00100	       POSSIBLE FORMS OF INTELLIGENCE - NATURAL AND ARTIFICIAL
00200	
00300	                by John McCarthy, Stanford University
00400	
00500	
00600		The likelihood and nature of interstellar communication would
00700	seem  to depend on the intellectual nature of the beings with whom we
00800	might communicate.   In the  study  of  artificial  intelligence,  we
00900	attempt  to  study intellectual mechanisms as independently as we can
01000	of the particular  ways  intellectual  activity  is  carried  out  by
01100	humans.   Therefore, as a specialist in artificial intelligence, I am
01200	glad to try to see what light that study can shed on the  problem  of
01300	intelligence in the universe.
01400	
01500		I  shall  begin  by  summarizing  the  history  of artificial
01600	intelligence, describing its present state, and  presenting  some  of
01700	the problems that currently baffle us.
01800	
01900		Artificial  intelligence - the study of how to make computers
02000	carry out activities presenting intellectual difficulties to humans -
02100	really  began  with the advent of the stored program digital computer
02200	in 1949.
02300	
02400		One of the earliest undertakings was to  make  the  computers
02500	play  games.   I want to mention the current levels of achievement in
02600	making computers play four different games.  The  games  in  question
02700	are  all  all  board  games,  they  are all played by players playing
02800	alternately and yet, as far as computer programs go,  we  have  quite
02900	different  levels of achievement.  The first game is called kalah.  A
03000	computer program written with a fair amount of effort over  a  couple
03100	of  years plays kalah better than any human players, at least that we
03200	have been able to find.  In fact we`ve learned to  play  the  game  a
03300	good  deal better from watching the computer play.  We were also able
03400	to solve the game in one of its common variants and prove that it was
03500	a win for the first player which had not been previously believed.
03600	
03700		The  leading  characteristic of kalah which makes it possible
03800	to write a better computer program than human players is that certain
03900	aspects  of intelligence seem to be irrelevant.  The position changes
04000	rapidly; there are no apparent strategic characteristics of positions
04100	that hold for a long time; there doesn't seem to be very much pattern
04200	recognition.  What seems important are minor strategems.  You want to
04300	capture  a  few  of the opponent's stones, so you figure out a little
04400	bit of--if I do this and he does that and I do this and he does that,
04500	aha,  I  win,  or  if he does that at this point I lose and so forth,
04600	i.e. you follow the move tree a little bit.  That`s how human  beings
04700	play  the game, and the machine is a lot better at tracing move trees
04800	than a human.  It should be mentioned, however, that the first  kalah
04900	programs  did  not  play  as well as people.  Two improvements in the
05000	early  programs  made  the  difference.   The  first  improvement  is
05100	somewhat  specific  to  kalah  and  consists of a variety of ways for
05200	determing that the game is over,  one  player  having  an  unbeatable
05300	advantage.   This  saves much branching at the ends of the move tree.
05400	The second--called the α-β  heuristic--applies  to  all  games  where
05500	players   move   alternately.    It   involves  not  examining  moves
05600	alternative to refuting moves.  (A refuting move is  one  that  shows
05700	that the opponent's move leading to the position is worse than one of
05800	his already examined moves.)
05900	
06000		The α-β heuristic is used by all human players, but  was  not
06100	identified   as  necessary  by  the  first  writers  of  game-playing
06200	programs.  This  illustrates  two  facts:  First,   an   intellectual
06300	mechanism  may  be  compulsory  for the problems; I cannot imagine an
06400	effective general game player--human, alien, or machine--that did not
06500	use α-β.  Second, many of our difficulties in artificial intelligence
06600	are a failure to recognize  mechanisms  that  will  be  obvious  once
06700	pointed out.
06800	
06900		The  next  level  of  performance  is  a more difficult game,
07000	checkers.  The checker program,  written  by  Arthur  Samuel  uses  a
07100	similar strategy to that used in the kalah program, and it also plays
07200	quite well.  It can beat most ordinary players and it played with the
07300	U.S. Champion and got one draw in 6 games; the other five he beat it.
07400	Since then the checker experts have learned a little bit  more  about
07500	how  the  program  plays  checkers and they beat it almost every time
07600	now.
