00100	COMMENT ⊗   VALID 00002 PAGES
00200	C REC  PAGE   DESCRIPTION
00300	C00001 00001
00400	C00002 00002	\\M0BASL30\M1BASI30\M2BASB30\M3MISC30\F0
00500	C00031 ENDMK
00600	C⊗;
     

00100	\\M0BASL30;\M1BASI30;\M2BASB30;\M3MISC30;\F0
00200	\JArtificial  Intelligence: A General  Survey by
00300	Professor  Sir  James Lighthill,  FRS, in  \F1Artificial  Intelligence: a
00400	paper symposium\F0, Science Research Council 1973.
00500	
00600		Professor   Lighthill  of  Cambridge   University  is   a  famous
00700	hydrodynamicist  with a recent  interest in  applications to biology.
00800	His review of  artificial intelligence  was at the  request of  Brian
00900	Flowers, then head of the  Science Research Council of Great  Britain, the
01000	main  funding body  for British  university scientific  research. Its
01100	purpose was to help the Science Research Council decide  requests for
01200	support of  work in  AI.  Lighthill  claims no  previous acquaintance
01300	with the  field,  but refers to a large number of authors whose works
01400	he consulted, though not to any specific papers.
01500	
01600		The \F1Lighthill Report\F0 is organized around a classification
01700	of AI research into three categories:
01800	
01900		Category A is \F1advanced automation\F0 or \F1applications\F0,
02000	and he approves of it in principle.  Included in A are
02100	some activities that are obviously applied but also activities like
02200	computer chess playing that are often done not for themselves but
02300	in order to study the structure of intelligent behavior.
02400	
02500		Category C comprises studies of the \F1central nervous system\F0
02600	including computer modeling in support of both neurophysiology and
02700	psychology.
02800	
02900		Category B is defined as "building robots" and "bridge" between
03000	the other two categories.  Lighthill defines a robot as a program or
03100	device built neither to serve a useful purpose nor to study the central
03200	nervous system, which obviously would exclude Unimates, etc. which are
03300	generally referred to as industrial robots.  Emphasizing the bridge aspect
03400	of the definition, Lighthill states as obvious that work in category B is
03500	worthwhile only in so far as it contributes to the other categories.
03600	
03700		If we take this categorization seriously, then most AI researchers
03800	lose intellectual contact with Lighthill immediately, because his three
03900	categories have no place for what is or should be our main scientific
04000	activity - \F2studying the structure of information and the structure of
04100	problem solving processes independently of applications and independently
04200	of its realization in animals or humans\F0.  This study is based on the
04300	following ideas:
04400	
04500		1. Intellectual activity takes place in a world that has a certain
04600	physical and intellectual structure: Physical objects exist, move about,
04700	are created and destroyed.  Actions that may be performed have effects that
04800	are partially known.  Entities with goals have available to them certain
04900	information about this world.  Some of this information may be built in,
05000	and some arises from observation, from communication, from reasoning, and
05100	by more or less complex processes of retrieval from information bases.
05200	Much of this structure is common to the intellectual position of animals,
05300	people, and machines which we may design, e.g. the effects of physical actions
05400	on material objects and also the information that may be obtained about
05500	these objects by vision.
05600	The general structure of the intellectual world is far from understood, and
05700	it is often quite difficult to decide how to represent effectively the information
05800	available about a quite limited domain of action even when we are quite
05900	willing to treat a particular problem in an \F1ad hoc\F0 way.
06000	
06100		2. The processes of problem solving depend on the class of problems
06200	being solved more than on the solver.  Thus playing chess seems to require
06300	look-ahead whether the apparatus is made of neurons or transistors.
06400	Isolation of the information relevant to a problem from the totality
06500	of previous experience is required whether the solver is man or machine,
06600	and so is the ability to divide a problem into weakly connected subproblems
06700	that can be thought about separately before the results are combined.
06800	
06900		3. Experiment is useful in determining what representations of
07000	information and what problem solving processes are needed to solve a
07100	given class of problems.  We can illustrate this point by an example from
07200	the \F1Lighthill Report\F0 which asserts (p. 15) that the heuristics of a chess
07300	program are embodied in the evaluation function.  This is plausible
07400	and was assumed by the first writers of chess programs.
