Date: Sat, 25 Oct 2003 12:34:11 -0700
Dear Konrad:
The raincoats problem, call it RP for short, is nowhere
nearly as trivial as one may think, and it certainly does not have an
obvious answer. Let me explain. For clarification, I will have to
restate some of the points which I had made earlier.
RP is a not a contrived or isolated example. On the
contrary, it is representative of many situations in which causality is
a factor. Its deep structure, shorn of surface details, is the
following. I activate an event AO and observe a consquent event, B. My
question is, did AO cause B and, if so, to what degree? I performed a
single experiment and have a single datapoint (AO,B). What is obvious is
that given just one datapoint ,no theory could be expected to answer the
question. But ,in reality, the problem is more complex. B is a
consequence not just of AO, the nominal event, but a multiplicity of
other events, call them collateral events, Al,A2,.... Some of these
events are visible to me and some are not. They form a network, call it
N(Al,A2,...), or N for short. In general, the consequent event B is
influenced by the collateral events Al, A2,... in different ways and to
different degrees. In the realm of law, N may be interpreted as a
network of agents.
I have partial information about N. Its principal
components are (a) world knowledge, and (b) case-based knowledge. In
the case of RP, an example of (a) is: rainy weather tends to increase
sales of raincoats.An example of (b) is: my competitor increased his
spending on advertising by 30% and his sales increased by l5%.
Based on my partial information about N, I may be able to
arrive at a conclusion like: it is likely that the 20% increase in
advertising played a significant role in increasing my sales by l0%.
The real question, then, is: can a theory of causality come up with
answers in this spirit? My position is that no bivalent -logic-based
theory of causality can do so because there is no machinery within such
theories to operate on imprecise information which is resident in world
knowledge and case-based knowledge. The reason is that such informatin
is for the most part perception-based ,and perceptions are intrinsically
imprecise in ways that put them well beyond the reach of
meaning-representation methods based on bivalent logic.
I hope that this will explain why the raincoats problem does
not have an obvious answer, and why it raises issues which are of basic
importance in the realms of law, medicine, AI and many other fields.
Cordial regards.
Lotfi
- --
Lotfi A. Zadeh
Professor in the Graduate School, Computer Science Division
Department of Electrical Engineering and Computer Sciences
University of California
Berkeley, CA 94720 -1776
Director, Berkeley Initiative in Soft Computing (BISC)
------- End of Forwarded Message
