Esta pesquisa visa a análise da contribuição de cinco variáveis de entrada e a otimização do desempenho termo-hidráulico de trocadores de calor com venezianas combinados com geradores de vórtices delta-winglets. O desempenho termohidráulico de duas geometrias distintas, aqui nomeadas por GEO1 e GEO2, foram avaliadas. Smoothing Spline ANOVA foi usado para avaliar a contribuição dos parâmetros de entrada na transferência de calor e perda de carga. Considerando aplicação automotiva, foram investigados números de Reynolds iguais a 120 e 240, baseados no diâmetro hidráulico. Os resultados indicaram que o ângulo de venezianas é o maior contribuidor para o aumento do fator de atrito para GEO1 e GEO2, para ambos os números de Reynolds. Para o número de Reynolds menor, o parâmetro mais importante em termos de transferência de calor foi o ângulo das venezianas para ambas as geometrias. Para o número de Reynolds maior, o ângulo de ataque dos geradores de vórtices posicionados na primeira fileira é o maior contribuidor para a tranfesferência de calor, no caso da geometria GEO1, enquanto que o ângulo de ataque dos geradores de vórtices na primeira fileira foi tão importante quanto os ângulos das venezianas para a geometria GEO2. Embora as geometrias analisadas possam ser consideradas como técnicas compostas de intensificação da transferência de calor, não foram observadas interações relevantes entre ângulo de venezianas e parâmetros dos geradores de vórtices. O processo de otimização usa NSGA-II (Non-Dominated Sorting Genetic Algorithm) combinado com redes neurais artificiais. Os resultados mostraram que a adição dos geradores de vórtices em GEO1 aumentaram a transferência de calor em 21% e 23% com aumentos na perda de carga iguais a 24,66% e 36,67% para o menor e maior números de Reynolds, respectivamente. Para GEO2, a transferência de calor aumentou 13% e 15% com aumento na perda de carga de 20,33% e 23,70%, para o menor e maior número de Reynolds, respectivamente. As soluções otimizadas para o fator de Colburn mostraram que a transferência de calor atrás da primeira e da segunda fileiras de geradores de vórtices tem a mesma ordem de magnitude para ambos os números de Reynolds. Os padrões de escoamento e as características de transferência de calor das soluções otimizadas apresentaram comportamentos vi particulares, diferentemente daqueles encontrados quando as duas técnicas de intensificação de transferência de calor são aplicadas separadamente. / This doctoral thesis focuses on screening analysis of five input parameters and heat transfer and pressure drop optimization of flat-tube multi-louvered fin heat exchangers combined with delta-winglet vortex generators. The thermal-hydraulic performance of two distinct geometries, GEO1 and GEO2, were evaluated. Smoothing Spline ANOVA was used to evaluate the contribution of the input parameters such as louver angle, angle of attack of the delta-winglet and streamwise position of the delta-winglet on heat transfer and pressure drop. Taking the automotive application into account, Reynolds numbers of 120 and 240, based on hydraulic diameter, were investigated. The results indicated that the louver angle is the main contributor to increase the Friction factor for GEO1 and GEO2 for both Reynolds numbers. For the lower Reynolds number, the most important heat transfer parameter was the louver angle for both geometries, while at the higher Reynolds number, the angles of attack of the first row of delta-winglets mostly contributed to GEO1, and the angle of attack of the first row of delta-winglets was as important as the louver angle for GEO2. Although those specific geometries can be considered a kind of compound enhancement technique, relevant interactions were not verified between louvers and delta-winglet vortex generators parameters. The surrogatebased optimization procedure uses NSGA-II method (Non-Dominated Sorting Genetic Algorithm) combined with artificial neural networks. The results showed that the addition of DWLs on GEO1 increased the heat transfer of 21.27% and 23.52% with associated pressure loss increasing of 24.66% and 36.67% for the lower and the higher Reynolds numbers, respectively. For GEO2, the heat transfer was increased 13.48% and 15.67% with an increase of the pressure drop of 20.33% and 23.70%, for the lower and the higher Reynolds numbers, respectively. The optimized solutions for the Colburn factor showed that heat transfer behind the second row of deltawinglets has the same order of magnitude of that behind the first row, for both Reynolds numbers. The flow patterns and heat transfer characteristics from optimized solutions presented some particular behavior, differently from the findings when those two heat transfer enhancement techniques are applied separately.

