COMMENT ⊗   VALID 00019 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00003 00002	notes[s83,jmc]		miscellaneous
C00004 00003	Elaboration Tolerance
C00007 00004	April 19, 1983.  For discussion with Kahn and Ohlander.
C00009 00005	(GOAL conclusion(R))→(GOAL condition(R))
C00011 00006	April 19, 1983
C00012 00007	April 19, 1983  The Army Conference on Application of Artificial Intelligence
C00014 00008	April 20, 1983  Morality of deterrence
C00015 00009	Language translation
C00017 00010	I have had a second thought about our discussion of getting AI capability into
C00019 00011	Modifying programs without reading them.
C00023 00012	Discussion with Suppes
C00024 00013	believing the consequences of one's beliefs
C00025 00014	1983 june 16
C00027 00015	dictated notes
C00035 00016	1983 June 20
C00047 00017	Corky
C00048 00018
C00049 00019	1983 june 25
C00051 ENDMK
C⊗;
notes[s83,jmc]		miscellaneous

Tasks:
blocks		Do example with circumscription
psycho		Really good example
		reread "Ascribing ... "
common		re-examine list of common sense abilities
		reformulate questions on mycin
Elaboration Tolerance
	Formalizations of general common sense facts should be, if possible, 
elaboration tolerant.  It should be possible to elaborate the description
they give to take into account more phenomena by adding sentences.  It should
not be necessary to take back existing sentences or even worse to change the
meanings of entities.

	This is rather unclear and I don't yet know how to state it more
clearly in general, but some examples can be given in which the idea is much
clearer.

	Consider a blocks world.  In the formulations common in AI, moving
a block from one location to another is a discrete act.  The preconditions
for and consequences of this action are described by suitable logical
sentences.  However, we sometimes also need to consider moving an object
as a continuous action capable, for example, of interruption.  In our
common sense thinking and in conversation we don't regard these two
notions of motion as contradictory and we can freely mix sentences
involving them.  Thus "He moved it from A to B and then while moving it
from B to C, he dropped it".  Other kinds of elaborations are possible,
e.g. "While he was moving the sugar from A to B, some of it spilled."

	Our formalizations should also tolerate such elaborations, and it
seems to me that neither the situation calculus  result(e,s)  formalism nor
STRIPS tolerates such elaborations, because neither has any way of going
from the discrete result of an event to continuous process.  However, it
seems that the situation calculus treatment can be modified to a more
elaboration tolerant form.
April 19, 1983.  For discussion with Kahn and Ohlander.

1.  ARPA or DEC
2.  CBCL, natural language
3.  Generally --the "new" ideas aren't that good.
4.  Changing programs without reading them.
5.  Other people's files,/National library compatible documents.
6.  Dialnet
7.  Common Sense
8.  Non-monotonic reasoning
9.  Standards
-----

Opel (if it was he) should be pilloried for advocating that IBM take a parasitic
attitude to artificial intelligence.  I - or Nils - should mention it in a AAAI
 speech advocating that IBM should pay extra (say about $150 million) when 
AI does pay off.
Palo Scientific Center
Harry Reinstein gave briefing.  boss↑2 = Horace Flatt
(GOAL conclusion(R))→(GOAL condition(R))

How do we get the backtracking after the


April 19, 1983  Principles of Rule-based expert systems by Buchanan and
Richard Duda

It would seem that data driven and goal driven systems are equivalent.  Each can
 model the other by reifying goals or conclusions.

Thus we model a goal driven system in a data driven system by allowing a 
sentence (goal G) and a data rule

Similarly we model a data driven system in a goal system by having a goal
(ASSERT X) and
April 19, 1983

Nevada (or much of it) has interior drainage.  Suppose the climate changed
so that Nevada got lots of rain.  How much water and how deep would it get
before it found a way out to the ocean?  Where?  Colorado river?  How much
rainfall would be needed?  -------- We might call this doctrine
"Dennettism" since "McCarthyism" has been used up.  ---------

Approximation by autonomous systems

He skied normally (for him) except that he bent his knees.  Thus we can use
circumscription to get consequences of counterfactuals.

