COMMENT ⊗   VALID 00005 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	The New York Academy of Sciences
C00012 00003		Now I come to the real central component of this talk which is not
C00023 00004		Alright, the next thing I would like to mention, I really got these
C00035 00005		Now something which is new and represents progress in artificial
C00044 ENDMK
C⊗;
The New York Academy of Sciences
Computer Culture
April 6, 1982
Programs Need Common Sense
John McCarthy

	JOHN MC CARTHY:  At Stanford we believe that
there is at most one computer science.  Okay, I want to talk
about programs needing common sense and I hope to convince
you that first of all that the point is non-trivial by
starting with the fact that there are many useful programs 
embodying a substantial amount of artificial intelligence
that don't have common sense and the one I will begin with is this one
of the early expert systems MYCIN, which is, let
me have the first (SLIDE).  Oh yeah, I should start by remarking
that this is an old topic for me, about in the 1958 conference I
gave my first paper on this subject.  Second (SLIDE) please.

	And, well I'll say a little more about that.
The notion of common sense and the programs need it is one which
has had it's ups?  That was an up and it's downs and it's now so to speak
in an up again, which is to say the field of artificial intelligence has
more or less come around to this notion.

	Well MYCIN is an expert system for advising physicians 
on how to deal with bacterial infections of the blood, and it's
a production system which means that it's basic structure is a
collection of rules which are pattern leads to an action.  So it
detects a pattern of some kind and then if it can instantiate
the pattern, that is if it can say, ah yes, this is the infection
that the patient has which corresponds to one of the variables
and the pattern then it can look up some action to be performed.  So
it's a large collection of such things by an expert system is meant
that it was developed by conducting dialogues with experts in this, in
the field of bacterial diseases of the blood.

	Then I want to take this word ontology and
say, well what is the ontology?  Now ontology is a branch 
of philosophy which has to do with those things that exist and when it was 
developed as a branch of philosophy then they were interested in questions like, 
well, does God exist, do angels exist, or things of that kind.  Quine the
Harvard philosopher put ontology on a scale and he says, well it's
the collection of values that the variable have, is the ontology
and in that's the sense in which we are interested n it here and the ontology of
MYCIN included diseases, so it has names of diseases, it has tests that can be 
performed, symptoms, and possible treatments but it doesn't know about
doctors, hospitals, life or death, it really doesn't even know about the 
patient.  It knows about symptoms and diseases.  It will ask you for his
name and age, etc., it doesn't give any prognosis because it doesn't know
or that may not be necessary but it doesn't do any reasoning involving time
or events or causality.

	Now the question is, does it need common sense?  Well it is impressive
that it doesn't and if you ask it questions within its domain or give it
treatments in its domain then it gets a score on these things higher than
medical students and higher than interns, and higher than just about 
anybody except experts in this particular field.

	On the other hand the doctors that I've talked to say, well it isn't
really ready to be used, you wouldn't see, wouldn't sell a MYCIN cassette for
a doctor to put in the microcomputer that this other salesman sold him for
doing his bidding, doing his billing and keeping his patient records in
and the reason, I haven't talked to them enough to really understand 
where it is, but the reason that they give at the present time is that it 
can't take into account and there is o way to make it take into account 
enough side issues.  For example, if the patient has blood in his urine, 
then it won't suspect well, wait a minute has this guy been in a fight?
Or something like that because it has this very limited this very limited
ontology.

	Well, the question of whether it needs common sense or merely needs
a few improvements in its ontology is not clear.  Let me have the next
(SLIDE) please.

	Now in our official intelligence one might distinguish several
levels of thought and the bottom level of thought is the one which MYCIN
has which are basically a collection of rules whereby pattern action rules.
Another level of thought which is so commonly used in artificial intelligence is,
involves goals with pre-conditions leading to sub-goals.  So you want, the
program is to achieve a certain goal so it attempts to instantiate this goal.
That is to match it against the particular facts of the situation can then
give rise to sub-goals and these sub-goals may be either achievable directly
or they may spawn more sub-goals and so forth.

