applications of
multi-objective optimization in water resources.

Carlos A. Coello Coello Multi-Objective Optimization

Historical Highlights of Multi-Objective Optimization

Applications
The first engineering application of multi-objective optimization
reported in the literature is a paper published by Lofti Zadeh in
the early 1960s (related to automatic control). However,
multi-objective optimization applications generalized until the
1970s.

Carlos A. Coello Coello Multi-Objective Optimization

A Taxonomy of Multi-Objective Optimization
Techniques

There have been several attempts to classify the many

multiobjective optimization techniques currently in use. First of
all, it is quite important to distinguish two stages in which the
solution of a multiobjective optimization problem can be
divided: the optimization of the several objective functions
involved and the process of deciding what kind of trade-offs
are appropriate from the decision maker perspective (the
so-called multicriteria decision making process).
Carlos A. Coello Coello Multi-Objective Optimization
A Taxonomy of Multi-Objective Optimization
Techniques

Cohon and Marks [1975] proposed one of the most popular

classifications of techniques within the Operations Research
community:
1 Generating techniques (a posteriori articulation of
preferences).
2 Techniques which rely on prior articulation of preferences
(non-interactive methods).
3 Techniques which rely on progressive articulation of
preferences (interaction with the decision maker).

Carlos A. Coello Coello Multi-Objective Optimization

A Taxonomy of Multi-Objective Optimization
Techniques

Other classifications are obviously possible (see for example

[Duckstein, 1984]). However, the classification proposed by
Cohon and Marks [1975] will be adopted here, because it
focuses the classification on the way in which each technique
handles the two problems of searching and making
(multi-criterion) decisions:
1 A priori Preference Articulation: make decisions before
searching (decide search).
2 A posteriori Preference Articulation: search before
making decisions (search decide).
3 Progressive Preference Articulation: integrate search
and decision making (decide search).

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

This group of techniques includes those approaches that

assume that either a certain desired achievable goals or a
certain pre-ordering of the objectives can be performed by the
decision maker prior to the search. Some representative
algorithms within this class are the following:
Global Criterion Method
Goal Programming
Goal-Attainment Method
Lexicographic Ordering
Min-Max Optimization
All of them will be briefly described next.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Global Criterion Method

In this method, the aim is to minimize a function which defines
a global criterion which is a measure of how close the decision
maker can get to the ideal vector ~f 0 . The most common form of
this function is [Osyczka, 1984]:
k
!p
X fi0 fi (~x )
f (~x ) = (12)
i=1
fi0
where k is the number of objectives.
It is normally suggested to use p = 1 or p = 2, but other values
of p can also be used. Obviously, the results differ greatly
depending on the value of p chosen. Thus, the selection of the
best p is an issue in this method, and it could also be the case
that any p could produce an unacceptable solution.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Global Criterion Method

Another possible measure of closeness to the ideal solution is
a family of Lp -metrics defined as follows:

k 
" #1/p
X p
 0
Lp (f ) = f f (x) , 1p (13)

i i 
i=1

In general, relative deviations of the form

fi0 fi (x)
(14)
fi0
are preferred over absolute deviations, because they have a
substantive meaning in any context.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Global Criterion Method

The relevant Lp metrics are
 p #1/p
k  0
"
 fi fi (~x ) 
X 
Lp (f ) = , 1p (15)
fi0
 
 
i=1

The value of p indicates the type of distance: for p = 1, all

deviations from fi are taken into account in direct proportion to
their magnitudes, which corresponds to group utility
[Duckstein, 1980].

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Global Criterion Method

Koski [1984] has suggested Lp -metrics with a normalized
vector objective function of the form

fi (~x ) minfi (~x )

xF
fi (~x ) = (16)
maxfi (~x ) minfi (~x )
xF xF

In this case, the values of every normalized function are limited

to the range [0,1].

