See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221911761

Artificial Neural Networks and Efficient Optimization

Techniques for Applications in Engineering

Chapter · April 2011

DOI: 10.5772/15293 · Source: InTech

CITATIONS READS

5 397

3 authors:

Rossana Moreno Santa Cruz Helton Maia Peixoto

Instituto Federal de Educação, Ciência e Tecnologia da Paraíba Universidade Federal do Rio Grande do Norte
18 PUBLICATIONS   105 CITATIONS    13 PUBLICATIONS   27 CITATIONS   

SEE PROFILE SEE PROFILE

Rafael Magalhães
Universidade Federal da Paraíba
11 PUBLICATIONS   34 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Design and Construction of SPR Optical Sensors View project

All content following this page was uploaded by Rossana Moreno Santa Cruz on 23 September 2014.

The user has requested enhancement of the downloaded file.

 3

Artificial Neural Networks and Efficient

Optimization Techniques for
Applications in Engineering
Rossana M. S. Cruz1, Helton M. Peixoto2 and Rafael M. Magalhães3
1Federal Institute of Education, Science and Technology of Paraiba, João Pessoa, PB,
2Federal University of Rio Grande do Norte, Natal, RN,
3Federal University of Paraiba, Rio Tinto, PB,

Brazil

1. Introduction
This chapter proposal describes some artificial neural network (ANN) neuromodeling
techniques used in association with powerful optimization tools, such as natural
optimization algorithms and wavelet transforms, which can be used in a variety of
applications in Engineering, for example, Electromagnetism (Cruz, 2009), Signal Processing
(Peixoto et al., 2009b) and Pattern Recognition and Classification (Magalhães et al., 2008).
The application of ANN models associated with RF/microwave devices (Cruz et al., 2009a,
2009b; Silva et al., 2010a) and/or pattern recognition (Lopes et al., 2009) becomes usual. In
this chapter, we present neuromodeling techniques based on one or two hidden layer
feedforward neural network configurations and modular neural networks − trained with
efficient algorithms, such as Resilient Backpropagation (RPROP) (Riedmiller & Braun, 1993),
Levenberg-Marquardt (Hagan & Menhaj, 1999) and other hybrid learning algorithms
(Magalhães et al., 2008), in order to find the best training algorithm for such investigation, in
terms of convergence and computational cost. The mathematical formulation and
implementation details of neural network models, wavelet transforms and natural
optimization algorithms are also presented.
Natural optimization algorithms, which are stochastic population-based global search
methods inspired in nature, such as genetic algorithm (GA) and particle swarm
optimization (PSO) are effective for optimization problems with a large number of design
variables and inexpensive cost function evaluation (Kennedy & Eberhart, 1995; R. Haupt &
S. Haupt, 2004). However, the main computational drawback for optimization of nonlinear
devices relies on the repetitive evaluation of numerically expensive cost functions (Haupt &
Werner, 2007; Rahmat-Samii, 2003). Finding a way to shorten the optimization cycle is
highly desirable. In case of GA, for example, several schemes are available in order to
improve its performance, such as: the use of fast full-wave methods, micro-genetic
algorithm, which aims to reduce the population size, and parallel GA using parallel
computation (R. Haupt & S. Haupt, 2004; Haupt & Werner, 2007). Therefore, this chapter
46 Artificial Neural Networks

also describes some hybrid EM-optimization methods, using continuous-GA and PSO
algorithms, blended with multilayer perceptrons (MLP) artificial neural network models.
These methods are applied to design spatial mesh filters, such as frequency selective
surfaces (FSSs). Moreover, the MLP model is used for fast and accurate evaluation of cost
function into continuous GA and PSO simulations, in order to overcome the computational
requirements associated with full wave numerical simulations.
Wavelets and artificial neural networks (ANN) have generated enormous interest in recent
years , both in science and practical applications (Peixoto et al., 2009c). The big advantage of
using wavelets is the fact of these functions make a local behavior, not only in the frequency
domain but also in the field space and time. ANNs are capable of learning from a training
set, which makes it useful in many applications, especially in pattern recognition. The
purpose of using wavelet transforms is to find an easier way to compress and extract the
most important features present in images, thereby creating a vector of descriptors that
should be used to optimize the pattern recognition by a neural network. The array of
descriptors contains elements whose values accurately describe the image content, and
should take up less space than a simple pixel by pixel representation. The greatest difficulty
found in this process is the generation of this vector, where the image content of interest
should be very well described, in order to show really relevant features. Thus, this chapter
proposal also shows a way to use a multiresolution technique, such as the wavelet
transforms, to perform filtering, compression and extraction of image descriptors for a later
classification by an ANN.
This chapter is organized in six sections. Section 2 presents the most important
fundamentals of artificial neural networks and the methodology used for the investigated
applications. In section 3, artificial neural networks are optimized using wavelet transforms
for applications in image processing (extraction and compression). Section 4 presents an
EM-optimization using artificial neural networks and natural optimization algorithms for
the optimal synthesis of stop-band filters, such as frequency selective surfaces. Section 5
shows a modular artificial neural network implementation used for pattern recognition and
classification. Finally, section 6 presents important considerations about the artificial neural
network models used in association with efficient optimization tools for applications in
Engineering.

2. Artificial Neural Networks

2.1 Fundamentals
Rosenblatt (1958) perceptron is the most used artificial neuron in neural network
configurations and is based on the nonlinear model proposed by McCulloch and Pitts
(1943). In this model, neurons are signal processing units composed by a set of input
connections (weights), an adder (for summing the input signals, weighted by the respective
synapses of a neuron, constituting a linear combiner) and an activation function, that can be
linear or nonlinear, as shown in Fig. 1(a). The input signals are defined as xi, i = 0, 1,…, Ni,
whose result corresponds to the level of internal activity of a neuron netj, as defined in (1),
where x0 = +1 is the polarization potential (or bias) of the neurons. The output signal yj is the
activation function response ϕ(⋅) to the activation potential netj, as shown in (2) (Silva et al.,
2010b).
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 47

Ni
net j = ∑ w ji ⋅ xi (1)
i =0

y j = ϕ (net j ) (2)

For a feedforward neural network (FNN), the artificial neurons are set into layers. Each
neuron of a layer is connected to those of the previous layer, as illustrated in Fig. 1(b). Signal
propagation occurs from input to output layers, passing through the hidden layers of the
FNN. Hidden neurons represent the input characteristics, while output neurons generate
the neural network responses (Haykin, 1999).
Modular artificial neural network is based on a principle commonly used: divided to
conquer. This concept aims to divide a large and complex task in a set of sub-tasks that are
easier to be solved. The modular artificial neural network could be defined, in summary, as
a set of learning machines, also called experts, whose decisions are combined to achieve a
better answer than the answers achieved individually, that is, a machine with a better
performance.
In the past few years, one of the main areas of learning machine is the characterization of
methods capable to design this kind of machines. There are two types of these machines:
static and dynamic structures. Modular neural networks, as seen in Fig. 1(c), is a dynamic
type. The input signal is used by the gating network to design the global response. An
advantage of modular artificial neural networks, when compared with other neural
networks, is the learning speed. The machine learning process is accelerated in case of
problems where it is observed a natural decomposition of data at simple functions. To
develop the modular machine architecture and to implement the experts, it is usual to apply
multilayer perceptrons (MLP) neural networks.

