Thesis (MScEng)--Stellenbosch University, 2012. / ENGLISH ABSTRACT: Construction projects are risky by nature, with many variables a ecting their outcome. A contingency cost and duration are allocated to the budget and schedule of a project to provide for the possible impact of risks. To enable the management of project-related risk on a portfolio level, contingency estimation must be performed consistently and objectively. The current project contingency estimation method used in the capital program management department of Eskom Distribution Western Cape Operating Unit is not standardised, and is based solely on expert opinion. The aim of the study was to develop a contingency estimation tool to decrease the in uence of subjectivity on contingency estimation methods throughout the project lifecycle so as to enable consistent project risk re ection on a portfolio level. From a review of contingency estimation approaches in literature, a hybrid method combining neural network analysis of systemic risks and expected value analysis of project-speci c risks was chosen. Interviews were conducted with project managers (regarding network asset construction projects completed in the last two nancial years) to distinguish systemic and project-speci c risk impact on cost and duration growth. Outputs from 22 interviews provided three data patterns for each of 89 projects. After interview data processing, 138 training patterns pertaining to 85 projects remained for neural network training, validation and testing. Six possible neural network inputs (systemic risk drivers) were selected as project de nition level, cost, duration, business category, voltage category and job category. A multilayer feedforward neural network was trained using a supervised training approach combining a multi-objective simulated annealing algorithm with the standard backpropagation algorithm. Neural network results were evaluated for di erent scenarios considering possible combinations of model input variables and number of hidden nodes. The best scenario (exclusion of business category input with nine hidden nodes) was chosen based on training and validation errors. Validation error levels are comparable to those of similar studies in the project management eld. The chosen scenario was shown to outperform multiple linear regression, but calculated R2 values were lower than anticipated. It is expected that neural network performance will further improve as additional training patterns become available. The trained neural network was combined with an expected value analysis tool (risk register format) to estimate contingency due to systemic risks alongside an estimation of contingency due to project-speci c risks. The project-speci c expected value method was modi ed by basing the contingency estimation on the expected number of realised risks according to a binomial scenario. A total cost distribution was included in tool outputs by assuming the contingency cost equal to the standard deviation of the cost estimate. To aid business integration of the developed tool, study outputs included the points in the project lifecycle model at which the tool should be applied, and the process by which tool outputs become inputs to the enterprise risk management system. By following this approach, systemic and project-speci c risks are contained in a single tool providing contingency cost and duration output on project level, while enabling integration with reporting on program, portfolio and enterprise level. / AFRIKAANSE OPSOMMING: Konstruksieprojekte het van nature 'n ho e risiko omdat hulle uitsette deur baie veranderlikes gea ekteer word. Gebeurlikheidsreserwes vir koste en tyd word toegeken aan die begroting en skedule van 'n projek om voorsiening te maak vir die moontlike gevolge van risiko's. Om die bestuur van projekverwante risiko op 'n portefeulje-vlak te vergemaklik, moet die beraming van gebeurlikheidsreserwes op 'n konsekwente en objektiewe manier uitgevoer word. Die huidige beramingsmetode vir projek gebeurlikheidsreserwes in die kapitaal programbestuur departement van Eskom Distribusie Wes-Kaap Bedryfseenheid is nie gestandardiseer nie, en word slegs gebaseer op deskundige opinie. Die doel van hierdie studie was om 'n gebeurlikheidsreserwe beramingsinstrument te ontwikkel wat die invloed van subjektiwiteit op beramingsmetodes verminder deur die hele projeklewensiklus, en sodoende die konsekwente weerspie eling van projekrisiko op 'n portefeulje-vlak, te bewerkstellig. Vanuit 'n studie van bestaande literatuur oor gebeurlikheidsreserwe-beraming, is 'n hibriede metode wat neurale netwerk analise van sistemiese risiko's en verwagte waarde analise van projek-spesi eke risiko's kombineer, gekies. Onderhoude is gevoer met projekbestuurders (rakende netwerk batekonstruksieprojekte wat voltooi is in die afgelope twee nansi ele jare) om te onderskei tussen die impak van sistemiese en projek-spesi eke risiko's op koste- en duurgroei. Uitsette van 22 onderhoude het drie datapatrone vir elk van 89 projekte verskaf. Na onderhouddata verwerk is, het 138 datapatrone vanuit 85 projekte oorgebly vir neurale netwerk opleiding, validasie en toetsing. Ses moontlike neurale netwerk insette (sistemiese risikodrywers) is gekies as projek de nisievlak, koste, duur, besigheidskategorie, spanningskategorie en werkskategorie. 