Imperfect and Unreliable Automation and Its Implications for Attention Allocation, Information Access and Situation Awareness Technical rept

摘要

In the first part of this report, four stages of information processing: attentional filtering, integration and inference, choice, and response execution, are outlined, each of which can be automated. Such automation can vary in its reliability. We distinguish between automation that is perfectly reliable, automation that fails 'catastrophically' and automation whose reliability is high, but understandably imperfect (e.g., drawing inference from inherently ambiguous data or noisy sensors). In the case of imperfect automation, we also distinguish between cases when the operator is and is not aware of the imperfection. We then describe the various human performance costs resulting from these different states and levels of unreliability, as they are relevant to the different stages of automation. We emphasize empirical data from automated attention filtering (Stage 1) in such systems as target cueing alarms, or intelligent information management. Many of these costs relate to the distribution of attention in the environment. In the second part of the report, we then describe a model of the influences on how pilots distribute and allocate visual attention in dynamic environments, in order to maintain situation awareness. The model incorporates bottom up influences on attention allocation related to event salience and information access effort, and top-down influences related to habit, to the anticipated probability of information and to the value of that information.