2.2 Methodology
Generally, the design of a neural network is composed by three main steps: configuration −
how layers are organized and connected; learning – how information is stored;
generalization − how neural network produces reasonable outputs for inputs not found in
the training (Haykin, 1999). In this work, we use feedforward and modular neural networks
associated with supervised learning to develop neural network models.
In the computational simulation of supervised learning process, a training algorithm is used
for the adaptation of neural network synaptic weights. The instantaneous error e(n), as
defined in (3), represents the difference between the desired response, d(n), and the neural
network output, z(n), at the iteration n, corresponding to the presentation of the nth training
pattern [x(n);(d(n)] (Silva et al., 2010b).

e(n) = z(n) − d(n) (3)

Supervised learning can be illustrated through the block diagram of Fig. 2(a) and has as
objective the minimization of the mean square error E(t), given in (4), where the index t
represents the number of training epochs (one complete presentation of all training
examples, n = 1, 2,…, N, where N is the total number of examples, called an epoch) (Silva et
al., 2010b).
48 Artificial Neural Networks

1 N 1
E(t ) = ∑ e(n)2
N n=1 2
(4)

Currently, there are several algorithms for the training of neural networks that use different
optimization techniques (Peixoto et al., 2009a). The most popular training algorithms are
those derived from backpropagation algorithm (Rumelhart et al., 1986). Among the family
of backpropagation algorithms, the RPROP shows to be very efficient in the solution of
complex electromagnetic problems. In this work, the stop criteria are defined in terms of the
maximum number of epochs and/or the minimum error and the activation function used
was the sigmoid tangent (Haykin, 1999).
After training, the neural network is submitted to a test, in order to verify its capability of
generalizing to new values that do not belong to the training dataset, for example, parts of
the region of interest where there is not enough knowledge about the modeled
device/circuit. Therefore, the neural network operates like a “black box” model that is
illustrated in Fig. 2(b) (Silva et al., 2010b).
Resilient backpropagation algorithm is a first-order local adaptive learning scheme. The
basic principle of RPROP is to eliminate the harmful influence of the partial derivative size
in the weight update. Only the sign of the derivative is considered to indicate the direction
of the weight update, Δwhp , as given in (5).

⎧ ∂E
⎪ −Δ hp (t ), if ∂w (t ) > 0
⎪ hp
⎪⎪ ∂E
Δwhp (t ) = ⎨ +Δ hp (t ), if (t ) < 0 (5)
⎪ ∂whp
⎪0, elsewhere
⎪
⎪⎩

The second step of RPROP algorithm is to determine the new update-values Δhp(t). This is
based on a sign-dependent adaptation process, similar to the learning-rate adaptation
shown by Jacobs (1988). The changes in the weight size are exclusively determined by a
weight ‘update-value’, Δhp, as given in (6).

⎧ + ∂E ∂E
⎪η ⋅ Δ hp (t − 1), ∂w (t − 1) ⋅ ∂w (t ) > 0
⎪ hp hp

⎪⎪ − ∂E ∂E
Δ hp (t ) = ⎨η ⋅ Δ hp (t − 1), (t − 1) ⋅ (t ) < 0 (6)
⎪ ∂whp ∂whp
⎪Δ (t − 1), elsewhere
⎪ hp
⎪⎩

Here, the following RPROP parameters were employed: η+ = 1.2 and η− = 0.5. The update-
values were restricted to the range 10-6 ≤ Δhp ≤ 50.
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 49

Fig. 1. (a) Nonlinear model of an artificial neuron; (b) FNN configuration with two hidden
layers; (c) Extended modular neural network configuration with K experts
50 Artificial Neural Networks

Fig. 2. (a) Block diagram of supervised learning; (b) neural network “black box” model

3. Artificial Neural Networks optimization using wavelet transforms

This section introduces the main concepts about wavelet transforms and some of their most
important features used to optimize artificial neural networks: the extraction of
characteristic information descriptors in order to improve training and pattern classification
mechanisms. Moreover, a model that can be used to solve various problems related to
pattern recognition is presented and a neural classifier is implemented to validate the
importance of such optimization. Transforms are widely used mathematical tools to
understand and analyze different signal behaviors. The objective of this analysis is to extract
important information (or features) that can essentially represent the signal from some
decomposition or transformation performed on it.

3.1 Wavelet transforms

Wavelet transforms have generated enormous interest from scientists, resulting in the
development of applications in various areas, such as computer vision (Wang et al., 2009),
seismology (Parolai et al., 2008), radar (Masnadi-Shirazi et al., 2009), astronomy (Ottensamer
et al., 2008), image compression (Bhatia et al., 2009), signal filtering (Vimal et al., 2009),
system optimization (Pinto, 2009) and many others. In general, the major advantage of using
wavelet transforms is the possibility of applying it to non-stationary signals, which allows
the study of function local behaviors, in both frequency and time-scale domains.

3.1.1 Advantages of use

Traditional methods of signal analysis based on Fourier transform can determine all the
frequencies present in the signal, however its relationship to the time domain does not exist.
To overcome this problem, it was created the Gabor transform (or STFT - Short Time Fourier
Transform); the main idea of this transform is to introduce a new measure of local frequency
as the local transformation observed the signal through a narrow window within which the
signal remains nearly stationary (Oliveira, 2007). The problems in time and frequency
domain resolution are result of a physical phenomenon known as Heisenberg's uncertainty
principle (it is impossible to know the exact frequency and time that a signal occurs). This
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 51

phenomenon is independent of the transformation used (Oliveira, 2007). Therefore, the

wavelet transform was developed as an alternative to the Gabor transform to solve the
resolution problem.
Wavelets are mathematical transformations that separate signals in different components
and extract each one of them with a resolution apropriated to its corresponding scale.
According to the transformation characteristics, there is the Continuous Fourier Transform
(CFT), that can be expressed as:

+∞
f ( t )e (
− j 2π ft )
F( w ) = ∫ dt (7)
−∞

Knowing the spectrum F(w) of a signal, it is possible to obtain it in time domain, using the
inverse transform concept, according to (8):

+∞
1 ( − j 2π ft )
f (t ) = ∫ F ( w )e dw (8)
−∞ 2π

On the other hand, the Continuous Wavelet Transform (CWT) is given by:

+∞
1 ∗ ⎛ t −τ ⎞
CWT (τ , a ) = ∫ f (t ) ψ ⎜ ⎟ dt (9)
−∞ a ⎝ a ⎠
and its corresponding inverse can be expressed according to (10):

+∞ +∞
1 ⎧ 1 ⎛ t − τ ⎞ ⎫ dadb
f (t ) = ∫ ∫ CWT ( a , b ) ⎨ ψ ⎜ ⎟⎬ 2 (10)
Cψ −∞ −∞ ⎩ a ⎝ a ⎠⎭ a

where Ψ(t ) is the mother wavelet,  and a are the translation and scale parameters,
respectively.

