COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002
C00031 ENDMK
C⊗;

June 19, 1981

MARSEILLES SPEECH TO FRENCH UNIVERSITY PRESIDENTS

Artificial intelligence is the science and technology of making computer
programs that behave intelligently.  Intelligent behavior is taken to be that
which requires intelligence in humans, leaving out things like arithmetic
calculation and simple data storage which are achievable by mechanisms too
simple to be called intelligent.

The definition is bad, because it depends on our intuitive notions of
intelligence in humans, but there are plenty of examples.  There are intellectual
tasks which have been programmed and others which we cannot program today and
where the difficulty is clearly an insufficient understanding of intellectual
mechanisms.

Our long term goal is to exceed human intelligence, but there are poorly understood
conceptual scientific problems in the way.  Most likely we will require one or
more geniuses of the level of Einstein before intelligence is understood well
enough to make programs with all the intellectual mechanisms used by humans.
Meanwhile, many discoveries have been made, many impressive and some useful 
programs have been written, and the prospects for both scientific and engineering
progress continue to be good.

Artificial intelligence research can be said to have begun in 1950 with Alan
Turing`s paper, "Computing Machinery and Intelligence," in which he 
discussed the problem of programming specific tasks, making programs that
could learn like a child.  He also answered many philosophical objections to 
the possibility of artificial intelligence.  Perhaps Turing's major conceptual
contribution was the idea of a universal computing machine that could do any
computation do-able by any machine.  This idea put AI firmly on the path of
programming computers rather than inventing machines.

The major American laboratories in artificial intelligence were established
between 1954 and 1964 at Carnegie-Mellon University, MIT, Stanford University,
and SRI.  Early research focussed on specific tasks such as playing chess and
proving mathematical theorems.  Attempts to achieve general problem-solving
ability were informative but not very successful.  In particular, efforts to
use general problem solvers to improve themselves were not successful.

In 1966 an MIT chess program won the Class D trophy in a tournament, and in
1978 another program won the Minnesota Open Tournament but was beaten for the 
state championship when the human players took it seriously enough to look for
its weaknesses.  The present best programs play some games at master level
but have endgame weaknesses that no one sees clearly how to remove.  The 
intellectual mechanisms we understand and can program suffice for much of chess.
and the micro-processor chess machines now marketed can beat most humans.

One important method of research in artificial intelligence is to conjecture
an intellectual mechanism, often by introspection, and to verify that a mechanism
has been identified and to test its power by writing a program.

Attempts to describe intellectual mechanisms are very old but were invariably
unsuccessful before experience with computers gave us a better idea of what 
is involved.  It isn't merely that the mechanisms proposed by pre-computer 
philosphers, linguists, and psychologists were incapable of exhibiting the behavior
they were designed to explain.  They simply didn't describe mechanisms at all,
so that it is impossible even to make computer experiments to test their theories.
Associationist psychology is perhaps the most obvious example.  Even today it is
common for a researcher to wrongly suppose that he has proposed an intellectual 
mechanism when he has merely stated a property he would like a mechanism to have.

Self-observation often fails to observe a mechanism used by the observer, and 
experiment often only shows that the self-observation has been inadequate without
telling what is missing.  Here is an example from game playing.  Turing, Shannon,
and no doubt others independently, proposed chess programs that could generate the 
"plausible moves" in a position, generate plausible replies to each move, generate
further replies, etc.  In order to prevent the search from taking too much time
for any computer, it is neecessary to stop going deeper according to some 
criterion of depth or the position being static and to evaluate the resulting
position according to some approximate evaluation function, taking into account
material, etc.  Values are then assigned to preceding positions according to a
rule that assumes that each side will move to the successor position that has 
the best value for his side.  Carrying the analysis back to the positions
immediately reachable by the machine then determines its move.  Early programs
operated in this way.

Figure 1 shows a "tree" of 17 positions that might be examined by such a program,
although the programs may examine between hundreds to hundreds of thousands
of positions.



                            Figure 1



The "minimax algorithm" described above will examine all 17 positions.  However,
the positions circled in red need never be examined if the program proceeds from
top to bottom.  Namely, as soon as the position labelled 5 and circled in blue has 
been reached, the program should know that its second move is not as good as its 
first move no matter what the opponent's other replies are.  Likewise, when the
blue circled 6 is reached, the program should conclude that its third move is
no good either.  This so-called 2B-heuristic was not noticed by the first
programmers of chess even though they included some champion players and is
implicitly used, i.e., the red circled moves would not be examined, even by 
beginners.  By better self-observation, variants of 2B were independently
discovered by several people, and it is used in all modern programs.

