DATAWorks Sessions Archive

Key outcomes for NASA's Quesst mission are noise dose and perceptual response data to inform regulators on their decisions regarding noise certification standards for the future of overland commercial supersonic flight. Dose-response curves are commonly utilized in community noise studies to describe the annoyance of a community to a particular noise source. The X-59 aircraft utilizes shaped-boom technology to demonstrate low noise supersonic flight. For X-59 community studies, the sound level from X-59 overflights constitutes the dose, while the response is an annoyance rating selected from a verbal scale, e.g., “slightly annoyed” and “very annoyed.” Dose-response data will be collected from individual flyovers (single event dose) and an overall response to the accumulation of single events at the end of the day (cumulative dose). There are quantifiable sources of error in the noise dose due to uncertainty in microphone measurements of the sonic thumps and uncertainty in predicted noise levels at survey participant locations. Assessing and accounting for error in the noise dose is essential to obtain an accurate dose-response model. There is also a potential for error in the perceptual response. This error is due to the ability of participants to provide their response in a timely manner and participant fatigue after responding to up to one hundred surveys over the course of a month. This talk outlines various challenges in estimating noise dose and perceptual response and the methods considered in preparation for X-59 community tests.

While the impact of local weather conditions on aircraft performance is well-documented, climate change has the potential to create long-term shifts in aircraft performance. Using just one metric, internal load capacity, we document operationally relevant performance changes for a UH-60L within the Indo-Pacific region. This presentation uses publicly available climate and aircraft performance data to create a representative analysis. The underlying methodology can be applied at varying geographic resolutions, timescales, airframes, and aircraft performance characteristics across the entire globe.

Ranking data are collected by presenting a respondent with a list of choices, and then asking what are the respondent's favorite, second favorite, and so on. The rankings may be complete, the respondent rank orders the complete list, or partial, only the respondent's favorite two or three, etc. Given a sample of rankings from a population, one goal may be to estimate the most favored choice from the population. Another may be to compare the preferences of one subpopulation to another. In this presentation I will introduce ranking data and probability models that form the foundation for statistical inference for them. The Plackette-Luce model will be the main focus. After that I will introduce a real data set containing ranking data assembled by the National Institute of Standards and Technology (NIST) based on the results of a national survey of first responders. The survey asked about how first responders use communication technology. With this data set, questions such as do rural and urban/suburban first responders prefer the same types communication devices, can be explored. I will conclude with some ideas for incorporating rankning data into test and evaluation settings.

The purpose of the Skyborg Data Pipeline is to allow for the rapid turnover of flight data collect during a test event, using collaborative easily access tool sets available in the AFOTEC Data Vault. Ultimately the goal of this data pipeline is to provide a working up to date dashboard that leadership can utilize shortly after a test event.

Post-hoc Uncertainty Quantification of Deep Learning Models Applied to Remote Sensing Image Scene Classification Steadily growing quantities of high-resolution UAV, aerial, and satellite imagery provide an exciting opportunity for global transparency and geographic profiling of activities of interest. Advances in deep learning, such as deep convolutional neural networks (CNNs) and transformer models, offer more efficient ways to exploit remote sensing imagery. Transformers, in particular, are capable of capturing contextual dependencies in the data. Accounting for context is important because activities of interest are often interdependent and reveal themselves in co-occurrence of related image objects or related signatures. However, while transformers and CNNs are powerful models, their predictions are often taken as point estimates, also known as pseudo probabilities, as they are computed by the softmax function. They do not provide information about how confident the model is in its predictions, which is important information in many mission-critical applications, and therefore limits their use in this space. Model evaluation metrics can provide information about the predictive model’s performance. We present and discuss results of post-hoc uncertainty quantification (UQ) of deep learning models, i.e., UQ application to trained models. We consider an application of CNN and transformer models to remote sensing image scene classification using satellite imagery, and compare confidence estimates of scene classification predictions of these models using evaluation metrics, such as expected calibration error, reliability diagram, and Brier score, in addition to conventional metrics, e.g. accuracy and F1 score. For validation, we use the publicly available and well-characterized Remote Sensing Image Scene Classification (RESISC45) dataset, which contains 31,500 images, covering 45 scene categories with 700 images in each category, and with the spatial resolution that varies from 30 to 0.2 m per pixel. This dataset was collected over different locations and under different conditions and possesses rich variations in translation, viewpoint, object pose and appearance, spatial resolution, illumination, background, and occlusion.

Post-hoc Uncertainty Quantification of Deep Learning Models Applied to Remote Sensing Image Scene Classification Steadily growing quantities of high-resolution UAV, aerial, and satellite imagery provide an exciting opportunity for global transparency and geographic profiling of activities of interest. Advances in deep learning, such as deep convolutional neural networks (CNNs) and transformer models, offer more efficient ways to exploit remote sensing imagery. Transformers, in particular, are capable of capturing contextual dependencies in the data. Accounting for context is important because activities of interest are often interdependent and reveal themselves in co-occurrence of related image objects or related signatures. However, while transformers and CNNs are powerful models, their predictions are often taken as point estimates, also known as pseudo probabilities, as they are computed by the softmax function. They do not provide information about how confident the model is in its predictions, which is important information in many mission-critical applications, and therefore limits their use in this space. Model evaluation metrics can provide information about the predictive model’s performance. We present and discuss results of post-hoc uncertainty quantification (UQ) of deep learning models, i.e., UQ application to trained models. We consider an application of CNN and transformer models to remote sensing image scene classification using satellite imagery, and compare confidence estimates of scene classification predictions of these models using evaluation metrics, such as expected calibration error, reliability diagram, and Brier score, in addition to conventional metrics, e.g. accuracy and F1 score. For validation, we use the publicly available and well-characterized Remote Sensing Image Scene Classification (RESISC45) dataset, which contains 31,500 images, covering 45 scene categories with 700 images in each category, and with the spatial resolution that varies from 30 to 0.2 m per pixel. This dataset was collected over different locations and under different conditions and possesses rich variations in translation, viewpoint, object pose and appearance, spatial resolution, illumination, background, and occlusion.

Post-hoc Uncertainty Quantification of Deep Learning Models Applied to Remote Sensing Image Scene Classification Steadily growing quantities of high-resolution UAV, aerial, and satellite imagery provide an exciting opportunity for global transparency and geographic profiling of activities of interest. Advances in deep learning, such as deep convolutional neural networks (CNNs) and transformer models, offer more efficient ways to exploit remote sensing imagery. Transformers, in particular, are capable of capturing contextual dependencies in the data. Accounting for context is important because activities of interest are often interdependent and reveal themselves in co-occurrence of related image objects or related signatures. However, while transformers and CNNs are powerful models, their predictions are often taken as point estimates, also known as pseudo probabilities, as they are computed by the softmax function. They do not provide information about how confident the model is in its predictions, which is important information in many mission-critical applications, and therefore limits their use in this space. Model evaluation metrics can provide information about the predictive model’s performance. We present and discuss results of post-hoc uncertainty quantification (UQ) of deep learning models, i.e., UQ application to trained models. We consider an application of CNN and transformer models to remote sensing image scene classification using satellite imagery, and compare confidence estimates of scene classification predictions of these models using evaluation metrics, such as expected calibration error, reliability diagram, and Brier score, in addition to conventional metrics, e.g. accuracy and F1 score. For validation, we use the publicly available and well-characterized Remote Sensing Image Scene Classification (RESISC45) dataset, which contains 31,500 images, covering 45 scene categories with 700 images in each category, and with the spatial resolution that varies from 30 to 0.2 m per pixel. This dataset was collected over different locations and under different conditions and possesses rich variations in translation, viewpoint, object pose and appearance, spatial resolution, illumination, background, and occlusion.

