You are on page 1of 14

3306

IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12, DECEMBER 2016

An Orthogonal Evolutionary Algorithm


With Learning Automata for
Multiobjective Optimization
Cai Dai, Yuping Wang, Senior Member, IEEE, Miao Ye, Xingsi Xue, and Hailin Liu

AbstractResearch on multiobjective optimization problems


becomes one of the hottest topics of intelligent computation. In
order to improve the search efficiency of an evolutionary algorithm and maintain the diversity of solutions, in this paper, the
learning automata (LA) is first used for quantization orthogonal
crossover (QOX), and a new fitness function based on decomposition is proposed to achieve these two purposes. Based on
these, an orthogonal evolutionary algorithm with LA for complex
multiobjective optimization problems with continuous variables
is proposed. The experimental results show that in continuous
states, the proposed algorithm is able to achieve accurate Paretooptimal sets and wide Pareto-optimal fronts efficiently. Moreover,
the comparison with the several existing well-known algorithms:
nondominated sorting genetic algorithm II, decompositionbased multiobjective evolutionary algorithm, decompositionbased multiobjective evolutionary algorithm with an ensemble
of neighborhood sizes, multiobjective optimization by LA, and
multiobjective immune algorithm with nondominated neighborbased selection, on 15 multiobjective benchmark problems,
shows that the proposed algorithm is able to find more accurate and evenly distributed Pareto-optimal fronts than the
compared ones.
Index TermsArtificial intelligence, evolutionary algorithm,
learning automata (LA), multiobjective optimization, quantization orthogonal crossover (QOX).

I. I NTRODUCTION
ULTIOBJECTIVE problems (MOPs) are typically
characterized by conflicting objectives, and the algorithms for MOPs should be able to [1]: 1) discover solutions as
close to the Pareto-optimal solutions as possible; 2) find solutions as diversely as possible in the obtained nondominated
front; and 3) find solutions distributed in the nondominated
front as evenly as possible. However, achieving these three

Manuscript received June 12, 2015; revised October 12, 2015; accepted
November 11, 2015. Date of publication December 17, 2015; date of current version November 15, 2016. This work was supported in part by the
National Natural Science Foundation of China under Grant 61472297 and
Grant 61272119, and in part by the Fundamental Research Funds for the
Central Universities under Grant BDZ021430. This paper was recommended
by Associate Editor Y. S. Ong. (Corresponding author: Yuping Wang.)
C. Dai, Y. Wang, M. Ye, and X. Xue are with the School of Computer
Science and Technology, Xidian University, Xian 710071, China (e-mail:
ywang@xidian.edu.cn).
H. Liu is with the School of Mathematics and Statistics, Guangdong
University of Technology, Guangzhou 510006, China.
This paper has supplementary downloadable multimedia material available
at http://ieeexplore.ieee.org provided by the authors.
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TCYB.2015.2503433

goals simultaneously is still a challenging task for multiobjective optimization algorithms. Among various multiobjective
optimization algorithms, multiobjective evolutionary algorithms (MOEAs), which make use of the population evolution
to get the optimal solutions, are a kind of effective methods for
solving MOPs. Nowadays, there exist many MOEAs [2][16],
such as multiobjective genetic algorithms (GAs) [3], multiobjective particle swarm optimization algorithms [5], [6],
multiobjective differential evolution (DE) algorithms [7], multiobjective immune clone algorithms [9], and group search
optimizer [12]. To enhance the performance of MOEAs, some
mathematical improvements, such as orthogonal experimental
design, are often applied [13].
In the evolutionary algorithm, an experiment refers to the
steps of generating some new solutions. Orthogonal experimental design is a method of generating uniformly distributed
multiple solutions, and it has been developed to sample
a small but representative set of combinations of components of variables. Orthogonal experimental design was first
introduced into the GA by Zhang and Leung [17] for discrete
optimization problems. Later, Leung and Wang [18] applied
quantization technique into orthogonal experiment design and
proposed quantization orthogonal crossover (QOX) for numerical optimization problems. Experimental studies [19], [20]
show QOX is an effective and efficient operator for numerical optimization problems. In order to limit the number of
potential offspring of each pair of parents and avoid a large
number of function evaluations during selection, QOX groups
variables into a few groups randomly and sequentially [18].
However, since randomly grouping the variables does not consider the relationship or connection of these variables, this
may reduce the possibility of QOX to generate the solutions
close to the global optimal solutions, and may degrade the
search efficiency of the evolutionary algorithm. To overcome
this drawback, we propose a variable grouping scheme which
considers the relationship or connection of these variables. As
a result, the solutions generated by QOX and the new grouping scheme are more possible to be close to the global optimal
solutions.
The variables can be grouped by a technique called
learning automata (LA) [21] which is derived from the
theory of probability and Markov process. LA model
was first proposed and developed in mathematical psychology by Bush and Mosteller [22], and then developed
by Atkinson et al. [23] and Tsetlin [24]. Recently,

c 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
2168-2267 
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

DAI et al.: OELA FOR MULTIOBJECTIVE OPTIMIZATION

LA method has been used in many engineering problems (e.g., control problem [25], data mining problem [26],
event patterns tracking [27], complex networks [28], function optimization [29], [30], vehicular sensor networks [31],
opportunistic [32], and economic emission dispatching [33])
to improve the performance of the algorithms, and
reinforcement learning has also been used in function
optimizations [34][37] to improve the performance of the
algorithms. Over the last two decades, LA method has also
been adopted to improve the performance of MOEAs. In [38],
LA is employed to divide the search space into separate
cubes. In this method, the multiobjective functions are aggregated into a single function by using a weighted-sum method.
However, because the method uses the weighted-sum method
to combine fitness functions, this method cannot find the solutions in nonconvex parts of the Pareto-optimal front. In [39],
LA and GAs are integrated to solve the MOP of generation dispatch in a six-bus power system. Liao and Wu [40]
proposed a method of multiobjective optimization by LA to
solve complex MOP. LA method is used to solve multiobjective generation dispatch in [41]. Horn and Oommen [42]
developed four LA-based algorithms to solve multiconstraint
assignment problems. In [43], LA method as a local search
technique is used in DE to solve numerical optimization
problems. In this paper, LA is employed to divide each
dimension into a certain number of parts, and each part
is called a cell. The value of a cell is updated based
on the current value of the cell and the number of current nondominated solutions whose corresponding component
locates in the cell. The reduction of the number of nondominated solutions will make the value of the cell become
smaller. The values of these cells are used to group variables
for QOX.
In this paper, an orthogonal evolutionary algorithm with
LA (OELA) for multiobjective optimization problems is presented. To be specific, an LA method, QOX, and a new fitness
function are used in this algorithm. LA is used to execute the
mutation operator and used as a tool to group the decision
variables for QOX. The resulted QOX is an improved QOX
which is more possible to generate good offspring than QOX,
and thus can enhance the performance of QOX and improve
the search efficiency of the algorithm. Moreover, a new fitness
function based on the decomposition of the objective space is
proposed and used to maintain the diversity and uniformity
of solutions and guide the obtained solutions approaching the
true Pareto front (PF).
The remainder of this paper is organized as follows:
Section II introduces the main concepts of the multiobjective
optimization problems, LA and QOX; Section III presents the
OELA, including the LA, QOX and the new fitness function;
while Section IV shows the experimental results of the proposed algorithm and the compared algorithms; Section V gives
the conclusion and future works.
II. P RELIMINARIES
In this section, the main concepts of the multiobjective
optimization problems, LA, and QOX are introduced.

