Elliptical and Archimedean Copulas
in Estimation of Distribution Algorithm with Model Migration
Martin Hyr
ˇ
s and Josef Schwarz
Faculty of Information Technology, Brno University of Technology, Bo
ˇ
zet
ˇ
echova 2, Brno, Czech Republic
Keywords:
Estimation of Distribution Algorithms, Copula Theory, Parallel EDA, Island-based Model, Multivariate
Copula Sampling, Migration of Probabilistic Models.
Abstract:
Estimation of distribution algorithms (EDAs) are stochastic optimization techniques that are based on building
and sampling a probability model. Copula theory provides methods that simplify the estimation of a probabil-
ity model. An island-based version of copula-based EDA with probabilistic model migration (mCEDA) was
tested on a set of well-known standard optimization benchmarks in the continuous domain. We investigated
two families of copulas – Archimedean and elliptical. Experimental results confirm that this concept of model
migration (mCEDA) yields better convergence as compared with the sequential version (sCEDA) and other
recently published copula-based EDAs.
1 INTRODUCTION
Estimation of distribution algorithms (EDAs) belong
to a new class of evolutionary optimization meth-
ods that explore the search space by estimating and
sampling an explicit probabilistic model of promis-
ing solutions. EDAs applied to discrete problems
are described in the well-known papers UMDA (Pe-
likan and M
¨
uhlenbein, 1999b), BMDA (Pelikan and
M
¨
uhlenbein, 1999a), MIMIC (De Bonet et al., 1997),
and BOA (Pelikan et al., 1999). Solutions of the op-
timization problems in the real value domain can be
found in (Larra
˜
naga and Lozano, 2001). A very mod-
ern and accessible survey of the EDAs algorithm is
presented in (Hauschild and Pelikan, 2011).
The main advantage of EDAs is its capacity to dis-
cover those variable linkages that yield a solution to a
complex optimization problem. On the one hand this
probability model-based approach has allowed EDAs
to be applied to large and complex problems. On
the other hand explicit probabilistic models are very
time consuming. That was the reason for implement-
ing various advanced EDAs to solve this problem.
The well-known enhancement approaches include the
parallelization of model building and sporadic model
building (Hauschild and Pelikan, 2011).
In the last ten years a new approach has appeared
to building an efficient probabilistic model that is
based on copula theory (Mai and Scherer, 2012).
Copulas are special probability distribution functions.
Due to their properties it is possible to use them
to model correlations within multivariate problems
the joint distribution is separated into the univariate
marginal distributions and into the correlation struc-
ture that is expressed by the copula function. Cop-
ula theory has very often been used in finance and
statistics works (Nelsen, 2006; Cherubini et al., 2004;
Aas et al., 2009) (e.g. modeling health insurance data
(Zimmer and Trivedi, 2006)).
Recently copulas have been utilized in the field of
the machine learning (Rey and Roth, 2012; P
´
oczos
et al., 2012). More recently the copula theory has
been applied to EDA probability models. The sim-
plest case is the application with bivariate (2D) copu-
las: (Wang et al., 2009) 2D Gaussian copula EDA,
(Wang et al., 2010b) 2D Clayton copula EDA,
(Wang et al., 2010a) – 2D Gumbel copula EDA.
In the case of multivariate models bivariate cop-
ulas are used as local building blocks in vari-
ous graph dependence structures: (Salinas-Guti
´
errez
et al., 2009) – MIMIC + Frank and Gaussian copula),
(M
´
endez and Landa, 2012) Bayesian network +
Archimedean copulas, (Salinas-Guti
´
errez et al., 2011)
D-vine + copulas, (Soto et al., 2012) C-vine, D-
vine + copulas.
The copula-based EDA starts with an estimation
of the marginal distribution of each variable, then us-
ing a proper copula, the joint distribution is estab-
lished. Given the margins and a copula, it is then pos-
sible to generate new solutions.
212
Hyrš, M. and Schwarz, J..
Elliptical and Archimedean Copulas in Estimation of Distribution Algorithm with Model Migration.
In Proceedings of the 7th International Joint Conference on Computational Intelligence (IJCCI 2015) - Volume 1: ECTA, pages 212-219
ISBN: 978-989-758-157-1
Copyright
c
2015 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
This paper deals with an efficient parallelization
of island-based mCEDA with the goal to increase
the convergence speed. After some experiments we
chose the island-based structure with bidirectional
ring topology. Instead of the often used migration of
individuals we instantiated the migration of probabil-
ity models. Note that the first experiments with this
new concept were published in (delaOssa et al., 2004;
Schwarz and Jaro
ˇ
s, 2008) for the optimization in the
discrete domain, the chief obstacle being an efficient
combination of the probabilistic models, especially
those having the dependence structure expressed by
a graph. That is why we focused on the migration of
probabilistic model parameters only.
