DATAWorks Sessions Archive

Recent research shows the effectiveness of machine learning on image classification and segmentation. The use of artificial neural networks (ANNs) on image datasets such as the MNIST dataset of handwritten digits is highly effective. However, when presented with a more complex image, ANNs and other simple computer vision algorithms tend to fail. This research uses Convolutional Neural Networks (CNNs) to determine how we can differentiate between ice and clouds in the imagery of the Arctic. Instead of using ANNs, where we analyze the problem in one dimension, CNNs identify features using the spatial relationships between the pixels in an image. This technique allows us to extract spatial features, presenting us with higher accuracy. Using a CNN named the Cloud-Net Model, we analyze how a CNN performs when analyzing satellite images. First, we examine recent research on the Cloud-Net Model’s effectiveness on satellite imagery, specifically from Landsat data, with four channels: red, green, blue, and infrared. We extend and modify this model, allowing us to analyze data from the most common channels used by satellites: red, green, and blue. By training on different combinations of these three channels, we extend this analysis by testing on an entirely different data set: GOES imagery. This gives us an understanding of the impact of each individual channel in image classification. By selecting images that exist in the same geographic location and containing both ice and clouds, such as the Landsat, we test GOES analyzing the CNN’s generalizability. Finally, we present CNN’s ability to accurately identify the clouds and ice in the GOES data versus the Landsat data.

AI and machine learning have become widespread throughout the defense, government, and commercial sectors. This has led to increased attention on the topic of trust and the role it plays in successfully integrating AI into highconsequence environments where tolerance for risk is low. Driven by recent successes of AI algorithms in a range of applications, users and organizations rely on AI to provide new, faster, and more adaptive capabilities. However, along with those successes have come notable pitfalls, such as bias, vulnerability to adversarial attack, and inability to perform as expected in novel environments. Many types of AI are data-driven, meaning they operate on and learn their internal models directly from data. Therefore, tracking how data were used to build data properties (e.g., training, validation, and testing) is crucial not only to ensure a high-performing model, but also to understand if the AI should be trusted. MLOps, an offshoot of DevSecOps, is a set of best practices meant to standardize and streamline the end-to-end lifecycle of machine learning. In addition to supporting the software development and hardware requirements of AI-based systems, MLOps provides a scaffold by which the attributes of trust can be formally and methodically evaluated. Additionally, MLOps encourages reasoning about trust early and often in the development cycle. To this end, we present a framework that encourages the development of AI-based applications that can be trusted to operate as intended and function safely both with and without human interaction. This framework offers guidance for each phase of the AI lifecycle, utilizing MLOps, through a detailed discussion of pitfalls resulting from not considering trust, metrics for measuring attributes of trust, and mitigations strategies for when risk tolerance is low.

Software processes define specific sequences of activities performed to effectively produce software, whereas tools provide concrete computational artifacts by which these processes are carried out. Tool independent modeling of processes and related practices enable quantitative assessment of software and competing approaches. This talk presents a framework to assess an approach employed in modern software development known as staged rollout, which releases new or updated software features to a fraction of the user base in order to accelerate defect discovery without imposing the possibility of failure on all users. The framework quantifies process metrics such as delivery time and product metrics, including reliability, availability, security, and safety, enabling tradeoff analysis to objectively assess the quality of software produced by vendors, establish baselines, and guide process and product improvement. Failure data collected during software testing is employed to emulate the approach as if the project were ongoing. The underlying problem is to identify a policy that decides when to perform various stages of rollout based on the software’s failure intensity. The illustrations examine how alternative policies impose tradeoffs between two or more of the process and product metrics.

Warhead design and performance optimization against a range of targets is a foundational aspect of the Department of the Army’s mission on behalf of the warfighter. The existing procedures utilized to perform this basic design task do not fully leverage the exponential growth in data science, machine learning, distributed computing, and computational optimization. Although sound in practice and methodology, existing implementations are laborious and computationally expensive, thus limiting the ability to fully explore the trade space of all potentially viable solutions. An additional complicating factor is the fast paced nature of many Research and Development programs which require equally fast paced conceptualization and assessment of warhead designs. By utilizing methods to take advantage of data analytics, the workflow to develop and assess modern warheads will enable earlier insights, discovery through advanced visualization, and optimal integration of multiple engineering domains. Additionally, a framework built on machine learning would allow for the exploitation of past studies and designs to better inform future developments. Combining these approaches will allow for rapid conceptualization and assessment of new and novel warhead designs. US overmatch capability is quickly eroding across many tactical and operational weapon platforms. Traditional incremental improvement approaches are no longer generating appreciable performance improvements to warrant investment. Novel next generation techniques are required to find efficiencies in designs and leap forward technologies to maintain US superiority. The proposed approach seeks to shift existing design mentality to meet this challenge.