3307

A. Multiobjective Optimization
A multiobjective optimization problem can be formulated
as follows [44]:

min F(x) = ( f1 (x), f2 (x), . . . , fm (x))


s.t.
gi (x) 0, i = 1, 2, . . . , q
(1)

hj (x) = 0, j = 1, 2, . . . , p
where x = (x1 , . . . , xn ) X Rn is called decision variable
and X is an n-dimensional decision space. fi (x)(i = 1, . . . , m)
is the ith objective to be minimized, gi (x)(i = 1, 2, . . . , q)
defines the ith inequality constraint and hj (x)( j = 1, 2, . . . , p)
defines the jth equality constraint. Furthermore, all the
constraints determine the set of feasible solutions which is
denoted by , and Y = {F(x)|x } is denoted as the objective space. To be specific, we try to find a feasible solution
x  to minimize each objective function fi (x)(i = 1, . . . , m)
in the vector-valued objective function F. Let x, z , x
is said to dominate z, if F(x) = F(z) and fi (x) fi (z) for
i = 1, 2, . . . , m. x is called Pareto-optimal if there is no other
x such that x dominates x . The set of all the Pareto-optimal
solutions is called the Pareto set (PS). The image of the PS on
the objective space, PF = {F(x)|x PS}, is called the PF [45].
B. Learning Automata
LA is used to execute the mutation operator and used as
a tool to group the decision variables for the QOX. In this
paper, LA consists of n automata, where n is the dimension of the optimization problem considered, and its role is
to search the optimal solutions on a specified dimension. The
ith automaton can be defined as i , Bi , Pi , where:
1) i = {xi } is the set of possible states on the ith dimension
and xi is the dimensional state on the ith dimension,
where xi [xmin,i , xmax,i ] with xmax,i and xmin,i being
the maximum and minimum values of the ith dimension,
respectively;
2) Bi = {bl, } denotes the set of possible actions which the
automaton can take on dimension i, i.e., bl, represents
an action to move a length on left path (when l = 1)
or right path (when l = 2) or own path (when l = 3),
where l represents a path label and length is no longer
than the path length. Also, each path will be defined
a path value, which represents the possibility of finding
a better solution if the automaton searches in this path.
This will be defined and explained later in detail;
3) Pi includes two probabilities p1 and p2 , where p1 represents the probability of path selection and p2 shows
the probability of moving the length of xi along the
selected path.
Because the dimensional states are continuous variables, the
search space contains an infinite number of possible states.
In order to achieve computational tractability, each dimension
can be divided into D cells. We denote the jth cell as ci,j of
the ith dimension, where i = 1, . . . , n, j = 1, . . . , D. Thus,
there are totally nD cells for an n-dimensional search space.
Moreover, the width of a cell in the ith dimension can be
computed by wi = (xmax,i xmin,i )/D, and the value of cell
ci,j is defined by the possibility of the ith component of an

3308

IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12, DECEMBER 2016

optimal solution locating in the cell ci,j and denoted by Vi,j .


The bigger the value of a cell is, the more possibility the
ith component of a solution locating in the cell will be an
optimal one.
1) Update Cell Values: In one generation, let S denote the
nondominated solution set of the current population and new
generated solutions, and the value of cell ci,j , denoted by Vi,j .
Liao and Wu [40] gave a method to update the values of Vi,j .
In this paper, the values of Vi,j is updated by the following
method:
 

(2)
Vi,j = 0.5V i,j + 0.5 x  xi ci,j , x S /|S|
where |S| indicates the cardinality of set S. As can be seen
from (2), the updated cell values are determined by the solutions already existed and newly generated, which means that
the information is used as more as possible. If one cell remains
few or no nondominated solutions for several generations, the
cell value of it will decrease fast, which will result in the cell
gradually losing its favor of being selected in the future search.
In this paper, cell values are used to guide the further search.
Moreover, cell values are also used to group the variables for
the QOX which will be described in the next section.
2) Path Value: Before a search action is implemented, we
have to estimate the probability of selecting the path, while
calculating this probability needs the information of the path
values [30], [40].
For a given solution x = (x1 , . . . , xn ), its ith dimensional
state xi will locate in a cell ci,j , where j = ceil((xi xmin,i )/wi )
and ceil(z) denotes the smallest integer greater than real
value z. After determining ci,j for xi , the three possible path
values, denoted by Ll,i , l = 1, 2, 3, represent the possibilities
of three possible actions (i.e., moving to the left path, moving
to the right path, staying at the current cell, respectively) and
are defined as follows:


(3)
L1,i = max Vi,k |1 k < j


L2,i = max Vi,k | j < k D
(4)
L3,i = Vi,j .
(5)
The bigger the value of Ll,i , the more possibility the ith dimensional state xi moves along the corresponding path. As can be
seen from these three formulas, the path value mainly depends
on the largest cell value in the ith dimension. Thus it enhances
the possibility that xi moves toward the cell with the largest
value.
3) Mutation Based on Cell Selection: Suppose that the
selection probability of mutation is pm . For a chosen solution
x = (x1 , . . . , xn ), the mutation operator is based on action
selection in this algorithm. For each dimension, action selection employs two probabilities p1 and p2 to make the state
variation, where p1 is used to select a path along which the
current dimensional state xi moves, while p2 is used to select
the cell in which xi will moves. Each path of three paths (i.e.,
the left path L1 , the right path L2 , and the own path L3 ) is
exclusively chosen based on the probabilities as follows:

eLl,i /a
p1 Ll,i =
3
, l = 1, 2, 3
Ll,k /a
k=1 e

(6)

Fig. 1. Three possible paths taken by a search starting at dimensional state


xi on the ith dimension.

where a > 0 is a parameter chosen to balance between


exploitation and exploration [46]. The bigger the parameter
is, the closer the selection probabilities of different paths
are. Conversely, a smaller parameter makes a greater difference in the selection probabilities of the different paths.
In this paper, since the main aim of the cell selection is
to carry out the mutation operation, the parameter is set to
a small value. Suppose that the left path is selected by the
roulette wheel selection among three paths, the probability p2
of xi moving to a specific cell ci,k in this path is calculated
as follows:

e0.5Vi,k /a
, 1 k < j.
p2 ci,k =

0.5Vi,s /a
1s<j e

(7)

When a cell of ( j 1) cells in the left path is selected by


the roulette wheel selection based on the probabilities, an
action is taken by moving the current dimensional state xi
to a random point (obeying uniform distribution) inside the
selected cell [denoted by the j0 th cell ( j0 {1, 2, . . . , D})].
Though (6) and (7) are the same as [30] and [40], we use
xi
= ( j0 1 + ) wi + xmin,i to generate the ith component
of an offspring of mutation, which is intuitively shown in
Fig. 1, where (0, 1] is a random number.
This mutation operator based on cell selection has the
following two properties.
1) High Search Efficiency: Since the bigger value of the
cell is, the greater selection probability of the cell will
be. Thus, offspring generated by the mutation operator have larger probability to locate in the regions of
nondominated solutions, which can improve the search
efficiency.
2) The Ability of the Global Search: For a current dimensional state xi , each cell of the ith dimension has the
chance of being chosen such that xi will move to
a new point of the chosen cell, which indicates that
the mutation operator has the ability of the global
search.
C. Quantization Orthogonal Crossover
In this paper, we propose a QOX with LA (QOXLA)
as a crossover operator to improve the performance
of the algorithm. The orthogonal design was developed
for generating the best combination of levels [47], [48].
Orthogonal crossover (OX) operator was first designed by
Zhang and Leung [17] for combinatorial optimization problem. Then, Leung and Wang [18] combined the quantization
technique with OX to design an QOX for numerical optimization problems with continuous variables [18]. To be specific,
given any two solutions x = (x1 , . . . , xn ) and z = (z1 , . . . , zn ),

DAI et al.: OELA FOR MULTIOBJECTIVE OPTIMIZATION

a domain [lb, ub] determined by these two solutions can be


defined as follows:

lb = [min(x1 , z1 ), . . . , min(xn , zn )]
(8)
ub = [max(x1 , z1 ), . . . , max(xn , zn )].
The domain of [lb, ub] is quantized into Q levels such that
the difference between any two successive levels is the same.
Specifically, the domain of the ith dimension is quantized into
i,1 , . . . , i,Q where

lbi ,

 j=1

|xi zi |
, 2 j Q 1 (9)
i,j = lbi + ( j 1)

