[pt] Esta dissertação trata da aplicação de modelos lineares mistos em dados provenientes de experimentos com restrição na aleatorização. O experimento utilizado neste trabalho teve como finalidade verificar quais eram os fatores de controle do processo de laminação a frio que mais afetavam a espessura do material utilizado na fabricação das latas para bebidas carbonatadas. A partir do experimento, foram obtidos dados para modelar a média e a variância da espessura do material. O objetivo da modelagem era identificar quais fatores faziam com que a espessura média atingisse o valor desejado (0,248 mm). Além disso, era necessário identificar qual a combinação dos níveis desses fatores que produzia a variância mínima na espessura do material. Houve replicações neste experimento, mas estas não foram executadas de forma aleatória, e, além disso, os níveis dos fatores utilizados não foram reinicializados, nas rodadas do experimento. Devido a estas restrições, foram utilizados modelos mistos para o ajuste da média, e da variância, da espessura, uma vez que com tais modelos é possível trabalhar na presença de dados auto-correlacionados e heterocedásticos. Os modelos mostraram uma boa adequação aos dados, indicando que para situações onde existe restrição na aleatorização, a utilização de modelos mistos se mostra apropriada. / [en] This dissertation presents an application of linear mixed models on data from an experiment with restriction in randomization. The experiment used in this study was aimed to verify which were the controlling factors, in the cold-rolling process, that most affected the thickness of the material used in the carbonated beverages market segment. From the experiment, data were obtained to model the mean and variance of the thickness of the material. The goal of modeling was to identify which factors were significant for the thickness reaches the desired value (0.248 mm). Furthermore, it was necessary to identify which combination of levels, of these factors, produced the minimum variance in the thickness of the material. There were replications of this experiment, but these were not performed randomly. In addition, the levels of factors used were not restarted during the trials. Due to these limitations, mixed models were used to adjust the mean and the variance of the thickness. The models showed a good fit to the data, indicating that for situations where there is restriction on randomization, the use of mixed models is suitable.

Nonlinear parametric inverse problems appear in many applications in science and engineering. We focus on diffuse optical tomography (DOT) in medical imaging. DOT aims to recover an unknown image of interest, such as the absorption coefficient in tissue to locate tumors in the body. Using a mathematical (forward) model to predict measurements given a parametrization of the tissue, we minimize the misfit between predicted and actual measurements up to a given noise level. The main computational bottleneck in such inverse problems is the repeated evaluation of this large-scale forward model, which corresponds to solving large linear systems for each source and frequency at each optimization step. Moreover, to efficiently compute derivative information, we need to solve, repeatedly, linear systems with the adjoint for each detector and frequency. As rapid advances in technology allow for large numbers of sources and detectors, these problems become computationally prohibitive. In this thesis, we introduce two methods to drastically reduce this cost. To efficiently implement Newton methods, we extend the use of simultaneous random sources to reduce the number of linear system solves to include simultaneous random detectors. Moreover, we combine simultaneous random sources and detectors with optimized ones that lead to faster convergence and more accurate solutions. We can use reduced order models (ROM) to drastically reduce the size of the linear systems to be solved in each optimization step while still solving the inverse problem accurately. However, the construction of the ROM bases still incurs a substantial cost. We propose to use randomization to drastically reduce the number of large linear solves needed for constructing the global ROM bases without degrading the accuracy of the solution to the inversion problem. We demonstrate the efficiency of these approaches with 2-dimensional and 3-dimensional examples from DOT; however, our methods have the potential to be useful for other applications as well. / Ph. D. / Medical image reconstruction presents huge computational challenges due to the quantity of data generated by modern equipment. Each stage of processing requires the solution of more than a thousand large, three-dimensional problems. Moreover, as rapid advances in technology allow for ever larger numbers of sources and detectors and using multiple frequencies, these problems become computationally prohibitive. In this thesis, we develop two computational methods to drastically reduce this cost and produce good images from measurements. First, we focus on efficiently estimating the absorption image while we reduce the cost of each optimization step by solving only for a few linear combinations of sources and of detectors. Second, we can replace the full mathematical model by a reduced mathematical model to drastically reduce the size of the linear systems in each optimization step while still producing good image reconstructions. However, the computation of this reduced model still poses a formidable cost. Hence, we propose to reduce the cost of building the reduced model by sampling the sources and detectors. Using this reduced model for image reconstruction does not degrade the accuracy of the solutions and the quality of the image reconstruction. We demonstrate the efficiency of these approaches with 2-dimensional and 3-dimensional examples from medical imaging. However, our methods have the potential to be useful for other applications as well.

