Made available in DSpace on 2016-04-29T14:23:48Z (GMT). No. of bitstreams: 1 Paulo Kuester Neto.pdf: 1284476 bytes, checksum: 99c3eadc17da7f6d51803dba0833899f (MD5) Previous issue date: 2009-04-24 / This project is part of the research line Collective Intelligence and Interactive Environments and aims to investigate modes of pattern recognition and classification in three-dimensional images using artificial neural networks. To achieve this, three-dimensional images will be submitted to a connection is system based on Artificial Neural Networks according to a back propagation algorithm used as the basis for training, in order to obtain patterns that are common among these images. This work aims to contribute to image analysis so that it can be applied to research, from forest mapping and virtual worlds construction to prognostics and/or diagnoses in health-related areas, in which, due to variances and imperfections in images that are said to be similar, it is not possible to use simple algorithms that recognize similarities between them. In light of the theoretical presuppositions discussed in chapter 2 and to the state-of-the-art approached in chapter 3, the characteristics, organization modes, learning algorithms and free parameters of this neural model that best adapt to the nature of the research are defined. The work must involve a simulation environment, the framework for neural models experimentation and results verification, chosen according to characteristics like reliability, viability and adequacy to hardware conditions and limitations. In addition, the environment must be capable of dealing with the research object, that is, the analysis and classification of three-dimensional forms and their recognition through adjustments to the parameters of the neural model. The research to be carried out was divided into two phases: the first one is network training, in which some images are arbitrarily chosen from an image base. These images share common characteristics that must be recognized to make adjustments to the Neural Network. In the second phase, after the stage of tests and training, the network must be capable of dealing with the rest of the selected image base. The system must also effectively deal with exceptions and variation in some characteristics, such as light, positioning and color. The challenge is making the neural network training be as generic as possible, so it can deal with these variations, offering a degree of reliability without substantial decrease in effectiveness / Este projeto se insere na linha de pesquisa Inteligência Coletiva e Ambientes Interativos, visando investigar os modos de reconhecimento e classificação de padrões em imagens tridimensionais utilizando Redes Neurais Artificiais. Para tanto, pretende-se submeter imagens tridimensionais a um sistema conexionista baseado em Redes Neurais Artificiais de acordo com um algoritmo de retro-propagação (backpropagation) como base para treinamento, buscando-se obter padrões comuns entre essas imagens. Este trabalho objetiva contribuir com a análise de imagens para aplicação em pesquisa, desde mapeamento florestal, construção de mundos virtuais até prognósticos e/ou diagnóstico em áreas relacionadas à saúde, em que, devido a variâncias e imperfeições em imagens ditas similares, não se aplicam a utilização de algoritmos simples que reconheçam semelhanças entre elas. De acordo com os pressupostos teóricos discutidos no capítulo 2 e o estado da arte no capítulo 3, definem-se características, modos de organização, algoritmos de aprendizagem e parâmetros livres desse modelo neural que melhor se adaptam a natureza da pesquisa. O trabalho deve envolver um ambiente de simulação, framework para experimentação dos modelos neurais e verificação de resultados, escolhido de acordo com características como confiabilidade, viabilidade e adequação as condições e limitações de hardware. O ambiente deve ser capaz de lidar ainda com o objeto de pesquisa, ou seja, a análise e a classificação de formas tridimensionais e seu reconhecimento através de ajustes nos parâmetros do modelo neural. A pesquisa a ser realizada foi dividida em duas fases, a primeira, de treinamento da rede, escolhendo arbitrariamente, a partir de um banco de imagens, algumas que compartilhem características comuns que devem ser reconhecidas para ajustes da Rede Neural. Na segunda fase, posterior a etapa de testes e treinamento, a rede deve ser capaz de lidar com o restante do banco de imagens selecionado. O sistema deve ainda ser efetivo ao lidar com exceções e variação em algumas características como luminosidade, posicionamento e cor. O desafio é tornar o treinamento da Rede Neural o mais genérico possível a fim de lidar com essas variações, oferecendo um grau de confiabilidade sem degradação substancial de efetividade