Q1

ubi ,
j=Q
where ubi = max(xi , zi ) and lbi = min(xi , zi ). In practice,
dimension n is often large and so is the scale of LM (Qn ).
However, each pair of parents should not produce too many
potential offspring in order to avoid a large number of function evaluations. For this purpose, the variables x1 , . . . , xn are
divided into G groups, where G is small and each group is
treated as one factor. In this way, the corresponding orthogonal array has a small number of combinations, and a small
number of potential offspring are generated.
III. O RTHOGONAL E VOLUTIONARY A LGORITHM
W ITH L EARNING AUTOMATA
OELA mainly consists of three parts: QOX with LA, and
the new fitness function, which will be introduced one by one
in this section.
A. Quantization Orthogonal Crossover
With Learning Automata
In [18], the variables (components) are sequentially and
randomly grouped, which does not consider the relationship
or connection of these variables. This may reduce the possibility of QOX to generate the solutions close to the global
optimal solutions, and may degrade the search efficiency of
the evolutionary algorithm. Note that LA can provide the
information that the biggest-value cell of each variable has
the most nondominated solutions. If the variables whose the
jth( j = 1, . . . , Q) levels ai,j simultaneously locate in the
biggest-value cells are grouped into a group. Many components of offspring generated by the QOX will simultaneously
locate in the biggest-value cells, which may make these offspring more likely be nondominated solutions. Thus, if we
use LA in this way by considering the relationship or connection of variables and combine LA with QOX, the search
efficiency can be improved. The detail of using LA is as follows. First, for each dimension, we find the biggest-value
cell, then these n cells in all dimensions form a vector
denoted as C = (c1,j1 , . . . , cn,jn ), where ci,ji {ci,1 , . . . , ci,D },
i = 1, . . . , n, and ji is the number of the cell with the
biggest cell value in the ith dimension. For a given pair of
solutions x and z, we get lb = [min(x1 , z1 ), . . . , min(xn , zn )]
and ub = [max(x1 , z1 ), . . . , max(xn , zn )] which are the lower
and upper bounds of x and z, respectively. To facilitate
grouping, a matrix R = [ri,j ]nQ (where i indicates the ith

3309

variable and j indicates the jth levels ai,j of the ith dimension, i = 1, . . . , n, j = 1, . . . , Q) is defined by the following
formula:

1
if ai,j ci,ji
(10)
ri,j =
0
otherwise
where ri,j = 1 indicates that the jth levels ai,j belongs to
the biggest-value cell of the ith dimension ci,ji ; ri,j = 0 indicates that ai,j does not belong to ci,ji . The column of R with
most elements 1 is first found, and then the rows (dimensions) corresponding to these elements 1 will be assigned to
a group. By repeating this process, the variables can be iteratively grouped. Suppose Q1 groups have been made at last.
There are two cases: 1) there is no ungrouped variable and
2) there are some ungrouped variables. For the first case, all
variables are divided into Q1 groups and let G = Q1 . For
the second case, the remaining variables will be randomly
grouped into (G Q1 ) and all variables will also be divided
into G groups. This grouping scheme is to make as many components of offspring generated by QOX as possible locate in C,
and enforce these offspring more likely to locate close to the
Pareto-optimal solutions. After grouping, the orthogonal array
LM (QG ) is used to choose a sample of M chromosomes as the
potential offspring. The improved QOX is denoted as QOXLA.
Example: Consider a 5-D optimization problem, let
the two parents be x = (0, 4, 2, 0, 1) and z =
(6, 1, 5, 3, 2), and they define the solution space [l, u] =
[(0, 1, 2, 3, 1), (6, 4, 5, 0, 2)]. We set G = 3, Q = 3, and
C = ([2.5, 4], [0, 2], [4, 6], [4, 1], [1.1, 1.9]). According
to (10), the matrix R can be calculated and the result is as
follows:

0 1 0
1 0 0

R=
0 0 1 .
1 0 0
0 1 0
Through the above grouping scheme, these five variables
are divided into three groups, i.e., (x1 , x5 ), (x2 , x4 ), and (x3 ).
By applying L32 (33 ), we can get the following nine potential
offspring: (0, 1, 2, 3, 1), (0, 2.5 3.5, 1.5, 1), (0, 4, 5, 0, 1),
(3, 1, 3.5, 3, 1.5), (3, 2.5, 5, 1.5, 1.5), (3, 4, 2, 0, 1.5),
(6, 1, 5, 3, 2), (6, 2.5, 2, 1.5, 2), and (6, 4, 3.5, 0, 2).
Since four components of both offspring (3, 1, 3.5, 3, 1.5)
and (3, 2.5, 5, 1.5, 1.5) locate in C, they are likely to locate
close to the Pareto-optimal solutions.
Comparing with other adaptive approaches (such as adaptive DE [7] and adaptive particle swarm optimization [5]),
the main advantages of the QOXLA are as follows:
1) QOXLA can predict the potential nondominated area more
exactly; 2) QOXLA can generate better diversity of nondominated solutions; 3) QOXLA takes the correlation of variables
into account; and 4) parameters of QOXLA are easily set.
The space complexity and the computational complexity of
QOXLA are O(2N n) and O(N n), respectively; the space
complexity and the computational complexity of adaptive
DE [7] are O(n) and O(N), respectively; the space complexity and the computational complexity of adaptive particle

3310

IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12, DECEMBER 2016

swarm optimization [5]) are O(3N n) and O(N), respectively (where N is the size of population). Thus, the cost of
QOXLA is larger than that of other two adaptive methods.
B. New Fitness Function Based on the Decomposition
of the Objective Space
A critical and difficult aspect of MOEAs is to find a suitable fitness function. The fitness value of each solution should
indicate the capability of various features of this member to
survive. In this paper, we propose a new fitness function which
is based on the decomposition of the objective space and is
helpful to maintain the diversity of the obtained solutions and
guide the obtained solutions converging to the true PF. The
objective space is uniformly divided into subspaces by a set
of weight vectors which are evenly distributed. The solutions
corresponding to one subspace are classified into one set by
the following equations:





 

i
i
j

,
Z = x  x POP,  F(x), = max  F(x),
1 jN

i = 1, . . . , N
(11)






i
j
Yi = F(x)|x ,  F(x), = max (F(x), )
1 jN

(12)
where POP = {x1 , x2 , . . . , xK1 } is the current population
with K1 solutions, W = {1 , 2 , . . . , N } is a set of weight
vectors, and (F(x), i ) is a specific aggregation function
which will be given in the following. For a given solution, suppose that the value (F(x), i ) is the largest among
{(F(x), j ), j = 1, . . . , N}, then the solution x belongs to the
set Z i . These K1 solutions are divided into N classes by (11)
and the objective space Y is divided into N subobjective spaces
Y1 , . . . , YN . To illustrate the idea intuitively, we give an example shown in Fig. 2. The objective space is evenly divided
into four parts by four weight vectors. The weight vector
1 corresponds to the subobjective space surrounded by the
space ABGF. The aggregation function plays a very important role in this model which directly determines whether the
objective space is evenly divided. The elliptic function and
ellipsoidal function which have a good performance in capturing nonconvex fronts [49] are used as the aggregation function
for two-objective and three-objective problem, respectively.
Consider an ellipse centered at origin whose semi major is
= (1 , 2 ), its elliptic function can be written as

g(F(x), ) = ( f1 (x)1 + f2 (x)2 )2 A2 + ( f1 (x)2 f2 (x)1 )2
(13)
where
A > 1, A is the major semi-axis of the ellipse
and A2 1/A is the eccentricity of the ellipse. An ellipsoid function whose center and semi major are origin and
= (1 , 2, 3 ), respectively, can be written as

g(F(x), ) = ( f1 (x)1 + f2 (x)2 + f3 (x)3 )2 A2
+ ( f1 (x)2 f2 (x)1 )2



2
+ ( f 2 (x)2 3 + f1 (x)1 3 12 + 22 f3 (x)) .
(14)

Fig. 2. Objective space (space ABCDEKIHGF) is evenly divided into four


parts [Y1 , Y2 , Y3 , Y4 (e.g., Y1 is the space ABGF)] by four weight vectors
(1 , 2 , 3 , and 4 ). (The PF is the curve ABCDE.)