07700	
07800		The game into which the most effort has been  put  is  chess,
07900	and  the best currrent chess program is Greenblatt's at MIT; this one
08000	plays class C chess, and he thinks he can get it up to playing  class
08100	B  chess  by  an  extension of his present methods. There are several
08200	rivals now: people who think they have  computer  programs  that  can
08300	play  as  well  as  Greenblatt's,  and  some  of them have proposed a
08400	computer chess tournnament.   Greenblatt  is,  like  all  good  chess
08500	players, coy about whether he will participate.  He wants to play his
08600	program against human players mainly, and he says  they  could  enter
08700	the  U.S.  Open  Tournament too. (When chess players say "open", they
08800	mean it; computer programs have been admitted provided they obey  all
08900	the rules including the time rules).
09000	
09100		In the chess programs, there is a  lot  more  going  on  than
09200	simple  tree  search,  which yields a very bad program even with α-β.
09300	Much of it is specific to chess: techniques  for  recognizing  threat
09400	situations,  techniques  for evaluating positions, and techniques for
09500	determining the effects of complicated exchanges of pieces.  Of  more
09600	general  significance  is  comparing the potential of a move with the
09700	requirements of a position; e.g.   if  in  a  certain  variant  being
09800	considered,  one  is down a bishop, one should not engage in plots to
09900	win a pawn.
10000	
10100		Jonathan  Ryder,  a  graduate  student at Stanford, who is an
10200	expert Go player, is writing a program to play Go.   It  still  plays
10300	extremely  badly, and this is because intellectual mechanisms that we
10400	understand and know how to make a computer carry out and  which  work
10500	fairly  well  in chess are very weak in Go.  The problem in Go is the
10600	large number of moves.  There  are  361  possible  first  moves,  360
10700	replies,  and 359 replies to that and so forth for a while, because a
10800	move consists of putting a stone on a 19x19  board.   Therefore,  the
10900	method  used in chess, following out the move tree, immediately fails
11000	in Go because the tree will be 361x360x359 etc. moves in size, and Go
11100	players  sometimes  look a fair distance ahead.  We examine this, and
11200	ask what is wrong; a human player certainly doesn`t look at all those
11300	alternatives.   Then  we  discover  that the human player divides the
11400	board up into regions on the basis of the stones that are already  on
11500	the board and he thinks about the regions separately.  "Is this group
11600	of stones safe or can it be captured by the opponent?"  In  order  to
11700	figure  this  out,  he does a local analysis.  He goes through a move
11800	tree alright, but the move tree only contains moves in the  immediate
11900	area  and only those moves which are suggested by certain principles.
12000	He may come to a conclusion that the group is safe provided he moves,
12100	but  if he doesn`t move, it will be captured and he will lose so many
12200	points.  He will  remember  this  while  he  thinks  about  what  his
12300	opportunities  are on the rest of the board and then after a while he
12400	will  begin  to  think  about  interactions   between   these   local
12500	situations.  He will note that a move at a certain place affects both
12600	situations and so forth.  As yet we do not fully  understand  how  to
12700	program  the  recognition  of  local  situations.   In  chess,  local
12800	situations also exist, and maybe we can`t get much beyond  the  level
12900	that  Greenblatt's  program  currently  achieves, without taking this
13000	into account, but in Go  their  recognition  is  essential  for  even
13100	moderately good play.  Ryder has had some success in this direction.
13200	
13300		Another  area  is  proving mathematical theorems by computer.
13400	At first this work was done in a large number of different formalisms
13500	but  now most of it is done in Robinson's resolution formalism of the
13600	predicate calculus.  There are quite  a  number  of  theorem  proving
13700	programs  that  work  by  resolution, and quite a number of different
13800	kinds of problems have been formalized in predicate calculus.