'n Multi-laag vooruitvoerende neurale netwerk is deur 'n opleidingonder- toesig benadering opgelei { 'n multi-doelwit gesimuleerde uitgloei ngsalgoritme gekombineer met die standaard agteruit-propagerende algoritme. Die resultate van die neurale netwerk is oorweeg vir verskillende scenario's rakende moontlike kombinasies van die aantal versteekte nodes en model insetveranderlikes. Die beste scenario (uitsluiting van besigheidskategorie inset met nege versteekte nodes) is gekies op grond van opleidings- en validasiefoute. Validasie foutvlakke is vergelykbaar met di e van soortgelyke studies in die projekbestuur veld. Daar is gewys dat die gekose scenario meervoudige line^ere regressie klop, maar met laer R2 waardes as wat verwag is. Dit word verwag dat die neurale netwerk beter sal presteer soos bykomende opleidingsdatapatrone beskikbaar word. Die opgeleide neurale netwerk is gekombineer met 'n verwagte waarde analise instrument (risiko-register formaat) om gebeurlikheidsreserwes as gevolg van sistemiese risiko's hand-aan-hand met gebeurlikheidsreserwes as gevolg van projekspesi eke risiko's, te beraam. Die projek-spesi eke verwagte waarde metode is aangepas deur gebeurlikheidsreserwe-beraming te baseer op die aantal verwagte gerealiseerde risiko's volgens 'n binomiaal scenario. 'n Totale koste-verdeling is ingesluit in modeluitsette deur aan te neem dat die gebeurlikheidsreserwe vir koste gelyk is aan die standaardafwyking van die kosteberaming. Om die besigheidsintegrasie van die ontwikkelde instrument te vergemaklik, het studie uitsette die punte in die projek lewensiklus waarby die instrument toegepas moet word, en die proses waardeur instrument uitsette omgesit word na insette vir die risikobestuur sisteem op ondernemingsvlak, ingesluit. Deur hierdie benadering te volg, word sistemiese en projek-spesi eke risiko's omvat in een instrument wat gebeurlikheidsreserwes vir koste en tyd op projekvlak verskaf. Die integrasie met verslagdoening op program-, portefeulje- en ondernemingsvlak word ook bewerkstellig.

Thesis (MSc)--Stellenbosch University, 2015. / ENGLISH ABSTRACT: This study has started as part of a research project at Stellenbosch University (SU) that aims at building a team of soccer-playing robots for the RoboCup Small Size League (SSL). In the RoboCup SSL the Decision- Making Module (DMM) plays an important role for it makes all decisions for the robots in the team. This research focuses on the development of some parts of the DMM for the team at SU. A literature study showed that the DMM is typically developed in a hierarchical structure where basic soccer skills form the fundamental building blocks and high-level team behaviours are implemented using these basic soccer skills. The literature study also revealed that strategies in the DMM are usually developed using a hand-coded approach in the RoboCup SSL domain, i.e., a specific and fixed strategy is coded, while in other leagues a Machine Learning (ML) approach, Reinforcement Learning (RL) in particular, is widely used. This led to the following research objective of this thesis, namely to develop basic soccer skills using RL for the RoboCup Small Size League. A second objective of this research is to develop a simulation environment to facilitate the development of the DMM. A high-level simulator was developed and validated as a result. The temporal-difference value iteration algorithm with state-value functions was used for RL, along with a Multi-Layer Perceptron (MLP) as a function approximator. Two types of important soccer skills, namely shooting skills and passing skills were developed using the RL and MLP combination. Nine experiments were conducted to develop and evaluate these skills in various playing situations. The results showed that the learning was very effective, as the learning agent executed the shooting and passing tasks satisfactorily, and further refinement is thus possible. In conclusion, RL combined with MLP was successfully applied in this research to develop two important basic soccer skills for robots in the RoboCup SSL. These form a solid foundation for the development of a complete DMM along with the simulation environment established in this research. / AFRIKAANSE OPSOMMING: Hierdie studie het ontstaan as deel van 'n navorsingsprojek by Stellenbosch Universiteit wat daarop gemik was om 'n span sokkerrobotte vir die RoboCup Small Size League (SSL) te ontwikkel. Die besluitnemingsmodule (BM) speel 'n belangrike rol in die RoboCup SSL, aangesien dit besluite vir die robotte in die span maak. Hierdie navorsing fokus op ontwikkeling van enkele komponente van die BM vir die span by SU. 'n Literatuurstudie het getoon dat die BM tipies ontwikkel word volgens 'n hiërargiese struktuur waarin basiese sokkervaardighede die fundamentele boublokke vorm en hoëvlak spangedrag word dan gerealiseer deur hierdie basiese vaardighede te gebruik. Die literatuur het ook getoon dat strategieë in die BM van die RoboCup SSL domein gewoonlik ontwikkel word deur 'n hand-gekodeerde benadering, dit wil s^e, 'n baie spesifieke en vaste strategie word gekodeer, terwyl masjienleer (ML) en versterkingsleer (VL) wyd in ander ligas gebruik word. Dit het gelei tot die navorsingsdoelwit in hierdie tesis, naamlik om basiese sokkervaardighede vir robotte in die RoboCup SSL te ontwikkel. 