COMMENT ⊗   VALID 00005 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	\input memo.tex[let,jmc]
C00059 00003	\noindent References:
C00067 00004	\smallskip\centerline{Copyright \copyright\ \number\year\ by John McCarthy}
C00068 00005		In this lecture, I shall recount some history of artificial
C00069 ENDMK
C⊗;
\input memo.tex[let,jmc]
\title{The Logic and Philosophy of Artificial Intelligence}

	The goal of artificial intelligence (AI) is machines
more capable than humans at solving problems and achieving goals
requiring intelligence.  There has been some useful success, but
the ultimate goal still requires major conceptual advances and
is probably far off.

	There are three ways of attacking the goal.  The first
is to imitate the human nervous system.  The second is to study
the psychology of human intelligence.  The third is to understand
the common sense world in which people achieve their goals and
develop intelligent computer programs.
This last is the computer science approach.  It
is the one that has had the most success so far, and it is the
one I will discuss in this lecture.

	Among AI researchers, opinions differ about how much to
emphasize mathematical logical languages for expressing what the
machine knows about the world.  Briefly, the advantage is that
facts can be learned independently of one another and of the
use to which they will be put.  The argument against it is that
much human reasoning doesn't seem very logical.  The current
expert system technology uses logical languages for expressing
much of their knowledge, but a lot is also built into computer
programs and also into the arrangement of the logical sentences.
I will emphasize the logical approach in this lecture.  As perhaps
befits the occasion, I will emphasize my own work.

	I became interested in artificial intelligence in
September 1948 when I attended some sessions of the Hixon
Symposium on Cerebral Mechanisms in Behavior (Jeffress 1951)
held at California Institute of Technology where I was starting
graduate work in mathematics.  The symposium
included the mathematician John von Neumann (one of the inventors of
the stored program computer), Warren McCulloch (who had shown
how networks of hypothetical neurons could be used for computing)
and many famous psychologists and neurophysiologists.  Von Neumann's
paper was entitled ``The General and Logical Theory of Automata'',
and there was much interest in comparing what was known about the
human brain with the new idea of an electronic computer.  At that
time, stored program computers were under construction, but the
first wasn't finished until the following year.

	When I recently re-examined the proceedings of the symposium,
I discovered that my memory was incorrect and
nothing had been said about trying
to make intelligent computers.  All the famous participants were
fully committed to their existing research programs in biology,
psychology and mathematics.  Von Neumann discussed two ideas that he later
developed more fully --- how to make reliable machines out of
unreliable components and how to make self-reproducing machines.
Although he had done important work in mathematical logic, he
didn't discuss, then or later, representing facts about the world
in logic.  Although he was actively involved in developing computers,
he didn't discuss programming them to behave intelligently.
Nevertheless, the conference oriented my own scientific
ambitions toward artificial intelligence.

	While I started thinking about AI in 1948, I didn't think
about programming computers or about representing facts in logic.  I
didn't know much about either logic or computers.  Instead I
successively pursued two ideas that I later came to regard as blind
alleys.  The first combined automata theory and information theory,
both newly fashionable.  It considered a brain as a finite automaton
connected to an environment also considered as an automaton.  To
represent the fact that the brain is uncertain about what the
environment is like, I considered an ensemble (i.e. a set with
probabilities) of environment automata.  Information theory applied to
this ensemble permitted defining an entropy at time 0 when the brain
was first attached to the environment and later when the system had
run for a while and the state of the brain was partially dependent on
which environment from the ensemble had been chosen.  The difference
of these entropies measured how much the brain had learned about the
environment.

	I came to believe that this was a bad idea and never
tried to publish it.  The trouble was that I couldn't see how
to represent specific facts about our own world in terms of an
ensemble of environments.  To anticipate later terminology, the
representation was epistemologically inadequate.  It couldn't
represent the information actually available.

	My next bad idea was more mathematical.  What is a problem,
i.e. one that we might want an intelligent machine to solve?  I
proposed that a ``well-defined problem'' is given by a method for
testing proposed solutions.  That's a good idea and has survived in
one form or another, but maybe it's obvious.  The method of testing
proposed solutions has to be well defined, so let it be given by a
Turing machine --- the abstract form of computer introduced by Alan
Turing in 1936.  Therefore, problem solving may be regarded as the
problem of inverting functions defined by Turing machines.  I wrote a
paper (McCarthy 1956) on this while working on a summer job with
Claude Shannon in 1952, and it was included in our Automata Studies
volume published as (Shannon and McCarthy 1956).  The trouble is that
this is also epistemologically inadequate.

