• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 90
  • 24
  • 4
  • 3
  • 3
  • 2
  • 2
  • 2
  • 2
  • 1
  • 1
  • 1
  • Tagged with
  • 151
  • 151
  • 138
  • 45
  • 29
  • 26
  • 26
  • 23
  • 22
  • 20
  • 20
  • 17
  • 16
  • 15
  • 15
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
121

Uma abordagem estatística para o modelo do preço spot da energia elétrica no submercado sudeste/centro-oeste brasileiro / A statistical approach to model the spot price of electric energy: evidende from brazilian southeas/middle-west subsystem.

Guilherme Matiussi Ramalho 20 March 2014 (has links)
O objetivo deste trabalho e o desenvolvimento de uma ferramenta estatistica que sirva de base para o estudo do preco spot da energia eletrica do subsistema Sudeste/Centro-Oeste do Sistema Interligado Nacional, utilizando a estimacao por regressao linear e teste de razao de verossimilhanca como instrumentos para desenvolvimento e avaliacao dos modelos. Na analise dos resultados estatsticos descritivos dos modelos, diferentemente do que e observado na literatura, a primeira conclusao e a verificacao de que as variaveis sazonais, quando analisadas isoladamente, apresentam resultados pouco aderentes ao preco spot PLD. Apos a analise da componente sazonal e verificada a influencia da energia fornecida e a energia demandada como variaveis de entrada, com o qual conclui-se que especificamente a energia armazenada e producao de energia termeletrica sao as variaveis que mais influenciam os precos spot no subsistema estudado. Entre os modelos testados, o que particularmente ofereceu os melhores resultados foi um modelo misto criado a partir da escolha das melhores variaveis de entrada dos modelos testados preliminarmente, alcancando um coeficiente de determinacao R2 de 0.825, resultado esse que pode ser considerado aderente ao preco spot. No ultimo capitulo e apresentada uma introducao ao modelo de predicao do preco spot, possibilitando dessa forma a analise do comportamento do preco a partir da alteracao das variaveis de entrada. / The objective of this work is the development of a statistical method to study the spot prices of the electrical energy of the Southeast/Middle-West (SE-CO) subsystem of the The Brazilian National Connected System, using the Least Squares Estimation and Likelihood Ratio Test as tools to perform and evaluate the models. Verifying the descriptive statistical results of the models, differently from what is observed in the literature, the first observation is that the seasonal component, when analyzed alone, presented results loosely adherent to the spot price PLD. It is then evaluated the influence of the energy supply and the energy demand as input variables, verifying that specifically the stored water and the thermoelectric power production are the variables that the most influence the spot prices in the studied subsystem. Among the models, the one that offered the best result was a mixed model created from the selection of the best input variables of the preliminarily tested models, achieving a coeficient of determination R2 of 0.825, a result that can be considered adherent to the spot price. At the last part of the work It is presented an introduction to the spot price prediction model, allowing the analysis of the price behavior by the changing of the input variables.
122

Estimação e teste de hipótese baseados em verossimilhanças perfiladas / "Point estimation and hypothesis test based on profile likelihoods"

