COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	Ideas and notes on conversations
C00046 ENDMK
C⊗;
Ideas and notes on conversations

1983 March 11

1. People concerned with concurrent programs prove fairness about
their programs.  Is it possible to include statements in the
program demanding fairness and leave it to the compiler to say
how this is to be accomplished?

2. Conversation with Les Lamport
Interval logic by Richard Schwartz and Michael Melliar-Smith.

Pierre Wolper on generating synchronization from temporal logic
specs.  Alan Emerson student of Ed Clark.

*Quantum 1988 Aug 5

Quantum mechanics on a graph.
We have a particle moving on a graph.
What's the momentum operator?  Remember the theorem that the
quantum mechanical commutation relation cannot be satisfied by
finite matrices.  What about some other kind of finite system?
We can try the condition that the derivatives of the wave
function moving into a vertex sum to zero.  However, there
may be other forms of dynamics that simply moving in a scalar
potential.  Is there a vertex condition, e.g. a linear relation
among the wave function (assumed continuous at the vertex)
and its derivatives.  I'm doubtful about a relation involving
the current as the main relation, because of the conjugates.
It should be a consequence of the main relation.  What would
a vector potential reduce to in this case?

*CAI - 1988 Aug 5
For CAI.  Besides the student and the teacher, there are other
simulated students with knowledge and personalities.  The real
student develops competitive (and possible co-operative) relations
with the simulated students.  The theory is that he will be
encouraged to learn by the possibility of beating them.  Since we
don't care about the self-esteem of a simulated student, we can
optimize the reinforcement of the real student.
Some aspects of this should be patentable.

*Elephant - 1988 Aug 7
Slogan: Programming without one hand tied behind one's back.
We want to be able to say, as a specification, or possibly
in a program, that when we return from a subroutine, the
ac should contain f(contents(ac,when the subroutine was
entered corresponding to the exit)).
We need to be able to refer to all past events.
We need sets (of past events) and maximal elements.
Palindrome patterns (added to regular exps),
(pushdown automata?) (return from subroutine).
when I entered.  The matching entry.
The times in Old Elephant are just labels for what is
accessed through them.  With enough language, the times
themselves may not be needed.
Consider a Spider program.
Is "the corresponding entry" the linguistic construction
we need or is it just something programmed with the
needed construction?

*spider - 1988 Aug 7
List heuristics for spider before trying to determine framework for
program.
We also need a list of concepts.
1. predicates on moves.
2. predicates on columns in a position
movable
a move to a column makes the column not movable, it may be immovable
directly, but may be still movable because there are other places
to put the covering card

heuristics: we prefer moves that don't make any column immovable.
we prefer moves from short columns.
If all moves make a some column immovable, we prefer covering a
high rank card.
We prefer the move that is part of the longer sequence.
If we would make a move that gets more information after
a given sequence of moves that don't provide information,
make, if possible, the information providing moves first.
At least if there is any possible outcome of the
information providing move sequence that would cause
us not to make the other sequence.
We prefer  move that leaves a card of a certain rank
available to a move that makes it unavailable - given
the choice and ceteris paribus.

1988 Oct 12
	In a spider solitaire position, let there be two moves A
and B that can be made in either order.  Suppose A does not turn
over a card, and we can determine that if we made A, then B would
be the next move made.  Suppose B does turn over a card.  Then B
should be done first, since it gains information.  The problem is
to express this rule in a general way so that a program could
take it into account in addition to its other rules.  Note that
explaining the principle requires little information about
spider.

Oct 17
	Ascending to the metalevel by making an assertion about
the quotation of any subexpression of anything that has come up
in memory should be worthwhile.

Oct 19
	Genius and scientist social service agency
One reads about geniuses who commit suicide, e.g. Turing.  Suppose
there were an agency that would try to help genius in difficulty.
That Turing was about to commit suicide, if he did, was probably
unknown, but his friends probably knew he was in difficulty.
They could have asked the agency to help.  Maybe he would have
rejected help, and maybe nothing could have been done.  On the
other hand, they might have saved him for many more years of
productive activity.