07500	Experiment showed, however, that the procedures that select what part of the
07600	move tree is examined are even more important, i.e. when the program errs
07700	it is usually because it didn't examine a line of play rather than because
07800	it mis-evaluated a final position.  Modern chess programs concentrate on this
07900	and often have simpler evaluators than the earlier programs.
08000	
08100		4. The experimental domain should be chosen to test the adequacy
08200	of representations of information and of problem solving mechanisms.  Thus
08300	chess has contributed much to the study of tree search; one Soviet computer
08400	scientist refers to chess as the \F1Drosophila\F0 of artificial intelligence.
08500	I think there is much more to be learned from chess, because master level
08600	play will require more than just improving the present methods of searching
08700	trees.  Namely, it will require the ability to identify, represent, and
08800	recognize the patterns of position and play that correspond to "chess ideas",
08900	the ability to solve some abstractions of positions (e.g. how to make use
09000	of a passed pawn and a seventh rank rook jointly) and to apply the result
09100	to actual positions.  It will probably also require the ability to analyze
09200	a problem into subproblems and combine the separate results.  (This ability
09300	is certainly required for a successful \F1Go\F0 program).
09400	
09500		Having ignored the possibility that AI has goals  of its own,
09600	Lighthill goes on  to document his claim that  it has not contributed
09700	to applications or to  psychology and physiology.   He exaggerates  a
09800	bit here,   it seems worthwhile  to spend some effort  disputing his
09900	claims that AI has not contributed to these other subjects.
10000	
10100		In my opinion,   AI's contribution to  practical applications
10200	has  been significant  but so  far mostly  peripheral to  the central
10300	ideas and  problems  of AI.   Thus  the  LISP language  for  symbolic
10400	computing was  developed for  AI use,   but  has had  applications to
10500	symbolic computations  in other areas, e.g.  physics.  Moreover, some
10600	ideas  from  LISP  such  as  conditional  expressions  and  recursive
10700	function definitions  have been used in  other programming languages.
10800	However,  the  ideas that have  been applied elsewhere  don't have  a
10900	specifically AI character  and might have been but  weren't developed
11000	without AI  in mind.  Other examples  include time-sharing, the first
11100	proposals for which had AI motivations and some techniques of picture
11200	processing that were first developed in AI laboratories and have been
11300	used elsewhere.   Even the current  work in automatic  assembly using
11400	vision  might have been developed  without AI in mind.   However, the
11500	Dendral work has always had a specifically AI character,  and many of
11600	the recent developments  in programming such as  PLANNER and CONNIVER
11700	have an AI motivation.
11800	
11900		AI's  contributions to  neurophysiology  have been  small and
12000	mostly of a negative character, i.e. showing that  certain mechanisms
12100	that neurophysiologists propose are not well defined or inadequate to
12200	carry  out the behavior they are supposed to  account for.  I have in
12300	mind Hebb's proposals in his book \F1The  Organization of Behavior\F0.
12400	No-one  today would believe  that the  gaps in  those ideas  could be
12500	filled without adding something much  larger than the original  work.
12600	Moreover, the  last 20  years experience  in programming machines  to
12700	learn  and solve problems  makes it implausible  that cell assemblies
12800	\F1per se\F0  would learn  much without  putting  in some  additional
12900	organization, and  physiologists today  would be unlikely  to propose
13000	such a theory.  However, merely showing that some things are unlikely
13100	to work is not a \F1positive\F0 contribution.
13200	I think there will be more interaction between AI and neurophysiology
13300	as soon as the neurophysiologists are in a position to compare
13400	information processing models of higher level functions with
13500	physiological data.  There is little contact at the nerve cell level,
13600	because, as Minsky showed in his PhD dissertation in 1954, almost any
13700	of the proposed models of the neuron is a universal computing element,
13800	so that there is no connection between the structure of the neuron and
13900	what higher level processes are possible.
14000	
14100		On the other  hand,  the  effects of artificial  intelligence
14200	research  on  psychology have  been  larger  as  attested by  various
14300	psychologists. First of all, psychologists have begun to use models in
14400	which  complex  internal  data structures  that  cannot  be  observed
14500	directly  are attributed to  animals and people.   Psychologists have
14600	come to use these models,  because they exhibit behavior  that cannot
14700	be exhibited by models conforming  to the tenets of behaviorism which
14800	essentially  allows  only connections  between  externally observable
14900	variables.   Information processing  models in  psychology have  also
15000	induced dissatisfaction  with psychoanalytic and  related theories of
15100	emotional behavior.  Namely,  these information processing models  of
15200	emotional states  can  yield predictions  that can  be compared  with
15300	experiment or experience in a more definite way than can the vague
15400	models of psychoanalysis and its offspring.