3.1.2 Discrete wavelets

The continuous wavelet transform is calculated by performing continuous translations and
scalings of a function over a signal. In practice, this transformation is not feasible because it
is necessary to make endless translations and scalings, requiring much time, effort and
computational redundancy. Discrete wavelets were introduced to overcome this problem
and therefore they are used in this work. They are not translated or scaled continuously but
in discrete steps, which is achieved from a modification of the continuous wavelet:

1 ⎛ t −τ ⎞
ψ s ,τ (t ) = ψ⎜ ⎟ (11)
s ⎝ s ⎠

1 ⎛ t − kτ 0 s0j ⎞
ψ j , k (t ) = ψ⎜ ⎟⎟ (12)
⎜ sj
s0j ⎝ 0 ⎠

where j and k are integers; s0 > 1 is a fixed dilation parameter; τ 0 is the translation factor,
which depends on the dilation factor.
52 Artificial Neural Networks

Generally, it is chosen s0 = 2 in order to have a sampling frequency called dyadic sampling

τ0 = 1 is chosen for the temporal sampling, also dyadic This can be shown in (13) (Oliveira,
2007):

(
ψ j , k (t ) = 2 jψ 2 j t − k ) (13)

When discrete wavelets are used to analyze a signal, the result is a series of wavelet
coefficients, also called the series of wavelet decomposition (Oliveira, 2007). As a wavelet
can be viewed as a bandpass filter, the scaled wavelet series can be seen as a set of bandpass
filters, with a Q factor (set of filters fidelity factor). In practice, there is a discretized wavelet,
with upper and lower limits for translations and scales. The wavelet discretization,
associated with the idea of passing the signal through a filter set, results in the well known
subband coding (Oliveira, 2007).

3.1.3 Multiresolution analysis

Mathematical transformations are used in a dataset to obtain additional information not
available in the primitive data model. For example, it may be necessary to use a
transformation that detects changes in color tones of a pixel neighborhood and its
corresponding spatial location, and even that efficiently transposes these changes in a
multiresolution space (Castleman, 1996). The multiresolution analysis using wavelet
transforms has become increasingly popular with the release of JPEG-2000 standard (Weeks,
2007) and consists of a signal processing strategy where it is used a set of specialized filters
to extract signal information, such as the range of frequencies present in it and their location
as a function of the signal duration at different resolutions (Castleman, 1996). A brief
description of the multiresolution analysis enables to display two functions responsible for
generating the entire wavelet system: the scale function and the primary wavelet (or mother
wavelet). The term mother comes from the fact that functions with different sizes are used in
the process of transformation and all of them are originated from a specific or mother
wavelet.
Scale function φ j , k and the primary wavelets ψ j , k are considered orthogonal to follow the
condition shown in (14):

+∞
∫ φ j , k ( x )ψ j , k ( x ) dx = 0 (14)
−∞

where j ∈ Z corresponds to the scale function parameter and k ∈ Z corresponds to the

k
translation of j in relation to the scale function and the primary wavelet, given by j = 0
2
and k = 0, respectively. Both scale and wavelet functions are defined in the set of real (R)
numbers, by scalings and translations of the mentioned functions. The translation parameter
corresponds to time information in the transform domain and the scaling parameter is the
process of compression and expansion of the signal (Mallat, 2009). In Fig. 3 is shown an
example of a scale function and a primary Haar wavelet.
Therefore, one can say that multiresolution analysis using discrete time wavelets
corresponds to successive band-pass filtering, through which signals are decomposed at
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 53

Fig. 3. Scale function and the primary Haar wavelet

each step, in terms of approximation and details. Fig. 4 illustrates this procedure being
applied to an input image, where filters are used in successive rows and columns, creating
the new scales. The reverse process, which performs the sum of subspaces, can reconstruct
the original image. The dilation equations expressed by h(k) represent the low-pass filters
that generate the approximations of the original image. However, the translation equations
g(k) represent the high-pass filters and are responsible for obtaining the details of the
original image. The decomposition of detail functions consists of details on vertical (high-
pass filter at the rows and low-pass filters at the columns), details on horizontal (low-pass
filter at the rows and high-pass filters at the columns), details on diagonal (high-pass filter at
the rows and columns).

Fig. 4. Wavelet decomposition

Then, using an input image with scale of (j+1), with m rows and n columns, it is shown in
Fig. 5 a two level image decomposition example.
The original image and its approximations are the lighter areas of the picture. The other
three remaining subpictures correspond to the three detail functions of the original image.
In the second step, the lightest part of the picture is decomposed again, generating a new
image approximation and three new detail subpictures. Thus, a twice smaller scale image
54 Artificial Neural Networks

Fig. 5. Multiresolution representation of an image

was generated. Fig. 6 shows an initial image and the degree of refinement obtained with the
wavelet transform. This form of decompose and reconstruct images can be implemented
quickly and effectively, because of the use of wavelet transforms.

Fig. 6. Multiresolution analysis of an image

3.2 Wavelet pre-processing

Pre-processing step aims to improve the performance of the next steps. Then, the refinement
of these data that will be used for training and classification are of fundamental importance.
Designing a neural classifier consists of choosing architecture, training algorithm, quality of
training dataset, among other aspects that must be optimized, in order to reduce the time
used in network training at the same time that the accuracy of the results is increased. This
optimization can be performed using wavelet transforms. The block diagram described in
Fig. 7 shows that the data input used by the neural network may be submitted to a wavelet
pre-processing or not. Here, both the situations are analyzed in parallel.

Fig. 7. Block diagram of a neural network system

Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 55

3.3 Neural network implementation

A simple but effective neural classifier has been implemented at this step, using the classic
aspects of this type of project. It is important to know that even robust classifiers as SVM
(Support Vector Machine) can not function properly without a pre-processing. Fig. 8 shows
a sample training set used in the experiment. This consists of 50 images containing the
vowels A, E, I, O, U, with slight 2D dislocation and no noise.

Fig. 8. Training dataset model

Fig. 9 shows a sample of the test dataset (composed by 25 images), where there is the
inclusion of uncorrelated noise to images up to four levels, aiming to difficult the neural
network generalization.