Yes / Aims: There is a clear association between alcohol use and offending behaviour and significant police time is spent on alcohol-related incidents. This study aimed to test the feasibility of a trial of screening and brief intervention in police custody suites to reduce heavy drinking and re-offending behaviour. Short summary: We achieved target recruitment and high brief intervention delivery if this occurred immediately after screening. Low rates of return for counselling and retention at follow-up were challenges for a definitive trial. Conversely, high consent rates for access to police data suggested at least some outcomes could be measured remotely. Methods: A three-armed pilot Cluster Randomised Controlled Trial with an embedded qualitative interview-based process evaluation to explore acceptability issues in six police custody suites (north east and south west of the UK). Interventions included: 1. Screening only (Controls), 2. 10 min Brief Advice 3. Brief Advice plus 20 min of brief Counselling. Results: Of 3330 arrestees approached: 2228 were eligible for screening (67%) and 720 consented (32%); 386 (54%) scored 8+ on AUDIT; and 205 (53%) were enroled (79 controls, 65 brief advice and 61 brief counselling). Follow-up rates at 6 and 12 months were 29% and 26%, respectively. However, routinely collected re-offending data were obtained for 193 (94%) participants. Indices of deprivation data were calculated for 184 (90%) participants; 37.6% of these resided in the 20% most deprived areas of UK. Qualitative data showed that all arrestees reported awareness that participation was voluntary, that the trial was separate from police work, and the majority said trial procedures were acceptable. Conclusion: Despite hitting target recruitment and same-day brief intervention delivery, a future trial of alcohol screening and brief intervention in a police custody setting would only be feasible if routinely collected re-offending and health data were used for outcome measurement. / NIHR School for Public Health Research (SPHR) (SPHR-SWP-ALC-WP2). Fuse is a UK Clinical Research Collaboration (UKCRC) Public Health Research Centre of Excellence. Funding for Fuse from the British Heart Foundation, Cancer Research UK, Economic and Social Research Council, Medical Research Council, the National Institute for Health Research, under the auspices of the UKCRC, is gratefully acknowledged.

The development of high-performing object detection models requires extensive and varied datasets with accurately annotated images, a process that is traditionally labor-intensive and prone to errors. To address these challenges, this report explores the generation of synthetic data using domain randomization techniques to train object detection models. We present a pipeline that integrates synthetic data creation in Unity, and the training of YOLOv8 object detection models. Our approach uses the Unity Perception package to produce diverse and precisely annotated datasets, overcoming the domain gap typically associated with synthetic data. The pipeline was evaluated through a series of experiments, analyzing the impact of various parameters such as background textures, and training arguments on model performance. The results demonstrate that models trained with our synthetic data can achieve high accuracy and generalize well to real-world scenarios, offering a scalable and efficient alternative to manual data annotation.

The chemical master equation, formulated on the Markov assumption of underlying chemical kinetics, offers an accurate stochastic description of general chemical reaction systems on the mesoscopic scale. The chemical master equation is especially useful when formulating mathematical models of gene regulatory networks and protein-protein interaction networks, where the numbers of molecules of most species are around tens or hundreds. However, solving the master equation directly suffers from the so called "curse of dimensionality" issue. This thesis first tries to study the numerical properties of the master equation using existing numerical methods and parallel machines. Next, approximation algorithms, namely the adaptive aggregation method and the radial basis function collocation method, are proposed as new paths to resolve the "curse of dimensionality". Several numerical results are presented to illustrate the promises and potential problems of these new algorithms. Comparisons with other numerical methods like Monte Carlo methods are also included. Development and analysis of the linear Shepard algorithm and its variants, all of which could be used for high dimensional scattered data interpolation problems, are also included here, as a candidate to help solve the master equation by building surrogate models in high dimensions. / Ph. D.

Composite endpoints are widely used in clinical trials. The outcome of a clinical trial can affect many individuals and it is therefore of importance that the methods used are as effective and correct as possible. Improvements of the standard method of testing composite endpoints have been proposed and in this thesis, the alternative method using nonparametric combination methodology is compared to the standard method. Performing a simulation study, the power of three combining functions (Fisher, Tippett and the Logistic) are compared to the power of the standard method. The performances of the four methods are evaluated for different compositions of treatment effects, as well as for independent and dependent components. The results show that using the nonparametric combination methodology leads to higher power in both dependent and independent cases. The combining functions are suitable for different compositions of treatment effects, the Fisher combining function being the most versatile. The thesis is written with support from Statisticon AB.

Software vulnerabilities in programs and malware deployments have been increasing almost every year since we started measuring them. Information about how to program securely, how malware shall be avoided and technological countermeasures for this are more available than ever. Still, the trend seems to favor the attacker. This thesis tries to visualize the effects of a selection of technological countermeasures that have been proposed by researchers. These countermeasures: non-executable memory, address randomization, system call interception and file integrity monitoring are described along with the attacks they are designed to defend against. The coverage of each countermeasure is then visualized with the help of attack trees. Attack trees are normally used for describing how systems can be attacked but here they instead serve the purpose of showing where in an attack a countermeasure takes effect. Using attack trees for this highlights a couple of important aspects of a security mechanism, such as how early in an attack it is effective and which variants of an attack it potentially defends against. This is done by the use of what we call defensive codes that describe how a defense mechanism counters a sub-goal in an attack. Unfortunately the whole process is not well formalized and depends on many uncertain factors.

