ATM (Asynchronous Transfer Mode) networks were deemed the best choice for multimedia communication. The traditional mode was replaced because ATM can provide varied traffic types and QoS (quality of service). Maintaining QoS, however, requires a flexible traffic control, including call admission control and congestion control. Traditional approaches fail to estimate the required bandwidth and cell loss rate precisely. To alleviate these problems, we employ AI methods to improve the capability of estimated bandwidth and predicted cell loss rate. This thesis aims to apply neural network techniques to ATM traffic control and consists of two parts. The first part concerns a neural-based call admission control, while the second part presents an intelligent congestion control for ATM networks. In the first part, we focus on the improvement of RBF (Radial basis function) networks and the design of a neural-based call admission control. RBF networks have been widely used for modeling a function from given input-output patterns. However, two difficulties are encountered with traditional RBF networks. One is that the initial configuration of a RBF network needs to be determined by a trial-and-error method. The other is that the performance suffers from some difficulties when the desired output has abrupt changes or constant values in certain intervals. We propose a novel approach to overcome these difficulties. New kernel functions are used for hidden nodes, and the number of nodes is determined automatically by an ART-like algorithm. Parameters and weights are initialized appropriately, and then tuned and adjusted by the gradient descent method to improve the performance of the network. Then, we employ ART-RBF networks to design and implement a call admission control. Traditional approaches fail to estimate appropriately the required bandwidth, leading to a waste of bandwidth or a high cell loss rate. To alleviate the problem, we employ ART-RBF networks to estimate the required bandwidth, and thus a new connection request can then be accepted or rejected. Because of the more accurate estimation on the required bandwidth, the proposed method can provide a better control on quality of service for ATM networks. In the second part, we propose a neural-fuzzy rate-based feedback congestion control for ATM networks. Traditional methods perform congestion control by monitoring the queue length. The source rate is decreased by a fixed rate when the queue length is greater than a predefined threshold. However, it is difficult to get a suitable rate according to the degree of traffic congestion. We employ a neural-fuzzy mechanism to control the source rate. Through learning, cell loss can be predicted from the current value and the derivative of the queue length. Then an explicit rate is calculated and the source rate is controlled appropriately. In summary, we have proposed improvements on architecture and performance of neural networks, and applied neural networks to traffic control for ATM networks. We have developed some control mechanisms which, through simulations, have been shown to be more effective than traditional methods.

Este projeto consiste em um sistema de navegação autônomo baseado em redes neurais nebulosas modulares capacitando o robô a alcançar alvos, ou pontos metas, em ambientes desconhecidos. Inicialmente, o sistema não tem habilidade para a navegação, após uma fase de experimentos com algumas colisões, o mecanismo de navegação aprimora-se guiando o robô ao alvo de forma eficiente. Uma arquitetura híbrida inteligente é apresentada para este sistema de navegação, baseada em redes neurais artificiais e lógica nebulosa. A arquitetura é hierárquica e costitiui-se de dois módulos responsáveis por gerar comportamentos inatos de desvio de obstáculos e de busca ao alvo. Um mecanismo de aprendizagem por reforço, baseada em uma extensão da lei de Hebb, pondera os comportamentos inatos conflitantes ajustando os pesos sinápticos das redes neurais nos instantes de captura do alvo e de colisão contra obstáculos. A abordagem consolidada em simulação é validada em ambientes reais neste trabalho. Para tanto, este sistema foi implementado e testado no simulador Saphira, ambiente de simulação que acompanha o robô Pioneer I e que denota um estágio anterior aos testes em ambientes reais por apresentar comportamentos do robô similares aos comportamentos do robô móvel. Modificações na arquitetura híbrida foram necessárias para adaptar o sistema de navegação simulado ao sistema incorporado no Pioneer I. Experimentos em ambientes reais demonstraram a eficiência e a capacidade de aprendizagem do sistema de navegação, validando a arquitetura híbrida inteligente para aplicação em robôs móveis / This project consists in a autonomous navigation system based on modular neuro-fuzzy networks that is able to guide the robot in unknown environments from a initial point to the goal. Initially, the system is not able to navigate, but after a trial and error period and some collisions, it improves in guiding the robot to the goal efficiently. A intelligent hybrid architecture is presented for this naviga tion system based on artificial neural networks and fuzzy logic. This architecture is hierarquical and consists in two modules that generate innate behaviors, like obstacles avoiding and target reaching. A reinforcement learning mecanism, based on the extended Hebb law, balances this conflicting innate behaviors adjusting the neural network synaptic weights as obstacle and collision avoidance and target reaching takes place. In this project, the approach is consolidated in simulation and validated in real environments. To this end, this system has been implemented by using Saphira simulator and Pioneer I simulation environment. This simulated evironment is a previous stage of tests performed real time and presents simulated robot behaviors similar to real mobile robot behaviors. The hybrid architecture was modified to adapt the simulated navigation system into Pioneer I software. Experiments in a real environments show the efficiency and learning capabilities of the navigation system, validating the intelligent hybrid architecture for mobile robots applications

