RPD-Grounded Simulated Users for Pre-screening AR Training Interface Designs: A Unity ML-Agents Approach
Files
Date
Authors
Supervisor
Item type
Degree name
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Evaluating augmented reality (AR) training interfaces typically requires recruiting domain practitioners, coordinating controlled studies, and incurring the logistical and ethical costs of human-subject testing. This paper presents a pre-screening framework that uses simulated users to surface interface design trade-offs before human trials begin. Agents are built with Unity ML-Agents and grounded in the Recognition-Primed Decision (RPD) model, with distinct reward structures derived from Cognitive Task Analysis of paramedic diagnostic reasoning encoding novice and expert decision strategies. We evaluate three interface configurations varying in button density, spatial layout, and labelling across both agent types, and characterize behaviour using behavioural interaction indicators computed from interaction logs and conceptually organized using a structure inspired by the NASA Task Load Index (NASA-TLX). Results show that dense dispersed layouts impose the highest temporal and search cost on novice agents, while deceptive clustering imposes the highest temporal and effort cost on expert agents. Search Complexity Indicator diverged between cohorts in a pattern consistent with RPD-predicted selective attention: expert agents narrowed search as complexity increased while novice agents maintained broad attention throughout. These patterns were consistent across behavioural indicator scores and raw interaction metrics including path entropy, inter-click intervals, and idle ratios. The framework requires no human participants, produces reproducible behavioural traces across conditions, and is intended as a complement to human-subject evaluation rather than a replacement. Validation against human self-reported ratings and objective performance data is identified as the essential next step.
