1

Multi-Objective Optimization Using Genetic Algorithms: A Tutorial

Abdullah Konak
1
, David W. Coit
2
, Alice E. Smith
3

1
Information Sciences and Technology, Penn State Berks-Lehigh Valley
2
Department of Industrial and Systems Engineering, Rutgers University
3
Department of Industrial and Systems Engineering, Auburn University

abstract Multi-objective formulations are a realistic models for many complex
engineering optimization problems. Customized genetic algorithms have been
demonstrated to be particularly effective to determine excellent solutions to these
problems. In many real-life problems, objectives under consideration conflict with
each other, and optimizing a particular solution with respect to a single objective
can result in unacceptable results with respect to the other objectives. A
reasonable solution to a multi-objective problem is to investigate a set of
solutions, each of which satisfies the objectives at an acceptable level without
being dominated by any other solution. In this paper, an overview and tutorial is
presented describing genetic algorithms developed specifically for these problems
with multiple objectives. They differ from traditional genetic algorithms by using
specialized fitness functions, introducing methods to promote solution diversity,
and other approaches.
1. Introduction
The objective of this paper is present an overview and tutorial of multiple-objective
optimization methods using genetic algorithms (GA). For multiple-objective problems, the
objectives are generally conflicting, preventing simultaneous optimization of each objective.
Many, or even most, real engineering problems actually do have multiple-objectives, i.e.,
minimize cost, maximize performance, maximize reliability, etc. These are difficult but realistic
problems. GA are a popular meta-heuristic that is particularly well-suited for this class of
problems. Traditional GA are customized to accommodate multi-objective problems by using
specialized fitness functions, introducing methods to promote solution diversity, and other
approaches.
There are two general approaches to multiple-objective optimization. One is to combine
the individual objective functions into a single composite function. Determination of a single
2
objective is possible with methods such as utility theory, weighted sum method, etc., but the
problem lies in the correct selection of the weights or utility functions to characterize the
decision-makers preferences. In practice, it can be very difficult to precisely and accurately
select these weights, even for someone very familiar with the problem domain. Unfortunately,
small perturbations in the weights can lead to very different solutions. For this reason and others,
decision-makers often prefer a set of promising solutions given the multiple objectives.
The second general approach is to determine an entire Pareto optimal solution set or a
representative subset. A Pareto optimal set is a set of solutions that are nondominated with
respect to each other. While moving from one Pareto solution to another, there is always a
certain amount of sacrifice in one objective to achieve a certain amount of gain in the other.
Pareto optimal solution sets are often preferred to single solutions because they can be practical
when considering real-life problems, since the final solution of the decision maker is always a
trade-off between crucial parameters. Pareto optimal sets can be of varied sizes, but the size of
the Pareto set increases with the increase in the number of objectives.
2. Multi-Objective Optimization Formulation
A multi-objective decision problem is defined as follows: Given an n-dimensional
decision variable vector x={x
1
,,x
n
} in the solution space X, find a vector x
*
that minimizes a
given set of K objective functions z(x
*
)={z
1
(x
*
),,z
K
(x
*
)}. The solution space X is generally
restricted by a series of constraints, such as g
j
(x
*
)=b
j
for j = 1, , m, and bounds on the decision
variables.
In many real-life problems, objectives under consideration conflict with each other.
Hence, optimizing x with respect to a single objective often results in unacceptable results with
respect to the other objectives. Therefore, a perfect multi-objective solution that simultaneously
optimizes each objective function is almost impossible. A reasonable solution to a multi-
objective problem is to investigate a set of solutions, each of which satisfies the objectives at an
acceptable level without being dominated by any other solution.
If all objective functions are for minimization, a feasible solution x is said to dominate
another feasible solution y ( x y ), if and only if, z
i
(x) z
i
(y) for i=1, , K and z
j
(x) < z
j
(y) for
least one objective function j. A solution is said to be Pareto optimal if it is not dominated by
any other solution in the solution space. A Pareto optimal solution cannot be improved with
respect to any objective without worsening at least one other objective. The set of all feasible
3
non-dominated solutions in X is referred to as the Pareto optimal set, and for a given Pareto
optimal set, the corresponding objective function values in the objective space is called the
Pareto front. For many problems, the number of Pareto optimal solutions is enormous (maybe
infinite).
The ultimate goal of a multi-objective optimization algorithm is to identify solutions in
the Pareto optimal set. However, identifying the entire Pareto optimal set, for many multi-
objective problems, is practically impossible due to its size. In addition, for many problems,
especially for combinatorial optimization problems, proof of solution optimality is
computationally infeasible. Therefore, a practical approach to multi-objective optimization is to
investigate a set of solutions (the best-known Pareto set) that represent the Pareto optimal set as
much as possible. With these concerns in mind, a multi-objective optimization approach should
achieve the following three conflicting goals:
1. The best-known Pareto front should be as close possible as to the true Pareto front. Ideally,
the best-known Pareto set should be a subset of the Pareto optimal set.
2. Solutions in the best-known Pareto set should be uniformly distributed and diverse over of
the Pareto front in order to provide the decision maker a true picture of trade-offs.
3. In addition, the best-known Pareto front should capture the whole spectrum of the Pareto
front. This requires investigating solutions at the extreme ends of the objective function
space.
This paper presents common approaches used in multi-objective genetic algorithms to
attain these three conflicting goals while solving a multi-objective optimization problem.
3. Genetic Algorithms
The concept of genetic algorithms (GA) was developed by Holland and his colleagues in
the 1960s and 1970s [18]. GA is inspired by the evolutionist theory explaining the origin of
species. In nature, weak and unfit species within their environment are faced with extinction by
natural selection. The strong ones have greater opportunity to pass their genes to future
generations via reproduction. In the long run, species carrying the correct combination in their
genes become dominant in their population. Sometimes, during the slow process of evolution,
random changes may occur in genes. If these changes provide additional advantages in the
challenge for survival, new species evolve from the old ones. Unsuccessful changes are
eliminated by natural selection.
4
In GA terminology, a solution vector xX is called an individual or a chromosome.
Chromosomes are made of discrete units called genes. Each gene controls one or more features
of the chromosome. In the original implementation of GA by Holland, genes are assumed to be
binary numbers. In later implementations, more varied gene types have been introduced.
Normally, a chromosome corresponds to a unique solution x in the solution space. This requires
a mapping mechanism between the solution space and the chromosomes. This mapping is called
an encoding. In fact, GA works on the encoding of a problem, not on the problem itself.
GA operates with a collection of chromosomes, called a population. The population is
normally randomly initialized. As the search evolves, the population includes fitter and fitter
solutions, and eventually it converges, meaning that it is dominated by a single solution. Holland
also presented a proof of convergence (the schema theorem) to the global optimum where
chromosomes are binary vectors.
GA use two operators to generate new solutions from existing ones: crossover and
mutation. The crossover operator is the most important operator of GA. In crossover, generally
two chromosomes, called parents, are combined together to form new chromosomes, called
offspring. The parents are selected among existing chromosomes in the population with
preference towards fitness so that offspring is expected to inherit good genes which make the
parents fitter. By iteratively applying the crossover operator, genes of good chromosomes are
expected to appear more frequently in the population, eventually leading to convergence to an
overall good solution.
The mutation operator introduces random changes into characteristics of chromosomes.
Mutation is generally applied at the gene level. In typical GA implementations, the mutation rate
(probability of changing the properties of a gene) is very small, typically less than 1%.
Therefore, the new chromosome produced by mutation will not be very different from the
original one. Mutation plays a critical role in GA. As discussed earlier, crossover leads the
population to converge by making the chromosomes in the population alike. Mutation
reintroduces genetic diversity back into the population and assists the search escape from local
optima.
Reproduction involves selection of chromosomes for the next generation. In the most
general case, the fitness of an individual determines the probability of its survival for the next
generation. There are different selection procedures in GA depending on how the fitness values
5
are used. Proportional selection, ranking, and tournament selection are the most popular
selection procedures. The procedure of a generic GA is given as follows:
Step 1. Set t =1. Randomly generate N solutions to form the first population, P
1
. Evaluate
the fitness of solutions in P
1
.
Step 2. Crossover: Generate an offspring population Q
t
as follows.
2.1. Choose two solutions x and y from P
t
based on the fitness values.
2.2. Using a crossover operator, generate offspring and add them to Q
t
.
Step 3. Mutation: Mutate each solution xQ
t
with a predefined mutation rate.
Step 4. Fitness Assignment: Evaluate and assign a fitness value to each solution xQ
t