Este projeto consiste em um sistema de navegação autônomo baseado em redes neurais nebulosas modulares capacitando o robô a alcançar alvos, ou pontos metas, em ambientes desconhecidos. Inicialmente, o sistema não tem habilidade para a navegação, após uma fase de experimentos com algumas colisões, o mecanismo de navegação aprimora-se guiando o robô ao alvo de forma eficiente. Uma arquitetura híbrida inteligente é apresentada para este sistema de navegação, baseada em redes neurais artificiais e lógica nebulosa. A arquitetura é hierárquica e costitiui-se de dois módulos responsáveis por gerar comportamentos inatos de desvio de obstáculos e de busca ao alvo. Um mecanismo de aprendizagem por reforço, baseada em uma extensão da lei de Hebb, pondera os comportamentos inatos conflitantes ajustando os pesos sinápticos das redes neurais nos instantes de captura do alvo e de colisão contra obstáculos. A abordagem consolidada em simulação é validada em ambientes reais neste trabalho. Para tanto, este sistema foi implementado e testado no simulador Saphira, ambiente de simulação que acompanha o robô Pioneer I e que denota um estágio anterior aos testes em ambientes reais por apresentar comportamentos do robô similares aos comportamentos do robô móvel. Modificações na arquitetura híbrida foram necessárias para adaptar o sistema de navegação simulado ao sistema incorporado no Pioneer I. Experimentos em ambientes reais demonstraram a eficiência e a capacidade de aprendizagem do sistema de navegação, validando a arquitetura híbrida inteligente para aplicação em robôs móveis / This project consists in a autonomous navigation system based on modular neuro-fuzzy networks that is able to guide the robot in unknown environments from a initial point to the goal. Initially, the system is not able to navigate, but after a trial and error period and some collisions, it improves in guiding the robot to the goal efficiently. A intelligent hybrid architecture is presented for this naviga tion system based on artificial neural networks and fuzzy logic. This architecture is hierarquical and consists in two modules that generate innate behaviors, like obstacles avoiding and target reaching. A reinforcement learning mecanism, based on the extended Hebb law, balances this conflicting innate behaviors adjusting the neural network synaptic weights as obstacle and collision avoidance and target reaching takes place. In this project, the approach is consolidated in simulation and validated in real environments. To this end, this system has been implemented by using Saphira simulator and Pioneer I simulation environment. This simulated evironment is a previous stage of tests performed real time and presents simulated robot behaviors similar to real mobile robot behaviors. The hybrid architecture was modified to adapt the simulated navigation system into Pioneer I software. Experiments in a real environments show the efficiency and learning capabilities of the navigation system, validating the intelligent hybrid architecture for mobile robots applications

Recent advances in the miniaturization, functionality, and integration of integrated circuits and packages, such as the system-on-package (SOP) methodology, require increasing use of microvias that generates vertical signal paths in a high-density multilayer substrate. A scanning projection excimer laser system has been utilized to fabricate the microvias. In this thesis, a novel technique implementing statistical experimental design and neural networks (NNs) is used to characterize and model the excimer laser ablation process for microvia formation. Vias with diameters from 10 50 micrometer have been ablated in DuPont Kapton(r) E polyimide using an Anvik HexScan(tm) 2150 SXE pulsed excimer laser operating at 308 nm. Accurate NN models, developed from experimental data, are obtained for microvia responses, including ablated thickness, via diameter, wall angle, and resistance. Subsequent to modeling, NNs and genetic algorithms (GAs) are utilized to generate optimal process recipes for the laser tool. Such recipes can be used to produce desired microvia responses, including open vias, specific diameter, steep wall angle, and low resistance. With continuing advancement in the use of excimer laser systems in microsystems packaging has come an increasing need to offset capital equipment investment and lower equipment downtime. In this thesis, an automated in-line failure diagnosis system using NNs and Dempster-Shafer (D-S) theory is implemented. For the sake of comparison, an adaptive neuro-fuzzy approach is applied to achieve the same objective. Both the D-S theory and neuro-fuzzy logic are used to develop an automated inference system to specifically identify failures. Successful results in failure detection and diagnosis are obtained from the two approaches. The result of this investigation will benefit both engineering and management. Engineers will benefit from high yield, reliable production, and low equipment down-time. Business people, on the other hand, will benefit from cost-savings resulting from more production-worthy (i.e., lower maintenance) laser ablation equipment.