	And there are programming languages, microplanner and prologue, these
are aout 1970, is when they were first proposed, microplanner a little earlier
microplanner is now deceased, pretty much not used.  Prologue is a great
enthusiasm in Europe and is currently being developed but there, well based
on this notion of goals and sub-goals.

	Now a higher level involves reasoning from explicitly given facts 
where these facts may be neutral as to purpose.  Now this orderly hierarchy can
say, well the higher the better, is spoiled the following fact.  At the lower
level you can simulate the higher levels or you can simulate as much of the higher 
levels as you feel are relevant to your purpose.  And as a matter o fact there are 
still lower levels which are machine language and in this respect one would
believe that computers and computer programs are like people, namely, we do some
of our thinking on a high level of taking into account facts and we do other
thinking on a basis of almost or a reflex action or things which are only
slightly above a reflex action.  A pattern being matched and giving rise to
a reflex action.

	You can write an interpreter for goal seeking out of reflexes or out 
of some lower level so and in fact it's quite likely that our own human higher
level processes are compounded out of lower level processes which sometimes
act directly and sometimes act through compounding.

	So it's not so easy to, this orderly hierarchy is somewhat spoiled by 
this kind of thin.  Alright, what more do I want to say aout them.  Oh yeah,
(SLIDE) please.
	Now I come to the real central component of this talk which is not
in as good shape as I would like.  Namely what I ought to tell you now is
well, what are the common sense facts and what are the common sense 
reasoning ability?  This problem was posed to some extent in this 1958
paper and has not been resolved and when I'll go through these things 
that I've been able to put down but I'll say that it's not a very good
list and let me discuss that point about why in 25 years not much,
insufficient progress has been made.  I think progress has been made but and
I'll come to that as to what areas I think progress has been made and the
basic reason is that hardly anybody is thinking about it.

	Now if you talk to the philosophers they will say, well yes that's
certainly something that we should be concerned with or that is in the
suitable subject matter of philosophy, but let me tell you an anecdote,
now let's see, there were a couple of well known philosophers should I
mention them by name or not?  Well sure why not?

	John Burwies and John Perry are writing a book on situations and 
John Burwies gave a series of seminars at Stanford that I went to on his theory
of seeing, so he was interested in sentences of the form, John sees Tom run,
or John sees that Tom ran and the distinctions between them and so forth,
and part way through one of these lectures it quite suddenly occured to me to
ask him, is there any distinction in your theory between seeing and hearing?
And he rather proudly said, no, John saw that Tom ran is similar in structure
to John heard that Tom ran and he was pleased with that and I was disappointed
with it because to me the distinctions between seeing and hearing are very
important if you are going to build robots.  Well that's not his goal, now
this I would like the philosophers to pay more attention to these problems
of detail and I would like to persuade them to do so by telling a story that
my father was a carpenter and he used to get the carpenters journal,
Union Journal and I remember being struck by one article in the carpenters 
union journal and the article said, you guys may not be working with steel
forms, that is forms into which concrete is poured but just remember that all
forms are in the jurisdiction of the carpenters union and comes the
depression we will want every bit of jurisdiction we have so I should like
to remind the philosophers that if the don't pay attention to these problems
they'll lose the jurisdiction.

	Alright.  Now let me say what little I know about what the problems are.
The, a key thing is to be able to express that facts that are basically neutral
as to purpose.  Suppose that we take the fact that when two objects collide
they make a noise.  Now we can imbed these facts in a program, like a program
for avoiding making a noise, but as it happens for many facts for us humans
are represented in a way which is quite neutral as to purpose.  You can even
use it for explaining the lack of a noise and, okay so that's one aspect of
the matter.

	Now a key matter concerns situations that are changin in time and the 
effects of events, events include actions and the goal here, or the requirement
is that we be able to represent what people actually know about such matters.
And this turns out to be very difficult, let me explain a notion here, we have
ways of representing situations, facts about this world and we can talk about
on the one hand what I will call metaphysically adequate representation, or
physically adequate representations and on the other epistomologically
adequate representations.

