This diploma thesis deals with the prediction of financial time series on capital markets using artificial intelligence methods. There are created several dynamic architectures of artificial neural networks, which are learned and subsequently used for prediction of future movements of shares. Based on the results an assessment and recommendations for working with artificial neural networks are provided.

Within this diploma thesis were evaluated factors, affecting consumption of drinking water from water supply system. Evaluation of time series of water consumption and chosen factors was performed using tools of mathematical statistics. In the last step was performed a dependence analysis of water consumption using artificial neuron network ANN. Diploma thesis was solved in cooperation with company Vodárenská akciová společnost, PLC and Severomoravské vodovody a kanalizace Ostrava, PCL.

The thesis deals with artificial neural networks theory. Subsequently, fuzzy sets are being described and fuzzy logic is explained. The hybrid neuro-fuzzy system stemming from ANFIS system is designed on the basis of artificial neural networks, fuzzy sets and fuzzy logic. The upper-mentioned systems' functionality has been demonstrated on an inverted pendulum controlling problem. The three controllers have been designed for the controlling needs - the first one is on the basis of artificial neural networks, the second is a fuzzy one, and the third is based on ANFIS system.  The thesis is aimed at comparing the described systems, which the controllers have been designed on the basis of, and evaluating the hybrid neuro-fuzzy system ANFIS contribution in comparison with particular theory solutions. Finally, some experiments with the systems are demonstrated and findings are assessed.

The thesis addresses the topic of Deep Neural Networks, in particular the methods regar- ding the field of Deep Learning, which is used to initialize the weight and learning process s itself within Deep Neural Networks. The focus is also put to the basic theory of the classical Neural Networks, which is important to comprehensive understanding of the issue. The aim of this work is to determine the optimal set of optional parameters of the algori- thms on various complexity levels of image recognition tasks through experimenting with created application applying Deep Neural Networks. Furthermore, evaluation and analysis of the results and lessons learned from the experimentation with classical and Deep Neural Networks are integrated in the thesis.

Modellbasierte virtuelle Sensoren (VS) ermöglichen die Messung von qualitätsbestimmenden Prozessparametern (bzw. Hilfsregelgrößen) dort, wo eine direkte Messung zu teuer oder gar nicht möglich ist. Für die adaptiven VS, die ihr internes Prozessmodell nach Data-Driven-Methode bilden (z. B. durch die Benutzung künstlicher neuronaler Netze (KNN)), besteht das Problem der Abschätzung der Prädiktionsstabilität. Aktuelle Lösungsansätze lösen dieses Problem nur für wenige KNN-Typen und erfordern enormen Entwurfs- und Rechenaufwand. In dieser Arbeit wird eine alternative Methode vorgestellt, welche für eine breite Klasse von KNN gilt und keinen hohen Entwurfs- und Rechenaufwand erfordert. Die neue Methode wurde anhand realer Anwendungsbeispiele getestet und hat sehr gute Ergebnisse geliefert. Für die nicht adaptiven virtuellen Sensoren wurde eine aufwandsreduzierte Adaption nach Smith-Schema vorgeschlagen. Dieses Verfahren ermöglicht die Regelung totzeitbehafteter und zeitvarianter Prozesse mit VS in einem geschlossenen Regelkreis. Im Vergleich zu anderen Regelungsstrategien konnte damit vergleichbare Regelungsqualität bei einem deutlich geringeren Entwurfsaufwand erzielt werden. / Model-based virtual sensors allow the measurement of parameters critical for process quality where a direct measurement is too expensive or not at all possible. For the adaptive virtual sensors built after data-driven method (e.g., by use of an ANN model) there is a problem of the prediction stability. Current solutions attempt to solve this problem only for a few ANN types and require a very high development effort. In this dissertation a new method for the solution of this problem is suggested, which is valid for a wide class of the ANNs and requires no high development effort. The new method was tested on real application examples and has delivered very good results. For the non-adaptive virtual sensors a simple adaptation mechanism was suggested. This technique allows the control of dead-time and time-variant processes in closed loop. Besides, in comparison to other control strategies the comparable results were achieved with smaller development effort.