The paper is organized as follows. In Section 2,
the basis of copula theory is given. In Section 3, the
utilization of copulas in EDA is described, and sam-
pling algorithms for copulas are presented. In Sec-
tion 4, the island-based model of evolution algorithm
with copula-model migration is described. Our exper-
iments are discussed in Section 5. The conclusions
are given in Section 6.
2 COPULA THEORY
The copula concept was introduced by (Sklar, 1959)
in order to separate the effect of dependence of vari-
ables from the effect of marginal distributions in a
joint distribution. A copula is a function which joins
the univariate distribution function and creates mul-
tivariate distribution functions. This approach allows
us to transform multivariate statistic problems into the
univariate problems with the relation represented by
just the copula.
Definition. A copula C is a multivariate probability
distribution function for which the marginal probabil-
ity distribution of each variable is uniform in [0; 1].
Definition. A copula is a function C : [0; 1]
d
[0; 1]
with the following properties:
1. C(u
1
, u
2
, . . . , u
d
) = 0 for at least one u
i
= 0
2. C(1, 1, . . . , 1, u
i
, 1, . . . , 1) = u
i
for all i = 1, 2, . . . , d
3. C(u
1
, u
2
, . . . , u
d
) is d-increasing (see lit. for de-
tails)
Theorem. Sklar’s theorem: Let F be a d-dimensional
distribution function with margins F
1
, . . . , F
d
. Then
there exists a d-dimensional copula C such that for
all (x
1
, . . . , x
d
) R
d
it holds that
F(x
1
, . . . , x
d
) = C (F
1
(x
1
), . . . , F
d
(x
d
)) (1)
If F
1
, . . . , F
d
are continuous, then C is unique. Con-
versely, if C is a d-dimensional copula and F
1
, . . . , F
d
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 1: Scatterplots of bivariate Archimedean copulas:
Clayton (left), Gumbel (middle) and Frank (right) with de-
pendence strengths 0.9 (top) and 0.3 (bottom).
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 2: Scatterplots of bivariate eliptical copulas: Gaus-
sian (top) and Student, ν = 2 (bottom), with dependence
strengths 0.5 (left), 0.3 (middle) and 0.9 (right).
are univariate distribution functions, then the func-
tion F defined via (1) is a d-dimensional distribution
function.
In this paper we focus on two big copula families
– Archimedean and elliptic.
Archimedean copulas are capable of capturing
wide ranges of dependence. The definition of the
Archimedean copula is based on the generator func-
tion. There are many existing Archimedean copulas
and many more that could be created. The three cop-
ulas, i.e. Clayton, Gumbel and Frank appear regu-
larly in statistics literature, they are popular because
they model different patterns of dependence and have
a relatively simple functional form. Fig. 1 shows scat-
terplots of these copulas.
The elliptical copulas are derived from the related
elliptical distribution. The first example of elliptical
copula is the Gaussian copula, which belongs to the
normal distribution, the second example is the Stu-
dent copula, which belongs to the t-distribution (see
Fig. 2).
Elliptical and Archimedean Copulas in Estimation of Distribution Algorithm with Model Migration
213
3 COPULA-BASED ESTIMATION
OF DISTRIBUTION
ALGORITHM
Estimation of distribution algorithms belongs to the
advanced evolutionary algorithms. Solving the nu-
merical optimization problem, vector x = (x
1
, . . . , x
d
)
of the optimal solution is searched out.
The core of the canonical EDA consists of three
main steps, see Algorithm 1.
Algorithm 1: The pseudocode of canonical EDA.
Generate initial population.
WHILE (termination criteria is false):
1. Select promising solutions into subpopulation
from the current population.
2. Create the probability model from the selected
subpopulation.
3. Sample the probability model and generate the
new population.
Step 1 is quite straightforward, the promising so-
lutions are stated using the standard selection trunca-
tion. In the case of copula-based EDA it is neces-
sary to choose the proper type of copula and derive
the copula parameters and the marginal distribution
parameters.
The principle of sampling schema for generating
the new individuals using the copula model is de-
scribed in Algorithm 2:
Algorithm 2: Sampling the copula and generating the new
individuals.
1. Obtain the random copula sample (u
1
, . . . , u
d
)
C, where u
i
[0; 1].
2. Derive the vector x of the searched solution using
inverse marginal distributions, x
i
= F
1
i
(u
i
).
3.1 Identification of Copula Probability
Model
The copula-based probability model includes two
parts: univariate marginal distributions and the cop-
ula function. Marginal distributions can be identified
separately for each variable and the copula includes
the correlation between variables.
For marginal distribution in each dimension i =
1, . . . , d we used normal distribution, which is param-
eterized by the mean value µ
i
and standard deviation
σ
i
.