15500	
15600		Contributions  of AI to  psychology are  further discussed in
15700	the paper  \F1Some Comments  on the  Lighthill Report\F0  by  N.   S.
15800	Sutherland which  was included  in the same  book with  the Lighthill
15900	report itself.
16000	
16100		Systematic  comment on  the main  section,   entitled \F1Past
16200	Disappointments\F0  is  difficult because  of  the  strange  way  the
16300	subject is divided up but here are some remarks:
16400	
16500		1. Automatic  landing systems for airplanes are  offered as a
16600	field in  which conventional  engineering techniques  have been  more
16700	successful than AI  methods.  Indeed, no-one would  advocate applying
16800	the scene analysis or tree search techniques developed in AI research
16900	to automatic landing  in the context in  which automatic landing  has
17000	been developed.  Namely, radio signals are available to determine the
17100	precise  position of  the airplane in  relation to  a straight runway
17200	which is  guaranteed clear  of  interfering objects.   AI  techniques
17300	would  be  necessary  to make  a  system  capable  of landing  on  an
17400	unprepared dirt strip with no radio aids which had to be located  and
17500	distinguished  from roads  visually  and  which  might have  cows  or
17600	potholes or muddy places on it.  The problem of automatically driving
17700	an automobile in an  uncontrolled environment is even more  difficult
17800	and will  definitely require AI  techniques, which, however,  are not
17900	nearly ready for a full solution of such a difficult problem.
18000	
18100		2.  Lighthill  is  disappointed  that  detailed  knowledge of
18200	subject matter has to be put in if programs  are to be successful
18300	in theorem proving, interpreting  mass spectra, and game playing.  He
18400	uses the word \F1heuristics\F0  in a non-standard way  for this.   He
18500	misses the fact that there are great  difficulties in finding ways of
18600	representing knowledge of  the world in computer programs and much AI
18700	research  and internal  controversy are  directed  to  this  problem.
18800	Moreover,  most  AI  researchers  feel that  more  progress  on  this
18900	\F1representation problem\F0 is essential before substantial progress
19000	can be made on the problem of automatic acquisition of knowledge.  Of
19100	course, missing  these particular points is a  consequence of missing
19200	the existence of  the AI  problem as distinct  from automation  and
19300	study of the central nervous system.
19400	
19500		3. A  further disappointment is  that chess  playing programs
19600	have only  reached an "experienced amateur" level  of play.  Well, if
19700	programs can't do better than that  by 1978, I shall lose a \F3B\F0250  bet
19800	and will  be disappointed  too though not  extremely surprised.   The
19900	present  level of  computer chess  is based  on the  incorporation of
20000	certain intellectual  mechanisms in the  programs.  Some  improvement
20100	can be made by further  refinement of the heuristics in the programs,
20200	but probably master  level chess  awaits the ability  to put  general
20300	configuration patterns into the programs in an easy and flexible way.
20400	I don't see how to set a date by which this problem must be solved in
20500	order to avoid disappointment in the field of artificial intelligence
20600	as a whole.
20700	
20800		4. Lighthill discusses the \F1combinatorial explosion\F0
20900	problem as though it were a relatively recent phenomenon that
21000	disappointed hopes that unguided theorem provers would be able to
21100	start from axioms representing knowledge about the world and solve
21200	difficult problems.  In fact, the \F1combinatorial explosion\F0
21300	problem has been recognized in AI from the beginning, and the usual
21400	meaning of \F1heuristic\F0 is a device for reducing this explosion.
21500	Regrettably, some people were briefly over-optimistic about what
21600	general purpose heuristics for theorem proving could do in problem
21700	solving.
21800	
21900	
22000	Did We Deserve It?
22100	
22200		Lighthill had  his shot  at AI and  missed, but  this doesn't
22300	prove  that  everything in  AI  is ok.    In my  opinion,  present AI
22400	research suffers  from some  major deficiencies apart  from the  fact
22500	that  any scientists  would  achieve more  if they  were  smarter and
22600	worked harder.