More examples: Sylvia Plath, Janis Joplin, Arthur Koestler?,
Ramanujan, the Santa Cruz 11 year old college graduate.

What about marginal geniuses, e.g. Gunkel?  Obviously that
depends on resources.

Undoubtedly the agency would have to do triage, deciding it
couldn't help.

The first step might be a survey of biographical information to
try to identify cases which might have been helped.

The help should be unpublicized in particular cases.

1988 Nov 22
NYT has article referring to Phys. Rev. Letters paper by Kip Thorne
and others suggesting time travel through wormholes may be possible.
I should explore the hypertime idea to see if there can be relaxation
to ordinary space-time solutions that correspond to time travel.

We also need to ask what kinds of differential equations permit
the evolution of intelligence.

1988 Nov 28
Maybe it will help with YSP to postulate that changes must have
causes.  If the gun becomes unloaded there must be a cause of that.

1988 Dec 5
is-giving(x) ⊃ gave(giver(x),gift(x),recipient(x))
but what about time(x)?

1988 dec 5
If you say, ``Ponce de Leon was searching for the fountain of youth'',
and your hearer doesn't know the fountain of youth is nonexistent,
you have an obligation to supply that fact.

1988 dec 11
What science tells us that is relevant to philosophy.

	Almost all philosophers today believe in the conclusions
of modern science.  However, it seems to me that they ignore
much of what science has to tell us when doing philosophy ---
even when doing philosophy of science.

	1. Science tells us that we have evolved by natural selection
in such a way that the phenomena directly observable to us are many
levels of complexity away from the basic phenomena of the universe.
What is observable to our senses is metaphysically and physically
accidental.  By the first I mean that there is no philosophical principle
that says that intelligence and curiosity can only evolve in
forms that have access to the fundamental structure of the world.
The second means that our particular senses are also an accident.
In particular, there is no reason why there can't be structures in
the world that are fundamentally unobservable.  There can also be,
and experimental science tells us there are, structures that are
ultimately observable but which remain unobserverved for millenia
in spite of our best efforts.

	Given these facts, regarding the world as a structure of
appearance seems scientifically perverse.  It would be particularly
perverse to build robots that regarded the world as a structure of
their appearances.  AI poses a particular problem to metaphysics.
Build a language for the robot to use to express what we tell it
about the world and what it discovers and which is sufficiently
general that it doesn't presuppose anything that scientific
research, human or its own, might subsequently overturn.  That is,
it may have beliefs that it comes to change, but its very structure
shouldn't preclude changes.  In other words, it should reach
a level where it doesn't have to be educated by brain surgery.

	2. Our morality has arisen by genetic and social evolution.
This includes a set of second order predicates on moral system,
giving a set of desiderata that are not simultaneously satisfiable.
For this reason we can accept or reject behaviors that are based
on our biology.  (Make this better or leave it out).

	3. Cryptography tells us that long texts are usually not
ambiguous.  This tells us that we can indeed get semantics from
syntax.

∂24-Dec-88  0948	JMC 	simplifying autoepistemic reasoning
To:   VAL@SAIL.Stanford.EDU, AIR@SAIL.Stanford.EDU,
      bmoore@SRI.COM
Bob Moore's ``If I had an elder brother I'd know it'' led
him to autoepistemic logic with its stable expansions
defined as fixed points.  Consider ``If I had an elder
brother, five minutes ago there would have been a sentence
in my database of the form  elder-brother(jmc) = x''.  Using
this fact does not involve a fixed point construction.

Also consider ``If I had an elder brother, it would have
been a logical consequence of my database of five minutes
ago''.  Using this involves a minimization rather than
a fixed point.

Subjectively, it seems to me that my own autoepistemic
reasoning is like these two examples and doesn't involve
fixed point considerations.

Comments?

1988 Dec 24
	Would the end of the cold war result in an increase in
other conflict that might even lead to nuclear wars?

Considerations:

1. If the Soviets would co-operate, there could be an effective
international police force.  This might lead to the evil of
world government.

