Made available in DSpace on 2014-06-12T15:52:45Z (GMT). No. of bitstreams: 2 arquivo4692_1.pdf: 1118830 bytes, checksum: 96894dd8a22373c59d67d3b286b6c902 (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2003 / Um dos principais obstáculos para o uso em larga escala das Redes Neurais é a dificuldade de definir valores para seus parâmetros ajustáveis. Este trabalho discute como as Redes Neurais de Funções Base Radial (ou simplesmente Redes RBF) podem ter seus parâmetros ajustáveis definidos por algoritmos genéticos (AGs). Para atingir este objetivo, primeiramente é apresentado uma visão abrangente dos problemas envolvidos e as diferentes abordagens utilizadas para otimizar geneticamente as Redes RBF. É também proposto um algoritmo genético para Redes RBF com codificação genética não redundante baseada em métodos de clusterização. Em seguida, este trabalho aborda o problema de encontrar os parâmetros ajustáveis de um algoritmo de aprendizagem via AGs. Este problema é também conhecido como o problema de seleção de modelos. Algumas técnicas de seleção de modelos (e.g., validação cruzada e bootstrap) são usadas como funções objetivo do AG. O AG é modificado para adaptar-se a este problema por meio de heurísticas tais como narvalha de Occam e growing entre outras. Algumas modificações exploram características do AG, como por exemplo, a abilidade para resolver problemas de otimização multiobjetiva e manipular funções objetivo com ruído. Experimentos usando um problema benchmark são realizados e os resultados alcançados, usando o AG proposto, são comparados com aqueles alcançados por outras abordagens. As técnicas propostas são genéricas e podem também ser aplicadas a um largo conjunto de algoritmos de aprendizagem

In this thesis, we introduce a radial basis function network approach to solve a color image classification problem in a real time industrial application. Radial basis function networks are employed to classify the images of finished wooden parts in terms of their color and species. Other classification methods are also examined in this work. The minimum distance classifiers are presented since they have been employed by the previous research. We give brief definitions about color space, color texture, color quantization, color classification methods. We also give an intensive review of radial basis functions, regularization theory, regularized radial basis function networks, and generalized radial basis function networks. The centers of the radial basis functions are calculated by the k-means clustering algorithm. We examine the k-means algorithm in terms of starting criteria, the movement rule, and the updating rule. The dilations of the radial basis functions are calculated using a statistical method. Learning classifier systems are also employed to solve the same classification problem. Learning classifier systems learn the training samples completely whereas they are not successful to classify the test samples. Finally, we present some simulation results for both radial basis function network method and learning classifier systems method. A comparison is given between the results of each method. The results show that the best classification method examined in this work is the radial basis function network method. / Master of Science

The Support Vector (SV) machine is a novel type of learning machine, based on statistical learning theory, which contains polynomial classifiers, neural networks, and radial basis function (RBF) networks as special cases. In the RBF case, the SV algorithm automatically determines centers, weights and threshold such as to minimize an upper bound on the expected test error. The present study is devoted to an experimental comparison of these machines with a classical approach, where the centers are determined by $k$--means clustering and the weights are found using error backpropagation. We consider three machines, namely a classical RBF machine, an SV machine with Gaussian kernel, and a hybrid system with the centers determined by the SV method and the weights trained by error backpropagation. Our results show that on the US postal service database of handwritten digits, the SV machine achieves the highest test accuracy, followed by the hybrid approach. The SV approach is thus not only theoretically well--founded, but also superior in a practical application.

The hybridization of soft computing methods of Radial Basis Function (RBF) neural networks, Multi Layer Perceptron (MLP) neural networks with back-propagation learning, fuzzy sets and rough sets are studied in the scope of this thesis. Conventional MLP, conventional RBF, fuzzy MLP, fuzzy RBF, rough fuzzy MLP, and rough fuzzy RBF networks are compared. In the fuzzy neural networks implemented in this thesis, the input data and the desired outputs are given fuzzy membership values as the fuzzy properties &ldquo / low&rdquo / , &ldquo / medium&rdquo / and &ldquo / high&rdquo / . In the rough fuzzy MLP, initial weights and near optimal number of hidden nodes are estimated using rough dependency rules. A rough fuzzy RBF structure similar to the rough fuzzy MLP is proposed. The rough fuzzy RBF was inspected whether dependencies like the ones in rough fuzzy MLP can be concluded.