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Compromise Programming
Using the global criterion method one non-inferior solution is
obtained. If certain parameters wi are used as weights for the
criteria, a required set of non-inferior solutions can be found.
Duckstein [1984] calls this method compromise
programming.

 f (~x ) f 0 p 1/p
  #
k
"
 i
wip 
X
i 
Lp (~x ) = (17)
 fi max fi0 

i=1

where wi are the weights, fi max is the worst value obtainable for
criterion i; fi (~x ) is the result of implementing decision ~x with
respect to the i th criterion.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Displaced Ideal
The Displaced Ideal technique [Zeleny, 1977] which proceeds
to define an ideal point, a solution point, another ideal point,
etc. is an extension of compromise programming.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Wierzbickis Method
Another variation of this technique is the method suggested by Wierzbicki
[1979, 1980] in which the global function has a form such that it penalizes the
deviations from the so-called reference objective. Any reasonable or
desirable point in the space of objectives chosen by the decision maker can
be considered as the reference objective.
Let ~f r = [f1r , f2r , . . . , fkr ]T be a vector which defines this point. Then the function
which is minimized has the form
k k
P(~x , ~f r ) =
X X
(fi (~x fir )2 + % (max(0, fi (~x fir )2 ) (18)
i=1 i=1

where % > 0 is a penalty coefficient which in this method can be chosen as

constant.
By minimizing (18) for the assumed point ~f r , a non-inferior solution which is
close to this point can be obtained. If for different points ~f r the procedure is
carried out, some representation of non-inferior solutions can be found.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

More information on the Global Criterion Method can be found

at the following references:
A. Osyczka, Multicriterion Optimization in Engineering
with FORTRAN programs, Ellis Horwood Limited, 1984.
M. Zeleny, Compromise Programming, in J. Cochrane
and M. Zeleny, editors, Multiple Criteria Decision
Making, pp. 262301. University of South Carolina Press,
Columbia, South Carolina, 1973.
M. Zeleny, Multiple Criteria Decision Making,
McGraw-Hill Book Company, New York, 1982.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Programming
Charnes and Cooper [1961] and Ijiri [1965] are credited with
the development of the goal programming method for a linear
model, and played a key role in applying it to industrial
problems. This was one of the earliest techniques specifically
designed to deal with multiobjective optimization problems.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Programming
In this method, the decision maker (DM) has to assign targets
or goals that wishes to achieve for each objective. These
values are incorporated into the problem as additional
constraints. The objective function then tries to minimize the
absolute deviations from the targets to the objectives. The
simplest form of this method may be formulated as follows:
k
X
min |fi (~x ) Ti | , subject to ~x F (19)
i=1

where Ti denotes the target or goal set by the decision maker

for the i th objective function fi (~x ), and F represents the feasible
region. The criterion, then, is to minimize the sum of the
absolute values of the differences between target values and
actually achieved values.
Carlos A. Coello Coello Multi-Objective Optimization
A Priori Preference Articulation
Goal Programming
A more general formulation of the goal programming objective
function is a weighted sum of the pth power of the deviation
|fi (~x ) Ti | [Haimes, 1975]. Such a formulation has been called
generalized goal programming [Ignizio, 1976; Ignizio, 1981].
Looking again to equation (19), the objective function is
nonlinear and the simplex method can be applied only after
transforming this equation into a linear form, thus reducing goal
programming to a special type of linear programming. In this
transformation, new variables di+ and di are defined such that
[Charnes, 1961]:

1
di+ = {|fi (~x ) Ti | + [fi (~x ) Ti ]}, (20)
2
1
di = {|fi (~x ) Ti | [fi (~x ) Ti ]}, (21)
2
Carlos A. Coello Coello Multi-Objective Optimization
A Priori Preference Articulation
Goal Programming
Adding and subtracting these equations, the following
equivalent linear formulation may be found:
k
X
min Z0 = (di+ + di ), (22)
i=1

subject to
~x F
fi (~x ) di+ + di = Ti (23)
di+ , di 0, i = 1, . . . , k
Since it is not possible to have both under- and
overachievements of the goal simultaneously, then at least one
of the deviational variables must be zero. In other words:

di+ di = 0 (24)
Carlos A. Coello Coello Multi-Objective Optimization
A Priori Preference Articulation

Goal Programming
Fortunately, this constraint is automatically fulfilled by the
simplex method because the objective function drives either di+
or di or both variables simultaneously to zero for all i.