A navegação autônoma é um dos problemas fundamentais da robótica móvel. Para um robô executar esta tarefa, é necessário determinar a região segura para a navegação. Este trabalho propõe um sistema de identificação de superfícies navegáveis baseado em visão computacional utilizando redes neurais artificiais. Mais especificamente, é realizado um estudo sobre a utilização de diferentes atributos de imagem, como descritores estatísticos e elementos de espaços de cores, para serem utilizados como entrada das redes neurais artificiais que tem como tarefa a identificação de superfícies navegáveis. O sistema desenvolvido utiliza resultados de classificação de múltiplas configurações de redes neurais artificiais, onde a principal diferença entre elas é o conjunto de atributos de imagem utilizados como entrada. Essa combinação de diversas classificações foi realizada visando maior robustez e melhor desempenho na identificação de vias em diferentes cenários / Autonomous navigation is a fundamental problem in mobile robotics. In order to perform this task, a robot must identify the areas where it can navigate safely. This dissertation proposes a navigable terrain identification system based on computer vision and neural networks. More specifically, it is presented a study of image attributes, such as statistical decriptors and elements of different color spaces, that are used as neural neworks inputs for the navigable surfaces identification. The system developed combines the classification results of multiple neural networks topologies with different image attributes. This combination of classification results allows for improved efficient and robustenes in different scenarios

Um dos problemas freqüentes em modelos de avaliação de imóveis é identificar quais de suas características devem ser levadas em consideração e o quanto cada uma destas influencia no valor final das propriedades. Além disso, é necessário um critério de avaliação bem estruturado, baseado em modelagem matemática adequada. A partir das constatações acima, foram estabelecidos os objetivos deste trabalho: após identificar as principais variáveis que interferem no valor das propriedades, avaliar o uso de Redes Neurais Artificiais para fins de avaliação e estudar a influência de uma medida de acessibilidade no valor de terrenos urbanos. Quanto às variáveis a serem empregadas nos modelos de avaliação, chegou-se a conclusão que um banco de dados misto, onde tanto variáveis de natureza espacial quanto física pudessem ser incluídas, parecia ser uma opção interessante. Desta forma, após a inclusão no banco de dados de uma variável de natureza espacial, a distância ao centro da cidade, este trabalho comparou dois métodos de avaliação: as Redes Neurais Artificiais e o modelo de regressão múltipla, este último muito usado na prática. Foram abordados dois estudos de caso, as cidades de Araçariguama e São Carlos. A primeira é uma cidade dormitório, de pequeno porte (cerca de 6000 habitantes), localizada nas proximidades da capital do estado, São Paulo. A segunda, por sua vez, é uma cidade de porte médio (cerca de 160000 habitantes), localizada no centro do estado e pólo industrial e tecnológico. A escolha destas cidades ofereceu a oportunidade de estudar a influência de uma variável como a acessibilidade em contextos diferentes. Os resultados obtidos para Araçariguama indicaram que a medida de acessibilidade empregada, distância ao centro, era uma das variáveis mais importantes na formação do preço de propriedades. Quanto aos resultados obtidos pelos modelos empregados, pôde-se observar que, ao utilizar regressão múltipla, o efeito da variável distância ao centro não pode ser estudado sozinho, pois esta variável interage com a área. A utilização de Redes Neurais, por sua vez, também forneceu estimativas adequadas de valor, verificando-se através dela que a acessibilidade apresentou um peso superior a 34% no valor final. Ao se comparar os dois métodos pode-se observar que as Redes Neurais (RN) demonstraram um desempenho superior, quando este foi avaliado pelo valor do erro relativo total. Com o objetivo de analisar a distribuição espacial dos erros, estes foram agrupados em cinco clusters, podendo-se constatar que os maiores erros fornecidos pelas RN se concentraram em um único bairro. Na análise para São Carlos pôde-se verificar também, através dos resultados obtidos por ambos os métodos, que a distância ao centro foi um dos fatores preponderantes na avaliação dos imóveis. A análise da distribuição espacial dos erros apontou uma concentração de erros maiores em um dos bairros, o que pôde ser observado para os dois métodos empregados e dois dos três conjuntos de dados. Um fato que chamou a atenção foi o de que para Araçariguama, que é uma cidade de porte menor, a importância relativa da variável acessibilidade foi maior que para a outra cidade. / A common problem in the use of land valuation models is the identification of the real estate features that should be incorporated in the models and how they influence the final property price. In addition, a well structured approach based in consistent mathematical models is also required. Based on the aforementioned assertions, the following objectives have been drawn for this work: after identifying the main variables that have a strong influence on land values, the use of Artificial Neural Networks (ANN) for land valuation have be tested and the influence of an accessibility measure on urban land values have been studied. Regarding the variables that should been part of the valuation models, we reached the conclusion that a mixed database containing physical and spatial attributes seemed to be an interesting option for this sort of problem. Therefore, after the addition of a spatial variable, the distance to the city center, to our database, two valuation methods have been compared: the ANN approach and a multiple regression model, the latter quite common in practice. Two case studies have been then analyzed: the cities of Araçariguama and São Carlos. The first one is a small bedroom town (around 6,000 inhabitants) not far from the state