Fig. 9. Test dataset model

Here, the neural network configuration used the following parameters: a
multilayernperceptron neural network; the Levenberg-Marquardt training algorithm;
number of epochs (150) or Error (10e-5) as stop criteria; two hidden layers and the sigmoid
tangent as activation function.
The results are shown in Table 01, according to two categories of pre-processing: without
wavelet - larger training set (designed pixel by pixel) and higher time to neural network
learning; with wavelet (Daubechies, level 2) - smaller and more efficient training dataset due
to the choice of better wavelet descriptors (higher intensity), which provides a faster
network learning. It is important to observe that the stop criterion was the number of epochs
while the error was kept around 10e-4 and the generalization result was about 92%, which
enabled to recognize images of the vowels a, e, i, o, u with small 2D dislocation levels (<5%)
and noise uncorrelated (<50%).

TT Error
PP CT GE
(min) (e-04)
Without
4096x50 4.27 1.8120 94%
wavelet
96%
With wavelet 1000x50 1.91 7.4728

Table 1. Results: PP - Pre-processing; TD – Training Dataset; TT –Time of Training; GE -

Generalization
The MLP Neural Network was used efficiently in the presented pattern classification,
generalizing in both cases above 94%. The extraction process of features was simplified and
improved by the use of wavelet transform at the pre-processing step in order to optimize all
56 Artificial Neural Networks

the classification step, since the time of training until the generalization of results. It is worth
to say that not only the method, but also its ability of reducing the matrix of data, preserving
the meaning of the information, was of extreme importance to reduce the computational
cost more than twice.

4. Hybrid EM-optimization method for optimal design of FSS

4.1 Frequency selective surfaces
Frequency selective surface (FSS) (Wu, 1995; Munk, 2000; Campos, 2009; Cruz, 2009) is a
two-dimensional periodic array of elements (patches or apertures) in a conducting screen,
which could be either freestanding or supported by dielectric substrates, as shown in Fig.
10(a). FSSs are used for a variety of applications, such as: hybrid radomes (Cruz et al., 2010),
polarizers and dichroic reflectors (Munk, 2000), waveguide filters (Monorchio, 2005),
artificial magnetic conductors (Genovesi et al., 2006) and microwave absorbers (Liu et al.,
2004). To provide high-performance filtering properties at microwave bands,
electromagnetic engineers have investigated various types of FSS devices: reconfigurable
FSS screens, fractal self-similar geometries (Cruz et al., 2009a), pixelized FSS screens,
multilayered FSS structures, as well as FSSs on anisotropic substrates (Silva et al., 2007;
Campos et al., 2001).

Fig. 10. (a) The conventional geometry of a FSS; (b) some patch element shapes; (c) thin
dipole unit cell configuration
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 57

There is no closed form solution directly from a given desired frequency response to the
corresponding FSS (Hwang et al, 1990). The analysis of scattering characteristics from FSS
devices requires the application of rigorous full-wave techniques. Besides that, due to the
computational complexity of using a full-wave simulator to evaluate the FSS scattering
variables, many electromagnetic engineers still use a trial-and-error process until to achieve
a given design criterion. Obviously, this procedure is very laborious and human dependent.
Therefore, optimization techniques are required to design practical FSSs with desired filter
specifications. Some authors have employed neural networks, PSO and GA for FSS design
and optimization (Cruz et al., 2009b; Silva et al., 2010a; 2010b).

4.2 Natural optimization algorithms

Natural optimization algorithms are stochastic population-based global search methods that
start with an initial population of candidate individuals for an optimal solution. Assuming
an optimization problem with Nvar input variables and Npop individuals, the population at
the kth iteration is a matrix P(k)Npop×Nvar of floating-point elements, denoted by pmk ,n , with
each row corresponding to an individual (Cruz, 2009; Silva et al., 2010b). Under GA and
PSO jargons, for example, individuals are named chromosomes and particles (or agents),
respectively. The implementation details of these algorithms are described.

4.2.1 Continuous Genetic Algorithm

Continuous genetic algorithm is very similar to the binary-GA but works with floating-point
variables. Continuous-GA chromosomes are defined by (15), where each one corresponds to
a vector with Nvar floating-point optimization variables. Each chromosome is evaluated by
means of its associated cost, which is computed through the cost function E given in (4)
(Cruz, 2009; Silva et al., 2010b).

chromosome (k ,m) = ⎡⎣ pmk ,1 , pmk ,2 , … , pmk , N var ⎤⎦ , m = 1, 2,… , Npop (15)

cos t(k ,m) = E ( chromosome (k ,m)) (16)

Based on the cost associated to each chromosome, the population evolves through
generations with the application of genetic operators, such as: selection, crossover and
mutation. Fig. 11(a) gives a “big picture” overview of continuous-GA. The mating step
includes roulette wheel selection presented in (Haupt & Werner, 2007; R. Haupt & S. Haupt,
2004). Population selection is performed after the Npop chromosomes are ranked from lowest
to highest costs. Then, the Nkeep most-fit chromosomes are selected to form the mating pool
and the rest are discarded to make room for the new offspring. Mothers and fathers pair in a
random fashion through the blending crossover method (R. Haupt & S. Haupt, 2004). Each
pair produces two offspring that contain traits from each parent. In addition, the parents
survive to be part of the next generation. After mating, a fraction of chromosomes in the
population will suffer mutation. Then, the chromosome variable selected for real-value
mutation is added to a normally distributed random number (Cruz, 2009; Silva et al., 2010b).

4.2.2 Particle Swarm Optimization

Particle swarm optimization algorithm was first formulated by Kennedy and Eberhart (1995).
The thought process behind the algorithm was inspired by the social behavior of animals, such
as bird flocking or fish schooling. PSO algorithm is similar to continuous-GA since it begins
58 Artificial Neural Networks

with a random population matrix. Unlike GA, PSO has no evolution operators such as
crossover and mutation. Each particle moves along the cost surface with an individual
velocity. A flow chart of the PSO algorithm is shown in Fig. 11(b). The implemented PSO
algorithm updates the velocities and positions of the particles based on the best local and
global solutions according to (17) and (18), respectively (Cruz, 2009; Silva et al., 2010b).

⎡
⎣
(
vmk +, n1 = C ⎢ r0 vmk ,n + Γ 1 ⋅ r1 ⋅ pm
local bes( k )
,n ) (
− pmk ,n + Γ 2 ⋅ r2 ⋅ pmglobal
,n
best ( k )
) ⎤
− pmk ,n ⎥
⎦
(17)

pmk +,n1 = pmk ,n + vmk +,n1 (18)

Here, vm,n is the particle velocity, pm,n is the particle variables, r0, r1 and r2 are independent
uniform random numbers; Γ1 and Γ2 are is the cognitive and social parameters, respectively,
localbest ( k )
pm ,n and pmglobalbest
,n
(k)
are the best local and global solutions, respectively; C is the
constriction parameter (Kennedy & Eberhart, 1995). If the best local solution has a cost less
than the best cost of the current global solution, then the best local solution replaces the best
global solution. PSO is a very simple natural optimization algorithm, easy to implement and
with few parameters to adjust.

