Coello Coello - Talk Lisbon 2019

Outline
Introduction
Historical Highlights
Some Applications
Little Known Approaches
The Algorithms that the World Forgot
Current State and Future Challenges
Evolutionary Multi-Objective Optimization:

Past, Present and Future
Carlos A. Coello Coello
ccoello@cs.cinvestav.mx
CINVESTAV-IPN
Evolutionary Computation Group (EVOCINV)
Computer Science Department
Av. IPN No. 2508, Col. San Pedro Zacatenco
México, D.F. 07360, MEXICO
Lisbon, Portugal, March 2019

Carlos A. Coello Coello EMOO: Past, Present and Future
Outline
Introduction
Some Applications
Outline of Topics
1 Introduction
2 Historical Highlights
3 Some Applications
4 Little Known Approaches
5 The Algorithms that the World Forgot
6 Current State and Future Challenges

Outline
Introduction
Some Applications
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.
Outline
Introduction
Some Applications
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)

Outline
Introduction
Some Applications
Notion of Optimality in MOPs
Having several objective functions, the notion of “optimum”

changes, because in MOPs, we are really trying 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.
Outline
Introduction
Some Applications
Notion of Optimality in MOPs
This notion was later generalized by Vilfredo Pareto (in 1896).

Although some authors call Edgeworth-Pareto optimum to this
notion, we will use the most commonly accepted term: Pareto
optimum.

Outline
Introduction
Some Applications
Pareto Optimality
Definition
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.

Outline
Introduction
Some Applications
Pareto Optimality
Explanation of the Definition

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. Unfortunately, this
concept almost always gives not a single solution, but rather 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 plot of the objective functions
whose nondominated vectors are in the Pareto optimal set is
called the Pareto front.

Outline
Introduction
Some Applications
Mathematical Programming Techniques
Currently, there are over 30 mathematical programming

techniques for multiobjective optimization. However, these
techniques tend to generate elements of the Pareto optimal set
one at a time. Additionally, most of them are very sensitive to
the shape of the Pareto front (e.g., they do not work when the
Pareto front is concave or when the front is disconnected).
Outline
Introduction
Some Applications
Evolutionary Algorithms
The idea of using techniques based on the emulation of the

mechanism of natural selection to solve problems can be
traced as long back as the 1930s. However, it was not until the
1960s that the three main techniques based on this notion were
developed: genetic algorithms [Holland, 1962], evolution
strategies [Schwefel, 1965] and evolutionary programming
[Fogel, 1966]. These approaches are now collectively
denominated evolutionary algorithms.
Outline
Introduction
Some Applications
Evolutionary Algorithms
Evolutionary algorithms seem particularly suitable to solve

multiobjective optimization problems, because they deal
simultaneously with a set of possible solutions (the so-called
population). This allows us to find several members of the
Pareto optimal set in a single run of the algorithm, instead of
having to perform a series of separate runs as in the case of
the traditional mathematical programming techniques.
Additionally, evolutionary algorithms are less susceptible to the
shape or continuity of the Pareto front (e.g., they can easily
deal with discontinuous or concave Pareto fronts), whereas
these two issues are a real concern for mathematical
programming techniques.
Outline
Introduction
Some Applications
The potential of evolutionary algorithms in multiobjective

optimization was hinted by Richard S. Rosenberg in his PhD
thesis from 1967, which is entitled “Simulation of genetic
populations with biochemical properties”.

Outline
Introduction
Some Applications
However, the first actual implementation of a multi-objective

evolutionary algorithm (MOEA) was John David Schaffer’s
Vector Evaluated Genetic Algorithm (VEGA), which dates back
to 1984. This was a naive multi-objective evolutionary
algorithm, whose selection mechanism did not rely on Pareto
optimality.

Outline
Introduction
Some Applications
In the old days, two types of approaches were normally

adopted with evolutionary algorithms:
1. Aggregating functions: They basically transform a
multi-objective optimization problem into a scalar
optimization problem. For example, aPlinear aggregating
function normally has the form: min ki=1 wi fi (~x ) where
wi ≥ 0 are the weighting coefficients representing the
relative importance of the k objective functions of our
problem.