based its objective function value and infeasibility.
Step 5. Selection: Select N solutions from Q
t
based on their fitness and assigned them
P
t+1
.
Step 6. If the stopping criterion is satisfied, terminate the search and return the current
population, else, set t=t+1 go to Step 2.
4. Multi-objective Genetic Algorithms
Being a population based approach, GA are well suited to solve multi-objective
optimization problems. A generic single-objective GA can be easily modified to find a set of
multiple non-dominated solutions in a single run. The ability of GA to simultaneously search
different regions of a solution space makes it possible to find a diverse set of solutions for
difficult problems with non-convex, discontinuous, and multi-modal solutions spaces. The
crossover operator of GA may exploit structures of good solutions with respect to different
objectives to create new non-dominated solutions in unexplored parts of the Pareto front. In
addition, most multi-objective GA do not require the user to prioritize, scale, or weigh
objectives. Therefore, GA has been the most popular heuristic approach to multi-objective design
and optimization problems. Jones et al. [25] reported that 90% of the approaches to multi-
objective optimization aimed to approximate the true Pareto front for the underlying problem. A
majority of these used a meta-heuristic technique, and 70% of all meta-heuristics approaches
were based on evolutionary approaches.
The first multi-objective GA, called Vector Evaluated Genetic Algorithms (or VEGA),
was proposed by Schaffer [44]. Afterward, several major multi-objective evolutionary algorithms
were developed such as Multi-objective Genetic Algorithm (MOGA) [13], Niched Pareto
6
Genetic Algorithm [19], Random Weighted Genetic Algorithm (RWGA)[39], Nondominated
Sorting Genetic Algorithm (NSGA) [45], Strength Pareto Evolutionary Algorithm (SPEA) [55],
Pareto-Archived Evolution Strategy (PAES) [27], Fast Non-dominated Sorting Genetic
Algorithm (NSGA-II) [9], Multi-objective Evolutionary Algorithm (MEA) [42], Rank-Density
Based Genetic Algorithm (RDGA) [32]. Note that although there are many variations of multi-
objective GA in the literature, these cited GA are well-known and credible algorithms that have
been used in many applications and their performances were tested in several comparative
studies.
Several survey papers [1-3, 12, 14, 22, 51, 54, 55] have been published on evolutionary
multi-objective optimization. Coello Coello lists more than 1800 references in his website [4].
Most survey papers on multi-objective evolutionary approaches introduce and compare different
algorithms. This paper takes a different course and focuses on important issues while designing a
multi-objective GA and describes common techniques used in multi-objective GA to attain the
three goals in multi-objective optimization.
4.1. Fitness Functions
4.1.1. Weighted Sum Approaches.
The classical approach to solve a multi-objective optimization problem is to assign a
weight w
i
to each normalized objective function ( )
i
z x so that the problem is converted to a
single objective problem with a scalar objective function as follows:

1 1 2 2
min ( ) ( ) ( )
k k
z w z w z w z = + + + x x x (1)