For assessing the parameters of the copula we
used the Kendall τ correlation coefficient.
In the case of Archimedean copulas the following
relations hold for the parameter θ (in the case of d-
variate copulas, d 3, we use average
¯
τ; for d = 2
the standard pairwise τ is used):
for the Clayton copula θ
Clayton
=
2τ
1τ
.
for the Gumbel copula θ
Gumbel
=
1
1τ
.
for the Frank copula (approximation) θ
Frank
.
=
(10
arcsin(τ)
1) e.
Elliptical copulas are parameterized by correlation
matrix R. Elements R
i j
are adapted from Kendall’s τ
for each pair of dimensions i, j using formula R
i j
=
sin
1
2
πτ
i j
.
3.2 Copula Sampling Algorithms
Now we specify step 1 from Alg. 2 for each type of
copulas in more details.
The algorithm for sampling Archimedean copu-
las uses a random value J, which is obtained from the
distribution given by the inverse of Laplace transform
L
1
of generator (Mai and Scherer, 2012; Aas, 2004;
Melchiori, 2006), see Alg. 3.
Algorithm 3: Archimedean copula sampling.
1. generate value J L
1
[ϕ(t)]
2. generate uniformly distributed random numbers
z
i
U (0, 1) (for i = 1, . . . , d)
3. return u
i
= ϕ
log(z
i
)
J
(for i = 1, . . . , d)
According to (Aas, 2004; Melchiori, 2006) the
value of J can be derived:
for the Clayton copula by Gamma distribution J
Gamma
1
θ
, θ
.
for the Gumbel copula by Levy skew alpha-Stable
distribution J Stable
1
θ
, 1,
cos
π
2θ
θ
, 0
.
for the Frank copula by logarithmic series distri-
bution J Logarithmic
1 e
θ
.
The sampling scheme for Gaussian and Student
copulas (Mai and Scherer, 2012) (see Alg. 4, 5) uses
the Cholesky decomposition of the given correlation
matrix R to obtain the lower triangular matrix L, such
that LL
T
= R. The Student copula is further specified
by degrees of freedom, we use ν = (N 1)d (where
N is population size and d number of dimensions).
ECTA 2015 - 7th International Conference on Evolutionary Computation Theory and Applications
214
Algorithm 4: Gaussian copula sampling.
1. compute L
2. generate random numbers z
i
No(0, 1) with stan-
dard normal distribution (for i = 1, . . . , d)
3. calculate s
i
=
i
j=1
L
i, j
z
j
(for i = 1, . . . , d)
4. return u
i
= Φ(s
i
) (for i = 1, . . . , d)
Algorithm 5: Student copula sampling.
1. compute L
2. generate V χ
2
(ν)
3. generate random numbers z
i
No(0, 1) with stan-
dard normal distribution (for i = 1, . . . , d)
4. calculate s
i
=
q
ν
V
i
j=1
L
i, j
z
j
(for i = 1, . . . , d)
5. return u
i
= t
ν
(s
i
) (for i = 1, . . . , d)
4 ISLAND-BASED COPULA-EDA
The principal motivation for the proposal of a concept
of copula-based EDA parallelization is to discover the
efficiency of the transfer of probabilistic parameters
instead of the traditional transfer of individuals. The
main goal is to improve algorithm convergence. In the
case of EDAs only a few papers deal with the discrete
probability model migration (delaOssa et al., 2004;
delaOssa et al., 2005) and (Hyr
ˇ
s and Schwarz, 2014)
in the case of copula-based EDA.
4.1 EDA with Migration
With the concordance of experimental work done in
(Schwarz and Jaro
ˇ
s, 2008), and according to our ex-
perimental results, we used the island-based commu-
nication model with bidirectional ring topology. This
topology provides good local interaction and in a few
steps allows the propagation of information along the
ring.
The evolution process on every island runs inde-
pendently. When the migration condition is met the
communication (transfer of model parameters) is ac-
tivated, see Alg. 6.
4.2 Model Combination
According to the island-based topology we have de-
composed the migration process into pairwise inter-
actions of two islands one of them is the resident
island specified by resident probabilistic model M
R
Algorithm 6 : The pseudocode of canonical EDA with
model migration.
1. Generate initial populations.
2. FOR each island DO IN PARALLEL:
3. WHILE (termination criteria is false):
4. Select promising individuals
5. Create probability model (see Sec. 3)
6. IF (sending condition):
7. Send model
8. WHILE (immigrant model received):
9. Combine models (see Sec. 4.2)
10. Sample new population from probability model
and the other one is the immigrant island whose prob-
abilistic model M
I
is transferred to a new resident
model.
The combination of the immigrant model with the
model of the resident island is described in more de-
tails. In general, the modification of the resident
model by the immigrant model can be formalized by
(Schwarz and Jaro
ˇ
s, 2008):
M
new
R
= (1 β)M
R
+ βM
I
(2)
where the coefficient β [0;1] specifies the influence
of the immigrant model.