Esta tese apresenta três ensaios sobre previsão de inflação e análise de dados em tempo real no Brasil. Utilizando uma curva de Phillips, o primeiro ensaio propõe um “modelo evolucionário” para prever inflação no Brasil. O modelo evolucionário consiste em uma combinação de um modelo não-linear (que é formado pela combinação de três redes neurais artificiais – RNAs) e de um modelo linear (que também é a referência para propósitos de comparação). Alguns parâmetros do modelo evolucionário, incluindo os pesos das combinações, evoluem ao longo do tempo segundo ajustes definidos por três algoritmos que avaliam os erros fora-da-amostra. As RNAs foram estimadas através de uma abordagem híbrida baseada em um algoritmo genético (AG) e em um algoritmo simplex de Nelder-Mead. Em um experimento de previsão fora-da-amostra para 3, 6, 9 e 12 passos à frente, o desempenho do modelo evolucionário foi comparado ao do modelo linear de referência, segundo os critérios de raiz do erro quadrático médio (REQM) e de erro absoluto médio (EAM). O desempenho do modelo evolucionário foi superior ao desempenho do modelo linear para todos os passos de previsão analisados, segundo ambos os critérios. O segundo ensaio é motivado pela recente literatura sobre análise de dados em tempo real, que tem mostrado que diversas medidas de atividade econômica passam por importantes revisões de dados ao longo do tempo, implicando importantes limitações para o uso dessas medidas. Elaboramos um conjunto de dados de PIB em tempo real para o Brasil e avaliamos a extensão na qual as séries de crescimento do PIB e de hiato do produto são revisadas ao longo do tempo. Mostramos que as revisões de crescimento do PIB (trimestre/trimestre anterior) são economicamente relevantes, embora as revisões de crescimento do PIB percam parte da importância à medida que o período de agregação aumenta (por exemplo, crescimento em quatro trimestres). Para analisar as revisões do hiato do produto, utilizamos quatro métodos de extração de tendência: o filtro de Hodrick-Prescott, a tendência linear, a tendência quadrática, e o modelo de Harvey-Clark de componentes não-observáveis. Todos os métodos apresentaram revisões de magnitudes economicamente relevantes. Em geral, tanto a revisão de dados do PIB como a baixa precisão das estimativas de final-de-amostra da tendência do produto mostraram-se fontes relevantes das revisões de hiato do produto. O terceiro ensaio é também um estudo de dados em tempo real, mas que analisa os dados de produção industrial (PI) e as estimativas de hiato da produção industrial. Mostramos que as revisões de crescimento da PI (mês/mês anterior) e da média móvel trimestral são economicamente relevantes, embora as revisões de crescimento da PI tornem-se menos importantes à medida que o período de agregação aumenta (por exemplo, crescimento em doze meses). Para analisar as revisões do hiato da PI, utilizamos três métodos de extração de tendência: o filtro de Hodrick-Prescott, a tendência linear e a tendência quadrática. Todos os métodos apresentaram revisões de magnitudes economicamente relevantes. Em geral, tanto a revisão de dados da PI como a baixa precisão das estimativas de final-de-amostra da tendência da PI mostraram-se fontes relevantes das revisões de hiato da PI, embora os resultados sugiram certa predominância das revisões provenientes da baixa precisão de final-de-amostra. / This thesis presents three essays on inflation forecasting and real-time data analysis in Brazil. By using a Phillips curve, the first essay presents an “evolutionary model” to forecast Brazilian inflation. The evolutionary model consists in a combination of a non-linear model (that is formed by a combination of three artificial neural networks - ANNs) and a linear model (that is also a benchmark for comparison purposes). Some parameters of the evolutionary model, including the combination weight, evolve throughout time according to adjustments defined by three algorithms that evaluate the out-of-sample errors. The ANNs were estimated by using a hybrid approach based on a genetic algorithm (GA) and on a Nelder-Mead simplex algorithm. In a 3, 6, 9 and 12 steps ahead out-of-sample forecasting experiment, the performance of the evolutionary model was compared to the performance of the benchmark linear model, according to root mean squared errors (RMSE) and to mean absolute error (MAE) criteria. The evolutionary model performed better than the linear model for all forecasting steps that were analyzed, according to both criteria. The second essay is motivated by recent literature on real-time data analysis, which has shown