where ( )
i
z x is the normalized objective function, ( )
i
z x and 1
i
w =

. This approach is called a

priori approach since the user is expected to provide the weights. Solving a problem with the
objective function (1) for a given weight vector
1 2
{ , , , }
k
w w w = w yields a single solution, and
if multiple solutions are desired, the problem should be solved multiple times with different
weight combinations. The main difficulty with this approach is selecting a weight vector for each
run. To automate this process, Hajela and Lin [17] proposed the weight-based genetic algorithm
for multi-objective optimization (WBGA-MO). In the WBGA-MO, each solution x
i
in the
population uses a different weight vector
1 2
{ , , , }
k
w w w =
i
w in the calculation of objective
function (1). The weight vector
i
w is embedded within the chromosome of solution x
i
.
7
Therefore, multiple solutions can be simultaneously searched in a single run. In addition, weight
vectors can be adjusted to promote diversity of the population.
Other researchers [39, 40] have proposed a multi-objective genetic algorithm based on a
weighted sum of multiple objective functions where a normalized weight vector w
i
is randomly
generated for each solution x
i
during the selection phase at each generation. This approach aims
to stipulate multiple search directions in a single run without using any additional parameters.
The main advantage of the weighted sum approach is a straightforward implementation.
Since a single objective is used in fitness assignment, a single objective GA can be used with
minimum modifications. In addition, this approach is computationally very efficient. The main
disadvantage of this approach is that not all Pareto-optimal solutions can be investigated when
the true Pareto front is non-convex. Therefore, the multi-objective genetic algorithms based on
the weighed sum approach have difficulty in finding solutions uniformly distributed over a non-
convex trade-off surface [54].
4.1.2. Altering Objective Functions.
As mentioned earlier, the VEGA [44] is the first GA used to approximate the Pareto
optimal set by a set of non-dominated solutions. In the VEGA, population P
t
is randomly divided
into K equal sized sub-populations; P
1
, P
2
, ..., P
K
. Then, each solution in subpopulation P
i
is
assigned a fitness value based on objective function z
i
. Solutions are selected from these
subpopulations using proportional selection for crossover and mutation. Crossover and mutation
are performed on the new population in the same way with the single objective GA. A similar
approach is to use only a single objective function which is randomly determined each time in
the selection phase [31].
These approaches are easy to implement and computationally as efficient as a single-
objective GA. The major drawback of objective switching is that the population tends to
converge to solutions which are very superior in one objective, but very poor at others.
4.1.3. Pareto-Ranking Approaches.
Pareto-ranking approaches explicitly utilize the concept of Pareto dominance in
evaluating fitness or assigning selection probability to solutions. The population is ranked
according to a dominance rule, and then each solution is assigned a fitness value based on its
rank in the population, not its actual objective function value. Note that herein all objectives are
assumed to be minimized. Therefore, a lower rank corresponds to a better solution in the
8
following discussions.
The first Pareto ranking technique was proposed by Goldberg [15] as follows:
Step 1. Set i=1 and TP=P
Step 2. Identify non-dominated solutions in TP and assigned them set to F
i
.
Step 3. Set TP = TP \ F
i
. If TP= go to Step 4, else set i=i+1 and go to Step 2.
Step 4. For every solution xP at generation t, assign rank
1
( , ) r t i = x if xF
i
.
In the procedure above, F
1
, F
2
, ... are called non-dominated fronts, and F
1
is the Pareto
front of population P. Fonseca and Fleming [13] used a slightly different rank assignment
approach follows:
2
( , ) 1 ( , ) r t nq t = + x x
where ( , ) nq t x is the number of solutions dominating solution x at generation t. This ranking
method penalizes solutions located in the regions of the objective function space which are
dominated (covered) by densely populated sections of the Pareto front. For example, in Figure 1b
solution i is dominated by solutions c, d and e. Therefore, it is assigned a rank of 4 although it is
in the same front with solutions f, g and h which are dominated by only a single solution.
The SPEA [55] uses a ranking procedure to assign better fitness values to non-dominated
solutions at underrepresented regions of the objective space. In the SPEA, an external list E of a
fixed size stores non-dominated solutions that have been investigated thus far during the search.
For each solution yE, a strength value is defined as,

( , )
( , )
1
P
np t
s t
N
=
+
y
y
where ( , ) np t y is the number solutions that y dominates in P. The rank r(y,t) of a solution yE is
assigned as
3
( , ) ( , ) r t s t = y y and the rank of a solution xP is calculated as,

3
,
( , ) 1 ( , )
E
r t s t

= +

y y x
x y

Figure 1c illustrates an example of the SPEA ranking method. In the former two methods,
all non-dominated solutions are assigned a rank of 1. This method, however, favors solution a (in
the figure) over the other non-dominated solutions since it covers the least number of solutions in
the objective function space. Therefore, a wide, uniformly distributed set of non-dominated
solutions is encouraged.
9
Accumulated ranking density strategy [32] also aims to penalize redundancy in the
population due to overrepresentation. This ranking method is given as,
4
,
( , ) 1 ( , )
P
r t r t

= +

y y x
x y

To calculate the rank of a solution x, the rank of the solutions dominating this solution
must be calculated first. Figure 1d shows an example of this ranking method (based on r
2
). Using
ranking method r
4
, solutions i, l and n are ranked higher than their counterparts at the same non-
dominated front since the portion of the trade-off surface covering them is crowded by three
nearby solutions c, d and e.
4.2. Diversity: Fitness Assignment, Fitness Sharing, and Niching.
Maintaining a diverse population is an important consideration in multi-objective GA to
obtain solutions uniformly distributed over the true Pareto font. Without taking any preventive
measures, the population tends to form relatively few clusters in multi-objective GA. This
phenomenon is called genetic drift, and several approaches are used to prevent genetic drift, as
follows.
4.2.1. Fitness Sharing
Fitness sharing aims to encourage the search in unexplored sections of a Pareto front by
artificially reducing fitness of solutions in densely populated areas. To achieve this goal, densely
populated areas are identified and a fair penalty method is used to penalize the solutions located
in such areas.
The idea of fitness sharing was first proposed by Goldberg and Richardson [16] in the
investigation of multiple local optima for multi-modal functions. Fonseca and Fleming [13] used
this idea to penalize clustered solutions with the same rank as follows.
Step 1. Calculate the Euclidean distance between every solution pair x and y in the
normalized objective space between 0 and 1 as

2
max min
1
( ) ( )
( , )
K
k k
k
k k
z z
dz
z z
=
| |
=
|

\ .

x y
x y (2)
where
max
k
z and
min
k
z are the maximum and minimum value of the objective
function ( )
k
z observed so far during the search, respectively.
Step 2. Based on these distances, calculate a niche count for each solution xP as
10

share
share
( , ) ( , )
( , )
( , ) max , 0
P
r t r t
d
nc t

=

=
`
)

y
y x
x y
x
where
share
is the niche size.
Step 3. After calculating niche counts, the fitness of each solution is adjusted as follows:

( , )
( , )
( , )
f t
f t
nc t
=
x
x
x

In the procedure above,
share
defines a neighborhood of solutions in the objective space
(Figure 1a). The solutions in the same neighborhood contribute to each others niche count.
Therefore, a solution in a crowded neighborhood will have a higher niche count reducing the
probability of selecting that solution as a parent. As a result, niching limits the proliferation of
solutions in one particular neighborhood of the objective function space.
Another alternative is to use the Hamming distance (the distance in the decision variable
space) between two solutions x and y which is defined as