capital, the city of São Paulo. The second one is a medium-sized city (around 160,000 inhabitants) located in the middle of the state and a technological and industrial center. The particularities of these two cities made possible a comparison of the influence that such a variable as accessibility could have on the land values under two different conditions. The results obtained for the city of Araçariguama indicated that the accessibility measure used, the distance from the city center, was one of the main variables influencing land prices. Although both models gave good estimates, their results were not exactly the same. While the influence of the variable distance to the city center could not be individually taken in the multiple regression model, because of its interaction with the variable area, the same variable has a strong weight on the ANN model, in which it appears as responsible for over 34% of the land value. The ANN performed better in a direct comparison of the two approaches, specially when looking to the total relative error. With the purpose of analyzing the spatial distribution of the estimation errors, they have been grouped into clusters, which have stressed that the worst cases are concentrated in a specific area of the city. Both methods showed that the distance to the city center has a strong influence on land values also in the city of São Carlos. The highest estimation errors were also concentrated in a specific neighborhood for two out of three data sets in both valuation methods. Another interesting outcome is the fact that the relative weight of the accessibility variable used was higher in Araçariguama than in São Carlos, although the former city is smaller than the latter.

Artificial neural networks (ANNs) have been used successfully in various practical problems. Though extensive improvements on different types of ANNs have been made to improve their performance, each ANN design still experiences its own limitations. The existing digital human models are mature enough to provide accurate and useful results for different tasks and scenarios under various conditions. There is, however, a critical need for these models to run in real time, especially those with large-scale problems like motion prediction which can be computationally demanding. For even small changes to the task conditions, the motion simulation needs to run for a relatively long time (minutes to tens of minutes). Thus, there can be a limited number of training cases due to the computational time and cost associated with collecting training data. In addition, the motion problem is relatively large with respect to the number of outputs, where there are hundreds of outputs (between 500-700 outputs) to predict for a single problem. Therefore, the aforementioned necessities in motion problems lead to the use of tools like the ANN in this work. This work introduces new algorithms for the design of the radial-basis network (RBN) for problems with minimal available training data. The new RBN design incorporates new training stages with approaches to facilitate proper setting of necessary network parameters. The use of training algorithms with minimal heuristics allows the new RBN design to produce results with quality that none of the competing methods have achieved. The new RBN design, called Opt_RBN, is tested on experimental and practical problems, and the results outperform those produced from standard regression and ANN models. In general, the Opt_RBN shows stable and robust performance for a given set of training cases. When the Opt_RBN is applied on the large-scale motion prediction application, the network experiences a CPU memory issue when performing the optimization step in the training process. Therefore, new algorithms are introduced to modify some steps of the new Opt_RBN training process to address the memory issue. The modified steps should only be used for large-scale applications similar to the motion problem. The new RBN design proposes an ANN that is capable of improved learning without needing more training data. Although the new design is driven by its use with motion prediction problems, the consequent ANN design can be used with a broad range of large-scale problems in various engineering and industrial fields that experience delay issues when running computational tools that require a massive number of procedures and a great deal of CPU memory. The results of evaluating the modified Opt_RBN design on two motion problems are promising, with relatively small errors obtained when predicting approximately 500-700 outputs. In addition, new methods for constraint implementation within the new RBN design are introduced. Moreover, the new RBN design and its associated parameters are used as a tool for simulated task analysis. This work initiates the idea that output weights (W) can be used to determine the most critical basis functions that cause the greatest reduction in the network test error. Then, the critical basis functions can specify the most significant training cases that are responsible for the proper performance achieved by the network. The inputs with the most change in value can be extracted from the basis function centers (U) in order to determine the dominant inputs. The outputs with the most change in value and their corresponding key body degrees-of-freedom for a motion task can also be specified using the training cases that are used to create the network's basis functions.