Michel Ferreira da Silva 20 May 2005 (has links)
Tratar a função de verossimilhança perfilada como uma verossimilhança genuína pode levar a alguns problemas, como, por exemplo, inconsistência e ineficiência dos estimadores de máxima verossimilhança. Outro problema comum refere-se à aproximação usual da distribuição da estatística da razão de verossimilhanças pela distribuição qui-quadrado, que, dependendo da quantidade de parâmetros de perturbação, pode ser muito pobre. Desta forma, torna-se importante obter ajustes para tal função. Vários pesquisadores, incluindo Barndorff-Nielsen (1983,1994), Cox e Reid (1987,1992), McCullagh e Tibshirani (1990) e Stern (1997), propuseram modificações à função de verossimilhança perfilada. Tais ajustes consistem na incorporação de um termo à verossimilhança perfilada anteriormente à estimação e têm o efeito de diminuir os vieses da função escore e da informação. Este trabalho faz uma revisão desses ajustes e das aproximações para o ajuste de Barndorff-Nielsen (1983,1994) descritas em Severini (2000a). São apresentadas suas derivações, bem como suas propriedades. Para ilustrar suas aplicações, são derivados tais ajustes no contexto da família exponencial biparamétrica. Resultados de simulações de Monte Carlo são apresentados a fim de avaliar os desempenhos dos estimadores de máxima verossimilhança e dos testes da razão de verossimilhanças baseados em tais funções. Também são apresentadas aplicações dessas funções de verossimilhança em modelos não pertencentes à família exponencial biparamétrica, mais precisamente, na família de distribuições GA0(alfa,gama,L), usada para modelar dados de imagens de radar, e no modelo de Weibull, muito usado em aplicações da área da engenharia denominada confiabilidade, considerando dados completos e censurados. Aqui também foram obtidos resultados numéricos a fim de avaliar a qualidade dos ajustes sobre a verossimilhança perfilada, analogamente às simulações realizadas para a família exponencial biparamétrica. Vale mencionar que, no caso da família de distribuições GA0(alfa,gama,L), foi avaliada a aproximação da distribuição da estatística da razão de verossimilhanças sinalizada pela distribuição normal padrão. Além disso, no caso do modelo de Weibull, vale destacar que foram derivados resultados distribucionais relativos aos estimadores de máxima verossimilhança e às estatísticas da razão de verossimilhanças para dados completos e censurados, apresentados em apêndice. / The profile likelihood function is not genuine likelihood function, and profile maximum likelihood estimators are typically inefficient and inconsistent. Additionally, the null distribution of the likelihood ratio test statistic can be poorly approximated by the asymptotic chi-squared distribution in finite samples when there are nuisance parameters. It is thus important to obtain adjustments to the likelihood function. Several authors, including Barndorff-Nielsen (1983,1994), Cox and Reid (1987,1992), McCullagh and Tibshirani (1990) and Stern (1997), have proposed modifications to the profile likelihood function. They are defined in a such a way to reduce the score and information biases. In this dissertation, we review several profile likelihood adjustments and also approximations to the adjustments proposed by Barndorff-Nielsen (1983,1994), also described in Severini (2000a). We present derivations and the main properties of the different adjustments. We also obtain adjustments for likelihood-based inference in the two-parameter exponential family. Numerical results on estimation and testing are provided. We also consider models that do not belong to the two-parameter exponential family: the GA0(alfa,gama,L) family, which is commonly used to model image radar data, and the Weibull model, which is useful for reliability studies, the latter under both noncensored and censored data. Again, extensive numerical results are provided. It is noteworthy that, in the context of the GA0(alfa,gama,L) model, we have evaluated the approximation of the null distribution of the signalized likelihood ratio statistic by the standard normal distribution. Additionally, we have obtained distributional results for the Weibull case concerning the maximum likelihood estimators and the likelihood ratio statistic both for noncensored and censored data.
123

A distribuição beta generalizada semi-normal / The beta generalized half-normal distribution

Rodrigo Rossetto Pescim 29 January 2010 (has links)
Uma nova família de distribuições denominada distribuição beta generalizada semi-normal, que inclui algumas distribuições importantes como casos especiais, tais como as distribuições semi-normal e generalizada semi-normal (Cooray e Ananda, 2008), é proposta neste trabalho. Para essa nova família de distribuições, foi realizado o estudo da função densidade probabilidade, função de distribuição acumulada e da função de taxa de falha (ou risco), que não dependeram de funções matemáticas complicadas. Obteve-se uma expressão formal para os momentos, função geradora de momentos, função densidade da distribuição de estatística de ordem, desvios médios, entropia, contabilidade e para as curvas de Bonferroni e Lorenz. Examinaram-se os estimadores de máxima verossimilhança dos parâmetros e deduziu- se a matriz de informação esperada. Neste trabalho é proposto, também, um modelo de regressão utilizando a distribuição beta generalizada semi-normal. A utilidade dessa nova distribuição é ilustrada através de dois conjuntos de dados, mostrando que ela é mais flexível na análise de dados de tempo de vida do que outras distribuições existentes na literatura. / A new family of distributions so-called beta generalized half-normal distribution, which includes some important distributions as special cases, such as the half-normal and generalized half-normal (Cooray and Ananda, 2008) distributions, is proposed in this work. For this new family of distributions, we studied the probability density function, cumulative distribution function and failure rate function (or hazard function), which did not depend on complicated mathematical functions. We obtained a formal expression for the moments, moment generating function, density function of order statistics distribution, mean deviation, entropy, reliability and Bonferroni and Lorenz curves. We examined maximum likelihood estimation of parameters and provided the information matrix. This work also proposed a regression model using the beta generalized half-normal distribution. The usefulness of the new distribution is illustrated through two data sets by showing that it is quite °exible in analyzing lifetime data instead other distributions in the literature.
124