2
1
1
( , ) ( )
M
i i
i
dx x y
M
=
=

x y (3)
in the calculation of niche count. Equation (3) is a measure of structural differences between two
solutions. Two solutions might be very close in the objective function space while they have very
different structural features. Therefore, fitness sharing based on the objective function space
may reduce diversity in the decision variable space. However, Deb and Goldberg [8] reported
that fitness sharing on the objective function space usually performs better than one based on the
decision variable space.
One of the disadvantages of the fitness sharing based on niche count is that the user has
to select a new parameter
share
. To address this problem, Deb and Goldberg [8] and Fonseca and
Fleming [13] developed systematic approaches to estimate and dynamically update
share
.
Another disadvantage of niching is computational effort to calculate niche counts. However,
benefits of fitness sharing surpass the burden of extra computational effort in many applications.
Miller and Shaw [36] proposed a dynamic niche sharing approach to increase effectiveness of
computing niche counts.
4.2.2. Crowding Distance
Crowding distance approaches aim to obtain a uniform spread of solutions along the best-
11
known Pareto front without using a fitness sharing parameter. For example, the NSGA-II [9] use
a crowding distance method as follows (Figure 2b):
Step 1. Rank the population and identify non-dominated fronts F
1
, F
2
, ..., F
R
. For each
front j=1, ..., R repeat Steps 2 and 3.
Step 2. For each objective function k, sort the solutions in F
j
in the ascending order. Let
l=|F
j
| and
[ , ] i k
x represent the i
th
solution in the sorted list with respect to the
objective function k. Assign
[1, ]
( )
k k
cd = x and
[ , ]
( )
k l k
cd = x , and for i=2, ..., l
assign

[ 1, ] [ 1, ]
[ , ] max min
( ) ( )
( )
k
k i k k i k
k i k
k k
z z
cd
z z
+

x x
x
Step 3. To find the total crowding distance cd(x) of a solution x, sum the solution
crowding distances with respect to each objective, i.e., ( ) ( )
k
k
cd cd =

x x .
The main advantage of the crowding approach described above is that a measure of
population density around a solution is computed without requiring a user-defined parameter. In
the NSGA-II, this crowding distance measure is used as a tie-breaker as in the selection phase
that follows. Randomly select two solutions x and y; if the solutions are in the same non-
dominated front, the solution with a higher crowding distance wins. Otherwise, the solution with
the lowest rank is selected.
4.2.3. Cell-Based Density
In this approach [26, 27, 32, 53], the objective space is divided into K-dimensional cells
(see Figure 2c). The number of solutions in each cell is defined as the density of the cell, and the
density of a solution is equal to the density of the cell in which the solution is located. This
density information is used to achieve diversity similarly to the fitness sharing approach. For
example, in the PAES [26, 27], between two non-dominated solutions, the one with a lower
density is preferable.
Lu and Yen [32, 53] developed an efficient approach to identify a solutions cell in case
of dynamic cell dimensions. In this approach, the width of a cell along the kth objective
dimension is
max min
( ) /
k k k
z z n where n
k
is the number cells dedicated the kth objective dimension
and
max
k
z and
min
k
z are the maximum and minimum values of the objective function k so far in the
search, respectively. Therefore, cell boundaries are updated when a new maximum or minimum
12
objective function value is discovered.
The main advantage of the cell based density approach is that a global density map of the
objective function space is obtained as a result of the density calculation. The search can be
encouraged toward sparsely inhabited regions of the objective function space based on this map.
The RDGA [32] uses a method based on this global density map to push solutions out of high
density areas to low density areas.
4.3. Elitisim
Elitism in the context of single-objective GA means that the best solution found so far
during the search has immunity against selection and always survives in the next generation. In
this respect, all non-dominated solutions discovered by a multi-objective GA are considered as
elite solutions. However, implementation of elitism in multi-objective optimization is not as
straightforward as in single objective optimization mainly due to the large number of possible
elitist solutions. Earlier multi-objective GA did not use elitism. However, most recent multi-
objective GA and their variations use elitism. As discussed in [6, 47, 55], multi-objective GA
using elitist strategies tend to outperform their non-elitist counterparts. Multi-objective GA use
two strategies to implement elitism [22]: (i) maintaining elitist solutions in the population, and
(ii) storing elitist solutions in an external secondary list and reintroducing them to the population.
4.3.1. Strategies to Maintain Elitist Solutions in the Population
Random selection does not ensure that a non-dominated solution will survive in the next
generation. A straightforward implementation of elitism in a multi-objective GA is to copy all
non-dominated solution in population P
t
to population P
t+1
, then fill the rest of P
t+1
by selecting
from the remaining dominated solutions in P
t
. This approach will not work when the total
number of non-dominated parent and offspring solutions is larger than N
P
. To address this
problem, several approaches have been proposed.
Konak and Smith [29, 30] proposed a multi-objective GA with dynamic population size
and a pure elitist strategy. In this multi-objective GA, the population includes only non-
dominated solutions. If the size of the population reaches an upper bound N
max
, N
max
-N
min

solutions are removed from the population giving consideration to maintaining the diversity of
the current non-dominated front. To achieve this, the Pareto domination tournament selection is
used as follows [19]. Two solutions are randomly chosen and the solution with the higher niche
count is removed since all solutions are non-dominated. A similar pure elitist multi-objective GA
13
with a dynamic population size has also been proposed [42].
The NSGA-II uses a fixed population size of N. In generation t, an offspring population
Q
t
of size N is created from parent population P
t
and non-dominated fronts F
1
, F
2
, ..., F
R
are
identified in the combined population P
t
Q
t
. The next population P
t+1
is filled starting from
solutions in F
1
, then F
2
, and so on as follows. Let k be the index of a non-dominated front F
k
that
|F
1
F
2
...F
k
| N and |F
1
F
2
... F
k
F
k+1
| > N. First, all solutions in fronts F
1
, F
2
, ..., F
k

are copied to P
t+1
, and then the least crowded (N-|P
t+1
|) solutions in F
k+1
are added to P
t+1
. This
approach makes sure that all non-dominated solutions (F
1
) are included in the next population if
|F
1
|N, and otherwise the selection based on a crowding distance will promote diversity.
4.3.2. Elitism with External Populations
When an external list is used to store elitist solutions, several issues must be addressed.
The first issue is which solutions are going to be stored in elitist list E. Most multi-objective GA
store non-dominated solutions investigated so far during the search [55], and E is updated each
time a new solution is created by removing elitist solutions dominated by the new solution or
adding the new solution if it is not dominated by any existing elitist solution. This is a
computationally expensive operation. Several data structures were proposed to efficiently store,
update, search in list E [11, 38]. Another issue is the size of list E. Since there might possibly
exist a very large number of Pareto optimal solutions for a problem, the elitist list can grow
extremely large. Therefore, pruning techniques were proposed to control the size of E. For
example, the SPEA uses the average linkage clustering method [37] to reduce the size of E to an
upper limit N when the number of the non-dominated solutions exceeds N as follows.
Step 1. Initially, assign each solution xE to a cluster c
i
,
1 2
{ , , , }
M
C c c c =
Step 2. Calculate the distance between all pairs of clusters c
i
and c
j
as follows