Starting in the later 1960s much AI research has concerned "expert programs"
that are written in consultation with experts in some field and try to incorporate
the knowledge that distinguishes experts from novices.  Chess programs are
examples, and one Russian computer scientist proposed that we should regard
chess as the Drosophila Melanogaster of artificial intelligence.  However, since
the early 1970s funding agencies often demand promises of practical applications 
so that most "expert system" research gives at least lip service to short
term goals.  (Somtimes this resembles deciding to do genetic research with
elephants because elephants are more useful than fruit flies.)

A typical expert system is MYCIN by E. Shortliffe of Stanford.  Its subject is
the diagnosis and treatment of bacterial infections of the bloodstream.  There
are many kinds of such infections, but most doctors are not very familiar with
them and would welcome advice from an expert.  MYCIN can conduct a dialog in 
limited English with a doctor about the symptoms of the patient, the blood tests
that have been performed, the antibiotics that have been tried, and the results
of these efforts.  As a result of the dialog, MYCIN may suggest possible
micro-organisms, tests that might be performed and antibiotics that might be
administered.  It always gives reasons for its advice, and it can answer
questions about why it has rejected certain alternatives.  Experts rate its
performance as superior to most medical students and doctors and equivalent
to that of the experts themselves.  I don't know how much it is used in practice.


Like the chess programs, MYCIN is very narrowly focussed.  It "knows" facts about
bacteria, tests, and treatments, but it probably doesn't know about doctors,
hospitals, patients, death or recovery.  It almost certainly couldn't answer whether
a bacterium is bigger than a man.  In short, present expert systems have
little general knowledge and almost completely lack what goes by the name of
"common sense."

The goal of providing computers with the facts and reasoning ability that make
up "common sense" was first proposed in 1958 but has proved extremely difficult.
Perhaps it is the core of the artificial intelligence problem.  While I don't
think we are very close to achieving common sense, it also seems to me that real
progress has been made and the prospects are good for the near future.  Progress
will be faster if scientists and their supporters will concentrate on identifying
fundamental problems rather than expecting to accomplish basic research only as a
byproduct of applied projects.  Also the problem of understanding common sense
seems to be more likely than any other to lead to a breakthrough to human-level
artificial intelligence.

An early approach to achieving common sense involves expressing general facts about
the common sense world in one of the languages of mathematical logic.  Facts about
a particular situation and the goals to be achieved are also to be expressed in 
this language, and a program is to be invoked that can deduce, according to the 
rules of logic, that a certain strategy of action is appropriate for achieving
the goals and that a certain action should be done to begin with.  As time passes,
new situations arise, and the sense organs of the computer (or robot controlled
by the computer) express their observations as new sentences which are then used 
in new deductions to carry out the plan.

This strategy of common sense through logic tends to split the artificial
intelligence problem into two parts, usually called heuristics and
epistemology.  The heuristic problem is that of search through spaces of possible
actions and their consequences for a strategy that will achieve the goal.  Chess
is the heuristic problem par excellence.  However, the epistemological problem is 
more basic and probably more difficult, and until it is solved for some class
of problems, it is usually impossible even to start on the heuristics.

In philosophy, epistemology is the study of knowledge, usually concentrating
on its justification and limits.  In AI epistemology is again the study of
knowledge, but this time concentrating on what partial knowledge of situations
can be obtained by a person or robot with given opportunities to observe, how
this knowledge can be represented in a computer and what rules determine what
further knowledge can be legitimately inferred from that already explicitly
present.  It is also necessary to have general knowledge about the likely
effects of actions and other events in situations with given properties.

The concrete study of common sense knowledge has only been undertaken recently
in connection with AI.  It might be described as a previously invisible branch
of philosophy.  Take, for example, what Pat Hayes calls common sense physics.  If
I spill my coffee cup on the podium, the people in the back of the hall will not
act to escape getting wet while the people in the front row might.  Their
"common sense physics" tells them about the likely dimensions of the event.
These calculations owe nothing to the science of hydrodynamics although a
hydrodynamicist might be able to use his knowledge to improve common sense
rules.

Formalizing common sense, i.e., devising a suitable logical language and 
expressing facts in it has proved extremely difficult.  Thus any
simple expression of laws determining the effects of moving objects
around a room is correct only in a limited class
of situations.  Also the computations required to solve the problems when
they were expressed in logic were too lengthy and grew with the
complexity of the situation and the number of laws represented in a discouraging
way.  Consequently, in the early 1970s many AI researchers abandoned representation 
in logic and devised their own more or less ad hoc systems for representing
knowledge.  These worked better on simple problems, but when they were
elaborated to express more complex information, it was observed
that bits of first order logic were being reinvented.

