perm filename LIGHT.RE3[2,JMC]2 blob
sn#092422 filedate 1974-03-21 generic text, type T, neo UTF8
00100 \\M0NGR25;\M1BDI25;\F0
00200 Draft for comments: This paper is in the file LIGHT.RE3[2,JMC]@SAIL
00300 It was solicited as a review by the AI Journal.
00400
00500 Artificial Intelligence: A General Survey by
00600 Professor Sir James Lighthill, FRS, in \F1Artificial Intelligence: a
00700 paper symposium\F0, Science Research Council 1973
00800
00900 \J Professor Lighthill of Cambridge University is a famous
01000 hydrodynamicist with a recent interest in applications to biology.
01100 His review of artificial intelligence was at the request of Brian
01200 Flowers, head of the Science Research Council of Great Britain, the
01300 main funding body for British university scientific research. Its
01400 purpose was to help the Science Research Council decide requests for
01500 support of work in AI. Lighthill claims no previous acquaintance
01600 with the field, but refers to a large number of authors whose works
01700 he consulted though not to any specific papers.
01800
01900 Unfortunately, workers in artificial intelligence lose
02000 intellectual contact with Professor Lighthill almost immediately,
02100 because he defines the field in such a way as to exclude our most
02200 important research goals. He does this by classifying work in
02300 artificial intelligence into three categories A, B and C. A stands
02400 for \F1advanced automation\F0 which he likes, C stands for
02500 \F1connections\F0 to psychology and neurophysiology which he also
02600 likes, and B stands for \F1bridge\F0 between the other two and also
02700 for \F1building robots\F0 both of which he doesn't like. The term
02800 \F1robot\F0 is defined in an idiosyncratic way to refer to any
02900 computer program or device which is neither aimed at an application
03000 nor intended to model the brain. He then states that activities in B
03100 can be justified only in so far as they make a connection between A
03200 and C.
03300
03400 Lighthill's ABC classification seems to exclude the
03500 possibility that there can be a science of intelligent behavior that
03600 can be studied apart from applications and apart from biology.
03700 However, for almost all workers in the field, the whole idea of
03800 artificial intelligence is that the relation between problems and
03900 problem solving methods and the relation between situations and the
04000 behavior that will achieve goals can be studied by theory and
04100 computer experiment as an independent subject.
04200
04300 He makes no argument for his classification, and gives no
04400 hint that anyone may think differently. This is somewhat puzzling
04500 since a number of the documents submitted by British AI workers for
04600 his consideration are quite explicit about the point. Perhaps
04700 ignoring this claim plays a tactical role in justifying his proposal
04800 that basic experimental research in AI be abandoned, because if AI
04900 research has scientific problems of its own, they should be pursued
05000 even though the level of funding may depend on the prospects for
05100 results at the present level of knowledge and talent. Whereas if the
05200 research is only a means toward solving some other scientific or
05300 practical problems, then the subject may be abandoned if there are
05400 more promising ways of solving the other problems.
05500
05600 Having ignored the possibility that AI has goals of its own,
05700 Lighthill goes on to document his claim that it has not contributed
05800 to applications or to psychology and physiology. He exaggerates a
05900 bit here, it seems worthwhile to spend some effort disputing his
06000 claims that AI has not contributed to these other subjects.
06100
06200 In my opinion, AI's contribution to practical applications
06300 has been significant but so far mostly peripheral to the central
06400 ideas and problems of AI. Thus the LISP language for symbolic
06500 computing was developed for AI use, but has had applications to
06600 symbolic computations in other areas, e.g. physics. Moreover, some
06700 ideas from LISP such as conditional expressions and recursive
06800 function definitions have been used in other programming languages.
06900 However, the ideas that have been applied elsewhere don't have a
07000 specifically AI character and might have been but weren't developed
07100 without AI in mind. Other examples include time-sharing, the first
07200 proposals for which had AI motivations and some techniques of picture
07300 processing that were first developed in AI laboratories and have been
07400 used elsewhere. Even the current work in automatic assembly using
07500 vision might have been developed without AI in mind. However, the
07600 Dendral work has always had a specifically AI character, and many of
07700 the recent developments in programming such as PLANNER and CONNIVER
07800 have an AI motivation.