'n Tweede doelwit was om 'n simulasie-omgewing te ontwikkel wat weer die ontwikkeling van die BM sou fasiliteer. Hierdie simulator is suksesvol ontwikkel en gevalideer. Die tydwaarde-verskil iterariewe algoritme met toestandwaarde-funksies is gebruik vir VL saam met 'n multi-laag perseptron (MLP) vir funksiebenaderings. Twee belangrike sokkervaardighede, naamlik doelskop- en aangeevaardighede is met hierdie kombinasie van VL en MLP ontwikkel. Nege eksperimente is uitgevoer om hierdie vaardighede in verskillende speelsituasies te ontwikkel en te evalueer. Volgens die resultate was die leerproses baie effektief, aangesien die leer-agent die doelskiet- en aangeetake bevredigend uitgevoer het, en verdere verfyning is dus moontlik. Die gevolgtrekking is dat VL gekombineer met MLP suksesvol toegepas is in hierdie navorsingswerk om twee belangrike, basiese sokkervaardighede vir robotte in die RoboCup SSL te ontwikkel. Dit vorm 'n sterk fondament vir die ontwikkeling van 'n volledige BM tesame met die simulasie-omgewing wat in hierdie werk daargestel is.

Made available in DSpace on 2014-06-11T19:30:50Z (GMT). No. of bitstreams: 0 Previous issue date: 2011-10-07Bitstream added on 2014-06-13T20:00:54Z : No. of bitstreams: 1 furini_ma_dr_ilha.pdf: 763780 bytes, checksum: c5cdfef6ac6feb2737b71cbcd978ec47 (MD5) / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / Neste trabalho é proposta a utilização da rede neural artificial (RNA) ARTMAP Nebulosa (fuzzy) no ajuste de parâmetros de controladores suplementares para o amortecimento de oscilações eletromecânicas de sistemas elétricos de potência, visando tornar este ajuste mais eficiente. Análises comparativas da atuação das redes neurais artificiais ARTMAP Nebulosa e Perceptron Multicamadas (PM) são realizadas para dois sistemas multimáquinas considerando o ajuste individual e coordenado dos controladores. Tais redes são utilizadas para o projeto dos controladores ESP (Estabilizadores de Sistemas de Potência) e POD (Power Oscillation Damping) acoplado ao dispositivo FACTS (Flexible Alternating Current Transmission Systems) UPFC (Unified Power Flow Controller). Será evidenciado que a RNA ARTMAP Nebulosa pode ser utilizada na melhora da estabilidade dinâmica, fornecendo resultados muito semelhantes aos da RNA Perceptron Multicamadas. Entretanto, é importante enfatizar que a vantagem da utilização da RNA ARTMAP Nebulosa está no fato da garantia da estabilidade e plasticidade associadas a um rápido treinamento, o que não ocorre com a RNA Perceptron Multicamadas / This work proposes the use of artificial neural network (ANN) Fuzzy ARTMAP to adjust the parameters of additional controllers to damp electromechanical oscillations in electric power systems in order to make this adjustment more efficient due to variations in load. Comparative analysis of the performance of artificial neural networks Fuzzy ARTMAP and Multilayer Perceptron are performed for two multimachine systems, considering individual and coordinated controller adjustment. Those networks are used for the design of Power System Stabilizers (PSS) and Power Oscillation Damping (POD) that is coupled to the FACTS (Flexible Alternating Current Transmission Systems) UPFC (Unified Power Flow Controller). It will be shown that the ANN Fuzzy ARTMAP can be used in the improvement of dynamic stability, providing very similar results to the ANN Multilayer Perceptron. However, it is important to emphasize that the advantage of using ANN Fuzzy ARTMAP is the guarantee of stability and plasticity associated with a fast training process which does not occur for the ANN Multilayer Perceptron

Made available in DSpace on 2014-06-11T19:30:52Z (GMT). No. of bitstreams: 0 Previous issue date: 2005-10-27Bitstream added on 2014-06-13T20:00:56Z : No. of bitstreams: 1 lopes_mlm_dr_ilha.pdf: 1509538 bytes, checksum: 3842df54e0429972a030219c885bd09a (MD5) / Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP) / Nos dias atuais, principalmente pelo fato de alguns sistemas serem desregulamentados, o estudo dos problemas de análise, planejamento e operação de sistemas de energia elétrica é de extrema importância para o funcionamento do sistema. Para isso é necessário que se obtenha, com antecedência, o comportamento da carga elétrica com o propósito de garantir o fornecimento de energia aos consumidores de forma econômica, segura e contínua. Este trabalho propõe o desenvolvimento de redes neurais artificiais utilizadas para resolver o problema de previsão de cargas elétricas. Para tanto, inicialmente, propôs-se a introdução de melhorias na rede neural feedforward com treinamento realizado utilizando o algoritmo retropropagação. Neste caso, foi desenvolvida/implementada a adaptação dos parâmetros de inclinação e translação da função sigmóide (função de ativação da rede neural). A inclusão desta nova estrutura de redes neurais produziu melhores resultados, se comparado à rede neural retropropagação convencional. Essas arquiteturas proporcionam bons resultados, porém, são estruturas de redes neurais que possuem o problema de convergência. O problema de previsão de cargas elétricas a curto-prazo necessita de uma rede neural que forneça uma saída de forma rápida e eficaz. No intuito de solucionar os problemas