For m-objective problems (m 2), a hyper-elliptic function


whose center and semi major are origin and = (1 , . . . , m ),
respectively, can be written as
g(F(x), ) =

 m


2 
i fi (x)

i=1

fi
(x) =


m

i1

A2 +


m

( fi
(x))2
i=2

j fj (x) fi1 (x)

j=i

m fm1 (x) m1 f m (x),

j=i

j2 , if i = 2, . . . , m 1
if i = m.

(15)
For two and three objective optimization problems, we
adopt g(F(x), i ) in (13) and (14) as (F(x), i ), respectively. When the solutions are classified by using (14), if a set
Z i (i = 1, . . . , N) is not empty, the fitness function value of
each solution x in Z i is calculated by the following formula:
 


fit(x) = Z i  + +  F(x), i
(16)
where x Z i ; if x is a nondominated solution of Z i , is equal
to 0, otherwise, is equal to 1; |Z i | indicates the number of
solutions in set Z i ; is a small positive constant to make
(F(x), i ) < 1. In this paper, our main objectives are to
improve the convergence and maintain the diversity by minimizing the fitness value of solutions. In particular, minimizing
|Z i | to make each subobjective space have more than one
solution can improve the diversity of the solutions of POP
in objective space, and minimizing to keep the nondominated solutions of Z i can improve the convergence. Moreover,
minimizing (F(x), i ) to make the angle between the objective vector of solution x and the weight vector i close to
0 can make the solutions in POP distribute more evenly in
objective space.
C. Selection Strategy
If an algorithm makes use of the neighborhood information of a solution to generate offspring, the quality of the
offspring will usually be better and the convergence to a good
PF will be speed up [50]. For this consideration, a method
of clarifying the neighborhood of the nondominated solutions
should be designed. In this paper, because the objective space
is divided into some subspaces by weight vectors, and the
decision space is accordingly divided into some subspaces.
Each subspace can be seen as a neighborhood. After that, for

DAI et al.: OELA FOR MULTIOBJECTIVE OPTIMIZATION

3311

each weight vector i , its T nearest neighbor weight vectors


are found, and the set of these T weight vectors are denoted
as NBi (i = 1, . . . , N). Then, for the subobjective space determined by weight vector i , we find the best solution by the
following formula:


 
(17)
xisos = argmin  F(z), i , z Zi .
Afterwards, the following procedure will be repeated NM
times to generate NM offspring of crossover: for each time,
one best solution among all xisos , still denoted by xisos , is first
selected by the binary tournament selection based on the fitness function (the smaller the fitness value of a solution is,
the bigger the probability of the solution will be selected),
and then its two neighborhood subobjective spaces, e.g., corresponding to weight vectors r1 and r2 in NBi , are randomly
selected. The best solutions in these two subdecision spaces,
r2
i
denoted by xr1
sos and xsos , respectively, as well as xsos , are
selected as parents for crossover. And a part of offspring is
generated by the following formula:


r2
(18)
x
xnew = xisos + 0.5 xr1
sos
sos .
i , two subobjective spaces
Note that for each solution xsos
corresponding to weight vectors r1 and r2 are randomly
selected, thus the offspring generated by (18) will have a good
diversity. Another part of offspring is generated by QOXLA as
j
i
and xsos )
follows: a pair of different solutions (e.g., xsos
are selected by the binary tournament selection as the parents. Then QOXLA is used to them to generate M offspring.
These offspring will also have a good diversity. Therefore,
the proposed algorithm will have a good global search
ability.

D. Update Strategy
The elite strategy is used in the algorithm when updating the
population. The update strategy uses the( + ) type of deterministic replacement where indicates the size of the parent
population and indicates the size of the descendant population. In this replacement strategy, the parent and descendant
populations are combined, then we select best individuals
which are kept to the next generation. This means that
individuals will be removed from the parent and descendant
populations. In this paper, and are equal to N. Deletion
rules are as follows.
Step 1: Let k = 1; if k > , stop; otherwise go to step 2.
Step 2: Find a set Z i whose size is the maximum among
{Z i (i = 1, . . . , N)}. Find a solution x Z i whose
(F(x), i ) is the maximum among Z i . Delete the
solution x from set Z i , then let Z i = Z i \x and k =
k + 1. Go to step 1.
E. Procedure of the Proposed Algorithm
Based on all above, a novel OELA is proposed. The
flowchart of OELA is shown in Fig. 3 and the procedure of
OELA is as follows.

Fig. 3.

Flowchart of OELA.

1) Initialization: Given N weight vectors (1 , 2 , . . . , N )


and let t = 0. Randomly generate an initial population POP(t)
with size N. Let oi = min{ fi (x)|x POP(t)}, 1 i m, and
o = (o1 , . . . , om ) be the center of the elliptic functions or
ellipsoidal functions. For each j {1, . . . , N}, find the index
set NBj = { j1 , . . . , jT } of the T closest weight vectors to j .
Let Vi,j = 1(i = 1, . . . , n, j = 1, . . . , D).
2) Optimization Loop: If the stop condition is not satisfied,
repeat the following.
First (generate new solutions), calculate the fitness value
of each solution in POP(t) by (10)(14), and set a temporary population pop with size 0. Apply the selection strategy
in Section III-C on POP(t) to select some parents, and use
these parents to generate offspring by (18) and QOXLA, then
these offspring further undergo the mutation by mutation operator in Section II-B3 and the offspring of mutation are put
into pop.
Second (update population), for each j = 1, . . . , m, if
oj < min{ fj (x)|x pop}, update oj as min{ fj (x)|x pop}; then
apply (2) on POP(t) pop to update Vi,j (i = 1, . . . , n; j =
1, . . . , D) and utilize the update strategy in Section III-D
on POP(t) pop to generate the next generation population
POP(t + 1). Let t = t + 1.
Else, stop and output the final N solutions of POP(t).
End if.
Notes:
1) In OELA, (18) is used as a main crossover operator to
generate offspring, while the QOXLA is only used to
generate a small number of offspring.
2) For problems whose variables are separable or independent, the group strategy for variables does not affect the
quality of the offspring generated by crossover operator QOXLA, thus, OELA can well deal with these
problems. For problems whose variables are not separable, first, Li and Zhang [15] solved some problems very
well with complicated PSs and nonseparable variables

