Jump over left menu
Comments on the Lighthill Report and the Sutherland Reply
Professor D Michie
Department of Machine Intelligence and Perception, Edinburgh.
Two contrary attitudes are common. In the first place there is a widespread, although mostly unconscious, desire to believe that a machine can be something more than a machine, and it is to this unconscious urge that the newspaper articles and headlines about mechanical brains appeal.
On the other hand, many people passionately deny that machines can ever think. They often hold this view so strongly that they are led to attack designers of high-speed automatic computing machines, quite unjustly, for making claims, which they do not in fact make, that their machines have human attributes.
M. V. Wilkes 'Can a machine think?' in Discovery May, 1953
Sir James Lighthill's report speaks of the ABC of the subject, categorising it as follows:
- A - Advanced Automation
- B - Building Robots
- C - Computer-based CNS research
The report regards A and C as worthy activities which, however, have made disappointing progress. B is regarded as unworthy, and as having made very disappointing progress indeed. B, it should be noted, is really used in the report to denote any experimental programming which lacks obvious application to either A or C. Thus computer chess is included in B whereas robot parcel-packing is put into A.
Most people in AI who have read the report have had the feeling that the above classification in misleading. Sir James has arrived at his position by interpreting AI as consisting merely of outgrowths from a number of established areas, viz.:
- A as an outgrowth from control theory,
- B as an outgrowth from science fiction,
- C as an outgrowth from neurobiology,
These interpretations are remote from those current in the field itself.
A number of questions accordingly pose themselves, including the following:-
- Was this report based on as thorough a survey as it should have been? In particular, was opportunity taken to invite the views of the international leaders of the field?
- How successful has the author been in overcoming the difficulties inherent in his inexperience of the field, and in putting aside his own professional biases?
- Has accepted practise been followed in documenting subjective opinions wherever possible and in providing factual sources and references which others can check?
- What is the validity of the A B C classification? Would the computing science community accept it?
- Are the report's assessments of work in the B category - Building Robots -
intended to apply to experimental robotics conducted in the United Kingdom?
If so, should not the author
- have said so plainly,
- have asked to see the experimental robotics work during his visit to Edinburgh?
International opinion not consulted
The first of these questions is so critical as to merit a brief note to the effect that the leading American workers, such as McCarthy, Minsky, Nilsson, Raphael and Robinson, were not in fact consulted. The appearance of their names in the list of fifty given in the report's third paragraph derives from the fact that the author has read scientific writings of theirs not that he invited their opinions. Since the field in question was pioneered in the United States of America, which supports to this day an effort on at least twenty times the scale of that in the United Kingdom, it is well to bear in mind this fact when assessing Sir James' evaluations.
Space does not allow a review of the remaining questions of the above list; a detailed critique is available elsewhere. (D. Michie (1972) On first looking into Lighthill's Artificial Intelligence report (unpublished)) Instead I will briefly indicate two themes which arise from Lighthill's implicit question: If you throw A and C away, what is left, if anything? Lighthill's answer is Building Robots. An alternative answer, which many will prefer, is Intelligence Theory. By this we mean attempts to systematise the design principles of intelligent systems wherever they may be found, whether in the A or C application areas.
Having fixed on Building robots, Lighthill paints a picture of this pursuit which must strike those actually engaged in experimental robotics as somewhat unfamiliar. In studies of actual robot work the role of the equipment is plainly seen as test gear for putting certain types of theoretical ideas to experimental test. A pertinent parallel is the building of wind-tunnels as an aid to aero-engineering - as illustrated in the figure below.
Broad subdivisions of two fields of enquiry according to theoretical-experimental and technological-biological classifications.
This figure brings into relief the reason why Building Robots is an unhelpful choice for the role of Bridge between A and C. It is surely more fruitful, if one seeks inter-disciplinary connections, to choose a common body of theory rather than to seize on a piece of laboratory equipment. One feels that Sir James would be among the first to agree that to speak of Building Wind-tunnels as the Bridge between aero-engineering and the study of bird flight, would direct attention away from the true bridge, namely the science of aerodynamics. The equivalent science in the case of AI is at a primitive stage. It is the hope of every AI professional to contribute in some way to bringing the required theory into being. This, as I see it, is the burden of Sutherland's re-definition, in his contribution to this symposium, of B as standing for Basic.