Modelos de regressão beta inflacionados / Inflated beta regression models

Raydonal Ospina Martinez 04 April 2008 (has links)
Nos últimos anos têm sido desenvolvidos modelos de regressão beta, que têm uma variedade de aplicações práticas como, por exemplo, a modelagem de taxas, razões ou proporções. No entanto, é comum que dados na forma de proporções apresentem zeros e/ou uns, o que não permite admitir que os dados provêm de uma distribuição contínua. Nesta tese, são propostas, distribuições de mistura entre uma distribuição beta e uma distribuição de Bernoulli, degenerada em zero e degenerada em um para modelar dados observados nos intervalos [0, 1], [0, 1) e (0, 1], respectivamente. As distribuições propostas são inflacionadas no sentido de que a massa de probabilidade em zero e/ou um excede o que é permitido pela distribuição beta. Propriedades dessas distribuições são estudadas, métodos de estimação por máxima verossimilhança e momentos condicionais são comparados. Aplicações a vários conjuntos de dados reais são examinadas. Desenvolvemos também modelos de regressão beta inflacionados assumindo que a distribuição da variável resposta é beta inflacionada. Estudamos estimação por máxima verossimilhança. Derivamos expressões em forma fechada para o vetor escore, a matriz de informação de Fisher e sua inversa. Discutimos estimação intervalar para diferentes quantidades populacionais (parâmetros de regressão, parâmetro de precisão) e testes de hipóteses assintóticos. Derivamos expressões para o viés de segunda ordem dos estimadores de máxima verossimilhança dos parâmetros, possibilitando a obtenção de estimadores corrigidos que são mais precisos que os não corrigidos em amostras finitas. Finalmente, desenvolvemos técnicas de diagnóstico para os modelos de regressão beta inflacionados, sendo adotado o método de influência local baseado na curvatura normal conforme. Ilustramos a teoria desenvolvida em um conjuntos de dados reais. / The last years have seen new developments in the theory of beta regression models, which are useful for modelling random variables that assume values in the standard unit interval such as proportions, rates and fractions. In many situations, the dependent variable contains zeros and/or ones. In such cases, continuous distributions are not suitable for modeling this kind of data. In this thesis we propose mixed continuous-discrete distributions to model data observed on the intervals [0, 1],[0, 1) and (0, 1]. The proposed distributions are inflated beta distributions in the sense that the probability mass at 0 and/or 1 exceeds what is expected for the beta distribution. Properties of the inflated beta distributions are given. Estimation based on maximum likelihood and conditional moments is discussed and compared. Empirical applications using real data set are provided. Further, we develop inflated beta regression models in which the underlying assumption is that the response follows an inflated beta law. Estimation is performed by maximum likelihood. We provide closed-form expressions for the score function, Fishers information matrix and its inverse. Interval estimation for different population quantities (such as regression parameters, precision parameter, mean response) is discussed and tests of hypotheses on the regression parameters can be performed using asymptotic tests. We also derive the second order biases of the maximum likelihood estimators and use them to define bias-adjusted estimators. The numerical results show that bias reduction can be effective in finite samples. We also develop a set of diagnostic techniques that can be employed to identify departures from the postulated model and influential observations. To that end, we adopt the local influence approach based in the conformal normal curvature. Finally, we consider empirical examples to illustrate the theory developed.
125

