Explanatory Diagnosis: Conjecturing Actions to Explain Observations

Our concern in this paper is with conjecturing diagnoses to explain what happened to a system, given a theory of system behaviour and some observed (aberrant) behaviour. We characterize what happened by introducing the notion of explanatory diagnoses in the language of the situation calculus. Explanatory diagnosesconjecture sequencesof actions to account for a change in system behaviour. We show that determining an explanatory diagnosis is analogous to the classical AI planning task. As such, we exploit previous results on goal regression in the situation calculus to show that determining an explanatory diagnosis can be achieved by regression followed by theorem proving in the database describing what is known of the initial state of our system. Further, we show that in the case of incomplete information, determining explanatory diagnoses is an abductive plan synthesis task.