3312

IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12, DECEMBER 2016

by using the DE [54], (18) is the variant of DE [15]


which is also effective for problems with nonseparable variables; second, QOXLA does not divide variables
into several groups randomly, instead, it groups variables
by considering the relationship or connection of variables. Thus, offspring generated by QOXLA are more
likely to locate close to the Pareto-optimal solutions,
and QOXLA can effectively handle the problems with
variable linkages; third, the selection strategy adopted is
helpful to enhance the global and local search abilities
of crossover. Thus, OELA can solve the problems with
either variable linkages or independent.
3) Complexity Analysis: In the proposed algorithm OELA,
2N solutions, Vi,j (i = 1, . . . , n; j = 1, . . . , D), NBi (i =
1, . . . , N) and the indices of solutions in Z i (i = 1, . . . , N) need
to be stored, thus the space complexity of OELA is O(2N
n) + O(n D) + O(N T) + O(N 2N) = O(N N) (in this
paper, n, T, D < N). Thus the space complexity is O(N 2 ). The
major computation in this algorithm involved is in the update
step of OELA. To update Vi,j (i = 1, . . . , n; j = 1, . . . , D),
O(N n) basic operations (i.e., +, , , , and comparison) are needed. To classify the population POP(k) pop,
O(2N N) basic operations are needed. To generate the next
generation population, at most O(N 2N) basic operations are
needed. Totally, the computational complexity of the algorithm
is O(N n) + O(2N N) + O(N 2N) = O(N 2 ). Moreover,
the proof of convergence of the proposed algorithm is put in
the supplementary material.
IV. E XPERIMENTAL S TUDY
In this section, OELA will be compared with six well
known algorithms: 1) decomposition-based multiobjective
evolutionary algorithm (MOEA/D [15]); 2) decompositionbased multiobjective evolutionary algorithm with an ensemble
of neighborhood sizes (MOEA/D-ENSs [16]); 3) nondominated sorting GA II (NSGAII [3]); 4) NSGAIII [11];
5) multiobjective optimization by LA (MOLA [40]); and
6) multiobjective immune algorithm with nondominated
neighbor-based selection (NNIA [9]) through experiments.
The experiments are conducted on 15 widely used and challenging enough multiobjective benchmark problems.
A. Test Problems
Among 15 benchmark problems, the first ten problems
F1F10 are chosen from CEC09 competition test suite [51].
These test problems have complicated PS shapes, and it has
been proved that these complicated PS shapes, as well as the
geometrical shapes of the PF, can cause much difficulty for
MOEAs [53]. Two test problems (DTLZ1 and DTLZ3) that
are marked as F11 and F12 here are chosen from the well
known DTLZ family [52]. The global PFs of these two problems are known analytically and they have 11nm+1 1 and
3nm+1 1 local PFs, respectively [52]. In addition, the three
three-objective problems (i.e., F11, F12, and F8) are extended
to five-objective problems which are marked as F13F15.
Moreover, the search spaces of test problems F1F15 are
[0, 1][1, 1]29 , [0, 1][1, 1]29 , [0, 1]30 , [0, 1][2, 2]29 ,

[0, 1][1, 1]29 , [0, 1][1, 1]29 , [0, 1][1, 1]29 , [0, 1]2
[2, 2]28 , [0, 1]2 [2, 2]28 , [0, 1]2 [2, 2]28 , [0, 1]20 ,
[0, 1]20 , [0, 1]20 , [0, 1]20 , and [0, 1]4 [2, 2]26 , respectively.
These fifteen benchmark problems are challenging enough
for MOEAs.
B. Parameter Settings
The experiments are carried out on a personal computer (Intel Xeon CPU 2.53 GHz, 3.98 GB RAM).
The solutions are all coded as real vectors. Polynomial
mutation [44] operators are applied directly to real vectors
in four algorithms: 1) NSGAII; 2) MOEA/D; 3) NNIA; and
4) MOEA/D-ENS. For crossover operators, simulated binary
crossover (SBX [44]) is used in NSGAII and NNIA, and
DE [54] is used in MOEA/D. The parameter configuration in
this paper is as follows: distribution index is 20 and crossover
probability is 1 in SBX operator; crossover rate is 0.5 and
scaling factor is 0.5 in DE operator; the parameter of the fitness function is set to 0.01; distribution index is 20 and
mutation probability is 0.1 in polynomial mutation operator;
quantization levels Q = 3 and there are G = 4 factors in
QOXLA operator, so that each pair of parents can produce
nine potential offspring (M = 9); parameters in LA: D = 20,
and the parameters of (6) and (7): = 0.2.
Population size of the algorithms is set to 105 for all test
problems on two or three objective problems, and 210 for fiveobjective problems. Weight vectors are generated by using the
method in [57] for MOEA/D and OELA. The number of the
weight vectors in the neighborhood in OELA are set to 10 for
two or three objective problems and 20 for five objective problems, respectively. The Techebycheff approach [57] is used for
MOEA/D as an aggregate function, while the elliptic function
and ellipsoidal function are used for OELA as an aggregate
function for two-objective problems, three-objective problems,
and five-objective
problems, respectively, and the eccentricities are set to 99/10. For NNIA, the size of active population
is 20 and the clone scale is 20. Each algorithm runs 20 times
independently on each test instance. The maximal number of
function evaluations is set to 100 000 for all test problems.
C. Performance Measures
In this paper, the following four performance metrics are
used to evaluate the performance of the different algorithms
quantitatively: 1) generational distance (GD) [55]; 2) inverted
GD (IGD) [55]; 3) hyper-volume indicator (HV) [58]; and
4) Wilcoxon rank-sum test [56]. GD measures how far the
known PF is away from the true PF. If GD is equal to 0, all
points of the obtained PF belong to the true PF. GD allows
us to observe whether the algorithm can converges to some
region in the true PF. IGD measures how far the true PF is
away from the obtained PF. If IGD is equal to 0, the obtained
PF contains every point of the true PF. IGD shows whether
points of the obtained PF are evenly distributed throughout the
true PF. Here, GD and IGD indicators are used simultaneously
to observe whether the solutions are distributed over the entire
PF. The HV measures both the convergence and diversity of
the obtained solutions. For the problem DTLZ1 (F11 and F13),

DAI et al.: OELA FOR MULTIOBJECTIVE OPTIMIZATION

3313

TABLE I
IGD, GD, AND HV O BTAINED BY OELA, MOEA/D, NSGAII, AND NNIA ON F1F10

the reference point is set to (0.5, . . . , 0.5); for other problems,


the reference point is set to (1, . . . , 1)
Wilcoxon rank-sum test [56] is used in the sense of statistics to compare the mean IGD, GD, and HV of the compared
algorithms. It tests whether the performance of OELA on each
test problem is better (+), same (=), or worse ()
than/as that of the compared algorithms at a significance
level of 0.05 by a two-tailed test. These results are given in
Tables I and II.
D. Comparisons of OELA With MOEA/D,
NSGAII, and NNIA
Tables I and II present the best, median, and worst values of
IGD, GD, and HV obtained by OELA, MOEA/D, NSGAII,

and NNIA, respectively. The best results obtained are highlighted bold in these tables. It can be seen from Tables I and II
that, for IGD metric, OELA outperforms MOEA/D and
NNIA on all 15 test problems, and outperforms NSGAII
on 14 test problems. NSGAII has the same performance as
OELA only on problem F3. For GD metric, OELA outperforms NSGAII on 13 test problems (F1F4, F6F8, and
F10F15), performs the same as NSGAII on one problem F9,
performs worse than NSGAII only on problem F5. OELA outperforms MOEA/D on 11 problems and performs same as
MOEA/D on two problems (F3 and F9), while performs worse
than MOEA/D on only two problems F1 and F5. OELA outperforms NNIA on 14 problems, and performs worse only on
problem F5. For HV metrics, OELA outperforms MOEA/D
and NNIA on 14 test problems, outperforms NSGAII on