22700	
22800		1. Much  work in  AI has  the "look  ma, no  hands"  disease.
22900	Someone programs  a computer  to do  something no  computer has  done
23000	before and writes a paper pointing out that the computer did it.  The
23100	paper is not directed to the identification and study of intellectual
23200	mechanisms and often contains no  coherent account of how the program
23300	works  at all.  As an  example, consider  that the  SIGART Newsletter
23400	prints the scores of the  games in the ACM Computer  Chess Tournament
23500	just as though the programs were human players and their innards were
23600	inaccessible.  We need to know why one program missed the right  move
23700	in a position  - what was it thinking  about all that time?   We also
23800	need  an  analysis of  what  class of  positions  the  particular one
23900	belonged to and how a  future program might recognize this class  and
24000	play better.
24100	
24200		2. A second disease is to work only on theories that can be
24300	expressed mathematically in the present state of knowledge.
24400	Mathematicians are often attracted to the artificial intelligence
24500	problem by its intrinsic interest.  Unfortunately for the mathematicians,
24600	however, many plausible mathematical theories with good theorems
24700	such as control theory or statistical decision theory have
24800	turned out to have little relevance to AI.  Even worse, the applicability
24900	of statistical decision theory to discriminating among classes of
25000	signals led to the mistaken identification of perception with
25100	discrimination rather than with description which so far has
25200	not led to much mathematics.
25300	More recently, however, problems of theorem proving and problems of
25400	representation have led to interesting mathematical problems in logic
25500	and mathematical theory of computation.
25600	
25700		3. Every now  and then, some AI scientist gets  an idea for a
25800	general  scheme of  intelligent behavior that  can be  applied to any
25900	problem provided the machine  is given the specific knowledge  that a
26000	human has about  the domain.  Examples of this  have included the GPS
26100	formalism, a simple predicate  calculus formalism, and more  recently
26200	the  PLANNER  formalism  and  perhaps   the  current  Carnegie-Mellon
26300	production formalism.  In the first and third  cases, the belief that
26400	any problem solving  ability and knowledge could  be fitted into  the
26500	formalisms led to published  predictions that computers would achieve
26600	certain  levels  of  performance  in certain  time  scales.    If the
26700	inventors of  the formalisms  had been  right about  them, the  goals
26800	might have  been achieved, but  regrettably they were  mistaken. Such
26900	general purpose formalisms will be  invented from time to time, and,
27000	most  likely, one of them will eventually prove adequate.
27100	However, it would be  a great relief to the rest of the workers in AI
27200	if the  inventors of  new general  formalisms  would express  their
27300	hopes in a more guarded form than has sometimes been the case.
27400	
27500		4. At present,  there does not exist  a comprehensive general
27600	review of  AI that discusses all the main approaches and achievements
27700	and issues.    Most likely,  this  is not  merely because  the  field
27800	doesn't have a  first rate reviewer at present,  but because the field
27900	is confused about what these  approaches and achievements and  issues
28000	are.   The production  of such  a review  will therefore  be a  major
28100	creative work and not merely a work of scholarship.
28200	
28300		5. While it is far beyond the scope of this review to try
28400	to summarize what has been accomplished in AI since Turing's 1950 paper,
28500	here is a five sentence try: Many approaches have been explored and
28600	tentatively rejected including automaton models, random search,
28700	sequence extrapolation, and many others.  Many heuristics have been
28800	developed for reducing various kinds of tree search; some of these are
28900	quite special to particular applications, but others are general.
29000	Much progress has been made in discovering how various kinds of
29100	information can be represented in the memory of a computer, but
29200	a fully general representation is not yet available.  The problem
29300	of perception of speech and vision has been explored and recognition
29400	has been found feasible in many instances.  A beginning has been made
29500	in understanding the semantics of natural language.
29600	
29700		These accomplishments notwithstanding, I think that artificial
29800	intelligence research has so far been only moderately successful;
29900	its rate of solid progress is perhaps greater than most social sciences
30000	and less than many physical sciences.  This is perhaps to be expected
30100	considering the difficulty of the problem.\.
30200	
30300	
30400	
30500					John McCarthy - 9 March 1974