	That stored program computers were the right vehicle for
developing artificial intelligence was still not apparent to me or to
any of the others who contributed to {\it Automata Studies}.

	The person to whom it was first apparent was Alan Turing.
His (1950) paper ``Computing Machinery and Intelligence'' was
in a philosophy journal and did't become well known
until it was reprinted by Edward R. Newman in his {\it World of
Mathematics} in 1956.  I had to come to the idea independently
and so did Allen Newell and Herbert Simon.  Newell was probably
the first person to make programming computers to behave
intelligently a career.  Turing died in 1954, before computers
were good enough to do much in the way of AI.
 
	My introduction to computers came in 1955,
when Nathaniel Rochester of IBM hired me for the summer
in his pioneering Information Research Department in
Poughkeepsie.  I learned to program the IBM 702, and
re-oriented my thinking about AI to making intelligent
computer programs.  I was an assistant professor of
mathematics at Dartmouth College then, and it occurred
to me that summer that there was enough interest in AI
to warrant a summer research project at Dartmouth
in 1956 that would invite everyone we knew who might
be interested.  That wasn't too many, and we were
able to invite a number of people that we only hoped
would be interested.  Claude Shannon, Marvin Minsky,
Oliver Selfridge and Nathaniel Rochester joined me
in making a proposal to the Rockefeller Foundation,
and they awarded us \$7500 to pay for some travel
and living expenses.

	We had to decide what to call the subject, and I chose
``artificial intelligence''.  The reason for choosing such a
bold title was the desired effect on participants.  When Shannon
and I collected papers for {\it Automata Studies}, it seemed to me
that we got too many papers treating the mathematical properties
of automata that were only peripherally relevant to making
intelligent automata.  The name ``artificial intelligence'' nailed
the flag to the mast.  Many people have criticized the name, but
I think that it was and still is important to have a name that
makes people compare their present projects with the ambitious
long term goal.

	To me the most interesting work presented at Dartmouth
was the Newell-Simon ``Logic Theory Machine''.
Their program simulated the processes that naive student
subjects went through in solving problems in elementary logic.
In the course of this, they also developed the IPL list processing
language and introduced recursive subroutines.
Other interesting projects were Arthur Samuel's program for the
game of checkers and Alex Bernstein's projected program for chess,
both IBM projects.  Marvin Minsky proposed a theorem prover for
plane geometry that used a ``diagram'' in the memory of the
computer to decide which sentences were plausible enough to try
to prove.  I had started thinking about the use of logic to
represent facts, but I didn't have much to present at the
Dartmouth meeting.

	After the meeting, Rochester decided to implement Minsky's
ideas of the geometry theorem prover and hired Herbert Gelernter to
do it, with me as a consultant.  I had been impressed by the Newell-Simon
idea of a list processing language, but their particular formalism
didn't appeal to me at all; IBM's projected Fortran had a much
more attractive style.  Therefore, I proposed to Gelernter that
he do his list processing for geometry in Fortran and proposed
some basic functions --- the $car$ and $cdr$ of LISP.
Gelernter and Carl Gerberich did it, and added something more ---
making $cons$ a function so it could be composed.  (I don't have
the space to describe the intellectual situation well enough to
convince you that this wasn't the one obvious thing to do.)

	I didn't become a full-time computer scientist till I
went to M.I.T. in Fall 1957.  By then I was convinced that logic
was the key formalism for artificial intelligence.  While Newell
and Simon (and several others soon after) had already written programs
to do logic, none of them were interested in using logic except
as a subject domain.  My conversion to logic was expressed in
my 1958 paper ``Programs with Common Sense'' (McCarthy 1959).

	In 1958 I spent another summer with Rochester's IBM group,
where Gelernter and Gerberich were finishing their plane geometry
program.  I became convinced that Fortran didn't permit the
expression of recursive list processing functions in a nice way
and started work on formulating LISP.  When I returned to M.I.T.
in Fall 1958, Minsky and I were given a big work room, a key punch, a
secretary, two programmers and six mathematics graduate students.
This was a very decisive and prompt action by Jerry Wiesner,
considering that we had asked for these things the preceding Spring in
order to form an Artificial Intelligence Project and hadn't even
prepared a written proposal.  Fortunately, the Research Laboratory of
Electronics at M.I.T. had just been given an open-ended Joint Services
Contract by the U.S. Armed Forces and hadn't yet committed all the
resources.  I think that such flexibility is one of the reasons the
U.S. started in AI ahead of other countries.  The Newell-Simon work
was also possible because of the flexible support the U.S. Air Force
provided the Rand Corporation.