We have proposed the following model combina-
tion rules according to (Fr
¨
uhwirth-Schnatter, 2006):
Learning the mean value µ
i
of each univariate
marginal distribution F
i
(x
i
)
µ
new
i
= (1 β)µ
R
i
+ βµ
I
i
(3)
Learning the standard deviation σ
i
of each uni-
variate marginal distribution F
i
(x
i
)
σ
new
i
=
r
(1 β)
µ
new
i
µ
R
i
2
+
σ
R
i
2
+
+β
µ
new
i
µ
I
i
2
+
σ
I
i
2
(4)
Learning the correlation matrix value R
i j
R
new
i j
= (1 β)R
R
i j
+ βR
I
i j
(5)
We have chosen the coefficient β as
β =
(
f it
R
f it
R
+ f it
I
f it
I
f it
R
0.1 otherwise
(6)
where f it
R
or f it
I
represents the fitness value of the
resident or the immigrant model.
Elliptical and Archimedean Copulas in Estimation of Distribution Algorithm with Model Migration
215
5 EXPERIMENT AND RESULTS
We used standard benchmark problems to compare
island-based copula-EDA (mCEDA) with the sequen-
tial one (sCEDA).
5.1 Benchmarks
The shifted variants of several well-known bench-
marks from the area of numerical optimization are
used. All these functions have been adapted such that
the task is minimization with optimal fitness value 0
and with shifted optima position.
Elliptic Function:
f (x) =
d
i=1
(10
6
)
i1
d1
x
2
i
, x
i
[100; 100]
Rastrigin’s Function:
f (x) =
d
i=1
x
2
i
10 cos(2πx
i
) + 10
, x
i
[5;5]
Ackley’s Function:
f (x) = 20 e
0.2
q
1
d
d
i=1
x
2
i
e
1
d
d
i=1
cos(2πx
i
)
+
20 + e, x
i
[32; 32]
Schwefel’s Problem 1.2:
f (x) =
d
i=1
i
j=1
x
i
2
, x
i
[100; 100]
Rosenbrock’s Function:
f (x) =
d1
i=1
100(x
2
i
x
i+1
)
2
+ (x
i
1)
2
,
x
i
[100; 100]
Summation Cancellation:
f (x) = 10
5
1
10
5
+
d
i=1
|
i
j=1
x
j
|
, x
i
[1; 1]
We used the shifted optima position in the form
f itness(x) = f (x
0
), where x
0
i
= x
i
0.25(x
max
x
min
), i = 1, . . . , d.
We have chosen the following settings:
Problem size: 10 variables/dimensions for all
problems.
Population size of each island: 500.
Selection: We used truncation selection, with a
selection proportion of 0.2, i.e. 100 individuals.
Number of islands: 10.
Migration rate: after every 20 generations.
Maximum number of fitness evaluations: 500000
(i.e. 100 generations for the island-based model
and 1000 generations for the sequential variant).
Number of independent runs: 20.
5.2 Results and Discussion
We carried out a comparison of two variants of
the copula-based EDA algorithm: sequential variant
sCEDA specified in Sec. 3 and island-based algorithm
with model migration mCEDA (Sec. 4) both of them
with the same classes of copulas.
We used the so-called weak model of paralleliza-
tion, the population size in sCEDA is equal to the pop-
ulation of one island in mCEDA. The total population
in mCEDA is thus ten-times bigger than in sCEDA.
To retain the same computational cost, we increased
the number of generations ten-times for sCEDA ac-
cording to mCEDA, thus the total computational cost
measured by fitness evaluations is the same.
In Tables 1–6 the convergence of the proposed
sCEDA and mCEDA algorithms is presented. The
mean values related to specified evaluation epochs are
listed.
It can be seen that mCEDA performs better than
sCEDA for most benchmarks. (Only in the case of
Ackley’s benchmark the sCEDA is better, the cause is
under our investigation.) The sCEDA is able to find
a relatively good local solution quite fast but then it
loses its diversity and no further improvement is ob-
tained. The mCEDA converges slowly but it has the
capability to find near optimal solution. We suppose
that this performance is caused by the phenomenon of
the proposed model migration.
In the case of Rosenbrock’s and Summation Can-
cellation problems, only mCEDA using elliptic copu-
las is able to achieve some progress during the evo-
lution process. Neither Archimedean copulas nor
sCEDA have this capacity.
Besides the comparison of sCEDA and mCEDA
the influence of each copula type in mCEDA is worth
discussing. In the case of Rosenbrock’s and Sum-
mation Cancellation, only elliptical copulas are able
to perform well. In the case of Rastrigin’s and El-
liptic functions, Archimedean copulas perform better
than elliptic ones. In the case of Ackley’s and Schwe-
fel’s 1.2 functions, there is no significant difference
between the two copula families.
