Exceptional Subclasses in Qualitative Probability
This work solves a specific technical limitation in default reasoning for AI, but it is incremental as it builds directly on an existing formalism.
The paper addresses System Z+'s inability to sanction inheritance across exceptional subclasses by proposing an extension that extracts additional conditions from defaults, without altering database consistency.
System Z+ [Goldszmidt and Pearl, 1991, Goldszmidt, 1992] is a formalism for reasoning with normality defaults of the form "typically if phi then + (with strength cf)" where 6 is a positive integer. The system has a critical shortcoming in that it does not sanction inheritance across exceptional subclasses. In this paper we propose an extension to System Z+ that rectifies this shortcoming by extracting additional conditions between worlds from the defaults database. We show that the additional constraints do not change the notion of the consistency of a database. We also make comparisons with competing default reasoning systems.