Researchers across the U.S. Army are conducting experiments on the implementation of emerging technologies on the battlefield. Key data points from these experiments include text comments on the technologies’ performances. Researchers use a range of Natural Language Processing (NLP) tasks to analyse such comments, including text summarization, sentiment analysis, and topic modelling. Based on the successful results from research in other domains, this research aims to yield greater insights by implementing military-specific language as opposed to a generalized corpus. This research is dedicated to developing a methodology to analyze text comments from Army experiments and field tests using topic models trained on an Army domain-specific corpus. The methodology is tested on experimental data agglomerated in the Forge database, an Army Futures Command (AFC) initiative to provide researchers with a common operating picture of AFC research. As a result, this research offers an improved framework for analysis with domain-specific topic models for researchers across the U.S. Army.

OED’s Overarching Tracker of DOT&E Actions distills information from DOT&E’s operational test reports and memoranda on test plan and test strategy approvals to generate informative metrics on the office’s activities. In FY22, DOT&E actions covered 68 test plans, 28 strategies, and 28 reports, relating to 74 distinct programs. This poster presents data from those documents and highlights findings on DOT&E’s effectiveness, suitability, and survivability determinations and other topics related to the state of T&E.

Machine learning (ML) systems demonstrate powerful predictive capability, but fielding such systems does not come without risk. ML can catastrophically fail in some scenarios, and in the absence of formal methods to validate most ML models, we require alternative methods to increase trust. While emerging techniques for uncertainty quantification and model explainability may seem to lie beyond the scope of many ML projects, they are essential tools for understanding deployment risk. This talk will share a practical workflow, useful tools, and lessons learned for ML development best practices. 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. SAND2023-11982A

DOE methods have evolved over the years, as have the needs and expectations of experimenters. Historically, the focus emphasized separating effects to reduce bias in effect estimates and maximizing hypotheses testing power, which are largely a reflection of the methodological and computational tools of their time. Often DOE in industry is done to predict product or process behavior under possible changes. We introduce Self-Validating Ensemble Models (SVEM), an inherently predictive algorithmic approach to the analysis of DOEs, generalizing the fractional bootstrap to make machine learning and bagging possible for small datasets common in DOE. In many DOE applications the number of rows is small, and the factor layout is carefully structured to maximize information gain in the experiment. Applying machine learning methods to DOE is generally avoided because they begin with a partitioning the rows into a training set for model fitting and a holdout set for model selection. This alters the structure of the design in undesirable ways such as randomly introducing effect aliasing. SVEM avoids this problem by using a variation of the fractionally weighted bootstrap to create training and validation versions of the complete data that differ only in how rows are weighted. The weights are reinitialized, and models refit multiple times so that our final SVEM model is a model average, much like bagging. We find this allows us to fit models where the number of estimated effects exceeds the number of rows. We will present simulation results showing that in these supersaturated cases SVEM outperforms existing approaches like forward selection as measured by prediction accuracy.

Rather than wait to the test and evaluation stage of a given system to evaluate safety, this talk proposes a technique which explicitly considers safety constraints during the learning process while providing probabilistic guarantees on performance subject to the operational environment's stochasticity. We provide evidence that such an approach results an overall safer system than their non-explicit counterparts in the context of wheeled robotic ground systems learning autonomous waypoint navigation from human demonstrations. Specifically, inverse reinforcement learning (IRL) provides a means by which humans can demonstrate desired behaviors for autonomous systems to learn environmental rewards (or inversely costs). The proposed presentation addresses two limitations of existing IRL techniques. First, previous algorithms require an excessive amount of data due to the information asymmetry between the expert and the learner. When a demonstrator avoids a state, it is not clear if it was because the state is sub-optimal or dangerous. The proposed talk explains how safety can be explicitly incorporated in IRL by using task specifications defined using linear temporal logic. Referred to as side information, this approach enables autonomous ground robots to avoid dangerous states both during training, and evaluation. Second, previous IRL techniques make the often unrealistic assumption that the agent has access to full information about the environment. We remove this assumption by developing an algorithm for IRL in partially observable Markov decision processes (POMDPs) which induces state uncertainty. The developed algorithm reduces the information asymmetry while increasing the data efficiency by incorporating task specifications expressed in temporal logic into IRL. The intrinsic nonconvexity of the underlying problem is managed in a scalable manner through a sequential linear programming scheme that guarantees local converge. In a series of examples, including experiments in a high-fidelity Unity simulator, we demonstrate that even with a limited amount of data and POMDPs with tens of thousands of states, our algorithm learns reward functions and policies that satisfy the safety specifications while inducing similar behavior to the expert by leveraging the provided side information.

Personal experience and anecdotal evidence suggest that presenting analyses to sponsors, especially technical sponsors, is improved by helping the sponsor understand how results were derived. Providing summaries of analytic results is necessary but can be insufficient when the end goal is to help sponsors make firm decisions. When time permits, engaging sponsors with walk-throughs of how results may change given different inputs is particularly salient in helping sponsors make decisions in the context of the bigger picture. Data visualizations and interactive software are common examples of what we call "decision tools" that can walk sponsors through varying inputs and views of the analysis. Given long-term engagement and regular communication with a sponsor, developing user-friendly decision tools is a helpful practice to support sponsors. This talk presents a methodology for building decision tools that combines leading practices in agile development and STEM education. We will use a Python-based app development tool called Streamlit to show implementations of this methodology.

Physical testing in the national security enterprise is often costly. Sometimes this is driven by hardware and labor costs, other times it can be driven by finite resources of time or hardware builds. Test engineers must make the most of their available resources to answer high consequence problems. Bayesian adaptive design of experiments (BADE) is one tool that should be in an engineer’s toolbox for designing and running experiments. BADE is sequential design of experiment approach which allows early stopping decisions to be made in real time using predictive probabilities (PP), allowing for more efficient data collection. BADE has seen successes in clinical trials, another high consequence arena, and it has resulted in quicker and more effective assessments of drug trials. 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 is to assess the robustness of PP in a BADE under different modeling assumptions, and to compare PP results to its frequentist alternative, conditional power (CP). Comparisons are made based on Type I error rates, statistical power, and time savings through average stopping time. Simulation results show PP has some robustness to distributional assumptions. PP also tends to control Type I error rates better than CP, while maintaining relatively strong power. While CP usually recommends stopping a test earlier than PP, CP also tends to have more inconsistent results, again showing the benefits of PP in a high consequence application. An application to a real problem from Sandia National Laboratories shows the large potential cost savings for using PP. The results of this study suggest BADE can be one piece of an evidence package during testing to stop testing early and pivot, in order to decrease costs and increase flexibility. 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.

The Chief Digital and AI Office (CDAO) Test & Evaluation Directorate is developing the Joint AI Test Infrastructure Capability (JATIC) program of record, which is an interoperable set of state-of-the-art software capabilities for AI Test & Evaluation. It aims to provide a provide a comprehensive suite of integrated testing tools which can be deployed widely across the enterprise to address key T&E gaps. In particular, JATIC will capabilities will support the assessment of AI system performance, cybersecurity, adversarial resilience, and explainability - enabling the end-user to more effectively execute their mission. It is a key component of the digital testing infrastructure that the CDAO will provide in order to support the development and deployment of data, analytics, and AI across the Department.

