A planned recognizer as an independent observer for autonomous systems

When an autonomous system fails, the first symptoms are often apparent well in advance. It is possible for a human observer to detect a potentially fatal problem long before it results in severe dysfunction. In this paper we propose an "event-recognition" module in the role of the independent observer. This module receives an abstracted list of actions from the control program and generates a causal interpretation of these actions. By comparing these observed action plans with a list of abstracted goals, it can analyze the behavior and detect flaws in reasoning even before they result in failure. This system, which functions as an adjunct to a knowledge based system in an autonomous environment, adds reliability by operating at a hierarchically abstracted level (much smaller size). It can also be used to accelerate the error-recovery process. In order to build such a plan recognizer, it is necessary to reason about events even before they are complete. The current temporal reasoning representations do not permit reasoning about ongoing events, and a more powerful representation is developed for this purpose.