that several measures of economic activities go through important data revisions throughout time, implying important limitations to the use of these measures. We developed a GDP real-time data set to Brazilian economy and we analyzed the extent to which GDP growth and output gap series are revised over time. We showed that revisions to GDP growth (quarter-onquarter) are economic relevant, although the GDP growth revisions lose part of their importance as aggregation period increases (for example, four-quarter growth). To analyze the output gap revisions, we applied four detrending methods: the Hodrick-Prescott filter, the linear trend, the quadratic trend, and the Harvey-Clark model of unobservable components. It was shown that all methods had economically relevant magnitude of revisions. In a general way, both GDP data revisions and the low accuracy of end-of-sample output trend estimates were relevant sources of output gap revisions. The third essay is also a study about real-time data, but focused on industrial production (IP) data and on industrial production gap estimates. We showed that revisions to IP growth (month-on-month) and to IP quarterly moving average growth are economic relevant, although the IP growth revisions become less important as aggregation period increases (for example, twelve-month growth). To analyze the output gap revisions, we applied three detrending methods: the Hodrick-Prescott filter, the linear trend, and the quadratic trend. It was shown that all methods had economically relevant magnitude of revisions. In general, both IP data revisions and low accuracy of end-of-sample IP trend estimates were relevant sources of IP gap revisions, although the results suggest some prevalence of revisions originated from low accuracy of end-of-sample estimates.

O objetivo deste trabalho foi desenvolver um modelo para reconhecer e reproduzir padrões de viagens encadeadas a pé. O processo de modelagem foi conduzido através da aplicação das técnicas das Redes Neurais Artificiais (RNAs), utilizando-se de uma rede estática MLP e de rede dinâmica Elman. A análise do desempenho do modelo foi baseada nos dados de uma pesquisa de Origem-Destino realizada, em 1987, pelo METRÔ-SP na Região Metropolitana de São Paulo. Na modelagem foi fixado o modo de viagem a pé, e, na abordagem seqüencial, padrões de viagens individuais foram representados em termos de dois componentes: duração da viagem e tipo de atividades. A análise foi realizada partindo da classificação geral e específica para cada segmento do encadeamento de viagens, o que permitiu a comparação dos resultados entre padrões de viagens observados e os reproduzidos pelas redes. Na classificação geral, cinco dos padrões previstos com maior freqüência pelas RNAs representaram em média 58,9% dos indivíduos no conjunto de dados usado para testar o desempenho do modelo. Para o vetor de duas e quatro viagens, as redes neurais reproduziram 50% das durações de viagem e 90% das atividades, tais como Trabalho e Escola. Embora esses resultados não pareçam muito robustos, não significa que eles estejam errados. As porcentagens acima representam a probabilidade de uma pessoa realizar viagens com aquelas durações ou tipo de atividades. / The main objective of this work was to develop a model for recognizing and reproduzing trip-chaining patterns by walk. The process of modeling was conducted applying the techniques of Artificial Neural Networks (ANNs), by using one of the static networks MLP and the Elman dynamic network. The analysis of the performance of the model was based on the origin-destination home-interview survey carried out by METRÔ-SP in São Paulo Metropolitan Area in 1987. The mode of trip by walk was fixed in the model, and, in the sequential approach, individual travel patterns were represented in terms of two components: trip duration and activity type. The analysis was accomplished starting from the general and specific classifications for each segment of the chained trips, which allowed the comparison of the results between the observed travel patterns and reproduced ones through ANNs. In general classification, 5 of the patterns most frequently predicted by the ANNs represented 58.9% of the individuals in the dataset used for testing the model performance. For the vectors of two and four trips, the neural networks reproduced 50% of trip durations and 90% of the activities, such as work and school. Although those results seem not so robust, it does not mean that they are wrong. The percentages above represent the probability of a person making trips with those durations or type of activities.

