00100	          ARTIFICIAL INTELLIGENCE AND NATURAL INTELLIGENCE

00200	
00300	by John McCarthy, Professor of Computer Science, Stanford University

00400	
00500	
00600		In this paper, I shall describe some of the work  done  under
00700	the flag of artificial intelligence (AI) and discuss its relevance to
00800	questions of natural intelligence.
00900	
01000		Interaction  between AI and the study of natural intelligence
01100	is greatly affected by the  primitive  state  of  knowledge  in  both
01200	fields.  There are many tasks performed by human intelligence that we
01300	have no clear idea of a mechanism for performing.   Thus we  know  no
01400	mechanism  that  will  locate the people in the visual field and give
01500	the present positions of their limbs.  We know no mechanism that will
01600	generate  from  a  photograph  of  a person a generalized idea of his
01700	appearance that will enable him to be recognized in the  future.   We
01800	also  don't know any way of representing a person's general knowledge
01900	of airplanes, travel agents, hotels,  etc.    that  will  enable  the
02000	planning  of  a trip.  Moreover, although we can mention a very large
02100	number of specific tasks that give us  difficulty  in  determining  a
02200	mechanism,  we  can't  yet divide the process of intelligent behavior
02300	into  a  comprehensive  set  of  subprocesses  that  can  be  studied
02400	separately.
02500	
02600		Therefore,  if AI research turns up a mechanism for doing one
02700	of these tasks, the psychologist will be aided in his search for  the
02800	mechanism used by humans or animals. Conversely, an AI researcher who
02900	is stumped in programming a task may  benefit  from  any  information
03000	obtainable  about how humans do it. The psychologist who has a theory
03100	of how a task is performed will benefit from trying  to  program  it.
03200	Experience indicates that unless this is done the theory is likely to
03300	remain so vague that it is uncertain whether a mechanism  has  really
03400	been  specified.  The AI researcher can often tell immediately that a
03500	proposed mechanism wouldn't work or even that no mechanism has really
03600	been specified.
03700	
03800		The most extensive work in correlating human performance with
03900	that of computer programs has been done at Carnegie-Mellon University
04000	by Allen Newell, Herbert Simon and  their  collaborators.   I  cannot
04100	here summarize this work or the conclusions they have reached.
04200	
04300		Outsiders often ask what formal definition of intelligence is
04400	used in AI research.   Unfortunately, we don't have one; AI  has  had
04500	to  be  defined  as  making  computers solve problems that people are
04600	called intelligent for solving.     In the short  run  this  is  good
04700	enough,  because  there  are  plenty of problems that clearly require
04800	intelligence for human solution that we are having  great  difficulty
04900	getting  machines  to  solve.       In  the  long  run, the adjective
05000	"intelligent" applied to behavior or systems may become  a  technical
05100	term  in a theory of intelligence, but no-one in AI has yet attempted
05200	to formulate a general theory of intelligence.    At least  there  is
05300	no such theory taken seriously by many workers in the field.
05400	
05500		Serious  work  in  AI didn't start until just after the first
05600	stored program computers.  Much of it then went into an area  I  call
05700	heuristics  which is concerned with searching spaces of possibilities
05800	for solutions to problems.  The first  problems  considered  included
05900	games  -  chess, checkers, etc., and mathematics - theorem proving by
06000	manipulation of formulas in such domains as  Euclidean  geometry  and
06100	symbolic logic.
06200	
06300		What was learned from this?
06400	
06500		Not  as much as one might hope but something.      We learned
06600	less than we should have, because the writers of  heuristic  programs
06700	haven't  been  very  analytical  about it.  Mostly they belong to the
06800	"look ma, no hands" school of artificial intelligence.   They write a
06900	program,  try it out, and if they have to write a paper, they tell us
07000	how the program works and how well it did.  They don't tell  us  much
07100	of  general  interest  about artificial intelligence learned from the
07200	particular task.  However, there is much more to be learned from  the
07300	programs than the authors have so far made explicit.