At George Mason University, we are developing swmfio, a Python package, for processing Space Weather Modeling Framework (SWMF) magnetosphere and ionosphere results, which is used to study the sun, heliosphere, and the magnetosphere. The SWMF framework centers around a high-performance magnetohydrodynamic (MHD) model, the Block Adaptive Tree Solar-wind Roe Upwind Scheme (BATS-R-US). This analysis uses swmfio and other methods, to trace magnetic field lines, compare the results, and identify why the methods differ. While the earth's magnetic field protects the planet from solar radiation, solar storms can distort the earth's magnetic field allowing solar storms to damage satellites and electrical grids. Being able to trace magnetic field lines helps us understand space weather. In this analysis, the September 1859 Carrington Event is examined. This event is the most intense geomagnetic storm in recorded history. We use three methods to trace magnetic field lines in the Carrington Event, and compare the field lines generated by the different methods. We consider two factors in the analysis. First, we directly compare methods by measuring the distances between field lines generated by different methods. Second, we consider how sensitive the methods are to initial conditions. We note that swmfio’s linear interpolation, which is customized for the BATS-R-US adaptive mesh, provides expected results. It is insensitive to small changes in initial conditions and terminates field lines at boundaries. We observe, that for any method, when the mesh size becomes large, results may not be accurate.

Over the past decade, Intrusion Detection Systems have become an important component of many organizations’ cyber defense and resiliency strategies. However, one of the greatest downsides of these systems is their reliance on known attack signatures for successful detection of malicious network events. When it comes to unknown attack types and zero-day exploits, modern Intrusion Detection Systems often fall short. Since machine learning algorithms for event classification are widely used in this realm, it is imperative to analyze the characteristics of network traffic that can lead to novelty detection using such classifiers. In this talk, we introduce a novel approach to identifying network traffic features that influence novelty detection based on survival analysis techniques. Specifically, we combine several Cox proportional hazards models to predict which features of a network flow are most indicative of a novel network attack and likely to confuse the classifier as a result. We also implement Kaplan-Meier estimates to predict the probability that a classifier identifies novelty after the injection of an unknown network attack at any given time. The proposed model is successful at pinpointing PSH Flag Count, ACK Flag Count, URG Flag Count, and Down/Up Ratio as the main features to impact novelty detection via Random Forest, Bayesian Ridge, and Linear SVR classifiers.

In today's Army, one of the fastest growing and most important areas in the effectiveness of our military is data science. One aspect of this field is image classification, which has applications such as target identification. However, one drawback within this field is that when an analyst begins to deal with a multitude of images, it becomes infeasible for an individual to examine all the images and classify them accordingly. My research presents a methodology for image classification which can be used in a military context, utilizing a typical unsupervised classification approach involving K-Means to classify a majority of the images while pairing this with user input to determine the label of designated images. The user input comes in the form of manual classification of certain images which are deliberately selected for presentation to the user, allowing this individual to select which group the image belongs in and refine the current image clusters. This shows how a semi-supervised approach to image classification can efficiently improve the accuracy of the results when compared to a traditional unsupervised classification approach.

The aim of the IARPA HIATUS program is to develop explainable systems for authorship attribution and author privacy preservation through the development of feature spaces which encode the distinguishing stylistic characteristics of authors independently of text genre, topic, or format. In this talk, I will discuss progress towards defining an evaluation framework for this task to provide robust insights into system strengths, weaknesses, and overall performance. Our evaluation strategy includes the use of an adversarial framework between attribution and privacy systems, development of a focused set of core metrics, analysis of system performance dependencies on key data factors, systematic exploration of experimental variables to probe targeted questions about system performance, and investigation of key trade-offs between different performance measures.

A critical change in how Test and Evaluation (T&E) supports capability delivery is needed to maintain our advantage over potential adversaries. Making this change requires a new paradigm in which T&E provides focused and relevant information supporting decision-making continually throughout capability development and informs decision makers from the earliest stages of Mission Engineering (ME) through Operations and Sustainment (O&S). This new approach improves the quality of T&E by moving from a serial set of activities conducted largely independently of Systems Engineering (SE) and ME activities to a new integrative framework focused on a continuum of activities termed T&E as a Continuum.

T&E as a Continuum has three key attributes – capability and outcome focused testing; an agile, scalable evaluation framework; and enhanced test design – critical in the conduct of T&E and improving capability delivery. T&E as a Continuum builds off the 2018 DoD Digital Engineering (DE) Strategy’s five critical goals with three key enablers – robust live, virtual, and constructive (LVC) testing; developing model-based environments; and a “digital” workforce knowledgeable of the processes and tools associated with MBSE, model-based T&E, and other model-based processes.

T&E as a Continuum improves the quality of T&E through integration of traditional SE and T&E activities, providing a transdisciplinary, continuous process coupling design evolution with VV&A.

The Mars Sample Return campaign aims at bringing rock and atmospheric samples from Mars to Earth through a series of robotic missions. These missions would collect the samples being cached and deposited on Martian soil by the Perseverance rover, place them in a container, and launch them into Martian orbit for subsequent capture by an orbiter that would bring them back. Given there exists a non-zero probability that the samples contain biological material, precautions are being taken to design systems that would break the chain of contact between Mars and Earth. These include techniques such as sterilization of Martian particles, redundant containment vessels, and a robust reentry capsule capable of accurate landings without a parachute. Requirements exist that the probability of containment not assured of Martian-contaminated material into Earth’s biosphere be less than one in a million. To demonstrate compliance with this strict requirement, a statistical framework was developed to assess the likelihood of containment loss during each sample return phase and make a statement about the total combined mission probability of containment not assured. The work presented here describes this framework, which considers failure modes or fault conditions that can initiate failure sequences ultimately leading to containment not assured. Reliability estimates are generated from databases, design heritage, component specifications, or expert opinion in the form of probability density functions or point estimates and provided as inputs to the mathematical models that simulate the different failure sequences. The probabilistic outputs are then combined following the logic of several fault trees to compute the ultimate probability of containment not assured. Given the multidisciplinary nature of the problem and the different types of mathematical models used, the statistical tools needed for analysis are required to be computationally efficient. While standard Monte Carlo approaches are used for fast models, a multi-fidelity approach to rare event probabilities is proposed for expensive models. In this paradigm, inexpensive low-fidelity models are developed for computational acceleration purposes while the expensive high-fidelity model is kept in the loop to retain accuracy in the results. This work presents an example of end-to-end application of this framework highlighting the computational benefits of a multi-fidelity approach. The decision to implement Mars Sample Return will not be finalized until NASA’s completion of the National Environmental Policy Act process. This document is being made available for information purposes only.

With solar energy being the most abundant energy source on Earth, it is no surprise that the reliance on solar photovoltaics (PV) has grown exponentially in the past decade. The increasing costs of fossil fuels have made solar PV more competitive and renewable energy more attractive, and the International Energy Agency (IEA) forecasts that solar PV's installed power capacity will surpass that of coal by 2027. Crucial to the management of solar PV power is the accurate forecasting of solar irradiance, which is heavily impacted by different types and distributions of clouds. Many studies have aimed to develop models that accurately predict the global horizontal irradiance (GHI) while accounting for the volatile effects of clouds; in this study, we aim to develop a statistical model that helps explain the relationship between various cloud properties and solar radiance reflected by clouds them-self. Using 2020 GOES-16 data from the GOES R-Series Advanced Baseline Imager (ABI), we investigated the effect that the cloud-optical depth, cloud top temperature, solar zenith angle, and look zenith angle had on cloud solar radiance while accounting for differing longitude and latitudes. Using these variables as the explanatory variables, we developed a linear model using multi-linear regression that, when tested on untrained data sets from different days (same time of day as the training set), results in a coefficient of determination (R^2) between .70-.75. Lastly, after analyzing the variables' degree of contribution to the cloud solar radiance, we presented error maps that highlight areas where the model succeeds and fails in prediction accuracy.