Fig. 11. Flow charts of the natural optimization algorithms: (a) GA; (b) PSO

4.3 EM-optimization method

The FSS design methodology, which blends MLP models, GA and PSO algorithms, is
applied for a thin dipole FSS synthesis, considering electromagnetic parameters such as
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 59

resonant frequency (fr) and –3dB bandwidth (BW) as a function of the substrate thickness (d)
and the periodicity of elements (t = tx = ty) corresponding to a square unit cell. The
dimensions of the thin dipole elements used in this work are shown in Fig. 10(c). The width
(W) and length (L) of the patch remain the same. The region of interest (or search space)
defined by desired input variables is a rectangle: 15.24 ≤ t ≤ 19.05 mm and 0.1 ≤ d ≤ 2.0 mm.
The FSS optimization problem consists of obtaining an optimal solution (d*; t*) in the search
space that results in a minimal thickness FSS with the desired specifications for resonant
frequency and bandwidth. After the definition of the FSS filter-type, design input variables
and search space, a parametric analysis is performed in order to observe the FSS EM-
behavior. At this step, the FSS transmission coefficients at 6.0−14.0 GHz band were obtained
by means of Method of Moments (MoM) simulations considering substrate anisotropy
(Campos et al., 2001). Representative EM-datasets must be obtained from the parametric
analysis for supervised learning of FSS synthesis MLP models.
The second step consists of modeling fr and BW responses by the synthesis MLP network
using the conventional neuromodeling technique (Zhang & Gupta, 2000). The third step just
consists of implementing the natural optimization algorithms for FSS optimal synthesis. In
particular, the continuous-GA and PSO algorithms were used.
Through the FSS neuromodeling, we avoid the very intensive application of the CPU for
MoM analysis in GA and PSO simulations. The ANN model was developed for FSS
synthesis using one hidden layer MLP configuration. The MLP learning processes were
carried out using the RPROP training algorithm, with standard parameters (Riedmiller &
Braun, 1993), in order to minimize the mean square error as defined in (4). In this case, the
RPROP training is performed until the mean square error reaches a minimum pre-
established value. The synthesis MLP network configuration (composed by three inputs: -1,
d, t; twenty hidden neurons and two outputs: fr, BW) simultaneously approaches both the
mappings fr(d,t) and BW(d,t). The MLP configuration is similar to that presented in Fig. 1(b).
To generate the synthesis training dataset were assumed the vectors given in (19) for the
design input variables d and t, adding up 72 training examples.

⎪⎧d = [ 0.1 0.3 0.5 0.7 0.9 1.0 1.2 1.4 1.5 1.6 1.8 2.0] mm
⎨ (19)
⎪⎩tx = ty = t = [ 15.24 15.87 16.51 17.14 18.41 19.05] mm

GA and PSO algorithms were implemented in Matlab® for optimal FSS synthesis. The
assumed design input variables (substrate thickness and periodicity) were symbolized by
k
dmk and t m to take into account the m-th individual at the k-th iteration. Given a desired FSS
filter specification, (frdesired, BWdesired), the aim is the minimization of the quadratic cost
function as defined in (20) in terms of absolute percent errors.

( ) ( ) ⎞⎟
2
⎛ fr k k
BWdesired − BW dmk , tmk
⎜ desired − fr dm , tm (20)
cos t( k , m) = ⎜ + ⎟ ,
⎜ f desired BWdesired ⎟
⎝ ⎠

m=1,2,…Npop
To evaluate the cost function shown in (20), the synthesis MLP approach the EM relations
( ) ( )
f r dmk , tmk and BW dmk , tmk . When the population evolves, each individual is constrained to
the region of interest using (21), where the dummy variable ξ can be replaced by d or t.
60 Artificial Neural Networks

( ( )
ξmk = min max ξmk , ξmin , ξmax ) (21)

The parameters used for continuous-GA and PSO simulations are shown in Table 2.

Continuous-GA PSO
initial population = 50 initial population = 25
maximum iteration number maximum iteration number = 200
(generations) = 500
constriction parameter C = 0.8
crossover probability = 0.9
cognitive parameter, Γ1 = 2
mutation rate = 0.01 cognitive parameter, Γ2 = 2
Table 2. GA and PSO parameters used in the simulations

Fig. 12. Tranmission coefficients of the thin dipole FSS: (a) first and (b) second examples
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 61

In this section, examples that use the hybrid EM-optimization technique for the optimal
synthesis of FSSs are presented. Random values for the resonant frequency and bandwidth
were chosen within the search space used for the MoM simulations of the synthesis MLP
network. The first example corresponds to output data (fr; BW) = (10.5; 1.5) GHz and the
second example corresponds to output data (fr; BW) = (11.0; 2.5) GHz.
Figs. 12(a) and 12(b) show the transmission coefficients obtained from MoM, GA and PSO
simulations for the first and second examples, respectively.
It is observed that although GA algorithm response is closer to MoM simulations, PSO
algorithm presents the best solutions of resonant frequency and bandwidth, compared to
the desired values. Table 3 shows the input/output values obtained from both algorithms.
The cost surface contours, the initial, intermediate, and final populations, as well as the best
path for both algorithms can be found in (Silva et al., 2010a; 2010b). One can also find a new
version of GA algorithm called improved GA in (Cruz, 2009).

First example investigated - (fr; BW) = (10.5; 1.5) GHz

Results
Parameters
GA PSO
Thickness of the substrate d (mm) 0.7685 0.5401
Periodicity of elements t (mm) 18.193 15.283
obtained fr (GHz) 10.7407 10.4917
obtained BW (GHz) 1.3827 1.8632
Second example investigated - (fr; BW) = (11.0; 2.5) GHz
Results
Parameters
GA PSO
Thickness of the substrate d (mm) 0.195 1.9457
Periodicity of elements t (mm) 15.826 19.006
obtained fr (GHz) 11.4444 11.1121
obtained BW (GHz) 2.2716 2.2924
Table 3. Input/output values obtained from GA and PSO algorithms

5. Hybrid and high processing pattern classification

In recent years, literature (Dieterrich, 1998) indicated some development directions in
machine learning which have a highlight in this research scene, including the committee
machines and also methods for scaling up supervised learning algorithms. Ensembles
studies, which deal with aggregation among many learning machines on the resolution of
others complex problems, had an increased effort with empiric data, formalization and new
methods (Valentini & Masulli, 2002; Kuncheva, 2004; Dimitrakakis & Bengio, 2005). New
training methods, such as hierarchy and parallel processing, have been developed and
evaluated in several aplications.
This section describes a pattern classification problem that deal with electric power line
systems. The quality of the energy provided by an electric system is one of the greatest
points of interest for concessionaires and electric energy consumers. Literature presents
distinct approches in the acquisition, characterization and classification of disturbs that may
occur in high potential transmission nets. Some pre-processing methods, like wavelet
transform, are used for characterization of voltage or current signals.
62 Artificial Neural Networks

Considering high performance and hybrid architectures neural networks as a develoment

pespective, this work treats a learning algorithm for extended modular neural networks.
The motivation is the unsatisfactory results in pattern classifications and function
approximations commonly used (Jacobs et al, 1991). In such cases, modular neural experts
are composed by a single linear neuron layer and training process is performed through
associative gaussian mixture models.
The architecture proposed for solution of this problem uses multilayer preceptrons experts,
as shown in Fig. 1(c). The modular neural network has K experts, L layers of q neurons with
linear or nonlinear activation function and bias, as well as a gating network with similar
architecture.