Sometimes it may be desirable to express preference for over-

or underachievement of a goal. Thus, it may be more desirable
to overachieve a targeted reliability figure than to underachieve
it. To express preference for deviations, the DM can assign
relative weights wi+ and wi to positive and negative deviations,
respectively, for each target Ti . If a minimization problem is
considered, choosing the wi+ to be larger than wi would be
expressing preference for underachievement of a goal.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Programming
In addition, goal programming provides the flexibility to deal
with cases that have conflicting multiple goals.

Essentially, the goals may be ranked in order of importance to

the problem solver. That is, a priority factor, pi (i = 1, . . . , k ) is
assigned to the deviational variables associated with the goals.
This variation of goal programming is called lexicographic
ordering by some authors (e.g., [Miettinen, 1999]).

It is worth noting that these factors pi are conceptually different

from weights.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Programming
The resulting optimization model becomes
k
X
min S0 = pi (wi+ di+ + wi di ), (25)
i=1

subject to

~x F
fi (~x ) di+ + di = Ti (26)
di , di 0, i = 1, . . . , k
+

Note that this technique yields a nondominated solution if the

goal point is chosen in the feasible domain [Duckstein, 1984].

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Programming
More information on Goal Programming can be found at the following
references:

A. Charnes, W. W. Cooper, R. J. Niehaus, and A. Stedry, Static and

dynamic assignment models with multiple objectives and some
remarks on organization design, Management Science,
15(8):B365B375, 1969.
S. Lee, Goal Programming for Decision Analysis, Auerbach,
Philadelphia, 1972.
J. P. Ignizio, Goal Programming and Extensions, Heath, Lexington,
Massachusetts, 1976.
D. F. Jones and M. Tamiz, Goal Programming in the Period
19902000, in M. Ehrgott and X. Gandibleux, editors, Multiple Criteria
Optimization. State of the Art. Annotated Bibliographic Surveys, pp.
129170. Kluwer Academic Publishers, Boston/Dordrecht/London,
2002.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Attainment Method

In this approach, a vector of weights w1 , w2 , . . . , wk relating the relative
under- or over-attainment of the desired goals must be elicited from the
decision maker in addition to the goal vector b1 , b2 , . . . , bk for the objective
functions f1 , f2 , . . . , fk . To find the best-compromise solution x , the following
problem is solved [Gembikci, 1974; Gembicki, 1975]:

Minimize (27)

subject to:

gj (~x ) 0; j = 1, 2, . . . , m
bi + wi fi (~x ); i = 1, 2, . . . , k (28)

where is a scalar variable unrestricted in sign and the weights

w1 , w2 , . . . , wk are normalized so that
k
X
|wi | = 1 (29)
i=1

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Attainment Method

If some wi = 0 (i = 1, 2, . . . , k ), it means that the maximum limit
of objectives fi (~x ) is bi .

It can be easily shown [Chen, 1994] that a set of nondominated

solutions can be generated by varying the weights, with wi 0
(i = 1, 2, . . . , k) even for nonconvex problems.

It should be pointed out that the optimum value of informs the

DM of whether the goals are attainable or not. A negative value
of implies that the goal of the decision maker is attainable
and an improved solution is then to be obtained. Otherwise, if
> 0, then the DMs goal is unattainable.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Goal Attainment Method

For more information on the Goal Attainment Method, see:
F. W. Gembicki, Vector Optimization for Control with
Performance and Parameter Sensitivity Indices, PhD
thesis, Case Western Reserve University, Cleveland, Ohio,
1974.
F. W. Gembicki and Y. Y. Haimes, Approach to
performance and sensitivity multiobjective
optimization: the goal attainment method, IEEE
Transactions on Automatic Control, AC-15:591593, 1975.
Y. L. Chen and C. C. Liu, Multiobjective VAR planning
using the goal-attainment method, IEE Proceedings on
Generation, Transmission and Distribution,
141(3):227232, May 1994.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Lexicographic Ordering
This is a peculiar method in which the aggregations performed
are not scalar. In this method, the objectives are ranked in
order of importance by the decision maker (from best to worst).