	In September 1958 we started on LISP, and the first interpreter
was working in early 1959.  The first idea was just a Fortran-like
language with list structures as data.  However, a look at the logical
structure of the process of differentiating algebraic expressions
as it is normally described to college freshmen showed that that
function could be described by a simple formula provided it could be
used recursively, provided conditional expressions were allowed, and
provided explicit erasure could be avoided by using garbage collection.
I had invented conditional expressions the previous year in connection
with a chess program written in Fortran.  Lambda expressions were another
tool whose use was suggested by the differentiation example.
In order to show mathematicians that a universal computer could be described
much more compactly in LISP than with Turing machines, I wrote the function
EVAL.  Steve Russell, one of our programmers, promptly turned it into
an interpreter for LISP.
(McCarthy 1977a) describes this history.

	Early in 1979 James Slagle started his work on
symbolic integration under Minsky's supervision.  About this time,
the promising IBM AI work was shut down, and IBM didn't really resume
AI work until 1983 and still plays a minor role.  The reason for
shutting it down was apparently a combination of uninformed scientific
criticism and public relations.  IBM wanted computers to have the
image of mere data processors --- nothing revolutionary was wanted.

	The work in AI also led to my first proposals for time-sharing
computer systems.  Early ideas about computers emphasized
running a program for a long time.  The programs were imagined to
be numerical.
  The process of developing the
program was considered auxiliary and so was any interaction with it.
It seemed to me that artificial intelligence required a quite different
approach.  Someone doing AI research might spend almost all of his
time developing the program and interacting with it.  He needed to
sit at a terminal in his very own office and interact at his own
pace and convenience, rather than sign up for half an hour at 2 am
or submit decks of cards.  Again it would require a lot of explanation
today to convince you that the idea wasn't obvious.  Not only wasn't
it not obvious, but it encountered considerable resistance.  For
example, the IBM developers of the 360 knew about the idea but
didn't believe it.  Fortunately, the Digital Equipment Corporation
developers of the PDP-6 computer (somewhat later) had been M.I.T.
students and did time-sharing from the start.

	Actually there were two approaches to providing on-line
computation.  One was time-sharing, in which a large computer switched
its time among users, and the other was the personal computer,
pioneered at M.I.T. Lincoln Laboratories by Wes Clark, who created
the TX-0 and TX-2 computers, both extremely expensive machines intended
to be used on-line by single users.  Personal computers didn't become
economical until the 1970s, and there is still a conflict between
providing on-line service to many users and providing expensive
personal computers to a few users.

\smallskip
\noindent Logic in Artificial Intelligence

	The 1959 ``Programs with Common Sense'' paper said:

\begingroup\narrower\narrower
% COMMON.TEX[E80,JMC]	TeX version Programs with Common Sense
The {\it advice taker} is a proposed program for solving problems by
manipulating sentences in formal languages.  The main difference between
it and other programs or proposed programs for manipulating formal
languages (the {\it Logic Theory Machine} of Newell, Simon and Shaw and
the Geometry Program of Gelernter) is that in the previous programs the
formal system was the subject matter but the heuristics were all embodied
in the program.  In this program the procedures will be described as much
as possible in the language itself and, in particular, the heuristics are
all so described.

	The main advantages we expect the {\it advice taker} to have is that
its behavior will be improvable merely by making statements to it,
telling it about its symbolic environment and what is wanted from it.  To
make these statements will require little if any knowledge of the program
or the previous knowledge of the {\it advice taker}.  One will be able to
assume that the {\it advice taker} will have available to it a fairly wide
class of immediate logical consequence of anything it is told and its
previous knowledge.  This property is expected to have much in common with
what makes us describe certain humans as having {\it common sense}.  We
shall therefore say that {\it a program has common sense if it automatically
deduces for itself a sufficiently wide class of immediate consequences of
anything it is told and what it already knows.}
\par\endgroup

	The main reasons for using logical sentences extensively in AI
are better understood by researchers today than in 1958.  Expressing
information in declarative sentences is far more flexible than
expressing it in segments of computer program or in tables.  Sentences
can be true in much wider contexts than specific programs can be useful.
The supplier of a fact does not have to understand much about how the
receiver functions or how or whether the receiver will use it.  The same
fact can be used for many purposes, because the logical consequences of
collections of facts can be available.

	The {\it advice taker} prospectus was ambitious in 1958, would
be considered ambitious today, and is still far from being immediately
realizable.  This is especially true of the goal of expressing the
heuristics guiding the search for a way to achieve the goal in the
language itself.

