Dear Members of the BISC [Berkeley Initiative in Soft Computing] Group:
Few concepts are as pervasive as the concept of causality.
Causality has a position of centrality in medicine and legal
reasoning. Causality is pervasive in everyday reasoning and
decision-making. But what is widely unrecognized is that in the
enormous literature of causality what cannot be found are
theories which work in real-world settings--settings in which
information is uncertain, imprecise, incomplete or partially
true. If you know of a working theory developed by yourself or
others, please bring it to my attention. I will admit that I am
wrong if you are right.
All theories of causality founder on the rocks of
multicausality. The problem is that in real-world settings
multicausality is the norm rather than exception. I find it
helpful to talk about multicausality in the context of a
prototypical example which I call the Raincoats Problem, or RP
for short. I am a manufacturer of raincoats. I would like to
increase my sales. To this end I increase the advertising budget
by 20%. Six months later sales have risen by 10%. Was the
increase in sales caused by the increase in the advertising
budget? Can a theory of causality come up with an answer to this
simple question? This is the litmus test. What is the problem?
The problem is that the increase in sales may have been caused
by a variety of causants other than the increase in the
advertising budget--causants such as rainy weather, improvement
in economic conditions, lowering price of raincoats, etc. Some
of the causants may be known and some not. Given this setting,
the question should be restated as: To what degree was the
increase in sales caused by the increase in the advertising
budget? It is this question that cannot be answered by existing
theories. In existing theories, causality is not a matter of
degree--as it should be. What should be underscored is that the
degree of strength of causality is not the same as the
probability of causality. What is the meaning of: The
probability that the increase in sales was caused by the
increase in the advertising budget, is 0.8?
In the case of RP, a theory of causality should suggest a
procedure for assessing the degree to which the increase in
sales was caused by the increase in the advertising budget. One
such procedure may involve interviewing all purchasers of my
raincoats, to identify in each case what led to the purchase of
a raincoat? The problem is that no such procedure can be
devised. The problem becomes more apparent when the advertising
budget is increased by 20% but the sales declined by 10%.
Consider the question: Was the 10% decline in sales caused by a
20% increase in the advertising budget? How should it be
interpreted? Can an existing theory of causality deal with this
question?
As an underlying issue, causality plays an important role in
political debates. Was the financial crisis caused by Wall
Street? To improve the economy, the Federal Reserve lowered the
interest rate from 1% to 0.8%. Six months later, the economic
activity rose by 5%. Was the increase in economic activity
caused by lowering the interest rate? Will a particular
initiative cause a decrease in unemployment? Is Obama's stimulus
program a success or failure? What lessons can be drawn from RP
to answer such questions?
A sobering thought is that no theory of causality can
answer such questions. Is Obama's stimulus program a success or
failure? Republicans argue that it is a failure because the
unemployment rate remains above 9%. The democrats can argue,
counterfactually, that it is a success because without the
stimulus the unemployment rate would be over 12%. The problem is
that counterfactual arguments are much less convincing than
factual arguments. Nevertheless, the fact remains that in the
instance of political debates, most causality assertions can
neither be proved nor disproved. It would be much more realistic
to accept that basically causality is a matter of degree. Once
this is accepted, debates will become less polarized. A
fundamental conclusion is that in realistic theories of
causality, causality should be a matter of degree. To put it
another way, realistic theories of causality should necessarily
be based on fuzzy logic. Having said that, a word of caution
is in order. Introduction of degrees into theories of causality
is an undertaking which is far from simple to formalize. Comments are welcome.
Regards to all,
Lotfi
-- Lotfi A. Zadeh Professor in the Graduate School Director, Berkeley Initiative in Soft Computing (BISC) Address: 729 Soda Hall #1776 Computer Science Division Department of Electrical Engineering and Computer Sciences University of California Berkeley, CA 94720-1776 zadeh@eecs.berkeley.edu Tel.(office): (510) 642-4959 Fax (office): (510) 642-1712 Tel.(home): (510) 526-2569 Fax (home): (510) 526-2433 URL: http://www.cs.berkeley.edu/~zadeh/ BISC Homepage URLs
No comments:
Post a Comment