encontrados com o algoritmo retropropagação foi desenvolvida/implementada uma rede neural baseada na arquitetura ART (Adaptive Rossonance Theory), denominada rede neural ART&ARTMAP nebulosa, aplicada ao problema de previsão de carga elétrica. Trata-se, por conseguinte, da principal contribuição desta tese. As redes neurais, baseadas na arquitetura ART, possuem duas características fundamentais que são de extrema importância para o desempenho da rede (estabilidade e plasticidade), que permite a implementação do treinamento de modo contínuo... / Nowadays due to the deregulamentation it is very important to study the problems of analyzing, planning and operation of electric power systems. For a reliable operation it is necessary to know previously the behavior of the load to guarantee the energy providing to the users with security and continuity and in an economic way. This work proposes to develop artificial neural networks to solve the problem of electric load forecasting. First, it is introduced some improvements on the feedforward neural network, with the training effectuated with the backpropagation algorithm. The improvement was the adaptation of the inclination and translation parameters of the sigmoid function (activation function of the neural network). The inclusion of this new structure provides better results if compared to the conventional backpropagation algorithm. These architectures provide good results, although they are structures that have some convergence problems. The short term electric load forecasting problem needs a neural network that provide a fast and efficient output. To solve this problem a neural network based on the ART (Adaptive Ressonance Theory), called_ fuzzy ART&ARTMAP applied to the load-forecasting problem, was developed and implemented._This is one of the contributions of this work. Neural networks based on the ART architecture have two important characteristics for the network performance, which are stability and plasticity, allowing the continuous training. The fuzzy ART&ARTMAP neural network reduces the imprecision of the results by a mechanism that separates the binary and analogical data and processing them separately. This represents a quality and an improvement on the results (reduction of the processing time and better precision), if compared to the neural network with backpropagation training (often considered as a benchmark in precision by the specialized...(Complete abastract click electronic access below)

La connaissance du fonctionnement hydrodynamique des karsts représente un enjeu planétaire pour la ressource en eau car ils alimentent en eau potable près de 25% de la population mondiale. Néanmoins, la complexité, l'anisotropie, l'hétérogénéité, la non-linéarité et l'éventuelle non-stationnarité de ces aquifères en font des objets encore largement sous-exploités du fait de la difficulté de caractériser leur morphologie et leur fonctionnement hydrodynamique. Dans ce contexte, le paradigme systémique permet d'apporter de nouvelles méthodes en étudiant ces hydrosystèmes au travers de la relation entre leurs signaux d'entrée (pluie) et de sortie (débit). Ainsi ce travail porte sur l'utilisation : i) d'analyses corrélatoires et spectrales pour caractériser la réponse des hydrosystèmes karstiques, ii) des réseaux de neurones pour étudier les relations linéaires et non-linéaires de ces hydrosystèmes. Pour ce faire, différents types de configuration de modèles par réseau de neurones sont explorés afin de comparer le comportement et les performances de ces modèles. On cherche à contraindre ces modèles pour les rendre interprétables en terme de processus hydrodynamiques en rapprochant le fonctionnement du modèle à celui du système naturel afin d'obtenir une bonne représentation et d'extraire des connaissances à partir des paramètres du modèle.Les résultats obtenus par les analyses corrélatoires et spectrales permettent d'orienter la configuration des modèles de réseaux de neurones. Appliqués à l'hydrosystème du Lez sur la période 1950-1967, les résultats montrent que les réseaux de neurones sont à même de modéliser les hydrosystèmes au fonctionnement non-linéaires. L'utilisation de deux hydrosystèmes variant dans le temps (la Durance en France et Fernow aux USA) tend à souligner la capacité des réseaux de neurones à modéliser efficacement les systèmes non stationnaires. Des méthodes d'ajustement en temps réel (adaptativité et assimilation de données) permettent d'accroître les performances des modèles par apprentissage statistique face à des modifications des entrées ou du système lui-même.Finalement, ces différentes méthodes d'analyse et de modélisation permettent d'améliorer la connaissance de la relation pluie-débit. Les outils méthodologiques réalisés dans cette thèse ont pu être développés à partir de l'application à l'hydrosystème du Lez dont le fonctionnement est étudié depuis des décennies. Cette méthodologie d'étude et de modélisation présente l'avantage d'être transposable à d'autres systèmes. / Improving knowledge of karst hydrodynamics represents a global challenge for water resource because karst aquifers provide approximately 25% of the world population in fresh water. Nevertheless, complexity, anisotropy, heterogeneity, non-linearity and possible non-stationarity of