Details

Autor(en) / Beteiligte

• Sanneman, Lindsay
• Shah, Julie A.

Titel

Validating metrics for reward alignment in human-autonomy teaming

Ist Teil von

• Computers in human behavior, 2023-09, Vol.146, p.107809, Article 107809

Ort / Verlag

Elsevier Ltd

Erscheinungsjahr

2023

Link zum Volltext

Quelle

Elsevier ScienceDirect Journals Complete

Beschreibungen/Notizen

• Alignment of human and autonomous agent values and objectives is vital in human-autonomy teaming settings which require collaborative action toward a common goal. In order to characterize the impact of alignment on team dynamics, a means to quantify and evaluate this alignment is critical. However, few alignment metrics have been proposed to date, and none have been validated through human-subject experiments. Individual and team goals can often be represented as reward functions the team must jointly optimize, so in this paper we focus on the alignment of these reward functions in particular. We propose four individual assessments and associated metrics for alignment between reward functions, along with three composite metrics. We then validate these metrics through a survey-based human-subject experiment. Results from a factor analysis of the data collected in the experiment provide empirical support for the existence of two latent constructs related to alignment, which we name “feature alignment” and “policy alignment,” and a subsequent partial least-squares analysis further supports the existence of a second-order formative construct, which we name “reward alignment.” These findings suggest the applicability of the proposed metrics for future evaluation of reward alignment, as well as its feature- and policy-based components. We also provide potential directions for future research into alignment in human-autonomy teaming. •There are two primary constructs which comprise reward alignment formatively.•These two constructs include “Feature Alignment” and “Policy Alignment”.•The FR, FS, PE, and BD metrics can be used to assess alignment of reward in teams.•The task and respondent burden should be considered when applying these metrics.