Orientador: Takaaki Ohishi / Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Eletrica e de Computação / Made available in DSpace on 2018-08-08T10:06:35Z (GMT). No. of bitstreams: 1 LeoneFilho_MarcosdeAlmeida_M.pdf: 1557959 bytes, checksum: 92dc63d4e3140cc61ba7900961c0e9fb (MD5) Previous issue date: 2006 / Resumo: Neste trabalho é proposta uma metodologia para previsão de séries temporais de carga de energia elétrica de curto prazo. Esta metodologia vem sendo muito utilizada no contexto da previsão de séries temporais e do reconhecimento de padrões. Os autores que propuseram esta metodologia a chamaram de "Ensembles". Este nome tenta explicar o é este modelo: uma combinação de partes que juntas formam um só modelo. Neste sentido, este nome expressa com relativa clareza qual é o principal aspecto desta metodologia, que no caso específico deste trabalho, é o de fazer várias previsões de uma mesma série temporal utilizando diferentes ferramentas que sozinhas são suficientemente competentes para prever a série temporal em questão, e em seguida combinar as soluções para, deste modo, tentar obter uma solução melhor do que quando é usada somente uma ferramenta. As ferramentas usadas para compor a previsão dos "Ensembles" finais são Redes Neurais Artificiais (RNAs) e Redes Neurais Nebulosas. Atualmente, estas redes são largamente utilizadas em problemas de previsão de séries temporais, principalmente quando o fator gerador destas séries é um sistema não-linear. Desta forma, isto as tornou candidatas potenciais para prever valores de uma série de cargas de energia elétrica, pois este tipo de série tem características essencialmente não-lineares. Sendo assim, foram utilizados quatro tipos de redes: RNAs MLPs, RNAs Recorrentes, RNAs de Base Radial e Redes Neurais Nebulosas tipo ANFIS. Com os modelos básicos de redes foram, utilizados Algoritmos Genéticos para evoluir os parâmetros destas redes e, assim, chegar a uma população de redes suficientemente competentes para fazer as previsões da série de cargas. Na próxima etapa, com os resultados das previsões da população de redes evoluídas foi feita a seleção dos melhores agrupamentos destas redes evoluídas e, como este processo requer a avaliação de diferentes configurações de modelos, esta seleção é baseada em Algoritmos Genéticos.Os resultados obtidos ao se utilizar "ensembles" mostraram que este modelo foi capaz de alcançar uma grande robustez na previsão, reduzindo os erros de previsão, suavizando os resultados de previsão e deixando o modelo menos suscetível a grandes erros quando surgem "outliers" no conjunto de dados / Abstract: This work proposes a methodology for short-term electric power load forecasting. This methodology is being widely used under the context of time series prediction and pattern recognition. It was named "ensembles" by the authors who developed it. This name carries the meaning of an assemblage of parts considered as forming a whole. Therefore, this name expresses rather clearly the main characteristic of this methodology, which under the framework of this study is to make several predictions of the same time series using various different tools in which every single one alone is sufficiently competent to predict the above mentioned time series. After that, the predictions are combined in order to achieve a better prediction compared to the one that is obtained if a single predictor is used. The tools implemented to form the final "ensembles" prediction are Artificial Neural Networks (ANNs) and Neuro-fuzzy Networks. Nowadays, these networks are being widely used in time series predictions problems, mainly when the factor that generates these series is a non-linear system. Hence, this fact has elected them as potential candidates to predict future values of an electric power load series because this series has essentially non-linear characteristics. As a result, four types of networks were utilized in this work: MLPs ANNs, Recurrent ANNs, Radial Basis ANNs and ANFIS type Neuro-fuzzy networks. So, with the basic networks models, Genetic Algorithms were applied to evolve the parameters of these networks and, as a consequence, a population of networks sufficiently capable of predicting future values of the load time series was built. On the next step, with the results obtained from the evolved population of networks, a selection of the most suitable results of the individual networks were made and, as soon as this process implies the evaluation of multiple different combinations of models, this methodology was based on Genetic Algorithms. Then, this selected networks were combined. The results when using "ensembles" revealed that this model was able to reach a great robustness in prediction tasks. In that sense, it was possible to reduce the level of prediction error, to smooth the resulting predictions and to make the model more stable reducing the possibilities of presenting high levels of errors when the used data set contains "outliers" / Mestrado / Energia Eletrica / Mestre em Engenharia Elétrica