L'imagerie par résonance magnétique fonctionnelle (IRMf) permet de mesurer l'activité cérébrale à travers le flux sanguin apporté aux neurones. Dans cette thèse nous évaluons la capacité de modèles biologiquement plausibles et issus de la vision par ordinateur à représenter le contenu d'une image de façon similaire au cerveau. Les principaux modèles de vision évalués sont les réseaux convolutionnels.Les réseaux de neurones profonds ont connu un progrès bouleversant pendant les dernières années dans divers domaines. Des travaux antérieurs ont identifié des similarités entre le traitement de l'information visuelle à la première et dernière couche entre un réseau de neurones et le cerveau. Nous avons généralisé ces similarités en identifiant des régions cérébrales correspondante à chaque étape du réseau de neurones. Le résultat consiste en une progression des niveaux de complexité représentés dans le cerveau qui correspondent à l'architecture connue des aires visuelles: Plus la couche convolutionnelle est profonde, plus abstraits sont ses calculs et plus haut niveau sera la fonction cérébrale qu'elle sait modéliser au mieux. Entre la détection de contours en V1 et la spécificité à l'objet en cortex inférotemporal, fonctions assez bien comprises, nous montrons pour la première fois que les réseaux de neurones convolutionnels de détection d'objet fournissent un outil pour l'étude de toutes les étapes intermédiaires du traitement visuel effectué par le cerveau.Un résultat préliminaire à celui-ci est aussi inclus dans le manuscrit: L'étude de la réponse cérébrale aux textures visuelles et sa modélisation avec les réseaux convolutionnels de scattering.L'autre aspect global de cette thèse sont modèles de “décodage”: Dans la partie précédente, nous prédisions l'activité cérébrale à partir d'un stimulus (modèles dits d’”encodage”). La prédiction du stimulus à partir de l'activité cérébrale est le méchanisme d'inférence inverse et peut servir comme preuve que cette information est présente dans le signal. Le plus souvent, des modèles linéaires généralisés tels que la régression linéaire ou logistique ou les SVM sont utilisés, donnant ainsi accès à une interprétation des coefficients du modèle en tant que carte cérébrale. Leur interprétation visuelle est cependant difficile car le problème linéaire sous-jacent est soit mal posé et mal conditionné ou bien non adéquatement régularisé, résultant en des cartes non-informatives. En supposant une organisation contigüe en espace et parcimonieuse, nous nous appuyons sur la pénalité convexe d'une somme de variation totale et la norme L1 (TV+L1) pour développer une pénalité regroupant un terme d'activation et un terme de dérivée spatiale. Cette pénalité a la propriété de mettre à zéro la plupart des coefficients tout en permettant une variation libre des coefficients dans une zone d'activation, contrairement à TV+L1 qui impose des zones d’activation plates. Cette méthode améliore l'interprétabilité des cartes obtenues dans un schéma de validation croisée basé sur la précision du modèle prédictif.Dans le contexte des modèles d’encodage et décodage nous tâchons à améliorer les prétraitements des données. Nous étudions le comportement du signal IRMf par rapport à la stimulation ponctuelle : la réponse impulsionnelle hémodynamique. Pour générer des cartes d'activation, au lieu d’un modèle linéaire classique qui impose une réponse impulsionnelle canonique fixe, nous utilisons un modèle bilinéaire à réponse hémodynamique variable spatialement mais fixe à travers les événements de stimulation. Nous proposons un algorithme efficace pour l'estimation et montrons un gain en capacité prédictive sur les analyses menées, en encodage et décodage. / Blood-oxygen-level dependent (BOLD) functional magnetic resonance imaging (fMRI) makes it possible to measure brain activity through blood flow to areas with metabolically active neurons. In this thesis we use these measurements to evaluate the capacity of biologically inspired models of vision coming from computer vision to represent image content in a similar way as the human brain. The main vision models used are convolutional networks.Deep neural networks have made unprecedented progress in many fields in recent years. Even strongholds of biological systems such as scene analysis and object detection have been addressed with enormous success. A body of prior work has been able to establish firm links between the first and last layers of