[pt] As Redes Neurais Artificiais (RNAs) começaram a ser desenvolvidas nos anos 1940. Porém, foi a partir dos anos 1980, com a popularização e o aumento de capacidade dos computadores, que as RNAs passaram a ter grande relevância. Também nos anos 1980, houve dois outros acontecimentos acadêmicos relacionados ao presente trabalho: (i) um grande crescimento do interesse de econometristas por modelos não lineares, que culminou nas abordagens econométricas para RNAs, no final desta década; e (ii) a introdução do Método Generalizado dos Momentos (MGM) para estimação de parâmetros, em 1982. Nas abordagens econométricas de RNAs, sempre predominou a estimação por Quasi Máxima Verossimilhança (QMV). Apesar de possuir boas propriedades assintóticas, a QMV é muito suscetível a um problema nas estimações em amostra finita, conhecido como sobreajuste. O presente trabalho estende o estado da arte em abordagens econométricas de RNAs, apresentando uma proposta alternativa à estimação por QMV que preserva as suas boas propriedades assintóticas e é menos suscetível ao sobreajuste. A proposta utiliza a estimação pelo MGM. Como subproduto, a estimação pelo MGM possibilita a utilização do chamado Teste J para verifificar a existência de não linearidade negligenciada. Os estudos de Monte Carlo realizados indicaram que as estimações pelo MGM são mais precisas que as geradas pela QMV em situações com alto ruído, especialmente em pequenas amostras. Este resultado é compatível com a hipótese de que o MGM é menos suscetível ao sobreajuste. Experimentos de previsão de taxas de câmbio reforçaram estes resultados. Um segundo estudo de Monte Carlo apontou boas propriedades em amostra finita para o Teste J aplicado à não linearidade negligenciada, comparado a um teste de referência amplamente conhecido e utilizado. No geral, os resultados apontaram que a estimação pelo MGM é uma alternativa recomendável, em especial no caso de dados com alto nível de ruído. / [en] Artificial Neural Networks (ANN) started being developed in the decade of 1940. However, it was during the 1980 s that the ANNs became relevant, pushed by the popularization and increasing power of computers. Also in the 1980 s, there were two other two other academic events closely related to the present work: (i) a large increase of interest in nonlinear models from econometricians, culminating in the econometric approaches for ANN by the end of that decade; and (ii) the introduction of the Generalized Method of Moments (GMM) for parameter estimation in 1982. In econometric approaches for ANNs, the estimation by Quasi Maximum Likelihood (QML) always prevailed. Despite its good asymptotic properties, QML is very prone to an issue in finite sample estimations, known as overfiting. This thesis expands the state of the art in econometric approaches for ANNs by presenting an alternative to QML estimation that keeps its good asymptotic properties and has reduced leaning to overfiting. The presented approach relies on GMM estimation. As a byproduct, GMM estimation allows the use of the so-called J Test to verify the existence of neglected nonlinearity. The performed Monte Carlo studies indicate that the estimates from GMM are more accurate than those generated by QML in situations with high noise, especially in small samples. This result supports the hypothesis that GMM is susceptible to overfiting. Exchange rate forecasting experiments reinforced these findings. A second Monte Carlo study revealed satisfactory finite sample properties of the J Test applied to the neglected nonlinearity, compared with a reference test widely known and used. Overall, the results indicated that the estimation by GMM is a better alternative, especially for data with high noise level.

Les protocoles avec alternance de traitements sont des protocoles expérimentaux à cas uniques utiles pour évaluer et pour comparer l’efficacité d’interventions. Pour l’analyse de ces protocoles, les meilleures pratiques suggèrent aux chercheurs et aux professionnels d’utiliser conjointement les analyses statistiques et visuelles, mais ces méthodes produisent des taux d’erreurs insatisfaisants sous certaines conditions. Dans le but de considérer cet enjeu, notre étude a examiné l’utilisation de réseaux de neurones artificiels pour analyser les protocoles avec alternance de traitements et a comparé leurs performances à trois autres approches récentes. Plus précisément, nous avons examiné leur précision, leur puissance statistique et leurs erreurs de type I sous différentes conditions. Bien qu’il ne soit pas parfait, le modèle de réseaux de neurones artificiels présentait en général de meilleurs résultats et une plus grande stabilité à travers les analyses. Nos résultats suggèrent que les réseaux de neurones artificiels puissent être des solutions prometteuses pour analyser des protocoles avec alternance de traitements. / Alternating-treatment designs are useful single-case experimental designs for the evaluation and comparison of intervention effectiveness. Most guidelines suggest that researchers and practitioners use a combination of statistical and visual analyses to analyze these designs, but current methods still produce inadequate levels of errors under certain conditions. In an attempt to address this issue, our study examined the use of artificial neural networks to analyze alternating-treatment designs and compared their performances to three other recent approaches. Specifically, we examined accuracy, statistical power, and type I error rates under various conditions. Albeit not perfect, the artificial neural networks model generally provided better and more stable results across analyses. Our results suggest that artificial neural networks are promising alternatives to analyze alternating-treatment designs.

Inverse, multidimensional input-output flow mapping is very important for use of valves in precision motion control applications. Due to the highly nonlinear characteristic and uncertain model structure of the cartridge valves, it is hard to formulate the modelling of their flow mappings into simple parameter estimation problems. This contribution conducts a comprehensive analysis and validation of three- and four-dimensional input-output-mapping approaches for a proportional pilot operated seat valves. Therefore, a virtual and a physical test-rig setup are utilized for initial measurement, implementation and assessment. After modeling and validating the valve under consideration, as a function of flow, pressure and temperature different mapping methods are investigated. More specifically, state of the art approaches, deep-learning methods and a newly developed approach (extPoly) are examined. Especially ANNs and Polynomials show reasonable approximation results even for more than two inputs. However, the results are strongly dependent on the structure and distribution of the input data points. Besides identification effort, the invertibility was investigated.