The optimum solution ~x is then obtained by minimizing the

objective functions, starting with the most important one and
proceeding according to the order of importance of the
objectives.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Lexicographic Ordering
Let the subscripts of the objectives indicate not only the objective function
number, but also the priority of the objective. Thus, f1 (~x ) and fk (~x ) denote the
most and least important objective functions, respectively. Then the first
problem is formulated as
Minimize f1 (~x ) (30)
subject to
gj (~x ) 0; j = 1, 2, . . . , m (31)
and its solution ~x1 and f1 = f (~x1 ) is obtained. Then, the second problem is
formulated as
Minimize f2 (~x ) (32)
subject to

gj (~x ) 0; j = 1, 2, . . . , m (33)
f1 (~x ) = f1 (34)

and the solution of this problem is obtained as ~x2 and f2 = f2 (~x2 ).

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Lexicographic Ordering
This procedure is repeated until all k objectives have been
considered. The i th problem is given by

Minimize fi (~x ) (35)

subject to

gj (~x ) 0; j = 1, 2, . . . , m (36)
fl (~x ) = fl , l = 1, 2, . . . , i 1 (37)

The solution obtained at the end, i.e., ~xk is taken as the desired
solution ~x of the problem.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Lexicographic Ordering
More information on Lexicographic Ordering can be found at:
G. V. Sarma, L. Sellami, and K. D. Houam, Application of
Lexicographic Goal Programming in Production
PlanningTwo case studies, Opsearch, 30(2):141162,
1993.
S. S. Rao, Multiobjective Optimization in Structural
Design with Uncertain Parameters and Stochastic
Processes, AIAA Journal, 22(11):16701678, November
1984.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
The idea of stating the min-max optimum and applying it to
multiobjective optimization problems was taken from game
theory, which deals with solving conflicting situations. The
min-max approach to a linear model was proposed by Jutler
[1967] and Solich [1969].

It was further developed by Osyczka [1978], Rao [1986] and

Tseng & Lu [1990].

The min-max optimum compares relative deviations from the

separately attainable minima.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
Lets consider the i th objective function for which the relative
deviation can be calculated from

0 |fi (~x ) fi0 |

zi (~x ) = (38)
|fi0 |

or from
00 |fi (~x ) fi0 )|
zi (~x ) = (39)
|fi (~x )|
It should be clear that for equations (38) and (39) it is
necessary to assume that for every i I (I = 1, 2, . . . , k) and for
every ~x F, fi (~x ) 6= 0.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
If all the objective functions are going to be minimized, then
equation (38) defines function relative increments, whereas if
all of them are going to be maximized, it defines relative
decrements. Equation (39) works conversely.

Let ~z (~x ) = [z1 (~x ), . . . , zi (~x ), . . . , zk (~x )]T be a vector of the

relative increments which are defined in Rk .
The components of the vector z(~x ) are evaluated from the
formula
0 00
iI zi (~x ) = max {zi (~x ), zi (~x )} (40)

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
Now the min-max optimum can be defined as follows [Osyczka,
1984]:
A point ~x F is min-max optimal, if for every ~x F the
following recurrence formula is satisfied:
Step 1:
v1 (~x ) = min{zi (~x )} (41)
xF iI

and then Ii = {i1 }, where i1 is the index for which the value of
z1 (~x ) is maximal.

If there is a set of solutions X1 F which satisfies Step 1, then

we continue to Step 2.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
Step 2:
 
~
v2 (x ) = min max {zi (~x )} (42)
xX1 iI,i6I1

and then I2 = {i1 , i2 }, where i2 is the index for which the value
of zi (x) in this step is maximal.

If there is a set of solutions Xr 1 F which satisfies Step r 1,

then we continue to Step r .

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
Step r :
 
~
vr (x ) = min max {zi (~x )} (43)
xXr 1 iI,i6Ir 1

and then Ir = {Ir 1 , ir }, where ir is the index for which the value
of zi (~x ) in the r th step is maximal.