For organizations to make data-driven decisions, they must be able to understand and organize their mission critical data.  Recently, the DoD, NASA and other federal agencies have declared their intention to become “data-centric” organizations, but transitioning from an existing mode of operation and architecture can be challenging.  Moreover, the DoD is pushing for artificial intelligence enabled systems (AIES) and wide scale digital transformation.  These concepts in the abstract seem straightforward, but because they can only evolve when people, processes, and technology change together, they have proven challenging in execution.  Since the structure and quality of an organization’s data limits what an organization can do with that data it is imperative to get data processes right before embarking on other initiatives that depend on quality data. Despite the importance of data quality, many organizations treat data architecture as an emergent phenomenon and not something to be planned or thought through holistically. In this discussion, panelists will explore what it means to be data-centric, what a data-centric architecture is, how it is different from the other data architectures, why an organization might prefer a data-centric approach, and the challenges associated with becoming data-centric.

STAT practitioners often find ourselves outsiders to the programs we assist. This session presents a case study that demonstrates some of the obstacles in communication of capabilities, purpose, and expectations that may arise due to approaching the project externally. Incremental value may open the door to greater collaboration in the future, and this presentation discusses potential solutions to provide greater benefit to testing programs in the face of obstacles that arise due to coming from outside the program team. DISTRIBUTION STATEMENT A. Approved for public release; distribution is unlimited. CLEARED on 5 Jan 2022. Case Number: 88ABW-2022-0002

The 75th Ranger Regiment is an elite Army Unit responsible for some of the most physically and mentally challenging missions. Entry to the unit is based on an assessment process called Ranger Regiment Assessment and Selection (RASP), which consists of a variety of tests and challenges of strength, intellect, and grit. This study explores the psychological and physical profiles of candidates who attempt to pass RASP. Using a Random Forest Artificial Intelligence model, and a penalized logistic regression model, we identify initial entry characteristics that are predictive of success in RASP. We focus on the differences between racial sub-groups and military occupational specialties (MOS) sub-groups to provide information for recruiters to identify underrepresented groups who are likely to succeed into the selection process.

The number of people using autonomous systems for everyday tasks has increased steadily since the 1960s and has dramatically increased with the invention of smart devices that can be controlled via smartphone. Within the defense community, automated systems are currently used to perform search and rescue missions and to assume control of aircraft to avoid ground collision. Until recently, researchers have only been able to gain insights on trust levels by observing a human’s reliance on the system, so it was apparent that researchers needed a validated method of quantifying how much an individual trusts the automated system they are using. IDA researchers developed the Trust of Automated Systems Test (TOAST scale) to serve as a validated scale capable of measuring how much an individual trusts a system. This presentation will outline the nine item TOAST scale’s understanding and performance elements, and how it can effectively be used in a defense setting. We believe that this scale should be used to evaluate the trust level of any human using any system, including predicting when operators will misuse or disuse complex, automated and autonomous systems.

Host-based sensors are standard tools for generating event data to detect malicious activity on a network. There is often interest in detecting activity using as few event classes as possible in order to minimize host processing slowdowns. Using DARPA’s Operationally Transparent Cyber (OpTC) Data Release, we consider the problem of detecting malicious activity using event counts aggregated over five-minute windows. Event counts are categorized by eleven features according to MITRE CAR data model objects. In the supervised setting, we use regression trees with all features to show that malicious activity can be detected at above a 90% true positive rate with a negligible false positive rate. Using forward and exhaustive search techniques, we show the same performance can be obtained using a sparse model with only three features. In the unsupervised setting, we show that the isolation forest algorithm is somewhat successful at detecting malicious activity, and that a sparse three-feature model performs comparably. Finally, we consider various search criteria for identifying sparse models and demonstrate that the RMSE criteria is generally optimal.

