Abstract: The backpropagation algorithm in general employs quadratic error function. In fact, most of the problems that involve minimization employ the Quadratic error function. With alternative error functions the performance of the optimization scheme can be improved. The new error functions help in suppressing the ill-effects of the outliers and have shown good performance to noise. In this paper we have tried to evaluate and compare the relative performance of complex valued neural network using different error functions. During first simulation for complex XOR gate it is observed that some error functions like Absolute error, Cauchy error function can replace Quadratic error function. In the second simulation it is observed that for some error functions the performance of the complex valued neural network depends on the architecture of the network whereas with few other error functions convergence speed of the network is independent of architecture of the neural network.
Abstract: The overall objective of this paper is to retrieve soil
surfaces parameters namely, roughness and soil moisture related to
the dielectric constant by inverting the radar backscattered signal
from natural soil surfaces.
Because the classical description of roughness using statistical
parameters like the correlation length doesn't lead to satisfactory
results to predict radar backscattering, we used a multi-scale
roughness description using the wavelet transform and the Mallat
algorithm. In this description, the surface is considered as a
superposition of a finite number of one-dimensional Gaussian
processes each having a spatial scale. A second step in this study
consisted in adapting a direct model simulating radar backscattering
namely the small perturbation model to this multi-scale surface
description. We investigated the impact of this description on radar
backscattering through a sensitivity analysis of backscattering
coefficient to the multi-scale roughness parameters.
To perform the inversion of the small perturbation multi-scale
scattering model (MLS SPM) we used a multi-layer neural network
architecture trained by backpropagation learning rule. The inversion
leads to satisfactory results with a relative uncertainty of 8%.
Abstract: Gabor-based face representation has achieved enormous success in face recognition. This paper addresses a novel algorithm for face recognition using neural networks trained by Gabor features. The system is commenced on convolving a face image with a series of Gabor filter coefficients at different scales and orientations. Two novel contributions of this paper are: scaling of rms contrast and introduction of fuzzily skewed filter. The neural network employed for face recognition is based on the multilayer perceptron (MLP) architecture with backpropagation algorithm and incorporates the convolution filter response of Gabor jet. The effectiveness of the algorithm has been justified over a face database with images captured at different illumination conditions.
Abstract: Classification of electroencephalogram (EEG) signals
extracted during mental tasks is a technique that is actively pursued
for Brain Computer Interfaces (BCI) designs. In this paper, we
compared the classification performances of univariateautoregressive
(AR) and multivariate autoregressive (MAR) models
for representing EEG signals that were extracted during different
mental tasks. Multilayer Perceptron (MLP) neural network (NN)
trained by the backpropagation (BP) algorithm was used to classify
these features into the different categories representing the mental
tasks. Classification performances were also compared across
different mental task combinations and 2 sets of hidden units (HU): 2
to 10 HU in steps of 2 and 20 to 100 HU in steps of 20. Five different
mental tasks from 4 subjects were used in the experimental study and
combinations of 2 different mental tasks were studied for each
subject. Three different feature extraction methods with 6th order
were used to extract features from these EEG signals: AR
coefficients computed with Burg-s algorithm (ARBG), AR
coefficients computed with stepwise least square algorithm (ARLS)
and MAR coefficients computed with stepwise least square
algorithm. The best results were obtained with 20 to 100 HU using
ARBG. It is concluded that i) it is important to choose the suitable
mental tasks for different individuals for a successful BCI design, ii)
higher HU are more suitable and iii) ARBG is the most suitable
feature extraction method.
Abstract: This paper presents a subjective job scheduler based
on a 3-layer Backpropagation Neural Network (BPNN) and a greedy
alignment procedure in order formulates a real-life situation. The
BPNN estimates critical values of jobs based on the given subjective
criteria. The scheduler is formulated in such a way that, at each time
period, the most critical job is selected from the job queue and is
transferred into a single machine before the next periodic job arrives.
If the selected job is one of the oldest jobs in the queue and its
deadline is less than that of the arrival time of the current job, then
there is an update of the deadline of the job is assigned in order to
prevent the critical job from its elimination. The proposed
satisfiability criteria indicates that the satisfaction of the scheduler
with respect to performance of the BPNN, validity of the jobs and the
feasibility of the scheduler.
Abstract: Since the presentation of the backpropagation algorithm, a vast variety of improvements of the technique for training a feed forward neural networks have been proposed. This article focuses on two classes of acceleration techniques, one is known as Local Adaptive Techniques that are based on weightspecific only, such as the temporal behavior of the partial derivative of the current weight. The other, known as Dynamic Adaptation Methods, which dynamically adapts the momentum factors, α, and learning rate, η, with respect to the iteration number or gradient. Some of most popular learning algorithms are described. These techniques have been implemented and tested on several problems and measured in terms of gradient and error function evaluation, and percentage of success. Numerical evidence shows that these techniques improve the convergence of the Backpropagation algorithm.