Surrogate modeling is an important class of techniques used to reduce the burden of resource-intensive computational models by creating fast and accurate approximations. In aerospace engineering, surrogates have been used to great effect in design, optimization, exploration, and uncertainty quantification (UQ) for a range of problems, like combustor design, spacesuit damage assessment, and hypersonic vehicle analysis. Consequently, the development, analysis, and practice of surrogate modeling is of broad interest. In this talk, several widely used surrogate modeling strategies are studied as archetypes in a discussion on parametric/nonparametric surrogate strategies, local/global model forms, complexity regularization, uncertainty quantification, and relative strengths/weaknesses. In particular, we consider several variants of two widely used classes of methods: polynomial chaos and Gaussian process regression. These surrogate models are applied to several synthetic benchmark test problems and examples of real high-speed flow problems, including hypersonic inlet design, thermal protection systems, and shock-wave/boundary-layer interactions. Through analysis of these concrete examples, we analyze the trade-offs that modelers must navigate to create accurate, flexible, and robust surrogates.

The future viability of DoD’s range enterprise depends on addressing dramatic changes in technology, rapid advances in adversary military capabilities, and the evolving approach the United States will take to closing kill chains in a Joint All Domain Operations environment. This recognition led DoD’s former Director of Operational Test and Evaluation (OT&E), the Honorable Robert Behler, to request that the National Academies of Science, Engineering and Medicine examine the physical and technical suitability of DoD’s ranges and infrastructure through 2035. The first half of this presentation will cover the highlights and key recommendations of this study, to include the need to create the “TestDevOps” digital infrastructure for future operational test and seamless range enterprise interoperability. The second half of this presentation looks at the legacy frameworks for the relationships of physical and virtual test capabilities, and how those frameworks are becoming outdated. This briefing explores proposals on how the interaction of operations, physical test capabilities, and virtual test capabilities need to evolve to support new paradigms of the rapidly evolving technologies and changing nature of multi-domain operations.

Florida International University (FIU) has developed an Advanced Automated Machine Learning System (AAMLS) under the sponsored research from Department of Defense - Test Resource Management Center (DOD-TRMC), to provide Artificial Intelligence based advanced analytics solutions in the area of Cyber, IOT, Network, Energy, Environment etc. AAMLS is a Rapid Modeling & Testing Tool (RMTT) for developing machine learning and deep learning models in few steps by subject matter experts from various domains with minimum machine learning knowledge using auto & optimization workflows. AAMLS allows analysis of data collected from different test technology domains by using machine learning / deep learning and ensemble learning approaches to generate models, make predictions, then apply advanced analytics and visualization to perform analysis. This system enables automated machine learning using AI based Advanced Analytics and the Analytics Control Center platforms by connecting to multiple Data Sources. Artificial Intelligence based Advanced Analytics Platform: This platform is the analytics engine of AAML which provides pre-processing, feature engineering, model building and predictions. Primary components of this platform include:

• Machine Learning Server: This module is deployed to build ML/DL models using the training data from the data sources and perform predictions/analysis of associated test data based on the AAML-generated ML/DL models.
• Machine Learning Algorithms: ML algorithms like Logistic Regression, Linear regression, Decision tree, Random Forest, One Class Support Vector Machine, Jaccard Similarity etc. are available for model building.
• Deep learning Algorithms: Deep learning algorithms such as Deep Neural Networks and Recurrent Neural Networks are available to perform classification & anomaly detection using the TensorFlow framework and the Keras API.

• Data Source: This module allows the user to connect to the existing data to the AAML system to perform analytics. These data sources may reside in a Network File Share, Database or Big Data Cluster.
• Model Development: This module provides the functionality to build ML/DL models with various AI algorithms. This is performed by engaging specific ML algorithms for five types of analysis: Classification, Regression, Time-Series, Anomaly Detection and Clustering
• Predictions: This module provides the functionality to predict the outcome of an analysis of an associated data set based on model built during Model Development.
• Manage Models and Predictions: These modules allow the user to manage the ML models that have been generated and resulting predictions of associated data sets.

Image classification is a critical part of gathering information on high-value targets. To this end, Convolutional Neural Networks (CNN) have become the standard model for image and facial classification. However, CNNs alone are not entirely effective at image classification, and especially human classification due to their lack of robustness and bias. Recent advances in CNNs, however, allow for data fusion to help reduce the uncertainty in their predictions. In this project, we describe a multimodal algorithm designed to increase confidence in image classification with the use of a joint fusion model with image and text data. Our work utilizes CNNs for image classification and bag-of-words for text categorization on Wikipedia images and captions relating to the same classes as the CIFAR-100 dataset. Using data fusion, we combine the vectors of the CNN and bag-of-words models and utilize a fully connected network on the joined data. We measure improvements by comparing the SoftMax layer for the joint fusion model and image-only CNN.

Circular Error Probable (CEP) is a measure of a weapon system’s precision developed based on the Bivariate Normal Distribution. Failing to understanding the theory behind CEP can result in misuse of equations developed to help estimation. Estimation of CEP is also much more straightforward given situations such as single samples where factors are not being manipulated. This brief aims to help build a theoretical understanding of CEP, and then presents a non-trivial example in which CEP is estimated via multilevel regression. The goal is to help build an understanding of CEP so it can be properly estimated in trivial (single sample) and non-trivial cases (e.g. regression and multilevel regression).

Decisions form the core of T&E: decisions about which tests to conduct and, especially, decisions on whether to accept or reject a system at its milestones. The traditional approach to acceptance is based on conducting tests under various conditions to ensure that key performance parameters meet certain thresholds with the required degree of confidence. In this approach, data is collected during testing, then analyzed with techniques from classical statistics in a post-action report. This work explores a new Bayesian paradigm for T&E based on one simple principle: maintaining a model of the probability distribution over system parameters at every point during testing. In particular, the Bayesian approach posits a distribution over parameters prior to any testing. This prior distribution provides (a) the opportunity to incorporate expert scientific knowledge into the inference procedure, and (b) transparency regarding all assumptions being made. Once a prior distribution is specified, it can be updated as tests are conducted to maintain a probability distribution over the system parameters at all times. One can leverage this probability distribution in a variety of ways to produce analytics with no analog in the traditional T&E framework. In particular, having a probability distribution over system parameters at any time during testing enables one to implement an optimal decision-making procedure using Bayesian Decision Theory (BDT). BDT accounts for the cost of various testing options relative to the potential value of the system being tested. When testing is expensive, it provides guidance on whether to conserve resources by ending testing early. It evaluates the potential benefits of testing for both its ability to inform acceptance decisions and for its intrinsic value to the commander of an accepted system. This talk describes the BDT paradigm for T&E and provides examples of how it performs in simple scenarios. In future work we plan to extend the paradigm to include the features, the phenomena, and the SME elicitation protocols necessary to address realistic T&E cases.