Control charts are often used to monitor the quality characteristics of a process over time to ensure undesirable behavior is quickly detected. The escalating complexity of processes we wish to monitor spurs the need for more flexible control charts such as those used in profile monitoring. Additionally, designing a control chart that has an acceptable false alarm rate for a practitioner is a common challenge. Alarm fatigue can occur if the sampling rate is high (say, once a millisecond) and the control chart is calibrated to an average in-control run length (ARL0) of 200 or 370 which is often done in the literature. As alarm fatigue may not just be annoyance but result in detrimental effects to the quality of the product, control chart designers should seek to minimize the false alarm rate. Unfortunately, reducing the false alarm rate typically comes at the cost of detection delay or average out-of-control run length (ARL1). Motivated by recent work on eigenvector perturbation theory, we develop a computationally fast control chart called the Eigenvector Perturbation Control Chart for nonparametric profile monitoring. The control chart monitors the l_2 perturbation of the leading eigenvector of a correlation matrix and requires only a sample of known in-control profiles to determine control limits. Through a simulation study we demonstrate that it is able to outperform its competition by achieving an ARL1 close to or equal to 1 even when the control limits result in a large ARL0 on the order of 10^6. Additionally, non-zero false alarm rates with a change point after 10^4 in-control observations were only observed in scenarios that are either pathological or truly difficult for a correlation based monitoring scheme.

Physical experiments in the national security arena, including nuclear deterrence, are often expensive and time-consuming resulting in small sample sizes which make it difficult to achieve desired statistical properties. Bayesian adaptive design of experiments (BADE) is a sequential design of experiment approach which updates the test design in real time, in order to optimally collect data. BADE recommends ending experiments early by either concluding that the experiment would have ended in efficacy or futility, had the testing completely finished, with sufficiently high probability. This is done by using data already collected and marginalizing over the remaining uncollected data and updating the Bayesian posterior distribution in near real-time. BADE has seen successes in clinical trials, resulting in quicker and more effective assessments of drug trials while also reducing ethical concerns. BADE has typically only been used in futility studies rather than efficacy studies for clinical trials, although there hasn’t been much debate for this current paradigm. BADE has been proposed for testing in the national security space for similar reasons of quicker and cheaper test series. Given the high-consequence nature of the tests performed in the national security space, a strong understanding of new methods is required before being deployed. The main contribution of this research was to reproduce results seen in previous studies, for different aspects of model performance. A large simulation inspired by a real testing problem at Sandia National Laboratories was performed to understand the behavior of BADE under various scenarios, including shifts to mean, standard deviation, and distributional family, all in addition to the presence of outliers. The results help explain the behavior of BADE under various assumption violations. Using the results of this simulation, combined with previous work related to BADE in this field, it is argued this approach could be used as part of an “evidence package” for deciding to stop testing early due to futility, or with stronger evidence, efficacy. The combination of expert knowledge with statistical quantification provides the stronger evidence necessary for a method in its infancy in a high-consequence, new application area such as national security. Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.

This paper presents an analysis of target location error (TLE) based on the Cox Ingersoll Ross (CIR) model. In brief, this model characterizes TLE as a function of range based the stochastic differential equation model dX(r) = a(b-X(r))dr + sigma *sqrt(X(r)) dW(r) where X(t) is TLE at range r, b is the long-term mean (terminal) of the TLE, a is the rate of reversion of X(r) to b, sigma is the process volatility, and W(t) is the standard Weiner process. Multiple flight test runs under the same conditions exhibit different realizations of the TLE process. This approach to TLE analysis models each flight test run as a realization the CIR process. Fitting a CIR model to multiple data runs then provides a characterization of the TLE system under test. This paper presents an example use of the CIR model. Maximum likelihood estimates of the parameters of the CIR model are found from a collection of TLE data runs. The resulting CIR model is then used to characterize overall system TLE performance as a function of range to the target as well as the asymptotic estimate of long-term TLE.

Designing a Bayesian assurance test plan requires choosing a test plan that guarantees a product of interest is good enough to satisfy consumer’s criteria but not ‘so good’ that it causes producer’s concern if they fail the test. Bayesian assurance tests are especially useful because they can incorporate previous product information in the test planning and explicitly control levels of risk for the consumer and producer. We demonstrate an algorithm for efficiently computing a test plan given desired levels of risks in binomial and exponential testing. Numerical comparisons with the Operational Characteristic (OC) curve, Probability Ratio Sequential Test (PRST), and a simulation-based Bayesian sample size determination approach are also considered.