,
,
1
( , )
| | | |
i j
i j
c c
c c
i j
d d
c c

=

x y
x y
Here, the distance ( , ) d x y can be calculated in the objective function space using
equation (2) or in the decision variable space using equation (3).
Step 3. Merge the cluster pair c
i
and c
j
with the minimum distance among all clusters into
a new cluster.
Step 4. If |C| N, go to Step 5, else go to Step 2.
14
Step 5. For each cluster, determine a solution with the minimum average distance to all
other solutions in the same cluster (called centroid solution). Keep the centroid
solutions for every cluster and remove other solutions from E.
The final issue is the selection of elitist solutions from E to be reintroduced to the
population. In [32, 53, 55], solutions for P
t+1
are selected from the combined population of P
t
and
E
t
. To implement this strategy, population P
t
and E
t
are combined together, a fitness value is
assigned to each solution in the combined population P
t
E
t
, and then, N solutions are selected
for the next generation P
t+1
based on the assigned fitness values. Another strategy is to reserve a
room for n elitist solutions in the next population [20]. In this strategy, N - n solutions are
selected from parents and newly created offspring and n solutions are selected from E
t
.
4.4. Constraint Handling
Most real-world optimization problems include constraints that must be satisfied. Single-
objective GA use four different constraint handling strategy: (i) discarding infeasible solutions,
(ii) reducing the fitness of infeasible solutions by using a penalty function, (iii) if possible,
customizing genetic operators to always produce feasible solutions, and (iv) repairing infeasible
solutions. Handling of constraints has not been adequately researched for multi-objective GA
[23]. For instance, all major multi-objective GA assumed problems without any constraints.
While constraint handling strategies (i), (iii), and (iv) are directly applicable in the multi-
objective case, implementation of penalty function strategies, which is by far the most frequently
used constraint handling strategy in single-objective GA, is not straightforward in multi-
objective GA, mainly due to fact that fitness assignment is usually based on the non-dominance
rank of a solution, not on its objective function values.
Jimenez et al. [24] proposed a niched selection strategy to address infeasibility in multi-
objective problems as follows:
Step 1. Randomly chose two solutions x and y from the population.
Step 2. If one of the solutions is feasible and the other one is infeasible, the winner is the
feasible solution, and stop. Otherwise, if both solutions are infeasible go to Step 3,
else go to step 4.
Step 3. In this case, solutions x and y are both infeasible. Then, select a random reference
set C among infeasible solutions in the population. Compare solutions x and y to
the solutions in reference set C with respect to their degree of infeasibility. In
15
order to achieve this, calculate a measure of infeasibility (e.g., the number of
constraints violated or total constraint violation) for solutions x, y, and in set C. If
one of solutions x and y is better and the other one is worse than the best solution
in C, with respect to the calculated infeasibility measure, then the winner is the
least infeasible solution. However, if there is a tie, that is both solutions x and y
are either better or worse than the best solution in C, then their niche counts in the
decision variable space (equation (3)) is used for selection. In this case, the
solution with the lower niche count is the winner.
Step 4. In this case, solutions x and y are both feasible. Then, select a random reference
set C among feasible solutions in the population. Compare solutions x and y to the
solutions in set C. If one of them is non-dominated in set C, and the other is
dominated by at least one solution, the winner is the former. Otherwise, there is a
tie between solutions x and y, and the niche count of the solutions are calculated
in the decision variable space. The solution with the smaller niche count is the
winner of the tournament selection.
The procedure above is a comprehensive approach to deal with infeasibility while
maintaining diversity and dominance of the population. Main disadvantages of this procedure are
its computational complexity and additional parameters such as the size of reference set C and
niche size. Modifications are also possible. In Step 4, for example, the niche count of the
solutions can be calculated in the objective function space instead of the decision variable space.
In Step 3, the solution with the least infeasibility can be declared as the winner without
comparing solutions x and y to a reference set C with respect to infeasibility. Such modifications
can reduce the computational complexity of the procedure.
Deb [9] proposed the constrain-domination concept and a binary tournament selection
method based on it, called a constrained tournament method. A solution x is said to constrain-
dominate a solution y if either of the following cases are satisfied:
Case 1: Solution x is feasible and solution y is infeasible.
Case 2: Solutions x and y are both infeasible; however, solution x has a smaller constraint
violation than y.
Case 3: Solutions x and y are both feasible, and solution x dominates solution y.
In the constraint tournament method, first non-constrain-dominance fronts F
1
, F
2
, F
3
, ....,
16
F
R
are identified in a similar way defined in [15], but by using the constrain-domination criterion
instead of the regular domination concept. Note that set F
1
corresponds to the set of feasible non-
dominated solutions in the population and front F
i
is more preferred than F
j
for i<j. In the
constraint tournament selection, two solutions x and y are randomly chosen from the population.
Between x and y, the winner is the one in a more preferred non-constrain-dominance front. If
solutions x and y are both in the same front, then the winner is decided based on niche counts or
crowding distances of the solution. The main advantages of the constrained tournament method
are that it requires fewer parameters and it can be easily integrated to multi-objective GA.
4.5. Parallel and Hybrid Multi-Objective GA
All comparative studies on multi-objective GA agree that elitism and diversity
preservation mechanisms improve performance of multi-objective GA. However, implementing
elitism and diversity preservation strategies usually require substantial computational effort and
computer memory. In addition, evaluation of objective functions may take considerable time in
real-life problems. Therefore, researchers have been interested in reducing execution time and