deep convolutional nets and brain regions: The first layer and V1 essentially perform edge detection and the last layer as well as inferotemporal cortex permit a linear read-out of object category. In this work we have generalized this correspondence to all intermediate layers of a convolutional net. We found that each layer of a convnet maps to a stage of processing along the ventral stream, following the hierarchy of biological processing: Along the ventral stream we observe a stage-by-stage increase in complexity. Between edge detection and object detection, for the first time we are given a toolbox to study the intermediate processing steps.A preliminary result to this was obtained by studying the response of the visual areas to presentation of visual textures and analysing it using convolutional scattering networks.The other global aspect of this thesis is “decoding” models: In the preceding part, we predicted brain activity from the stimulus presented (this is called “encoding”). Predicting a stimulus from brain activity is the inverse inference mechanism and can be used as an omnibus test for presence of this information in brain signal. Most often generalized linear models such as linear or logistic regression or SVMs are used for this task, giving access to a coefficient vector the same size as a brain sample, which can thus be visualized as a brain map. However, interpretation of these maps is difficult, because the underlying linear system is either ill-defined and ill-conditioned or non-adequately regularized, resulting in non-informative maps. Supposing a sparse and spatially contiguous organization of coefficient maps, we build on the convex penalty consisting of the sum of total variation (TV) seminorm and L1 norm (“TV+L1”) to develop a penalty grouping an activation term with a spatial derivative. This penalty sets most coefficients to zero but permits free smooth variations in active zones, as opposed to TV+L1 which creates flat active zones. This method improves interpretability of brain maps obtained through cross-validation to determine the best hyperparameter.In the context of encoding and decoding models, we also work on improving data preprocessing in order to obtain the best performance. We study the impulse response of the BOLD signal: the hemodynamic response function. To generate activation maps, instead of using a classical linear model with fixed canonical response function, we use a bilinear model with spatially variable hemodynamic response (but fixed across events). We propose an efficient optimization algorithm and show a gain in predictive capacity for encoding and decoding models on different datasets.

No description available.

Cette thèse propose des approches neuromimétiques d'identification et de commande avec des applications directes au Filtre Actif Parallèle (FAP) et au Moteur Synchrone à Aiment Permanent (MSAP). Une structure neuronale complète a été développée pour réaliser toutes les fonctionnalités d'un FAP pour compenser des harmoniques de courant. La phase instantanée et les composantes symétriques d'un système triphasé de tensions ou de courants ont été estimées avec une boucle à verrouillage de phase neuronale. L'identification des harmoniques de courant a été réalisée avec des réseaux de neurones de type Adaline opérant dans les différents repères. Plusieurs schémas de commande ont été développés pour réinjecter les courants de compensation à l'aide d'un onduleur. Ils sont basés sur des techniques neuromimétiques, sur la logique floue, ou sur leur association. Une approche neuronale a été développée pour commander une MSAP à distribution quelconque avec des contraintes prédéterminées réduisant les ondulations du couple. Elle consiste en des schémas de commande directe en couple ou en vitesse pour obtenir les courants statoriques optimaux qui donnent exactement le couple électromagnétique (ou la vitesse) désiré et qui réduisent au maximum les pertes par effet Joule. Ces commandes intègrent deux blocs neuronaux, l'un dédié au calcul des courants optimaux et l'autre pour assurer leur génération à travers un onduleur de tension. Toutes les approches neuromimétiques ont été validées par des tests de simulation et des essais expérimentaux. Des comparaisons avec les méthodes de commande classique démontrent des caractéristiques supérieures en termes de performance et de robustesse. / This thesis proposes Artificial Neural Networks (ANN) approaches for the