13900	
14000		The most spectacular result of computer theorem  proving  was
14100	the  solution  of  a  known  unsolved  problem  in lattice theory.  A
14200	mathematician looking at  computer  output  noted  that  one  of  the
14300	computer-proved  formulas implied the resolution of the conjecture. I
14400	don't know if the program could have proved the conjecture  directly.
14500	Moreover,  lattice theory seems to be particularly suited to computer
14600	theorem  provers  because  much  of  it  is   formula   manipulation;
14700	complicated  structures  of  lemmas  and  the use of examples is less
14800	essential than in other branches of mathematics.
14900	
15000		Much work has also been done in computer interaction with the
15100	real  world.   For example, a computer is equipped with an artificial
15200	arm and artificial eye, the artificial eye being a television camera.
15300	Programs  can  be  written  to  assemble  objects  out  of parts, for
15400	example.  It turns out that the most difficult problems for this  are
15500	in  the  area  of  vision. This involves going from a TV image in the
15600	computer (in our case a 256x333 array of 4 bit numbers) to a list  of
15700	the  objects in the scene with their positions and attitudes, so that
15800	the programs that control the artificial arm will  be  able  to  know
15900	where to reach out and grasp and pick something up and move it.
15950	
16000		At present there are several approaches,  all  of  which  are
16100	clearly  quite  limited  even  in  their potential accomplishment let
16200	alone in their present accomplishment. One approach  involves  scenes
16300	that  are  composed  of  objects  with  flat  faces.   Thus there are
16400	definite edges between the faces and there exist  programs  that  can
16500	find  these edges and even distinguish edges of objects from edges of
16600	shadows, which is not so trivial since sometimes the shadows are more
16700	prominent  than  the  edges  of  the  objects.   In  any case, block-
16800	stacking is possible; you can pick up a  batch  of  blocks  and  make
16900	towers  out of them, and we hope to be able to do some more ambitious
17000	things shortly.
17050	
17100		Another   approach   being   followed  at  Stanford  Research
17200	Institute involves  recognizing  objects  in  a  different  way:  The
17300	program  divides  the  region  into a large number of subregions, say
17400	100x100, it characterizes each subregion, and then it joins  together
17500	adjacent subregions that have the same characterization: for example,
17600	about the same color and the same shade if color is something  you're
17700	using  or  the  same shade, the same brightness in any case.  Thus it
17800	builds larger regions and finds the boundaries of the  regions  group
17900	these  regions  together to make objects.  This has some potentiality
18000	for dealing with curved objects.  Still other possibilities are being
18100	pursued.
18150	
18200		I think this is enough about the concrete accomplishments  of
18300	artificial intelligence work.
18400	
18500	
18600		We  can  complement  the   above   look   at   the   concrete
18700	accomplishments  of artificial intelligence research by taking a look
18800	at the current research problems.  In my opinion, there are two  main
18900	classes of problems in AI. The first is the discovery or invention of
19000	intellectual  mechanisms  and  their   implementation   by   computer
19100	programs.   (Mostly the mechanisms are discovered in human behavior -
19200	more often by introspection than by formal psychological  experiment,
19300	but sometimes mechanisms are invented that have no obvous counterpart
19400	in human or animal behavior.  α-β  is  an  example  of  a  discovered
19500	mechanism,  and  associative  memories based on hash-addressing is an
19600	example of an invented one).  Mechanisms are often found  by  looking
19700	into  the  reasons  for  the  disappointing  performance of a program
19800	containing mechanisms previously thought adequate to solve a  certain
19900	class  of problems. Often there is no glory in this work, because the
20000	new mechanisms,  once  found,  are  considered  obvious.   Particular
20100	intellectual  mechanisms are often called heuristics, and their study
20200	is called heuristic programming.
20300	
20400		The second class of problems is more basic, but this fact has
20500	only  recently  become  apparent  to  very many workers in the field.