Outline
Introduction
Some Applications
In fact, linear aggregating approaches are the oldest mathematical

programming methods proposed to solve multi-objective optimization
problems, since they can be derived from the (famous) Kuhn-Tucker
conditions for nondominated solutions. Linear aggregating functions
are considered “evil” by most EMO researchers because of their
limitations (they cannot properly generate nonconvex portion of the
Pareto front). Note however, that nonlinear aggregating functions
do not have this limitation.
Outline
Introduction
Some Applications
2. Lexicographic ordering: In this method, the user is asked

to rank the objectives in order of importance. The optimum
solution is then obtained by minimizing the objective
functions, starting with the most important one and
proceeding according to the assigned order of importance
of the objectives.

Outline
Introduction
Some Applications
David Goldberg’s seminal book on genetic algorithms (published in

1989) introduced the notion of Pareto ranking: individuals in a MOEA
must be selected based on Pareto dominance, such that all
nondominated individuals are considered equally good among
themselves. He also pointed out the importance of maintaining
diversity as to allow the generation of several (different)
nondominated solutions in a single run. Fitness sharing was
proposed for that sake.
Outline
Introduction
Some Applications
The basic expression adopted in fitness sharing is the following:
( α
d
1 − σshij , dij < σshare
φ(dij ) = (4)
0, otherwise
where α = 1, dij indicates the distance between solutions i and j, and
σshare is the niche radius (or sharing threshold). By using this
parameter, the fitness of the individual i is modified as:
fi
fsi = PM (5)
j=1 φ(dij )
where M is the number of individuals that are located in the
neighborhood of the i-th individual.
Outline
Introduction
Some Applications
Several other approaches have been proposed to handle

diversity:
Clustering
Use of entropy
Adaptive grids
Crowding
Performance indicators
Parallel coordinates

Outline
Introduction
Some Applications
Since Goldberg’s early proposal, MOEAs consist of two basic

components:
A selection mechanism that normally (but not necessarily)
incorporates Pareto optimality.
A density estimator, which is responsible for maintaining
diversity, and therefore, keeping the MOEA from
converging to a single solution.

Outline
Introduction
Some Applications
Carlos M. Fonseca and Peter J. Fleming proposed the

Multiobjective Optimization Genetic Algorithm (MOGA) in 1993.
MOGA would remain as one of the most popular (and effective)
multi-objective evolutionary algorithms during many years.
MOGA uses a selection operator based on Pareto ranking and
fitness sharing to maintain diversity.
Outline
Introduction
Some Applications
Jeffrey Horn proposed the Niched Pareto Genetic Algorithm

(NPGA) in 1994. The NPGA uses a tournament selection
based on Pareto ranking and fitness sharing to maintain
diversity. However, not all the population needs to be ranked,
since only a sample is used each time. The NPGA is fast and
very competitive.
Outline
Introduction
Some Applications
Kalyanmoy Deb proposed the Nondominated Sorting Genetic

Algorithm (NSGA) in 1994. The NSGA adopts a Pareto ranking
selection based on layers of dominance and dummy fitnesses.
Fitness sharing is adopted to maintain diversity. The NSGA
was slow and not very effective.
Outline
Introduction
Some Applications
In 1998, Eckart Zitzler introduces the Strength Pareto

Evolutionary Algorithm (SPEA) at a conference. The next year,
this approach is published in the IEEE Transactions on
Evolutionary Computation. SPEA uses an external archive to
retain the nondominated solutions found during the search.
This notion of elitism would soon become popular. A new wave
of algorithms was about to arrive.
Outline
Introduction
Some Applications
In 1999, Joshua D. Knowles introduces the Pareto Archive

Evolution Strategy (PAES). The next year, this approach is
published in the Evolutionary Computation journal. PAES uses
a (1+1) evolution strategy combined with a very clever external
archive which is responsible of both storing nondominated
solutions and distributing them in a uniform way (in objective
function space).
Outline
Introduction
Some Applications

Outline
Introduction
Some Applications
In 1999, Deb published in the Evolutionary Computation journal

a methodology to build multiobjective test problems that soon
became very popular.