	Let me give you an example of a representation that is physically 
adequate metaphysically adequate but no epistomologically adequate.  We can
represent a situation or La Plas (sp?) claimed that by giving the positions and
velocities of all of the molecules in the world.  Okay?  He said, if you
tell me the positions and velocities of all the particles I will predict
the future.

	Now leaving aside difficulties from coming from more modern physics
than LaPlas knew we also have the simple fact that there are 10 to the 23rd
molcules in a mole, now LaPlas didn't know that but he knew there were a
lot so he would have recognized this problem alright.  Now if we want to
predict the behavior of an actual object like a person we cannot get that
information about what the positions and velocities of all the molecules are
nor could we use it if we had it.

	On the other hand this scientific theory or th molecular theory is 
enormously useful, it can be used very directly to derive the loss of motion
of gases, the effects of heating gases, and so forth and it does interact with
our common sense knowledge but we have, we require a way of representing this
common sense knowledge.  Let me give you another example.
	
	Suppose that you see me waving this glass around and you think I'm
getting too excited and that I might drop it.  Well then suppose in fact
you see me drop it?  Then the people, suppose it was full, the people in 
the third row would not even move, they would not feel threatened by this
glass of water, people right next, the guy right next to me might jump.
He might feel it but it wouldn't be because they have set up an initial 
condition for the equations of hydro-dynamics of the water spilling and the
glass, the mechanical equations of the glass following is because they have
this common sense physical knowledge.

	Now we can represent, we know how to represent the initial conditions
for partial differential equations but we do not know how to represent in the
computer the common sense knowledge of the initial conditions in a way that 
would cause a robot sitting in the third row to sit still and a robot sitting in
the first row to try to get out of the way.

	Okay, another area had to do with goals and sub-goals.  It turns out 
that we've made a little progress n representing situations changing in time
in the following very simple case.  The simple case is where you have a
situation and then when you perform an action it gives rise to a new 
situation.  So you say, well gee that should  be fairly general, but what it
emphatically does not include and which no one in AI has been able to represent 
common sense knowledge about are concurrent events, while one thing is 
happening another thing is going on.  Now there is a theory of concurrent programs
but it turns out in its present form not to be usable for this purpose.
	Alright, the next thing I would like to mention, I really got these
out of order, are information about space and objects and substances  and
extended objects and this raises some very serious problems of partial 
knowledge.  Namely we have knowledge of other people, let us say man
sitting in the first row nearest me and I have some knowledge of him as a three
dimensional object which will cause me to have some expectations as to what
he won't do, I think he'll sit there for almost the rest of the lecture.
On the other hand much of him is occluded by the podium and his back is
occluded by his front so I can't see his back and I sort of round him off
and more or less expect him to be rounded off. But in any case what we're
talking about there is extremely partial knowledge and furthermore we're
talking about two different kinds that is appropriate to us and may be
appropriate to us and may be appropriate to a robot.  While I'm looking at
him I can answer certain questions and if I close my eyes or look away then I can 
answer some questions but by no means all the questions I can answer looking 
at him directly.  So people make this distinction between what what they can do
when they are directly perceiving something and what they can do from the
information they store about it.

	Now of course we may be abe to get around this computer program since
it's not very difficult to store a full picture but on the other hand it may be
still worthwhile even if you have a stored picture to consider that some things
are done with symbolic information which has been extracted like he's 
wearing a tie, and therfore they'll let him into this fancy restaurant if
that were the issue and so forth and so on which I can remember with my eyes
closed although I couldn't draw the tie, well I can't even draw it with my
eyes open but that's another matter.  Alright, so, well these are the major
topics, I've listed a couple of others but I see I more or less mentioned them
out of order.  Oh yeah, back to topic five, namely with other people and robots
and their beliefs and goals and intentions.  Let's go to the next (SLIDE).

	Let's see actually, yeah, well do I want to skip around in these
slides, yeah, let me go to the very las (SLIDE).  Is that the last slide?
Yeah, alright.  The last slide has to do with goals and intentions with a
topic which is, hasn't, isn't exactly the same topic or it doesn't come quite
under the heading and this has to do with the ascription of mental qualities to
material objects.