Many scientific and engineering experiments are developed to study specific questions of interest. Unfortunately, time and budget constraints make operating these controlled experiments over wide ranges of conditions intractable, thus limiting the amount of data collected. In this presentation, we discuss a Bayesian approach to identify the most informative conditions, based on the expected information gain. We will present a framework for finding optimal experimental designs that can be applied to physics-based models with high-dimensional inputs and outputs. We will study a real-world example where we aim to infer the parameters of a chemically reacting system, but there are uncertainties in both the model and the parameters. A physics-based model was developed to simulate the gas-phase chemical reactions occurring between highly reactive intermediate species in a high-pressure photolysis reactor coupled to a vacuum-ultraviolet (VUV) photoionization mass spectrometer. This time-of-flight mass spectrum evolves in both kinetic time and VUV energy producing a high-dimensional output at each design condition. The high-dimensional nature of the model output poses a significant challenge for optimal experimental design, as a surrogate model is built for each output. We discuss how accurate low-dimensional representations of the high-dimensional mass spectrum are necessary for computing the expected information gain. Bayesian optimization is employed to maximize the expected information gain by efficiently exploring a constrained design space, taking into account any constraint on the operating range of the experiment. Our results highlight the trade-offs involved in the optimization, the advantage of using optimal designs, and provide a workflow for computing optimal experimental designs for high-dimensional physics-based models.

Statistical analysis is typically conducted using either a frequentist or Bayesian approach. But what is the impact of choosing one analysis method over another? This presentation will compare the results of both linear and logistic regression using Bayesian and frequentist methods. The data set combines information on simulated diffusion of material and anticipated background signal to imitate sensor output. The sensor is used to estimate the total concentration of material, and a threshold will be set such that the false alarm rate (FAR) due to the background is a constant. The regression methods are used to relate the probability of detection, for a given FAR, to predictor variables, such as the total amount of material released. The presentation concludes with a comparison of the similarities and differences between the two methods given the results.

SIPET, established in 2021, is a deputate within the office of the Director, Operational Test and Evaluation (DOT&E). DOT&E created SIPET to codify and implement the Director’s strategic vision and keep pace with science and technology to modernize T&E tools, processes, infrastructure, and workforce. That is, The mission of SIPET is to drive continuous innovation to meet the T&E demands of the future; support the development of a workforce prepared to meet the toughest T&E challenges; and nurture a culture of information exchange across the enterprise and update policy and guidance. SIPET proactively identifies current and future operational test and evaluation needs, gaps, and potential solutions in coordination with the Services and agency operational test organizations. Collaborating with numerous stakeholders, SIPET develops and refines operational test policy guidance that support new test methodologies and technologies in the acquisition and test communities. SIPET, in collaboration with the T&E community, is leading the development of the 2022 DOT&E Strategy Update Implementation Plan (I-Plan). I-Plan initiatives include:

• Test the Way We Fight – Architect T&E around validated mission threads and demonstrate the operational performance of the Joint Force in multi-domain operations.
• Accelerate the Delivery of Weapons that Work – Embrace digital technologies to deliver high-quality systems at more dynamic rates.
• Increase the Survivability of DOD in Contested Environments – Identify, assess, and act on cyber, electromagnetic, spectrum, space, and other risks to DOD mission – at scale and at speed.
• Pioneer T&E of Weapons Systems Built to Change Over Time – Implement fluid and iterative T&E across the entire system lifecycle to help assure continued combat credibility as the system evolves to meet warfighter needs.
• Foster an Agile and Enduring T&E Enterprise Workforce – Centralize and leverage efforts to assess, curate, and engage T&E talent to quicken the pace of innovation across the T&E enterprise.

Testing and assuring responsible use of AI/ML enabled capabilities is a nascent topic in the DOD with many efforts being spearheaded by CDAO. In general, black box models tend to suffer from consequences related to edge cases, emergent behavior, misplaced or lack of trust, and many other issues, so traditional testing is insufficient to guarantee safety and responsibility in the employment of a given AI enabled capability. Focus of this concern tends to fall on well-publicized and high-risk capabilities, such as AI enabled autonomous weapons systems. However, while AI/ML enabled capabilities supporting personnel processes and systems, such as algorithms used for retention and promotion decision support, tend to carry low safety risk, many concerns, some of them specific to the personnel space, run the risk of undermining the DOD’s 5 ethical principles for RAI. Examples include service member privacy concerns, invalid prospective policy analysis, disparate impact against marginalized service member groups, and unintended emergent service member behavior in response to use of the capability. Eroding barriers to use of AI/ML are facilitating an increasing number of applications while some of these concerns are still not well understood by the analytical community. We consider many of these issues in the context of an IDA ML enabled capability and propose mechanisms to assure stakeholders of the adherence to the DOD’s ethical principles.

In its mission to expand knowledge and improve aviation, NASA conducts research to address sonic boom noise, the prime barrier to overland supersonic flight. For half a century civilian aircraft have been required to fly slower than the speed of sound when over land to prevent sonic boom disturbances to communities under the flight path. However, lower noise levels may be achieved via new aircraft shaping techniques that reduce the merging of shockwaves generated during supersonic flight. As part of its Quesst mission, NASA is building a piloted, experimental aircraft called the X-59 to demonstrate low noise supersonic flight. After initial flight testing to ensure the aircraft performs as designed, NASA will begin a national campaign of community overflight tests to collect data on how people perceive the sounds from this new design. The data collected will support national and international noise regulators’ efforts as they consider new standards that would allow supersonic flight over land at low noise levels. This presentation provides an overview of the community test campaign, including the scope, key objectives, stakeholders, and challenges.

We describe recent advances in Gaussian process emulation, which allow us to both save computation time and to apply inference algorithms that previously were too expensive for operational use. Specific examples are given from the Earth-orbiting Orbiting Carbon Observatory and the future Surface Biology and Geology Missions, dynamical systems, and other applications. While Gaussian processes are a well-studied field, there are surprisingly important choices that the community has not paid so much attention to this far, including dimension reduction, kernel parameterization, and objective function selection. This talk will highlight some of those choices and help understand what practical implications they have.

A key component of global and national security in the nuclear weapons age is the proliferation of nuclear weapons technology and development. A key component of enforcing this non-proliferation policy is developing an awareness of the scientific research being pursued by other nations and organizations. To support non-proliferation goals and contribute to nuclear science research, we trained a RoBERTa deep neural language model on a large set of U.S. Department of Energy Office of Science and Technical Information (OSTI) research article abstracts and then finetuned this model for classification of scientific abstracts into 60 disciplines, which we call NukeLM. This multi-step approach to training improved classification accuracy over its untrained or partially out-of-domain competitors. While it is important for classifiers to be accurate, there has also been growing interest in ensuring that classifiers are well-calibrated with uncertainty quantification that is understandable to human decision-makers. For example, in the multiclass problem, classes with a similar predicted probability should be semantically related. Therefore, we also introduced an extension of the Bayesian belief matching framework proposed by Joo et al. (2020) that easily scales to large NLP models, such as NukeLM, and better achieves the desired uncertainty quantification properties.

In space mission formulation, it is critical to link the scientific phenomenology under investigation directly to the spacecraft design, mission design, and the concept of operations. With many missions of discovery, the large uncertainty in the science phenomenology and the operating environment necessitates mission architecture solutions that are robust and resilient to these unknowns, in order to maximize the probability of achieving the mission objectives. Feasible mission architectures are assessed against performance, cost, and risk, in the context of large uncertainties. For example, despite Cassini observations of Enceladus, significant uncertainties exist in the moon’s jet properties and the surrounding Enceladus environment. Orbilander or any other mission to Enceladus will need to quantify or bound these uncertainties in order to formulate a viable design and operations trade space that addresses a range of mission objectives within the imposed technical and programmatic constraints. Uncertainty quantification (UQ), utilizes a portfolio of stochastic, data science, and mathematical methods to characterize uncertainty of a system and inform risk and decision-making. This discussion will focus on a formulation of a UQ workflow and an example of an Enceladus mission development use case.