April 19, 1983  The Army Conference on Application of Artificial Intelligence
to Battlefield Information Management

0.  DARPA
1.  Technological surprise?
2.  Intellectual mech.
	Human level
	Expert Systems
	Present limitations
	Scientific preparation for
	  a.  research
	  b.  practitioner

Vision --reduce end spoofing

CBCL -- Battlefield CL






1.  DARPA support -- hope it's now becoming useful
		     our responsibility to help.

2.  history; 1950 Turing, computers, distractions
				     neural nets
			             discrimination

3.  Intellectual mechanisms		robotics
    Intellectual problems		$10M  84
    Theory and experiment		$22M  85
     a.  basic experiment		Patience
     b.  applications			Natural selection
					Think of it as evolution in action

4.  Human level AI -- when
					drosophila
         SURPRISE

5.  Expert systems, Mycin, production rule based systems.

					ADA vs LISP?

6   Need for common sense

	CBCL -- military communication lang.
	semantics of natural languages

7.  education 
	for research
	for practitioners
April 20, 1983  Morality of deterrence
		--necessity of objectivity
		--religious style of argument
Language translation

Automatic language translation was effectively dropped in the middle of 
1960s as too ambitious for the linguistics, artificial intelligence, and
computer capabilities of the time.  With improvements in these sciences
and technologies it may be worthwhile to try again with more modest
goals.  The languages of primary interest are Russian as before but also
Japanese.  A worthwhile goal would be a reader's workstation that could
accept a microfiche text, do the necessary OCR, referring doubtful
characters to the user and do a rough translation.  If the translation of
the sentence was doubtful the user could ask for dictionary entries for particular
words.  He could also ask for the reasoning behind a proposed translation
and ask for alternate translation.  The workstation would be useful even if
it required that the user have taken a one semester course in the grammar of
the language and the use of the workstation, because many times as many people
will do this as will study for the several years required to acquire a fluent
reading knowledge of either Japanese or Russian.
I have had a second thought about our discussion of getting AI capability into
industry.  This is to ask what kind of capability we want industry to have.
Is it to make more SRI's or BBN's with basic research capability and the ability
to make high quality unsolicited research proposals?  If so this can be done in the 
same way that SRI and BBN improved their capability -- a steady diet of contracts to
a capable group.  The alternate goal is to transfer the technology of expert systems
so that operational systems can be contracted at reasonable cost.  Judging from
the Army Conference and other indications, a fair number of companies claim this
capability, perhaps more than the present market or the people supply will support.
Natural selection will prune and improve this group.
Modifying programs without reading them.

	To an increasing extent the job of a programmer involves modifying
existing programs rather than writing them from scratch.  Even when a
program is written from scratch, this is because a program that performs
most of the desired functions is too incomprehensible to modify.

	If we make an analogy between modifying a computer program and
educating a human, then present day programming corresponds to education
by brain surgery.

	There are programming languages and styles of using them that come
closer to our English language ability to modify the behavior of another
person without fully understanding the present behavior.  Here are some
examples of modifications possible in English which we would like to be
able to do with computer languages.  We use an airline reservation system
as vehicle for our example.

April 21, 1983

When can we modify without looking.

1.  When the program can optimize or satisfice subject to constraints, we
can always add a new constraint.  The task remains definite even though
there is a risk of error if the constraint set becomes incompatible.

	This works provided the desired constraints can be expressed in
language not referring to the internal names of the program.

	2.  In general, modification without understanding requires the
program and its modifications refer only to the external world.  We can
minimize the understanding required if the amount of innards is minimized.
Thus we minimize jargon, and this might be a technical notion of jargon.
We can use abbreviations defined in the program so long as the full
expressions are available for making modifications.

	3.  Circumscriptions will be needed in compiling the program.
Perhaps when such a circumscription reaches a disjunction, the user must
be questioned as to which alternative he prefers.  This won't always be
true.

	4.  We require creativity when new entities must be introduced to
solve the problem.  Such aspects of programs cannot be modified without
reading unless the compiler can understand the program and supply the
creativity.