Now logic is coming back into fashion but with several changes in the way it
is treated.  First, it is now recognized that whether logical formulas are used
within the program or not, it is worthwhile to think about programs at
a "logical level" of abstraction, as Allen Newell expresses it.  This involves
formulating in logic what information the program can represent and what
general facts are available to it for drawing further conclusions.  Many 
general facts about programs can be understood this way without going into detail.
Second, problem solving methods in logic, whose strategies depend on the domain,
often reduce the feared "combinatorial explosion."  Third, Alain Colmerauer's
PROLOG and variant programming languages based on logic are increasingly studied and 
used in AI.  Fourth, the recent development of formalized non-monotonic
reasoning promises to make possible expressing common sense knowledge of the
world in a compact way.

Formal deduction in mathematical logic is monotonic in the sense that
any conclusion P that follows from a set A of premisses also follows from
any more inclusive set of premisses.  Indeed, the same proof that proves P from
A will serve to prove it from B.

Ordinary reasoning does not have this property.  If I tell you that my
plane is scheduled to leave Marseilles at 9 p.m., you will conclude that
I must leave Luminy well before that 9 p.m.  However, if I add the
non-contradictory statement that the departure of the plane has been delayed 
for three hours, you will no longer draw the conclusion.

The fact is that the conclusion that I must leave Luminy well before 9 p.m.
does not follow deductively from the fact that my plane is scheduled to leave 
at that time.  Many other facts, indeed more than we usually think of, would
have to be added before the conclusion would be a rigorous consequence of the
scheduled departure combined with common knowledge about travel.  The step by
which we conclude that I should depart before 7:30 p.m. is only partly logical
deduction.  The other part is some kind of default reasoning.  One should
leave well before the scheduled departure of one's plane unless there is some
reason to do otherwise that follows from other known facts.  This is rather
vague as stated, but we can now make it precise enough so that computers can be
programmed to do such reasoning.

For the benefit of any mathematicians present I shall give one formalization of 
such reasoning.  Let P be a predicate symbol, and we want to say that the
extension of P is as small as is compatible with the sentence A(P).  This
can be expressed by the formula

     A(P)∧∀ Phi.A(Phi)∧∀x.(Phi(x)⊃P(x))⊃∀x.(P(x)⊃Phi(x))

If P(x) is interpreted to mean that x is a reason for not going to the
airport and A(P) is the conjunction of the known facts, then (1) expresses
that I should go to the airport on time unless some reason not to
follows from A(P).

The study of formalized non-monotonic reasoning is just beginning, and there
is reason to hope that it will greatly advance the epistemological part
of AI.  Some interesting results have been obtained here at Luminy by
two students of Professor Colmerauer.

Artificial intelligence research has recently become a part of a larger entity
called cognitive science that includes much of cognitive psychology, 
linguistics, and philosophy, and AI is partly responsible for new trends in
these fields.

In psychology, the demise of behaviorism has been hastened by the availability
of models of behavior that include important internal states, i.e., are
not just stimulus-response relations.

In linguistics AI has shown the utility of information about semantics and 
about the particular situation in parsing sentences.  Besides that it is
becoming increasingly clear that the semantics of the information
conveyed in natural language can and must be studied entirely apart from
syntactic questions.

In philosophy the new understanding of mechanism has thrown new light on 
questions of reductionism and the criteria for ascribing mental qualities such as 
belief both to humans and to machines.  The requirement for formalizing
common sense knowledge leads to the conclusion that "naive" views of the
world have much more content requiring study than philosophers have supposed.

Two recent books are Daniel Dennett's BRAINSTORM and Aaron Sloman's THE
COMPUTER REVOLUTION IN PHILOSOPHY.

Since you are university presidents, it seems to me that I should say something
about the position of AI research in universities.  In most American universities,
AI resides in computer science departments.  As in any new field, AI researchers
have a wide variety of backgrounds, including mathematics, psychology,
philosophy, and engineering.  The main material requirement is a good
time-sharing computer system with the languages, such as LISP, that are widely
used in AI research.  In America and some other countries the PDP-10
computer and its variant the DEC system 20 have become standard for AI
research, and this has greatly helped progress by facilitating exchanges of
programs and letting people bring their programs when they move.

The ARPA net which connects many American AI research computers electronically
has also been a unifying force.  The PDP-10 is becoming obsolete because
of its limited address, but I believe a new standard will develop, and
AI research in other countries will be aided if its research groups
can choose the computers most appropriate for their work unhampered by
bureaucratic considerations or by fruitless attempts to satisfy the
needs of physical sciences for numerical calculation and information for
symbolic computation by the same computer system.