Orientador: Flávio Ferrari Aragon / Banca: Adriano Wagner Ballarin / Banca: Luís Fernando Nicolosi Bravin / Banca: Rui Vieira de Moraes / Banca: Sandra Fiorelli de Almeida P. Simeão / Resumo: Nas Ciências Agronômicas, encontram-se várias situações em que são observadas diversas variáveis respostas nas parcelas ou unidades experimentais. Nestas situações, um caso de interesse prático à experimentação agronômica é o que considera a construção de regiões de similaridade entre as parcelas para a discriminação entre os grupos experimentais e ou para a classificação de novas unidades experimentais em uma dessas regiões. Os métodos de classificação ou discriminação exigem, para sua utilização prática, uma quantidade considerável de retenção de informação da estrutura de variabilidade dos dados e, principalmente, alta fidedignidade e competência nas alocações de novos indivíduos nos grupos, mostradas nas distribuições corretas destes indivíduos. Existem vários procedimentos para medir o grau de decisão correta (acurácia) das informações fornecidas pelos métodos classificatórios. Praticamente, a totalidade deles utilizam a probabilidade de classificação errônea como o indicador de qualidade, sendo alguns destes freqüentistas (probabilidade estimada pela freqüência relativa de ocorrências - métodos não paramétricos) e outros baseados nas funções densidade de probabilidade das populações (métodos paramétricos). A principal diferença entre esses procedimentos é a conceituação dada ao cálculo da probabilidade de classificação errônea. Pretende-se, no presente estudo, apresentar alguns procedimentos para estimar estas probabilidades, desenvolver um software para a obtenção das estimativas considerando a distância generalizada de Mahalanobis como o procedimento relativo à da função densidade de probabilidade para populações com distribuição multinormal . Este software será de acesso livre e de fácil manuseio para pesquisadores de áreas aplicadas, completado com o manual do usuário e com um exemplo de aplicação envolvendo divergência genética de girassol. / Abstract: In the Agronomical Sciences, mainly in studies involving biomass production and rational use of energy, there are several situations in which several variable answers in the parts or experimental units are observed. In these situations, a case of practical interest to the agronomical experimentation is that one which considers the construction of similarity regions among parts and or the classification of new experimental units. The classification methods demand, for their utilization, a considerable quantity for utilization of their information retention of data and, mostly, high fidelity and competence in the new individual allocations. There are several procedures to measure accuracy degree of the information supplied by the discrimination method. Practically all of them use the miss-classification probability (erroneous classification) like the quality indicator. The main difference among these evaluation methods is the characterization of the miss-classification probability. Therefore, the aim is to present some estimate procedures of the missclassification probabilities involving repetition frequency and distribution methods and to develop a software to obtain their estimate, which is accessible and easy handling for researchers of applied areas, complementing the study with user's manual and examples in the rational energy application and biomass energy. / Doutor

In this thesis, we explore two different approaches to causal inferences. The traditional approach models the theoretical relationship between the outcome variables and their explanatory variables, i.e., the science, at the same time as the systematic differences between treated and control subjects are modeled, i.e., the assignment mechanism. The alternative approach, based on Rubin's Causal Model (RCM), makes it possible to model the science and the assignment mechanism separately in a two-step procedure. In the first step, no outcome variables are used when the assignment mechanism is modeled, the treated students are matched with similar control students using this mechanism, and the models for the science are determined. Outcome variables are only used in the second step when these pre-specified models for the science are fitted. In the first paper, we use the traditional approach to evaluate whether a husband is more prone to quit smoking when his wife quits smoking than he would have been had his wife not quit. We find evidence that this is the case, but that our analysis must rely on restrictive assumptions. In the subsequent two papers, we use the alternative RCM approach to evaluate if a Harvard freshman who does not smoke (observed potential outcome) is more prone to start smoking when he shares a suite with at least one smoker, than he would have been had he shared a suite with only smokers (missing potential outcomes). We do not find evidence that this is the case, and the small and insignificant treatment effect is robust against various assumptions that we make regarding covariate adjustments and missing potential outcomes. In contrast, we do find such evidence when we use the traditional approach previously used in the literature to evaluate peer effects relating to smoking, but the treatment effect is not robust against the assumptions that we make regarding covariate adjustments. These contrasting results in the two latter papers allow us to conclude that there are a number of advantages with the alternative RCM approach over the traditional approaches previously used to evaluate peer effects relating to smoking. Because the RCM does not use the outcome variables when the assignment mechanism is modeled, it can be re-fit repeatedly without biasing the models for the science. The assignment mechanism can then often be modeled to fit the data better and, because the models for the science can consequently better control for the assignment mechanism, they can be fit with less restrictive assumptions. Moreover, because the RCM models two distinct processes separately, the implications of the assumptions that are made on these processes become more transparent. Finally, the RCM can derive the two potential outcomes needed for drawing causal inferences explicitly, which enhances the transparency of the assumptions made with regard to the missing potential outcomes. / Diss. Stockholm : Handelshögskolan, 2006 S. 1-13: sammanfattning, s. [15]-161: 4 uppsatser