2. An outlet for aggression may be better than letting it build
up.  Soccer riots, pro and anti abortion riots, riots between
Armenians and Azeris, protestants and catholics in northern
Ireland, Bloods and Crips.  The riots have less chance of turning
into wars if they are about something that doesn't interest most
of the population.  Also if the main conflicts are between
rioters and police rather than between opposing gangs of rioters.

3. The big danger appears to occur when intellectuals, i.e. people
with access to media, become involved as participants.  Perhaps
there doesn't need to be gang violence, i.e. even the most
violent can be restrained.

4. A scientific study of the matter is worthwhile.

5. Perhaps there ought to be a peace movement.

1989 Jan 5

I now suspect that a limitation of philosophy is an insistence on
complete theories.  A causally complete theory is one that predicts
the consequences of the future of any initial conditions, e.g.
gravitational astronomy.  The theory of clocked digital circuitry
is an incompletable theory, because flipflops can receive
ambiguous inputs.

1989 Jan 15
	Names provide access to their values.  However, they have
two other equally important roles.  Names
can be substituted for variables.  In general expressions without
looking down (on their property lists) for their values.  Moreover,
occurrences of names can be tested for equality, and this also
leads to further action without looking at the values associated
with the names.  This is at variance with some remark about names
that I vaguely recall Newell making.

1989 Jan 15
	Parametrizing circumscription.
	Suppose we define circum1(A) = (λ P;Z)circum(A;P;Z).  We
can then consider two minimal models by writing
circum1(P1;Z1) ∧ circum1(P2;Z2).  We can also write such things
as (∀P Z)(circum1(P;Z) ⊃ ...) and (∃P Z)(circum1(P;Z) ∧ ...).
This came up in Kurt Konolige's seminar in response to some remarks
that default logic could do some things circumscription couldn't.

1989 Feb 4

Some Linguistic Problems from the Standpoint of Artificial Intelligence

A relatively decontextual formula is one in which the subformulas have
the same meanings in their contexts as they have in the context of
the formula as a whole.

1989 Feb 5

A possible project is a database for an expert system for foreign
travel.
1. travel documents
2. list countries to be visited
3. get back union of requirements
4. purpose of trip, activities in countries
5. citizenship, previous citizenship
6. must send passport and application to consulate in good time
7. check for strike at Heathrow
8. which airports are subject to most delays
9. How to combine visits to minimize travel and cost.
How far did the Xerox project get?

1989 Mar 6

The causes of national poverty include
tyranny,
corruption,
incompetence,
bad decisions on top,
ignorance,
lack of technological knowledge.
It appears that tyranny worsens and democracy alleviates the other
problems.  Lack of technology is a secondary problem for two reasons.
(a) The USSR is supposedly 5 years behind the West in technology.
However, it is 50 years behind in standard of living.
(b) The lag in technology is is mostly a consequence of the tyranny
with its resulting corruption and incompetence.

1989 Mar 8
	We consider the ``attache case problem'' as it relates to
approximate theories.
Here are some facts.
1. I have just parked my car at A.
2. I must go to a meeting at C for which I am already late.
3. After that I want to be at my office at B.
4. I will need my heavy attache case at B.
5. B is between A and C, but is somewhat out of the direct path.
6. I have a bike in a bike locker at D near B.  D is more on
the direct path from A to C than is B.
7. It is inconvenient to carry the attache case on the bike.

Solution: Take the case to bike locker and store it in the locker
while going to C on the bike.  This solution involves using the
bike locker for the nonstandard purpose of storing the attache
case.  It didn't come to mind immediately.

Here are some stages in the reasoning (all done while driving and
parking the car).

Highest level plan:
Go from A to C.
Attend meeting.
Go from C to B.

Reaction: What about attache case?  It imposes a kind of constraint.

Alternatives:	(a) Take it to C, but it's heavy.
		(b) Drop it off at B, but B is somewhat out of the
way.
		(c) Leave it in the car and come back for it, but this
involves an extra trip.