The response of a karst aquifer to storm events is often faster and more severe than that of a non-karst aquifer. This distinction is often problematic for planners and municipalities, because karst flooding does not typically occur along perennial water courses; thus, traditional flood management strategies are usually ineffective. The City of Bowling Green (CoBG), Kentucky is a representative example of an area plagued by karst flooding. The CoBG, is an urban karst area (UKA), that uses Class V Injection Wells to lessen the severity of flooding. The overall effectiveness, siting, and flooding impact of Injection Wells in UKA’s is lacking; their influence on groundwater is evident from decades of recurring problems in the form of flooding and groundwater contamination. This research examined Class V Injection Wells in the CoBG to determine how Injection Well siting, design, and performance influence urban karst hydrology. The study used high-resolution monitoring, as well as hydrologic modeling, to evaluate Injection Well and spring responses during storm and baseflow conditions. In evaluating the properties of the karst aquifer and the influences from the surrounding environment, a relationship was established between precipitation events, the drainage capacity of the Injection Wells, and the underlying karst system. Ultimately, the results from this research could be used to make sound data-driven policy recommendations and to inform stormwater management in UKAs.

Excessive erosion and fine sediment delivery to river corridors and receiving waters degrade aquatic habitat, add to nutrient loading, and impact infrastructure. Understanding the sources and movement of sediment within watersheds is critical for assessing ecosystem health and developing management plans to protect natural and human systems. As our changing climate continues to cause shifts in hydrological regimes (e.g., increased precipitation and streamflow in the northeast U.S.), the development of tools to better understand sediment dynamics takes on even greater importance. In this research, advanced geomatics and machine learning are applied to improve the (1) monitoring of streambank erosion, (2) understanding of event sediment dynamics, and (3) prediction of sediment loading using meteorological data as inputs. Streambank movement is an integral part of geomorphic changes along river corridors and also a significant source of fine sediment to receiving waters. Advances in unmanned aircraft systems (UAS) and photogrammetry provide opportunities for rapid and economical quantification of streambank erosion and deposition at variable scales. We assess the performance of UAS-based photogrammetry to capture streambank topography and quantify bank movement. UAS data were compared to terrestrial laser scanner (TLS) and GPS surveying from Vermont streambank sites that featured a variety of bank conditions and vegetation. Cross-sectional analysis of UAS and TLS data revealed that the UAS reliably captured the bank surface and was able to quantify the net change in bank area where movement occurred. Although it was necessary to consider overhanging bank profiles and vegetation, UAS-based photogrammetry showed significant promise for capturing bank topography and movement at fine resolutions in a flexible and efficient manner. This study also used a new machine-learning tool to improve the analysis of sediment dynamics using three years of high-resolution suspended sediment data collected in the Mad River watershed. A restricted Boltzmann machine (RBM), a type of artificial neural network (ANN), was used to classify individual storm events based on the visual hysteresis patterns present in the suspended sediment-discharge data. The work expanded the classification scheme typically used for hysteresis analysis. The results provided insights into the connectivity and sources of sediment within the Mad River watershed and its tributaries. A recurrent counterpropagation network (rCPN) was also developed to predict suspended sediment discharge at ungauged locations using only local meteorological data as inputs. The rCPN captured the nonlinear relationships between meteorological data and suspended sediment discharge, and outperformed the traditional sediment rating curve approach. The combination of machine-learning tools for analyzing storm-event dynamics and estimating loading at ungauged locations in a river network provides a robust method for estimating sediment production from catchments that informs watershed management.

