VAL@SAIL.STANFORD.EDU.UUCP (04/23/87)
SHOULD JOHN MCCARTHY AND ED FEIGENBAUM TALK TO EACH OTHER? Thursday, April 23, 4:15pm Bldg. 160, Room 161K Matt Ginsberg In this talk, I discuss one possible way to bridge the apparently widening gap between the "neats" and the "scruffies" in AI. According to Kuhn, a necessary step in resolving the differences between the two camps is that one attack problems of interest to the other. I attempt to do this by suggesting that the scruffy programs are doing essentialy two things: a recognizable approximation to first-order inference (such as MYCIN's backward chaining), and some sort of bookkeeping with the results returned (e.g., manipulation of certainty factors). Formalizing this bookkeeping is attractive for a variety of reasons: it will allow precise statements to be made about what the scruffies' programs are doing, and may lead to more effective implementations of their ideas. There are also advantages for the neats, since understanding some of the proposed extensions to first-order inference in this fashion appears to lead to computationally tractable algorithms for some simple non-mononotonic logics. If time permits, I will present a formalization which appears to have the properties described in the previous paragraph.