COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	ARTIFICIAL INTELLIGENCE AND THE DEPARTMENT OF DEFENSE
C00003 00003	INTRODUCTION
C00015 ENDMK
C⊗;
ARTIFICIAL INTELLIGENCE AND THE DEPARTMENT OF DEFENSE


Abstract: The Department  of Defense has been supporting  research in
artificial  intelligence  for   a  number  of  years.    This  report
describes the scientific problem of artificial intelligence,  reviews
the progress made both in  DoD supported work and elsewhere, outlines
the  present  state  of  the  artificial  intelligence  problem,  and
suggests  where  the  research  may  go  in  the  near  term  future.
Potential  applications of  concern to  DoD are  identified  and some
milestone goals aimed at realizing these applications are set forth.
INTRODUCTION

	The phrase "artificial intelligence" was coined in 1954 to
describe research aimed at making machines behave intelligently.
The long term goal set forth at that time was to make machines as
intelligent as humans, and the short term goals were to make
computers perform specific tasks that require intelligence in
humans.  This is still the general structure of much of the work
that goes under the name "artificial intelligence", but it would now
be necessary to add another goal.  This is the study of the
structure of intelligent behavior, because it is increasingly
evident that intelligent behavior has a structure that depends on
the problem being attacked only in a general way and is almost
independent of whether the problem solver is man or machine.  In
other words, there are certain general structures and mechanisms of
intelligence that can be studied scientifically.  From this point of
view, the word "artificial" might be omitted from the term, and we
could look for a reasonably graceful term that had the meaning of
"intelligencology". Finding such a term might be nice, because
"artificial intelligence" grates on many people, because it suggests
science fiction.  It has turned out, however, that it is necessary
to have a term that keeps peoples attention on the main goal of
eventually making computer programs capable of carrying out every
important intellectual activity carried out by people.  Otherwise,
research may wander into byways unrelated to this main goal.

	In our view, the problem of intelligence has a certain core
surrounded by a number of shells.  The "shell problems" are
important, but they depend on the "core problems" for their
solutions, whereas the "core problems" interlock with each other but
can be attacked substantially independently of the "shell problems".

	Some of the core problems are:

	1. Heuristics: The efficient search through spaces of
alternatives for entities that will help solve problems.

	2. Representation: The structure of the information relevant
to a problem that an intelligent entity can obtain about the world.
How this information can be represented in the memory of a computer.

	3. Modelling: How the information relevant to a problem can
be separated from all the information at the machine's disposal
posed as an "isolated" problem whose solution can be translated into
action in the real world.

	4. Automatic programming: How to go from a collection of
information that is implicitly sufficient to determine what to do in
most cases of a problem to an efficient program that can perform
tasks quickly.

	Some problems that we hope are shell problems are:

	1. Language: How information is or can be coded for
transmission from one intelligent being to another in such a way
that neither being need understand the complete mental state of the
other.

	2. Uncertainty: How to act in conditions of uncertainty.

	3. Perception: How information is obtained from the external
world.

	This classification into core and shell problems is not
intended to imply that the shell problems are less important than
the core problems. In some of the most immediate application areas,
shell problems predominate, because the associated core areas are
well enough understood for the particular application.

	We shall deal with the structure of artificial intelligence
as it is presently understood further in a later section.  However,
AI specialists know and the lay reader should understand that even
the basic structure of artificial intelligence is enmeshed in
controversy, and no view can be given that would satisfy all workers
in the field - much less our external critics.


Some history

	While the idea of artificial intelligence may be traced far
back in literature, the intellectual ancestry of present research
goes only to A.M. Turing's 1950 paper "Computing Machinery and
Intelligence" published in Mind, a British philosophical journal.
The first research group to work continuously on the problem was the
Carnegie-Mellon group starting in 1954, and the M.I.T. group was
organized in 1957.  ARPA support of AI research started in 1963.
Foreign work started later, but there are now active groups with
considerable accomplishments in Britain and Japan.  There has been
isolated good work in the Soviet Union, but systematic research in
AI has just gotten started there.  Other countries are also
beginning to be active.

	Early work concentrated mainly on specific sample tasks with
the idea that the experience gained on these tasks could be
generalized.  These tasks included game playing (e.g. chess),
mathematical theorem proving, the transformation of mathematical
expressions, and tree searching. Codification of the experience
gained on these specific problems into general principles proved
difficult and not a lot of success was obtained. What success there
was is reflected in Minsky's 1960 paper "Steps towards Artificial
Intelligence".  

	In our opinion, the inability to generalize the early
results has two causes.  First, the importance of explicit work on
this problem and the importance of doing the specific experiments in
such a way as to get generalizable results was not recognized.  This
is sometimes called the "look ma, no hands" school of artificial
intelligence.  A computer is made to do a task no computer has done
before, and the paper merely reports the success with no attempt to
get generalizable information.  Secondly, however, there does not
yet exist an accepted general framework for intelligent systems into
which particular behaviors can be fitted.  More specifically, the
lack of development of a theory of representation of information has
meant that each programmer devises the representation used by his
own program, and the program itself does not determine the
representation. Therefore, there is little contact between the
different programs and they cannot be unified.  If the lay reader
finds this a bit vague, he can console himself with the fact that
the experts are also confused about it. If we understood it better,
maybe we could get better generalizations.

	The use of the languages of modern mathematical logic in
artificial intelligence has led to results of general importance.
In the first place, proving theorems in predicate calculus has
provided a good domain for heuristic programming.  The "resolution"
formalism of J. Allen Robinson has provided the basis for much work
in making better theorem provers by doing resolution more cleverly.
Secondly, the formalism of predicate calculus has been used to
express facts about the world.  The study of how facts about
problems can be expressed starts with McCarthy's 1959 paper
"Programs with Common Sense".  However, some early optimism that one
could simply express facts in predicate calculus and find the
solutions to problems by exercising general resolution type theorem
provers has been disappointed.  There are two difficulties: First,
the heuristics used must depend on the particular predicates, and
this means that either the programs must be quite specialized or
that ways must be found to express "how to think" in the predicate
calculus itself.  Second, much important information is not easily
expressed in predicate calculus, and it seems that methods have to
be found for "plausible reasoning" and "jumping to conclusions" as
well as for carrying out deduction.