	Now we're going to quarrel about this in the subsequent panel but what
I want to mention is in this framework of the things that we need common sense 
knowledge about.  Now what I've written up there is an actual quotation from
the instructions from for an electric blanket or at least what I remember of
that quotation.  If you put the blanket control on the windowsill it will
think the room is cold and there was more about if you put it on the radiator
it will think the room is too warm and so forth and so on, and there was more
and the manufacturer put those quotes on think so he was being a little careul
but what he thought that by using this mental terminology he was making it more
likely that the purchasers of the electric blanket would be satisfied with it.
That is in other words would understand it well enough to be satisfied with it.

	Now what I believe is this, that we are going to find ourselves 
dealing in our daily lives with increasingly complex computer systems and that is 
and the example I will give there is a sort of banking system and in this we
will find ourselves using our, the abilities we have that lead us to ascribe
mental qualities to other human beings in order to ascribe mental qualities to
the materials of devices.

	The example ther is, second example is, suppose that you were dealing 
with a baking program and you have on this home terminal, dial this bank and asked
it for a loan and it gave you this loan and then later you discover that a suit
has been initiated against you for going into excessive debt and it says, it
thinks I promised to tell it if I incur any additional debt and so forth, 
that is if you mortgage your car twice is the feable(?) example I had in mind.

	Now I believe that this will, that this knowledge will under some
circumstances correctly describe what you know about this computer program.  Now 
what I'd like to do is I guess is to go back to that previous (SLIDE).  Okay, so
we have to deal with other people and their belief and knowledge and goals
and intentions.  An example is that one program may want to have in it a fact
that says that there is a program in the computer at the Lawrence Berkeley
Laboratory that knows the populations of American cities from the 1970 
census.  That is you may want to have another program know what other programs
know.  Or it may want to say that, know that travel agents know airline schedules,
or at least they know how to find out about airline schedules.  Well that's
another aspect of common sense that we shall have to formalize.

	Okay, now I've been talking about the complements that of common sense
knowledge now I'd like to talk aout common sense reasoning.

	Okay, the direct approach to artificial intelligence which I propose in 
this 1958 paper involved representing these facts in language of matthematical
logic.  That is involve representing or common sense knowledge of the world in
languages of mathematical logic and deciding what to do by logical deduction
from these facts of general common sense knowledge, facts about the goals that 
were to be achieved or say one fact, identifying a goal to be achieved and
observed facts of particular situations.  So that one might say is the brute
force approach to, the brute force logicians approach to artificial intelligence.
And it has turned out that this approach has got to be modified and the 
modification has to do with what is called, non-monotonic reasoning and that's
a technical term and let me describe it for this amount.  Logical reasoning
has the following properties, suppose you start with some collection of 
facts, A.  Let's call this collection of facts A, and you deduce some conclusion
P, and now suppose somebody give you additional fact not contradicting the
first one, then you can still deduce a conclusion P.  In logic this is true 
because a proof in logic consists of a sequence of statemtns each of which is
either one of your premises or an axiom or is a consequence of previous statements
in a proof according to an allowed rule of inference.  So if something is
provable from given premises then it will remain provable if you add more
premises.  The only thing that can happen if you add more premises that's  bad
is that you may make the whole system inconsistent so that you can prove one
equals, so that you can prove everything.  If you don't do that then this is ...

	Now ordinary reasoning does not have this monotonicity property, 
namely if you, suppose I tell you that I have a car then you may think well 
reason that it's appropriate to ask me for a ride.  However, if I tell you that
add the fact not contradicting the first that the car is in the shop you may
draw the reverse conclusion, and then if I add the further fact that the car
is due out in ten minutes you'll change your mind again, when I add the further
fact that I've promised rides already to five people you'll change your mind
but my car is a big one and holds ten people and so you can change your mind
as you add facts.

	Now something which is new and represents progress in artificial
intelligence is that since 1978, it has people have begun to think about 
systems for formalizing non-monotonic reasoning.  That is for making computers 
that will involve making computers do it and we've been thinking about that...