Submitted by Automa??o e Estat?stica (sst@bczm.ufrn.br) on 2017-11-06T19:32:32Z No. of bitstreams: 1 ElizabethVivianaCabreraAvila_DISSERT.pdf: 25950680 bytes, checksum: d7419fec7a225d199d438d7c2c6390dd (MD5) / Approved for entry into archive by Arlan Eloi Leite Silva (eloihistoriador@yahoo.com.br) on 2017-11-20T19:57:50Z (GMT) No. of bitstreams: 1 ElizabethVivianaCabreraAvila_DISSERT.pdf: 25950680 bytes, checksum: d7419fec7a225d199d438d7c2c6390dd (MD5) / Made available in DSpace on 2017-11-20T19:57:50Z (GMT). No. of bitstreams: 1 ElizabethVivianaCabreraAvila_DISSERT.pdf: 25950680 bytes, checksum: d7419fec7a225d199d438d7c2c6390dd (MD5) Previous issue date: 2017-07-05 / Coordena??o de Aperfei?oamento de Pessoal de N?vel Superior (CAPES) / Dentro da ?rea de Mecatr?nica, principalmente em CAD (Computer Aided Desing) e Vis?o Rob?tica, s?o desenvolvidas uma s?rie de aplica??es que necessitam da an?lise de objetos n?o r?gidos ou deform?veis atrav?s de representa??es computacionais dos mesmos. Esta pesquisa mostra como medir similaridade de objetos deform?veis utilizando a representa??o atrav?s de nuvens de pontos tridimensionais. Basicamente, s?o consideradas tr?s nuvens de pontos do objeto analisado: uma sem altera??es, outra que representa o grau de m?xima deforma??o e uma terceira que descreve a deforma??o de interesse, no momento de execu??o de alguma aplica??o. Estudamos dois m?todos alternativos para medir similaridade baseadas em medidas de dist?ncias, com as respectivas verifica??es da precis?o e tempos. O primeiro m?todo ? baseado no c?lculo da dist?ncia de Mahalanobis e, no segundo, ? usada a dist?ncia de Hausdorff ap?s uma etapa pr?via de registro e alinhamento dos dados. Foram realizados experimentos e an?lises considerando algumas partes do corpo humano, onde se evidencia que a dist?ncia de Mahalanobis tem o menor tempo de execu??o, sendo fact?vel em tempo real. V?rias aplica??es nas ?reas acima mencionadas podem se basear nos resultados obtidos nesta disserta??o para determinar os n?veis de deforma??o de objetos deform?veis restritos. / Within the area of Mechatronics, mainly in CAD (Computer Aided Desing) and Robotic Vision, many applications are developed that require the analysis of non-rigid or deformable objects through computational representations of them. This master thesis proposes an approach to measure similarity of deformable objects using three-dimensional points clouds of them. Basically, three point clouds of the analyzed object are considered: one without changes, another representing the degree of maximum deformation and a third that describes the deformation of interest, at the time of application execution. Here are presented two alternatives to measure similarity based on Distance measures, with the respective accuracy and time checks. The first method is based on the Mahalanobis distance computation and, in the second, the Hausdorff distance is used after a registration and alignment steps of the data. The experiments are developed considering some parts of the human body, its evidents that the analysis with Mahalanobis distance has the shortest execution time, being feasible in real time. Several applications in the above mentioned areas can be based on the results obtained in this dissertation to determine the deformation levels of restricted deformable objects.

In this work, we propose a novel method for a supervised dimensionality reduc- tion, which learns weights of a neural network using an evolutionary algorithm, CMA-ES, optimising the success rate of the k-NN classifier. If no activation func- tions are used in the neural network, the algorithm essentially performs a linear transformation, which can also be used inside of the Mahalanobis distance. There- fore our method can be considered to be a metric learning algorithm. By adding activations to the neural network, the algorithm can learn non-linear transfor- mations as well. We consider reductions to low-dimensional spaces, which are useful for data visualisation, and demonstrate that the resulting projections pro- vide better performance than other dimensionality reduction techniques and also that the visualisations provide better distinctions between the classes in the data thanks to the locality of the k-NN classifier. 1