these aquifers makes them underexploited objects due to the difficulty to characterize their morphology and hydrodynamics. In this context, the systemic paradigm proposes others methods by studying these hydrosystems through input-output (rainfall-runoff) relations.This work covers the use of: i) correlation and spectral analysis to characterize response of karst aquifers, ii) neural networks to study and model linear and non-linear relations of these hydrosystems. In order to achieve this, different types of neural networks model configurations are explored to compare behavior and performances of these models. We are looking to constrain these models to make them interpretable in terms of hydrodynamic processes by making the operation of the model closer to the natural system in order to obtain a good representation and extract knowledge from the model parameters.The results obtained by correlation and spectral analysis are used to manage the configuration of neural networks models. Applied on the Lez hydrosystem over the period 1950-1967, results show that neural networks models are capable to model non-linear operation of the karst.Application of neural modelling on two non stationary hydrosystems (Durance in France and Fernow in the the USA) proved the ability of neural networks to model satisfactorily non-stationary conditions. Moreover, two real-time adjustment methods (adaptativity and data assimilation) enhanced the performance of neural network models face to changing conditions of the inputs or of the system itself.Finally, these various methods to analyze and model allow improving knowledge of the rainfall-runoff relationship. Methodological tools developed in this thesis were developed thanks to the application on Lez hydrosystem which has been studied for decades. This study and modeling methodology have the advantage of being applicable to other systems provided the availability of a sufficient database.

This study focuses on improving the quality of statistical speech synthesis based on sinusoidal models. Vocoders play a crucial role during the parametrisation and reconstruction process, so we first lead an experimental comparison of a broad range of the leading vocoder types. Although our study shows that for analysis / synthesis, sinusoidal models with complex amplitudes can generate high quality of speech compared with source-filter ones, component sinusoids are correlated with each other, and the number of parameters is also high and varies in each frame, which constrains its application for statistical speech synthesis. Therefore, we first propose a perceptually based dynamic sinusoidal model (PDM) to decrease and fix the number of components typically used in the standard sinusoidal model. Then, in order to apply the proposed vocoder with an HMM-based speech synthesis system (HTS), two strategies for modelling sinusoidal parameters have been compared. In the first method (DIR parameterisation), features extracted from the fixed- and low-dimensional PDM are statistically modelled directly. In the second method (INT parameterisation), we convert both static amplitude and dynamic slope from all the harmonics of a signal, which we term the Harmonic Dynamic Model (HDM), to intermediate parameters (regularised cepstral coefficients (RDC)) for modelling. Our results show that HDM with intermediate parameters can generate comparable quality to STRAIGHT. As correlations between features in the dynamic model cannot be modelled satisfactorily by a typical HMM-based system with diagonal covariance, we have applied and tested a deep neural network (DNN) for modelling features from these two methods. To fully exploit DNN capabilities, we investigate ways to combine INT and DIR at the level of both DNN modelling and waveform generation. For DNN training, we propose to use multi-task learning to model cepstra (from INT) and log amplitudes (from DIR) as primary and secondary tasks. We conclude from our results that sinusoidal models are indeed highly suited for statistical parametric synthesis. The proposed method outperforms the state-of-the-art STRAIGHT-based equivalent when used in conjunction with DNNs. To further improve the voice quality, phase features generated from the proposed vocoder also need to be parameterised and integrated into statistical modelling. Here, an alternative statistical model referred to as the complex-valued neural network (CVNN), which treats complex coefficients as a whole, is proposed to model complex amplitude explicitly. A complex-valued back-propagation algorithm using a logarithmic minimisation criterion which includes both amplitude and phase errors is used as a learning rule. Three parameterisation methods are studied for mapping text to acoustic features: RDC / real-valued log amplitude, complex-valued amplitude with minimum phase and complex-valued amplitude with mixed phase. Our results show the potential of using CVNNs for modelling both real and complex-valued acoustic features. Overall, this thesis has established competitive alternative vocoders for speech parametrisation and reconstruction. The utilisation of proposed vocoders on various acoustic models (HMM / DNN / CVNN) clearly demonstrates that it is compelling to apply them for the parametric statistical speech synthesis.