07400	
07500		First  of  all,  we  learned  what  was  easy  and  what  was
07600	difficult.  Thus a game called kalah is quite easy (we got a computer
07700	program  (Russell  1965)  better  than any human player); checkers is
07800	harder, but Samuel's program (Samuel 196x) got some  draws  with  the
07900	U.S.   champion  till he learned its ways; chess is harder still (the
08000	best program so far (Greenblatt 1967) plays class C chess);  and  the
08100	Japanese  game  of  go  is so hard that the best program (Ryder 1971)
08200	can't beat a novice.
08300	
08400		The reason seems  to  be  that  these  games  differ  in  the
08500	intellectual  mechanisms  required for successful play. When the game
08600	requires intellectual mechanisms that we understand how  to  program,
08700	then  the  programs  do  well.  If, on the other hand, the mechanisms
08800	used in high quality human play are not well understood, then  it  is
08900	hard to make a good program.  Experiments in programming computers to
09000	play  games  help  us  identify  and  verify  our  understanding   of
09100	intellectual mechanisms.
09200	
09300		Thus, in kalah there is no long range strategy  -  there  are
09400	only  tactical gimmicks for winning a few points. Moreover, there are
09500	only a few moves in any position and it is hard to eliminate some  of
09600	them  as  without interest. Once these mechanisms are programmed, the
09700	machine's speed wins. Checkers requires more of the player.    In the
09800	first  place,  there  is  some  strategy  -  if I can hold two of the
09900	opponent's men with one of mine and we both king  all  our  remaining
10000	men I will win by outnumbering him on the rest of the board.    I can
10100	win a game using this principle without being able to look  ahead  on
10200	the   move   tree   to  the  point  where  my  victory  is  explicit.
10300	Nevertheless, there are relatively  few  moves  and  a  computer  can
10400	overcome  our  limited  ability  to program strategy by looking ahead
10500	very far. Chess has many more moves in a given position than checkers
10600	so  the  computer  cannot  look very far ahead unless it can severely
10700	restrict the class of moves worth looking at.     Humans do  this  by
10800	making  plans  and  by  considering  only moves that further plans or
10900	refute them.  None of  the  present  chess  programs  generate  plans
11000	worthy  of  the  name.  The game of go introduces still another major
11100	complication.    The number of moves in a position  is  usually  more
11200	than  100  so that straight look-ahead is of no use whatsoever. Human
11300	play depends on our ability to analyze the position into subpositions
11400	and  to  examine  first  the  move trees of the subpositions and then
11500	their interaction. This analysis  is  rather  subtle,  and  we  can't
11600	program it well yet.
11700	
11800		Another  advantage  that  human  players of chess and go have
11900	over present computer programs lies in the way they consider the tree
12000	of  legal  moves.  Almost no present programs associate corresponding
12100	moves on different branches of the tree even though  moving  a  given
12200	piece  to  a  given  square  may  have similar properties in distinct
12300	positions.  In the case of good moves, this is not  hard  to  do.   A
12400	move  that  has  been  shown  to be good in one position may be tried
12500	first in other positions. However, consider  the  move  P-R3  in  the
12600	initial position in chess. Under the name of "killer heuristic", this
12700	mechanism has been used to good effect in  a  chess  mating  program.
12800	Present  programs  find that it has no special virtues just as humans
12900	do and don't make it.  However, after  the  program  has  decided  to
13000	investigate,  say,  P-K4 and an opponent's reply, the program repeats
13100	its investigation of P-R3.  In fact, the program  may  consider  P-R3
13200	fifty  times  in deciding on its first move.  The obvious solution is
13300	to put on a list of bad moves and not consider it further, but events
13400	may  occur  that will make it relevant, so it has to be arranged that
13500	certain events  will  trigger  renewed  consideration  of  P-R3.   No
13600	program  along these lines has yet been written, and we still observe
13700	that chess programs waste most of their time considering  moves  that
13800	any human chess player can tell you are not worthy of consideration.