	Let me give one other example, and that has to do with ordinary language
non-monotonic reasoning, this is an example that Marvin has talked aout but
I'll put it in a slightly different terms than he has.

	The example is this, suppose I ask you to make me a bird cage and you
make this bird cage and I complain and I say, you wasted mondy constructing
this bird cage you put a top on it and my bird is a penguin and then you
would feel that I had wronged you, that I ought to have told you that the
bird couldn't fly and that you're entitled to assume that a bird that I talk 
about can fly.

	On the other hand if I told you to make me a bird cage and you made it 
without a top and my bird was a canary then I would feel wronged justifiably.
In other words it is a speech convention that if a bird can fly it does not
have to be mentioned whereas if a bird cannot fly then this does have to be 
mentioned.  Okay, well you might say that's merely because most birds can fly but 
that's not necessarily the convenient way to look at it.  One convenient way to
look at it is to take it as a convention and that turns out to be one use for
non-monotonic reasoning because if we have something that will draw a conclusion 
that the bird can fly, that is an AI system if a bird is mentioned, but will
not draw this conclusion and will not be put out so to speak or not arrive at
a contradiction if we add the additional fact that the bird in question is a
penguin then it has to use these forms of non-monotonic reasoning.

	Okay, next (SLIDE).  Okay, this area of thinking computers have common
sense has a lot of unsolved problems in it and the pont is that some of them
have been identified explicitly, now I've lost track of time I have to confess,
oh yes, I know what time I was supposed to have started and if I started when
I was supposed to then I'm three minutes overdue.  Three minutes perhaps to
sum up.  Maybe it says eight minutes.

	Okay, I'm going to just mention one idea that will lead into the panel a
little bit.  This is the notion of ambiguity tolerance, namely, the phrase was
used in a book by Burt Dreyfus (wrote) a long time ago called, What Computers
Can't Do and he said, computers don't have it but it isn't quite the same
thing but anyway I'll tell you what I know about it by giving an example.

	Suppose a law is passedd that it's illegal, it's a crime to attempt to
bribe a public official and the District Attorney is a very modern District
Attorney and he gets this programmer, a tech knowledge (?) for example, to 
build him an expert system and the expert system is one which advises him on
whether to seek an indictment in a given case and also what crime to charge.
And among things you want to put in not facts about this new crime of attempting
to bribe a public official so you put in these facts and for 10 or 20 years some 
people are indicted, some people are acquitted and some people are convicted and 
so forth and finally some smart lawyer offers the following defense for his client.

	He said, you've indeed proved that my client offered the man $5,000 to
see if he could get out of his drunk driving conviction but what you haven't
proved is that my client knew he was the Commissioner of Motor Vehicles, my
client may have thought that he was just a lawyer who would invent a new 
defense.  So in attempting to bribe a public official it is essential that you
know the individual is a public official in order to be considered to have
made the attempt.

	Another lawyer offers the following defense, philosophers oould call that 
the de dicto defense, offers the deray defense.  It's true my client offered
him $5,000 to fix his drunk driving conviction but in fact the Governor had
never properly signed his commission as Commissioner of Motor Vehicles, so that
this man who my client attempted to bribe was not actually a public official
even though my client thought so.

	The third lawyer offers a defense, he says it's true that my client
put an advertisement in the Criminal Gazette that he would pay $5,000 to any
public official who would fix his drunk driving conviction, so does 
attempting to bribe, but you haven't proved that there was any public official
who was capable of fixing his drunk driving conviction, so does attempting to
bribe a public official require that there be a public official whom he was
attempting to bribe?

	Okay, now these puzzles are known, this kind of puzzle is known 
both to philosophers and lawyers, the interesting thing from the AI point of
view is not the resolution of the puzzles but the ten or twenty years that
went by before the problem was recognized, because if artificial intelligence
has to solve all of the problems of philosophy before it can get started, in
particular all of the problems of ambiguity of language before it was going
to get started we may have to wait yet another 2,000 years before we're even
ready to begin on artificial intelligence and therefore, we need this notion
of ambiguity tolerance, we need to be able to use concepts that are
ambiguous.

Thank you.