Idea: use bike, because it shortens trip.

Reaction:	(a) Nice, but don't want to carry case on bike.
		(b) Seems to require dropping case of at B.

Bright idea:	Lock case in locker.

Remark: Whle the reasoning is quite schematic, no formal model seems
to be formed in advance.  However, at each stage of elaboration,
there is a sensation of having completely understood the possibilities.

This leads to a notion of a trivial game against nature.  Small number
of moves for each player and short limit on length of game.  Perfect
information.  It is easy to figure out the possibilities.  However,
evaluation of endpoints might be problematical or left open - anyway
not given in advance for all end nodes.  Many nodes are immediately
choppable or even not generated by a primed move generator.

1989 March 8
Causes as objects.
	These may be the best example of useful entities with limited
meaning.  A cause is an object in a context dependent approximate
theory.  Thus there isn't even a general approximate theory in
which most causes exist.

Example: A cracked car window is replaced by a window made of ordinary
window glass.  A baseball smashes the window and injures someone.
What was the cause of the injury?  People give ready answers to
the question that they can't sustain when the context is widened.
Here it might be (a) from doctor's context: blunt object striking
face. (b) from police context: the baseball. (c) in car repair
context: using the wrong glass in replacing the window.  There are
combined contexts in which there is more than one cause and contexts
in which assigning a cause is not a meaningful mental operation, e.g.
in the universal wave function context.

The cause of a specific event is a case, possibly leading to a
general theory of the cause of this kind of event.  Perhaps it
can be regarded as an identifying instance of a conjectured
natural kind---in this case a natural law.

1989 March 14
Reasons, known and unknown
	Suppose Fred was shot and is dead.  The statement that shooting
kills is just default, so Fred might not have died from the shooting
and might have had a heart attack.  However, we want to presume that
it was the shooting that killed Fred, since we don't know about a
heart attack.

1989 March 19
fragments from a notebook, perhaps concerning Dennett's Intentional Stance
... Here is where the philosophers are missing some science.
122 - cond.(itions?) a,b,c
Under what conditions can we usefully build propositions satisfying
Dennett's a, b, and c into a robot?
136 - If you imagine that a text can have a radically different
interpretation.
---
Suppose the robot ascribes beliefs to Chinese and dogs in the first
order language we have given it.  Where does it become brittle?
145 - design
---
The fact that a signal is on a certain wire or neuron replaces some
of its syntax.
Consider a thermostat on a buss.  Its inputs and outputs require more
coding.  Suppose that it is to work with a variety of mechanisms
that might turn on the furnace and a variety of ways of determining
the temperature.  Then it needs more of a language.

March 24
	We need a program to construct models of theories, chiefly
unintended models that will help people check their axiomatizations.
Grading the cs323 exam suggests this.  If they could interact with
such a model builder, they would learn to axiomatize.  Maybe cs323
next year should use ekl, and maybe we should keep SAIL alive through
winter quarter.

	Maybe Barwise and Etchemendy can be interested in
collaborating on the model builder with giving as an example.

	Exams have too much randomness.  Let's try to make an exam
like an adventure game.  The examinee is trying to achieve a goal
and learns from his failures.  Mere slips lose only time.

march 24
wants(p,x,p') p wants x for p'

holds(wants(p,x,p'),s) ∧ holds(has(p,x),s) ⊃ next s = result(gives(p,p',x),s)

holds(wants(p,x,p'),s) ∧ ¬holds(has(p,x),s) ∧ ∃p''(friend(p'',p)
∧ ¬∀p'''¬wants(p'',p''',x)) ⊃ ∃!p''(friend(p'',p) ∧
¬∀p'''¬wants(p'',p''',x)) ∧ next s = result(ask(p',p'',x),s))

holds(wants(p'',p',x),result(ask(p',p'',x),s))

holds(has(p,x),s) ⊃ holds(willget(p,x),s)

holds(wants(p,p',x),s) ∧ next(s) = result(ask(p,p'',x),s)
∧ holds(willget(p'',x),next s) ⊃ holds(willget(p,x),s)

It looks like giving objects and exchanging them is a good basis
for cs323.  There is an excellent opportunity to emphasize
elaboration tolerance.