Discussion with Suppes
The epistemology of quantum mechanics should be based on amplitudes
rather than on probabilities.

Dirac
von Neumann
The Lifshitz book on quantum electrodynamics points out that the
notion of wave function isn't relativistically invariant.  Are
the conceptual foundations of relativistic quantum mechanics
different from those of non-relativistic QM?
believing the consequences of one's beliefs
Circumscriptions are required here.

1. It believes any consequence of its beliefs unless ...

2. If I can do the inference, he can unless ...
1983 june 16
mail file notes:

1. Get dfh to make a file of the comments I have received about Altos.

2. p. msg.msg[1,jmc]/268p DAM to be disagreed with
msg.msg[1,jmc]/271p from kdf
p. 275 from Leslie Pack

3. 272 from Stallman about car rental

4. Fran for secretarial

5. 298 TOB complaint about purchasing

6. 308 Nilsson about AAAI.

7. send Feigenbaum parallel lisp stuff and cover page of Taiwan
edition of AI Handbook.  Also I can join his project for my work
on parallel lisp.

Correctness of a program for computing  pi  and   e.  pi  and  e 
are real numbers - not floating point numbers.  State and prove
correctness of algorithm for computing  n  digits.
dictated notes
1. Servan-Schreiber has a brown nose.

2. Notes for paper on ontology and reification
We begin with mathematical examples pointing out that the phrase
"elementary theory of groups" has a double meaning.
It means "simple", but it also means that the individuals are
only elements of the group and not subgroups, cosets, etc.  The
elementary theory of blocks only has blocks as individuals, not
sets of blocks or towers.  There is no obvious AI reason to
prefer elementary theories, although we could imagine that
it might allow some decision procedures.

The reification of names to get the unique names hypothesis as
a consequence of circ. provides one good example.

Individual concepts and propositions with a reference to my
paper.

Reifying possible worlds for the benefit of S and P and wise men.

3. the paper on circ. needs a self-critical remark about prev-can-fly
and other purely technical predicates being a blemish that can
perhaps be removed by reformulating the theory in some way.

4. Mr.Hug should be inspected to see if it gives us more common
sense problems that can be spelled out.

Actions by groups of robbers and whether the robber who emptied the
cash register was the same one who pushed Hug down the elevator  shaft.
The story shold also be checked for possible circumscriptions and
non-monotonicities.  Presumably it would be a blemish in a news story
if the early paragraphs were misleading as to the story as a whole.

5. The defect with Yoram's treatment of the Berliner position is
that it proves a win by getting a function that decreases rather
than some analog of the situation calculus that tells what happens.

We need to consider the trivial case of concurrent events where
there is no interaction and the time relations are clear, e.g.
I will surely be done with my part before you are done with yours.

June 18 note: Consider a generalized situation that has components
corresponding to states of parts of the system.  When we say that
s  occurs, we do not claim that each part is in the prescribed
state at a common time  t.  Instead we merely assert that each
part takes the right value at different times that are not causally
related, i.e. on a space-like surface.

6. McDermott and Doyle err in trying to make arbitrary sets of axioms
interpretable.  An interpretation that only works in favorable
cases that occur in practice will make a more realistic theory.
There is no human intuitiion about interpretation of arbitrary
sets of axioms and this is true also of circ.

7. Ask Diana to send Dialnet papers to Richard C. T. Lee at Tsinghua
University in Taiwan.

8. The papers on planning volume 3 of Handbook of AI suggest a
more systematic approach.  There should be some kind of propositional
planning problem.  A system in which the preconditions are boolean
functions and the actions are assignements can readily be formalized
in first order logic using states.  It's like STRIPS.  The propositional
planning problem may reduce to a known decision problem or even
satisfiability.  We may get something in propositional tense logic
or propositional dynamic logic.

Other predicate calculus axiomatized systems may admit propositional
approximations based on a finite number of instantiations.  An interesting
class is monotonic in that if the propositional problem is solvable,
then the more general problem is solvable.  We also want that if
the problem is solvable in general then some propositional
approximation is solvable.  Where non-monotonic reasoning is
involved, this may not be true.  The ideal case is when the relevant
set of instantiations can be determined in advance.