3314

IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12, DECEMBER 2016

TABLE II
IGD, GD, AND HV O BTAINED BY OELA, MOEA/D, NSGAII, AND NNIA ON F11F15

13 test problems, while performs the same as NSGAII on two


test problems F5 and F9, performs the same as MOEA/D on
problem F6, and performs the same as NNIA on problem F5.
These results indicate OELA outperforms all three compared
algorithms.
From Tables I and II, we can see that the median values
of IGD obtained by OELA are smaller than those obtained
by other three algorithms for all fifteen problems, especially,
e.g., for each problem, the median value of IGD obtained
by OELA is smaller at least 32% than that obtained by
MOEA/D. These indicate that for these test problems, the
solutions obtained by OELA have better coverage than those
obtained by other three algorithms. For GD metric, we can
also see that the median values of GD obtained by OELA are
smaller than those obtained by NSGAII and NNIA on all test
problems except for F5, and the median value of GD obtained
by OELA is smaller than that obtained by MOEA/D on all test
problems except for F1 and F5. These illustrate that the solutions obtained by OELA have a better convergence than those
obtained by other three compared algorithms. For HV metric, the median values of HV obtained by OELA are bigger
than those obtained by other three algorithms, and are better at least 2% than those obtained by other three algorithms.
These demonstrate that the solutions obtained by OELA have
better diversities than those obtained by MOEA/D, NSGAII,
and NNIA. In addition, especially for test problems F9 and
F10, they have 1118 1 and 318 1 local PFs, respectively.
All local PFs are parallel to the global PF and an MOEA can
get stuck at any of these local PFs before converging to the
global PF [52].
To visually compare the performance of the four algorithms,
the solutions obtained by them on these test problems are
shown in Fig. 4. From this figure, and the values of IGD,

we can see that for two-objective problems F1F7, OELA can


maintain a good diversity of solutions on the PF, and there
are no gaps (i.e., concave corners) existing on the obtained
PF except for F5 and F6, however, there are some gaps existing on the PF obtained by the NSGAII, NNIA, and MOEA/D
on problems F4F7; for the three-objective problems F8F12,
OELA can maintain a good diversity of solutions which
converge to the true PF on F8, F9 and F11, F12, while
the diversities and convergences of solutions obtained by
NSGAII, NNIA, and MOEA/D on problems F8F12 are
not good.
From Fig. 4 and the values of IGD, GD, and HV, we can
obtain that only OELA can well converge to the true PF and
maintain well diversity of the obtained solutions for F11, F12,
while other three algorithms cannot well converge to the true
PF for these two problems. Moreover, for three five-objective
problems F13F15, the performance of OELA is also better
than that of MOEA/D, NSGAII, and NNIA.
E. Comparisons of OELA With MOEA/D-ENS,
MOLA, and NSGAIII
MOEA/D-ENS is an improvement of MOEA/D.
MOEA/D-ENS uses an ensemble of different neighborhood
sizes with online self-adaptation to improve the performance
of MOEA/D. MOLA is a novel method of multiobjective
optimization by LA. In order to compare OELA with
MOEA/D-ENS and MOLA, we used F1F10, DTLZ1 and
DTLZ2 as the test problems. To make a fair comparison
and use the results of MOEA/D-ENS [16] and MOLA [40]
directly, population size of OELA is set to 600 and
1000 for two and three objectives problems, respectively,
and the maximal number of function evaluations is set to

DAI et al.: OELA FOR MULTIOBJECTIVE OPTIMIZATION

Fig. 4.

3315

Solutions obtained by MOEA/D, NSGAII, NNIA, and OELA on F1F12.

300 000 for these ten test problems, moreover, the experimental results of MOEA/D-ENS and MOLA are obtained
directly from [16] and [40], respectively.

The mean and standard deviation values of IGD obtained


by OELA and MOEA/D-ENS are shown in Table III. Table III
shows that for problems F1 and F9, the mean values of IGD

3316

IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12, DECEMBER 2016

TABLE III
IGD O BTAINED BY OELA, MOEA/D-ENS, AND MOLA

obtained by OELA are slightly larger than those obtained by


MOEA/D-ENS, which indicates that for problems F1 and F9,
the solutions obtained by MOEA/D-ENS have better coverage
than those obtained by OELA. However, for the remaining
eight problems, the mean values of IGD obtained by OELA are
smaller than those obtained by MOEA/D-ENS, which indicate that for problems F2F8 and F10, the quality of the
final solutions obtained by OELA is significantly better than
that obtained by MOEA/D-ENS. It also can be seen from
Table III that, for problems F1F9, DTLZ1 and DTLZ2, the
mean values of IGD obtained by OELA are smaller than those
obtained by MOLA except for problem F9, which indicate
that the quality of the final solutions obtained by OELA is
significantly better than that obtained by MOLA on almost
problems.
Moreover, reference-point-based nondominated sorting
approach is also used to compare with OELA on test problems DTLZ1 and DTLZ2. For DTLZ1 and DTLZ2, the
median IGD values obtained by NSGAIII are 1.38 103
and 1.307 103 [11], respectively; the median IGD values obtained by OELA are 1.21 103 and 2.42 103 ,
respectively. These indicates that both NASGIII and OELA are
good at maintaining the diversity of the obtained solutions for
continue problems.

F. Major Contribution of LA to OELA


The major role of LA in the proposed algorithm OELA is
to provide information to group variables for QOX. The
improvement QOX (QOXLA) can enhance the exploitation of
OELA. To identify this, Fig. 5 shows the evolution of the average IGD and GD values of the current population obtained by
OELA, OELA2, and NFE on problems DTLZ1 and DTLZ3
(where, when QOXLA in OELA is replaced by QOX, the
resulted algorithm is denoted as OELA2, and OELA2 without QOX is denoted as NFE). These results indicate that
OELA converges, in terms of the number of the function
evolutions, much faster than OELA2 and NFE in minimizing the IGD and GD values on DTLZ1 and DTLZ3, and

OELA2 converges much faster than NFE. The result of these


comparisons indicates that, QOX can enhance the exploitation
of the proposed algorithm, and QOXLA has better exploitation than QOX. In other words, LA can help QOX to improve
the exploitation.

G. Sensitivity Analysis to Parameters


In LA, the number D of cells in each dimension is a major
parameter, and the number T of neighbors and the major semiaxis of the ellipse (ellipsoidal) are two main control parameters
of the proposed algorithm OELA. For QOX, the parameter
Q is a major parameter, however, big values of Q will cause
a large number of function evaluations. So we do not use other
value of Q.
Note that if the value of D is too small, LA cannot effectively group the variables, while if the value of D is too large,
LA will cost much space and computation; If the value of
T is too small, the global search ability of OELA will be
reduced, and if the value of T is too large, the local search
ability of OELA will be reduced and OELA will make a lot
of comparisons; if the value of the major semi-axis is too
small, OELA cannot solve the nonconvex problems well, and
if the value of the major semi-axis is too large, OELA may
result in many local optimal solutions. Thus, it is reasonable
to set a range for the values of these parameters in the experiments of the sensitivity analysis. To test the sensitivity to the
parameters D, T and the major semi-axis values of the ellipse
and ellipsoidal, respectively, we set the major semi-axis values of the ellipse and ellipsoidal from 2 to 50 on a typical
problem F1, the number of cells D in each dimension from
5 to 39, and the number of neighbors T from 5 to 40, respectively. Other parameters are same as those in Section IV-B. The
ranges of parameters adopted above are large enough to test
the sensitivity. Fig. 6 shows the variations of the values of
IGD with the different major semi-axis values, different values of D and T on a typical problem F1. It can be seen from
Fig. 6 that, for different values of the three parameters, the
values of IGD only have a little bit variation. This indicates

DAI et al.: OELA FOR MULTIOBJECTIVE OPTIMIZATION

3317

Fig. 5.

Evolution of the mean of GD and IGD values on problems DTLZ1 and DTLZ3.

Fig. 6.

