DATAWorks Sessions Archive

As advanced technologies and capabilities are enabling machines to engage in tasks that only humans have done previously, new challenges have emerged for the rigorous testing and evaluation (T&E) of human-machine teaming (HMT) concepts. We differentiate the distinction between a HMT and a human using a tool, and new challenges are enumerated: Agents’ mental models are opaque, machine-to-human communications need to be evaluated, and self-tasking and autonomy need to be evaluated. We argue that a focus on mission outcomes cannot fully characterize team performance due to the increased problem space evaluated and that the T&E community needs to develop and refine new metrics for agents of teams and teammate interactions. Our IDA HMT framework outlines major categories for HMT evaluation, emphasizing team metrics and parallelizing agent metrics across humans and machines. Major categories are tied to the literature and proposed as a starting point for additional T&E metric specification for robust evaluation.

Human-system interaction is a critical yet often neglected aspect of the system development process. It is mostly commonly incorporated into system performance assessments late in the design process leaving little opportunity for any substantive changes to be made to ensure satisfactory system performance achieved. As a result, workarounds and compromises become a patchwork of “corrections” that end up in the final fielded system. But what if mission outcomes, the work context, and performance expectations can be articulated earlier in the process, thereby influencing the development process throughout? This presentation will discuss how a formative method from the field of cognitive systems engineering, cognitive work analysis, can be leveraged to derive design requirements compatible with traditional systems engineering processes. This method establishes not only requirements from which system designs can be constructed, but also how system performance expectations can be more acutely defined a priori to guide the validation and verification process. Cognitive work analysis methods will be described to highlight how ‘cognitive work’ and ‘information relationship’ requirements can be derived and will be showcased in a case-study application of building a decision support system for future human spaceflight operations. Specifically, a description of the testing campaign employed to verify and validate the fielded system will be provided. In summary, this presentation will cover how system requirements can be established early in the design phase, guide the development of design solutions, and subsequently be used to assess the operational performance of the solutions within the context of the work domain it is intended to support.

The Collaborative Human AI Red Teaming (CHART) project is an effort to develop an AI Collaborator which can help human test engineers quickly develop test plans for AI systems. CHART was built around processes developed for cybersecurity red-teaming. Using a goal-focused approach based upon iteratively testing and attacking a system then updating the testers model to discover novel failure modes not discovered by traditional T&E processes. Red teaming is traditionally a time intensive process which requires subject matter expert to study the system they are testing for months in order to develop attack strategies. CHART will accelerate this process by guiding the user through the process of diagraming the AI system under test and drawing upon a pre-established body of knowledge to identify the most probably vulnerabilities. CHART was provided internal seedling funds during FY20 to perform a feasibility study of the technology. During this period the team developed a taxonomy of AI vulnerabilities and an ontology of AI irruptions. Irruptions being events (either caused by a malicious actor or unintended consequences) which trigger the vulnerability and lead to an undesirable result. Using this taxonomy we built a threat modeling tool that allows users to diagram their AI system and identifies all the possible irruptions which could occur. This initial demonstration was based around two scenarios. An smartphone-based ECG system for telemedicine and a UAV trained reinforcement learning to avoid mid-air collisions. In this talk we will first discuss how Red Teaming differs from adversarial machine learning and traditional testing and evaluation. Next, we will provide an overview of how industry is approaching the problem of AI Red Teaming and how our approach differs. Finally, we will discuss how we developed our taxonomy of AI vulnerabilities, how to apply goal-focused testing to AI systems, and our strategy for automatically generating test plans.

Combinatorial methods have attracted attention as a means of providing strong assurance at reduced cost. Combinatorial testing takes advantage of the interaction rule, which is based on analysis of thousands of software failures. The rule states that most failures are induced by single factor faults or by the joint combinatorial effect (interaction) of two factors, with progressively fewer failures induced by interactions between three or more factors. Therefore if all faults in a system can be induced by a combination of t or fewer parameters, then testing all t-way combinations of parameter values is pseudo-exhaustive and provides a high rate of fault detection. The talk explains background, method, and tools available for combinatorial testing. New results on using combinatorial methods for oracle-free testing of certain types of applications will also be introduced

Combinatorial methods have attracted attention as a means of providing strong assurance at reduced cost. Combinatorial testing takes advantage of the interaction rule, which is based on analysis of thousands of software failures. The rule states that most failures are induced by single factor faults or by the joint combinatorial effect (interaction) of two factors, with progressively fewer failures induced by interactions between three or more factors. Therefore if all faults in a system can be induced by a combination of t or fewer parameters, then testing all t-way combinations of parameter values is pseudo-exhaustive and provides a high rate of fault detection. The talk explains background, method, and tools available for combinatorial testing. New results on using combinatorial methods for oracle-free testing of certain types of applications will also be introduced

