In this work, we propose and investigate a series of methods to predict stock market movements. These methods use stock market technical and macroeconomic indicators as inputs into different machine learning classifiers. The objective is to survey existing domain knowledge, and combine multiple techniques into one method to predict daily market movements for stocks. Approaches using nearest neighbor classification, support vector machine classification, K-means classification, principal component analysis and genetic algorithms for feature reduction and redefining the classification rule were explored. Ten stocks, 9 companies and 1 index, were used to evaluate each iteration of the trading method. The classification rate, modified Sharpe ratio and profit gained over the test period is used to evaluate each strategy. The findings showed nearest neighbor classification using genetic algorithm input feature reduction produced the best results, achieving higher profits than buy-and-hold for a majority of the companies.

Hardware implementations of artificial neural networks (ANNs) have become feasible due to the advent of persistent 2-terminal devices such as memristor, phase change memory, MTJs, etc. Hybrid memristor crossbar/CMOS systems have been studied extensively and demonstrated experimentally. In these circuits, memristors located at each cross point in a crossbar are, however, stacked on top of CMOS circuits using back end of line processing (BOEL), limiting scaling. Each neuron’s functionality is spread across layers of CMOS and memristor crossbar and thus cannot support the required connectivity to implement large-scale multi-layered ANNs. This work proposes a new fine-grained 3D integrated circuit technology for ANNs that is one of the first IC technologies for this purpose. Synaptic weights implemented with devices are incorporated in a uniform vertical nanowire template co-locating the memory and computation requirements of ANNs within each neuron. Novel 3D routing features are used for interconnections in all three dimensions between the devices enabling high connectivity without the need for special pins or metal vias. To demonstrate the proof of concept of this fabric, classification of binary images using a perceptron-based feed forward neural network is shown. Bottom-up evaluations for the proposed fabric considering 3D implementation of fabric components reveal up to 19x density, 1.2x power benefits when compared to 16nm hybrid memristor/CMOS technology.

At Amgen, the Automated Vision Inspection (AVI) systems capture the movement of unwanted particles in Amgen's drug product containers. For quality inspection, the AVI system must detect these undesired particles using a high speed spin-stop agitation process. To better understand the fluid movements to swirl the particles away from the walls, Computational Fluid Dynamics (CFD) is used to analyze the nature of the two phase flow of air and a liquid solution. Several 2-D and 3-D models were developed using Fluent to create simulations of Amgen's drug product containers for a 1 mL syringe, 2.25 mL syringe, and a 5 mL cartridge. Fluid motion and potential bubble formations were studied within the liquid/gas domain inside the container by varying parameters such as viscosity, angular velocity, and surface tension. Experiments were conducted using Amgen's own equipment to capture the images of the spin-stop process and validate the models created in Fluent. Observations were made to see the effects of bubble formation or splashing during spin-down to rest. The numerical and experimental results showed favorable comparison when measuring the meniscus height or the surface profile between the air and liquid. Also, at high angular velocity and dynamic viscosity, the container experiences instabilities and bubble formations. These studies indicate that CFD can be used as an useful and important tool to study fluid movement during agitation and observe any undesirable results for quality inspection.

With recent advances in technology, Vehicular Ad-hoc Networks (VANETs) have grown in application. One of these areas of application is Vehicle Safety Communication (VSC) technology. VSC technology allows for vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications that enhance vehicle safety and driving experience. However, these newly developing technologies bring with them a concern for the vehicular privacy of drivers. Vehicles already employ the use of pseudonyms, unique identifiers used with signal messages for a limited period of time, to prevent long term tracking. But can attackers still attack vehicular privacy even when vehicles employ a pseudonym change strategy? The major contribution of this paper is a new attack model that uses long-distance pseudonym changing and short-distance non-changing protocols to associate vehicles with their respective pseudonyms.

Quantitative structure-activity relationship (QSAR) models based on machine learning algorithms are powerful tools to expedite drug discovery processes and therapeutics development. Given the cost in acquiring large-sized training datasets, it is useful to examine if QSAR analysis can reasonably predict drug activity with only a small-sized dataset (size < 100) and benchmark these small-dataset QSAR models in application-specific studies. To this end, here we present a systematic benchmarking study on small-dataset QSAR models built for prediction of effective Wnt signaling inhibitors, which are essential to therapeutics development in prevalent human diseases (e.g., cancer). Specifically, we examined a total of 72 two-dimensional (2D) QSAR models based on 4 best-performing algorithms, 6 commonly used molecular fingerprints, and 3 typical fingerprint lengths. We trained these models using a training dataset (56 compounds), benchmarked their performance on 4 figures-of-merit (FOMs), and examined their prediction accuracy using an external validation dataset (14 compounds). Our data show that the model performance is maximized when: 1) molecular fingerprints are selected to provide sufficient, unique, and not overly detailed representations of the chemical structures of drug compounds; 2) algorithms are selected to reduce the number of false predictions due to class imbalance in the dataset; and 3) models are selected to reach balanced performance on all 4 FOMs. These results may provide general guidelines in developing high-performance small-dataset QSAR models for drug activity prediction.