[pt] O sistema hidroelétrico é responsável por 83,7% da energia elétrica gerada no país. Assim sendo, a geração de energia elétrica no Brasil depende basicamente das vazões naturais que afluem aos aproveitamentos hidroelétricos distribuídos por doze bacias hidrográficas no país. Sendo o Operador Nacional do Sistema Elétrico (ONS) o órgão responsável por elaborar a previsão e a geração de cenários de vazões naturais médias diárias, semanais e mensais para todos os locais de aproveitamentos hidroelétricos do Sistema Interligado Nacional (SIN), a qualidade da previsão da vazão natural é de suma importância para este órgão. A qualidade dessa previsão impacta diretamente no planejamento e em programas de operação do SIN, tal como o Programa Mensal de Operação - PMO. Mesmo com a melhoria na qualidade da previsão de vazões por meio da criação e adoção dos mais diversos modelos determinísticos e estocásticos nos últimos anos, os erros de previsão são, ainda, significativos. Deste modo, o objetivo principal desta dissertação foi propor um novo modelo capaz de proporcionar um significativo ganho de qualidade na previsão de vazões nas regiões dos aproveitamentos hidrelétricos das bacias hidrográficas do país. O modelo proposto, baseado em redes neurais, tem como ferramenta primordial a utilização de transformadas wavelets, que filtram os dados históricos de vazões, ou seja, as entradas das redes neurais de previsão, dividindo esses dados de entrada (sinais) em diversas escalas, no intuito de que as redes neurais possam melhor analisá-los. Para verificar a eficácia do modelo proposto, aqui denominado MIP (Modelo Inteligente de Previsão), procedeu-se um estudo de caso que realiza a previsão de vazões naturais incrementais médias diárias e semanais no trecho incremental entre as Usinas Hidroelétricas (UHE) Porto Primavera, Rosana e Itaipu da Bacia do Rio Paraná, chegando-se a um erro de aproximadamente 3,5% para previsão de vazões um dia à frente, 16% para 12 dias à frente, e 9% para previsão média semanal. Esta dissertação objetiva, também, investigar a eficácia do uso de informações das precipitações observadas e previstas na previsão de vazão, em conjunção com o uso do histórico de vazões. / [en] The hydroelectricity system is responsible for 83.7% of the electric energy generated at Brazil. Therefore, the generation of electric power in Brazil depends basically on the natural flow rates distributed by twelve basins in the country. The quality of prediction of natural flow is of crucial importance for the Brazilian governmental agency, ONS (from the portuguese language Electrical National Operator System), responsible for preparing the forecast and the generation of scenarios of daily, weekly and monthly average natural streamflows of all places of hydroelectric exploitations of SIN (from the portuguese language National Linked System). The quality of that forecast impacts directly in the planning and operation programs of SIN, for example, the PMO (from the portuguese language Monthly Operation Program). Even with the improvement in the quality of river flow forecasts through the creation and adoption of the various deterministic and stochastic models in recent years, the errors of forecasting are still significant. Thus, the main goal of this dissertation was proposing a new model capable of providing a significant improvement in Streamflow forecasts in regions of exploitations of hydroelectric basins of the country. The proposed model, based on neural networks, has the primary tool the use of wavelet transforms, to filter streamflows historical data, or the entries of predict neural networks, dividing the input data (signals) in several scales, in order that the neural networks can better analyse them. In order to check the effectiveness of the proposed model, here called MIP (from the portuguese language Forecast Intelligent Model), it was developed a case study to forecast daily and weekly average of natural incremental streamflows between the Hydroelectric Plants: Porto Primavera, Rosana e Itaipu belonging to the the Parana River Basin. The model reaches up an error of about 3,5% to estimates of streamflows one day ahead, 16% to 12 days ahead, and 9% for average weekly forecast. This thesis aims to also investigate the effectiveness of the use of information of observed and predicted rainfall in the forecast flow, in conjunction with the use of the historical streamflows.