resource requirement of multi-objective GA using advanced data structures. One of the latest
trends in this avenue is parallel and distributed processing. Several recent papers [5, 48-50]
presented parallel implementation of multi-objective GA over multiple processors.
Hybridization of GA with local search algorithms is frequently applied in single-
objective GAs. This approach is usually referred to as a memetic algorithm. Generally, a local
search algorithm proceeds as follows.
Step 1. Start with an initial solutions x.
Step 2. Generate a set of neighbor solutions around solution x using a simple perturbation
rule.
Step 3. If the best solution in the neighborhood set is better than x, replace x with this
solution and go to Step 2, else stop.
A local search algorithm is particularly effective in finding local optima if the solution
space around the initial solution is convex. This is usually difficult to achieve using standard GA
operators. In hybridization of multi-objective GA with local search algorithms, important issues
are: (i) selecting a solution to apply the local search and (ii) identifying a solution in the
neighborhood as the new best solution when multiple non-dominated local solutions exist.
Several approaches have been proposed to address these two issues as follows.
17
Paquete and Stutzle [41] described a bi-objective GA where a local search is used to
generate initial solutions by optimizing only one single objective. Deb and Goel [7] applied a
local search to only final solutions. In Ishibuchi and Muratas approach [20], a local search
procedure is applied to each offspring generated by crossover, using the same weight vector of
the offsprings parents to evaluate neighborhood solutions. Similarly, Ishibuchi [21] also used
the weighted sum of the objective functions to evaluate solution during the local search.
However, the local search is selectively applied to only promising solutions, and weights are also
randomly generated, instead of using the parents weight vector. Knowles and Corne [28]
presented a memetic version of the PAES, called M-PAES. The PAES uses the dominance
concept to evaluate solutions. Therefore, in M-PAES, a set of local non-dominated solutions is
used as a comparison set for solutions investigated during the local search. When a new solution
is created in the neighborhood, it is only compared with this local non-dominated set and
necessary updates are performed. The local search is terminated after a maximum number of
local solutions are investigated or a maximum number of local moves are performed without any
improvement. Tan et al. [46] proposed applying a local search procedure to only solutions that
are located apart from others. In addition, the neighborhood size in the local search depends on
the density or crowdedness of solutions. Being selective in applying a local search, this strategy
is computationally efficient and also aims to main diversity.
5. Multi-objective GA for Reliability Optimization
Many engineering problems have multiple objectives, including engineering system
design and reliability optimization. There have been several interesting and successful
implementations of multi-objective GA for this class of problems. A few successful examples are
described in the following paragraphs.
Marseguerra, Zio and Podofillini [33] determine optimal surveillance test intervals using
multi-objective GA with the goal of improving reliability and availability. Their research
implemented a multi-objective GA which transparently and explicitly accounts for the
uncertainties in the parameters. The objectives considered were the inverse of the expected
system failure probability and the inverse of its variance. These are used to drive the genetic
search toward solutions which are guaranteed to give optimal performance with high assurance,
i.e., low estimation variance. They successfully applied their procedure to a complex system, a
residual heat removal safety system for a boiling water reactor.
18
Martorell et al. [35] studied the selection of technical specifications and maintenance
activities at nuclear power plants to increase reliability, availability and maintainability (RAM)
of safety-related equipment. However, to improve RAM, additional limited resources (e.g. costs,
task force, etc.) are required creating a multi-objective problem. They demonstrated the viability
and significance of their proposed approach using multi-objective GA for an emergency diesel
generator system.
Additionally, Martorell et al. [34] considered the optimal allocation of more reliable
equipment, testing and maintenance activities to assure high RAM levels for safety-related
systems. For these problems, the decision-maker encounters a multi-objective optimization
problem where the parameters of design, testing and maintenance are decision variables.
Solutions were obtained by using both single-objective GA and multi-objective GA, which were
demonstrated to solve the problem of testing and maintenance optimization based on
unavailability and cost criteria.
Sasaki and Gen [43] introduce a multi-objective problem which had fuzzy multiple
objective functions and constraints with GUB (Generalized Upper Bounding) structure. They
solved this problem by using a new hybridized GA. This approach leads to a flexible optimal
system design by applying fuzzy goals and fuzzy constraints. A new chromosome representation
was introduced in their work. To demonstrate the effectiveness of their method, a large-scale
optimal system reliability design problem was analyzed.
Reliability allocation to minimize total plant costs, subject to an overall plant safety goal,
is presented by Yang [52]. For their problem, design optimization is needed to improve the
design, operation and safety of new and/or existing nuclear power plants. They presented an
approach to determine the reliability characteristics of reactor systems, subsystems, major
components and plant procedures that are consistent with a set of top-level performance goals.
To optimize the reliability of the system, the cost for improving and/or degrading the reliability
of the system are also included in the reliability allocation process creating a multi-objective
problem. GA was applied to the reliability allocation problem of a typical pressurized water
reactor.
Elegbede and Adjallah [10] present a methodology to optimize the availability and the
cost of repairable parallel-series systems. It is a multi-objective combinatorial optimization,
modeled with continuous and discrete variables. They transform the problem into a single
19
objective problem and used traditional GA.