	The formalism given in (McCarthy 1959) was just a sketch of
a theory of the achievement of goals by sequences of actions.  It took
a more definite form in the {\it situation calculus} of a 1964 Stanford
report and was published in (McCarthy and Hayes 1969).  The basic
idea is to use the formula
%
$$s' = result(e,s)$$
%
to represent the new situation $s'$ that results when the event $e$
occurs in situation $s$.  The events most studied are actions, and
there are usually conditions that $s$ has to satisfy before we can
infer much about $s'$.  The situation calculus embodies a special
case of reasoning about actions and other events.  First, the events
can be regarded as discrete; they occur in one situation and result
in another, and we don't need to reason about what happens during
the event.  Second, we consider only one event occurring at a time;
concurrent events are not analyzed.

	One feature of situations was emphasized conceptually but
didn't play much role in the actual axiomatizations.  Situations were
regarded as infinitely detailed, e.g. a block on the table was in
a particular location and had a detailed shape and distribution of
material, perhaps down to an atomic level.  Therefore, the formalism
did not provide for knowing a situation exactly but only for knowing
facts about a situation that partially characterized it.  In this
respect situation calculus differed essentially from the mathematical
models used in physics and discussed in most philosophy of science, e.g.
in gravitational astronomy.  In physics models, it is customary to decide
what planets are to be taken into account and whether to represent
them as mass points or whether to consider (say) some moments of their
mass distributions.  In contrast to this, situation calculus was intended
to provide a models to which new detail could be added at any time.
I contend that this open-endedness is an essential characteristic
of the common sense information situation, whether the reasoning is
done by people or by machines.  I now refer to entities like situations
that have infinite detail as rich entities and contrast them with
discrete entities that can be completely described.  Reasoning involves
using discrete entities to approximated rich entities, but the level
of approximation can change during a reasoning process.

	The situation calculus was used in Cordell Green's (1969) PhD
thesis along with resolution theorem proving and a method of ``answer
extraction'' he devised.  However, he and his colleagues at SRI found
that their theorem prover did too much search to be practical.  My
opinion was that this was because they didn't have any way of using
heuristic facts to control the search, but I didn't have a proposal
of how to do it.  I have more ideas about that now, but they still
don't amount to a definite proposal for controlling reasoning
with facts.

	In the event, Fikes and Nilsson (1971), went to a restricted
formalism called STRIPS in which resolution theorem proving was used
for reasoning about properties of single situations, whereas going
from one situation to the next was done by a program that interpreted
the action descriptions directly.  The consequence was a faster
program, but it didn't allow sentences that involved more than one
situation.  Because the control problem still isn't solved, the
practical AI systems that reason about actions all use restricted
languages.

	In the late 1970s, the introduction of formalized nonmonotonic
reasoning revolutionized the use of logic in AI.  See (McCarthy 1977b,
1980, 1986), (Reiter 1980), (McDermott and Doyle 1980).  Traditional
logic is monotonic in the following sense.  If a sentence $p$ is
inferred from a collection $A$ of sentences, and $B$ is a more
inclusive set of sentences (symbolically $A ⊂ B$), then $p$ can be
inferred from $B$.

	If the inference is logical deduction, then exactly the same
proof that proves $p$ from $A$ will serve as a proof from $B$. If the
inference is model-theoretic, i.e.  $p$ is true in all models of $A$,
then $p$ will be true in all models of $B$, because the models of $B$
will be a subset of the models of $A$.  So we see that the monotonic
character of traditional logic doesn't depend on the details of the
logical system but is quite fundamental.

	While much human reasoning corresponds to that of traditional
logic, some important human common sense reasoning is not monotonic.  We
reach conclusions from certain premisses that we would not reach if
certain other sentences were included in our premisses.  For example,
learning that I own a car, you conclude that it is appropriate on a
certain occasion to ask me for a ride, but when you learn the further
fact that the car is in the garage being fixed you no longer draw that
conclusion.  Some people think it is possible to try to save
monotonicity by saying that what was in your mind was not a general rule
about asking for a ride from car owners but a probabilistic rule.  So
far it has not proved possible to try to work out the detailed
epistemology of this approach, i.e.  exactly what probabilistic
sentences should be used.  In fact it seems that the probabilistic
reason would end up using nonmonotonic techniques.
Anyway AI has moved to directly formalizing
nonmonotonic logical reasoning.