From the observations it follows that the success
rate of the both versions of mCEDA is almost iden-
tical. But in the case of Archimedean mCEDA the
drawback appears in tendency to become stuck on lo-
cal optima, see Table 3.
In Tables 7–9 we arranged a comparison (mean
fitness values) of mCEDA using the Frank copula
(mCEDA-F) and the Gaussian copula (mCEDA-G)
(as members of Archimedean and elliptic families)
with the other published algorithms that used differ-
ent versions of copulas. The comparison is done for
the same number of fitness evaluations for the same
subset of benchmarks with 10 dimensions.
In Table 7 a comparison with the algorithm using
the Copula Bayesian network (M
´
endez and Landa,
ECTA 2015 - 7th International Conference on Evolutionary Computation Theory and Applications
216
Table 1: Experimental results (mean fitness) for Rastrigin’s
function.
fit. eval. Clayton Gumbel Frank Gauss Student
island based
100000 5.78e+00 1.42e-01 9.88e-02 2.47e+01 2.49e+01
200000 3.70e-05 1.04e-07 3.92e-07 1.16e+00 1.70e+00
300000 2.79e-11 2.72e-14 2.23e-13 2.41e-01 4.26e-01
400000 0.00e+00 0.00e+00 0.00e+00 5.02e-02 9.60e-02
500000 0.00e+00 0.00e+00 0.00e+00 1.71e-04 5.00e-02
sequential
100000 2.56e-01 4.65e-01 8.67e-01 2.36e+00 2.75e+00
200000 2.56e-01 4.65e-01 6.52e-01 2.36e+00 2.75e+00
300000 2.56e-01 4.65e-01 6.52e-01 2.36e+00 2.75e+00
400000 2.56e-01 4.65e-01 6.52e-01 2.36e+00 2.75e+00
500000 2.56e-01 4.65e-01 6.52e-01 2.36e+00 2.75e+00
Table 2: Experimental results (mean fitness) for Rosen-
brock’s function.
fit. eval. Clayton Gumbel Frank Gauss Student
island based
100000 8.40e+00 8.40e+00 8.57e+00 8.13e+00 8.27e+00
200000 7.83e+00 7.89e+00 8.00e+00 6.79e+00 6.59e+00
300000 7.66e+00 7.76e+00 7.86e+00 6.46e+00 6.16e+00
400000 7.53e+00 7.62e+00 7.69e+00 6.25e+00 5.86e+00
500000 7.50e+00 7.60e+00 7.66e+00 6.20e+00 5.82e+00
sequential
100000 8.27e+00 8.13e+00 8.14e+00 3.28e+04 7.63e+04
200000 8.27e+00 8.13e+00 8.14e+00 3.28e+04 7.61e+04
300000 8.27e+00 8.13e+00 8.14e+00 3.28e+04 7.61e+04
400000 8.27e+00 8.13e+00 8.14e+00 3.28e+04 7.61e+04
500000 8.27e+00 8.13e+00 8.14e+00 3.28e+04 7.61e+04
Table 3: Experimental results (mean fitness) for Summation
Cancellation function.
fit. eval. Clayton Gumbel Frank Gauss Student
island based
100000 1.00e+05 1.00e+05 1.00e+05 9.99e+04 9.98e+04
200000 9.99e+04 9.99e+04 9.99e+04 9.43e+04 9.64e+04
300000 9.99e+04 9.97e+04 9.97e+04 7.66e+04 6.89e+04
400000 9.97e+04 9.95e+04 9.92e+04 5.23e+04 4.53e+04
500000 9.97e+04 9.93e+04 9.90e+04 4.46e+04 4.40e+04
sequential
100000 1.00e+05 1.00e+05 1.00e+05 9.99e+04 1.00e+05
200000 1.00e+05 1.00e+05 1.00e+05 9.99e+04 1.00e+05
300000 1.00e+05 1.00e+05 1.00e+05 9.99e+04 1.00e+05
400000 1.00e+05 1.00e+05 1.00e+05 9.99e+04 1.00e+05
500000 1.00e+05 1.00e+05 1.00e+05 9.99e+04 1.00e+05
2012) after 100,000 evaluations is shown. For the
case of Rastrigin’s, Ackley’s and Schwefel’s 1.2 func-
tions mCEDA-F and mCEDA-G are evidently better,
for Rosenbrock’s function the results are comparable.