identification and the control of an Active Power Filter (APF) and a Permanent-Magnet Synchronous Motor (PMSM). A completed neural architecture was developed for an APF for harmonic currents compensation. The instantaneous phase and the symmetrical components of a three-phase voltage or current were estimated with a neural phase Jock loop. The harmonic terms were identified by Adaline neural networks that estimate the instantaneous powers within different reference frames. Several intelligent techniques, based on neural networks, fuzzy logic or their association, were developed to control the inverter used to inject the harmonic currents phase-opposite. An original neural approach was also carried out for reducing the torque ripple of a non-sinusoidal PMSM. It consists in a direct torque or in a speed control schemes that elaborate the optimal stator currents which exactly give a desired electromagnetic torque or speed and which minimize the ohmic losses. The control schemes integrate two neural networks, one to calculate the optimal currents and one to ensure their generation through an inverter. The neural network approaches were all evaluated by simulated and experimental tests. The results confirm their excellent characteristics in terms of both performance and robustness. Comparisons with conventional methods prove their superiority.

Two types of artificial neural networks, Multilayer Perceptron (MLP) and Self-organizing Feature Map (SOM), were employed to detect mastitis for robotic milking stations using the preprocessed data relating to the electrical conductivity and milk yield. The SOM was developed to classify the health status into three categories: healthy, moderately ill and severely ill. The clustering results were successfully evaluated and validated by using statistical techniques such as K-means clustering, ANOVA and Least Significant Difference. The result shows that the SOM could be used in the robotic milking stations as a detection model for mastitis. For developing MLP models, a new mastitis definition based on higher EC and lower quarter yield was created and Principle Components Analysis technique was adopted for addressing the problem of multi-colinearity existed in the data. Four MLPs with four combined datasets were developed and the results manifested that the PCA-based MLP model is superior to other non-PCA-based models in many respects such as less complexity, higher predictive accuracy. The overall correct classification rate (CCR), sensitivity and specificity of the model was 90.74 %, 86.90 and 91.36, respectively. We conclude that the PCA-based model developed here can improve the accuracy of prediction of mastitis by robotic milking stations.

為了瞭解影響演化性神經網路（ENT）預測表現的四項重要的機制：輸入資料性質、訓練樣本大小、網路搜尋密度以及控制模型複雜度，進而找出能使ENT充分發揮效果的組合。在本論文中首先設計ENT在模擬資料上的實驗，探討上述四項機制個別對預測表現的影響，再依照實驗結果的建議，設計能讓ENT發揮功效的組合，並以實際金融高頻率資料：香港恆生指數在一九九八年十二月報酬率為標的，探討模擬資料的結果在實際金融資料需要調整的部份。實驗結果顯示，當輸入資料經過線性過濾後，搭配大樣本訓練、高搜尋強度與適當地模型複雜度控制，會是能讓神經網路提高預測能力的組合。在實際金融資料的實驗當中同時發現，資料中偶而出現特別高或特別低的變化，會對ENT的預測表現有相當程度的影響。 / In this thesis, Evolutionary Neural Trees (ENTs) are applied to forecast the artificial data generated by financial and chaos models — iid random, linear process (Auto Regressive-Moving Average；ARMA), nonlinear processes (AutoRegressive Conditional Heteroskedasticity；ARCH, General AutoRegressive Conditional Heteroskedasticity；GARCH, Bilinear), mixed linear and nonlinear process (AR and GARCH). Experiments of the artificial data were conducted to understand the characteristics of ENTs mechanism. – data pre-processing procedures, search intensity, sample size and complexity regularization. From the experiment results of artificial data, the combination of pure linear or nonlinear time series, large sample size, intensive search and simple neural trees are suggested for the parameters setting of ENTs. And for the sake of computational burden, we have a trade-off between search intensity and sample size. Ten experiments are designed for ENTs modeling on the high-frequency stock returns of Heng Sheng stock index on December, 1998, in order to have an efficient combination of the factors of ENTs. The results show that ENTs would perform more efficiently if data are pre-processed by a linear filter, for ENTs will concentrate on searching in the space of nonlinear signals. Also, as is well demonstrated in this study, the infrequent bursts (outliers) appearing in the data set can be very disturbing for the ENTs modeling.