	Formalized nonmonotonic reasoning is under rapid development
and many kinds of systems have been proposed.  I shall concentrate on
an approach called circumscription.  It has met with wide acceptance
and is perhaps the most actively pursued at present.  The idea is to
single out among the models of the collection of sentences being
assumed some ``preferred'' or ``standard'' models.  The preferred
models are those that satisfy a certain minimum principle.  What
should be minimized is not yet decided in complete generality and may
be problem dependent.  However, many domains that have been studied
yield quite general theories using minimizations of abnormality or of
the set of some kind of entity.  The idea is not completely
unfamiliar.  For example, Ockham's razor ``Do not multiply entities
beyond necessity'' leads to such minimum principles if one tries
to formalize it in logic.

	Minimization in logic is another example of an area of mathematics
being discovered in connection with applications rather than via
the normal internal development of mathematics.  Of course, the reverse
is happening on an even larger scale; many logical concepts developed
for purely mathematical reasons turn out to have AI importance.

	As a more concrete example of nonmonotonic reasoning, consider
the conditions under which a boat may be used to cross a river.  We all
know of certain things that might be wrong with a boat, e.g. a leak, no
oars or motor or sails depending on what kind of a boat it is.  It would
be reasonably convenient to list some of them in a set of axioms.
However, besides those that we can expect to list in advance, human
reasoning will admit still others, should they arise, but cannot be
expected to think of them in advance, e.g. a fence down the middle of
the river.  This is handled using circumscription by minimizing the set
of ``things that prevent the boat from crossing the river'', i.e. the
set of obstacles to be overcome.  If the reasoner knows of none in a
particular case, he or it will conjecture that the boat can be used, but
if he learns of one, he will get a different result when he minimizes.

	This illustrates the fact that nonmonotonic reasoning
is conjectural rather than rigorous.  Indeed it has been shown
that certain mathematical logical systems cannot be rigorously extended,
i.e. that they have a certain kind of completeness.

	It is as misleading to conduct a discussion of this kind
entirely without formulas as it would be to discuss the foundations of
physics without formulas.  Unfortunately, many people are unaware of this
fact.  Therefore, we present a formalization by Vladimir Lifschitz
(1987) of a simple example called ``The Yale shooting problem''.  Drew
McDermott (1987), who has become discouraged about the use of logic in
AI and especially about the nonmonotonic formalisms, invented it as a
challenge.  (The formal part is only one page, however).  Some earlier
styles of axiomatizing facts about change don't work right on this
problem.  Lifschitz's method works well here, but I think it
will require further modification.

	In an initial situation there is an unloaded gun and a person
Fred.  The gun is loaded, then there is a wait, and then the gun
is pointed at Fred and fired.  The desired conclusion is the death
of Fred.  Informally, the rules are (1) that a living person remains alive until
something happens to him, (2) that loading causes a gun to become loaded,
(3) that a loaded gun remains loaded until something unloads it, (4) that
shooting unloads a gun and (5) that shooting a loaded gun at a person
kills him.  We are intended to reason as follows.  Fred will remain alive
until the gun is fired, because nothing can be inferred to happen to
him.  The gun will remain loaded until it is fired, because nothing
can be inferred to happen to it.  Fred will then die when the gun is
fired.  The nonmonotonic part of the reasoning is minimizing ``the
things that happen'' or assuming that ``nothing happens without a reason''.

	The logical sentences are intended to express the above 5 premisses,
but they don't explicitly say that no other phenomenon occurs.  For example,
it isn't asserted that Fred isn't wearing a bullet proof vest, nor are
any properties of bullet proof vests mentioned.  Nevertheless, a human
will conclude that unless some unmentioned aspect of the situation
is present, Fred will die.  The difficulty is that the sentences admit
an {\it unintended minimal model}, to use the terminology of mathematical logic.
Namely, it might happen that for some unspecified reason the gun becomes
unloaded during the wait, so that Fred remains alive.  The way nonmonotonic
formalisms, e.g. circumscription and Reiter's logic of defaults,
were previously used to formulate
the problem, minimizing ``abnormality'' results in two possibilities,
not one.  The unintended possibility is that the gun mysteriously
becomes unloaded.

\noindent Lifschitz's Cauality Axioms for the Yale Shooting Problem

	Lifschitz's axioms use the situation calculus
but introduce a predicate $causes$ as an undefined notion.

	We quote from (Lifschitz 1987).