Table 4: Experimental results (mean fitness) for Elliptic
function.
fit. eval. Clayton Gumbel Frank Gauss Student
island based
100000 2.76e-01 3.35e-01 3.42e-01 2.78e+01 4.25e+01
200000 4.94e-08 5.78e-08 8.14e-08 5.28e+00 4.67e-01
300000 6.75e-15 1.02e-14 1.46e-14 1.94e+00 1.82e-01
400000 3.82e-17 4.27e-17 1.21e-18 1.39e+00 1.32e-01
500000 3.82e-17 4.27e-17 1.21e-18 9.63e-01 1.17e-01
sequential
100000 6.46e-16 7.23e-16 6.68e-16 1.30e+03 2.45e+03
200000 6.46e-16 7.23e-16 6.68e-16 1.30e+03 2.45e+03
300000 6.46e-16 7.23e-16 6.68e-16 1.30e+03 2.45e+03
400000 6.46e-16 7.23e-16 6.68e-16 1.30e+03 2.45e+03
500000 6.46e-16 7.23e-16 6.68e-16 1.30e+03 2.45e+03
Table 5: Experimental results (mean fitness) for Ackley’s
function.
fit. eval. Clayton Gumbel Frank Gauss Student
island based
100000 1.16e-02 1.17e-02 1.35e-02 1.27e-02 1.33e-02
200000 4.91e-06 4.39e-06 5.85e-06 5.65e-06 5.53e-06
300000 2.01e-09 2.16e-09 2.51e-09 2.36e-09 2.00e-09
400000 8.67e-13 8.40e-13 1.15e-12 8.48e-13 7.40e-13
500000 1.33e-15 1.69e-15 2.93e-15 9.77e-16 1.51e-15
sequential
100000 0.00e+00 0.00e+00 0.00e+00 3.21e-02 8.60e-02
200000 0.00e+00 0.00e+00 0.00e+00 3.21e-02 8.60e-02
300000 0.00e+00 0.00e+00 0.00e+00 3.21e-02 8.60e-02
400000 0.00e+00 0.00e+00 0.00e+00 3.21e-02 8.60e-02
500000 0.00e+00 0.00e+00 0.00e+00 3.21e-02 8.60e-02
Table 6: Experimental results (mean fitness) for Schwefel’s
1.2 function.
fit. eval. Clayton Gumbel Frank Gauss Student
island based
100000 1.78e-03 1.79e-03 2.20e-03 2.85e-03 2.54e-03
200000 1.03e-10 9.30e-11 1.11e-10 2.63e-10 2.48e-10
300000 5.09e-17 7.49e-17 1.08e-16 2.07e-16 2.85e-16
400000 5.09e-17 7.49e-17 1.08e-16 2.07e-16 2.85e-16
500000 5.09e-17 7.49e-17 1.08e-16 2.07e-16 2.85e-16
sequential
100000 6.15e-16 5.39e-16 6.00e-16 5.47e+01 2.23e+02
200000 6.15e-16 5.39e-16 6.00e-16 5.46e+01 2.23e+02
300000 6.15e-16 5.39e-16 6.00e-16 5.46e+01 2.23e+02
400000 6.15e-16 5.39e-16 6.00e-16 5.46e+01 2.23e+02
500000 6.15e-16 5.39e-16 6.00e-16 5.46e+01 2.23e+02
In Table 8 a comparison with the other suite of
algorithms (Zhao and Wang, 2012; Jia et al., 2013)
is carried out on the level of 300,000 evaluations.
In the case of Rosenbrock’s function the results are
Elliptical and Archimedean Copulas in Estimation of Distribution Algorithm with Model Migration
217
Table 7: Comparison (mean fitness) of mCEDA with Cop-
ula Bayesian Network (CBN) from (M
´
endez and Landa,
2012).
Rastr. Ack. Schw. 1.2 Rosen.
CBN 2.39e+00 3.71e-02 2.23e+01 1.05e+01
mCEDA-F 9.88e-02 1.35e-02 2.20e-03 8.57e+00
mCEDA-G 2.47e+01 1.27e-02 2.85e-03 8.13e+00
Table 8: Comparison (mean fitness) of mCEDA with: Cop-
ula EDA (cE), Copula EDA of Dynamic K-S test (cE-KS)
from (Zhao and Wang, 2012); Clayton (Cl), Gumbel (Gu),
Sn-EDA (Sn) from (Jia et al., 2013).
Ellip./sphere Rastrigin’s Rosenbrock’s
cE 4.62e-08 6.45e-08 6.52e+00
cE-KS 1.16e-08 2.60e-08 7.05e+00
Cl 1.45e-07 7.00e-08 8.36e+00
Gu 3.59e-09 5.49e-09 6.62e+00
Sn 1.22e-09 9.52e-09 6.54e+00
mCEDA-F 1.46e-14 2.23e-13 7.86e+00
mCEDA-G 1.94e+00 2.41e-01 6.46e+00
Table 9: Comparison (mean fitness) of mCEDA with
MIMIC
Gaussian
Gaussian
and TREE
Gaussian
Gaussian
copula model from
(Salinas-Guti
´
errez et al., 2011).
