David B. Kaber and Jennifer M. Riley Mississippi State University
Adaptive automation (AA) can be defined as the dynamic allocation of tasks or control functions between human operators and automated control systems over time, based on the state of the human-task- environment system. Research interests in AA are expanding beyond the realm of psychomotor tasks, such as tracking in aircraft piloting, to cognitive tasks involving strategizing for automated system performance, including AA in air traffic control ( Hilburn & Jorna, 1997). This interest has been motivated by the success of AA in psychomotor tasks for addressing issues such as increased monitoring and miscalculated trust in automated system reliability, degradation of skills, and attentional resource overload due to static automation. All of these issues may potentially impede human cognition and decision-making under static automation as well; thus, AA may be helpful for cognitive task performance.
Aside from a need to examine AA in the context of different types of tasks, research questions remain as to how AA can be most effectively encouraged in human-machine systems for promoting overall performance. Different mechanisms for triggering dynamic control allocations between human and computer servers have been proposed and some empirically examined in the contexts of monitoring and tracking tasks. These mechanisms include operator performance monitoring, operator workload monitoring, systems activities monitoring for critical events, and system/operator behavior modeling. Of these mechanisms, operator workload (arousal) monitoring has been predominantly examined in experimental settings. Specifically, physiological measures including heart rate variability and electroencephalogram signals have been related to control allocations for increasing or decreasing operator arousal levels (e.g., Byrne & Parasuraman, 1996). However, the relationship of other objective measures of workload, including secondary task performance, to effective facilitation of AA has not been investigated for manipulating operator cognitive workload or indicating the potential for out-of-the-loop performance problems. Potential advantages of secondary task workload methods for AA include reliable indication of resources expended by primary task processing and applicability across a broad range of tasks. Further, such measures directly represent trade-offs in cognitive resources and changes in task performance. With these characteristics in mind, secondary tasks may be useful as effective predictors for AA based on operator primary task workload.
In potential applications of AA based on operator workload, the question can be raised as to who should decide whether, and when, automation should be invoke--the human or the computer. Some have expressed concern that humans may not be the best judges of task allocations because they are limited in decision making by individual perceptions and professionalism and may not be qualified to make control allocation decisions. Others acknowledge that while it may not be optimal for the human to have total control of automation decisions, problems can arise when they are completely removed from the loop. Furthermore, if machine control of task allocations is exclusive and impervious to human override, system safety may be comprised in situations where artificial intelligence regarding task circumstances is not superior to that of the human operator.
The question of who makes allocation decisions needs to be resolved in various contexts, including psychomotor and cognitive task performance. One potential solution may be to design systems for computer decision-making regarding function allocation based on operator workload, but allow the human to invoke automation upon system suggestions. Methods that utilize both human and computer servers in different ways for allocation decisions need to be empirically investigated.