Estimação via EM e diagnóstico em modelos misturas assimétricas com regressão

Louredo, Graciliano Márcio Santos 26 February 2018 (has links)
Submitted by Geandra Rodrigues (geandrar@gmail.com) on 2018-04-10T15:11:39Z No. of bitstreams: 1 gracilianomarciosantoslouredo.pdf: 1813142 bytes, checksum: b79d02006212c4f63d6836c9a417d4bc (MD5) / Approved for entry into archive by Adriana Oliveira (adriana.oliveira@ufjf.edu.br) on 2018-04-11T15:25:36Z (GMT) No. of bitstreams: 1 gracilianomarciosantoslouredo.pdf: 1813142 bytes, checksum: b79d02006212c4f63d6836c9a417d4bc (MD5) / Made available in DSpace on 2018-04-11T15:25:36Z (GMT). No. of bitstreams: 1 gracilianomarciosantoslouredo.pdf: 1813142 bytes, checksum: b79d02006212c4f63d6836c9a417d4bc (MD5) Previous issue date: 2018-02-26 / FAPEMIG - Fundação de Amparo à Pesquisa do Estado de Minas Gerais / O objetivo deste trabalho é apresentar algumas contribuições para a melhoria do processo de estimação por máxima verossimilhança via algoritmo EM em modelos misturas assimétricas com regressão, além de realizar neles a análise de influência local e global. Essas contribuições, em geral de natureza computacional, visam à resolução de problemas comuns na modelagem estatística de maneira mais eficiente. Dentre elas está a substituição de métodos utilizados nas versões dos algoritmos GEM por outras que reduzem o problema aproximadamente a um algoritmo EM clássico nos principais exemplos das distribuições misturas de escala assimétricas de normais. Após a execução do processo de estimação, discutiremos ainda as principais técnicas existentes para o diagnóstico de pontos influentes com as adaptações necessárias aos modelos em foco. Desejamos com tal abordagem acrescentar ao tratamento dessa classe de modelos estatísticos a análise de regressão nas distribuições mais recentes na literatura. Também esperamos abrir caminho para o uso de técnicas similares em outras classes de modelos. / The objective of this work is to present some contributions to improvement the process of maximum likelihood estimation via the EM algorithm in skew mixtures models with regression, as well as to execute in them the global and local influence analysis. These contributions, usually with computational nature, aim to solving common problems in statistical modeling more efficiently. Among them is the replacement of used methods in the versions of the GEM algorithm by other techniques that reduce the problem approximately to a classic EM algorithm in the main examples of skew scale mixtures of normals distributions. After performing the estimation process, we will also discuss the main existing techniques for the diagnosis of influential points with the necessaries adaptations to the models in focus. We wish with this approach to add for the treatment of this statistical model class the regression analysis in the most recent distributions in the literature. We too hope to paving the way for use of similar techniques in other models classes.
126

Analyse statistique de processus stochastiques : application sur des données d’orages / Inference for some stochastic processes : with application on thunderstorm data

Do, Van-Cuong 19 April 2019 (has links)
Les travaux présentés dans cette thèse concernent l'analyse statistique de cas particuliers du processus de Cox. Dans une première partie, nous proposons une synthèse des résultats existants sur le processus power-law (processus d'intensité puissance), synthèse qui ne peut être exhaustive étant donné la popularité de ce processus. Nous considérons une approche bayésienne pour l'inférence des paramètres de ce processus qui nous conduit à introduire et à étudier en détails une distribution que nous appelons loi H-B. Cette loi est une loi conjuguée. Nous proposons des stratégies d'élicitation des hyperparamètres et étudions le comportement des estimateurs de Bayes par des simulations. Dans un deuxième temps, nous étendons ces travaux au cas du processus d’intensité exponentielle (exponential-law process). De la même façon, nous définissons et étudions une loi conjuguée pour l'analyse bayésienne de ce dernier. Dans la dernière partie de la thèse, nous considérons un processus auto-excité qui intègre une covariable. Ce travail est motivé, à l'origine, par un problème de fiabilité qui concerne des données de défaillances de matériels exposés à des environnements sévères. Les résultats sont illustrés par des applications sur des données d'activités orageuses collectées dans deux départements français. Enfin, nous donnons quelques directions de travail et perspectives de futurs développements de l'ensemble de nos travaux. / The work presented in this PhD dissertation concerns the statistical analysis of some particular cases of the Cox process. In a first part, we study the power-law process (PLP). Since the literature for the PLP is abundant, we suggest a state-of-art for the process. We consider the classical approach and recall some important properties of the maximum likelihood estimators. Then we investigate a Bayesian approach with noninformative priors and conjugate priors considering different parametrizations and scenarios of prior guesses. That leads us to define a family of distributions that we name H-B distribution as the natural conjugate priors for the PLP. Bayesian analysis with the conjugate priors are conducted via a simulation study and an application on real data. In a second part, we study the exponential-law process (ELP). We review the maximum likelihood techniques. For Bayesian analysis of the ELP, we define conjugate priors: the modified- Gumbel distribution and Gamma-modified-Gumbel distribution. We conduct a simulation study to compare maximum likelihood estimates and Bayesian estimates. In the third part, we investigate self-exciting point processes and we integrate a power-law covariate model to this intensity of this process. A maximum likelihood procedure for the model is proposed and the Bayesian approach is suggested. Lastly, we present an application on thunderstorm data collected in two French regions. We consider a strategy to define a thunderstorm as a temporal process associated with the charges in a particular location. Some selected thunderstorms are analyzed. We propose a reduced maximum likelihood procedure to estimate the parameters of the Hawkes process. Then we fit some thunderstorms to the power-law covariate self-exciting point process taking into account the associated charges. In conclusion, we give some perspectives for further work.
127