Outline
Introduction
Some Applications
In 2000, Kalyanmoy Deb and his students introduced the

Nondominated Sorting Genetic Algorithm-II (NSGA-II). Two
years later, this approach is published in the IEEE Transactions
on Evolutionary Computation. The NSGA-II is very fast and
effective. It uses a crowded-comparison operator which takes
into consideration both the nondomination rank of an individual
and its crowding distance. It also uses a plus selection (i.e.,
parents and children are merged together and the best
individuals from their union are selected). For more than 10
years, NSGA-II was considered the reference approach with
respect to which other MOEAs had to be compared.

Outline
Introduction
Some Applications
In 2001, Zitzler and his colleagues introduced SPEA2, which

has 3 main differences with respect to its predecessor: (1) it
incorporates a fine-grained fitness assignment strategy which
takes into account for each individual the number of individuals
that dominate it and the number of individuals by which it is
dominated; (2) it uses a nearest neighbor density estimation
technique which guides the search more efficiently, and (3) it
has an enhanced archive truncation method that guarantees
the preservation of boundary solutions.
Outline
Introduction
Some Applications
In 2001, Coello Coello and Toscano-Pulido introduced the first

microgenetic algorithm for multi-objective optimization. The
microGA uses a population of only 4 individuals combined with
3 forms of elitism. It is a very fast algorithm and highly
competitive. Its main defect was its high number of parameters
(eight).
Outline
Introduction
Some Applications
In 2002, Zitzler and his colleagues published a conference

paper in which they indicated the limitations of many of the
performance measures (“metrics”) normally adopted to validate
multi-objective evolutionary algorithms in a quantitative way
(many of them are not Pareto compliant). An extended version
of this paper was published in 2003 in the IEEE Transactions on
Evolutionary Computation.
Outline
Introduction
Some Applications
In 2002, Deb and his colleagues introduce a new set of test

functions that are scalable and that have true Pareto fronts that
can be obtained analytically. This set is named DTLZ after the
first letter of the last names of its creators.
Outline
Introduction
Some Applications
In 2002, Laumanns and his colleagues introduced a relaxed

form of Pareto dominance called -dominance as an archiving
technique. -dominance allows to control the granularity of the
approximation of the Pareto front obtained. As a consequence,
it is possible to accelerate convergence using this mechanism
(if we are satisfied with a very coarse approximation of the
Pareto front).
Outline
Introduction
Some Applications
In 2003, Toscano Pulido and Coello Coello introduced the

microgenetic algorithm 2, which was the first attempt to
produce a parameterless MOEA. The µGA2 only requires one
parameter: the size of the external archive (i.e., how many
nondominated solutions are required). The algorithm is even
able to stop by itself (by adopting a scheme to detect
convergence).
Outline
Introduction
Some Applications
In 2004, Zitzler introduces the Indicator-Based Evolutionary

Algorithm (IBEA) in which he proposes to use a selection
mechanism based on a performance indicator. A few years
later, many people would adopt this concept, using the
hypervolume (the hypervolume of a set of solutions measures
the size of the portion of objective space that is dominated by
those solutions collectively).
Outline
Introduction
Some Applications
Many-Objective Optimization
In 2005, research on many-objective optimization (problems
having 4 or more objectives) flourishes. Proposals to deal with
these problems include: dimensionality reduction, relaxed
forms of Pareto dominance and alternative ranking schemes.
This remains as a hot research topic in spite of the fact that

some people have questioned if scalability in objective function
space is really a source of difficulty.

Outline
Introduction
Some Applications
In 2005, a new procedure to design test problems was

introduced by Huband et al. This set is known as WFG
(Walking-Fish Group) and is very popular today.