``Our axioms for the shooting problem can be classified into three groups.
The first group describes the initial situation:
%
$$holds(alive,S0),\eqno(Y1.1)$$
$$¬ holds(loaded,S0).\eqno(Y1.2)$$
%
The second group tells us how the fluents are affected by actions:
%
$$causes(load,loaded,true),\eqno(Y2.1)$$
$$causes(shoot,loaded,false),\eqno(Y2.2)$$
$$causes(shoot,alive,false).\eqno(Y2.3)$$
%
These axioms describe the effects of 
{\it successfully performed} actions;
they do not say {\it when} an action can be successful.
This information is supplied separately:
%
$$precond(loaded,shoot).\eqno(Y2.4)$$
%
The last group consists of two axioms of a more general nature. We use
the abbreviations:
%
$$success(a,s) ≡ ∀f(precond(f,a) ⊃ holds(f,s)),$$
$$affects(a,f,s) ≡ success(a,s) ∧ ∃v\;causes(a,f,v).$$
%
One axiom describes how the value of a fluent changes after an action affecting
this fluent is carried out:
%
$$success(a,s) ∧ causes(a,f,v) ⊃ (holds(f,result(a,s)) ≡ v=true).\eqno(Y3.1)$$
%
(Recall that $v$ can take on two values here, $true$ and $false$; the
equivalence in $Y3.1$ reduces to $holds(f,result(a,s))$ in the first case and
to the negation of this formula in the second.)
If the fluent is not affected then its value remains the same:
%
$$¬affects(a,f,s) ⊃ (holds(f,result(a,s)) ≡ holds(f,s)).\eqno(Y3.2)\hbox{''}$$

	Minimizing $causes$ and $precond$ makes the right thing happen.
While these axioms and {\it circumscription policy} solve this problem, it
remains to be seen whether we can write a large body of common sense
knowledge in the formalism without getting other unpleasant surprises.
Another current question is whether we can get by with axioms about the
external world only or whether the axioms must contain information about
the purposes of the reasoning in order to determine the preferred models.
Moreover, there are many more possibilities to explore for the formal
minimum principle required for common sense reasoning.

\medskip
\noindent {\bf Conclusions and Remarks}

	1. Progress in using logic to express facts about the world has
always been slow.  Aristotle didn't invent any formalisms.  Leibniz,
who wanted to replace argument by calculation in human affairs, didn't
invent propositional calculus, although it is technically far easier
than the infinitesimal calculus of which he was a co-inventor with
Newton.  Boole, who invented propositional calculus, and who called
his book {\it The Laws of Thought} didn't invent predicate calculus.
Frege and his successors saw no need or possibility of formalizing
nonmonotonic reasoning.  It seems to me that almost any of these
ideas would have been accepted by preceding innovators, e.g. Leibniz
would have accepted propositional calculus, had it been suggested.
Therefore, we must conclude that we humans find it difficult to
formulate many facts about our thought processes that are apparent
when suggested.

	Even with formalized nonmonotonic reasoning there are obstacles
to expressing the general facts about the common sense world in an
epistemologically adequate and elaboration tolerant way.  I take
this as a sign that major innovations are yet to come, and I will
have some suggestions in my more technical lecture.

	I have always wondered why science doesn't progress more
rapidly than it does.  Progress in some sciences is limited by
instrumentation and experimental technique.  For example, this
is true of molecular biology.  In other sciences, progress is
limited by mistaken ideas that direct attention away from the
ideas that lead to progress.  When this occurs, it is often hard
to say why a discovery wasn't made 20 or more years earlier than
it was.  For example, there is no technical reason why nonmonotonic
reasoning couldn't have been formalized in the 1920s.

	2. Artificial intelligence research took form in the late
1950s in ways that were quite different from the ideas of senior
scientists who considered the problem.  I, and I think most other AI
researchers, avoided our seniors, rather than either following them or
opposing them.  This was possible, because it turned out that we were
not dependent on them either for research support or academic
position.  I suspect this would not be good advice in general.
However, it seems to have turned out well in artificial intelligence
in that the ideas related to AI proposed by most of them --- von
Neumann, Wiener and McCulloch have not so far been fruitful.  In my
opinion, only Turing proposed the line of research that has given us
the main results in AI that have been achieved so far.  Of course, the
older proposals have not been refuted, and maybe new people will find
them fruitful.  Connectionism might be regarded as one such gamble.

	3. My work in AI has been in accordance with a certain
philosophical point of view.  To what extent the philosophy has
influenced the research is harder to say.  My subjective impression
is that the philosophy has been important.

		a. Our human knowledge of the world does not consist
of knowledge of sense data or even summaries of it.  Material objects
and human purposes, for example, are far more stable than our sense
impressions of them.  This isn't peculiar to the human situation;
it will also be true for any robots we might build.  It is possible
that babies learn from their experience
that material objects exist, but it's also possible
that this organization of the world is in our genes.  Animals that
presume material objects and some other things are likely to have
an evolutionary advantage.  From the AI point of view, this suggests
that we build into our programs presumptions of structures that exist
in the real world.  If we are going to use logical sentences to
represent knowledge, then there should be variables ranging over material
objects, and also variables ranging over other entities like actions
and beliefs.