If there is a set of solutions Xk1 F which satisfies Step

k 1, then we continue to Step k .

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
Step k:

vk (~x ) = min zi (~x ) for i I and i 6 Ik1 (44)

~x Xk1

where v1 (~x ), . . . , vk (~x ) is the set of optimal values of

fractional deviations ordered non-increasingly.

This optimum can be described in words as follows. Knowing

the extremes of the objective functions which can be obtained
by solving the optimization problems for each criterion
separately, the desirable solution is the one which gives the
smallest values of the relative increments of all the objective
functions.

Carlos A. Coello Coello Multi-Objective Optimization

A Priori Preference Articulation

Min-Max Optimization
The point ~x F which satisfies the equations of Steps 1 and 2
may be called the best compromise solution considering all the
criteria simultaneously and on equal terms of importance.

It should be noticed that even when these equations look quite

complicated, in many optimization models, only the first step of
this process is necessary to determine the optimum.

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

These techniques do not require prior preference information

from the DM. Some of the techniques included in this category
are among the oldest multiobjective optimization approaches
proposed. The reason is that the main idea of these
approaches follows directly from the Kuhn-Tucker conditions for
noninferior solutions [Cohon and Marks, 1975].

The two most representative algorithms within this class are the
following:
Linear Combination of Weights
The -Constraint Method
Both of them will be briefly described next.

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

Linear Combination of Weights

Zadeh [1963] was the first to show that the third of the
Kuhn-Tucker conditions for noninferior solutions implies that
these noninferior solutions might be found by solving a scalar
optimization problem in which the objective function is a
weighted sum of the components of the original vector-valued
function.
Carlos A. Coello Coello Multi-Objective Optimization
A Posteriori Preference Articulation

Linear Combination of Weights

That is, the solution to the following problem is, in general,
noninferior:
Xk
min wi fi (~x ) (45)
i=1

subject to:
~x F (46)
where wi 0 for all i and is strictly positive for at least one
objective.

The noninferior set and the set of noninferior solutions can be

generated by parametrically varying the weights wi in the
objective function. This was initially demonstrated by Gass and
Saaty [1955] for a two-objective problem.

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

Linear Combination of Weights

Note that the weighting coefficients do not reflect proportionally
the relative importance of the objectives, but are only factors
which, when varied, locate points in the Pareto optimal set.

For the numerical methods that can be used to seek the

minimum of equation (45), this location depends not only on wi
values, but also on the units in which the functions are
expressed.

It is easy to prove that the use of a linear combination of

weights is unable to generate non-convex portions of a Pareto
front, regardless of the weight values that we use.

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

Linear Combination of Weights

This is an example of a case in which the true Pareto front
cannot be completely covered using a linear combination of
weights.

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

The -Constraint Method

This method also follows directly from the Kuhn-Tucker
conditions for noninferior solutions. The third Kuhn-Tucker
condition for optimality can be rewritten as:
k
X m
X
wr fr (~x ) + wl fl (~x ) i gi (~x ) = 0 (47)
l=1, l6=r i=1

Since only relative values of the weights are of significance, the

r th objective can be selected so that wr = 1. The previous
condition defined in equation (47) then becomes:
k
X m
X
fr (~x ) wl f (~x ) i gi (~x ) = 0 (48)
l=1, l6=r i=1

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

The -Constraint Method

This rewritten condition allows the second term to be
interpreted as a weighted sum of the gradients of k 1
lower-bound constraints, since there is a plus sign before the
summation. This interpretation implies that noninferior solutions
can be found by solving:

minfr (~x ) (49)

subject to:

fl (~x ) l for l = 1, 2, . . . , k and l 6= r (50)

where l are assumed values of the objective functions that

must not be exceeded.

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

The -Constraint Method

The idea of this method is to minimize one (the most preferred
or primary) objective function at a time, considering the other
objectives as constraints bound by some allowable levels l . By
varying these levels l , the noninferior solutions of the problem
can be obtained.