13900	
14000		One   may   draw   another  conclusion  relevant  to  natural
14100	intelligence.       This is that the solution methods used  for  some
14200	important  classes  of problems are objective; they depend neither on
14300	heredity  nor  environment,  because  since  they  are  required  for
14400	success,  they  must  either evolve with the species or be learned by
14500	the individual. This point is illustrated by the αβ-heuristic used in
14600	game playing. When the possibility of computer game playing was first
14700	discussed (Shannon 1950), it was proposed to use a minimax  procedure
14800	for  selecting the right move to make: the move tree is examined to a
14900	certain depth terminating in a move of the machine's.    Then  values
15000	are  assigned to all terminal positions.  Then values are assigned to
15100	the  positions  immediately  preceding  the  terminal  positions   by
15200	assigning to such a position the best value of the terminal positions
15300	to which moves exist from it.  Values are assigned to  the  preceding
15400	positions  by  taking  the  worst  value  of  a  position immediately
15500	accessible since the move at that  point  belongs  to  the  machine's
15600	opponent.    This procedure is continued until a value is assigned to
15700	the initial position.    The first game playing  programs  used  this
15800	procedure  or  one equivalent to it in amount of work.    However, in
15900	1956 I noticed that if the program has found a  move  that  gains  an
16000	amount  α and in the course of analyzing another move it has found an
16100	opponents reply that holds the gain to less than α, then there is  no
16200	point  in examining other moves of the opponent.  The αβ-heuristic is
16300	a generalization of this observation which was  put  into  its  final
16400	form   by   Edwards   and  Hart  (1960),  and  which  was  discovered
16500	independently by Brudno (1963).  A little thought  convinces  one  of
16600	two  points:    First, every human uses αβ, including the programmers
16700	whose programs don't. Second, good game playing requires its  use  or
16800	something equivalent, whether the player be man, machine, or Martian.
16900	
17000		Some  of  the  aspects  of  game-playing  have  more  general
17100	significance  in  artificial  intelligence  and  in  the   study   of
17200	intelligence  in general.  For example, the analysis of problems into
17300	subproblems  that  can  be  studied  separately  at  first  with  the
17400	interactions having to be combined later is a general phenomenon. The
17500	depth first and breadth first  methods  of  tree  searching  and  the
17600	alpha-beta  method  of  cutting off tree search are relevant to other
17700	problems than games. (alpha-beta is relevant to games against  nature
17800	arising from the search for worst case solutions to problems).    For
17900	an extended discussion of tree search see (Nilsson 1971)  and  for  a
18000	discussion  of  particular  problems  that  have  been worked on, see
18100	(Slagle 1971).
18200	
18300		Another moral that arises from the experiments with different
18400	games  is  the  invalidity  of  identifying  a  specific intellectual
18500	mechanism with intelligence.  The prize example of this would be  the
18600	identification of intelligence in rats with maze learning ability. We
18700	can easily make a maze running computer program better than  any  rat
18800	or  human,  but  we  would not be inclined to regard it as especially
18900	intelligent, because we would not have to put into  it  many  of  the
19000	known intellectual mechanisms.
19100	
19200		So  far  it  has usually turned out that once an intellectual
19300	mechanism is discovered it seems obvious  even  though  people  wrote
19400	programs  for years without including it.  This leads us to hope that
19500	the source of difficulty in understanding intelligent behavior is bad
19600	methodology,  and  once  we  know what to look for, the mechanisms at
19700	least of conscious thought will not be  difficult  to  observe.    At
19800	present,  it  seems  that  most  theories  of cognitive processes are
19900	clearly wrong in that either no mechanism is described or, if it were
20000	programmed, it would not produce the results ascribed to it.
20100	
20200		We  can also derive some instructive lessons from the work in
20300	computer theorem proving.    Some of the mathematicians who  did  the
20400	first  work  in  computer  theorem proving were very optimistic about
20500	what theorems the programs they proposed would be able to prove.  See
20600	(Davis  and  Putnam  1958).    In  fact, the performance of the first
20700	programs was extremely disappointing, and after ten years  of  effort
20800	programs  are  just on the verge of being useful in a few very formal
20900	and unintuitive branches of mathematics. Even the present programs do
21000	not  use  examples well to guide the formal reasoning and do not have
21100	mechanisms that allow the methods of reasoning to depend properly  on
21200	the nature of what is to be proved.