(a) Sensitivity of the major semi-axis of the ellipse on F1. (b) Sensitivity of T on F1. (c) Sensitivity of D on F1.

that the proposed algorithm OELA is not sensitive to the three


parameters.
H. Discussion
Through the above numerical experiment results, we can
get that the performance of OELA is better than NSGAII,
NNIA, MOEA/D, MOEA/D-ENS, and MOLA on most problems. Especially, for problems with many local PFs (i.e.,
DTLZ1 and DTLZ3), the convergence speed of OELA is much
faster. The main reasons of OELA with good performance are
that: 1) LA improves the performance of QOX; 2) LA as the
mutation operator enhances the local search and convergence
ability; 3) QOXLA can predict the potential nondominated
area more exactly, generate better diversity of nondominated
solutions, and take the correlation of variables into account;
4) the new fitness function maintains the diversity and helps
to converge to the true PF; and 5) the selection strategy helps
crossover operator (18) to make the global search and the local
search.
However, if the PF is not a continuous curve of surface
(i.e., it consists of several subcurves or subsurfaces), the proposed algorithm may not find solutions at the boundary of the
subcurves or subsurfaces.
V. C ONCLUSION
This paper designs an OELA which consists of a new
fitness function based on the decomposition of the objective space and a QOX with LA. The goal of the proposed
algorithm is to try to find an optimal solution for each subobjective space. To this end, the fitness function which uses the

weight vector and aggregate function is proposed to evenly


decompose the objective space so that the diversity and the
coverage of the obtained solutions to the true PF can be
well maintained. In addition, a QOX with learning automaton which improves the search efficiency by using the past
information is also utilized in the algorithm, where the variables are purposely grouped by using the past information so
as to quickly find optimal solutions. The experiment results
show that, with these schemes, the proposed algorithm is able
to search well in continuous domain, and achieve relatively
accurate PSs and wide PFs efficiently. Moreover, compared
with six well known algorithms MOEA/D, MOEA/D-ENS,
NSGAII, NSGAIII, MOLA, and NNIA, the simulation results
show that OELA is able to find much better spread of solutions and these solutions have better convergence to the true
PF. Especially for test problems with many local PFs (e.g.,
DTLZ1 and DTLZ3), OELA can more quickly converge to
the true PF than the compared algorithms. In the future, LA
and QOX will be used for solving many-objective optimization
problems.

R EFERENCES
[1] C. A. C. Coello, D. A. Van Veldhuizen, and G. B. Lamont, Evolutionary
Algorithms for Solving Multi-Objective Problems. New York, NY,
USA: Kluwer, May 2002.
[2] A. M. Zhou et al., Multiobjective evolutionary algorithms: A survey
of the state of the art, Swarm Evol. Comput., vol. 1, no. 1, pp. 3249,
Mar. 2011.
[3] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, A fast and elitist
multiobjective genetic algorithm: NSGA-II, IEEE Trans. Evol. Comput.,
vol. 6, no. 2, pp. 182197, Apr. 2002.

3318

[4] N. Chen et al., An evolutionary algorithm with double-level archives


for multiobjective optimization, IEEE Trans. Cybern., vol. 45, no. 9,
pp. 18511863, Sep. 2015.
[5] M. Kiani and S. H. Pourtakdoust, State estimation of nonlinear
dynamic systems using weighted variance-based adaptive particle swarm
optimization, Appl. Soft Comput., vol. 34, pp. 117, Sep. 2015.
[6] B. Xue, M. J. Zhang, and W. N. Browne, Particle swarm optimization
for feature selection in classification: A multi-objective approach, IEEE
Trans. Cybern., vol. 43, no. 6, pp. 16561671, Dec. 2013.
[7] Q. H. Lin et al., A novel hybrid multi-objective immune algorithm with
adaptive differential evolution, Comput. Oper. Res., vol. 62, pp. 95111,
Oct. 2015.
[8] S. Bechikh, A. Chaabani, and L. B. Said, An efficient chemical reaction
optimization algorithm for multiobjective optimization, IEEE Trans.
Cybern., vol. 45, no. 10, pp. 20512064, Oct. 2015.
[9] R. H. Shang, L. C. Jiao, F. Liu, and W. P. Ma, A novel immune clonal
algorithm for MO problems, IEEE Trans. Evol. Comput., vol. 16, no. 1,
pp. 3550, Feb. 2012.
[10] H. Xia, J. Zhuang, and D. H. Yu, Combining crowding estimation in
objective and decision space with multiple selection and search strategies
for multi-objective evolutionary optimization, IEEE Trans. Cybern.,
vol. 44, no. 3, pp. 378393, Mar. 2014.
[11] K. Deb and H. Jain, An evolutionary many-objective optimization
algorithm using reference-point-based nondominated sorting approach,
Part I: Solving problems with box constraints, IEEE Trans. Evol.
Comput., vol. 18, no. 4, pp. 577601, Aug. 2014.
[12] Z.-H. Zhan et al., Multiple populations for multiple objectives:
A coevolutionary technique for solving multiobjective optimization
problems, IEEE Trans. Cybern., vol. 43, no. 2, pp. 445463, Apr. 2013.
[13] M. S. Snchez, M. C. Ortiz, and L. A. Sarabia, Selection of nearly
orthogonal blocks in ad-hoc experimental designs, Chemometr. Intell.
Lab. Syst., vol. 133, pp. 109120, Apr. 2014.
[14] P. Yang, K. Tang, and X. F. Lu, Improving estimation of distribution
algorithm on multimodal problems by detecting promising areas, IEEE
Trans. Cybern., vol. 45, no. 8, pp. 14381449, Aug. 2015.
[15] H. Li and Q. F. Zhang, Multiobjective optimization problems with
complicated Pareto sets, MOEA/D and NSGA-II, IEEE Trans. Evol.
Comput., vol. 13, no. 2, pp. 284302, Apr. 2009.
[16] S.-Z. Zhao, P. N. Suganthan, and Q. F. Zhang, Decomposition-based
multiobjective evolutionary algorithm with an ensemble of neighborhood sizes, IEEE Trans. Evol. Comput., vol. 16, no. 3, pp. 442446,
Jun. 2012.
[17] Q. F. Zhang and Y.-W. Leung, Orthogonal genetic algorithm for multimedia multicast routing, IEEE Trans. Evol. Comput., vol. 3, no. 1,
pp. 5362, Apr. 1999.
[18] Y.-W. Leung and Y. P. Wang, An orthogonal genetic algorithm with
quantization for global numerical optimization, IEEE Trans. Evol.
Comput., vol. 5, no. 1, pp. 4153, Feb. 2001.
[19] Q. X. Feng, S. Y. Liu, G. Q. Tang, L. Q. Yong, and J. K. Zhang,
Biogeography-based optimization with orthogonal crossover, Math.
Probl. Eng., vol. 2013, 2013, Art. ID 353969.
[20] B. Suman, N. Hoda, and S. Jha, Orthogonal simulated annealing for
multiobjective optimization, Comput. Chem. Eng., vol. 34, no. 10,
pp. 16181631, Oct. 2010.
[21] K. S. Narendra and M. A. L. Thathachar, Learning AutomataAn
Introduction. Englewood Cliffs, NJ, USA: Prentice Hall, 1989.
[22] R. R. Bush and F. Mosteller, Stochastic Models for Learning. New York,
NY, USA: Wiley, 1958.
[23] R. C. Atkinson, G. H. Bower, and E. J. Crothers, An Introduction to
Mathematical Learning Theory. New York, NY, USA: Wiley, 1965.
[24] M. L. Tsetlin, Automaton Theory and Modeling of Biological Systems
(Mathematics in Science and Engineering), vol. 102. New York, NY,
USA: Academic Press, 1973.
[25] M. Shirali, N. Shirali, and M. R. Meybodi, Sleep-based topology control in the ad hoc networks by using fitness aware learning automata,
Comput. Math. Appl., vol. 64, no. 2, pp. 137146, Jul. 2012.
[26] M. R. Aghaebrahimi, S. H. Zahiriand, and M. Amiri, Data mining
using learning automata, World Acad. Sci. Eng. Technol., vol. 49,
pp. 308311, Feb. 2009.
[27] W. Jiang, C.-L. Zhao, S.-H. Li, and L. S. Chen, A new learning automata based approach for online tracking of event patterns,
Neurocomputing, vol. 137, pp. 205211, Aug. 2014.
[28] A. Rezanian, M. Rahmati, and M. R. Meybodi, Sampling from complex
networks using distributed learning automata, Phys. A Stat. Mech. Appl.,
vol. 396, pp. 224234, Feb. 2014.