Establishing and maintaining successful communication stands as a critical prerequisite for achieving the goals of inducing and studying advanced computation in small-scale living neuronal networks. The following work establishes a novel and effective method for communicating arbitrary "sensory" input information to cultures of living neurons, living neuronal networks (LNNs), consisting of approximately 20 000 rat cortical neurons plated on microelectrode arrays (MEAs) containing 60 electrodes. The sensory coding algorithm determines a set of effective codes (symbols), comprised of different spatio-temporal patterns of electrical stimulation, to which the LNN consistently produces unique responses to each individual symbol. The algorithm evaluates random sequences of candidate electrical stimulation patterns for evoked-response separability and reliability via a support vector machine (SVM)-based method, and employing the separability results as a fitness metric, a genetic algorithm subsequently constructs subsets of highly separable symbols (input patterns). Sustainable input/output (I/O) bit rates of 16-20 bits per second with a 10% symbol error rate resulted for time periods of approximately ten minutes to over ten hours. To further evaluate the resulting code sets' performance, I used the system to encode approximately ten hours of sinusoidal input into stimulation patterns that the algorithm selected and was able to recover the original signal with a normalized root-mean-square error of 20-30% using only the recorded LNN responses and trained SVM classifiers. Response variations over the course of several hours observed in the results of the sine wave I/O experiment suggest that the LNNs may retain some short-term memory of the previous input sample and undergo neuroplastic changes in the context of repeated stimulation with sensory coding patterns identified by the algorithm.

Η παρούσα διατριβή ασχολείται με την μελέτη και την εκπαίδευση Τεχνητών Νευρωνικών Δικτύων (ΤΝΔ) με μεθόδους Βελτιστοποίησης και τις εφαρμογές αυτών. Η παρουσίαση των επιμέρους θεμάτων και αποτελεσμάτων της διατριβής αυτής οργανώνεται ως εξής : Στο κεφάλαιο 1 παρέχουμε τους βασικούς ορισμούς και περιγράφουμε τη δομή και τη λειτουργία των ΤΝΔ. Στη συνέχεια, παρουσιάζουμε μια συντομή ιστορική αναδρομή, αναφέρουμε μερικά από τα πλεονεκτήματα της χρήσης των ΤΝΔ και συνοψίζουμε τους κύριους τομείς όπου τα ΤΝΔ εφαρμόζονται. Τέλος, περιγράφουμε τις βασικές κατηγορίες μεθόδων εκπαίδευσης. Το κεφάλαιο 2 αφιερώνεται στη μαθηματική θεμελίωση της εκπαίδευσης ΤΝΔ. Περιγράφουμε τη γνωστή μέθοδο της οπισθοδρομικής διάδοσης του σφάλματος (Backpropagation) και δίνουμε αποδείξεις σύγκλισης για μια κλάση μεθόδων εκπαίδευσης που