20600	These are the  problems  of  what  an  intelligent  being,  human  or
20700	machine,  can know about the world and how this information should be
20800	represented in the memory of the computer.  In the early game playing
20900	and  theorem  proving programs, it was possible for the programmer to
21000	devise an ad hoc representation  of  what  was  believed  to  be  all
21100	relevant information. Present programs for game playing are all based
21200	on these ad hoc representations  of  positions.    Any  strategic  or
21300	tactical concepts are represented by features of the program.
21400	
21500		When  we  want to design a program that has anything like the
21600	human generality in reacting to real-world  situations  that  include
21700	observing   the   physical  world,  receiving  information  about  it
21800	expressed in a natural language, deciding whether enough  information
21900	is  available  to take successful action in a given situtation and if
22000	not deciding how to get more - then, the ad hoc  representations  are
22100	inadequate.   Then  we  must  equip our program with some ideas about
22200	what the world is like in general (metaphysics) and some ideas  about
22300	what   knowledge   is   available   and  how  more  can  be  obtained
22400	(epistemology).  The  problems  are  those  of  the   above-mentioned
22500	branches  of  philosophy,  but  when we look at what the philosophers
22600	have done we are disappointed. Almost all of what they have  proposed
22700	is  too vague; e.g. we cannot program a computer to look at the world
22800	in the way recommended by (say) Wittgenstein.  Also  much  philosophy
22900	seems  clearly wrong; the recommended ways of getting information (to
23000	the extent that they are precise) just won't  work.   The  positivist
23100	philosophers  seem  to have thrown out real problems in their efforts
23200	to clean out meaningless ideas.
23300	
23400		Recently, there as  been  a  beginning  in  making  a  formal
23500	language capable of expressing what human beings know and robots need
23600	to know about real world  situations  in  order  to  take  successful
23700	action.   This   has   involved   expressing  in  first  order  logic
23800	descriptions of situations and the effects of taking actions in them.
23900	The  desire  has  been  to formalize the situation well enough so the
24000	fact that a certain strategy is appropriate to  realizing  a  certain
24100	goal  is  a  logical consequence of the description of the particular
24200	situation and general information  about  the  effects  of  different
24300	kinds  of  actions.  Moderate  success  has  been  achieved,  and the
24400	research continues.  We hope the results will be of philosophical  as
24500	well  as   artificial  intelligence  interest.   Within  the last two
24600	years, a new formalism has been developed by  Hewitt,  Winograd,  and
24700	others  at  the  M.I.T.  Artificial  Intelligence Laboratory based on
24800	Hewitt's  language   PLANNER.    This   formalism   represents   much
24900	information  as  procedures  and  has  been particularly effective in
25000	translating information originally expressed in natural language into
25100	computer   form.    It  also  gives  a  new  approach  to  expressing
25200	generalizations  that  have  exceptions  not  all  of  which  can  be
25300	presented along with the original generalization.
25400	
25500		The  epistemological problem is also very acute in connection
25600	with programs that physically manipulate the world on  the  basis  of
25700	visual information.
25750	
25800		Now I would like to make some connections  with  interstellar
25900	communication.   The  first  conclusion  is  that  the  mechanisms of
26000	intelligence are mostly objective and are not  dependent  on  whether
26100	it's  a human being or a machine or a Martian doing the thinking.  If
26200	you are going to play chess well, then you have to carry out  certain
26300	processes,  processes  of  search, processes of factoring a situation
26400	into its subparts, and it would seem that these  are  independent  of
26500	who  you  are.  If  you  want to discover physics for yourself on the
26600	basis of experiment, again the intellectual procedures that you  have
26700	to go through seem to be determined by the nature of the problem to a
26800	large extent, rather than by who you are.  This suggests to  me  that
26900	we  should  expect  to find other intelligences in the universe using
27000	similar procedures to the ones which we use and which we  would  like
27100	to  program our machines to use.  Of course, they may think slower or
27200	faster than we  do, and  they  may  have  pursued  certain  areas  of
27300	knowledge to a lesser or greater extent than we have.