5.1 The hybrid algorithm

The modified modular artificial neural network is trained with a developed algorithm
adapted according to (Jacobs et al., 1991), for the model of Gaussian mixing associative and
also the error backpropagation algorithm, by including the calculation of the descend
gradient. The main steps of this algorithm is in this section; more details can be found in
(Magalhães et al., 2008).
In this description, some conventions are utilized: the stucture has MLP experts (indexed by
i = 1, ..., k) with LEspi layers (indexed by l = 1, ..., LEspi), with q(l)Espi neurons in each layer
(indexed by j = 1, ..., q(l)Espi). Also, there is a MLP gating network with LPas layers (indexed by
l = 1, ..., LPas), with q(l)Pas neurons in each layer (indexed by j = 1, ..., q(l)Pas). The neurons
activation functions in all networks can be linear or non-linear. The processing algorithm
can be divided in three steps.

5.1.1 First step

The first step of the algorithm is the calculation of a priori probability associated with the i-th
output layer of the gating network, as expressed by (22):

gi ( n ) =
(
exp ui( l ) (n) ) (22)
∑
K
j =1
exp ( ui( l ) (n) )
where ui(l)(n) is the i-th output neuron of the l-th layer of the gating network and is giving
by:

(l)
uil (n) = ϕ pi (
v(pil ) (n) ) (23)

(i ) (i )
where ϕ pi is the derivative function from the activation potential υ pi , from the i-th neuron
of the l-th layer of the gating network, as shown in (24):

q ( l ) Pas
v(pil ) (n) = ∑ aij( l )uil − 1 (n) (24)
j =0

where aij is the weight associated to the j-th input neuron i of the l-th layer of the gating
network.
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 63

5.1.2 Second step

The second step of the algorithm is responsible for evaluating the values of posteriori
probabilities hi(n), associated to the i-th output neuron of the output gating network. The
posteriori probabilities can be expressed as:

−1
||d(n) − y e( K ) (n)||2 )
gi (n)exp(
hi (n) = K 2 i
(25)
−1
∑ g j (n)exp( 2 ||d(n) − y ei( K ) (n)||2 )
j =1

where d(n) is the desired response and yei(k)(n) is the actual response of the i-th neuron of the
l-th layer of the K-th expert, as shown in (26):

y ( l( K) ) (n) = ϕ ((lK) ) ( v( l( K) ) (n)) (26)

ei ei ei

where φ(l) is the activation potential derivative v(l) from the i-th neuron, l-th layer and k-th
expert.

5.1.3 Third step

The principal point in this algorithm is the incrementation in synaptic weights of the
modular network, that is done with the multiple layers. The synaptic weights from the
networks experts are updated according to:

we((lK) ) j (n + 1) = we((lK) ) j (n) + ηδ e((lK) ) (n) y (jl −1) (n) (27)

i i i

where η is the learning rate, and the gradient δ ei( i()k ) for the output layer neurons is obtained
by:

δ e((lK) ) (n) = hi (n)ei (n)( l )ϕ e'((lK)) ( v(el( K) ) (n)) (28)

i i i

where ei is the difference between di and ye(K). The gradient for the neurons of hidden layers
is computed as:

q( l +1) Espi
( l + 1)
δ e((lK) ) (n) = ϕ e'((lK)) ( v(el( K) ) (n))
i i i
∑ δ m( l + 1) (n)wme (K )(n) (29)
m=1 i

The synaptic weight update of the gating network is accomplished according to (30):

a (pli )j (n + 1) = a (pli )j (n ) + ηδ p(li ) (n ) y (jl −1) (n )

(30)
where the output layer gradient is obtained by:

δ pi( l ) (n) = [ hi (n) − gi (n)]ϕ pi'( l ) ( v(pil ) (n)) (31)

The error is the difference between hi and gi. The gradient of hidden layers is calculated by:
64 Artificial Neural Networks

q( l +1)Pas
( l + 1)
δ pi( l ) (n) = ϕ pi'( l ) ( v( l ) ∑ δ m( l + 1) (n)ampi (n)) (32)
m=1

Thus, the error backpropagates from the gating network to its hidden layers.

5.2 The application and results

As mentioned before, the application implemented was a pattern classification for power
line disturbances. The implementation methodology, data files and information were the
same presented in (Medeiros et al., 2007). The application evaluated the performance of an
intelligent system classifier, in this case, a modular artificial neural network, in electric
disturbances classification. The approach is performed into four main steps: getting the
signal, pre-processing, definition and classification of descriptors.
First, the electrical signals were obtained using the oscillograph network of São Francisco
Hydro Electric Company (CHESF) and also from the simulation via Transient Alternative
Program (ATP). The electric network consists of 370 oscillographs operation with a
sampling rate ranging between 20 and 256 samples/cycle. Here, the signals were collected
in voltage line levels of 69, 230 and 500 kV, with a rate of 128 samples/cycle during 14
cycles. The pre-processing stage aimed to suggest descriptors that characterize the signal
variations when diverted from a certain standard. The third step, which deals with the
descriptor definitions, is performed by the decomposition of signals from the previous step.
Once obtained the descriptors, four disturbances classes were defined as: Voltage Sag,
Voltage Swell, Harmonics and Transitories. The last step (classification) is performed by the
application of classifiers based on modular artificial neural networks. Several different
architectures were tested, as shown in Table 4.