It is important to be aware of the fact that a preliminary analysis

is required to identify proper starting values for l . To get
adequate l values, single-objective optimizations are normally
carried out for each objective function in turn by using
mathematical programming techniques.

Carlos A. Coello Coello Multi-Objective Optimization

A Posteriori Preference Articulation

The -Constraint Method

This figure illustrates the way in which the -Constraint Method
works.

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

These techniques normally operate in three stages [Cohon and

Marks, 1975]:
1 Find a nondominated solution
2 Get the reaction of the DM regarding this nondominated
solution, and modify the preferences of the objectives
accordingly
3 Repeat the two previous steps until the DM is satisfied or
no further improvement is possible.

Next, we will briefly describe 3 methods that adopt progressive

preference articulation:
Probabilistic Trade-Off Development Method
STEP Method
Sequential Multiobjective Problem Solving Method
Carlos A. Coello Coello Multi-Objective Optimization
Progressive Preference Articulation

Probabilistic Trade-Off Development Method

The main motivation for the development of this method (also known as
PROTRADE) was to be able to handle risk in the development of the
objective trade-offs, and at the same time being able to accommodate the
preferences of the DM in a progressive manner [Goicochea, 1976].

In this case, it is assumed that our multiobjective optimization problem has a

probabilistic objective function and probabilistic constraints [Goicochea,
1977]. According to a 12-step algorithm, an initial solution is found using a
surrogate objective function, and then a multiattribute utility function is formed
leading to a new surrogate objective function and a new solution. The
solution is checked to see if it is satisfactory to the decision maker. The
process is repeated until a satisfactory solution is reached, as described in
[Goicochea, 1979; Goicochea, 1982].

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

Probabilistic Trade-Off Development Method

The results of the multiobjective optimization provide not only levels of
attainment of the objective function elements (as in the goal attainment
method [Gembicki, 1974]), but also the probabilities of reaching those levels.
The technique is interactive, which means that the DM formulates a
preference function in a progressive manner, after a trial process [Duckstein,
1984].

One interesting aspect of this approach is that the DM actually ranks

objectives in order of importance (a multi-attribute utility function is used to
assist the DM in the articulation of preferences) at the beginning of the
process, and later uses pairwise comparisons to reconcile these preferences
with the real (observed) behavior of the attributes. This allows not only an
interactive participation of the DM, but it also allows to gain knowledge about
the trade-offs of the problem.

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

Probabilistic Trade-Off Development Method

For more information about this method, see:
A. Goicoechea, L. Duckstein, and M. Fogel,
Multi-objective programming in watershed
management: A study of the Charleston watershed,
Water Resources Research, 12(6):10851092, December
1976.
A. Goicoechea, L. Duckstein, and M. Fogel, Multiple
objectives under uncertainty: An illustrative
application of PROTRADE, Water Resources Research,
15(2):203210, April 1979.
A. Goicoechea, D. R. Hansen, and L. Duckstein,
Multiobjective Analysis with Engineering and Business
Applications, John Wiley and Sons, New York, 1982.

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

STEP Method
This method (also known as STEM) is an iterative technique
based on the progressive articulation of preferences. The basic
idea is to converge toward the best solution in the min-max
sense, in no more than k steps, being k the number of
objectives. This technique, which is mostly useful for linear
problems, starts from an ideal point and proceeds in six steps.

One criticism to this technique is the fact that it assumes that a

best-compromise solution does not exist if it is not found after
executing all the k steps of its iterative process. This does not
give any clue to the DM of what to do [Cohon and Marks, 1975].

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

STEP Method
Another problem with the STEP Method is that it does not explicitly capture
the trade-offs between the objectives. The weights in no way reflect a value
judgment on the part of the DM. The weights are artificial quantities,
generated by the analyst to reflect deviations from an ideal solution, which is
itself an artificial quantity. This definition of the weights serves to obscure
rather than capture the normative nature of the multiobjective optimization
problems [Cohon and Marks, 1975].