Due to small Tactical Data Link testing windows, only commonly used messages are tested resulting in the evaluation of only a small subset of all possible Link 16 messages. To increase the confidence that software design and implementation issues are discovered in the earliest phases of government acceptance testing, Marine Corps Tactical Systems Support Activity (MCTSSA) Instrumentation and Data Management Section (IDMS) successfully implemented an extension of the traditional form of Design of Experiments (DOE), called Combinatorial Testing (CT). CT was utilized to reduce the human bias and inconsistencies involved in Link 16 testing and replace them with a thorough test that can validate a system’s ability to properly consume all of the possible valid combinations of Link 16 message field values. MCTSSA’s unique team of subject matter experts was able to bring together the tenants of virtualization, automation, C4I Air systems testing, tactical data link testing, and Design of Experiments methodology, to invent a testing paradigm that will exhaustively evaluate tactical Air systems. This presentation will give an overview of how CT was implemented for the test.

As systems and missions become increasingly complex, the roles of humans throughout the mission life cycle is evolving. In areas, such as maintenance and repair, hands-on tasks still dominate, however, new technologies have changed many tasks. For example, some critical human tasks have moved from manual control to supervisory control, often of systems at great distances (e.g., remotely piloting a vehicle, or science data collection on Mars). While achieving mission success remains the key human goal, almost all human performance metrics focus on failures rather than successes. This talk will examine the role of humans in creating mission success as well as new approaches for system validation testing needed to keep up with evolving systems and human roles.

More often than not, the data we analyze for the military is plagued with statistical issues. Multicollinearity, small sample sizes, quasi-experimental designs, and convenience samples are some examples of what we commonly see in military data. Many of these complications can be resolved either in the design or analysis stage with appropriate statistical procedures. But, to keep our work useful, usable, and transparent to the military leadership who sponsors it, we must strike the elusive balance between explaining and justifying our design and analysis techniques and not inundating our audience with unnecessary details. It can be even more difficult to get military leadership to understand the statistical problems and solutions so well that they are enthused and supportive of our approaches. Using literature written on the subject as well as a variety of experiences, we will showcase several examples, as well as present ideas for keeping our clients actively engaged in statistical methodology discussions.

More often than not, the data we analyze for the military is plagued with statistical issues. Multicollinearity, small sample sizes, quasi-experimental designs, and convenience samples are some examples of what we commonly see in military data. Many of these complications can be resolved either in the design or analysis stage with appropriate statistical procedures. But, to keep our work useful, usable, and transparent to the military leadership who sponsors it, we must strike the elusive balance between explaining and justifying our design and analysis techniques and not inundating our audience with unnecessary details. It can be even more difficult to get military leadership to understand the statistical problems and solutions so well that they are enthused and supportive of our approaches. Using literature written on the subject as well as a variety of experiences, we will showcase several examples, as well as present ideas for keeping our clients actively engaged in statistical methodology discussions.

Communication is critical both for analysts and decision-makers who rely on analysis to inform their choices. The Service Academies are responsible for educating men and women who may serve in both roles over the course of their careers. Analysts must be able to summarize their results concisely and communicate them to the decision-maker in a way that is relevant and actionable. Decision-makers understand that analytical results may carry with them uncertainty and be able to incorporate this uncertainty properly when evaluating different options. This panel explores the role of the US Service Academies in preparing their students for both roles. Featuring representatives from the US Air Force Academy, the US Naval Academy, and the US Military Academy, this panel will cover how future US Officers are taught to use and communicate with data. Topics include developing and motivating numerical literacy, understandings of uncertainty, how analysts should frame uncertainty to decision-makers, and how decision-makers should understand information presented with uncertainty. Panelists will discuss what they think the academies do well and areas that are ripe for improvement.

This tutorial will show how to compare and choose experimental designs based on multiple criteria. Answers to questions like “Which Design of Experiments (DOE) is better/best?” will be answered by looking at both data and graphics that show the relative performance of the designs based on multiple criteria, including; power of the designs for different model terms, how well the designs minimize predictive variance across the design space, to what level are model terms confounded or correlated, what are the relative efficiencies that measure how well coefficients are estimated or how well predictive variance is minimized. Many different case studies of screening, response surface, and screening augmented to response surface designs will be compared. Designs with both continuous and categorical factors, and with constraints on the experimental region will also be compared.

In the operational testing of DoD weapons systems, modeling and simulation (M&S) is often used to supplement live test data in order to support a more complete and rigorous evaluation. Before the output of the M&S is included in reports to decision makers, it must first be thoroughly verified and validated to show that it adequately represents the real world for the purposes of the intended use. Part of the validation process should include a statistical comparison of live data to M&S output. This presentation includes an example of one such validation analysis for a tactical missile system. In this case, the goal is to validate a lethality model that predicts the likelihood of destroying a particular enemy target. Using design of experiments, along with basic analysis techniques such as the Kolmogorov-Smirnov test and Poisson regression, we can explore differences between the M&S and live data across multiple operational conditions and quantify the associated uncertainties.