The thesis reports a contribution to the development of neural-like network- based element-free methods for the numerical simulation of some non-Newtonian fluid flow problems. The numerical approximation of functions and solution of the governing partial differential equations are mainly based on radial basis function networks. The resultant micro-macroscopic approaches do not require any element-based discretisation and only rely on a set of unstructured collocation points and hence are truly meshless or element-free. The development of the present methods begins with the use of the multi-layer perceptron networks (MLPNs) and radial basis function networks (RBFNs) to effectively eliminate the volume integrals in the integral formulation of fluid flow problems. An adaptive velocity gradient domain decomposition (AVGDD) scheme is incorporated into the computational algorithm. As a result, an improved feed forward neural network boundary-element-only method (FFNN- BEM) is created and verified. The present FFNN-BEM successfully simulates the flow of several Generalised Newtonian Fluids (GNFs), including the Carreau, Power-law and Cross models. To the best of the author's knowledge, the present FFNN-BEM is the first to achieve convergence for difficult flow situations when the power-law indices are very small (as small as 0.2). Although some elements are still used to discretise the governing equations, but only on the boundary of the analysis domain, the experience gained in the development of element-free approximation in the domain provides valuable skills for the progress towards an element-free approach. A least squares collocation RBFN-based mesh-free method is then developed for solving the governing PDEs. This method is coupled with the stochastic simulation technique (SST), forming the mesoscopic approach for analyzing viscoelastic flid flows. The velocity field is computed from the RBFN-based mesh-free method (macroscopic component) and the stress is determined by the SST (microscopic component). Thus the SST removes a limitation in traditional macroscopic approaches since closed form constitutive equations are not necessary in the SST. In this mesh-free method, each of the unknowns in the conservation equations is represented by a linear combination of weighted radial basis functions and hence the unknowns are converted from physical variables (e.g. velocity, stresses, etc) into network weights through the application of the general linear least squares principle and point collocation procedure. Depending on the type of RBFs used, a number of parameters will influence the performance of the method. These parameters include the centres in the case of thin plate spline RBFNs (TPS-RBFNs), and the centres and the widths in the case of multi-quadric RBFNs (MQ-RBFNs). A further improvement of the approach is achieved when the Eulerian SST is formulated via Brownian configuration fields (BCF) in place of the Lagrangian SST. The SST is made more efficient with the inclusion of the control variate variance reduction scheme, which allows for a reduction of the number of dumbbells used to model the fluid. A highly parallelised algorithm, at both macro and micro levels, incorporating a domain decomposition technique, is implemented to handle larger problems. The approach is verified and used to simulate the flow of several model dilute polymeric fluids (the Hookean, FENE and FENE-P models) in simple as well as non-trivial geometries, including shear flows (transient Couette, Poiseuille flows)), elongational flows (4:1 and 10:1 abrupt contraction flows) and lid-driven cavity flows.

The thesis compares several non-linear regression methods on synthetic data sets gen- erated using standard benchmarks for a continuous black-box optimization. For that com- parison, we have chosen the following regression methods: radial basis function networks, Gaussian processes, support vector regression and random forests. We have also included polynomial regression which we use to explain the basic principles of regression. The com- parison of these methods is discussed in the context of black-box optimization problems where the selected methods can be applied as surrogate models. The methods are evalu- ated based on their mean-squared error and on the Kendall's rank correlation coefficient between the ordering of function values according to the model and according to the function used to generate the data. 1