Intelligent fault detection systems using machine learning can be applied to learn to spot anomalies in signals sampled directly from machinery. As a result, expensive repair costs due to mechanical breakdowns and potential harm to humans due to malfunctioning equipment can be prevented. In recent years, Autoencoders have been applied for fault detection in areas such as industrial manufacturing. It has been shown that they are well suited for the purpose as such models can learn to recognize healthy signals that facilitate the detection of anomalies. The content of this thesis is an investigation into the applicability of Autoencoders for fault detection in mobile robotics by assigning anomaly scores to sampled torque signals based on the Autoencoder reconstruction errors and the Mahalanobis distance to a known distribution of healthy errors. An experiment was carried out by training a model with signals recorded from a four-wheeled mobile robot executing a pre-defined diagnostics routine to stress the motors, and datasets of healthy samples along with three different injected faults were created. The model produced overall greater anomaly scores for one of the fault cases in comparison to the healthy data. However, the two other cases did not yield any difference in anomaly scores due to the faults not impacting the pattern of the signals. Additionally, the Autoencoders ability to isolate a fault to a location was studied by examining the reconstruction errors faulty samples determine whether the errors of signals originating from the faulty component could be used for this purpose. Although we could not confirm this based on the results, fault isolation with Autoencoders could still be possible given more representative signals.

Les analyses effectuées dans le cadre de ce mémoire ont été réalisées à l'aide du module MatchIt disponible sous l’environnent d'analyse statistique R. / Statistical analyzes of this thesis were performed using the MatchIt package available in the statistical analysis environment R. / L’estimation sans biais de l’effet causal d’une intervention nécessite la comparaison de deux groupes homogènes. Il est rare qu’une étude observationnelle dispose de groupes comparables et même une étude expérimentale peut se retrouver avec des groupes non comparables. Les chercheurs ont alors recours à des techniques de correction afin de rendre les deux groupes aussi semblables que possible. Le problème consiste alors à choisir la méthode de correction appropriée. En ce qui nous concerne, nous limiterons nos recherches à une famille de méthodes dites d’appariement. Il est reconnu que ce qui importe lors d’un appariement est l’équilibre des deux groupes sur les caractéristiques retenues. Autrement dit, il faut que les variables soient distribuées de façon similaire dans les deux groupes. Avant même de considérer la distribution des variables entre les deux groupes, il est nécessaire de savoir si les données en question permettent une inférence causale. Afin de présenter le problème de façon rigoureuse, le modèle causal contrefactuel sera exposé. Par la suite, les propriétés formelles de trois méthodes d’appariement seront présentées. Ces méthodes sont l’appariement par la distance de Mahalanobis, de l’appariement par le score de propension et de l’appariement génétique. Le choix de la technique d’appariement appropriée reposera sur quatre critères empiriques dont le plus important est la différence des moyennes standardisées. Les résultats obtenus à l’aide des données de l’Enquête longitudinale et expérimentale de Montréal (ÉLEM) indiquent que des trois techniques d’appariement, l’appariement génétique est celui qui équilibre mieux les variables entre les groupes sur tous les critères retenus. L’estimation de l’effet de l’intervention varie sensiblement d’une technique à l’autre, bien que dans tous les cas cet effet est non significatif. Ainsi, le choix d’une technique d’appariement influence l’estimation de l’effet d’une intervention. Il est donc impérieux de choisir la technique qui permet d’obtenir un équilibre optimal des variables selon les données à la disposition du chercheur. / The unbiased estimate of the causal effect of an intervention requires the comparison of two homogeneous groups. It is rare that an observational study has comparable groups and even an experiment may end up with non-comparable groups. The researchers then used correction techniques to make the two groups as similar as possible. The problem then is to choose the appropriate correction method. In our case, we will restrict our research to a family of so-called matching methods. It is recognized that what matters in a match is the balance between the two groups on selected characteristics. In other words, it is necessary that the variables are distributed similarly in both groups. Even before considering the distribution of variables between the two groups, it is necessary to know whether the data in question allow for causal inference. To present the problem rigorously, the counterfactual causal model will be exposed. Thereafter, the formal properties of three matching methods will be presented. Those methods are the Mahalanobis matching, the propensity score matching and genetic matching. The choice of the appropriate matching technique is based on four empirical criteria which the most important is the standardized mean difference. Results obtained using data from the Montréal Longitudinal and Experimental Study indicate that of the three matching techniques, genetic matching is the one that better balance the variables between groups on all criteria. The estimate of the effect of intervention varies substantially from one technique to another, although in all cases this effect is non significant. Thus, the selection of a matching technique influences the estimation of the effect of an intervention. Therefore, it is imperative to choose the technique that provides an optimal balance of the variables based on data available to the researcher.