Las Redes Neuronales Artificiales (RNA) se han desarrollado en el campo de ingeniería estructural cada vez más con el paso de los años, esta herramienta trata de simular el comportamiento de las neuronas biológicas permitiendo adaptarse a cualquier entorno y resolver diferentes tipos de problemas. En la presente investigación, se aplica al diseño estructural de viviendas de albañilería confinada a estructuras regulares con una geometría rectangular en planta. La aplicación de la red neuronal en este campo nos permite ahorrar tiempo y costos de diseño, así mismo, solo se requiere de personas con un conocimiento básico de manejo de computadoras o aplicativos móviles para la operación de la red neuronal. De este modo, es más sencillo poder otorgar diseños estructurales preliminares a usuarios con escasos recursos económicos que deseen construir viviendas de albañilería confinada. En la presente investigación, se aplican las redes neuronales para el diseño estructural de viviendas de albañilería confinada de 1 a 4 pisos ubicadas en los distritos de Lima con perfil de suelo tipo S1. Para ello, se realizó el diseño de 33 viviendas que cumplen con las especificaciones del Reglamento Nacional de Edificaciones (E020, E030 y E070), estos diseños se usaron como modelos de entrenamiento para el proceso de aprendizaje de la red neuronal y de esta manera se obtuvo un modelo de RNA capaz de diseñar viviendas de albañilería confinada con un error menor al 10%. / Artificial Neural Networks (ANN) have been developed in the field of structural engineering more and more over the years, this tool tries to simulate the behavior of biological neurons allowing to adapt to any environment and solve different types of problems. In the present research, it is applied to the structural design of masonry houses confined to regular structures with a rectangular geometry in plan. The application of the neuronal network in this field allows us to save time and design costs. Likewise, it only requires people with basic knowledge of computer handling or mobile applications for the operation of the neuronal network. In this way, it is easier to provide preliminary structural designs to users with limited economic resources who wish to build confined masonry housing. In the present investigation, the neural networks are applied for the structural design of confined masonry houses from 1 to 4 floors located in in Lima districts with soil profile type S1. For this, the design of 33 houses that meet the specifications of the National Building Regulations (E020, E030 and E070) was carried out, these designs were used as training models for the learning process of the neural network and in this way obtained an RNA model capable of designing confined masonry houses with an error of less than 10%. / Tesis

Los sobrecostos en la construcción de edificaciones de concreto armado representan pérdidas de un 28% de la inversión (Flyvbjerg, 2002), esto se debe a que las secciones de los elementos estructurales son sobredimensionadas y generan mayor costo en el concreto y acero. Por ello, se realizó la presente investigación en la que se empleó una metodología capaz de optimizar las dimensiones de los elementos estructurales (columnas, vigas y placas) en edificios multifamiliares regulares de 6 pisos. La metodología empleada se basó en el uso de redes neuronales del tipo feedforward, en la que se estableció como variables de entrada, los datos preliminares que se tienen de una edificación y como variables de salida las dimensiones de cada elemento estructural. Para ello, se elaboraron 30 edificios de 6 pisos como base de datos y en cada uno de estos se realizaron las verificaciones de derivas según la Norma Técnica Peruana E 0.30 y la resistencia de cada elemento estructural. De la base de datos se usaron 22 como entrenamiento y 8 para la validación interna de la red neuronal. La estructura de la red neuronal se estableció luego de ejecutar 10 diferentes redes neuronales y se seleccionó la red con un coeficiente de correlación más homogéneo y cercano a 1, en esta investigación fue de 0.98. Finalmente, se realizó la comparación del volumen de concreto que se emplea en una edificación dimensionada con métodos convencionales con el uso del software ETABS y los obtenidos con el uso de la metodología empleando redes neuronales artificiales, según esto, se calculó la diferencia de concreto entre ambos casos. Con los resultados obtenidos se comprobó que la metodología aplicada en esta investigación brinda un ahorro eficaz cercano al 10%. / Cost overruns in the construction of reinforced concrete buildings represent losses of 28% of the investment (Flyvbjerg, 2002), this is due to the fact that the sections of the structural elements are oversized and generate higher costs in concrete and steel. Therefore, the present research was carried out using a methodology capable of optimizing the dimensions of structural elements (columns, beams and slabs) in regular 6-story multifamily buildings. The methodology used was based on the use of feedforward neural networks, in which the preliminary data of a building were established as input variables and the dimensions of each structural element as output variables. For this purpose, 30 6-story buildings were prepared as a database and in each one of them the drift verifications were performed according to the Peruvian Technical Standard E 0.30 and the resistance of each structural element. From the database, 22 were used for training and 8 for the internal validation of the neural network. The structure of the neural network was established after running 10 different neural networks and the network with the most homogeneous correlation coefficient close to 1 was selected; in this research it was 0.98. Finally, a comparison was made between the volume of concrete used in a building dimensioned with conventional methods with the use of ETABS software and those obtained with the use of the methodology employing artificial neural networks, according to this, the difference of concrete between both cases was calculated. With the results obtained, it was proved that the methodology applied in this research provides an effective saving close to 10%. / Tesis