Neste trabalho, um novo enfoque para detecção e diagnóstico de falhas (DDF) em robôs manipuladores é apresentado. Um robô com falhas pode causar sérios danos e pode colocar em risco o pessoal presente no ambiente de trabalho. Geralmente, os pesquisadores têm proposto esquemas de DDF baseados no modelo matemático do sistema. Contudo, erros de modelagem podem ocultar os efeitos das falhas e podem ser uma fonte de alarmes falsos. Aqui, duas redes neurais artificiais são utilizadas em um sistema de DDF para robôs manipuladores. Um perceptron multicamadas treinado por retropropagação do erro é usado para reproduzir o comportamento dinâmico do manipulador. As saídas do perceptron são comparadas com as variáveis medidas, gerando o vetor de resíduos. Em seguida, uma rede com função de base radial é usada para classificar os resíduos, gerando a isolação das falhas. Quatro algoritmos diferentes são empregados para treinar esta rede. O primeiro utiliza regularização para reduzir a flexibilidade do modelo. O segundo emprega regularização também, mas ao invés de um único termo de penalidade, cada unidade radial tem um regularização individual. O terceiro algoritmo emprega seleção de subconjuntos para selecionar as unidades radiais a partir dos padrões de treinamento. O quarto emprega o mapa auto-organizável de Kohonen para fixar os centros das unidades radiais próximos aos centros dos aglomerados de padrões. Simulações usando um manipulador com dois graus de liberdade e um Puma 560 são apresentadas, demostrando que o sistema consegue detectar e diagnosticar corretamente falhas que ocorrem em conjuntos de padrões não-treinados. / In this work, a new approach for fault detection and diagnosis in robotic manipulators is presented. A faulty robot could cause serious damages and put in risk the people involved. Usually, researchers have proposed fault detection and diagnosis schemes based on the mathematical model of the system. However, modeling errors could obscure the fault effects and could be a false alarm source. In this work, two artificial neural networks are employed in a fault detection and diagnosis system to robotic manipulators. A multilayer perceptron trained with backpropagation algorithm is employed to reproduce the robotic manipulator dynamical behavior. The perceptron outputs are compared with the real measurements, generating the residual vector. A radial basis function network is utilized to classify the residual vector, generating the fault isolation. Four different algorithms have been employed to train this network. The first utilizes regularization to reduce the flexibility of the model. The second employs regularization too, but instead of only one penalty term, each radial unit has a individual penalty term. The third employs subset selection to choose the radial units from the training patterns. The forth algorithm employs the Kohonen’s self-organizing map to fix the radial unit center near to the cluster centers. Simulations employing a two link manipulator and a Puma 560 manipulator are presented, demonstrating that the system can detect and isolate correctly faults that occur in nontrained pattern sets.

Processor architectures are becoming increasingly complex and hence architects have to evaluate a large design space consisting of several parameters, each with a number of potential settings. In order to assist in guiding design decisions we develop simple and accurate models of the superscalar processor design space using a detailed and validated superscalar processor simulator. Firstly, we obtain precise estimates of all significant micro-architectural parameters and their interactions by building linear regression models using simulation based experiments. We obtain good approximate models at low simulation costs using an iterative process in which Akaike’s Information Criteria is used to extract a good linear model from a small set of simulations, and limited further simulation is guided by the model using D-optimal experimental designs. The iterative process is repeated until desired error bounds are achieved. We use this procedure for model construction and show that it provides a cost effective scheme to experiment with all relevant parameters. We also obtain accurate predictors of the processors performance response across the entire design-space, by constructing radial basis function networks from sampled simulation experiments. We construct these models, by simulating at limited design points selected by latin hypercube sampling, and then deriving the radial neural networks from the results. We show that these predictors provide accurate approximations to the simulator’s performance response, and hence provide a cheap alternative to simulation while searching for optimal processor design points.