本研究採用Lee 和Poon 所提出的隱藏常態變數模型來估計混合連續與間斷型變數之參數估計，並估計其馬式距離。此外，並利用穩健估計來估計混合型資料參數及其馬式距離，可在有離群值時解決最大蓋似估計的不穩定。 / Poon and Lee (1987) applied normal latent variable model to deal with the parameters estimation for the data mixed with continuous and discrete variables and Bedrick et al. (2000) used this idea to evaluate the Mahalanobis distance. In this thesis, we extend a similar idea to robustly estimate Multivariate Data Mixed with Continuous and Discrete Variables with the same model. Furthermore, we evaluate the Mahalanobis distance which can determine similarity of variables. The proposed method can overcome the unreliability of MLE while there exist outliers in the data.

QBEST, a novel statistical method, can be applied to the problem of estimating the No Observed Adverse Effect Level (NOAEL or QNOAEL) of a New Molecular Entity (NME) in order to anticipate a safe starting dose for beginning clinical trials. The NOAEL from QBEST (called the QNOAEL) can be calculated using multiple disparate studies in the literature and/or from the lab. The QNOAEL is similar in some ways to the Benchmark Dose Method (BMD) used widely in toxicological research, but is superior to the BMD in some ways. The QNOAEL simulation generates an intuitive curve that is comparable to the dose-response curve. The NOAEL of ellagic acid (EA) is calculated for clinical trials as a component therapeutic agent (in BSN476) for treating Chikungunya infections. Results are used in a simulation based on nonparametric cluster analysis methods to calculate confidence levels on the difference between the Effect and the No Effect studies. In order to evaluate the statistical power of the algorithm, simulated data clusters with known parameters are fed into the algorithm in a separate study, testing the algorithm’s accuracy and precision “Around the Compass Rose” at known coordinates along the circumference of a multidimensional data cluster. The specific aims of the proposed study are to evaluate the accuracy and precision of the QBEST Simulation and QNOAEL compared to the Benchmark Dose Method, and to calculate the QNOAEL of EA for BSN476 Drug Development.

由績效觀點，品質(適質)與成本(適量)，在概念上是完全一致的。因此，績效的管理，應以品質與成本作為其目標達成與否的衡量標準。本研究以績效觀點來解決公司面臨到品質與成本的兩難的問題。隨著電子產品的功能多樣化，發熱問題卻接踵而來，發熱密度的不斷提昇，對於散熱設計的需求也越來越受到重視。本研究以電源轉換器為對象，其目前已設計完成且已通過美國UL安規認證，但因為其溫升及其變異很大，因此降低電源轉換器的溫升及其變異是一急需解決的問題，以期能找出穩健於不可控因子，使產品變異小且各零件溫升與損失均能降至最低的最適外部零件組合。透過了田口與實驗設計的方法規劃及進行實驗並收集數據。引用加權SN比(multi-response S/N ratio)的方法，分別透過(1)管制圖法及(2)資料包絡法的CCR保證領域法(指CCR-AR模型)來決定加權SN比的權數，以決定可控因子及其水準值。對矩陣實驗的數據利用MTS ( M a h a l o n o b I s － Taguchi System)來篩選研究問題中較重要的特性變數，再針對篩選結果中較重要的特性變數的數據分別利用(1)倒傳遞類神經網路結合資料包絡法及(2)資料包絡法結合主成份分析法兩種分析方法，得到外殼鑽孔形狀與矽膠片大小的最佳因子組合。由改善後的確認實驗結果得知，雖然平均溫升下降的程度不大，然而大部份量測點的溫升標準差都顯著變小了，因此本研究在降低該電源轉換器溫升變異的效果顯著。

Cluster and discriminant analysis belong to basic classification methods. Using cluster analysis can be a disordered group of objects organized into several internally homogeneous classes or clusters. Discriminant analysis creates knowledge based on the jurisdiction of existing classes classification rule, which can be then used for classifying units with an unknown group membership. The aim of this thesis is a comparison of discriminant analysis and different methods of cluster analysis. To reflect the distances between objects within each cluster, squeared Euclidean and Mahalanobis distances are used. In total, there are 28 datasets analyzed in this thesis. In case of leaving correlated variables in the set and applying squared Euclidean distance, Ward´s method classified objects into clusters the most successfully (42,0 %). After changing metrics on the Mahalanobis distance, the most successful method has become the furthest neighbor method (37,5 %). After removing highly correlated variables and applying methods with Euclidean metric, Ward´s method was again the most successful in classification of objects (42,0%). From the result implies that cluster analysis is more precise when excluding correlated variables than when leaving them in a dataset. The average result of discriminant analysis for data with correlated variables and also without correlated variables is 88,7 %.