ABSTRACT OF THE DISSERTATION OF Yonglian Wang, for Doctor of Philosophy degree in Electrical and Computer Engineering, presented on May 19, 2009, at Southern Illinois University- Carbondale. TITLE: SPEECH RECOGNITION UNDER STRESS MAJOR PROFESSOR: Dr. Nazeih M. Botros In this dissertation, three techniques, Dynamic Time Warping (DTW), Hidden Markov Models (HMM), and Hidden Control Neural Network (HCNN) are utilized to realize talker-independent isolated word recognition. DTW is a technique utilized to measure the distance between two input patterns or vectors; HMM is a tool utilized to model speech signals using stochastic process in five states to compare the similarity between signals; and HCNN calculates the errors between actual output and target output and it is mainly built for the stress compensated speech recognition. When stress (Angry, Question and Soft) is induced into the normal talking speech, speech recognition performance degrades greatly. Therefore hypothesis driven approach, a stress compensation technique is introduced to cancel the distortion caused by stress. The database for this research is SUSAS (Speech under Simulated and Actual Stress) which includes five domains encompassing a wide variety of stress, 16,000 isolated-word speech signal samples available from 44 speakers. Another database, called TIMIT (10 speakers and 6300 sentences in total) is used as a minor in DTW algorithm. The words used for speech recognition are speaker-independent. The characteristic feature analysis has been carried out in three domains: pitch, intensity， and glottal spectrum. The results showed that speech spoken under angry and question stress indicates extremely wide fluctuations with average higher pitch, higher RMS intensity, and more energy compared to neutral. In contrast, the soft talking style has lower pitch, lower RMS intensity, and less energy compared to neutral. The Linear Predictive Coding (LPC) cepstral feature analysis is used to obtain the observation vector and the input vector for DTW, HMM, and stress compensation. Both HMM and HCNN consist of training and recognition stages. Training stage is to form references, while recognition stage is to compare an unknown word against all the reference models. The unknown word is recognized by the model with highest similarity. Our results showed that HMM technique can achieve 91% recognition rate for Normal speech; however, the recognition rate dropped to 60% for Angry stress condition, 65% for Question stress condition, and 76% for Soft stress condition. After compensation was applied for the cepstral tilts, the recognition rate increased by 10% for Angry stress condition, 8% for Question stress condition, and 4% for Soft stress condition. Finally, HCNN technique increased the recognition rate to 90% for Angry stress condition and it also differentiated the Angry stress from other stress group.

