教学课程资源库 更多>>

关键词： Biostatistics   Statistical physics   Computer science  

摘要： The essence of Bayesian data analysis is to ascertain posterior distributions. Posteriors generally do not have closed-form expressions for direct computation in practical applications. Analysts, therefore, resort to Markov Chain Monte Carlo (MCMC) methods for the generation of sample observations that approximate the desired posterior distribution. Standard MCMC methods simulate sample values from the desired posterior distribution via random proposals. As a result, the mechanism used to generate the proposals inevitably determines the efficiency of the algorithm. One of the modern MCMC techniques designed to explore the high-dimensional space more efficiently is Hamiltonian Monte Carlo (HMC), based on the Hamiltonian differential equations. Inspired by classical mechanics, these equations incorporate a latent variable to generate MCMC proposals that are likely to be accepted. This dissertation discusses how such a powerful computational approach can be used for implementing statistical models. Along this line, I created a unified computational procedure for using HMC to fit various types of statistical models. The procedure that I proposed can be applied to a broad class of models, including linear models, generalized linear models, mixed-effects models, and various types of semiparametric regression models. To facilitate the fitting of a diverse set of models, I incorporated new parameterization and decomposition schemes to ensure the numerical performance of Bayesian model fitting without sacrificing the procedure's general applicability. As a concrete application, I demonstrate how to use the proposed procedure to fit a multivariate generalized additive model (GAM), a nonstandard statistical model with a complex covariance structure and numerous parameters. Byproducts of the research include two software packages that all practical data analysts to use the proposed computational method to fit their own models. The research's main methodological contribution is

关键词： Industrial engineering   Operations research   Computer science   Mechanical engineering  

摘要： Immediately after a disaster, humanitarian efforts are hindered by limited logistical resources and infrastructure failures. To address these challenges, we propose several decision frameworks for scheduling and routing of trucks and uncrewed aerial vehicles (UAVs or drones). Through numerical analysis, we provide managerial insights into the impacts of disaster-specific geographical and system parameters on operational performance metrics. In the first framework, we consider the decision problem of replenishing warehouses with emergency supplies following a disaster. We define this problem as a variant of truckload open vehicle routing problems and present two alternative formulations. We designed two column-generation-based solution approaches and several path generation algorithms to solve the resulting models. From numerical experiments, we find that the proposed solution approaches outperform commercial solver in terms of solution time. Due to post-disaster infrastructure failures, obtaining precise demand information and meeting demands using only trucks become all but impossible. To that end, in our second framework, we consider a two-echelon vehicle routing problem, where we assume that trucks transport two sets of UAVs and stop at predefined (satellite) locations to work as UAV launching pads. The first set of dispatched UAVs provide temporary telecommunication signals and obtain demand information, while the second set of UAVs make deliveries to satisfy the demand. We present a two-stage decision approach and design column-generation-based decomposition heuristics to solve the resulting models. From numerical experiments, we find that the proposed approach significantly reduces the solution time with a slight reduction in solution quality. Using a case study with post-disaster scenarios in Puerto Rico, we evaluate the changes in the model outcomes with the variation in several system parameters. In the third framework, we consider a two-echelon vehicle rou

