automation, trust, robotics, human-machine interaction
The present work examines whether user's trust of and reliance on automation, were affected by the manipulations of user's perception of the responding agent. These manipulations included agent reliability, agent type, and failure salience. Previous work has shown that automation is not uniformly beneficial; problems can occur because operators fail to rely upon automation appropriately, by either misuse (overreliance) or disuse (underreliance). This is because operators often face difficulties in understanding how to combine their judgment with that of an automated aid. This difficulty is especially prevalent in complex tasks in which users rely heavily on automation to reduce their workload and improve task performance. However, when users rely on automation heavily they often fail to monitor the system effectively (i.e., they lose situation awareness - a form of misuse). However, if an operator realizes a system is imperfect and fails, they may subsequently lose trust in the system leading to underreliance. In the present studies, it was hypothesized that in a dual-aid environment poor reliability in one aid would impact trust and reliance levels in a companion better aid, but that this relationship is dependent upon the perceived aid type and the noticeability of the errors made. Simulations of a computer-based search-and-rescue scenario, employing uninhabited/unmanned ground vehicles (UGVs) searching a commercial office building for critical signals, were used to investigate these hypotheses. Results demonstrated that participants were able to adjust their reliance and trust on automated teammates depending on the teammate's actual reliability levels. However, as hypothesized there was a biasing effect among mixed-reliability aids for trust and reliance. That is, when operators worked with two agents of mixed-reliability, their perception of how reliable and to what degree they relied on the aid was effected by the reliability of a current aid. Additionally, the magnitude and direction of how trust and reliance were biased was contingent upon agent type (i.e., 'what' the agents were: two humans, two similar robotic agents, or two dissimilar robot agents). Finally, the type of agent an operator believed they were operating with significantly impacted their temporal reliance (i.e., reliance following an automation failure). Such that, operators were less likely to agree with a recommendation from a human teammate, after that teammate had made an obvious error, than with a robotic agent that had made the same obvious error. These results demonstrate that people are able to distinguish when an agent is performing well but that there are genuine differences in how operators respond to agents of mixed or same abilities and to errors by fellow human observers or robotic teammates. The overall goal of this research was to develop a better understanding how the aforementioned factors affect users' trust in automation so that system interfaces can be designed to facilitate users' calibration of their trust in automated aids, thus leading to improved coordination of human-automation performance. These findings have significant implications to many real-world systems in which human operators monitor the recommendations of multiple other human and/or machine systems.
Doctor of Philosophy (Ph.D.)
College of Sciences
Length of Campus-only Access
Doctoral Dissertation (Open Access)
Ross, Jennifer, "Moderators Of Trust And Reliance Across Multiple Decision Aids" (2008). Electronic Theses and Dissertations. 3754.