COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.require "let.pub" source
C00008 00003	grumble about not acknowledging that I have been pursuing this
C00015 ENDMK
C⊗;
.require "let.pub" source
∂AIL Professor Patrick Hayes↓Department of Computer Science
↓University of Essex↓Colchester, Essex↓ENGLAND∞

Dear Pat:

	I agree with the general trend of your %2Naive Physics
Manifesto%1 as you might expect from my IJCAI paper which
makes some of the same points.  It even makes the point that
the situation calculus is not epistemologically adequate
for describing concurrent action.  (I wonder if our 1969 paper
makes that point; I had examples like the one you attribute
to Burstall in mind even then).  I also like the idea of sticking
to first order logic for epistemological purposes.

	I now think that the lack of criteria for success may
be one of the major reasons that few people have entered  the
field of formalization of common sense.  I think this is not
so difficult as your paper makes out.  The basic criterion
is that desired conclusions should follow from reasonable
axiomatizations (that meet as many as possible of your criteria)
by proofs that are regarded as reasonable formalizations of
informal reasoning.  It would be very useful to have some
good challenge problems, e.g. in the following areas:
concurrent events with weak interaction, sources of knowledge,
construction and destruction of objects, shape.

	I think any problem that can be described verbally
and solved mentally is a good candidate.  The trouble is that
much intelligent physical action involves using the world
as a representation of itself.
Here are some concrete problems I have considered:

1. The three wise men.  I enclose two memos.

2. Mr. S and Mr. P.  Described in one of the memos.

3. Trading by telephone.  This one has not been given a precise
English formulation.  The idea is that it involves purpose and
concurrent action, but doesn't involve common sense physics or
using the world as a self-representation, i.e. doesn't involve
using observation as a logical step in a proof.  I enclose
a summary of Robert E. Filman's thesis which is concerned with
using observation as a reasoning step.

4. Elaborating the blocks world beyond what has been discussed
in the AI literature.

5. The McCarthy airline reservation system is somewhat related
to this.

	Would you be still interested in joining the AI and
philosophy working group at the Center for Advanced Study in
the Behavioral Sciences (at Stanford) for the academic year
1979-80?  I would now like to have you, because your current
work is very directly relevant - even if still more programmatic
than concrete.  Gardner Lindzey is on a trip, and I won't be
able to ask him till September, and Preston Cutler, the second
man there, doubts that we could squeeze you in, but I would
like to try.

.reg

	I don't think the right way to handle defaults is by
having the language talk about its own expressions.  Even a
dog can handle defaults, so the mechanism should be an elementary
one.  From this point of view, even circumscription, though
more elementary than metamathematics, is suspiciously high
order.

	I'm not sure whether circumscription or the metamathematics
you suggest is right for the example of the Texan shaking your
belief in the number of lines.
grumble about not acknowledging that I have been pursuing this
approach since 1959

the ratio of facts to concepts needs to be larger than two, i.e.
the average concept should do somewhat more than connect two
others.  I don't claim this is a precise idea, because counting
facts is not a well-defined operations, i.e. how do
we make sure that p≡q and ¬(¬p∨¬q) are counted as two facts each.
Anyway I have my doubts about density being the right criterion.

I don't understand why comprehension axioms should be the typical
existence axiom wanted by naive physics.  They never assert physical
existence, only the existence of sets corresponding to properties, which 
may be empty.

agree about meaning depending on the whole formalism

regard Wilks's example as showing that haggling over "the meaning"
is likely to be unproductive

agree that %2piece of wood%1 and "being wet" are theoretical constructs

I would prefer that papers I send you be referred to by name and
computer file if they haven't been published in a journal or proceedings
rather than just as private communications, e.g. I sent you
MINIMA[S77,JMC]@SU-AI.

The proliferation of concepts (p. 19) can fail to slow down if
the formulation is bad - forcing special concepts when general
concepts could have been formulated.

Would you buy (p.20) two orders of magnitude?

measuring scales (p.21) is untypically easy

The key question about shape (p. 22) is to what extent a verbal
theory of shape is possible.  To what extent can people reason
verbally about shape without the objects in front of them
for observation and even manipulation.  Certainly some verbal
reasoning is possible, but most of  our conclusions about
shape depend on observation and non-verbal internal models.

homology theory and differential geometry (p.22) are
entirely irrelevant to common sense physics.  Exercise:
say why.

From the beginning I have pointed out the limitations of the
situation calculus especially as epistemology.  We seem to
be still committed to it metaphysically.  It is a bit unfair
for you to write as though I were still committed to it
and to put yourself (p.23) in the place of a revolutionary when
like me, you only have an aspiration for the revolution.
As to the metaphysical adequacy, note that it makes sense to
ask "Where was A at the moment B arrived in San Francisco".
There just may be no need or possibility of answering it.

I enclose a memo aspiring to a "Relativistic theory of automata"
which attempts to get away from the single state even as a
metaphysical object.
It seems possible, but it isn't clear that it is worthwhile.  An
epistemological formalism of partial states (not necessarily local
in space) is also needed.  Note that you and I may be interacting
on the telephone between Essex and Stanford,
and Brooker and Knuth may also be conversing by telephone.  Further
suppose that both conversations are about a joint effort to draft
a computer science curriculum.  We need partial states in order
to describe separately the progress of Hayes-McCarthy and Brooker-Knuth,
but this separation is not spatial.

Weren't histories (p.24) also mentioned in our joint paper?

I don't see why "every patch of space" (p.25) shouldn't be
counted as a "well-defined place".  One needn't have constants
in the language for every well-defined place.

"Opaque to the causal consequences"?  The Hiroshima bomb exploded
at an altitude of 2500 feet, but its causal consequences were not
confined to that place.

I hope "structured formalization" (p.31) doesn't turn out to be as empty
a phrase as "structured programming".

The sentence qf(x) ≡ qf1(x) ∨ ... ∨ qfn(x)
may be obtainable from some kind of circumscription, but it won't be
identical with the one in my circumscription paper.

I don't understand (p.32) what you mean by "comprehension axiom".
What you say just doesn't seem to match the usage of the term
in set thoery.

Now I understand the way you are using comprehension.  You mean
using it to define a set of points in 3-dimensional space by
a property.  This is far from what Frege or logicians generally
meant by "unrestricted comprehension".

(McCarthy 1959) makes similar arguments.  p.34