DATAWorks Sessions Archive

This brief talk will focus on the process of human-machine trust in context of automated intelligence tools. The trust process is multifaceted and this talk will define concepts such as trust, trustworthiness, trust behavior, and will examine how these constructs might be operationalized in user studies. The talk will walk through various aspects of what might make an automated intelligence tool more or less trustworthy. Further, the construct of transparency will be discussed as a mechanism to foster shared awareness and shared intent between humans and machines.

Materials scientists, computer scientists and statisticians at LANL have teamed up to investigate how to make near real time decisions during fast-paced experiments. For instance, a materials scientist at a beamline typically has a short window in which to perform a number of experiments, after which they analyze the experimental data, determine interesting new experiments and repeat. In typical circumstances, that cycle could take a year. The goal of this research and development project is to accelerate that cycle so that interesting leads are followed during the short window for experiments, rather than in years to come. We detail some of our UQ work in materials science, including emulation, sensitivity analysis, and solving inverse problems, with an eye toward real-time decision making in experimental settings.

In support of the Navy’s effort to obtain improved outcomes through data-driven decision making, The Center for Naval Analyses’ Data Science Program (CNA/DSP) supports the performance-to plan(P2P) forum, which is co chaired by the Vice Chief of Naval Operations and the Assistant Secretary of the Navy (RD&A). The P2P forum provides senior Navy leadership forward looking performance forecasts, which are foundational to articulating Navy progress toward readiness and capability goals. While providing analytical support for this forum, CNA/DSP leveraged machine learning techniques, including Random Forests and Artificial Neural Networks, to develop improved estimates of future maintenance durations for the Navy. When maintenance durations exceed their estimated timelines, these delays can affect training, manning, and deployments in support of operational commanders. Currently, the Navy creates maintenance estimates during numerous timeframes including the program objective memorandum (POM) process, the Presidential Budget (PB), and at contract award leading to evolving estimates over time. The limited historical accuracy for these estimates, especially with the POM and PB estimates, have persisted over the last decade. These errors have led to a gap between planned funding and actual costs in addition to changes in the assets available for operational commanders each year. The CNA/DSP prediction model reduces the average error in forecasted maintenance duration days from 128 days to 31 days for POM estimates. Improvements in duration accuracy for the PB and contract award time frames were also achieved using similar ML processes. The data curation for these models involved numerous data sources of varying quality and required significant feature engineering to provide usable model inputs that could allow for forecasts over the Future Years Defense Program (FYDP) in order to support improved resource allocation and scheduling in support of the optimized fleet response training plan (OFRTP).

There are many testing situations that historically involve a large number of runs. The use of experimental design methods can reduce the number of runs required to obtain the information desired. Example applications include wind tunnel test campaigns, computational experiments and live fire tests. In this work we present three case studies conducted under the auspices of the Science of Test Research Consortium comparing the information obtained via a historical experimental approach with the information obtained via an experimental design approach. The first case study involves a large scale wind tunnel experimental campaign. The second involves a computational fluid dynamics model of a missile through various speeds and angles of attack. The third case involves ongoing live fire testing involve hot surface testing. In each case, results suggest a tremendous opportunity to reduce experimental test efforts without losing test information.

The use of statistically rigorous methods to support testing at Arnold Engineering Development Complex (AEDC) has been an area of focus in recent years. As part of this effort, the use of Design of Experiments (DOE) has been introduced for calibration of AEDC wind tunnels. Historical calibration efforts used One- Factor-at-a-Time (OFAT) test matrices, with a concentration on conditions of interest to test customers. With the introduction of DOE, the number of test points collected during the calibration decreased, and were not necessary located at historical calibration points. To validate the use of DOE for calibration purposes, the 4-ft Aerodynamic Wind Tunnel 4T was calibrated using both DOE and OFAT methods. The results from the OFAT calibration were compared to model developed from the DOE data points and it was determined that the DOE model sufficiently captured the tunnel behavior within the desired levels of uncertainty. DOE analysis also showed that within Tunnel 4T, systematic errors are insignificant as indicated by agreement noted between the two methods. Based on the results of this calibration, a decision was made to apply DOE methods to future tunnel calibrations, as appropriate. The development of the DOE matrix in Tunnel 4T required the consideration of operational limitations, measurement uncertainties, and differing tunnel behavior over the performance map. Traditional OFAT methods allowed tunnel operators to set conditions efficiently while minimizing time consuming plant configuration changes. DOE methods, however, require the use of randomization which had the potential to add significant operation time to the calibration. Additionally, certain tunnel parameters, such as variable porosity, are only of interest in a specific region of the performance map. In addition to operational concerns, measurement uncertainty was an important consideration for the DOE matrix. At low tunnel total pressures, the uncertainty in the Mach number measurements increase significantly. Aside from introducing non-constant variance into the calibration model, the large uncertainties at low pressures can increase overall uncertainty in the calibration in high pressure regions where the uncertainty would otherwise be lower. At high pressures and transonic Mach numbers, low Mach number uncertainties are required to meet drag count uncertainty requirements. To satisfy both the operational and calibration requirements, the DOE matrix was divided into multiple independent models over the tunnel performance map. Following the Tunnel 4T calibration, AEDC calibrated the Propulsion Wind Tunnel 16T, Hypersonic Wind Tunnels B and C, and the National Full-Scale Aerodynamics Complex (NFAC). DOE techniques were successfully applied to the calibration of Tunnel B and NFAC, while a combination of DOE and OFAT test methods were used in Tunnel 16T because of operational and uncertainty requirements over a portion of the performance map. Tunnel C was calibrated using OFAT because of operational constraints. The cost of calibrating these tunnels has not been significantly reduced through the use of DOE, but the characterization of test condition uncertainties is firmly based in statistical methods.

The System Usability Scale (SUS) was developed by John Brooke in 1986 “to take a quick measurement of how people perceived the usability of (office) computer systems on which they were working.” The SUS is a 10-item, generic usability scale that is assumed to be system agnostic, and it results in a numerical score that ranges from 0-100. It has been widely employed and researched with non-military systems. More recently, it has been strongly recommended for use with military systems in operational test and evaluation, in part because of its widespread commercial use, but largely because it produces a numerical score that makes it amendable to statistical operations. Recent lessons learned with SUS in operational test and evaluation strongly question its use with military systems, most of which differ radically from non-military systems. More specifically, (1) usability measurement attributes need to be tailored to the specific system under test and meet the information needs of system users, and (2) a SUS numerical cutoff score of 70—a common benchmark with non-military systems—does not accurately reflect “system usability” from an operator or test team perspective. These findings will be discussed in a psychological and human factors measurement context, and an example of system-specific usability attributes will be provided as a viable way forward. In the event that the SUS is used in operational test and evaluation, some recommendations for interpreting the outcomes will be provided.

Two responses to an expensive, time consuming test on a final product will be referred to as “adhesion” and “strength”. A screening test was performed on compounds that comprise the final product. These screening tests are multivariate profile measurements. Previous models to predict the expensive, time consuming test lacked accuracy and precision. Data visualization was used to guide a statistical engineering model that makes use of multiple statistical techniques. The modeling approach raised some interesting statistical questions for partial least square models regarding over-fitting and cross validation. Ultimately, the model interpretation and the visualization both make engineering sense and led to interesting insights regarding the product development program and screening compounds.