"It's One Thing After Another": Supporting the Management of Frequent and Nested Interruptions in High-Tempo Multi-Agent Operations

Abstract

The growing number of semi-autonomous machine agents in many safety-critical application domains brings with it an increase in the frequency with which human supervisory controllers need to interrupt ongoing tasks and lines of reasoning to handle unexpected and/or time-critical problems and requests. These disruptions may occur at inopportune times, such as being interrupted again immediately after completing a previous interrupting task (serial interruption) or while still handling a previous interruption (nested interruption). Frequent interruptions can lead to errors and delays which threaten safety in time-sensitive event-driven domains such as aviation and medicine. Successful teaming of operators with multiple machine agents therefore requires a better understanding of, and support for attention allocation and interruption management (IM). The goals of this dissertation were to 1) identify and analyze the challenges that operators encounter at various stages of interruption management (IM) when handling frequent and nested interruptions, and 2) develop and evaluate a set of candidate displays to address the observed challenges.

To this end, three human-subject experiments were conducted. The first two experiments focused on identifying the difficulties faced by operators when detecting, interpreting, and switching between frequent and nested interruptions in a supervisory command and control task. Frequent and nested interruption notifications were less likely to be acknowledged, compared to less frequent and non-nested ones. Participants also struggled with the appropriate scheduling of incoming tasks and took longer to switch to nested interrupting tasks of higher urgency, compared to both single and serial interruptions. The longer switch time resulted from delays at the earlier stages of detection and interpretation of notifications as well as a resistance to switch away from the ongoing task, even for highly urgent interrupting tasks. In the third and final study, two candidate displays were developed and tested to address issues with poor scheduling of pending tasks. The first display involved automatic sorting of incoming task notifications by level of urgency; the second candidate consisted of a color- and location-based visualization of the relative urgency levels for the ongoing and interrupting task to support task prioritization and switching. The visualization of relative task urgency improved overall performance, led to decision-making accuracy, and resulted in more efficient prioritization of ongoing and interrupting tasks. At the same time, it involved a greater risk of failing to notice misclassifications made by imperfectly reliable automation.

The theoretical contribution of this research is a better understanding of the process of interruption management. More specifically, challenges and performance breakdowns experienced in the detection, interpretation, and integration of frequent and nested interruptions were identified. In contrast to what is suggested by current IM models, our findings show that interruption management is not a linear process, but one where behavior and performance at one stage depends on anticipated and experienced difficulties at both earlier and subsequent stages. In addition to identifying and analyzing challenges with handling frequent interruptions, this work also addresses said challenges and provides empirically based guidance on the design of interruption-resilient interfaces. From an applied perspective, findings from this line of work will help reduce the attentional demands and improve the safety and performance of human-machine teams, and the well-being of human operators in a variety of complex event-driven application domains like aviation and healthcare.