9. Propositional STRIPS will provide a good example of reification
if we axiomatize it using situations.  Moreover, it will be decidable,
since if there are only a finite number of propositional variables,
there are only a finite number of states.  Can propositional STRIPS
be jazzed up to give a propositional ABSTRIPS?

To some extent propositional STRIPS is just a certain way of
presenting a finite automaton.  This may not be quite true since
the number of certain kinds of statements can increase, i.e. the
state may not simply be a Boolean vector.  We can call it finite
planning rather than boolean planning, and there may be some
reasonably effective algorithms.  Don't forget that there
there may be constraints coming form such facts as that a block
can be -on- at most one other.

9.complete aside.  It would be nice if Levchenko (if that's the
name of the Russian who defected in Japan) had had some evidence.
Publishing what would be wanted would be nice and would make the
KGB nervous.

10. Was it really Doyle who did the non-monotonic logic in the
Handbook starting on page 114?  On p. 115 there are some examples
of default reasoning that should be looked at as examples of circ.


1983 June 20
Invention while running and being passed by a bicycle.  Consider a pedal
powered walking machine, perhaps with two legs, perhaps with four.  Fancy
materials are used so that the legs are very light.  The human control
is supplemented by a computer so that the center of gravity is
kept at nearly constant height.  It should be able to go over
rough terrain and up stairs but with the efficiency of a bicycle.
This assumes that the main reasoning running is hard work is the
effort required to start and stop the heavy legs.  Perhaps energy
generated by pedalling should be stored so that greater instantaneous
effort can be used to overcome obstacles.  A task for Macready.

June 21
Here are some design principles for a new publication system.  There
are two motives for an author to publish - 
professional: to get the ideas out and/or to be recognized.
financial: to get royalties.
A publication system needs to allow for both motivations and to treat
properly cases in which either motive is dominant.  The main motive
for scientific publication is professional.  Even if an individual
is selfish, the money he makes from publication is either nothing or
small compared to the increment to his income from the professional
recognition he hopes to obtain.  If this were the sole motivation
the following kind of publication system would be appropriate.
There is a single grand publication system consisting of some terabytes
of disk at the Library of Congress.  Anyone can publish anything by
paying the cost of keeping it on the disk, and these costs are already
trivial compared to the cost of the time required to write something.
Anyone can read anything that has been published and must pay only
the costs of transmission.  These are presently non-trivial but
affordable, although I don't have precise estimates.  Formerly it
would have been worthwhile to have regional disk libraries for frequently
used items in order to reduce long distance transmission costs, but
this is probably no longer true for a new system.

	Of course, many
kinds of backup files will be required for safety.  We consider technical
safety (preventing machine malfunction, deteriorationof the
storage medium, natural disaster, war, or
other catastrophe from destroying the files) and also political catastrophe
(someone in control of the Government wishing to edit the files).  The
protection against the latter is that backup copies can be kept by
any organization willing to bear the cost.  Fantasy: shoot thousands
of backup copies into space on orbits that will cause them to return
and safely re-enter the atmosphere at random times in the future.
In any case it is possible to provide a much higher degree of backup
than exists with paper systems.

	A large fraction of publication is professionally motivated,
but it would be wrong to depend on professional motivation for all
publication.  The reason is that receiving professional rewards for
publication depends on that publication meeting the approval of the
institution that pays the salary of the writer.  In most cases this
approval is partially or mainly mediated by public opinion among
other professionals in the area.  Diversity and liberty require that
it also be possible to appeal directly to the public and be
rewarded by what the public pays to read.  This can easily be
administered through the computer system provided members of the
public don't make bootleg copies.  If the authors' fees are a
small part of the cost of using the system, there will be little
motivation to cheat, but some kinds of publication may require
large authors' fees.  Another system that can work is that the
library pays authors according to the amount that their work is
read and that readers are not directly charged.  Both systems should
be used.  The latter seems to require a fixed charge per unit
of material or some rate determined by the library administration.