The aim of this dissertation is to develop a new algorithm to embed a watermark in JPEG compressed images, using encoding methods. This encompasses the embedding of proprietary information, such as identity and authentication bitstrings, into the compressed material. This watermark encoding scheme involves combining entropy coding with homophonic coding, in order to embed a watermark in a JPEG image. Arithmetic coding was used as the entropy encoder for this scheme. It is often desired to obtain a robust digital watermarking method that does not distort the digital image, even if this implies that the image is slightly expanded in size before final compression. In this dissertation an algorithm that combines homophonic and arithmetic coding for JPEG images was developed and implemented in software. A detailed analysis of this algorithm is given and the compression (in number of bits) obtained when using the newly developed algorithm (homophonic and arithmetic coding). This research shows that homophonic coding can be used to embed a watermark in a JPEG image by using the watermark information for the selection of the homophones. The proposed algorithm can thus be viewed as a ‘key-less’ encryption technique, where an external bitstring is used as a ‘key’ and is embedded intrinsically into the message stream. The algorithm has achieved to create JPEG images with minimal distortion, with Peak Signal to Noise Ratios (PSNR) of above 35dB. The resulting increase in the entropy of the file is within the expected 2 bits per symbol. This research endeavor consequently provides a unique watermarking technique for images compressed using the JPEG standard. / Dissertation (MEng)--University of Pretoria, 2008. / Electrical, Electronic and Computer Engineering / unrestricted

L'apprentissage robotique est incroyablement prometteur pour l'intelligence artificielle incarnée, avec un apprentissage par renforcement apparemment parfait pour les robots du futur: apprendre de l'expérience, s'adapter à la volée et généraliser à des scénarios invisibles. Cependant, notre réalité actuelle nécessite de grandes quantités de données pour former la plus simple des politiques d'apprentissage par renforcement robotique, ce qui a suscité un regain d'intérêt de la formation entièrement dans des simulateurs de physique efficaces. Le but étant l'intelligence incorporée, les politiques formées à la simulation sont transférées sur du matériel réel pour évaluation; cependant, comme aucune simulation n'est un modèle parfait du monde réel, les politiques transférées se heurtent à l'écart de transfert sim2real: les erreurs se sont produites lors du déplacement des politiques des simulateurs vers le monde réel en raison d'effets non modélisés dans des modèles physiques inexacts et approximatifs. La randomisation de domaine - l'idée de randomiser tous les paramètres physiques dans un simulateur, forçant une politique à être robuste aux changements de distribution - s'est avérée utile pour transférer des politiques d'apprentissage par renforcement sur de vrais robots. En pratique, cependant, la méthode implique un processus difficile, d'essais et d'erreurs, montrant une grande variance à la fois en termes de convergence et de performances. Nous introduisons Active Domain Randomization, un algorithme qui implique l'apprentissage du curriculum dans des espaces de tâches non structurés (espaces de tâches où une notion de difficulté - tâches intuitivement faciles ou difficiles - n'est pas facilement disponible). La randomisation de domaine active montre de bonnes performances sur le pourrait utiliser zero shot sur de vrais robots. La thèse introduit également d'autres variantes de l'algorithme, dont une qui permet d'incorporer un a priori de sécurité et une qui s'applique au domaine de l'apprentissage par méta-renforcement. Nous analysons également l'apprentissage du curriculum dans une perspective d'optimisation et tentons de justifier les avantages de l'algorithme en étudiant les interférences de gradient. / Robotic learning holds incredible promise for embodied artificial intelligence, with reinforcement learning seemingly a strong candidate to be the \textit{software} of robots of the future: learning from experience, adapting on the fly, and generalizing to unseen scenarios. However, our current reality requires vast amounts of data to train the simplest of robotic reinforcement learning policies, leading to a surge of interest of training entirely in efficient physics simulators. As the goal is embodied intelligence, policies trained in simulation are transferred onto real hardware for evaluation; yet, as no simulation is a perfect model of the real world, transferred policies run into the sim2real transfer gap: the errors accrued when shifting policies from simulators to the real world due to unmodeled effects in inaccurate, approximate physics models. Domain randomization - the idea of randomizing all physical parameters in a simulator, forcing a policy to be robust to distributional shifts - has proven useful in transferring reinforcement learning policies onto real robots. In practice, however, the method involves a difficult, trial-and-error process, showing high variance in both convergence and performance. We introduce Active Domain Randomization, an algorithm that involves curriculum learning in unstructured task spaces (task spaces where a notion of difficulty - intuitively easy or hard tasks - is not readily available). Active Domain Randomization shows strong performance on zero-shot transfer on real robots. The thesis also introduces other variants of the algorithm, including one that allows for the incorporation of a safety prior and one that is applicable to the field of Meta-Reinforcement Learning. We also analyze curriculum learning from an optimization perspective and attempt to justify the benefit of the algorithm by studying gradient interference.

This thesis studies inference to the complier treatment effect denoted LATE. The standard approach is to base the inference on the two-stage least squares (2SLS) estimator and asymptotic Neyman inference, i.e., the t-test. The paper suggests a Fisher Randomization Test based on the t-test statistic as an alternative to the Neyman inference. Based on the setup with a randomized experiment with noncompliance, for which one can identify the LATE, I compare the two approaches in a Monte Carlo (MC) simulations. The results from the MC simulation is that the Fisher randomization test is not a valid alternative to the Neyman’s test as it has too low power.