21300	
21400		PATTERN RECOGNITION, PATTERN DESCRIPTION AND PATTERN MATCHING
21500	
21600		Another  problem that was tackled soon after the availability
21700	of computers was pattern recognition.  As usually posed in the 1950s,
21800	the  problem  was  to  choose which of a number of classes a stimulus
21900	belongs to.   Classifying dot arrays as representing letters  of  the
22000	alphabet  was  much  studied.     This  problem  lent  itself well to
22100	mathematical analysis, because one could imagine that the image scene
22200	represented  and ideal signal to which noise had been added.    Other
22300	approaches represented the signals as points in  a  many  dimensional
22400	space and the signals in a class as a point set in this space, and it
22500	was hoped that the different classes could be  separated  by  passing
22600	hyperplanes  between them or in some other mathematically simple way.
22700	Algorithms for learning how to classify signals were also susceptible
22800	to  mathematical  treatment.      Initially,  some  people  were very
22900	optimistic about these methods and even  proposed  applying  them  to
23000	situations  where  heuristic  programming  was being applied, but the
23100	long run success was  quite  limited.     In  the  first  place,  the
23200	mathematical  problems  were  applicable only to simple problems, but
23300	even more seriously it has turned  out  that  classification  is  the
23400	wrong  problem;  a  robot  or a person that has to look at parts on a
23500	table and assemble a given structure out  of  suitable  parts  cannot
23600	succeed if its pattern handling ability is limited to classifying the
23700	scene as a whole  into  a  finite  number  of  categories.   Instead,
23800	programs  have had to be written that describe the scene as a list of
23900	objects and describe each  object  as  made  up  of  parts.   Objects
24000	bounded  by  flat  faces  can be described, at least in principle, by
24100	present programs, but we still have no satisfactory ideas on  how  to
24200	describe irregular objects such as plants or heads with hair on them.
24300	Clearly, it is unreasonable to describe each hair, but what  kind  of
24400	approximate description should be used?
24500	
24600	
24700		It seems to me that the psychological work of  Jerome  Bruner
24800	on concepts described in his book "A Study of Thinking" is similar in
24900	spirit to that of the pattern recognizers and is subject to  some  of
25000	the  same  criticisms.   Bruner  considered a concept to be a Boolean
25100	combination of elementary concepts  and  studied  experimentally  the
25200	strategies  used  by  human  subjects  in  inducing  a  concept  when
25300	presented with a sequence of  stimuli  and  told  whether  they  were
25400	examples of the unknown concept. As a counter-example to this concept
25500	of concept consider the following description of  a  class  of  plane
25600	figures:  it  consists  of four line segments, one vertical and three
25700	horizontal, the left ends of the horizontals coinciding with the top,
25800	middle,  and  bottom  of the vertical. the top and bottom horizontals
25900	are the same length but shorter than the  vertical,  and  the  middle
26000	horizontal  is  no  longer  than  the  other  two.    We  have here a
26100	description  of  the  letter  E.     However,  the  concept  is   not
26200	representable as a Boolean combination of properties of the figure as
26300	a whole.    It  is,  however,  quite  naturally  representable  by  a
26400	predicate calculus formula such as
26500	
26600		H(x)   ∧   H(y)   ∧   H(z)   ∧   V(w)   ∧   left(x)=top(w)  ∧
26700	left(y)=middle(w) ∧  left(z)=bottom(w)  ∧  length(y)  ≤  length(x)  =
26800	length(z) < length(w)
26900	
27000	where it must be stated in some added formalism that an E consists of
27100	a quadruplet (x,y,z,w) where  x,  y,  z,  and  w  satisfy  the  above
27200	formula.
27300	
27400		More  recently,  ideas of pattern matching originating in the
27500	formal linguistics have been applied to pattern description.   It  is
27600	too  early  to say how successful this will be, but some extension of
27700	the ideas to take into account partial  occlusion  of  an  object  by
27800	others seems to be required.