Schwefel’s 1.2 Elliptic
MIMIC 9.96e-01 1.15e+00
TREE 7.74e-01 3.99e-01
mCEDA-F 2.20e-03 3.42e-01
mCEDA-G 2.85e-03 2.78e+01
comparable, for Rastrigin’s and the Sphere functions
(the Sphere function is a simplified version of Ellip-
tic problem) our algorithm mCEDA-F is better, but
mCEDA-G is worse.
In Table 9 a comparison with the algorithm using
the MIMIC
Gaussian
Gaussian
and TREE
Gaussian
Gaussian
copula models
(Salinas-Guti
´
errez et al., 2011) after 100,000 evalua-
tions is shown. In the case of Schwefel’s 1.2 function
mCEDA-F and mCEDA-G achieve evidently better
behavior. In the case of the Elliptic problem mCEDA-
F is better than TREE and MIMIC versions, on the
other hand mCEDA-G is worse than the TREE and
MIMIC versions. Unfortunately (Salinas-Guti
´
errez
et al., 2011) have used 12 dimensions and the vari-
able domain is narrower than in our case, so the mu-
tual comparison is only partly true.
6 CONCLUSION
In this paper we have introduced the utilization
of multivariate elliptic and Archimedean copulas
as cases of the probability model in the Estima-
tion of Distribution Algorithm with model migration
(mCEDA). We have presented the main theoretical
basis and an effective approach of constructing and
sampling both classes of copulas.
In order to illustrate the performance of the island-
based mCEDA algorithm, a few known benchmarks
of optimization were used. From the experimental re-
sults it follows:
1. mCEDA with model migration performs evi-
dently better than the sequential sCEDA.
2. mCEDA using the elliptic copulas performs better
than the Archimedean version of mCEDA.
We have also compared the performance of our
mCEDA algorithm with other published copula EDA
algorithms (M
´
endez and Landa, 2012; Zhao and
Wang, 2012; Jia et al., 2013; Salinas-Guti
´
errez et al.,
2011). Our mCEDA algorithm is evidently better in
most cases.
Our future research will be focused on the uti-
lization of different implementations of univariate
marginal distributions and their modification during
the evolution process. An additional problem seems
to be the tuning of the whole learning process during
model combination in the migration mode.
ACKNOWLEDGEMENTS
This work was supported by the Brno University of
Technology project FIT-S-14-2297.
REFERENCES
Aas, K. (2004). Modelling the dependence structure of fi-
nancial assets: A survey of four copulas.
Aas, K., Czado, C., Frigessi, A., and Bakken, H. (2009).
Pair-copula constructions of multiple dependence. In-
surance: Mathematics and Economics, 44(2):182
198.
Cherubini, U., Luciano, E., and Vecchiato, W. (2004). Cop-
ula Methods in Finance. John Wiley & Sons, Hobo-
ken, NJ.
De Bonet, J. S., Isbell, C. L., and Viola, P. A. (1997).
MIMIC: Finding optima by estimating probability
densities. In Advances in Neural Information Pro-
cessing Systems, volume 9, pages 424–430. The MIT
Press, Cambridge.
delaOssa, L., G
´
amez, J. A., and Puerta, J. M. (2004). Migra-
tion of probability models instead of individuals: An
alternative when applying the island model to edas.
In Parallel Problem Solving from Nature - PPSN VIII,
volume 3242 of LNCS of Lecture Notes in Computer
Science, pages 242–252. Springer.
ECTA 2015 - 7th International Conference on Evolutionary Computation Theory and Applications
218
delaOssa, L., G
´
amez, J. A., and Puerta, J. M. (2005). Im-
proving model combination through local search in
parallel univariate edas. In Congress on Evolutionary
Computation, volume 2, pages 1426–1433. IEEE.
Fr
¨
uhwirth-Schnatter, S. (2006). Finite Mixture and Markov
Switching Models. Springer, New York.
Hauschild, M. and Pelikan, M. (2011). An introduction
and survey of estimation of distribution algorithms.
Swarm and Evolutionary Computation, 1(3):111
128.
Hyr
ˇ
s, M. and Schwarz, J. (2014). Multivariate gaussian
copula in estimation of distribution algorithm with
model migration. In 2014 IEEE Symposium on Foun-
dations of Computational Intelligence Proceedings,
pages 114–119, Piscataway. Institute of Electrical and
Electronics Engineers.
Jia, B., Wang, L., and Cui, Z. (2013). Copula for estima-
tion of distribution algorithm based on goodness-of-fit
test. In Journal of Theoretical and Applied Informa-
tion Technology, number 3, pages 1128–1132.