χρησιμοποιούν μονοδιάστατες ελαχιστοποιήσεις. Στο τέλος του κεφαλαίου παρουσιάζουμε κάποια θεωρητικά αποτελέσματα σχετικά με την ικανότητα των ΤΝΔ να προσεγγίζουν άγνωστες συναρτήσεις. Στο κεφάλαιο 3 προτείνουμε μια νέα κλάση μεθόδων εκπαίδευσης ΤΝΔ και αποδεικνύουμε ότι αυτές έχουν την ιδιότητα της ευρείας σύγκλισης , δηλαδή συγκλίνουν σε ένα ελάχιστο της αντικειμενικής συνάρτησης σχεδόν από οποιαδήποτε αρχική συνθήκη. Τα αποτελέσματα μας δείχνουν ότι η προτεινόμενη τεχνική μπορεί να βελτιώσει οποιαδήποτε μέθοδο της κλάσης της οπισθοδρομικής διάδοσης του σφάλματος. Στο επόμενο κεφάλαιο παρουσιάζουμε τη γνωστή μέθοδο Quick-Prop και μελετάμε τις ιδιότητες σύγκλισής της. Με βάση το θεωρητικό αποτέλεσμα που προκύπτει, κατασκευάζουμε μια νέα τροποποίηση της μεθόδου Quick-Prop, που έχει την ιδιότητα της ευρείας σύγκλισης και βελτιώνει σημαντικά την κλασίκη Quick-Prop μέθοδο. Στα επόμενα δύο κεφάλαια μελετάμε την εκπαίδευση ΤΝΔ με μεθόδους ολικής Βελτιστοποίησης. Πιο συγκεκριμένα, στο Κεφάλαιο 5 προτείνουμε και μελετάμε διεξοδικά μια νέα κλάση μεθόδων που είναι ικανές να εκπαιδεύσουν ΤΝΔ με περιορισμένα ακέραια βάρη. Στη συνέχεια, επεκτείνουμε τις μεθόδους αυτές έτσι ώστε να υλοποιούνται σε παράλληλους υπολογιστές και να εκπαιδεύουν ΤΝΔ με χρήση συναρτήσεων κατωφλιών. Το κεφάλαιο 6 πραγματεύεται την εφαρμογή γνωστών μεθόδων όπως οι Γενετικοί Αλγόριθμοι, η μέθοδος της προσομοιωμένης ανόπτησης ( Simulated Annealing ) και η μέθοδος βελτιστοποίησης με σμήνος σωματιδίων (Particle Swarm Optimization) στην εκπαίδευση ΤΝΔ. Επίσης, παρουσιάζουμε νέους μετασχηματισμούς της αντικειμενικής συνάρτησης με σκοπό την σταδιακή εξάλειψη των τοπικών ελαχίστων της. Στο κεφάλαιο 7 κάνουμε μια σύντομη ανασκόπηση της στοχαστικής μεθόδου της πιο απότομης κλίσης (stochastic gradient descent) για την εκπαίδευση ΤΝΔ ανά πρότυπο εισόδου και προτείνουμε μια νέα τέτοια μέθοδο . Η νέα μέθοδος συγκρίνεται με άλλες γνωστές μεθόδους και τα πειράματά μας δείχνουν ότι υπερτερεί. Η παρουσίαση του ερευνητικού έργου για αυτή τη διατριβή ολοκληρώνεται με το Κεφάλαιο 8, όπου προτείνουμε και μελετάμε εκτενώς μη μονότονες μεθόδους εκπαίδευσης ΤΝΔ. Η τεχνική που προτείνουμε μπορεί να εφαρμοστεί σε κάθε μέθοδο της κλάσης της οπισθοδρομικής διάδοσης του σφάλματος με αποτέλεσμα η τροποποιημένη μέθοδος να έχει την ικανότητα , πολλές φορές, να αποφεύγει τοπικά ελάχιστα της αντικειμενικής συνάρτησης. Η παρουσίαση της διατριβής ολοκληρώνεται με το κεφάλαιο 9 και δύο Παραρτήματα. Το Κεφάλαιο 9 περιέχει τα γενικά συμπεράσματα της διατριβής. Στο παράρτημα Α παρουσιάζουμε συνοπτικά μερικά από τα προβλήματα εκπαίδευσης που εξετάσαμε στα προηγούμενα κεφάλαια και τέλος στο Παράρτημα Β δίνουμε την απόδειξη της μεθόδου της οπισθοδρομικής διάδοσης του σφάλματος. / -