27900	
28000	
28100	LEARNING
28200	
28300		Considerable effort in AI has been put into  making  programs
28400	that learn from their experience.  Samuel's checker program contained
28500	two forms of learning.   The first was rote  learning  in  which  the
28600	values assigned to positions by look-ahead were stored so that if the
28700	program encountered the position anywhere  on  the  move  tree  in  a
28800	subsequent  game, the value could be used without further look-ahead.
28900	The second form of learning is of the best values of the  parameters
29000	that  determine  how  positions are evaluated.  This is done by going
29100	through book games and adjusting the parameters so that they  propose
29200	the  moves  that  the  experts  consider  good  as often as possible.
29300	Similar learning schemes have been used by others.
29400	
29500		The big problem in  learning  is  generality.  Thus  Samuel's
29600	checker  program is a sucker for the following stratagem provided you
29700	can make the situation come up in the game.  Suppose there are  white
29800	men at squares 20 and 28 and a black king at 19. Suppose further that
29900	black has one more man than white on the rest of the board.   If  the
30000	configuration  on  squares 19, 20, and 28 is allowed to persist while
30100	the remaining men on each side become  kings,  then  black  will  win
30200	because  his  king holds two white men leaving him with a majority on
30300	the rest of the board which will enable him to  exchange  white  down
30400	and  win.   Therefore,  white has to use his forces to drive away the
30500	king on 19 and release its men to become  kings.    Samuel's  program
30600	may  happen  to  do this, because it wants to advance its men, but it
30700	will be just as happy  to  advance  other  men  if  this  is  easier.
30800	Look-ahead  doesn't  help  because  the  ultimate disaster is too far
30900	away.
31000	
31100		Not only doesn't the program not know about  this  stratagem,
31200	but  it  can't  even  be  told,  because no setting of its parameters
31300	corresponds to this knowledge.  Of course, the program can easily  be
31400	modified  to  use  this  stratagem, but this is like educating humans
31500	through brain surgery.
31600	
31700		This suggests  trying  to  divide  the  problem  of  computer
31800	learning  into  two  parts.   First comes the problem of representing
31900	information in the computer in such a way that small changes  in  the
32000	behavior   of   the  program  correspond  to  small  changes  in  the
32100	representation.   It  is  also  desirable  that  these   changes   in
32200	representation  be  expressable  as  additions  without  requiring  a
32300	knowledge of the complete structure of the program, for certainly one
32400	human  can  teach  another  checker  stratagems  without  knowing the
32500	structure of his brain.  Only after we had a teachable program  would
32600	we try to make one that could learn for itself.
32700	
32800		These  ideas have led to a project called the Advice Taker of
32900	making a program that could be told anything that can be expressed in
33000	ordinary  language.   This  leads  to  the  epistemological  problems
33100	discussed in the next section.
33200	
33300	
33400	EPISTEMOLOGY
33500	
33600		The preceding discussion  has  concerned  programs  that  use
33700	particular  intellectual  mechanisms  to  solve particular classes of
33800	problems.   What must we do to  give  a  machine  the  same  flexible
33900	ability  to attack a wide class of problems that a human has.   There
34000	is not general agreement about this among AI researchers, so I  shall
34100	give  my  own  views  and their consequences for the study of natural
34200	intelligence.
34300	
34400		Our first step is to try to break the problem of AI into  two
34500	parts,  the  epistemological  part  and  the  heuristic  part.    The
34600	epistemological part is the problem of how to represent in the memory
34700	of  the  computer what the program "believes" about the present state
34800	of the world, its "laws of motion", and the facts that determine  the
34900	consequences of the actions it may take. The heuristic part of the AI
35000	problem starts from such a representation and is concerned  with  the
35100	procedures  that decide what to do to achieve a goal given the facts.
35200	Most of the  work  in  artificial  intelligence  including  the  work
35300	described  in  the  previous  parts  of this paper have concerned the
35400	heuristic problem.