Mar 25
Another example of minor ingenuity.  If worst comes to worst, I'll leave
the door key to the car at home and the ignition key in the car.

Apr 26
The interesting case of approximate theory is when the weak theory
cannot predict the future of the variables of the weak theory as
well as can the stronger theory.  The flip-flop is a good example.
An example closer to common sense reasoning might be better, e.g.
the behavior of a clerk in a store.  Buying is often regarded as
a unitary act without regard for details of the clerk.

May 20
making a program into a variable or other constant into a variable
makes the context more general.  It then becomes a parameter of
various predicates.

May 21
1. In search of partially extensional expressions, I noticed that
``Mike's telephone number'' in ``Pat knows Mike's telephone number''
is extensional in Mike, even though it is not extensional in
``Mike's telephone number''.  Thus if Mike is Mary's husband and
has the same telephone number, it follows from the above that
``Pat knows Mary's husband's telephone number'' even though ``Pat
knows Mary's telephone number'' doesn't follow.  The formalism
in ``First order theories of individual concepts and propositions''
doesn't provide for this.

2. I was thinking about how to publish this short observation.  The
correct place is ``AI Letters'', a publication that doesn't exist.
Here's what it should be like.

	a. It appears monthly both on-line and on-paper.

	b. It has length restrictions but no shortness restrictions.

	c. The editor can accept items on his own and so can certain
other editors.  It is improper to submit an item to more than one
editor simultaneously, but it is ok to try to get another editor to
accept something that one has rejected, provided that the second
editor is informed of the rejection.  The editors can argue about
whether one of them is accepting too much junk.

	d. Submissions are electronic in TEX form.  There may be
some restrictions associated with a requirement that an item should be
TEXable when combined with other items.  Thus it should not reset
any TEX global variables and it should not force page boundaries.
The text should be readable in TEX source form, e.g. a reader should
not have to mentally execute the author's macros.  Some Latex form
might be appropriate.

	e. Electronic publication should be free and one form
might be as a Usenet newsgroup.  However, other electronic
publications should also be allowed.

	f. A C program to TEX articles in a way acceptable to X-Windows
should exist.

	g. Libraries will ordinarily receive the paper version.

	h. The publisher should be an organization like AAAI or a
publisher like Morgan Kaufman.  If the latter, they should be
familiar with TEX and have access to electronic mail both for
submissions and distribution.

May 23
controllable chaos
What is the simplest system for which controllability can be studied?
How about Feigenbaum's  x → ax(1-x)?
PDE systems will have the property that parts of it are separately
controllable.  In the weather one may be able to create storms that
will collide, and in celestial mechanics one can arrange multiple
collisions.  Are the strings evidence of control?
Chudnovsky, Gosper, Vardi?

June 11
*Causes
The cause of x was y.  Maybe there is an automata theory of causes.
We have interacting automata.  Some subautomata are known and
some are unknown.  The causes of events are the actions of
the unknown automata.
	Causes are said to involve law-like behavior.  However,
the instance ``x caused y'' may be known w/o the general law
being known or even conjectured.

June 29
Summarize ebos ideas and send to Ken Thompson

July 11
val
parsing the past
As you remember, I decided that Elephant requires regarding the
past as a sequence of events and that referring to the past
involves parsing this sequence, i.e. doing pattern matching on
it.  However, if we have continuous time, this isn't quite right,
because between any two events there are as many clock ticks as
you like.  This introduces the idea of pattern matching
continuous pasts.  Maybe the patterns should allow for
overlapping.  Have you heard of continuous pattern matching?
It should keep the mathematicians cheerfully occupied for
a while.