IEEE TRANSACTIONS ON CYBERNETICS, VOL. 46, NO. 12, DECEMBER 2016

[29] B. Moradabadi and H. Beigy, A new real-coded Bayesian optimization


algorithm based on a team of learning automata for continuous optimization, Genet. Program. Evol. Mach., vol. 15, no. 2, pp. 169193,
2014.
[30] Q. H. Wu and H. L. Liao, Function optimisation by learning automata,
Inf. Sci., vol. 220, pp. 379398, Jan. 2013.
[31] N. Kumar, J.-H. Lee, and J. J. P. C. Rodrigues, Intelligent mobile video
surveillance system as a Bayesian coalition game in vehicular sensor networks: Learning automata approach, IEEE Trans. Intell. Transp. Syst.,
vol. 16, no. 3, pp. 11481161, Jun. 2014.
[32] M. Ghasemi, M. Abdolahi, M. Bag-Mohammadi, and A. Bohlooli,
Adaptive multi-flow opportunistic routing using learning automata,
Ad Hoc Netw., vol. 25, pp. 472479, Feb. 2014.
[33] H. L. Liao, Q. H. Wu, Y. Z. Li, and L. Jiang, Economic emission dispatching with variations of wind power and loads using multi-objective
optimization by learning automata, Energy Convers. Manage., vol. 87,
pp. 990999, Nov. 2014.
[34] F. M. Pour and A. A. Gharaveisi, Opposition-based discrete action
reinforcement learning automata algorithm case study: Optimal design
of a PID controller, Turk. J. Electr. Eng. Comput. Sci., vol. 21, no. 6,
pp. 16031614, 2013.
[35] M. A. Khamis and W. Gomaa, Adaptive multi-objective reinforcement learning with hybrid exploration for traffic signal control based on
cooperative multi-agent framework, Eng. Appl. Artif. Intell., vol. 29,
pp. 134151, Mar. 2014.
[36] A. Moustafa and M. J. Zhang, Multi-objective service composition
using reinforcement learning, in Proc. 11th Int. Conf. Serv. Orient.
Comput. (ICSOC), vol. 8274. Berlin, Germany, Dec. 2013, pp. 298312.
[37] H. L. Liao and Q. H. Wu, Multi-objective optimization by reinforcement learning, in Proc. IEEE Congr. Evol. Comput., Barcelona, Spain,
2010, pp. 33743381.
[38] M. R. Aghaebrahimi, S. H. Zahiriand, and M. Amiri, A new method for
multiobjective optimization based on learning automata, World Acad.
Sci. Eng. Technol., vol. 25, pp. 312315, 2009.
[39] A. Karakas, C. Kocatepe, and F. X. Li, A new multi-objective optimization technique for generation dispatch, in Proc. North Amer. Power
Symp. (NAPS), Starkville, MS, USA, Oct. 2009, pp. 16.
[40] H. L. Liao and Q. H. Wu, Multi-objective optimization by learning
automata, J. Global Optim., vol. 55, no. 2, pp. 459487, Feb. 2013.
[41] A. Karakas, C. Kocatepe, and F. X. Li, Using learning automata for
multi-objective generation dispatch considering cost, voltage stability
and power losses, Turk. J. Electr. Eng. Comput. Sci., vol. 19, no. 6,
pp. 913927, Nov. 2011.
[42] G. Horn and B. J. Oommen, Solving multiconstraint assignment problems using learning automata, IEEE Trans. Syst., Man, Cybern. B,
Cybern., vol. 40, no. 1, pp. 618, Feb. 2010.
[43] A. Sengupta, T. Chakraborti, A. Konar, E. Kim, and A. K. Nagar, An
adaptive memetic algorithm using a synergy of differential evolution
and learning automata, in Proc. IEEE Congr. Evol. Comput., Brisbane,
QLD, Australia, Jun. 2012, pp. 18.
[44] K. Miettinen, Nonlinear Multiobjective Optimization. Boston, MA,
USA: Kluwer, 1999.
[45] K. Deb, Multiobjective Optimization Using Evolutionary Algorithms.
New York, NY, USA: Wiley, 2001.
[46] M. A. L. Thathachar and P. S. Sastry, Networks of Learning
Automata: Techniques for Online Stochastic Optimization. Dordrecht,
The Netherlands: Kluwer, 2004.
[47] D. C. Montgomery, Design and Analysis of Experiments, 3rd ed.
New York, NY, USA: Wiley, 1991.
[48] C. R. Hicks, Fundamental Concepts in the Design of Experiments,
4th ed. New York, NY, USA: Saunders College, 1993.
[49] O. Tekinalp and G. Karsli, A new multiobjective simulated annealing
algorithm, J. Global Optim., vol. 39, no. 1, pp. 4977, Sep. 2007.
[50] J. Dubois-Lacoste, M. Lpez-Ibez, and T. Sttzle, Combining
two search paradigms for multi-objective optimization: Two-phase
and Pareto local search, Universite Libre de Bruxelles, Institut de
Recherches Interdisciplinaires et de Dveloppements en Intelligence
Artificielle, Brussels, Belgium, Tech. Rep. TR/IRIDIA/2012-004,
Mar. 2012.
[51] Q. F. Zhang et al., Multiobjective optimization test instances for the
CEC 2009 special session and competition, School Comput. Sci. Electr.
Eng., Univ. Essex, Colchester, U.K., Tech. Rep. CES-487, 2008.
[52] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, Scalable
multi-objective optimization test problems, in Proc. Congr. Evol.
Comput. (CEC), vol. 1. Honolulu, HI, USA, 2002, pp. 825830.

DAI et al.: OELA FOR MULTIOBJECTIVE OPTIMIZATION

[53] S. Huband, P. Hingston, L. Barone, and L. While, A review of multiobjective test problems and a scalable test problem toolkit, IEEE Trans.
Evol. Comput., vol. 10, no. 5, pp. 477506, Oct. 2006.
[54] R. Storn and K. Price, Differential evolutionA simple and efficient heuristic for global optimization over continuous space, J. Global
Optim., vol. 11, no. 4, pp. 341359, Dec. 1997.
[55] C. A. C. Coello and N. C. Corts, Solving multiobjective optimization
problems using an artificial immune system, Genet. Program. Evol.
Mach., vol. 6, no. 2, pp. 163190, Jun. 2005.
[56] R. G. D. Steel, J. H. Torrie, and D. A. Dickey, Principles and
Procedures of Statistics: A Biometrical Approach. New York, NY,
USA: McGraw-Hill, 1997.
[57] Q. F. Zhang and H. Li, MOEA/D: A multiobjective evolutionary algorithm based on decomposition, IEEE Trans. Evol. Comput., vol. 11,
no. 6, pp. 712731, Dec. 2007.
[58] E. Zitzler and L. Thiele, Multiobjective evolutionary algorithms:
A comparative case study and the strength Pareto approach, IEEE
Trans. Evol. Comput., vol. 3, no. 4, pp. 257271, Nov. 1999.
[59] A. O. Allen, Probability, Statistics, and Queuing Theory With Computer
Science Applications, 2nd ed. Boston, MA, USA: Academic Press, 1990.

Cai Dai received the Ph.D. degree in computer science and technology from the School of Computer
Science and Technology, Xidian University, Xian,
China, in 2014.
His current research interests include multiobjective optimization and evolutionary algorithms.

3319

Yuping Wang (SM11) received the Ph.D. degree


from the Department of Mathematics, Xian Jiaotong
University, Xian, China, in 1993.
He is a Full Professor with the School
of Computer Science and Technology, Xidian
University, Xian. His current research interests
include optimization methods, theory and application, evolutionary computation, data mining, and
machine learning.

Miao Ye, photograph and biography not available at the time of publication.

Xingsi Xue, photograph and biography not available at the time of publication.

Hailin Liu, photograph and biography not available at the time of publication.

You might also like