Optimisation of adaptive localisation techniques for cognitive radio

Thomas, Robin Rajan 06 August 2012 (has links)
Spectrum, environment and location awareness are key characteristics of cognitive radio (CR). Knowledge of a user’s location as well as the surrounding environment type may enhance various CR tasks, such as spectrum sensing, dynamic channel allocation and interference management. This dissertation deals with the optimisation of adaptive localisation techniques for CR. The first part entails the development and evaluation of an efficient bandwidth determination (BD) model, which is a key component of the cognitive positioning system. This bandwidth efficiency is achieved using the Cramer-Rao lower bound derivations for a single-input-multiple-output (SIMO) antenna scheme. The performances of the single-input-single-output (SISO) and SIMO BD models are compared using three different generalised environmental models, viz. rural, urban and suburban areas. In the case of all three scenarios, the results reveal a marked improvement in the bandwidth efficiency for a SIMO antenna positioning scheme, especially for the 1×3 urban case, where a 62% root mean square error (RMSE) improvement over the SISO system is observed. The second part of the dissertation involves the presentation of a multiband time-of arrival (TOA) positioning technique for CR. The RMSE positional accuracy is evaluated using a fixed and dynamic bandwidth availability model. In the case of the fixed bandwidth availability model, the multiband TOA positioning model is initially evaluated using the two-step maximum-likelihood (TSML) location estimation algorithm for a scenario where line-of-sight represents the dominant signal path. Thereafter, a more realistic dynamic bandwidth availability model has been proposed, which is based on data obtained from an ultra-high frequency spectrum occupancy measurement campaign. The RMSE performance is then verified using the non-linear least squares, linear least squares and TSML location estimation techniques, using five different bandwidths. The proposed multiband positioning model performs well in poor signal-to-noise ratio conditions (-10 dB to 0 dB) when compared to a single band TOA system. These results indicate the advantage of opportunistic TOA location estimation in a CR environment. / Dissertation (MEng)--University of Pretoria, 2012. / Electrical, Electronic and Computer Engineering / unrestricted
128

[en] A POISSON-LOGNORMAL MODEL TO FORECAST THE IBNR QUANTITY VIA MICRO-DATA / [pt] UM MODELO POISSON-LOGNORMAL PARA PREVISÃO DA QUANTIDADE IBNR VIA MICRO-DADOS

JULIANA FERNANDES DA COSTA MACEDO 02 February 2016 (has links)
[pt] O principal objetivo desta dissertação é realizar a previsão da reserva IBNR. Para isto foi desenvolvido um modelo estatístico de distribuições combinadas que busca uma adequada representação dos dados. A reserva IBNR, sigla em inglês para Incurred But Not Reported, representa o montante que as seguradoras precisam ter para pagamentos de sinistros atrasados, que já ocorreram no passado, mas ainda não foram avisados à seguradora até a data presente. Dada a importância desta reserva, diversos métodos para estimação da reserva IBNR já foram propostos. Um dos métodos mais utilizado pelas seguradoras é o Método Chain Ladder, que se baseia em triângulos run-off, que é o agrupamento dos dados conforme data de ocorrência e aviso de sinistro. No entanto o agrupamento dos dados faz com que informações importantes sejam perdidas. Esta dissertação baseada em outros artigos e trabalhos que consideram o não agrupamento dos dados, propõe uma nova modelagem para os dados não agrupados. O modelo proposto combina a distribuição do atraso no aviso da ocorrência, representada aqui pela distribuição log-normal truncada (pois só há informação até a última data observada); a distribuição da quantidade total de sinistros ocorridos num dado período, modelada pela distribuição Poisson; e a distribuição do número de sinistros ocorridos em um dado período e avisados até a última data observada, que será caracterizada por uma distribuição Binomial. Por fim, a quantidade de sinistros IBNR foi estimada por método e pelo Chain Ladder e avaliou-se a capacidade de previsão de ambos. Apesar da distribuição de atrasos do modelo proposto se adequar bem aos dados, o modelo proposto obteve resultados inferiores ao Chain Ladder em termos de previsão. / [en] The main objective of this dissertation is to predict the IBNR reserve. For this, it was developed a statistical model of combined distributions looking for a new distribution that fits the data well. The IBNR reserve, short for Incurred But Not Reported, represents the amount that insurers need to have to pay for the claims that occurred in the past but have not been reported until the present date. Given the importance of this reserve, several methods for estimating this reserve have been proposed. One of the most used methods for the insurers is the Chain Ladder, which is based on run-off triangles; this is a format of grouping the data according to the occurrence and the reported date. However this format causes the lost of important information. This dissertation, based on other articles and works that consider the data not grouped, proposes a new model for the non-aggregated data. The proposed model combines the delay in the claim report distribution represented by a log normal truncated (because there is only information until the last observed date); the total amount of claims incurred in a given period modeled by a Poisson distribution and the number of claims occurred in a certain period and reported until the last observed date characterized by a binomial distribution. Finally, the IBNR reserve was estimated by this method and by the chain ladder and the prediction capacity of both methods will be evaluated. Although the delay distribution seems to fit the data well, the proposed model obtained inferior results to the Chain Ladder in terms of forecast.
129