Outline
Introduction
Some Applications
In 2005, the first version of the S Metric Selection Evolutionary Multiobjective

Algorithm (SMS-EMOA) SMS-EMOA is published in a conference. The
journal version of this algorithm was published in 2007. This MOEA is based
on the hypervolume performance measure. This initiated a new research
trend, due to the limitations of Pareto dominance in the presence of a large
number of objectives.

Outline
Introduction
Some Applications
In 2005, Joshua D. Knowles and Evan J. Hughes introduced

parEGO, which is a MOEA that only performs 250 fitness
function evaluations. The approach, however, could only be
used with problems of low dimensionality.

Outline
Introduction
Some Applications
In 2006, Zhang and Li proposed a MOEA based on

decomposition (MOEA/D). This approach decomposes a
multi-objective optimization problem into a number of scalar
optimization subproblems and optimizes them simultaneously.

Outline
Introduction
Some Applications
In 2014, the NSGA-III is published in a two-parts paper that

appeared in the IEEE Transactions on Evolutionary
Computation. Its main aim is to deal with many-objective
optimization problems. The NSGA-III uses reference points and
adopts a decomposition approach.
Outline
Introduction
Some Applications
Some Sample Applications
Chung et al. (2003,2004) used a variation of the micro-GA

hybridized with a gradient method for the multi-objective design
low-boom supersonic business jets.

Outline
Introduction
Some Applications
Benedetti et al. (2006) used the NSGA-II with k optimality to

design the tire-suspension system of a racing car. The problem
has 24 objectives, from which 18 are in conflict with each other.
Some designs that outperformed the reference car design were
obtained.
Outline
Introduction
Some Applications
Shimoyama (2006) adopted a multiobjective genetic algorithm

coupled to a robust design optimization approach called
“design for six sigma”, to design a Mars exploratory airplane.
Outline
Introduction
Some Applications
Rivas-Dávalos and Irving (2005) adopted SPEA2 for power

distribution system planning. Two objectives were minimized:
economical cost and energy non-supplied.

Outline
Introduction
Some Applications
Little Known Multi-Objective Evolutionary Algorithms
There have been many multi-objective evolutionary algorithms

that have been pretty much ignored:
The Nash Genetic Algorithm (Sefrioui, 1996).
The Maximin Fitness Function (Balling, 2000).
The Incrementing Multi-Objective Evolutionary Algorithm
(IMOEA) (Tan et al., 2001).
The Constraint Method-Based Evolutionary Algorithm
(CMEA) for multi-objective optimization (Ranjithan et al.,
2001).
The Orthogonal Multi-Objective Evolutionary Algorithm
(OMOEA) (Zeng et al., 2004).
Outline
Introduction
Some Applications
Number of papers published per year (early 2019)

Outline
Introduction
Some Applications
Research representative of the current trends
New indicator-based MOEAs (based on hypervolume,

IGD+, Deltap , R2, etc.).
Use of surrogate and parallel methods to deal with
computationally expensive objective functions.
Methods to generate weights for dealing with inverted test
problems using decomposition-based MOEAs.
New applications in a wide variety of domains (e.g.,
software engineering, architecture, robotics, etc.)

Outline
Introduction
Some Applications
Future Challenges
Schemes for dealing with large scale MOEAs (e.g., with

10,000 decision variables)
Efficient and effective hybridization schemes that combine
MOEAs with mathematical programming techniques.
New benchmarks that better represent the complexities of
real-world problems.
Challenging applications (e.g., preliminary design, in which
we aim to find quickly good trade-offs rather than optimal
solutions).
More theoretical work is needed (e.g., there is almost no
theoretical work on parallel MOEAs).
Outline
Introduction
Some Applications
Some Criticism
A lot of “research by analogy” that lacks scientific value

and relevance now appears even in top journals.
“New” research topics that are at the very least
questionable (e.g., multimodal multi-objective optimization).
Ironically, it is now more difficult to publish ideas that are
really new, since, for many reviewers, validation
procedures are now more important than the scientific
contributions of a paper.