27400		There  are  much  greater  possibilities  for  difference  in
27500	motivational  structure. Most computer programs are not properly said
27600	to have motivational structures; they just run.  However, many of our
27700	attempts  to  make  intelligent programs involve interpreting certain
27800	expressions in the computer memory as goals.   The  program  compares
27900	the  present  situation with the goal which suggests possible actions
28000	the predicted results of which  are  also  compared  with  the  goal.
28100	Subgoals  are  generated  subordinate  to the main goal and these are
28200	pursued.  In my opinion, such programs may properly be said  to  have
28300	motivational  structures,  and  we  may  also  try to interpret human
28400	behavior in terms of formal goal structures.  However,  the  programs
28500	we  have written so far and even programs that we contemplate writing
28600	in  the  future  to  achieve  our  purposes  seem  to  have   simpler
28700	motivational structures than have human beings.
28800	
28900		Human  beings often change their goals in stronger sense than
29000	programs change subgoals in pursuit of a main goal.  In fact, a human
29100	does  not  have  a  main  goal  in the sense of some function that he
29200	conducts his life so as to  optimize.  We  often  find  ourselves  in
29300	states  of  relative  goallessness,  where  what we want is to find a
29400	goal.
29450	
29500		To  take  another example, a dog's motivational structure can
29600	be described approximately in the following way: sometimes the dog is
29700	hungry  and wants to eat, sometimes he is thirsty and wants to drink,
29800	sometimes he is driven by sex.  When none of these things  apply,  he
29900	is  content to lie down and rest.  One could imagine that  some  very
30000	intelligent organism could have a similar motivational  structure  to
30100	the  dog,  it solves problems as they present themselves but does not
30200	do much when no problem presents itself.  On the  other  hand,  other
30300	aliens  might  be  motivated  by curiousity. They want to find out as
30400	much about the universe as a whole as possible, which they  would  do
30500	by  pursuing  science and also by exploration.  We can imagine beings
30600	motivated by a drive to expand; they might want to convert as much of
30700	the universe as possible into their own substance.
30750	
30800		It seems to me that all of these are possible stable forms of
30900	motivational  structure  for an intelligent being, though it's not so
31000	easy to see which of them could have evolved.
31050	
31100		With regard to  social  organization,  there  are  also  many
31200	possibilities.    Thus,  we tend to presume civilizations composed of
31300	many independent beings with distinct  individual  goals  interacting
31400	with  each other, but this is not enivitable. Science fiction writers
31500	have imagined a single intelligcnce that had  incorporated  more  and
31600	more material of its original planet into its structure and which has
31700	goals of expansion or curiousity.
31850	
31900		Some  ideas  about  the communicativeness of extra-terrestial
32000	intelligences can be obtained by speculating about  our  own  future.
32100	Of  course many people are trying to plan our future for the next 100
32200	years or even 1000, but it seems  to  me  that  these  plans  have  a
32300	certain  arrogance.  Suppose  we  look back and ask what attention we
32400	today pay to plans made for us, especially in the  technological  and
32500	economic area, by people 100 years ago.  It turns out that we're  not
32600	even very curious as to what they thought the future 100 years  hence
32700	would  be  like,  and  I think that our descendants will be similarly
32800	uninterested in what we predict and plan for them.  I think  planning
32900	should  be limited to one doubling time of our technology in the area
33000	in question.  However, speculations  about  our  own  future  may  be
33100	relevant  to  other  technologies  that we might encounter.   Some of
33200	these questions concern  artificial  intelligence  itself.    Namely,
33300	will  we  develop computer programs that are more intelligent than we
33400	are?
33450	
33500		I would say very likely yes, unless we decide not to for some
33600	reason.  I won't promise you a date, because there are, as I see  it,
33700	some  big  problems  that  we  don't have any idea how to solve.  The
33800	situation may even be  as  primitive  as  the  space  program  before
33900	Newton.     Nevertheless,  there  don't  seem  imitations  on machine
34000	intelligence short of human intelligence, and if we can get the human
34100	level of intelligence we can get much more intellectual work than our
34200	whole race does just by using faster computers and  larger  memories.