Net MOD-0 MOD-1 MOD-2 MOD-3

Number of 3 3 3 3
Experts
Expert 10:3:4 10:5:4 10:10:4 10:15:4
Architecture
Gating 10:5:4 10:5:4 10:10:4 10:15:4
Architecture
100
Classification (%) 98,46 99,48 100

Table 4. Modular Neural Network Architecture

The classification step used two datasets, one for the training at computer, and the other for
validation at multiprocessor on chip, utilizing a FPGA system architecture. The training set
consisted of 800 patterns formed by 4 output classes. To validate the modular neural
network, 344 input patterns were used with their respective expected responses, consisting
only of data obtained from the oscillographs.
Table 4 shows that the modular neural network with the proposed algorithm reaches a high
amount of accuracy, approximately 100%. Besides, the utilization of two processors on a
FPGA system increased the optimization speed-up by 47%, while using four processors, the
speed-up was increased by 87%.
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 65

6. Conclusion
This chapter described some of the neuromodeling methodology used for applications in
various areas of Engineering, in particular, EM-optimization, Signal Processing and Pattern
Classification and Recognition. Some original contributions were shown, such as the hybrid
EM-optimization for optimal design of FSS, the artificial neural network optimization using
wavelet transforms and the modular neural network used to pattern classification and
recognition.
The choice of the activation function of a FNN or a modular neural network strongly
influences the neural model performance. However, there is not an universal activation
function that can be used to solve any kind of nonlinear modeling problem. Using
additional information from the hybrid neural models, as well as sharing the training
process with modular neural networks, increased the efficiency and therefore the
generalization ability and model reliability of the resulting neural models.
The fast and accurate obtained results for all the applications demonstrated the
improvements with the utilization of these models, and proved that the MLP network
global approximations are able to generalize. In addition, the idea of blending artificial
neural networks and natural optimization algorithms, as well as mathematical transforms
(for EM-optimization and pattern recognition, respectively) shows to be very interesting.
PSO algorithm needs less individuals and reaches the best solution at a few iterations,
compared to GA algorithm. The PSO implementation is simpler than GA, since it is not
required the presence of genetic operatiors such as crossover and mutation. Therefore, PSO
algorithm has proved to be an interesting EM-optimization tool, with few parameters to
adjust and low computational cost.
The characteristics of ANN models (precision, CPU efficiency and flexibility) can be
perfectly used in association with these optimization techniques in order to develop
powerful soft computing tools. A good point in combining these complementary tools is the
possibilty of multiprocessing application, using parallel processors and computers, not only
to increase performance and execution speed, but also to enable an improved type of
competition and collaboration. Each individual tool allows solving the problem or part of
the problem and, at the same time, they can collaborate one each other to improve the
solution in a higher level. This is the intelligent computing that this chapter wants to apply
in efficient engineering.

7. References
Beale, E. M. L. (1972). A derivation of conjugate gradients, In: Numerical Methods for
Nonlinear Optimization, F. A. Lootsma, (Ed.), Academic Press, London.
Bhatia, P.; Bharti, S. & Gupta, R. (2009). Comparative analysis of image compression
techniques: a case study on medical images, In: International Conference on Advances
in Recent Technologies in Communication and Computing, pp. 820-822.
Campos, A. L. P. de S. (2009). Superfícies seletivas em frequência: análise e projeto, IFRN, 198 p.,
Natal, RN, Brazil.
Campos, A. L. P. S.; D’Assunção, A. G. & Melo, M. A. B. (2001). Investigation of scattering
parameters for frequency selective surfaces using thin dipoles on anisotropic layers.
Proceedings of the SBMO/IEEE MTT-S International Microwave and Optoelectronics
Conference, Vol. 1, pp. 413-416.
66 Artificial Neural Networks

Castleman, K. R. (1996). Digital Imaging Processing. Prentice Hall, 1st Ed, Englewood Cliffs,
New Jersey.
Cruz, R. M. S. (2009). Análise e otimização de superfícies seletivas de frequência utilizando
redes neurais artificiais e algoritmos de otimização natural, Ph.D. Dissertation,
Federal University of Rio Grande do Norte, Natal, RN, Brazil.
Cruz, R. M. S.; Santos, A. F. & D’Assunção, A. G. (2010). Spectral analysis of electromagnetic
waves in uniaxial anisotropic dielectric materials, Proceedings of 14º SBMO –
Simpósio Brasileiro de Microondas e Optoeletrônica and 9º CBMag – Congresso
Brasileiro de Eletromagnetismo, MOMAG2010, pp. 440-444, Vila Velha, ES, Brazil.
Cruz, R. M. S.; Silva, P. H. da F. & D’Assunção, A. G. (2009a). Neuromodeling stop-band
properties of Koch island patch elements for FSS filter design. In: Microwave and
Optical Technology Letters, Vol. 51, No. 12, pp. 3014-3019.
__________. (2009b). Synthesis of crossed dipole frequency selective surfaces using genetic
algorithms and artificial neural networks, Proceedings of the 2009 International Joint
Conference on Neural Networks, pp. 627-633, ISBN 978-1-4244-3553-1, Atlanta, GA,
USA.
Dietterich, T. G. (1998). Machine-learning research: four current directions, The AI Magazine,
Vol. 4, pp. 97–136.
Dimitrakakis, C. & Bengio, S. (2005). Online adaptive policies for ensemble classifiers.
Neurocomputing, Vol. 64, pp. 211-221.
Genovesi, S. et al. (2006). Particle swarm optimization of frequency selective surfaces for the
design of artificial magnetic conductors, In: IEEE Antennas and Propagation Society
International Symposium, Vol. 9, No. 14, pp. 3519–3522.
Hagan, M. T. & Menhaj, M. (1999). Training feed-foward networks with the Marquardt
algorithm. IEEE Transactions on Neural Networks. Vol. 5, No. 6, pp. 989-993.
Haupt, R. L. & Haupt, S. (2004). Practical genetic algorithms, Wiley & Sons, Inc., 2nd Ed, New
Jersey.
Haupt, R. L. & Werner, D. H. (2007). Genetic algorithms in electromagnetic, John Willey & Sons,
Inc., 1st Ed, New York.
Hwang, J. N.; Chan, C. H. & Marks, R. J. (1990). Frequency selective surface design based on
iterative inversion of neural networks. Proceedings of the International Joint Conference
on Neural Networks, Vol. 1, pp. 39-44.
Haykin, S. (1999). Neural networks: a comprehensive foundation, Prentice-Hall, 2nd Ed, Upper
Saddle River, New Jersey.
Jacobs, R. A.; Jordan, M. I.; Nowlan, S. J. & Hinton, G. E. (1991). Adaptive mixture of local
experts. Neural Computation, MIT Press, Vol. 3, pp. 79-87.
Kennedy, L. & Eberhart, R. C. (1995). Particle swarm optimization. Proceedings of IV IEEE
Conference On Neural Networks, Piscataway, New Jersey.
Kuncheva, L. I. (2004). Combining pattern classifiers: methods and algorithms, John Wiley &
Sons, Inc., New Jersey.
Liu, J. C. et al. (2004). Implementation of broadband microwave absorber using FSS screens
coated with Ba(MnTi)Fe10O19 ferrite, In: Microwave and Optical Technology Letters,
Vol. 41, No. 4, pp. 323−326.
Lopes, D. C. et al. (2009). Implementation of a Modular Neural Network in a Multiple
Processor System on FPGA to Classify Electric Disturbance. In: Industrial Electronics,
2009. IECON '09. 35th Annual Conference of IEEE, Porto, Portugal.
Artificial Neural Networks and Efficient
Optimization Techniques for Applications in Engineering 67