One of the challenges to conducting research in the Air Force Research Laboratory is that many of our equipment controllers cannot be directly connected to our internal networks, due to older or specialized operating systems and the need for administrative privileges for proper functioning. This means that the current data collection process is often highly manual, with users documenting experiments in physical notebooks and transferring data via CDs or portable hard drives to connected systems for sharing or further processing. In the Materials & Manufacturing Directorate, we have developed a unique approach to seamlessly integrate our labs for more efficient data collection and transfer, which is specifically designed to help users ensure that data is findable for future reuse. In this talk, we will highlight our two enabling tools: NORMS, which assists users to easily generate metadata for direct association with data collected in the lab to eliminate physical notebooks; and Spike, which automates one-way data transfer from isolated systems to databases mirrored on other networks. In these databases, metadata can be used for complex search queries and data is automatically shared with project members without requiring additional transfers. The impact of this solution has been significantly faster data availability (including searchability) to all project members: a transfer and scanning process that used to take 3 hours can now take a few minutes. Future use cases will also enable Spike to transfer data directly into cloud buckets for in situ analysis, which would streamline collaboration with partners.

Today there are an increasing number of sensors on the battlefield. These sensors collect data that includes, but is not limited to, images, audio files, videos, and text files. With today’s technology, the data collection process is strong, and there is a growing opportunity to leverage multiple streams of data, each coming in different forms. This project aims to take multiple types of data, specifically images and audio files, and combine them to increase our ability to detect and recognize objects. The end state of this project is the creation of an algorithm that utilizes and merges voice recordings and images to allow for easier recognition. Most research tends to focus on one modality or the other, but here we focus on the prospect of simultaneously leveraging both modalities for improved entity resolution. With regards to audio files, the most successful deconstruction and dimension reduction technique is a deep auto encoder. For images, the most successful technique is the use of a convolutional neural network. To combine the two modalities, we focused on two different techniques. The first was running each data source through a neural network and multiplying the resulting class probability vectors to capture the combined result. The second technique focused on running each data source through a neural network, extracting a layer from each network, concatenating the layers for paired image and audio samples, and then running the concatenated object through a fully connected neural network.

Standard deep learning models for classification and regression applications are ideal for capturing complex system dynamics. Unfortunately, their predictions can be arbitrarily inaccurate when the input samples are not similar to the training data. Implementation of distance aware uncertainty estimation can be used to detect these scenarios and provide a level of confidence associated with their predictions. We present results using Deep Gaussian Process Approximation (DGPA) methods for 1) anomaly detection at Spallation Neutron Source (SNS) accelerator and 2) uncertainty aware surrogate model for the Fermi National Accelerator Lab (FNAL) Booster Accelerator Complex.

The United States Military possesses special forces units that are entrusted to engage in the most challenging and dangerous missions that are essential to fighting and winning the nations wars. Entry into special forces is based on a series of assessments called Special Forces Assessment and Selection (SFAS), which consists of numerous challenges that test a soldiers mental toughness, physical fitness, and intelligence. Using logistic regression, random forest classification, and neural network classification, the researchers in this study aim to create a model that both accurately predicts whether a candidate passes SFAS and which variables are significant indicators of passing selection. Logistic regression proved to be the most accurate model, while also highlighting physical fitness, military experience, and intellect as the most significant indicators associated with success.

With the rise in availability of larger datasets, there is a growing need of tools and methods to help inform data-driven decisions. Data that vary over a continuum, such as time, exist in a wide array of fields, such as defense, finance, and medicine. One such class of methods that addresses data varying over a continuum is functional data analysis (FDA). FDA methods typically make three assumptions that are often violated in real datasets: all observations exist over the same continuum interval (such as a closed interval [a,b]), all observations are regularly and densely observed, and if the dataset consists of multiple covariates, the covariates are independent of one another. We look to address violation of the latter two assumptions. In this talk, we will discuss methods for analyzing functional data that are irregularly and sparsely observed, while also accounting for dependencies between covariates. These methods will be used to estimate the reconstruction of partially observed multivariate functional data that contain measurement errors. We will begin with a high-level introduction of FDA. Next, we will introduce functional principal components analysis (FPCA), which is a representation of functions that our estimation methods are based on. We will discuss a specific approach called principal components analysis through conditional expectation (PACE) (Yao et al, 2005), which computes the FPCA quantities for a sparsely or irregularly sampled function. The PACE method is a key component that allows us to estimate partially observed functions based on the available dataset. Finally, we will introduce multivariate functional principal components analysis (MFPCA) (Happ & Greven, 2018), which utilizes the FPCA representations of each covariate’s functions in order to compute a principal components representation that accounts for dependencies between covariates. We will illustrate these methods through implementation on simulated and real datasets. We will discuss our findings in terms of the accuracy of our estimates with regards to the amount and portions of a function that is observed, as well as the diversity of functional observations in the dataset. We will conclude our talk with discussion on future research directions.

When doing multivariate data analysis, one common obstacle is the presence of incomplete observations, i.e., observations for which one or more covariates are missing data. Rather than deleting entire observations that contain missing data, which can lead to small sample sizes and biased inferences, data imputation methods can be used to statistically “fill-in” missing data. Imputing data can help combat small sample sizes by using the existing information in partially complete observations with the end goal of producing less biased and higher confidence inferences.

In aerospace applications, imputation of missing data is particularly relevant because sample sizes are small and quantifying uncertainty in the model is of utmost importance. In this paper, we outline the benefits of a fully Bayesian imputation approach which samples simultaneously from the joint posterior distribution of model parameters and the imputed values for the missing data. This approach is preferred over multiple imputation approaches because it performs the imputation and modeling steps in one step rather than two, making it more compatible with complex model forms. An example of this imputation approach is applied to the NASA Instrument Cost Model (NICM), a model used widely across NASA to estimate the cost of future spaceborne instruments. The example models are implemented in Stan, a statistical-modeling tool enabling Hamiltonian Monte Carlo (HMC).

Planetary Protection is the practice of protecting solar system bodies from harmful contamination by Earth life and protecting Earth from possible life forms or bioactive molecules that may be returned from other solar system bodies. Microbiologists and engineers at NASA’s Jet Propulsion Laboratory (JPL) design microbial reduction and sterilization protocols that reduce the number of microorganisms on spacecraft or eliminate them entirely. These protocols are developed using controlled experiments to understand the microbial reduction process. Many times, a phenomenological model (such as a series of differential equations) is posited that captures key behaviors and assumptions of the process being studied. A Sterility Assurance Level (SAL) – the probability that a product, after being exposed to a given sterilization process, contains one or more viable organisms – is a standard metric used to assess risk and define cleanliness requirements in industry and for regulatory agencies. Experiments performed to estimate the SAL of a given microbial reduction or sterilization protocol many times have large uncertainties and variability in their results even under rigorously implemented controls that, if not properly quantified, can make it difficult for experimenters to interpret their results and can hamper a credible evaluation of risk by decision makers. In this talk, we demonstrate how Bayesian statistics and experimentation can be used to quantify uncertainty in phenomenological models in the case of microorganism survival under short-term high heat exposure. We show how this can help stakeholders make better risk-informed decisions and avoid the unwarranted conservatism that is often prescribed when processes are not well understood. The experiment performed for this study employs a 6 kW infrared heater to test survivability of heat resistant Bacillus canaveralius 29669 to temperatures as high as 350 °C for time durations less than 30 sec. The objective of this study was to determine SALs for various time-temperature combinations, with a focus on those time-temperature pairs that give a SAL of 10^-6. Survival ratio experiments were performed that allow estimation of the number of surviving spores and mortality rates characterizing the effect of the heat treatment on the spores. Simpler but less informative fraction-negative experiments that only provide a binary sterile/not-sterile outcome were also performed once a sterilization temperature regime was established from survival ratio experiments. The phenomenological model considered here is a memoryless mortality model that underlies many heat sterilization protocols in use today. This discussion and poster will outline how the experiment and model were brought together to determine SALs for the heat treatment under consideration. Ramifications to current NASA planetary protection sterilization specifications and current missions under development such as Mars Sample Return will be discussed. This presentation/poster is also relevant to experimenters and microbiologists working on military and private medical device applications where risk to human life is determined by sterility assurance of equipment.