Visual saliency computation is about detecting and understanding salient regions and elements in a visual scene. Algorithms for visual saliency computation can give clues to where people will look in images, what objects are visually prominent in a scene, etc. Such algorithms could be useful in a wide range of applications in computer vision and graphics. In this thesis, we study the following visual saliency computation problems. 1) Eye Fixation Prediction. Eye fixation prediction aims to predict where people look in a visual scene. For this problem, we propose a Boolean Map Saliency (BMS) model which leverages the global surroundedness cue using a Boolean map representation. We draw a theoretic connection between BMS and the Minimum Barrier Distance (MBD) transform to provide insight into our algorithm. Experiment results show that BMS compares favorably with state-of-the-art methods on seven benchmark datasets. 2) Salient Region Detection. Salient region detection entails computing a saliency map that highlights the regions of dominant objects in a scene. We propose a salient region detection method based on the Minimum Barrier Distance (MBD) transform. We present a fast approximate MBD transform algorithm with an error bound analysis. Powered by this fast MBD transform algorithm, our method can run at about 80 FPS and achieve state-of-the-art performance on four benchmark datasets. 3) Salient Object Detection. Salient object detection targets at localizing each salient object instance in an image. We propose a method using a Convolutional Neural Network (CNN) model for proposal generation and a novel subset optimization formulation for bounding box filtering. In experiments, our subset optimization formulation consistently outperforms heuristic bounding box filtering baselines, such as Non-maximum Suppression, and our method substantially outperforms previous methods on three challenging datasets. 4) Salient Object Subitizing. We propose a new visual saliency computation task, called Salient Object Subitizing, which is to predict the existence and the number of salient objects in an image using holistic cues. To this end, we present an image dataset of about 14K everyday images which are annotated using an online crowdsourcing marketplace. We show that an end-to-end trained CNN subitizing model can achieve promising performance without requiring any localization process. A method is proposed to further improve the training of the CNN subitizing model by leveraging synthetic images. 5) Top-down Saliency Detection. Unlike the aforementioned tasks, top-down saliency detection entails generating task-specific saliency maps. We propose a weakly supervised top-down saliency detection approach by modeling the top-down attention of a CNN image classifier. We propose Excitation Backprop and the concept of contrastive attention to generate highly discriminative top-down saliency maps. Our top-down saliency detection method achieves superior performance in weakly supervised localization tasks on challenging datasets. The usefulness of our method is further validated in the text-to-region association task, where our method provides state-of-the-art performance using only weakly labeled web images for training.

Cette thèse s’inscrit dans le domaine de la navigation robotique bio-inspirée en environnement réel et implique la capacité pour un robot mobile à se déplacer de manière autonome dans un monde a priori dynamique et inconnu. Les travaux décrits au cours de ce manuscrit s’attacheront à montrer comment, en partant des travaux académiques réalisés par l’équipe Neurocybernétique du laboratoire ETIS, il a été possible de concevoir le robot mobile Diya One capable de naviguer de manière autonome dans de grands environnements intérieurs. Depuis une vingtaine d’année, l’équipe Neurocybernétique élabore des modèles de navigation bio-inspirée. De précédents travaux ont montré qu’un modèle de cellules de lieu, enregistrées chez le rat, permet à un robot mobile d’apprendre des comportements de navigation robustes, tels qu’une ronde ou un retour au nid, à partir d’associations entre lieu et action. L’apprentissage et la reconnaissance d’un lieu ne reposent alors que sur des informations visuelles. Cependant, trois problèmes critiques ne permettent pas de naviguer dans de grands environnements : 1- l’ambiguïté de certaines situations visuelles (ou alias perceptif), 2- l’apprentissage sur le long terme et 3- la sensibilité aux conditions environnementales. L’ajout d’autres modalités constitue une solution efficace pour augmenter la robustesse de la localisation. L’équipe a développé plusieurs modèles basés sur la proprioception du robot afin desuppléer, dans les cas limites, les modèles purement visuels. La principale limitation des approches proprioceptives est qu’elles sont soumisesà l’accumulation d’erreurs. Il est donc nécessaire de recalibrer périodiquement les modèles. Fusionner des modalités allothétiques et idiothétiques semblent être une bonne stratégie pour obtenir une estimation fiable de la localisation du robot. Les champs de neurones dynamiques (DNF) ou continous attractor neural network (CANN) constituent un puissant candidat pour mettre en œuvre le type de mémoire requis pour la construction de cellules de lieu. Nous présenterons un premier modèle de fusion utilisant les champs de neurones dynamiques pour maintenir l’orientation du robot puis un second généralisant le principe du modèle de fusion pour la construction de cellules de lieux multimodales.