	This may seem obvious, but it is in contrast with some views.
For example, some people have proposed that since the experience of
a human or robot is representable by a sequence of sense impressions,
then intelligence might consist of the ability to predict the future
of a sequence from its value up to a given point.  This presumes
nothing about the world and therefore might be preferred by an
extremely cautious philosopher.  Efforts to program computers to
predict sequences have not been informative; nothing that wasn't
obvious resulted from the experiments, and the experimenters were
at a loss how to proceed further.

	I suppose the positivistic emphasis on sense data is a residue
of the early twentieth century reaction against nineteenth century
idealistic philosophy.  This philosophy involved many vaguely postulated
entities, and the reaction against it took the form of admitting only
the most directly observable entities as basic.

	b. Our human ascription of mental qualities, e.g. beliefs, to
each other is warranted by the usefulness of the ascriptions in
understanding, predicting and affecting other people's behavior.  We
don't know whether the tendency to do this is genetic, but one
explanation of autistic children might be a failure of such a
mechanism to develop normally.  It is legitimate to ascribe mental
qualities to any system where it helps explain its behavior.  This
idea is developed as philosophy in (Dennett 1971 and 1987) and for AI
in (Newell 1980) and (McCarthy 1979a).  It still isn't generally
accepted in philosophy.

	c. Many qualities, including especially mental
qualities, that we ascribe to various systems are meaningful
only in terms of approximate theories that enable us to
understand aspects of phenomena but which cannot be given
precise definitions in terms of the state of the world.
See (McCarthy 1979a).

	d. Today artificial intelligence is far from being able
to produce systems of human capability in general reasoning.
This has led to a modesty that one might also recommend to
philosophers.  An AI system involving some concept like {\it belief}
can't purport to use ``the true'' completely general concept
of belief, because there is no agreement about that.  Indeed
at the level of precision required for computational implementation,
there aren't even any candidates.  Therefore, AI has to get
by with limited, approximate concepts.  But maybe that's all
humans have either.  Maybe the philosophers' attempts to understand
belief in general merely force them to construct a concept
rather than discover it.  Maybe there is no completely general
concept of belief that corresponds to what humans actually
do or what computers can be programmed to do.

\noindent References:

\noindent
{\bf Dennett, D.C. (1971)}: ``Intentional Systems'', {\it Journal of Philosophy}
vol. 68, No. 4, Feb. 25.
% logic.2[w87,jmc]

\noindent
{\bf Dennett, D. C. (1987)}:
{\it The Intentional Stance},
M.I.T. Press.

\noindent
{\bf Fikes, R, and Nils Nilsson, (1971)}:
``STRIPS: A New Approach to the Application of 
Theorem Proving to Problem Solving,'' {\it Artificial Intelligence}, Volume 2,
Numbers 3,4, January,
pp. 189-208.

\noindent
{\bf Green, C., (1969)}:
``Application of Theorem Proving to Problem Solving. In IJCAI-1, pp. 219-239.

\noindent
{\bf Jeffress, Lloyd A. (ed.) (1951)}: {\it Cerebral Mechanisms
in Behavior: The Hixon Symposium}, Wiley.

\noindent
{\bf Lifschitz, Vladimir (1987)}:
``Formal Theories of Action'', 
in: F. M. Brown (ed.),
{\it The Frame Problem in Artificial Intelligence}, Morgan Kaufmann, pp. 35--58.
Reprinted in: M. L. Ginsberg (ed.), {\it Readings in Nonmonotonic Reasoning},
Morgan-Kaufmann, pp. 410--432.

\noindent
{\bf McCarthy, John (1956)}:  ``The Inversion of Functions Defined by Turing
Machines,'' in {\it Automata Studies, Annals of Mathematical Study No. 34,}
Princeton, pp. 177-181.