35500	
35600		The epistemological problem for AI is presently  in  a  state
35700	where it gets more complicated the more we look at it.  Here are some
35800	of the considerations:
35900	
36000	
36100		1.  Laplace proposed to represent the state of the  world  by
36200	specifying  the  positions  and  velocities  of all the particles and
36300	offered to predict the future given the force law.  Modern physics is
36400	similar;   given   the  wave  function,  it  offers  to  predict  the
36500	probabilities of the various future states.  However, even if Laplace
36600	were  correct  and  we  knew  the  force  law  we  could  not use his
36700	representation to describe, say, the situation  in  a  room  full  of
36800	people.   In  the first place, our information about the situation is
36900	not given in terms of the positions and velocities of particles, and,
37000	even  if  it  were,  we  couldn't  do the calculations to predict the
37100	future. Therefore, we must ask ourselves what information is actually
37200	available  about  situations  and what laws allow us to determine the
37300	effects of our actions.  This information must  include  descriptions
37400	and  physical locations of objects including animate objects, beliefs
37500	and purposes of humans and other  animate  objects,  and  qualitative
37600	forms  of  the  laws  of  motion such as the fact that the chalk will
37700	break if I throw it on the floor.  Included also is  the  information
37800	that  permits me to draw conclusions about partially occluded objects
37900	such as the facts that enable me to infer something about the back of
38000	a person's head seeing the front of it.
38100	
38200		2.  The  main  tool  that has been used so far for expressing
38300	this  kind  of  information  formally  has  been  mathematical  logic
38400	especially  predicate  calculus.  A number of programs such as (Green
38500	1969) have been used for  solving  various  problems  expressing  the
38600	particular   data  in  the  form  of  predicate  calculus  sentences.
38700	(McCarthy and Hayes 1969)  contains  a  discussion  of  some  of  the
38800	problems that arise and proposes methods for their solution.
38900	
39000		3.  The  use  of  mathematical  logic for expressing ordinary
39100	reasoning encounters a number of difficulties one of which is  called
39200	the  qualification  problem.  Namely, all logical systems constructed
39300	so far have the property that if a sentence p follows from a set A of
39400	sentences,  then p also follows from any set B of sentences such that
39500	A ⊂ B.  This does not seem to be so in ordinary language.  Suppose  a
39600	person  wants  to  go to Fresno from San Francisco and after thinking
39700	about the circumstances he has described, we conclude that  he  ought
39800	to  take  a  certain  flight.   Then he tells us that he is afraid to
39900	travel by airplane.   Thus, the conclusion that he  ought  to  go  by
40000	airplane  follows from the facts originally stated but not from these
40100	facts supplemented by an additional fact.  In fact,  human  processes
40200	of reasoning allow jumping  to  certain  conclusions  provided  these
40300	conclusions do not contradict known facts.  Efforts are being made to
40400	make computer programs reason in  such  ways.   In  particular,  Carl
40500	Hewitt's  (1971)  programming language PLANNER makes it convenient to
40600	write programs that will assume something unless they can  prove  the
40700	contrary when this is desirable.
40800	
40900	
41000	CONCLUSIONS
41100	
41200		Experience with trying to make computer programs that  behave
41300	intelligently has led me to the following conclusions which may be of
41400	interest to students of natural intelligence.
41500	
41600		1.  Many intellectual mechanisms are required by the problems
41700	they  are used to solve.  Whether the mechanism is reached by genetic
41800	selection, learning in the environment, or by conscious reasoning, it
41900	will be the same.
42000	
42100		2.  Before  an  organism  or  a machine can learn a fact or a
42200	behavior it must have a mechanism whereby this fact or  behavior  can
42300	be represented in its memory structure.
42400	
42500		3. Likewise, it is worthwhile to distinguish what an organism
42600	knows from what it  can  say.     I  liked  Dr.    Premack's  careful
42700	investigation  to determine that the chimpanzee knows about something
42800	before trying to teach it to communicate about it.
42900	
43000		4.  To the extent  that  we  can  identify  the  mind  as  an
43100	information  structure we can study it independently of the hardware.