Despite the extent of a theoretical framework in financial market studies, a vast majority of the traders, investors and computer scientists have relied only on technical and timeseries data for predicting future prices. So far, the forecasting models have rarely incorporated macro-economic and market fundamentals successfully, especially with short-term predictions ranging less than a month. In this investigation on the predictability of certain financial markets, an attempt has been made to incorporate a un-exampled and encompassing set of parameters into an Artificial Neural Network prediction system. Experiments were carried out on three market instruments ??? namely currency exchange rates, share prices and oil prices. The choice of parameters for inclusion or exclusion, and the time frame adopted for the experimental sets were derived from the market literature. Good directional prediction accuracies were achieved for currency exchange rates and share prices with certain parameters as inputs, which consisted of predicting short-term movements based on past movements. These predictions were better than the results produced by a traditional least square prediction method. The trading strategy developed based on the predictions also achieved a higher percentage of winning trades. No significant predictions were observed for oil prices. These results open up questions in the microstructure of the markets and provide an insight into the inputs required for market forecasting in the corresponding time frame, for future investigation. The study concludes by advocating the use of trend based input parameters and suggests ways to improve neural network forecasting models.

This thesis analyses, develops and implements a very fast on-line parameter identification algorithm for both rotor and stator resistances of a rotor flux oriented induction motor drive, with the best possible convergence results using artificial neural networks and fuzzy logic systems. The thesis focuses mainly on identifying the rotor resistance, which is the most critical parameter for RFOC. Limitations of PI and fuzzy logic based estimators were identified. Artificial neural network based estimators were found to track the rotor and stator resistances of the drive accurately and fast. The rotor flux of the induction motor estimated with a classical voltage model was the key input of the rotor resistance estimator. Because, pure digital integrators were unable to play this role, an alternative rotor flux synthesizer using a programmable cascaded filter was developed. This rotor flux synthesizer has been used for all of the resistance estimators. It was found that the error in rotor resistance estimation using an ANN was contributed to by error in the stator resistance (caused by motor heating). Several stator resistance estimators using the stator current measurements were developed. The limitations of a PI and a fuzzy estimator for stator resistance estimation were also established. A new stator resistance identifier using an ANN was found to be much superior to the PI and fuzzy estimators, both in terms of dynamic estimation times and convergence problems. The rotor resistance estimator developed for this thesis used a feedforward neural network and the stator resistance estimator used a recurrent neural network. Both networks exhibited excellent learning capabilities; the stator resistance estimator network was very fast as it had a feedback input. A speed estimator was also developed with the state estimation principles, with the updated motor parameters supplied by the ANN estimators. Analysis for speed sensorless operation has shown that the stator and rotor resistances could be updated on-line.

就目前所知，這是第一篇應用人工類神經網路在股票風格投資方面的研究。類神經網路在樣本內與樣本外的分類正確率皆優於區別分析，而且類神經網路在樣本內的訓練範例中達成了百分之百的分類正確率。此外，我們也解決了傳統方法無法展示股票風格動態的問題。 檢視各種風格投資策略在台灣股票市場的績效表現之後，我們以神經網路為基礎，提出一個簡單而容易實行的投資策略。由這個策略的表現可以說明，即使在考慮了風險因素之後，積極的風格投資策略的確可以增加投資組合的績效表現。 / This is the first study of applying artificial neural networks (ANN) to classify and identify the equity styles. Regarding the accuracy, ANN outperforms discriminant analysis (DA) in all pure samples from 1987 to 1997. The ANN also commits the 100% classification accuracy for the in-sample training samples. In addition, the problem that traditional approach couldn't show equity style dynamics was solved with ANN and DA. The performances of style investing strategies were examined in Taiwan stock market. The proposed strategy is easily implemented by constructing portfolios based on the return, which neural networks forecasted. There is good evidence to show this simple strategy could enhance profit on the return and risk adjusted basis. This gives one evidence to illustrate that active style investing would add value.