<p>Molecular dynamics (MD) simulations performed on two-phase simulation cells were used to compute the Austenite (FCC) / Ferrite (BCC) boundary mobility in pure iron (Fe) over the temperature range of 600K - 1400K. An embedded atom method interatomic potential was used to model Fe and the driving force for interface motion is the free energy difference between the two phases, which was computed as a function of temperature using a thermodynamic integration technique. For low index FCC/BCC crystallographic orientations, no interface motion was observed. But for slight misorientations steps were introduced at the interphase and sufficient mobility was observed over MD time scales. A new interphase mechanism was found that instead of the moving of structure disconnection by diffusion control, growing of misfit dislocations in each steps were observed (interphase control). The interphase velocity could reach 2 m/s and the mobility at 1000K was approximately 0.001 mol-m/J-s. In agreement with previous MD studies of grain boundary mobility, we found that the activation energy for the austenite-ferrite boundary mobility was much lower than the values found from previous experiments.</p> / Master of Applied Science (MASc)

<p>Today's highly integrated supply chains are exposed to various types of risks which disrupt the normal flow of goods or services within a supply chain network. Since most of these individual risks are interconnected, a mitigation strategy to tackle one risk may result in the exacerbation of another.</p> <p>Risk dependencies have been modelled using two approaches in the financial insurance literature : (i) random variables, and (ii) copulas. In this dissertation these studies are reviewed and extended. Also, applications for these models for different supply chain network configurations are presented. Then, a Poisson process model for risk propagation is proposed. Unlike the existing models, the transition rate of the proposed model not only expresses the time dependency, but also captures other possible dependencies in the network. Finally, the thesis is summarized and general directions and suggestions for future research on risk dependency and propagation modelling are provided.</p> / Master of Science (MSc)

The direct and adjoint methods are to linearize the time-averaged solution of bounded dynamical systems about one or more design parameters. Hence, such methods are one way to obtain the gradient necessary in locally optimizing a dynamical system’s time-averaged behavior over those design parameters. However, when analyzing nonlinear systems whose solutions exhibit chaos, standard direct and adjoint sensitivity methods yield meaningless results due to time-local instability of the system. The present work proposes a new method of solving the direct and adjoint linear systems in time, then tests that method’s ability to solve instances of the Lorenz system that exhibit chaotic behavior. Promising results emerge and are presented in the form of a regression analysis across a parametric study of the Lorenz system.

Recent advances in cloud-based big-data technologies now makes data driven solutions feasible for increasing numbers of scientific computing applications. One such data driven solution approach is machine learning where patterns in large data sets are brought to the surface by finding complex mathematical relationships within the data. Nowcasting or short-term prediction of rainfall in a given region is an important problem in meteorology. In this thesis we explore the nowcasting problem through a data driven approach by formulating it as a machine learning problem. State-of-the-art nowcasting systems today are based on numerical models which describe the physical processes leading to precipitation or on weather radar extrapolation techniques that predict future radar precipitation maps by advecting from a sequence of past maps. These techniques, while they can perform well over very short prediction horizons (minutes) or very long horizons (hours to days), tend not to perform well over medium horizons (1-2 hours) due to lack of input data at the necessary spatial and temporal scales for the numerical prediction methods or due to the inability of radar extrapolation methods to predict storm growth and decay. Given that water must first concentrate in the atmosphere as water vapor before it can fall to the ground as rain, one goal of this thesis is to understand if water vapor information can improve radar extrapolation techniques by giving the information needed to infer growth and decay. To do so, we use the GPS-Meteorology technique to measure the water vapor in the atmosphere and weather radar reflectivity to measure rainfall. By training a machine learning nowcasting algorithm using both variables and comparing its performance against a nowcasting algorithm trained on reflectivity alone, we draw conclusions as to the predictive power of adding water vapor information. Another goal of this thesis is to compare different machine learning techniques, viz., the random forest ensemble learning technique, which has shown success on a number of other weather prediction problems, and the current state-of-the-art machine learning technique for images and image sequences, convolutional neural network (CNN). We compare these in terms of problem representation, training complexity, and nowcasting performance. A final goal is to compare the nowcasting performance of our machine learning techniques against published results for current state-of-the-art model based nowcasting techniques.

Die vollständige Ausnutzung des Leichtbaupotentials bei der Dimensionierung von mehrschichtigen endlosfaserverstärkten Strukturbauteilen erfordert die Bereitstellung von geeigneten Optimierungswerkzeugen, da bei der Auslegung eine große Anzahl von Entwurfsvariablen zu berücksichtigen sind. In dieser Arbeit werden Optimierungsalgorithmen und -strategien zur Lösung wissenschaftlicher Fragestellungen für industrielle Anwendungen bei der Konstruktion von entsprechenden Faserkunststoffverbunden entwickelt und bewertet. Um das breite Anwendungsspektrum aufzuzeigen, werden drei unterschiedliche repräsentative Problemstellungen bearbeitet. Dabei wird für Mehrschichtverbunde die Festigkeitsoptimierung hinsichtlich eines bruchtypbezogenen Versagenskriteriums vorgenommen, ein Dämpfungsmodell zur Materialcharakterisierung entworfen sowie eine bivalente Optimierungsstrategie zur Auslegung von gewickelten Hochdruckbehältern erstellt. Die Grundlage der entwickelten Methoden bilden dabei jeweils stochastische naturanaloge Optimierungsheuristiken, da die betrachteten Aufgabenstellungen nicht konvex sind und derartige Verfahren flexibel eingesetzt werden können. / The full utilization of the light weight potential in the dimensioning of multilayer fiber reinforced composites requires suitable optimization tools, since a large number of design variables has to be taken into account. In this work, optimization algorithms and strategies for the solution of scientific questions for industrial applications are developed and evaluated in the design of corresponding fiber-plastic composites. In order to show the wide range of applications, three different representative topics have been chosen. It will carry out a strength optimization for multilayer composites with regard to a type-related failure criterion, devolop a damping model for material characterization and established a bivalent optimization strategy for the design of wound high-pressure vessels. The developed methods are based on stochastic natural-analog optimization heuristics, since the considered tasks are not convex and such methods can be used in a very flexible manner.