Abstract: This paper reports the study results on neural network
training algorithm of numerical optimization techniques multiface
detection in static images. The training algorithms involved are scale
gradient conjugate backpropagation, conjugate gradient
backpropagation with Polak-Riebre updates, conjugate gradient
backpropagation with Fletcher-Reeves updates, one secant
backpropagation and resilent backpropagation. The final result of
each training algorithms for multiface detection application will also
be discussed and compared.
Abstract: Gas chromatography (GC) is the most widely used
technique in analytical chemistry. However, GC has high initial cost
and requires frequent maintenance. This paper examines the
feasibility and potential of using a neural network model as an
alternative whenever GC is unvailable. It can also be part of system
verification on the performance of GC for preventive maintenance
activities. It shows the performance of MultiLayer Perceptron (MLP)
with Backpropagation structure. Results demonstrate that neural
network model when trained using this structure provides an
adequate result and is suitable for this purpose. cm.
Abstract: In this paper we present, propose and examine
additional membership functions for the Smoothing Transition
Autoregressive (STAR) models. More specifically, we present the
tangent hyperbolic, Gaussian and Generalized bell functions.
Because Smoothing Transition Autoregressive (STAR) models
follow fuzzy logic approach, more fuzzy membership functions
should be tested. Furthermore, fuzzy rules can be incorporated or
other training or computational methods can be applied as the error
backpropagation or genetic algorithm instead to nonlinear squares.
We examine two macroeconomic variables of US economy, the
inflation rate and the 6-monthly treasury bills interest rates.
Abstract: In this paper the neural network-based controller is
designed for motion control of a mobile robot. This paper treats the
problems of trajectory following and posture stabilization of the
mobile robot with nonholonomic constraints. For this purpose the
recurrent neural network with one hidden layer is used. It learns
relationship between linear velocities and error positions of the
mobile robot. This neural network is trained on-line using the
backpropagation optimization algorithm with an adaptive learning
rate. The optimization algorithm is performed at each sample time to
compute the optimal control inputs. The performance of the proposed
system is investigated using a kinematic model of the mobile robot.
Abstract: Minimization methods for training feed-forward networks with Backpropagation are compared. Feedforward network training is a special case of functional minimization, where no explicit model of the data is assumed. Therefore due to the high dimensionality of the data, linearization of the training problem through use of orthogonal basis functions is not desirable. The focus is functional minimization on any basis. A number of methods based on local gradient and Hessian matrices are discussed. Modifications of many methods of first and second order training methods are considered. Using share rates data, experimentally it is proved that Conjugate gradient and Quasi Newton?s methods outperformed the Gradient Descent methods. In case of the Levenberg-Marquardt algorithm is of special interest in financial forecasting.
Abstract: Nonlinear system identification is becoming an important tool which can be used to improve control performance. This paper describes the application of adaptive neuro-fuzzy inference system (ANFIS) model for controlling a car. The vehicle must follow a predefined path by supervised learning. Backpropagation gradient descent method was performed to train the ANFIS system. The performance of the ANFIS model was evaluated in terms of training performance and classification accuracies and the results confirmed that the proposed ANFIS model has potential in controlling the non linear system.
Abstract: The aim of this paper is to rank the impact of Object
Oriented(OO) metrics in fault prediction modeling using Artificial
Neural Networks(ANNs). Past studies on empirical validation of
object oriented metrics as fault predictors using ANNs have focused
on the predictive quality of neural networks versus standard
statistical techniques. In this empirical study we turn our attention to
the capability of ANNs in ranking the impact of these explanatory
metrics on fault proneness. In ANNs data analysis approach, there is
no clear method of ranking the impact of individual metrics. Five
ANN based techniques are studied which rank object oriented
metrics in predicting fault proneness of classes. These techniques are
i) overall connection weights method ii) Garson-s method iii) The
partial derivatives methods iv) The Input Perturb method v) the
classical stepwise methods. We develop and evaluate different
prediction models based on the ranking of the metrics by the
individual techniques. The models based on overall connection
weights and partial derivatives methods have been found to be most
accurate.
Abstract: Security has been an important issue and concern in the
smart home systems. Smart home networks consist of a wide range of
wired or wireless devices, there is possibility that illegal access to
some restricted data or devices may happen. Password-based
authentication is widely used to identify authorize users, because this
method is cheap, easy and quite accurate. In this paper, a neural
network is trained to store the passwords instead of using verification
table. This method is useful in solving security problems that
happened in some authentication system. The conventional way to
train the network using Backpropagation (BPN) requires a long
training time. Hence, a faster training algorithm, Resilient
Backpropagation (RPROP) is embedded to the MLPs Neural
Network to accelerate the training process. For the Data Part, 200
sets of UserID and Passwords were created and encoded into binary
as the input. The simulation had been carried out to evaluate the
performance for different number of hidden neurons and combination
of transfer functions. Mean Square Error (MSE), training time and
number of epochs are used to determine the network performance.
From the results obtained, using Tansig and Purelin in hidden and
output layer and 250 hidden neurons gave the better performance. As
a result, a password-based user authentication system for smart home
by using neural network had been developed successfully.
