A Multi-Timescale Analysis of Reward Functions Learned from Human-Automation Collaboration
Abstract
We have developed a system combining mobile research video games with mobile, continuous sensing to study human-autonomy teaming. We propose that two of the primary benefits of such an approach, relative to the more traditional, laboratory-based approach, are that more people can be assessed for longer periods of time. An implication of this proposal is that analyses of the multi-timescale properties of human-autonomy teaming should reveal potentially valuable insights. Therefore, we conducted a preliminary analysis of 45 subjects who have completed a 180-day study. We operationalized human-autonomy teaming in terms of a reward function and assessed the ways in which reward functions varied at multiple timescales and how mobile sensing data streams related to decision-making at different timescales. Our findings clearly indicated the presence of individual differences in the importance of multiple timescales, suggesting that, indeed, the traditional laboratory approach to studying human-autonomy teaming may overlook important information. We argue that future work should find ways to combine the strengths of approaches such as ours with the strengths of more traditional data collection techniques.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 26, 2023
- Accession Number
- AD1217509
Entities
People
- Alexander Danvers
- Andrew Campbell
- Evan C. Carter
- Javier Mendoza
- Lydia Tapia
- Matthias Mehl
- Torin Adamson
- Yazied Hansan
Organizations
- United States Army Research Laboratory