June 21
Pat Winston sent me M.I.T. AIM 678 (September 1982) entitled
"Learning by augmenting rules and accumulating censors".  He remarked
in the letter "I did not realize my censor thing was another
perspective on nonmonotonic logic until your talk".

The following remarks precede detailed reading of Winston's paper.
His censors seem to correspond to what I have called inhibition.
Statements contain free variables that are implicitly universally
quantified and are labelled with labels that contain the variables.
Another statement can have the conclusion that a certain label
is inhibited for certain arguments.  This is non-monotonic, because
when we add a new statement that inhibites another we lose conclusions
that might have been made.  Allowing variables in the labels is
crucial.  Without them we cannot inhibit an ostrich's ability to
fly without killing that ability for all birds.
Cross-inhibitions are possible, and much of the problems of the
McDermott-Doyle non-monotonic logic might be interpreted as arising
from the desire to have interpretations for all possible patterns
of cross-inhibition.  The examples that motivate inhibition will be
satisfied by a system that forbids cross inhibition, i.e. we demand
that there be no cycles of inhibition.  Therefore, cross inhibition
should be allowed only if someone can show good reasons for it.

The relation between systems with inhibition and circumscription isn't
immediately obvious.  It may be possible to translate one to the other.
The inhibition logics (to coin a term) may be more natural than
circumscription, because we don't have to invent predicates like
prev-can-fly  whose significance is purely technical.  We need to
explore whether the absence of a bridge can be handled by inhibition.
Remark: if we omit the labels we can still imagine that each sentence
is implicitly labelled by its quotation.  This would require inhibition
statements to explicitly quote the statement they are inhibiting.

It would seem that any system of inhibition statements can be replaced
by an ordinary collection of sentences which will have the same models.
However, doing so loses the non-monotonic character of the system, since
adding a sentence in the inhibitory system corresponds to modifying
a sentence in the translation.

It seems clear now that inhibition is weaker than circumscription.
In particular inhibition can't do induction and perhaps it can't
even handle disjunction.  The latter might be handled by allowing
some cross inhibition.

We might want to combine circumscription and inhibition in some way.

Actually more is required of the inhibit rules.  We prefer minimal
models of the inhibit predicate, so that unless an inhibition can
be proved, we assume the rule is uninhibited.  This seems to put
inhibition logic directly under circumscription.  The translation
works as follows.  Suppose a sentence  P{x,y}  is labeled  foo(x,y).  We
remove the label and write  ¬inhibit(foo(x,y)) ⊃ P{x,y}  instead
of the previous  P{x,y}.  The parts of sentences with conclusion
inhibit(foo(x,y))  are left as before.  Then we circumscribe the
wff  inhibit(x)  keeping all the other sentences as the constraint.
At least in the absence of cross inhibition, the circumscription
gives exactly the interpretation we want.  We can hope to derive
a theorem that when there is no cross-inhibition, there is a unique
minimal model.  Presumably this corresponds to some kind of
generalized induction as discussed by the recursion theorists.
Corky

Higher order objects via lazy evaluation.
covers Edinburgh LCF (has general notion of formula uses Igarashi)

minimization schema =? complete recursive program

first order axioms for lazy data domains.  Derived rule of induction
that permits things proved for finite data domains to be extended
to infinite domains provided formula passes syntactic test.
1983 june 25

1. Represent generalized situations which are elements of the
cartesian product of the local situation spaces.  The generalization
is that there is no requirement of distant simultaneity, i.e. a
we can predict a generalized situation without having to show
that the different local components take the prescribed values
at the same time.  There probably should be a requirement that the
different local situations be causally independent, i.e. nothing
in one of them influences the others.

2. Consider the missionaries and cannibals problem with the addition
of a small mischievous monkey.  It is impossible to fully consider
all the varied things the monkey might do that might interfere with
the solution of the problem.  Human reasoning results in some kinds
of incomplete prediction of the consequences of the actions.  It is
presumed that interference by the monkey will be countered and the
damage to the plan repaired without any listing of everything the
monkey might try.  This is just an example of the incompleteness
that must be tolerated in intelligent reasoning.

incompleteness tolerance
elaboration tolerance
ambiguity tolerance