Être capable de produire et commercialiser rapidement un produit est un enjeu majeur pour la survie de Partnering. En plus des capacités de navigation et de localisation, un robot commercialisable requiert un ensemble de comportements indispensables à la mobilité, à la sécurité (loi de contrôle, évitement des obstacles et des trous) et à son autonomie (gestion d’énergie et retour à la station de recharge). Pour aboutir à cette première solution, nous avons suivit une démarche ascendante (bottom-up) défendue par la robotique comportementale. Nous avons développé progressivement la complexité du robot au travers de comportements élémentaires intégrés dans une architecture de contrôle régissant à chaque instant l’expression de ces comportements et la sélection des actions à exécuter.Ce mémoire est découpée en deux parties. Une première partie industrielle relevant d’objectifs à court terme, consistant à mettre en place, à partir des modèles existants développés par l’équipe Neurocybernétique, l’architecture comportementale de la première version du robot Diya One. Puis, une seconde partie plus fondamentale dans laquelle nous traiterons de la réalisation demodèles de fusion multimodale. Ces modèles seront ajoutés incrémentalement au robot afin d’améliorer progressivement ses capacités de navigation. / This thesis falls into the field of navigation in bio-inspired robotics in real environment and implies the ability for a mobile robot to move autonomously in a world a priori dynamic and unknown. The work described in this manuscript will show how, starting from the academic work carried out by the Neurocybernetics team of the ETIS laboratory, it was possible to design the mobile robot Diya One able to navigate autonomously in large indoor environments. For the past 20 years, the Neurocybernetics team has been developing bio-inspired navigation models. Previous work has shown that a model of place cells, recorded in the rat, allows a mobile robot to learn robust navigation behaviors, such as a round or a homing, from associations between place and action. Learning and the recognition of a place are based only on visual information. However, three critical problems do not allow to navigate in large environments: 1- the ambiguity of certain visual situations (or perceptual alia), 2- long-term learning, and 3-sensitivity to environmental conditions. The addition of other modalities is an effective solution for increasing the robustness of the location. The team has developed several models based on the proprioception of the robot in order to compensate, in limiting cases, for purely visual models. The main limitation of proprioceptive approaches is that, based on the proprioception of the robot, they are subject to the accumulation of errors. It is therefore necessary to periodically recalibrate the models. Merging allothetic and idiothetic modalities seems to be a good strategy for obtaining a reliable estimate of the robot’s location. Dynamic neural network (DNF) or continous attractor neural network (CANN) constitute a powerful candidate to implement the type of memory required for the construction of place cells. We present a first fusion model using dynamic neuron fields to maintain the orientation of the robot and then a second one generalizing the principle of fusion model for the construction of cells of multimodal places. Being able to produce and market quickly a product is a major challenge for Partnering’s survival. In addition to navigation and location capabilities, a marketable robot requires a set of behaviors that are essential to mobility, security (control law, avoidance of obstacles and holes) and its autonomy (energy management and return to the charging station). To arrive at this first solution, we followed a bottom-up approach defended by behavioral robotics. We have progressively developed the complexity of the robot through elementary behaviors integratedin a control architecture governing at each moment the expression of these behaviors and the selection of the actions to be executed. This manuscript is divided into two parts. A first industrial part with short-term objectives, consisting of implementing the behavioral architecture of the first version of the Diya One robot, based on the existing models developed by the Neurocybernetics team. Then, a second more theoretical part in which we will deal with the realization of multimodal fusion models. These models will be added incrementally to the robot in order to gradually improve its navigationcapabilities.

A number of researchers have shown that ear recognition is a viable alternative to more common biometrics such as fingerprint, face and iris because the ear is relatively stable over time, the ear is non-invasive to capture, the ear is expressionless, and both the ear’s geometry and shape have significant variation among individuals. Researchers have used different approaches to enhance ear recognition. Some researchers have improved upon existing algorithms, some have developed algorithms from scratch to assist with recognizing individuals by ears, and some researchers have taken algorithms tried and tested for another purpose, i.e., face recognition, and applied them to ear recognition. These approaches have resulted in a number of state-of-the-art effective methods to identify individuals by ears. However, most ear recognition research has been done using ear images that were captured in an ideal setting: ear images have near perfect lighting for image quality, ears are in the same position for each subject, and ears are without earrings, hair occlusions, or anything else that could block viewing of the entire ear. In order for ear recognition to be practical, current approaches must be improved. Ear recognition must move beyond ideal settings and demonstrate effectiveness in an unconstrained environment reflective of real world conditions. Ear recognition approaches must be scalable to handle large groups of people. And, ear recognition should demonstrate effectiveness across a diverse population. This dissertation advances ear recognition from ideal settings to real world settings. We devised an ear recognition framework that outperformed state-of-the-art recognition approaches using the most challenging sets of publicly available ear images and the most voluminous set of unconstrained ear images that we are aware of. We developed a Convolutional Neural Network-based solution for ear normalization and description, we designed a two-stage landmark detector, and we fused learned and handcrafted descriptors. Using our framework, we identified some individuals that are wearing earrings and that have other occlusions, such as hair. The results suggest that our framework can be a gateway for identification of individuals in real world conditions.