07900
08000 AI's contributions to neurophysiology have been small and
08100 mostly of a negative character, i.e. showing that certain mechanisms
08200 that neurophysiologists propose are not well defined or inadequate to
08300 carry out the behavior they are supposed to account for. I have in
08400 mind Hebb's proposals in his book \F1The Organization of Behavior\F0.
08500 No-one today would believe that the gaps in those ideas could be
08600 filled without adding something much larger than the original work.
08700 Moreover, the last 20 years experience in programming machines to
08800 learn and solve problems makes it implausible that cell assemblies
08900 \F1per se\F0 would learn much without putting in some additional
09000 organization, and physiologists today would be unlikely to propose
09100 such a theory. However, merely showing that some things are unlikely
09200 to work is not a \F1positive\F0 contribution.
09300 I think there will be more interaction between AI and neurophysiology
09400 as soon as the neurophysiologists are in a position to compare
09500 information processing models of higher level functions with
09600 physiological data. There is little contact at the nerve cell level,
09700 because, as Minsky showed in his PhD dissertation in 1954, almost any
09800 of the proposed models of the neuron is a universal computing element,
09900 so that there is no connection between the structure of the neuron and
10000 what higher level processes are possible.
10100
10200 On the other hand, the effects of artificial intelligence
10300 research on psychology have been larger as attested by various
10400 psychologists. First of all, psychologists have begun to use models in
10500 which complex internal data structures that cannot be observed
10600 directly are attributed to animals and people. Psychologists have
10700 come to use these models, because they exhibit behavior that cannot
10800 be exhibited by models conforming to the tenets of behaviorism which
10900 essentially allows only connections between externally observable
11000 variables. Information processing models in psychology have also
11100 induced dissatisfaction with psychoanalytic and related theories of
11200 emotional behavior. Namely, these information processing models of
11300 emotional states can yield predictions that can be compared with
11400 experiment or experience in a more definite way than can the vague
11500 models of psychoanalysis and its offspring.
11600
11700 Contributions of AI to psychology are further discussed in
11800 the paper \F1Some Comments on the Lighthill Report\F0 by N. S.
11900 Sutherland which was included in the same book with the Lighthill
12000 report itself.
12100
12200 Systematic comment on the main section, entitled \F1Past
12300 Disappointments\F0 is difficult because of the strange way the
12400 subject is divided up but here are some remarks:
12500
12600 1. Automatic landing systems for airplanes are offered as a
12700 field in which conventional engineering techniques have been more
12800 successful than AI methods. Indeed, no-one would advocate applying
12900 the scene analysis or tree search techniques developed in AI research
13000 to automatic landing in the context in which automatic landing has
13100 been developed. Namely, radio signals are available to determine the
13200 precise position of the airplane in relation to a straight runway
13300 which is guaranteed clear of interfering objects. AI techniques
13400 would be necessary to make a system capable of landing on an
13500 unprepared dirt strip with no radio aids which had to be located and
13600 distinguished from roads visually and which might have cows or
13700 potholes or muddy places on it. The problem of automatically driving
13800 an automobile in an uncontrolled environment is even more difficult
13900 and will definitely require AI techniques, which, however, are not
14000 nearly ready for a full solution of such a difficult problem.
14100
14200 2. Lighthill is disappointed that detailed knowledge of
14300 subject matter has to be put in if programs are to be successful
14400 in theorem proving, interpreting mass spectra, and game playing. He
14500 uses the word \F1heuristics\F0 in a non-standard way for this. He
14600 misses the fact that there are great difficulties in finding ways of
14700 representing knowledge of the world in computer programs and much AI
14800 research and internal controversy are directed to this problem.
14900 Moreover, most AI researchers feel that more progress on this
15000 \F1representation problem\F0 is essential before substantial progress
15100 can be made on the problem of automatic acquisition of knowledge. Of
15200 course, missing these particular points is a consequence of missing
15300 the existence of the AI problem as distinct from automation and
15400 study of the central nervous system.
15500
15600 3. A further disappointment is that chess playing programs
15700 have only reached an "experienced amateur" level of play. Well, if
15800 programs can't do better than that by 1978, I shall lose 250 pounds
15900 and will be disappointed too though not extremely surprised. The
16000 present level of computer chess is based on the incorporation of
16100 certain intellectual mechanisms in the programs. Some improvement
16200 can be made by further refinement of the heuristics in the programs,
16300 but probably master level chess awaits the ability to put general
16400 configuration patterns into the programs in an easy and flexible way.