Magalhães, R. M. et al. (2008). Application of a Hybrid Algorithm in the Modular Neural

Nets Trainning with Multilayers Specialists in Electric Disturbance Classication. In:
INES 2008 12th International Conference on Intelligent Engineering Systems, Miami,
Florida.
Mallat, S. (2009). A Wavelet Tour of Signal Processing, Elsevier Inc, 3rd ed.
Masnadi-Shirazi, R.; Mohseni, A. & Sheikhi, M. A. (2009). Efficient compression of wavelet
packet OFDM radar signals.
McCculloch, W. S. & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous
activity. Bulletin of Mathematical Biophysics, Vol. 5, pp. 115-133.
Medeiros, M. F. et al. (2007). Influence of signal processing in the Efficiency of algorithms
Based on Neural Networks for Disturbanc Classification, Computacional Intelligence
in Image and Signal Processing, pp. 95 -100.
Monorchio, A. et al. (2005). Design of waveguide filters by using genetically optimized
frequency selective surfaces, In: IEEE Microwave Wireless Component Letters, Vol. 15,
No. 6, pp. 407–409.
Munk, B. A. (2000). Frequency selective surfaces - theory and design, John Wiley & Sons, Inc.,
ISBN 0-471-37047-9, New York, USA.
Oliveira, H. M. (2007). Análise de Fourier e Wavelets: sinais estacionários e não
estacionários. Ed. Universítária UFPE, 1ª Ed, Recife.
Ottensamer, J.; Bobin, J. & Starck, R. (2008). Compressed sensing in astronomy, In: IEEE
Journal of Selected Topics in Signal Processing, Vol. 2, No. 5, pp. 718-726.
Parolai, J. J.; Galiana-Merino, J. L. & Rosa-Herranz, S. (2008). Seismic P phase picking using a
kurtosis-based criterion in the stationary wavelet domain, In: IEEE Transactions on
Geoscience and Remote Sensing, Vol. 46, No. 11, pp. 3815-3826.
Peixoto, H. M.; Cruz, R. M. S.; Guerreiro, A. M. G.; Dória Neto, A. D. & D’Assunção, A. G.
(2009a). A comparison of multilayer perceptrons training algorithms for the
optimization of frequency selective surfaces, Proceedings of 8th International
Information and Telecommunication Technologies Symposium, Florianópolis, SC, Brazil.
Peixoto, H. M.; Guerreiro, A. M. G. & Dória Neto, A. D. (2009b). Image processing for eye
detection and classification of the gaze direction, Proceedings of the 2009 International
Joint Conference on Neural Networks, pp. 2475-2480, ISBN 978-1-4244-3553-1, Atlanta,
GA, USA.
Peixoto, H. M.; Rego, J. B. A.; Guerreiro, A. M. G. & Dória Neto, A. D. (2009c). Otimizando
redes neurais artificiais com transformadas wavelets, Anais do IX Congresso Brasileiro
de Redes Neurais – Inteligência Computacional, Ouro Preto, MG, Brazil.
Pinto, L. F. Q.; Silveira, L. G. Q. S.; Junior, F. M. & Assis, E. L. (2009). Analysis and
optimization of wavelet-coded communication systems, Vol. 8, No. 2, pp. 563-657.
Rahmat-Samii, Y. (2003). Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) in
Engineering Electromagnetics, Proceedings of 17TH International Conference on Applied
Electromagnetics and Communications, Dubrovnik, Croatia.
Riedmiller, M. & Braun, H. (1993). A direct adaptative method for faster backpropagation
learning: the RPROP algorithm. Proceedings of IEEE International Conference on
Neural Networks, pp. 586–591, San Francisco, USA.
Rosenblatt, F. (1958). The Perceptron: A probabilistic model for information storage and
organization in the brain. Psychological Review, Vol. 65, No. 6, pp. 386-408. Retrieved
September 19, 2010, from
 68 Artificial Neural Networks

http://www.ling.upenn.edu/courses/Fall_2007/cogs501/Rosenblatt1958.pdf.
Rumelhart, D. E.; Hinton, G. E. & Williams, R. J. (1986). Learning internal representations by
error propagation. In: Parallel Distributed Processing: Explorations in the
microstructure of cognition, D. E. Rumelhart & J. L. McClelland (Ed.), MIT Press, Vol.
1, pp. 318-362, Cambridge, MA.
Silva, P. H. da F. & Campos, A. L. P. S. (2008). Fast and accurate modeling of frequency
selective surfaces using a new modular neural network configuration of multilayer
perceptrons, Microwaves, Antennas e Propagation, IET, Vol. 2, No. 5, pp. 503–511.
Silva, P. H. da F.; Cruz, R. M. S & D’Assunção, A. G. (2010a). Blending PSO and ANN for
optimal design of FSS filters with Koch island patch elements, IEEE Transactions on
Magnetics, Vol. 46, No. 8.
__________. (2010b). Neuromodeling and natural optimization of nonlinear devices and
circuits, In: System and Circuit Design for Biologically-Inspired Intelligent Learning,
Turgay Temel (Ed.), IGI Global, ISBN 9781609600181, Hershey PA. No prelo.
Silva, P. H. da F.; Lacouth, P.; Fontgalland, G.; Campos, A. L. P. S. & D'Assunção, A. G.
(2007). Design of frequency selective surfaces using a novel MoM-ANN-GA
technique. Proceedings of the SBMO/IEEE MTT-S International Microwave and
Optoelectronics Conference, Vol. 1, pp. 275-279.
Valentini, G. & Masulli, F. (2002). Ensemble of learning machines, in Neural Nets WIRN
Vietri-02, Series Lectures Notes in Computer Sciences, Springer-Verlag.
Vimal, P. S.; Kumar, R. & Arumugabathan, C. (2009). Wavelet based ocular artifact removal
from EEG signals using arma method and adaptive filtering, pp. 667-671.
Wang, X. Q. G.; Liu, D. G. Z.; Li, Z. & Wang, H. (2009). Object categorization using
hierarchical wavelet packet texture descriptors, In: 11th IEEE International
Symposium on Multimedia, pp. 44-51.
Weeks, M. (2007). Digital Signal Processing Using MATLAB and Wavelets, Infinity Science
Press LLC Hingham, Georgia State University, Massachusetts.
Wu, T. K. (1995). Frequency selective surface and grid array, John Wiley & Sons, Inc., ISBN 0-
471-31189-8, New York, USA.
Zhang, Q. J. & Gupta, C. (2000). Neural networks for RF and microwaves design, Artech House,
Norwood, MA.

View publication stats