Larra
˜
naga, P. and Lozano, J. A. (2001). Estimation of Dis-
tribution Algorithms: A New Tool for Evolutionary
Computation. Kluwer Academic Publishers, Norwell,
MA, USA.
Mai, J. and Scherer, M. (2012). Simulating Copulas:
Stochastic Models, Sampling Algorithms, and Appli-
cations, volume 4 of Series in quantitative finance.
Imperial College Press.
Melchiori, M. R. (2006). Tools for sampling multivariate
archimedean copulas. YieldCurve, April.
M
´
endez, M. and Landa, R. (2012). An EDA based on
bayesian networks constructed with archimedean cop-
ulas. In 2012 Fourth World Congress on Nature
and Biologically Inspired Computing (NaBIC), pages
188–193.
Nelsen, R. B. (2006). An Introduction to Copulas. Springer
Series in Statistics. Springer New York.
Pelikan, M., Goldberg, D., and Cant
´
u-Paz, E. (1999). BOA:
The bayesian optimization algorithm. In Proceedings
of the Genetic and Evolutionary Computation Confer-
ence (GECCO-99), volume I, pages 525–532 also Il-
liGAL Report no. 99003.
Pelikan, M. and M
¨
uhlenbein, H. (1999a). The bivariate
marginal distribution algorithm. In Advances in Soft
Computing, pages 521–535. Springer London.
Pelikan, M. and M
¨
uhlenbein, H. (1999b). Marginal dis-
tributions in evolutionary algorithms. In In Proceed-
ings of the International Conference on Genetic Algo-
rithms Mendel 98, pages 90–95.
P
´
oczos, B., Ghahramani, Z., and Schneider, J. (2012).
Copula-based kernel dependency measures. In Lang-
ford, J. and Pineau, J., editors, Proceedings of the
29th International Conference on Machine Learning
(ICML-12), pages 775–782, New York, NY, USA.
ACM.
Rey, M. and Roth, V. (2012). Copula mixture model for
dependency-seeking clustering. In Langford, J. and
Pineau, J., editors, Proceedings of the 29th Interna-
tional Conference on Machine Learning (ICML-12),
pages 927–934, New York, NY, USA. ACM.
Salinas-Guti
´
errez, R., Hern
´
andez-Aguirre, A., and Villa-
Diharce, E. R. (2009). Using copulas in estimation of
distribution algorithms. In MICAI 2009: Advances in
Artificial Intelligence, volume 5845 of Lecture Notes
in Computer Science, pages 658–668. Springer Berlin
Heidelberg.
Salinas-Guti
´
errez, R., Hern
´
andez-Aguirre, A., and Villa-
Diharce, E. R. (2011). Estimation of distribution algo-
rithms based on copula functions. In Proceedings of
the 13th Annual Conference Companion on Genetic
and Evolutionary Computation, GECCO ’11, pages
795–798, New York, NY, USA. ACM.
Schwarz, J. and Jaro
ˇ
s, J. (2008). Parallel bivariate marginal
distribution algorithm with probability model migra-
tion. In Linkage in Evolutionary Computation, volume
157 of Studies in Computational Intelligence, pages
3–23. Springer Berlin Heidelberg.
Sklar, A. (1959). Fonctions de r
´
epartition
`
a n dimensions et
leurs marges. Publications de l’Institut de Statistique
de l’Universit
´
e de Paris, 8:229–231.
Soto, M., Gonz
´
alez-Fern
´
andez, Y., and Ochoa, A. (2012).
Modeling with copulas and vines in estimation of dis-
tribution algorithms. CoRR, abs/1210.5500.
Wang, L.-F., Guo, X., Zeng, J.-C., and Hong, Y. (2010a).
Using gumbel copula and empirical marginal distri-
bution in estimation of distribution algorithm. In
Advanced Computational Intelligence (IWACI), 2010
Third International Workshop on, pages 583–587.
IEEE.
Wang, L.-F., Zeng, J.-C., and Hong, Y. (2009). Estima-
tion of distribution algorithm based on copula theory.
In Evolutionary Computation, 2009. CEC ’09. IEEE
Congress on, pages 1057–1063.
Wang, L.-F., Zeng, J.-C., Hong, Y., and Guo, X. (2010b).
Copula estimation of distribution algorithm sampling
from clayton copula. Journal of Computational Infor-
mation Systems, 6(7):2431–2440.
Zhao, H. and Wang, L. (2012). Marginal distribution in
copula estimation of distribution algorithm based dy-
namic K-S test. In IJCSI International Journal of
Computer Science Issues, number 3, pages 507–514.
Zimmer, D. M. and Trivedi, P. K. (2006). Using trivari-
ate copulas to model sample selection and treatment
effects. Journal of Business & Economic Statistics,
24(1):63–76.
Elliptical and Archimedean Copulas in Estimation of Distribution Algorithm with Model Migration
219