\noindent
{\bf McCarthy, John (1959)}: ``Programs with Common Sense'', in {\it
Proceedings of the Teddington Conference on the Mechanization of
Thought Processes}, Her Majesty's Stationery Office, London.
%  common[e80,jmc],
% common.tex[e80,jmc]

\noindent
{\bf McCarthy, John and P.J. Hayes (1969)}:  ``Some Philosophical Problems from
the Standpoint of Artificial Intelligence'', in D. Michie (ed), {\it Machine
Intelligence 4}, American Elsevier, New York, NY.
%  phil.tex[ess,jmc] with slight modifications

\noindent
{\bf McCarthy, John (1977a)}:
``History of LISP'', in Proceedings of the ACM Conference on the
History of Programming Languages, Los Angeles.
%  lisp[f77,jmc]

\noindent
{\bf McCarthy, John (1977b)}:
``Epistemological Problems of Artificial Intelligence'', {\it Proceedings
of the Fifth International Joint Conference on Artificial 
Intelligence}, M.I.T., Cambridge, Mass.
%  ijcai.c[e77,jmc]

\noindent
{\bf McCarthy, John (1979a)}:
``Ascribing Mental Qualities to Machines'' in {\it Philosophical Perspectives 
in Artificial Intelligence}, Ringle, Martin (ed.), Harvester Press, July 1979.
%  .<<aim 326, MENTAL[F76,JMC],
% mental.tex[f76,jmc]>>

\noindent
{\bf McCarthy, John (1979b)}: 
``First Order Theories of Individual Concepts and Propositions'', 
in Michie, Donald (ed.) {\it Machine Intelligence 9}, (University of
Edinburgh Press, Edinburgh).
%  .<<aim 325, concep.tex[e76,jmc]>>

\noindent
{\bf McCarthy, John (1980)}: 
``Circumscription - A Form of Non-Monotonic Reasoning'', {\it Artificial
Intelligence}, Volume 13, Numbers 1,2, April.
%  .<<aim 334, circum.new[s79,jmc], cirnew.tex[s79,jmc]>>

\noindent
{\bf McCarthy, John (1982)}: ``Common Business Communication Language'', in
{\it Textverarbeitung und B\"urosysteme}, Albert Endres and J\"urgen Reetz, eds.
R. Oldenbourg Verlag, Munich and Vienna 1982.
%  cbcl[f75,jmc]

\noindent
{\bf McCarthy, John (1983)}: ``Some Expert Systems Need Common Sense'',
in {\it Computer Culture: The Scientific, Intellectual and Social Impact
of the Computer}, Heinz Pagels, ed.
 vol. 426, Annals of the New York Academy of Sciences.
%paper
%presented at New York Academy of Sciences Symposium.
%  common[e83,jmc]
%common.tex[e83,jmc]

\noindent
{\bf McCarthy, John (1986)}:
``Applications of Circumscription to Formalizing Common Sense Knowledge''
{\it Artificial Intelligence}, April 1986
%  circum.tex[f83,jmc]

\noindent
{\bf McCarthy, John (1987)}:
``Generality in Artificial Intelligence'', {\it Communications of the ACM}.
Vol. 30, No. 12, pp. 1030-1035
% genera[w86,jmc]

\noindent
{\bf McCarthy, John (1987)}:
``Mathematical Logic in Artificial Intelligence'', in
 {\it Daedalus}, vol. 117, No. 1, American Academy of Arts and Sciences,
Winter 1988.

\noindent
{\bf McDermott, D. and J. Doyle, (1980)}:
``Non-Monotonic Logic I,'' {\it Artificial Intelligence\/},
Vol. 13, N. 1

\noindent
{\bf Newell, Allen (1981)}: ``The Knowledge Level,'' {\it AI Magazine\/},
Vol. 2, No. 2.

\noindent
{\bf Reiter, R.A. (1980)}: ``A Logic for default reasoning,'' {\it Artificial 
Intelligence\/}, 13 (1,2), 81-132.

\noindent
{\bf Robinson, J. Allen (1965)}: ``A Machine-oriented Logic Based
on the Resolution Principle''. {\it JACM}, 12(1), 23-41.

\noindent
{\bf Shannon, Claude and John McCarthy (eds.) (1956)}:
{\it Automata Studies}, Annals of Mathematics Study 34,
Princeton University Press.

\noindent
{\bf Turing, A.M. (1950)} ``Computing Machinery and Intelligence''. {\it Mind}
{\bf 59}, 433-460.
\smallskip\centerline{Copyright \copyright\ \number\year\ by John McCarthy}
\smallskip\noindent{This draft of kyoto.2[e88,jmc]\ TEXed on \jmcdate\ at \theTime}
\vfill\eject\end
	In this lecture, I shall recount some history of artificial
intelligence and its precursors, tell about some of the concepts
that have been discovered and give some opinions about the direction
in which further progress may be made.

	Making intelligent machines is difficult.  We humans are not
very good at observing and understanding our own intellectual
processes or at understanding the informatic situations in which we
find ourselves.  However, there's hope.