Improving term structure measurements by incorporating steps in a multiple yield curve framework

Villwock, Gustav, Rydholm, Clara January 2022 (has links)
By issuing interest rate derivative contracts, market makers such as large banks are exposed to undesired risk. There are several methods for banks to hedge themselves against this type of risk; one such method is the stochastic programming model developed by Blomvall and Hagenbjörk (2022). The effectiveness of their model relies on accurate pricing of interest rate derivatives and risk factor analysis, both of which are derived from a term structure. Blomvall and Ndengo (2013) present a discretized multiple yield curve framework for term structure measurement that allows for price deviations. The model uses regularization to deal with noise inherent in market price observations, where the regularization counteracts oscillations in the term structure and retains the smoothness of the curve by penalizing the first and second-order derivatives. Consequently, the resulting model creates a trade-off between a smooth curve and market price deviations. Changes in policy rates adjusted by a country’s central bank significantly impact the financial market and its actors. In this thesis, the model developed by Blomvall and Ndengo (2013) was further extended to include these steps in conjunction with monetary policy meetings. Two models were developed to realize the steps in the risk-free curve. The first model introduced an additional deviation term to allow for a shift in the curve. In the second model, the weights in the regularization were adjusted to allow for rapid changes on days surrounding the closest monetary policy meeting. A statistical test was conducted to determine the performance of the two models. The test showed that the model with adjusted regularization outperformed the model with an additional deviation term as well as a benchmark model without steps. However, both step models managed to reduce in-sample pricing errors, while the model with an additional deviation term performed worse than the benchmark model for out-of-sample data, given the current parameter setting. Other parameter combinations would potentially result in different outcomes, but it remains conjectural.
130

Exploring Normalizing Flow Modifications for Improved Model Expressivity / Undersökning av normalizing flow-modifikationer för förbättrad modelluttrycksfullhet