关键词： Computer Engineering   Computer science   Artificial intelligence  

摘要： Wireless communications systems design and its performance depend on the wireless fading channels, which are often characterized using a Rician or Nakagami probability density function. A Rician or Nakagami parameter describes the fading severity in a fading channel and is used in the system design and performance evaluation. Therefore, the estimation of the fading parameter is important in wireless communications research and development. Traditionally, a fading parameter equation, maximum likelihood estimation, or Moment-based estimation is used to estimate the parameters. This work explores the use of deep learning for K and m parameter estimation. The signals are generated and used to train a convolutional neural network (CNN) to estimate the Rician K-factor and Nakagami-m parameter from a waveform signal in a fading channel. In the first part, we generate random channels of different Rician K-factors with different Gaussian noise levels. Then, we use a convolutional neural network model to estimate the Rician K-factor. Several scenarios are considered, signals with different signal-to-noise ratios (SNR), dataset sizes, time window sizes, and LTE signal. In the Nakagami-m parameter estimation part, we generate different waveform signals with different parameters of Nakagami-m. Next, we classify the Nakagami-m parameters using the convolutional neural network. We estimate the Nakagami-m parameters with different SNR levels, dataset sizes, LTE signal, and time window sizes. In the final part, we generate signals of different levels of signal-to-noise ratio with the Rayleigh channel. Moreover, we identify the different levels of SNR using the CNN model. We study SNR levels with different dataset sizes, time window sizes, and LTE signal. Numerical results in all three parts demonstrate its good performance in estimating the Rician K-factors, Nakagami-m parameters, and different SNR levels.

关键词： Hispanic-Serving Institutions   Latinx students   STEM   Computer science  

摘要： Computer science remains one of the least diverse disciplines within STEM. As society becomes more dependent on technology, there should be more diverse representation in the people producing technology. Latinx are underrepresented in the technology sector. In order to understand the lack of diversity in the technology sector it is important to look at diversity efforts in computer science programs at higher education institutions. The purpose of this study was to understand the experiences of Latinx students in computer science programs at Hispanic-Serving Institutions. Focusing on Hispanic-Serving institutions is important given that they educate a large percentage of Latinx college students. The research question guiding this study was: What barriers and supports do Latinx students in computer science experience at Hispanic-Serving Institutions? Data for this study were collected through semi-structured interviews with twelve students from three higher education institutions. The analysis revealed four themes: Navigating College as a First-Generation College Student, Disconnections in Computer Science/Feelings of Inadequacy, On-Campus Relationships and Professional Development for Computer Science Students. This study illustrates how faculty, staff, and peers can best support Latinx students in computer science. The knowledge students shared in this study provides insights into how to create a more inclusive environment in STEM that will promote the success of Latinx computer science students.

关键词： Sonar   Unmanned aerial vehicles   Software   Sensors   Shorelines   Aerospace engineering   Computer science   Robotics  

摘要： Bathymetric mapping is an important tool for reservoir management, typically completed before reservoir construction. Historically, bathymetric maps were produced by interpolating between points measured at a relatively large spacing throughout a reservoir, typically on the order of a few, up to 10, meters or more depending on the size of the reservoir. These measurements were made using traditional survey methods before the reservoir was filled, or using sonar surveys after filling. Post-construction issues such as sedimentation and erosion can change a reservoir, but generating updated bathymetric maps is difficult as the areas of interest are typically in the sediment deltas and other difficult-to-access areas that are often above water or exposed for part of the year. We present a method to create complete reservoir bathymetric maps, including areas above the water line, using small unmanned aerial vehicle (sUAV) photogrammetry combined with multi-beam sonar data--both established methods for producing topographic models. This thesis presents methods to create accurate above-water shoreline models using images from sUAVs, processed using a commercial software package and a method to accurately knit sonar and Structure from Motion (SfM) data sets by matching slopes. The models generated by both approaches are point clouds, which consist of points representing the ground surface in three-dimensional space. Generating models from sUAV-captured images requires ground control points (GCPs), i.e., points with a known location, to anchor model creation. We explored issues with ground control spacing, masking water regions (or omitting water regions) in the images, using no GCPs, and incorrectly tagging a GCP. To quantify the effect these issues had on model accuracy, we computed the difference between generated clouds and a reference point cloud to determine the point cloud error. We found that the time required to place GCPs was significantly more than the time requir