Goodness-of-fit (GOF) testing is used in many applications, including statistical hypothesis testing to determine if a set of data come from a hypothesized distribution. In addition, combined probability tests are extensively used in meta-analysis to combine results from several independent tests to asses an overall null hypothesis. This paper summarizes a study conducted to determine which GOF and/or combined probability test(s) can be used to determine if a set of data with relative small sample size comes from the standard uniform distribution, U(0,1). The power against different alternative hypothesis of several GOF tests and combined probability methods were examined. The GOF methods included: Anderson-Darling, Chi-Square, Kolmogorov-Smirnov, Cramér-Von Mises, Neyman-Barton, Dudewicz-van der Meulen, Sherman, Quesenberry-Miller, Frosini, and Hegazy-Green; while thecombined probability test methods included: Fisher’s Combined Probability Test, Mean Z, Mean P, Maximum P, Minimum P, Logit P, and Sum Z. While no one method was determined to provide the best power in all situations, several useful methods to support model validation were identified.

Statistical tolerance intervals of the form (1−α, P) provide bounds to capture at least a specified proportion P of the sampled population with a given confidence level 1−α. The quantity P is called the content of the tolerance interval and the confidence level 1−α reflects the sampling variability. Statistical tolerance intervals are ubiquitous in regulatory documents, especially regarding design verification and process validation. Examples of such regulations are those published by the Food and Drug Administration (FDA), the Environmental Protection Agency (EPA), the International Atomic Energy Agency (IAEA), and the standard 16269-6 of the International Organization for Standardization (ISO). Research and development in the area of statistical tolerance intervals has undoubtedly been guided by the needs and demands of industry experts. Some of the broad applications of tolerance intervals include their use in quality control of drug products, setting process validation acceptance criteria, establishing sample sizes for process validation, assessing biosimilarity, and establishing statistically-based design limits. While tolerance intervals are available for numerous parametric distributions, procedures are also available for regression models, mixed-effects models, and multivariate settings (i.e., tolerance regions). Alternatively, nonparametric procedures can be employed when assumptions of a particular parametric model are not met. Tools for computing such tolerance intervals and regions are a necessity for researchers and practitioners alike. This was the motivation for designing the R package ‘tolerance,’ which not only has the capability of computing a wide range of tolerance intervals and regions for both standard and non-standard settings, but also includes some supplementary visualization tools. This session will provide a high-level introduction to the ‘tolerance’ package and its many features. Relevant data examples will be integrated with the computing demonstration, and specifically designed to engage researchers and practitioners from industry and government. A recently-launched Shiny app corresponding to the package will also be highlighted.

Abstract. While fault testing a system with many factors each appearing at some number of levels, it may not be possible to test all combinations of factor levels. Most faults are caused by interactions of only a few factors, so testing interactions up to size t will often find all faults in the system without executing an exhaustive test suite. Call an assignment of levels to t of the factors a t-way interaction. A covering array is a collection of tests that ensures that every t-way interaction is covered by at least one test in the test suite. Locating arrays extend covering arrays with the additional feature that they not only indicate the presence of faults but locate the faulty interactions when there are no more than d faults in the system. If an array is (d, t)-locating, for every pair of sets of t-way interactions of size d, the interactions do not appear in exactly the same tests. This ensures that the faulty interactions can be differentiated from non-faulty interactions by the results of some test in which interactions from one set or the other but not both are tested. When the property holds for t-way interaction sets of size up to d, the notation (d, t ¯ ) is used. In addition to fault location, locating arrays have also been used to identify significant effects in screening experiments. Locating arrays are fairly new and few techniques have been explored for their construction. Most of the available work is limited to finding only one fault (d = 1). Known general methods require a covering array of strength t + d and produce many more tests than are needed. In this talk, we present Partitioned Search with Column Resampling (PSCR), a computational search algorithm to verify if an array is (d, t ¯ )-locating by partitioning the search space to decrease the number of comparisons. If a candidate array is not locating, random resampling is performed until a locating array is constructed or an iteration limit is reached. Algorithmic parameters determine which factor columns to resample and when to add additional tests to the candidate array. We use a 5 × 5 × 3 × 2 × 2 full factorial design to analyze the performance of the algorithmic parameters and provide guidance on how to tune parameters to prioritize speed, accuracy, or a combination of both. Last, we compare our results to the number of tests in locating arrays constructed for the factors and levels of real-world systems produced by other methods.