Classification models for binary outcomes are in widespread use across a variety of industries. Results are commonly summarized in a misclassification table, also known as an error or confusion matrix, which indicates correct vs incorrect predictions for different circumstances. Models are developed to minimize both false positive and false negative errors, but the optimization process to train/obtain the model fit necessarily results in cost-benefit trades. However, how to obtain an objective assessment of the performance of a given model in terms of predictive capability or benefit is less well understood, due to both the rich plethora of options described in literature as well as the largely overlooked influence of noise factors, specifically class imbalance. Many popular measures are susceptible to effects due to underlying differences in how the data are allocated by condition, which cannot be easily corrected. This talk considers the wide landscape of possibilities from a statistical robustness perspective. Results are shown from sensitivity analyses for a variety of different conditions for several popular metrics and issues are highlighted, highlighting potential concerns with respect to machine learning or ML-enabled systems. Recommendations are provided to correct for imbalance effects, as well as how to conduct a simple statistical comparison that will detangle the beneficial effects of the model itself from those of imbalance. Results are generalizable across model type.

Situation Awareness (SA) plays a key role in decision making and human performance; higher operator SA is associated with increased operator performance and decreased operator errors. In the most general terms, SA can be thought of as an individual’s “perception of the elements in the environment within a volume of time and space, the comprehension of their meaning, and the projection of their status in the near future.” While “situational awareness” is a common suitability parameter for systems under test, there is no standardized method or metric for quantifying SA in operational testing (OT). This leads to varied and suboptimal treatments of SA across programs and test events. Current measures of SA are exclusively subjective and paint an inadequate picture. Future advances in system connectedness and mission complexity will exacerbate the problem. We believe that technological improvements will necessitate increases in the complexity of the warfighters’ mission, including changes to team structures (e.g., integrating human teams with human-machine teams), command and control (C2) processes (e.g., expanding C2 frameworks toward joint all-domain C2), and battlespaces (e.g., overcoming integration challenges for multi-domain operations). Operational complexity increases the information needed for warfighters to maintain high SA, and assessing SA will become increasingly important and difficult to accomplish. IDA’s Test science team has proposed a piecewise approach to improve the measurement of situation awareness in operational evaluations. The aim of this presentation is to promote a scientific understanding of what SA is (and is not) and encourage discussion amongst practitioners tackling this challenging problem. We will briefly introduce Endsley’s Model of SA, review the trade-offs involved in some existing measures of SA, and discuss a selection of potential ways in which SA measurement during OT may be improved.

Design thinking is a problem-solving approach that promotes the principles of human-centeredness, iteration, and diversity. The poster provides a five-step framework for how to incorporate these design principles when building an operational test. In the first step, test designers conduct research on test users and the problems they encounter. In the second step, designers articulate specific user needs to address in the test design. In the third step, designers generate multiple solutions to address user needs. In the forth step, designers create prototypes of their best solutions. In the fifth step, designers refine prototypes through user testing.

Data literacy, the ability to read, write, and communicate data in context, is fundamental for military organizations to create a culture where data is appropriately used to inform both operational and non-operational decisions. However, oftentimes organizations outsource data problems to outside entities and rely on a small cadre of data experts to tackle organizational problems. In this talk we will argue that data literacy is not solely the role or responsibility of the data expert. Ultimately, if experts develop tools and analytics that Army decision makers cannot use, or do not effectively understand the way the Army makes decisions, the Army is no more data rich than if it had no data at all. While serving on a sabbatical as the Chief Data Scientist for Joint Special Operations Command, COL Nick Clark (Department of Mathematical Sciences, West Point), noticed that a lack of basic data literacy skills was a major limitation to creating a data centric organization. As a result of this, he created 10 hours of training focusing on the fundamentals of data literacy. After delivering the course to JSOC, other DoD organizations began requesting the training. In response to this, a team from West Point joined with Army Talent Management Task Force to create mobile training teams. The teams have now delivered the training over 30 times to organizations ranging from tactical units up to strategic level commands. In this talk, we discuss what data literacy skills should be taught to the force and highlight best practices in educating soldiers, civilians, and contractors on the basics of data literacy. We will finally discuss strategies for assessing organizational Data Literacy and provide a framework for attendees to assess their own organizations data strengths and weaknesses.

Stealthy communication allows the transfer of information while hiding not only the content of that information but also the fact that any hidden information was transferred. One way of doing this is embedding information into network covert channels, e.g., timing between packets, header fields, and so forth. We describe our work on an integrated system for the design, analysis, and testing of such communication. The system consists of two main components: the analytical component, the NExtSteP (NRL Extensible Stealthy Protocols) testbed, and the emulation component, consisting of CORE (Common Open Research Emulator), an existing open source network emulator, and EmDec, a new tool for embedding stealthy traffic in CORE and decoding the result. We developed the NExtSteP testbed as a tool to evaluate the performance and stealthiness of embedders and detectors applied to network traffic. NExtSteP includes modules to: generate synthetic traffic data or ingest it from an external source (e.g., emulation or network capture); embed data using an extendible collection of embedding algorithms; classify traffic, using an extendible collection of detectors, as either containing or not containing stealthy communication; and quantify, using multiple metrics, the performance of a detector over multiple traffic samples. This allows us to systematically evaluate the performance of different embedders (and embedder parameters) and detectors against each other. Synthetic data are easy to generate with NExtSteP. We use these data for initial experiments to broadly guide parameter selection and to study asymptotic properties that require numerous long traffic sequences to test. The modular structure of NExtSteP allows us to make our experiments increasingly realistic. We have done this in two ways: by ingesting data from captured traffic and then doing embedding, classification, and detector analysis using NExtSteP, and by using EmDec to produce external traffic data with embedded communication and then using NExtStep to do the classification and detector analysis. The emulation component was developed to build and evaluate proof-of-concept stealthy communications over existing IP networks. The CORE environment provides a full network, consisting of multiple nodes, with minimal hardware requirements and allows testing and orchestration of real protocols. Our testing environment allows for replay of real traffic and generation of synthetic traffic using MGEN (Multi-Generator) network testing tool. The EmDec software was created with the already existing NRL-developed protolib (protocol library). EmDec, running on CORE networks and orchestrated using a set of scripts, generates sets of data which are then evaluated for effectiveness by NExtSteP. In addition to evaluation by NExtSteP, development of EmDec allowed us to discover multiple novelties that were not apparent while using theoretical models. We describe current status of our work, the results so far, and our future plans.