关键词： Software engineering   Computer science   Engineering  

摘要： Context: Heterogeneous computing, i.e., computing performed on processors of different types - such as combination of CPUs and GPUs, or CPUs and FPGAs - has shown to be a feasible path towards higher performance and less energy consumption. However, this approach imposes a number of challenges on the software side that must be addressed in order to achieve the aforementioned advantages. Objective: The objective of this thesis is to improve the process of software deployment on heterogeneous platforms. Through a detailed analysis of the state-of-the-art and state-of-the-practice, we aim to provide a reasoning framework for engineers to migrate software to be executed on such platforms. Method: To achieve our goal, we conducted: (i) a literature review in the form of a systematic mapping study on software deployment on heterogeneous platforms; (ii) a multiple case study in industry that highlights the main challenges and concerns in the state-of-the-practice in the area; and (iii) a study in which we propose and evaluate a decision framework to guide engineers in migrating software for execution on heterogeneous platforms, with a case study in the automotive domain. Results: In the mapping study, we provided a thorough classification of the identified concerns and approaches to deploying software on heterogeneous platforms. Among other findings, we discovered a lack of holistic approaches that include development processes, as well as few validation studies in industrial contexts. In the second study, we discovered and analyzed common practices and challenges that companies face when using heterogeneous platforms. One of such challenges is related to the lack of approaches that cover the software development lifecycle. In the third study, we proposed a decision framework that guides engineers in the process of reasoning for migrating software for execution on heterogeneous platforms. It consists of five stages (assessing, re-architecting, developing, deploying, evalua

关键词： Computer science   Artificial intelligence  

关键词： Computer science   Biomedical engineering   Electrical engineering  

关键词： Computer science   Engineering   Mathematics  

摘要： Deep learning methods have achieved great successes in recent years for excellently performing many tasks which humans excel at like speech and pattern recognition. However, many machine learning-based models, including neural-network models, are found vulnerable to different kinds of attacks which can compromise the systems in which they are used for classification or other kinds of decision-making tasks. The existence of these kinds of attacks reveals many vulnerabilities of systems based on machine-learning models and discovering them also helps us understand neural networks in better ways. Furthermore, current research in defending these kinds of attacks is still in its infancy. This work presents a study of different attacks and defense techniques used for deep learning models that may be used in building many life and safety-critical systems. We provide a brief mathematical background of the algorithms needed in performing the attacks and understand their nature and survey defense mechanisms to make neural networks robust to those kinds of attacks. The report mainly focuses on the deep-learning paradigm of machine learning which primarily concerns with deep neural networks. This study aims to shed light on the counter-intuitive nature of deep-learning models through the help of various works done in the current literature.

关键词： Electrical engineering   Computer science   Systems science  

摘要： Data traffic demands are increasing annually. The networks need to respond to these increased demands. To do end, there are a few directions for the network operators or researchers: first is to increase the size of the networks or increasing the equipment. This direction in some cases is inevitable, however it some aspects of this strategy are inconsistent with current user expectations. Using more equipment increases the costs of the network and puts extra cost on the end users and the networks operators. Also, it does not guarantee any improvement in the quality of service which is another major expectation of the end users. A second approach being undertaken by the research community is to invent more efficient equipment such as, less noisy amplifiers, optical wavelength convertors, etc. This has been of the major direction to have better networks. Finally, adjusting the communication and networking protocols and standards have always been one of the major considerations because of its low cost and effectiveness. Optical networks currently form the major parts of the current communication infrastructures. Especially in tier one and tier 2 networks. Wavelength division multiplexing and Elastic Optical Networks are two existing Networking framework for optical networks. In this research by focusing on the Network layer, we review the major algorithms existing in these networks, and then some new strategies to overcome their main issues are presented along with experimental results. The main challenge behind resource allocation for different traffic demands is called routing , modulation level and spectrum allocation (RMSA). This problem introduces more flexibilities compared to WDM similar problem. At the same time, this higher flexibility increases the complexity of any algorithm to solve RMSA. In this research, first we proposed a sub-tree model to serve multicast demands. In this model we assume the flexibility of choosing routing, spectrum and modulation level