Outline
Introduction
Some Applications
To know more about evolutionary

multi-objective optimization
Please visit the new webpage of the EMOO repository:
https://emoo.cs.cinvestav.mx/
Outline
Introduction
Some Applications
To know more about evolutionary

multi-objective optimization
The EMOO repository currently contains:

Over 12,000 bibliographic references including 300 PhD
theses, 52 Masters theses, over 5770 journal papers and
over 4380 conference papers.
Contact information of 79 EMOO researchers.
Public domain implementations of SPEA, NSGA, NSGA-II,
the microGA, MOGA, -MOEA, MOPSO and PAES, among
others.

Coello Coello - Talk Lisbon 2019

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Coello Coello - Talk Lisbon 2019

Uploaded by

Copyright:

Available Formats

Outline

Evolutionary Multi-Objective Optimization:

Carlos A. Coello Coello

Lisbon, Portugal, March 2019

4 Little Known Approaches

5 The Algorithms that the World Forgot

6 Current State and Future Challenges

Carlos A. Coello Coello EMOO: Past, Present and Future

Most problems in nature have several (possibly conflicting)

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

Carlos A. Coello Coello EMOO: Past, Present and Future

Notion of Optimality in MOPs

Having several objective functions, the notion of “optimum”

Notion of Optimality in MOPs

This notion was later generalized by Vilfredo Pareto (in 1896).

Carlos A. Coello Coello EMOO: Past, Present and Future

Carlos A. Coello Coello EMOO: Past, Present and Future

Explanation of the Definition

Carlos A. Coello Coello EMOO: Past, Present and Future

Mathematical Programming Techniques

Currently, there are over 30 mathematical programming

The idea of using techniques based on the emulation of the

Evolutionary algorithms seem particularly suitable to solve

The potential of evolutionary algorithms in multiobjective

Carlos A. Coello Coello EMOO: Past, Present and Future

However, the first actual implementation of a multi-objective

Carlos A. Coello Coello EMOO: Past, Present and Future

In the old days, two types of approaches were normally

Carlos A. Coello Coello EMOO: Past, Present and Future

In fact, linear aggregating approaches are the oldest mathematical

2. Lexicographic ordering: In this method, the user is asked

Carlos A. Coello Coello EMOO: Past, Present and Future

David Goldberg’s seminal book on genetic algorithms (published in

Several other approaches have been proposed to handle

Carlos A. Coello Coello EMOO: Past, Present and Future

Since Goldberg’s early proposal, MOEAs consist of two basic

Carlos A. Coello Coello EMOO: Past, Present and Future

Carlos M. Fonseca and Peter J. Fleming proposed the

Jeffrey Horn proposed the Niched Pareto Genetic Algorithm

Kalyanmoy Deb proposed the Nondominated Sorting Genetic

In 1998, Eckart Zitzler introduces the Strength Pareto

In 1999, Joshua D. Knowles introduces the Pareto Archive

Carlos A. Coello Coello EMOO: Past, Present and Future

In 1999, Deb published in the Evolutionary Computation journal

Carlos A. Coello Coello EMOO: Past, Present and Future

In 2000, Kalyanmoy Deb and his students introduced the

Carlos A. Coello Coello EMOO: Past, Present and Future

In 2001, Zitzler and his colleagues introduced SPEA2, which

In 2001, Coello Coello and Toscano-Pulido introduced the first

In 2002, Zitzler and his colleagues published a conference

In 2002, Deb and his colleagues introduce a new set of test

In 2002, Laumanns and his colleagues introduced a relaxed

In 2003, Toscano Pulido and Coello Coello introduced the

In 2004, Zitzler introduces the Indicator-Based Evolutionary

This remains as a hot research topic in spite of the fact that

Carlos A. Coello Coello EMOO: Past, Present and Future

In 2005, a new procedure to design test problems was

Carlos A. Coello Coello EMOO: Past, Present and Future

In 2005, the first version of the S Metric Selection Evolutionary Multiobjective

Carlos A. Coello Coello EMOO: Past, Present and Future

In 2005, Joshua D. Knowles and Evan J. Hughes introduced

Carlos A. Coello Coello EMOO: Past, Present and Future