34300	The  question  is how will we use this artificial intelligence.  Will
34400	we limit it, will we in some sense merge with it as has been proposed
34500	in  some  science  fiction stories in which a human being adds to his
34600	own intellectual and physical powers.  Just as our cells are replaced
34700	every  7  years, after some time of interaction with these additional
34800	powers, the intelligence would  have  to  be  described  as  residing
34900	mainly  in  the  artifact  and only slightly in the original carcass.
35000	Well, I don't know; it's very difficult to predict what we  will  do.
35100	However, imagine that we either directly or through machines acquired
35200	some much higher level of intelligence: then we would have to ask the
35300	following  questions:   Is the universe still interesting to us; i.e.
35400	does it still have structure yet to be discovered or will we discover
35500	the  fundamental structure of the universe once and for all?  Another
35600	question concerns exploration.  Is it true that when you've seen  one
35700	galaxy you've seen them all?  Are all high intelligences alike or are
35800	they sufficiently diferent to be interesting to each  other?    In  a
35900	sense,  two  computers  are  not interesting to each other--they have
36000	very little to say to  each  other  unless  they  have  substantially
36100	different  data  bases.    Any computer can run the programs that any
36200	other can run  and  they  might  have  perhaps  different  data  sets
36300	associated  with  them,  but  there  isn't the same reason for mutual
36400	cooperation among computers as there is among human beings.
36500	
36600		To   summarize:   artificial   intelligence   is    possible,
36700	intelligent  entities  in  the  universe  may  be  similar  in  their
36800	intellectual properties, because  the  methods  of  intelligence  are
36900	determined  by the problems. However, intelligent entities may differ
37000	considerably in their motivational structures.  The possibilities for
37100	interstellar  communication  depend  partly  on  this  and  partly on
37200	presently undecideable objective  questions  that  determine  whether
37300	intelligent entities will have much of mutual interest to say to each
37400	other.
37500	
37600	
37700	APPENDIX -- SOME CONSIDERATIONS OF INTERSTELLAR TRAVEL
37800	
37900	
38000		In  connection  with  the   consideration   of   interstellar
38100	communication,  we  ought also to consider the extent to which we and
38200	other intelligences might travel or even expand to occupy other solar
38300	systems.  My conclusion is that such travel and expansion is feasible
38400	in times that are long compared with present human lifetimes but very
38500	short  in  comparison  with  the length of time life can exist on our
38600	planet.
38700	
38800		For this purpose, we shall postulate a rocket that  expels  a
38900	working  fluid  using  energy  derived from a nuclear reactor.   This
39000	includes as extreme cases the photon rocket and  the  nuclear  rocket
39100	that expels only the fission products.
39200	
39300		When we attempt to optimize the design of such a rocket so as
39400	to arrive at a given other star  in  the  shortest  time  with  given
39500	initial  and  final  masses,  we  find that we have to compromise two
39600	considerations.  The first consideration is that the mass we expel to
39700	get  our  velocity  must be carried along so that we want to minimize
39800	it.  If we only take this into account, we would want to maximize the
39900	exhaust  velocity,  and the best rocket would appear to be the photon
40000	rocket in which the exhaust velocity is the velocity of light.
40100	
40200		However, we must also take into account  the  power  handling
40300	capability  of  the rocket system.  Here we get the most acceleration
40400	for a given power if the exhaust velocity is as low as possible.  The
40500	power limitation may come from the power that can be generated by the
40600	system or from the ability to dissipate  waste  energy.    From  this
40700	point  of  view,  the  photon  rocket is the worst possible, and some
40800	people have attempted  to  refute  the  possibility  of  interstellar
40900	travel by noting that the photon rocket is the best according to mass
41000	considerations and then pointing out that the dissipation  of  energy
41100	required  is  impossible.   (see  J.  R. Newman, Scientific American,
41200	Vol.210,no.2,  Feb.  1964,  pp.141-146.)  The  best  performance   is
41300	obtained  by  compromising these considerations, and this requires us
41400	to vary the exhaust velocity during the mission.