43200	This is fortunate since anatomy and physiology have not been able  to
43300	give the student of intelligence.
43400	
43500		5. The question of whether an idea for a mechanism of solving
43600	a class of problems is  well  defined  or  will  work  is  often  not
43700	obvious.    Attempting  to  write  a  computer  program to embody the
43800	mechanism will often decide the matter.
43900	
44000		6. I don't know to what extent behaviorism is  still  a  live
44100	issue of view in psychology, but it has never been a live possibility
44200	in artificial intelligence.    Computers  have  no  interesting  pure
44300	stimulus-response  properties  above  the level of, "if you press the
44400	button labelled ' START' on the  printer,  a  yellow  light  goes  on
44500	labelled  'PRINTER  ON'".   All  reasoning about intelligent behavior
44600	involves states of the machine that are not directly observable.
44700	
44800		7. Many of the most  important  intellectual  mechanisms  are
44900	observable  by  analyzing  how  people  perform  complex  tasks.  The
45000	easiest technique is to observe oneself.  Many errors will be  caught
45100	by  programming  a  computer  to  carry  out the task in the proposed
45200	manner.
45300	
45400	
45500	REFERENCES
45600	Brudno
45700	Bruner, J.S., Goodnow, J.J. and Austin, C.A. (1956) A Study of
45800		Thinking. New York: John Wiley.
45900	Davis,  M.  and  Putnam,  H.  (1960)  A Computing Procedure for
46000		Quantification Theory. J. ACM, 7, 201.
46100	Gelernter, H.(1959) Realization of  a Geometry Theorem Proving Machine.
46200		Proc. ICIP. Paris: UNESCO House.
46300	Greenblatt,R.D.,  Eastlake,  D.E. and Crocker, S.D. (1967) The Greenblatt
46400		Chess Program. Proc.  FJCC.  Washington:  Thompson  Book  Co.
46500	Hart, T.P. and Edwards, D.dJ. (1961) The Tree Prune (TP) Algorithm.
46600		M.I.T.  AI  Project:  Memo  No.  30.
46700	Hewitt, Carl (1967) Description and Theoretical Analysis (using Schemas)
46800		of Planner: A Language for Proving Theorems and Manipulating
46900		Models in a Robot. Ph.D. Thesis, Massachusetts Institute of
47000		Technology. McCarthy, J. (1959) Programs with  Common  Sense.
47100	McCarthy, J. (1959) Programs with Common Sense. Mehanisation of Thought
47200		Processes, Volume I. London:HMSO.
47300	McCarthy, J. and Hayes, P. (1969) Philosophical Problems from the
47400		Standpoint of Artificial Intelligence.  Machine  Intelligence
47500	        4, pp. 463-502 (eds. Meltzer, B.  and  Michie,  D.).  Edinburgh:
47600		Edinburgh University Press.
47700	Newell
47800	Nilsson, N.J. (1971) Problem-Solving Methods in Artificial
47900		Intelligence.  New York: McGraw-Hill.
48000	Russell, S.R. (1963) Improvements in LISP Debugging.  Stanford Artificial
48100		Intelligence Project: Memo No. 10.
48200	Russell, R. (1964) Kalah - The Game and the Program. Stanford Artificial
48300		Intelligence   Project:  Memo  No.  22.
48400	Russell, R. (1964) Improvements to the Kalah Program. Stanford Artificial
48500		Intelligence Project: Memo No. 23.
48600	Russell, R. (19  ) Personal Communication.
48700	Ryder, J.L. (1971) Heuristic Analysis of  Large  Trees as Generated in the
48800		Game of Go. Ph.D. Thesis, Stanford University.
48900	Samuel, A.L.
49000	Shannon, C. (1950) Programming a Digital Computer for Playing Chess.
49100		Philosophy Magazine. 41, 356-375.
49200	Simon
49300	Slagle, J.R., ed. (1971) Artificial Intelligence: The Heuristic Programming
49400		Approach.  New  York:  McGraw-Hill.
49500	Turing, A.M. (1950) Computing Machinery and Intelligence.  Mind. 59, 433-
49600		460.