Modern data analysis is typically quite computational. Correspondingly, sharing scientific and statistical work now often means sharing code and data in addition writing papers and giving talks. This type of code sharing faces several challenges. For example, it is often difficult to take code from one computer and run it on another due to software configuration, version, and dependency issues. Even if the code runs, writing code that is easy to understand or interact with can be difficult. This makes it difficult to assess third-party code and its findings, for example, in a review process. In this talk we describe a combination of two computing technologies that help make analyses shareable, interactive, and completely reproducible. These technologies are (1) analysis containerization, which leverages virtualization to fully encapsulate analysis, data, code and dependencies into an interactive and shareable format, and (2) code notebooks, a literate programming format for interacting with analyses. This talks reviews both the problems at the high-level and also provides concrete solutions to the challenges faced. In addition to discussing reproducibility and data/code sharing generally, we will touch upon several such issues that arise specifically in the defense and aerospace communities.

In this talk we’ll focus on exploring the motivation for using cloud computing for Computational Fluid Dynamics (CFD) for Federal Government Test & Evaluation. Using examples from automotive, aerospace and manufacturing we’ll look at benchmarks for a number of CFD codes using CPUs (x86 & Arm) and GPUs and we’ll look at how the development of high-fidelity CFD e.g. WMLES, HRLES, is accelerating the need for access to large scale HPC. The onset of COVID-19 has also meant a large increase in the need for remote visualization with greater numbers of researchers and engineering needing to work from home. This has also accelerated the adoption of the same approaches needed towards the pre- and post-processing of peta/exa-scale CFD simulation and we’ll look at how these are more easily accessed via a cloud infrastructure. Finally, we’ll explore perspectives on integrating ML/AI into CFD workflows using data lakes from a range of sources and where the next decade may take us.

A strength of the Bayesian paradigm is that it allows for the explicit use of all available information—to include subject matter expert (SME) opinion and previous (possibly dissimilar) data. While frequentists are constrained to only including data in an analysis (that is to say, only including information that can be observed), Bayesians can easily consider both data and SME opinion, or any other related information that could be constructed. This can be accomplished through the development and use of priors. When prior development is done well, a Bayesian analysis will not only lead to more direct probabilistic statements about system performance, but can result in smaller standard errors around fitted values when compared to a frequentist approach. Furthermore, by quantifying the uncertainty surrounding a model parameter, through the construct of a prior, Bayesians are able to capture the uncertainty across a test space of consideration. This presentation develops a framework for thinking about how different priors can be used throughout the continuum of testing. In addition to types of priors, how priors can change or evolve across the continuum of testing—especially when a system changes (e.g., is modified or adjusted) during phases of testing—will be addressed. Priors that strive to provide no information (reference priors) will be discussed, and will build up to priors that contain available information (informative priors). Informative priors—both those based on institutional knowledge or summaries from databases, as well as those developed based on previous testing data—will be discussed, with a focus on how to consider previous data that is dissimilar in some way, relative to the current test event. What priors might be more common in various phases of testing, types of information that can be used in priors, and how priors evolve as information accumulates will all be discussed.

In the acquisition and testing world, data analysts repeatedly encounter certain categories of data, such as time or distance until an event (e.g., failure, alert, detection), binary outcomes (e.g., success/failure, hit/miss), and survey responses. Analysts need tools that enable them to produce quality and timely analyses of the data they acquire during testing. This poster presents four web-based apps that can analyze these types of data. The apps are designed to assist analysts and researchers with simple repeatable analysis tasks, such as building summary tables and plots for reports or briefings. Using software tools like these apps can increase reproducibility of results, timeliness of analysis and reporting, attractiveness and standardization of aesthetics in figures, and accuracy of results. The first app models reliability of a system or component by fitting parametric statistical distributions to time-to-failure data. The second app fits a logistic regression model to binary data with one or two independent continuous variables as predictors. The third calculates summary statistics and produces plots of groups of Likert-scale survey question responses. The fourth calculates the system usability scale (SUS) scores for SUS survey responses and enables the app user to plot scores versus an independent variable. These apps are available for public use on the Test Science Interactive Tools webpage https://new.testscience.org/interactive-tools/.