COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	counter[f85,jmc]		The AI uses of counterfactuals
C00013 ENDMK
C⊗;
counter[f85,jmc]		The AI uses of counterfactuals

	Human speech and writing abounds in counterfactual conditional
sentences.  Philosophers have studied them with the object of finding
their truth conditions.  We wish to investigate their possible usefulness
for AI.  Perhaps they can be regarded as a frill of natural language
and avoided in AI.  Some of their uses can certainly be replaced by
suitable uses of ordinary conditionals.

	Consider the example from (Ginsberg 1985) ``If I had some oars,
I'd be able to cross the river''.  We can consider replacing this by
the ordinary conditional ``If I get some oars, I'll be able to cross
the river''.  Are these equivalent?  Can this replacement avoid the
problematical counterfactual?

	The answer seems to be that for some purposes the replacement
works fine, and these seem to be the case of greatest use in AI ---
at least in the kind of AI program that has been written so far.
Certainly the replacement is ok for straightforward planning.

	However, suppose one is considering the non-monotonic
predicate completion step leading to the conclusion ``I could cross
the river only if I had some oars''.  Hmm, this looks ok too.

	What about, ``

****
Ginsberg makes "what if" questions the main use of counterfactuals.
This is like a counterfactual with a kind of universal conclusion,
e.g. ``If I had some oars, what would follow?''.  It seems to me
that the universal conclusion counterfactuals cannot replace all
others, because the conclusion offered is often needed to specify
the scope of the approximate theory.  Consider Lewis's ``If Otto
had come to the party it would have been a bad party'' and
compare it to ``What if Otto had come to the party?''.


oct 14

draft beginning of note

THE AI USES OF COUNTERFACTUALS

	Humans use counterfactual conditional sentences a lot, and
it may be that AI systems should also use them.  There is an extensive
philosophical literature concerning counterfactuals (see Lewis 197x)
that is mainly concerned with how they should be assigned meanings.
Most of this literature seems to take for granted the assumption that
this is the only important question about them --- surely a mistake
from the AI point of view.  This leads the philosophers to base
their discussion on examples that do not arise naturally.  It is
my contention that an AI discussion of counterfactuals must begin
by discussing how they are useful in problem-solving thought or
human thought in general.

	It would seem that much human use of counterfactuals is of little
use for AI.  We humans spend much time in vain wishes and regrets.  ``If
only I really understood quantum mechanics, I'd be able to see if my ideas
about causality can be applied to it.''  ``If only I hadn't wasted time on
politics when I was a student, I'd really have learned quantum
mechanics''.  Even if we suppose that there is no intention to further
study quantum mechanics or to avoid politics or other time-wasters
in the future, a human will still think such sentences and even utter
them.  So what are the uses of counterfactuals, perhaps even the above
examples.

	One more example.  Consider Quine's chestnuts, ``If Ceasar had
been in command in Korea he would have used the atomic bomb'' and
``If Ceasar had been in command in Korea he would have used catapults''.
We can imagine the first arising in a discussion of the Korean war or
of history and used to throw light on either.  The second is only a
joke derivative from the first.  If a theory of counterfactuals puts
the two on an equal footing, then I contend there is something wrong
with the theory.  Anyway that's the intuition that we shall try to
formalize in this paper.

	Suppose someone offers the following argument on behalf of
fewer politically motivated restrictions on the American military
establishment and fewer scruples about weapons in the doctrine taught
in the service academies and includes the first of the above
counterfactuals.

		Ceasar in Gaul had no orders from Rome that would have inhibited
	what weapons he could use and no scruples about using what he had.
	The concept that use of certain weapons might lead to world disaster
	had not arisen.  By the time of the Korean war such inhibitions and
	scruples had arisen, and they might have prevented a decisive U.N.
	victory.  By the time of Vietnam, such inhibitions and scruples had
	become vastly stronger in America, and they might be taken as the
	reason for our defeat.

What role might the counterfactual play?

	It seems to provide a ``model'', Ceasar in command in Korea
that one can run to get information about the effects of policies.
This seems too elaborate a use for the present state of AI.  Perhaps
we can find a simpler use of a counterfactual.

	``If we had some ham, we could have ham and eggs.''

	This is more promising.  The hearer might either say
whether he would like ham and eggs or express an opinion on whether
ham can be obtained or suggest a substitute.  However, it looks like
we could replace the counterfactual conditional by an ordinary
conditional without losing anything, namely

	``If we can get some ham, we can have ham and eggs.''

	How about

	``If it had rained, you would have been sorry you didn't bring
your umbrella''?

	``If there had been another train coming on the other track,
your beating the first train to the crossing would have been fatal.''

Both of these involve combining features of the actual situation
with something that might have been the case in order to show the
inadequacy of strategy.  It can also be used to show the correctness
of a strategy as in

	``If it had rained, you'd have been glad I brought my umbrella''.

One use of counterfactuals is to learn more from an experience
than can be learned without them.

	Consider

	``If Oswald hadn't shot Kennedy someone else would have''.

It is used to make an assertion about fate or an assertion about a
plot.  The plotter might assert

	``If Oswald hadn't got him, my other man would have''.

The above are all concerned with causality.  How about

	``If wishes were horses beggars would ride''.

This is usually said in connection with an expression of a wish that
the hearer considers vain.  It is advice to think more to the point.

Perhaps many counterfactuals are used in the larger scale exploration
and explanation of phenomena.  If ducks were the size of turkeys, they
wouldn't be able to fly.