Integrating objective and subjective hazard risk in decision-aiding system design

A generalized model is presented to incorporate objective (hard) and subjective (soft) hazard information in automated decision-aiding systems. The model may be used with more than one hazard, of more than one type, in a given problem. Uncertainties in state measurements, dynamics, hazard extent, and hazard severity are included, as is consideration of the fact that different operators may have different concepts of what is an acceptable or unacceptable risk. By examining the tradeoffs created by these uncertainties, appropriate decision thresholds can be selected. Using an aviation case study, information gained from observation of aircraft behavior in the presence of weather was used to develop a model of weather as a soft hazard. This information could then be used in a decision aid to provide feedback on route acceptability. Introduction Real-time decision aiding and alerting systems are often used to assist human operators in controlling processes efficiently and in preventing undesirable incidents from occurring (such as a collision in a vehicle control application, or exceeding temperature limits in process control). There are a number of types of real-time decision aids, ranging from process status displays, to planning tools, to safetyand time-critical warning systems. However, all decision aids can be broadly classified as either active or passive. Active systems generate discrete decisions or commands that are communicated to the operator with the intent of modifying the process’ future state trajectory (e.g., a traffic conflict resolution command). Passive systems provide process and environment state information to the operator (e.g., depicting precipitation levels on a weather radar display) without explicit decisions being made by the automation. Thus, an active system acts as an automated decision maker (which may agree or disagree with the human operator’s decisions), while a passive system acts as an automated decision supporter. To date, the use of active systems has generally been restricted to cases in which there is a clear definition of hazardous states. For example, traffic collision risk can be defined in concrete, objective terms (e.g., no closer than 500 ft separation between aircraft), which then is translated into algorithms and decision thresholds. This can be classified as a case of objective assessment of hazard risk. Due to sensor and prediction errors, there still may be uncertainty in whether a decision to change the process’ trajectory is needed. These uncertainties, however, can also be objectively estimated and used when defining decision thresholds to balance false alarms and missed detections and optimize system performance from the human operator’s perspective. In cases in which the distinction between hazard and non-hazard is less distinct (i.e., the hazard risk is subjective), decision aids typically display the state information but leave the