[mod.ai] Seminar - More Agents are Better Than One

VAL@SAIL.STANFORD.EDU (Vladimir Lifschitz) (10/20/86)

		MORE AGENTS ARE BETTER THAN ONE

   		      Michael Georgeff

     		Artificial Intelligence Center
		      SRI International

		  Thursday, October 23, 4pm
			  MJH 252

A recent paper by Steve Hanks and Drew Mcdermott shows how some
previous "solutions" to the frame problem turn out to be inadequate,
despite appearances otherwise.  They use a simple example -- come to
be called the "Yale Shooting Problem" -- for which it is impossible to
derive some expected results -- in this case, that the target of a
shooting event ceases living.  Such difficulties, they suggest, call
into question the utility of nonmonotonic logics for solving the frame
problem. 

In this talk, we describe a theory of action suited to multiagent
domains, and show how this formulation avoids the problems raised by
Hanks and McDermott.  In particular, we show how the Yale Shooting
Problem can be solved using a generalized form of the situation
calculus for multiagent domains, together with notions of causality
and independence.  The solution does not rely on complex
generalizations of nonmonotonic logics or circumscription, but instead
uses traditional circumscription.  We will also argue that most
problems traditionally viewed as involving a single agent are better
formulated as multiagent problems, and that the frame problem, as
usually posed, is not what we should be attempting to solve.