VAL@SU-AI.ARPA.UUCP (05/14/86)
NON-MONOTONICITY IN PROBABILISTIC LOGIC
Benjamin Grosof
Computer Science Department, Stanford University
Thursday, May 15, 4pm
MJH 252
I will discuss how to formalize the notion of non-monotonicity in
probabilistic reasoning, using the framework of Probabilistic Logic
(cf. Nils Nilsson). I will give some motivating examples of types of
non-monotonic probabilistic reasoning that seem to be found in
practice. There seems to be a relationship to default inheritance,
i.e. prioritized defaults of the type used in classic example of
whether birds and ostriches fly. Next, I introduce the idea of
maximizing conditional independence, which can be thought of as
maximizing irrelevance. This can be described more simply in terms of
non-monotonic reasoning on Graphoids (cf. Judea Pearl).
I conjecture that an important type of non-monotonicity in probabilistic
reasoning may be concisely expressed in terms of conditional
independence and Graphoids. Finally, I pose as an open question how
to formulate in the above terms the non-monotonic behavior of
maximizing entropy, a widely-used technique in probabilistic
reasoning.