z
2
a
1
b
1
c
1
d
1
e
1
f
2
g
2
h
2
i
2
j
3
k
3
l
3
m
4
n
4
z
1
z
2
a
1
b
1
c
1
d
1
e
1
f
2
g
2
h
2
i
4
j
5
k
4
l
6
m
5
n
8
z
1
a
2/15
b
7/15
c
5/15
d
4/15
e
3/15
f
g
h
i
j
k
l
m
n
z
2
a
1
b
1
c
1
d
1
e
1
f
2
g
2
h
2
i
4
j
6
k
5
l
9
m
9
n
17
z
1 z
1
z
2
(a) (b)
(c) (d)
F
2
F
1
F
3
F
4

Figure 1. Ranking methods used in multi-objective GA.


z
2
z
1
(b)
a
a
x
a
cd
1
(x)
cd
2
(x)
a
a
a
a
a
z
2
z
1
(c)
a
a
x
a
a
a
a
a
a
0 0
0
3 0
1
1
1
1
0 1
2
z
2
z
1
(a)
a
a
x
a
a
a
a
a
a

share

Figure 2. Diversity methods used in multi-objective GA.
20
REFERENCES
[1] Coello, C.A.C., A comprehensive survey of evolutionary-based multiobjective
optimization techniques, Knowledge and Information Systems 1(3) (1999) 269-308.
[2] Coello, C.A.C. An updated survey of evolutionary multiobjective optimization
techniques: state of the art and future trends. in Proceedings of the 1999. Congress on
Evolutionary Computation-CEC99, 6-9 July 1999. 1999. Washington, DC, USA: IEEE.
[3] Coello, C.A.C., An updated survey of GA-based multiobjective optimization techniques,
ACM Computing Surveys 32(2) (2000) 109-143.
[4] Coello Coello, C., List of References on Evolutionary Multiobjective Optimization,
http://www.lania.mx/~ccoello/EMOO/EMOObib.html
[5] de Toro, F., Ortega, J., Fernandez, J., and Diaz, A. PSFGA: a parallel genetic algorithm
for multiobjective optimization. in Proceedings 10th Euromicro Workshop on Parallel,
Distributed and Network-based Processing, 9-11 Jan. 2002. 2002. Canary Islands, Spain:
IEEE Comput. Soc.
[6] Deb, K., Multi-Objective Optimization using Evolutionary Algorithms, John Wiley &
Sons, Ltd, 2001.
[7] Deb, K. and Goel, T. A hybrid multi-objective evolutionary approach to engineering
shape design. in Evolutionary Multi-Criterion Optimization. First International
Conference, EMO 2001, 7-9 March 2001. 2001. Zurich, Switzerland: Springer-Verlag.
[8] Deb, K. and Goldberg, D.E. An investigation of of niche an species fromation in genetic
function optimization. in The Third International Conference on Genetic Algorithms.
1989. George Mason University.
[9] Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T., A fast and elitist multiobjective
genetic algorithm: NSGA-II, IEEE Transactions on Evolutionary Computation 6(2)
(2002) 182-197.
[10] Elegbede, C. and Adjallah, K., Availability allocation to repairable systems with genetic
algorithms: A multi-objective formulation, Reliability Engineering and System Safety
82(3) (2003) 319-330.
[11] Fieldsend, J.E., Everson, R.M., and Singh, S., Using unconstrained elite archives for
multiobjective optimization, IEEE Transactions on Evolutionary Computation 7(3)
(2003) 305-323.
[12] Fonseca, C.M. and Fleming, P.J. Genetic algorithms for multiobjective optimization:
formulation, discussion and generalization. in Proceedings of ICGA-93: Fifth
International Conference on Genetic Algorithms, 17-22 July 1993. 1993. Urbana-
Champaign, IL, USA: Morgan Kaufmann.
[13] Fonseca, C.M. and Fleming, P.J. Multiobjective genetic algorithms. in IEE Colloquium
on `Genetic Algorithms for Control Systems Engineering' (Digest No. 1993/130), 28 May
1993. 1993. London, UK: IEE.
21
[14] Fonseca, C.M. and Fleming, P.J., Multiobjective optimization and multiple constraint
handling with evolutionary algorithms. I. A unified formulation, IEEE Transactions on
Systems, Man & Cybernetics, Part A (Systems & Humans) 28(1) (1998) 26-37.
[15] Goldberg, D.E., Genetic Algorithms in Search, Optimization, and Machine Learning,
Addison-Wesley, Reading, MA, 1989.
[16] Goldberg, D.E. and Richardson, J. Genetic algorithms with sharing for multimodal
function optimization. in Genetic Algorithms and their Applications: Proceedings of the
Second International Conference on Genetic Algorithms, 28-31 July 1987. 1987.
Cambridge, MA, USA: Lawrence Erlbaum Associates.
[17] Hajela, P. and Lin, C.-Y., Genetic search strategies in multicriterion optimal design,
Structural Optimization 4(2) (1992) 99-107.
[18] Holland, J.H., Adaptation in Natural and Artificial Systems, University of Michigan
Press, Ann Arbor, 1975.
[19] Horn, J., Nafpliotis, N., and Goldberg, D.E. A niched Pareto genetic algorithm for
multiobjective optimization. in Proceedings of the First IEEE Conference on
Evolutionary Computation. IEEE World Congress on Computational Intelligence, 27-29
June 1994. 1994. Orlando, FL, USA: IEEE.
[20] Ishibuchi, H. and Murata, T. Multi-objective genetic local search algorithm. in
Proceedings of IEEE International Conference on Evolutionary Computation, 20-22 May
1996. 1996. Nagoya, Japan: IEEE.
[21] Ishibuchi, H., Yoshida, T., and Murata, T., Balance between genetic search and local
search in memetic algorithms for multiobjective permutation flowshop scheduling, IEEE
Transactions on Evolutionary Computation 7(2) (2003) 204-223.
[22] Jensen, M.T., Reducing the run-time complexity of multiobjective EAs: The NSGA-II
and other algorithms, IEEE Transactions on Evolutionary Computation 7(5) (2003) 503-
515.
[23] Jimenez, F., Gomez-Skarmeta, A.F., Sanchez, G., and Deb, K. An evolutionary algorithm
for constrained multi-objective optimization. in Proceedings of 2002 World Congress on
Computational Intelligence - WCCI'02, 12-17 May 2002. 2002. Honolulu, HI, USA:
IEEE.
[24] Jimenez, F., Verdegay, J.L., and Gomez-Skarmeta, A.F. Evolutionary techniques for
constrained multiobjective optimization problems. in Workshop on Multi-Criterion
Optimization Using Evolutionary Methods GECCO-1999. 1999.
[25] Jones, D.F., Mirrazavi, S.K., and Tamiz, M., Multiobjective meta-heuristics: an overview
of the current state-of-the-art, European Journal of Operational Research 137(1) (2002)
1-9.
[26] Knowles, J. and Corne, D. The Pareto archived evolution strategy: a new baseline
algorithm for Pareto multiobjective optimisation. in Proceedings of the 1999. Congress
on Evolutionary Computation-CEC99, 6-9 July 1999. 1999. Washington, DC, USA:
IEEE.
22