Abstract: In a handwriting recognition problem, characters can
be represented using chain codes. The main problem in representing
characters using chain code is optimizing the length of the chain
code. This paper proposes to use randomized algorithm to minimize
the length of Freeman Chain Codes (FCC) generated from isolated
handwritten characters. Feedforward neural network is used in the
classification stage to recognize the image characters. Our test results
show that by applying the proposed model, we reached a relatively
high accuracy for the problem of isolated handwritten when tested on
NIST database.
Abstract: Although backpropagation ANNs generally predict
better than decision trees do for pattern classification problems, they
are often regarded as black boxes, i.e., their predictions cannot be
explained as those of decision trees. In many applications, it is
desirable to extract knowledge from trained ANNs for the users to
gain a better understanding of how the networks solve the problems.
A new rule extraction algorithm, called rule extraction from artificial
neural networks (REANN) is proposed and implemented to extract
symbolic rules from ANNs. A standard three-layer feedforward ANN
is the basis of the algorithm. A four-phase training algorithm is
proposed for backpropagation learning. Explicitness of the extracted
rules is supported by comparing them to the symbolic rules generated
by other methods. Extracted rules are comparable with other methods
in terms of number of rules, average number of conditions for a rule,
and predictive accuracy. Extensive experimental studies on several
benchmarks classification problems, such as breast cancer, iris,
diabetes, and season classification problems, demonstrate the
effectiveness of the proposed approach with good generalization
ability.
Abstract: In this paper we present an autoregressive model with
neural networks modeling and standard error backpropagation
algorithm training optimization in order to predict the gross domestic
product (GDP) growth rate of four countries. Specifically we propose
a kind of weighted regression, which can be used for econometric
purposes, where the initial inputs are multiplied by the neural
networks final optimum weights from input-hidden layer after the
training process. The forecasts are compared with those of the
ordinary autoregressive model and we conclude that the proposed
regression-s forecasting results outperform significant those of
autoregressive model in the out-of-sample period. The idea behind
this approach is to propose a parametric regression with weighted
variables in order to test for the statistical significance and the
magnitude of the estimated autoregressive coefficients and
simultaneously to estimate the forecasts.
Abstract: This paper aims to develop a NOx emission model of
an acid gas incinerator using Nelder-Mead least squares support
vector regression (LS-SVR). Malaysia DOE is actively imposing the
Clean Air Regulation to mandate the installation of analytical
instrumentation known as Continuous Emission Monitoring System
(CEMS) to report emission level online to DOE . As a hardware
based analyzer, CEMS is expensive, maintenance intensive and often
unreliable. Therefore, software predictive technique is often
preferred and considered as a feasible alternative to replace the
CEMS for regulatory compliance. The LS-SVR model is built based
on the emissions from an acid gas incinerator that operates in a LNG
Complex. Simulated Annealing (SA) is first used to determine the
initial hyperparameters which are then further optimized based on the
performance of the model using Nelder-Mead simplex algorithm.
The LS-SVR model is shown to outperform a benchmark model
based on backpropagation neural networks (BPNN) in both training
and testing data.
Abstract: In this paper, an improved technique for contingency
ranking using artificial neural network (ANN) is presented. The
proposed approach is based on multi-layer perceptrons trained by
backpropagation to contingency analysis. Severity indices in dynamic
stability assessment are presented. These indices are based on the
concept of coherency and three dot products of the system variables.
It is well known that some indices work better than others for a
particular power system. This paper along with test results using
several different systems, demonstrates that combination of indices
with ANN provides better ranking than a single index. The presented
results are obtained through the use of power system simulation
(PSS/E) and MATLAB 6.5 software.
Abstract: Many studies have focused on the nonlinear analysis
of electroencephalography (EEG) mainly for the characterization of
epileptic brain states. It is assumed that at least two states of the
epileptic brain are possible: the interictal state characterized by a
normal apparently random, steady-state EEG ongoing activity; and
the ictal state that is characterized by paroxysmal occurrence of
synchronous oscillations and is generally called in neurology, a
seizure.
The spatial and temporal dynamics of the epileptogenic process is
still not clear completely especially the most challenging aspects of
epileptology which is the anticipation of the seizure. Despite all the
efforts we still don-t know how and when and why the seizure
occurs. However actual studies bring strong evidence that the
interictal-ictal state transition is not an abrupt phenomena. Findings
also indicate that it is possible to detect a preseizure phase.
Our approach is to use the neural network tool to detect interictal
states and to predict from those states the upcoming seizure ( ictal
state). Analysis of the EEG signal based on neural networks is used
for the classification of EEG as either seizure or non-seizure. By
applying prediction methods it will be possible to predict the
upcoming seizure from non-seizure EEG.
We will study the patients admitted to the epilepsy monitoring
unit for the purpose of recording their seizures. Preictal, ictal, and
post ictal EEG recordings are available on such patients for analysis
The system will be induced by taking a body of samples then
validate it using another. Distinct from the two first ones a third body
of samples is taken to test the network for the achievement of
optimum prediction. Several methods will be tried 'Backpropagation
ANN' and 'RBF'.