Multimodal Feature Selection to Unobtrusively Model Trust, Workload, and Situation Awareness

Buchner, Savannah

Graduate Thesis Or Dissertation

Multimodal Feature Selection to Unobtrusively Model Trust, Workload, and Situation Awareness Public Deposited

Analytics

Download PDF

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/xs55md57f

Abstract

Effective human-autonomy teaming is an increasingly important challenge to ensure mission success in operational environments, such as future deep space missions. Modeling an operator’s cognitive state, such as trust, situation awareness, and mental workload, can improve performance by informing the autonomous system about the humans that they interact with. Subjective questionnaires are often used to measure these states; however, they are obtrusive and impractical for operational contexts. The use of embedded measures (e.g., measures from actions that naturally occur while completing a task) and physiological measures (e.g., heart rate monitoring, respiration, or eye-based measures) may be an effective way to unobtrusively understand cognitive states.However, these signals may be ambiguous as they can be tied to multiple cognitive states simultaneously. Models were developed to estimate trust from a variety of predictor variables, as well as models to simultaneously estimate situation awareness and workload. In order to generate these models, data from 15 subjects was collected during a spacecraft piloting and docking simulation. A LASSO-based algorithm was developed to select the important features for these models. From these features, multivariate regression models were generated, and predictive capabilities were assessed through cross validation. Having the use of multiple categories of features (e.g., embedded measures and physiological measures) allowed for increase model performance and a viable way to estimate cognitive states over models without multiple categories available. Additionally, simultaneously fitting situation awareness and workload generated more parsimonious and operationally feasible models than models that were fit to a single cognitive state. The developed algorithm, use of multiple features, and understanding of importance of simultaneously fitting can be used for generating models for future human-autonomy teaming research.

Creator