MVILAIN@G.BBN.COM (Marc Vilain) (02/17/88)
BBN Science Development Program
AI Seminar Series Lecture
PROBLEMS IN PREDICTION AND CAUSAL REASONING
Tom Dean
Brown University
(tld%cs.brown.edu@RELAY.CS.NET)
BBN Labs
10 Moulton Street
2nd floor large conference room
10:30 am, Friday February 19
Causal reasoning involving incomplete information constitutes a topic of
growing interest in AI. Despite the enthusiasm and the volume of paper
devoted to the topic, there are still very few well defined problems. In this
talk, we will consider three problems corresponding to variations on what is
commonly referred to as THE prediction problem. In the first problem, all
events are known, but their order of occurrence is not. The task is simply to
determine what facts persist over what intervals of time. The general problem
is NP-hard, and, hence, the solutions we propose involve polynomial
approximations. In the second problem, all of the events are not known.
Here, the task is to account for the possible impact of unknown events on the
persistence of facts over intervals of time. Our solution, involving a
probabilistic theory of causation, introduces a number of problems of its own,
and, in the process of dealing with these new problems, we introduce a third
prediction problem involving unexplained but contingent events. Our analysis
of this third problem leads us to a new view of prediction which has many
elements of what is commonly referred to as explanation. We provide a precise
characterization of this problem and then consider the consequences of our new
view of prediction for existing formal accounts of causation and temporal
inference.
-------