For more details of this technique see:

R. Benayoun, J. Montgolfier, J. Tergny, and O. Laritchev, Linear
programming with multiple objective functions: Step Method
(STEM), Mathematical Programming, 1(3):366375, 1971.
J. L. Cohon and D. H. Marks, A Review and Evaluation of
Multiobjective Programming Techniques, Water Resources
Research, 11(2):208220, April 1975.

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

Sequential Multiobjective Problem Solving Method

This method (also known as SEMOPS) was proposed by
Monarchi et al. [1973] and it basically involves the DM in an
interactive fashion in the search for a satisfactory course of
action.

A surrogate objective function is used based on the goal and

aspiration levels of the DM. The goal levels are conditions
imposed on the DM by external forces, and the aspiration levels
are attainment levels of the objectives which the DM personally
desires to achieve. One would say, then, that goals do not
change once they are stated, but that the aspiration levels may
change during the iteration process.

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

Sequential Multiobjective Problem Solving Method

Operationally, SEMOPS is a three-step procedure involving
setup, iteration, and termination. Setup involves structuring a
principal problem and a set of auxiliary problems with a
surrogate objective function.

The iteration step involves cycling between an optimization

phase (by the analyst), and an evaluation phase (by the DM)
until a satisfactory solution is reached, if it exists.

The procedure terminates when either a satisfactory solution is

found, or the DM concludes that none of the nondominated
solutions obtained are satisfactory and gives up in the search.

Carlos A. Coello Coello Multi-Objective Optimization

Progressive Preference Articulation

Sequential Multiobjective Problem Solving Method

More information on this method can be found in:
D. E. Monarchi, C. C. Kisiel, and L. Duckstein, Interactive
multiobjective programming in water resources: a
case study, Water Resources Research, 9(4):837850,
August 1973.
A. Goicoechea, D. R. Hansen, and L. Duckstein,
Multiobjective Analysis with Engineering and Business
Applications, John Wiley and Sons, New York, 1982.

Carlos A. Coello Coello Multi-Objective Optimization

Current State

Currently, there are some 30 mathematical programming

techniques for nonlinear multi-objective optimization. However,
they have several limitations. For example, some of them
require that the objectives (and the constraints) are
differentiable. Other approaches cannot be applied to
disconnected or to non-convex Pareto fronts. Additionally, most
of them generate a single solution per algorithmic execution.

Carlos A. Coello Coello Multi-Objective Optimization

Current State

This has motivated the use of metaheuristics (particularly,

bio-inspired metaheuristics).

A metaheuristic is a high-level search procedure that applies

some form of rule or set of rules based on some source of
knowledge, in order to explore the search space in a more
efficient way.

Carlos A. Coello Coello Multi-Objective Optimization

Current State

From the many metaheuristics currently available, one particular class has
become very popular in the last 20 years: bio-inspired metaheuristics.

Bio-inspired metaheuristics use rules that are inspired on some biological

metaphore (e.g., in the case of evolutionary algorithms, the inspiration is
Darwins survival of the fittest principle). Most bio-inspired metheuristics are
stochastic search techniques (e.g., evolutionary algorithms, particle swarm
optimization, ant colony optimization, etc.).

Carlos A. Coello Coello Multi-Objective Optimization

Current State

Most bio-inspired metaheuristics operate on a set of solutions (normally

called population) at each iteration. A clever use of this population in
multi-objective optimization, allows the generation of several elements of the
Pareto optimal set in a single algorithmic execution.

Also, bio-inspired metaheuristics require little information about the domain

(e.g., they dont require derivatives) and are less susceptible to the shape or
continuity of the Pareto front.

Carlos A. Coello Coello Multi-Objective Optimization

Current State

In spite of their several advantages, bio-inspired metaheuristics also have

some disadvantages. One of them is that they cannot guarantee convergence
to the true Pareto front of a problem in most practical cases. Another one
(which is more relevant in practical applications) is that their computational
cost is normally significantly higher than that of mathematical programming
techniques. This is due to their stochastic nature, which requires sampling
several solutions to find an appropriate search direction. This may be
unaffordable in some applications (e.g., in aeronautical engineering).

Carlos A. Coello Coello Multi-Objective Optimization