Traditional methods for reliability analysis are challenged in developmental testing (DT) as systems become increasingly complex and DT programs become shorter and less predictable. Bayesian statistical methods, which can combine data across DT segments and use additional data to inform reliability estimates, can address some of these challenges. However, Bayesian methods are not widely used. I will present the results of a study aimed at identifying effective practices for the use of Bayesian reliability analysis in DT programs. The study consisted of interviews with reliability subject matter experts, together with a review of relevant literature on Bayesian methods. This analysis resulted in a set of best practices that can guide an analyst in deciding whether to apply Bayesian methods, in selecting the appropriate Bayesian approach, and in applying the Bayesian method and communicating the results.

Traditional methods for reliability analysis are challenged in developmental testing (DT) as systems become increasingly complex and DT programs become shorter and less predictable. Bayesian statistical methods, which can combine data across DT segments and use additional data to inform reliability estimates, can address some of these challenges. However, Bayesian methods are not widely used. I will present the results of a study aimed at identifying effective practices for the use of Bayesian reliability analysis in DT programs. The study consisted of interviews with reliability subject matter experts, together with a review of relevant literature on Bayesian methods. This analysis resulted in a set of best practices that can guide an analyst in deciding whether to apply Bayesian methods, in selecting the appropriate Bayesian approach, and in applying the Bayesian method and communicating the results.

A shortfall of the Derringer and Suich (1980) desirability function for multi-objective optimization has been a lack of inferential methods to quantify uncertainty. Most articles for addressing uncertainty involve robust methods, providing a point estimate that is less affected by variation. Few articles address confidence intervals or bands but not specifically for the widely used Derringer and Suich method. 8 methods are presented to construct 100(1-alpha) confidence intervals around Derringer and Suich desirability function optimal values. First order and second order models using bivariate and multivariate data sets are used as examples to demonstrate effectiveness. The 8 proposed methods include a simple best/worst case method, 2 generalized methods, 4 simulated surface methods, and a nonparametric bootstrap method. One of the generalized methods, 2 of the simulated surface methods, and the nonparametric method account for covariance between the response surfaces. All 8 methods seem to perform decently on the second order models; however, the methods which utilize an underlying multivariate-t distribution, Multivariate Generalized (MG) and Multivariate t Simulated Surface (MVtSSig) are recommended methods from this research as they perform well with small samples for both first order and second order models with coverage only becoming unreliable at consistently non-optimal solutions. MG and MVtSSig inference could also be used in conjunction with robust methods such as Pareto Front Optimization to help ascertain which solutions are more likely to be optimal before constructing confidence interval.

Artificial intelligence (AI) and Machine Learning (ML) are increasingly being examined for their utility in many domains. AI/ML solutions are being proposed for a broad set of applications, including surrogate modeling, anomaly detection, classification, image segmentation, control, etc. A lot of effort is being put into evaluating these solutions for robustness, especially in the context of safety critical applications. While traditional methods of verification and validation continue to be necessary, challenges exist in many safety critical applications given the limited ability to gather data covering all possible conditions, and limited ability to conduct experiments. This presentation will discuss potential approaches for testing and evaluating machine learning algorithms in such applications, as well as metrics for this purpose.

The Test Science Team facilitates data-driven decision-making by disseminating various testing and analysis methodologies. One way they disseminate these methodologies is through the annual workshop, DATAWorks; another way is through the website, TestScience.org. The Test Science website includes video training, interactive tools, a related research library as well as the DATAWorks Archive. "Introducing TestScience.org", a presentation at DATAWorks, could include a poster and an interactive guided session through the site content. The presentation would inform interested DATAWorks attendees of the additional resources throughout the year. It could also be used to inform the audience about ways to participate, such as contributing interactive Shiny tools, training content, or research. "Introducing TestScience.org" would highlight the following sections of the website: 1. The DATAWorks Archives 2. Learn (Video Training) 3. Tools (Interactive Tools) 5. Research (Library) 6. Team (About and Contact) Incorporating into DATAWorks an introduction to TestScience.org would inform attendees of additional valuable resources available to them, and could encourage broader participation in Testscience.org, adding value to both the DATAWorks attendees and the TestScience.org efforts.

BACKGROUND Post-prandial glucose response resulting from a mixed meal tolerance test is evaluated from trajectory data of measured glucose, insulin, C-peptide, GLP-1 and other measurements of insulin sensitivity and β-cell function. In order to compare responses between populations or different composition of mixed meals, the trajectories are collapsed into the area under the curve (AUC) or incremental area under the curve (iAUC) for statistical analysis. Both AUC and iAUC are coarse distillations of the post-prandial curves and important properties of the curve structure are lost. METHODS Visual Basic Application (VBA) code was written to automatically extract seven different key calculus-based curve-shape properties of post-prandial trajectories (glucose, insulin, C-peptide, GLP-1) beyond AUC. Through two-sample t-tests, the calculus-based markers were compared between outcomes (reactive hypoglycemia vs. healthy) and against demographic information. RESULTS Statistically significant p-values (p < .01) between multiple curve properties in addition to AUC were found between each molecule studied and the health outcome of subjects based on the calculus-based properties of their molecular response curves. A model was created which predicts reactive hypoglycemia based on individual curve properties most associated with outcomes. CONCLUSIONS There is a predictive power using response curve properties that was not present using solely AUC. In future studies, the response curve calculus-based properties will be used for predicting diabetes and other health outcomes. In this sense, response-curve properties can predict an individual's susceptibility to illness prior to its onset using solely mixed meal tolerance test results.

As the Test and Evaluation community increasingly relies on Modeling and Simulation (M&S) to supplement live testing, M&S validation has become critical for ensuring credible weapon system evaluations. System-level evaluations of Armored Fighting Vehicles (AFV) rely on the Advanced Joint Effectiveness Model (AJEM) and Full-Up System Level (FUSL) testing to assess AFV vulnerability. This report reviews one of the primary methods that analysts use to validate AJEM, called the Component Damage Vector (CDV) Method. The CDV method compares components that were damaged in FUSL testing to simulated representations of that damage from AJEM.

Significant literature has been published in recent years calling for updated and new T&E infrastructure to allow for the credible, verifiable assessment of DoD AI-enabled capabilities (AIECs) . However, existing literature falls short in providing the detail necessary to justify investments in specific DoD Enterprise T&E infrastructure. The goal of this study was to collect data about current DoD programs with AIEC and corresponding T&E infrastructure to identify high priority investments by tracing AIECs to tailored, specific recommendations for enterprise AI T&E infrastructure. The study is divided into six bins of research, itemized below. This presentation provides an interim study update on the current state of programs with AIEC across DoD.

• Goals : State specific enterprise AI T&E normative goal(s) and timeline, including rationale.
• Demand: Generate T&E requirements and a demand forecast by querying existing and anticipated AI programs across the Department.
• Supply Baseline: Catalog existing and planned AI T&E activities and infrastructure at the enterprise, Service, and Program levels, including existing resourcing levels.
• Gaps: Identify specific gaps based on this supply and demand information.
• Responsibilities: Clarify roles and responsibilities across OSD T&E stakeholders.
• Actions: Identify differentiated lines of effort—including objectives, milestones, and cost estimates—that create a cohesive plan to achieve enterprise AI T&E goals.

The DoD maintains a broad array of systems, each one sustained by an often complex supply chain of components and suppliers. The ways that these supply chains are interlinked can have major implications for the resilience of the defense industrial base as a whole, and the readiness of multiple weapon systems. Finding opportunities to improve overall resilience requires gaining visibility of potential weak links in the chain, which requires integrating data across multiple disparate sources. By using open-source data pipeline software to enhance reproducibility, and flexible network analysis methods, multiple stovepiped data sources can be brought together to develop a more complete picture of the supply chain across systems.