[27] Knowles, J.D. and Corne, D.W., Approximating the nondominated front using the Pareto
archived evolution strategy, Evolutionary Computation 8(2) 149-172.
[28] Knowles, J.D. and Corne, D.W. M-PAES: a memetic algorithm for multiobjective
optimization. in Proceedings of 2000 Congress on Evolutionary Computation, 16-19 July
2000. 2000. La Jolla, CA, USA: IEEE.
[29] Konak, A. and Smith, A.E. Multiobjective optimization of survivable networks
considering reliability. in The 10th International Conference on Telecommunication
Systems. 2002. Naval Postgraduate School, Monterey, CA,.
[30] Konak, A. and Smith, A.E., Capacitated Network Design Considering Survivability: An
Evolutionary Approach, Journal of Engineering Optimization 36(2) (2004) 189-205.
[31] Kursawe, F. A variant of evolution strategies for vector optimization. in Parallel Problem
Solving from Nature. 1st Workshop, PPSN 1 Proceedings, 1-3 Oct. 1990. 1991.
Dortmund, West Germany: Springer-Verlag.
[32] Lu, H. and Yen, G.G., Rank-density-based multiobjective genetic algorithm and
benchmark test function study, IEEE Transactions on Evolutionary Computation 7(4)
(2003) 325-343.
[33] Marseguerra, M., Zio, E., and Podofillini, L., Optimal reliability/availability of uncertain
systems via multi-objective genetic algorithms, IEEE Transactions on Reliability 53(3)
424-434.
[34] Martorell, S., Sanchez, A., Carlos, S., and Serradell, V., Alternatives and challenges in
optimizing industrial safety using genetic algorithms, Reliability Engineering & System
Safety 86(1) (2004) 25-38.
[35] Martorell, S., Villanueva, J.F., Carlos, S., Nebot, Y., Sanchez, A., Pitarch, J.L., and
Serradell, V., RAMS+C informed decision-making with application to multi-objective
optimization of technical specifications and maintenance using genetic algorithms,
Reliability Engineering and System Safety 87(1) (2005) 65-75.
[36] Miller, B.L. and Shaw, M.J. Genetic algorithms with dynamic niche sharing for
multimodal function optimization. in Proceedings of the 1996 IEEE International
Conference on Evolutionary Computation, ICEC'96, May 20-22 1996. 1996. Nagoya,
Jpn: IEEE, Piscataway, NJ, USA.
[37] Morse, J.N., Reducing the size of the nondominated set: pruning by clustering,
Computers & Operations Research 7(1-2) (1980) 55-66.
[38] Mostaghim, S., Teich, J., and Tyagi, A. Comparison of data structures for storing Pareto-
sets in MOEAs. in Proceedings of 2002 World Congress on Computational Intelligence -
WCCI'02, 12-17 May 2002. 2002. Honolulu, HI, USA: IEEE.
[39] Murata, T. and Ishibuchi, H. MOGA: multi-objective genetic algorithms. in Proceedings
of 1995 IEEE International Conference on Evolutionary Computation, 29 Nov.-1 Dec.
1995. 1995. Perth, WA, Australia: IEEE.
[40] Murata, T., Ishibuchi, H., and Tanaka, H., Multi-objective genetic algorithm and its
applications to flowshop scheduling, Computers & Industrial Engineering 30(4) 957-968.
23
[41] Paquete, L. and Stutzle, T. A two-phase local search for the biobjective traveling
salesman problem. in Evolutionary Multi-Criterion Optimization. Second International
Conference, EMO 2003. Proceedings, 8-11 April 2003. 2003. Faro, Portugal: Springer-
Verlag.
[42] Sarker, R., Liang, K.-H., and Newton, C., A new multiobjective evolutionary algorithm,
European Journal of Operational Research 140(1) (2002) 12-23.
[43] Sasaki, M. and Gen, M., A method of fuzzy multi-objective nonlinear programming with
GUB structure by Hybrid Genetic Algorithm, International Journal of Smart Engineering
System Design 5(4) (2003) 281-288.
[44] Schaffer, J.D. Multiple Objective optimization with vector evaluated genetic algorithms.
in International Conference on Genetic Algorithm and their applications. 1985.
[45] Srinivas, N. and Deb, K., Multiobjective Optimization Using Nondominated Sorting in
Genetic Algorithms, Journal of Evolutionary Computation 2(3) (1994) 221-248.
[46] Tan, K.C., Lee, T.H., and Khor, E.F., Evolutionary algorithms with dynamic population
size and local exploration for multiobjective optimization, IEEE Transactions on
Evolutionary Computation 5(6) (2001) 565-588.
[47] Van Veldhuizen, D.A. and Lamont, G.B., Multiobjective evolutionary algorithms:
analyzing the state-of-the-art, Evolutionary Computation 8(2) 125-147.
[48] Van Veldhuizen, D.A., Zydallis, J.B., and Lamont, G.B., Considerations in engineering
parallel multiobjective evolutionary algorithms, IEEE Transactions on Evolutionary
Computation 7(2) (2003) 144-173.
[49] Wilson, L.A., Moore, M.D., Picarazzi, J.P., and Miquel, S.D.S. Parallel genetic algorithm
for search and constrained multi-objective optimization. in Proceedings. 18th
International Parallel and Distributed Processing Symposium, 26-30 April 2004. 2004.
Santa Fe, NM, USA: IEEE Comput. Soc.
[50] Xiong, S. and Li, F. Parallel strength Pareto multiobjective evolutionary algorithm. in
Proceedings of the Fourth International Conference on Parallel and Distributed
Computing, Applications and Technologies, 27-29 Aug. 2003. 2003. Chengdu, China:
IEEE.
[51] Xiujuan, L. and Zhongke, S., Overview of multi-objective optimization methods, Journal
of Systems Engineering and Electronics 15(2) (2004) 142-146.
[52] Yang, J.-E., Hwang, M.-J., Sung, T.-Y., and Jin, Y., Application of genetic algorithm for
reliability allocation in nuclear power plants, Reliability Engineering & System Safety
65(3) 229-238.
[53] Yen, G.G. and Lu, H., Dynamic multiobjective evolutionary algorithm: adaptive cell-
based rank and density estimation, IEEE Transactions on Evolutionary Computation 7(3)
(2003) 253-274.
[54] Zitzler, E., Deb, K., and Thiele, L., Comparison of multiobjective evolutionary
algorithms: empirical results, Evolutionary Computation 8(2) 173-195.
24
[55] Zitzler, E. and Thiele, L., Multiobjective evolutionary algorithms: a comparative case
study and the strength Pareto approach, IEEE Transactions on Evolutionary Computation
3(4) (1999) 257-271.