16500 I don't see how to set a date by which this problem must be solved in
16600 order to avoid disappointment in the field of artificial intelligence
16700 as a whole.
16800
16900 4. Lighthill discusses the \F1combinatorial explosion\F0
17000 problem as though it were a relatively recent phenomenon that
17100 disappointed hopes that unguided theorem provers would be able to
17200 start from axioms representing knowledge about the world and solve
17300 difficult problems. In general, the \F1combinatorial explosion\F0
17400 problem has been recognized in AI from the beginning, and the usual
17500 meaning of \F1heuristic\F0 is a device for reducing this explosion.
17600 Regrettably, some people were briefly over-optimistic about what
17700 general purpose heuristics for theorem proving could do in problem
17800 solving.
17900
18000
18100 Did We Deserve It?
18200
18300 Lighthill had his shot at AI and missed, but this doesn't
18400 prove that everything in AI is ok. In my opinion, present AI
18500 research suffers from some major deficiencies apart from the fact
18600 that any scientists would achieve more if they were smarter and
18700 worked harder.
18800
18900 1. Much work in AI has the "look ma, no hands" disease.
19000 Someone programs a computer to do something no computer has done
19100 before and writes a paper pointing out that the computer did it. The
19200 paper is not directed to the identification and study of intellectual
19300 mechanisms and often contains no coherent account of how the program
19400 works at all. As an example, consider that the SIGART Newsletter
19500 prints the scores of the games in the ACM Computer Chess Tournament
19600 just as though the programs were human players and their innards were
19700 inaccessible. We need to know why one program missed the right move
19800 in a position - what was it thinking about all that time? We also
19900 need an analysis of what class of positions the particular one
20000 belonged to and how a future program might recognize this class and
20100 play better.
20200
20205 2. A second disease is to work only on theories that can be
20210 expressed mathematically in the present state of knowledge.
20215 Mahtematicians are often attracted to the artificial intelligence
20220 problem,...
20225
20300 2. Every now and then, some AI scientist gets an idea for a
20400 general scheme of intelligent behavior that can be applied to any
20500 problem provided the machine is given the specific knowledge that a
20600 human has about the domain. Examples of this have included the GPS
20700 formalism, a simple predicate calculus formalism, and more recently
20800 the PLANNER formalism and perhaps the current Carnegie-Mellon
20900 production formalism. In the first and third cases, the belief that
21000 any problem solving ability and knowledge could be fitted into the
21100 formalisms led to published predictions that computers would achieve
21200 certain levels of performance in certain time scales. If the
21300 inventors of the formalisms had been right about them, the goals
21400 might have been achieved, but regrettably they were mistaken. Such
21500 general purpose formalisms will be invented from time to time, and,
21600 most likely, one of them will eventually prove adequate.
21700 However, it would be a great relief to the rest of the workers in AI
21800 if the inventors of new general formalisms would express their
21900 hopes in a more guarded form than has sometimes been the case.
22000
22100 3. At present, there does not exist a comprehensive general
22200 review of AI that discusses all the main approaches and achievements
22300 and issues. Most likely, this is not merely because the field
22400 doesn't have a first rate reviewer at present, but because the field
22500 is confused about what these approaches and achievements and issues
22600 are. The production of such a review will therefore be a major
22700 creative work and not merely a work of scholarship.
22800
22900 4. While it is far beyond the scope of this review to try
23000 to summarize what has been accomplished in AI since Turing's 1950 paper,
23100 here is a five sentence try: Many approaches have been explored and
23200 tentatively rejected including automaton models, random search,
23300 sequence extrapolation, and many others. Many heuristics have been
23400 developed for reducing various kinds of tree search; soee of these are
23500 quite special to particular applications, but others are general.
23600 Much progress has been made in discovering how various kinds of
23700 information can be represented in the memory of a computer, but
23800 a fully general representation is not yet available. The problem
23900 of perception of speech and vision has been explored and recognition
24000 has been found feasible in many instances. A beginning has been made
24100 in understanding the semantics of natural language.
24200
24300
24400 John McCarthy - 9 March 1974