Jul 14
Peter Rathman came by to ask me to replace Papadimitriou on his
committee.
Peter Rathman mentioned combining databases, and we discussed
localizing inconsistencies.  VAL and I later discussed the matter
at lunch.  I proposed that nonmonotonicity replaces inconsistencies
with abnormalities.  The example PKR and I discussed was where
one db lists  retired(Smith)  and another says  current(Smith)  and
there is a general axiom  ∀x¬(retired(x) ∧ current(x)).  I
proposed replacing the first two sentences by
says(A,retired Smith)  and  says(B,current Smith).  VAL suggested
that this is unnecesarily complicated.  Using  ¬abA(Smith) ⊃ retired(Smith)
and ¬abB(Smith) ⊃ current(Smith) works even better.
He seems to be right, but it's a good problem.

We humans are good at localizing inconsistencies.  The problem is:
How much reification and nonmonotonic reasoning is required to localize
inconsistencies, e.g. by replacing them with disjunctions?

jul 17
discussed Elephant with Ray Perrault.  Question came up about whether
the distinction between illocutionary and perlocutionary is real.
I said that whether it is real or not in philosophy I want it
in Elephant.  The example was informing a person of a fact,
e.g. the cancellation of a flight.  We take  inform  as
a perlocutionary act meaning to make known.  Suppose the
program is asked to inform the passenger of a cancellation.
Ordinarily it will take the usual illocutionary act of  telling
him as having the effect of informing him.  However, suppose
the program has a reason to know that the telling it just
did was ineffective.  Then, if it is programmed to  inform
and not just  tell,  it will look for another way of
making sure the passenger knows the flight was cancelled.
Perhaps only a reservation system with some intelligence
can be asked to do it.

Side remark: maybe the most immediate form of intelligence and
planning we shall want occurs with goals that have normal
ways of achieving them but which ways are sometimes ineffective.
It must then plan another way of achieving the goal.
Because the routine works almost all the time, the AI
will be a useful frill.

Returning to the problem of whether the illocutionary-perlocutionary
distinction is real.  The way we philosophers shouuld think about
it is to ask in what context is it real.  What is the most
general context in which it is real?  This will be better than
just searching for cases in which it isn't.

July 26
	Consider contexts in which there are no false beliefs.
The agents either know the value (or truth value) of some
expression, or else know that they don't know it.  We are
often in that condition, an it should have a good theory.
Unfortuntely, once we admit nonmonotonic reasoning, it
seeems that we go outside that domain.

July 26, NOTES[1,JMC]/2P/3116L connotative circumscription, Jun Arima, ICOT

Jul 29, Natural kinds
Natural Kinds and Biological Taxa by John Dupre,  Philosophical Review,
XC, No. 1 (january 1981)
explains Putnam and Kripke's notion of natural kinds.  It turns out
to be different and not as useful as the notion I thought it was.
It involves and essence of the kind that might be determined
by science.  My idea is based on the fact that in the common
sense world, kinds are often discrete.  Therefore, to recognize
objects of a kind requires only some of the properties, because
objects that have these properties will also have the other
properties, not for any semantic reason, but just because that's
what our world is like.
My notion, unlike Putnam and Kripke's, covers the phenomenon
of a child's learning about a kind.  He needs very little
establish the kind in his mind.  He is then prepared to
learn more about it.  It isn't until he is ten or so
that definitions as characterizations even interest him.
Perhaps I should call them discrete kinds.
%
The nicest discrete kinds are those like biological species
that are sharply separated from others.  It is a great
convenience to the child sent to the store to buy lemons
that there isn't a continuum of fruit between lemons and
oranges.  However, consider hills.  There is a continuum
between, hills, mountains, bumps on hills and other
eminences to small or insufficiently separated from
other eminences to be called hills.  (Meissner considers
that Cho Oyu is insufficiently separated from Everest).
Nevertheless, the child does learn more about hills,
and so is justified in using ``hill'' as a natural
kind word.  Perhaps the child's point of view is that
``hill'' describes something about which he knows a
little and will learn more later.

1989 July 30
to make with Fischer-Teknik
1. Fischer Teknik manufacturing machine
2. Pendulum clock