Juschak, Marcel January 2023 (has links)
Normalizing flows represent a class of generative models that exhibit a number of attractive properties, but do not always achieve state-of-the-art performance when it comes to perceived naturalness of generated samples. To improve the quality of generated samples, this thesis examines methods to enhance the expressivity of discrete-time normalizing flow models and thus their ability to capture different aspects of the data. In the first part of the thesis, we propose an invertible neural network architecture as an alternative to popular architectures like Glow that require an individual neural network per flow step. Although our proposal greatly reduces the number of parameters, it has not been done before, as such architectures are believed to not be powerful enough. For this reason, we define two optional extensions that could greatly increase the expressivity of the architecture. We use augmentation to add Gaussian noise variables to the input to achieve arbitrary hidden-layer widths that are no longer dictated by the dimensionality of the data. Moreover, we implement Piecewise Affine Activation Functions that represent a generalization of Leaky ReLU activations and allow for more powerful transformations in every individual step. The resulting three models are evaluated on two simple synthetic datasets – the two moons dataset and one generated from a mixture of eight Gaussians. Our findings indicate that the proposed architectures cannot adequately model these simple datasets and thus do not represent alternatives to current stateof-the-art models. The Piecewise Affine Activation Function significantly improved the expressivity of the invertible neural network, but could not make use of its full potential due to inappropriate assumptions about the function’s input distribution. Further research is needed to ensure that the input to this function is always standard normal distributed. We conducted further experiments with augmentation using the Glow model and could show minor improvements on the synthetic datasets when only few flow steps (two, three or four) were used. However, in a more realistic scenario, the model would encompass many more flow steps. Lastly, we generalized the transformation in the coupling layers of modern flow architectures from an elementwise affine transformation to a matrixbased affine transformation and studied the effect this had on MoGlow, a flow-based model of motion. We could show that McMoGlow, our modified version of MoGlow, consistently achieved a better training likelihood than the original MoGlow on human locomotion data. However, a subjective user study found no statistically significant difference in the perceived naturalness of the samples generated. As a possible reason for this, we hypothesize that the improvements are subtle and more visible in samples that exhibit slower movements or edge cases which may have been underrepresented in the user study. / Normalizing flows representerar en klass av generativa modeller som besitter ett antal eftertraktade egenskaper, men som inte alltid uppnår toppmodern prestanda när det gäller upplevd naturlighet hos genererade data. För att förbättra kvaliteten på dessa modellers utdata, undersöker detta examensarbete metoder för att förbättra uttrycksfullheten hos Normalizing flows-modeller i diskret tid, och därmed deras förmåga att fånga olika aspekter av datamaterialet. I den första delen av uppsatsen föreslår vi en arkitektur uppbyggt av ett inverterbart neuralt nätverk. Vårt förslag är ett alternativ till populära arkitekturer som Glow, vilka kräver individuella neuronnät för varje flödessteg. Även om vårt förslag kraftigt minskar antalet parametrar har detta inte gjorts tidigare, då sådana arkitekturer inte ansetts vara tillräckligt kraftfulla. Av den anledningen definierar vi två oberoende utökningar till arkitekturen som skulle kunna öka dess uttrycksfullhet avsevärt. Vi använder så kallad augmentation, som konkatenerar Gaussiska brusvariabler till observationsvektorerna för att uppnå godtyckliga bredder i de dolda lagren, så att deras bredd inte längre begränsas av datadimensionaliteten. Dessutom implementerar vi Piecewise Affine Activation-funktioner (PAAF), vilka generaliserar Leaky ReLU-aktiveringar genom att möjliggöra mer kraftfulla transformationer i varje enskilt steg. De resulterande tre modellerna utvärderas med hjälp av två enkla syntetiska datamängder - ”the two moons dataset” och ett som genererats genom att blanda av åtta Gaussfördelningar. Våra resultat visar att de föreslagna arkitekturerna inte kan modellera de enkla datamängderna på ett tillfredsställande sätt, och därmed inte utgör kompetitiva alternativ till nuvarande moderna modeller. Den styckvisa aktiveringsfunktionen förbättrade det inverterbara neurala nätverkets uttrycksfullhet avsevärt, men kunde inte utnyttja sin fulla potential på grund av felaktiga antaganden om funktionens indatafördelning. Ytterligare forskning behövs för att hantera detta problem. Vi genomförde ytterligare experiment med augmentation av Glow-modellen och kunde påvisa vissa förbättringar på de syntetiska dataseten när endast ett fåtal flödessteg (två, tre eller fyra) användes. Däremot omfattar modeller i mer realistiska scenarion många fler flödessteg. Slutligen generaliserade vi transformationen i kopplingslagren hos moderna flödesarkitekturer från en elementvis affin transformation till en matrisbaserad affin transformation, samt studerade vilken effekt detta hade på MoGlow, en flödesbaserad modell av 3D-rörelser. Vi kunde visa att McMoGlow, vår modifierade version av MoGlow, konsekvent uppnådde bättre likelihood i träningen än den ursprungliga MoGlow gjorde på mänskliga rörelsedata. En subjektiv användarstudie på exempelrörelser genererade från MoGlow och McMoGlow visade dock ingen statistiskt signifikant skillnad i användarnas uppfattning av hur naturliga rörelserna upplevdes. Som en möjlig orsak till detta antar vi att förbättringarna är subtila och mer synliga i situationer som uppvisar långsammare rörelser eller i olika gränsfall som kan ha varit underrepresenterade i användarstudien.

Page generated in 0.2362 seconds