Neste trabalho, um novo enfoque para detecção e diagnóstico de falhas (DDF) em robôs manipuladores é apresentado. Um robô com falhas pode causar sérios danos e pode colocar em risco o pessoal presente no ambiente de trabalho. Geralmente, os pesquisadores têm proposto esquemas de DDF baseados no modelo matemático do sistema. Contudo, erros de modelagem podem ocultar os efeitos das falhas e podem ser uma fonte de alarmes falsos. Aqui, duas redes neurais artificiais são utilizadas em um sistema de DDF para robôs manipuladores. Um perceptron multicamadas treinado por retropropagação do erro é usado para reproduzir o comportamento dinâmico do manipulador. As saídas do perceptron são comparadas com as variáveis medidas, gerando o vetor de resíduos. Em seguida, uma rede com função de base radial é usada para classificar os resíduos, gerando a isolação das falhas. Quatro algoritmos diferentes são empregados para treinar esta rede. O primeiro utiliza regularização para reduzir a flexibilidade do modelo. O segundo emprega regularização também, mas ao invés de um único termo de penalidade, cada unidade radial tem um regularização individual. O terceiro algoritmo emprega seleção de subconjuntos para selecionar as unidades radiais a partir dos padrões de treinamento. O quarto emprega o mapa auto-organizável de Kohonen para fixar os centros das unidades radiais próximos aos centros dos aglomerados de padrões. Simulações usando um manipulador com dois graus de liberdade e um Puma 560 são apresentadas, demostrando que o sistema consegue detectar e diagnosticar corretamente falhas que ocorrem em conjuntos de padrões não-treinados. / In this work, a new approach for fault detection and diagnosis in robotic manipulators is presented. A faulty robot could cause serious damages and put in risk the people involved. Usually, researchers have proposed fault detection and diagnosis schemes based on the mathematical model of the system. However, modeling errors could obscure the fault effects and could be a false alarm source. In this work, two artificial neural networks are employed in a fault detection and diagnosis system to robotic manipulators. A multilayer perceptron trained with backpropagation algorithm is employed to reproduce the robotic manipulator dynamical behavior. The perceptron outputs are compared with the real measurements, generating the residual vector. A radial basis function network is utilized to classify the residual vector, generating the fault isolation. Four different algorithms have been employed to train this network. The first utilizes regularization to reduce the flexibility of the model. The second employs regularization too, but instead of only one penalty term, each radial unit has a individual penalty term. The third employs subset selection to choose the radial units from the training patterns. The forth algorithm employs the Kohonen’s self-organizing map to fix the radial unit center near to the cluster centers. Simulations employing a two link manipulator and a Puma 560 manipulator are presented, demonstrating that the system can detect and isolate correctly faults that occur in nontrained pattern sets.

<p>Математички модели за естимацију тешко мерљивих величина називају<br />се софт сензорима. Процес формирања софт сензора није тривијалан и<br />квалитет естимације тешко мерљиве величине директно зависи од<br />начина формирања. Недостаци постојећих алгоритама за формирање<br />спречавају аутоматску конфигурацију софт сензора. У овом раду су<br />реализовани нови алгоритми који имају за сврху аутоматизацију<br />конфигурације софт сензора. Реализовани алгоритми решавају<br />проблеме проналаска оптималног сета улаза у софт сензор и кашњења<br />сваког од њих као и одабира структуре и начина обуке софт сензора<br />заснованих на вештачким неуронским мрежама са радијално базираним<br />функцијама.</p> / <p>Matematički modeli za estimaciju teško merljivih veličina nazivaju<br />se soft senzorima. Proces formiranja soft senzora nije trivijalan i<br />kvalitet estimacije teško merljive veličine direktno zavisi od<br />načina formiranja. Nedostaci postojećih algoritama za formiranje<br />sprečavaju automatsku konfiguraciju soft senzora. U ovom radu su<br />realizovani novi algoritmi koji imaju za svrhu automatizaciju<br />konfiguracije soft senzora. Realizovani algoritmi rešavaju<br />probleme pronalaska optimalnog seta ulaza u soft senzor i kašnjenja<br />svakog od njih kao i odabira strukture i načina obuke soft senzora<br />zasnovanih na veštačkim neuronskim mrežama sa radijalno baziranim<br />funkcijama.</p> / <p>Mathematical models that are used for estimation of variables that can not be<br />measured in real time are called soft sensors. Creation of soft sensor is a<br />complex process and quality of estimation depends on the way soft sensor is<br />created. Restricted applicability of existing algorithms is preventing automatic<br />configuration of soft sensors. This paper presents new algorithms that are<br />providing automatic configuration of soft sensors. Presented algorithms are<br />capable of determing optimal subset of soft sensor inputs and their time<br />delays, as well as optimal architecture and automatic training of the soft<br />sensors that are based on artificial radial basis function networks.</p>