Encouragement of research in machine intelligence
On this note I would like to leave Sir James Lighthill's interesting and imaginative review and to mention an assessment of a more home-spun quality: the report of SRC Computing Science Committee's long-range panel, published in Computing Science Review. This panel, composed of computer professionals, considered the machine-oriented part of Artificial Intelligence (ie the A+B part) and recommended that special encouragement should be given to this field. However, it is evident to those who work in the field that it would be helpful if a clear and concise statement were given of its goals and methodology. The style of Sir James Lighthill's report suggests that there is a lack of understanding in some quarters, and without this there is a reluctance to recommend significant expenditure. The status and position of the subject are particularly clear at the moment and it is, therefore, opportune that a statement should be made to avoid any further misunderstandings.
The subject, in so far as it comes within the Computing Science Committee's realm of interest, is concerned with machines, and in particular computers, displaying characteristics which would be identified in a human being as intelligent behaviour. Perhaps the characteristics which are most important are those of learning and problem solving. The applied benefits which may be gained from work in this field could bring considerable economic benefit to the country. They are two-fold:
- To relieve the burden at present on the systems analyst and programmer in implementing applications;
- To enable new and more complex applications to be undertaken in this country in competition with work elsewhere.
These are the long-term advantages and to this end work is proceeding on a number of detailed problems, including the following:
- Automatic assembly and other robotic applications
Mass spectogram analysis
Chemical synthesis planning
Assembly line balancing
- Language-understanding systems
Semi-automatic programming (ie teachable systems) and ultra-high level programming languages (like PLANNER, SAIL, CONNIVER)
Group I represents useful applications. Group II represents the subject's own special contribution, independent of specific applications to computer science. This lies in making it more possible for the user to get computing systems to understand what he means.
Many good scientists have been involved in this field and their work has resulted in the development of techniques and methods of wider use, for example:
- List-processing was originally devised by Newell, Shaw and Simon for AI work and first implemented in their IPL language.
- The incorporation of conditional expressions into ALGOL 60 was McCarthy's suggestion derived from his work on LISP, itself inspired by the needs of Artificial Intelligence work.
- The POP-2 language, now implemented on 5 main hardware ranges was specifically developed for AI work, but subsequently shown to be of wider utility.
- In fifteen years of struggle towards language understanding, striking advances have been scored (Bobrow, Winograd, Woods).
- Some of the search and associative techniques used by programmers and operations research workers have been initiated in AI, and assimilated without awareness of their origin.
The problems that have been mentioned above are practical problems. Abstracting from these, and observing the methods of solution, workers in the field have been able to define general principles for intelligent systems. This work has made some progress and the following theorems and methods have been developed.
Some of these are reviewed in a Nature article Machines and the theory of intelligence 23 Feb. 1973
- Theorems of minimality and completeness of various algorithms for heuristically guided search.
- Methods of pruning search trees in special situations: Plausibility analysis; alpha-beta pruning.
- Recursive formation of sub-problems as in Newell and Simon's General Problem Solver.
- Application of theorem-proving ideas in problem-solving.
- Studies of problem-representation.
- Various methods of feature extraction and interpretation for visual data.
- Use of semantics to disambiguate linguistic analysis.
- Matching of descriptions represented as directed graphs (eg hierarchical synthesis).
- Adaptive learning via parameter-optimisation.
- Rote-learning techniques.
- Formation of new concepts from examples and counter-examples.
- Inductive generalisation.
Even so incomplete a list as the above puts into perspective the importance of examining particular problems in depth (such as chess-playing or those involving robots) so as to investigate how to bring the above functions to bear in an integrated fashion. They are but experiments which may be used to derive or test theories. At this early stage of innovation the overwhelming benefit to be derived from a given experimental study lies in its role as a forcing function for new programming techniques and tools. The field is so difficult and the choice of the right problem at the right moment so much part of the art of enquiry that this should be left to the research workers themselves. They should be judged by their success or otherwise in advancing the state of computer programming, and in introducing and testing computer languages of greater expressive power.
Footnote on Sutherland's commentary
Sutherland's otherwise admirable analysis contains two expressions of view with which exception must be taken, namely (1) that AI should not be handled by the Engineering Board of SRC and (2) that AI research in Britain is in a bad way.
- A reasonable approach would surely be to distinguish A-oriented and C-oriented poles of the subject and to provide for the first under the Engineering Board and for the second under the Science Board. Since important contributions continue to be made by computer scientists ignorant of psychology and brain science, and by psychologists ignorant of computer science, it would avoid embarrassment, and reflect scientific reality, to make separate provision.
- Sutherland's proposition that AI research in Britain is in a bad way deserves to be vigorously challenged. But it is inappropriate for me, as founder of the longest-established British research group, to be the one to do this. A better corrective can be obtained from assessments by authoritative outside observers, such as that by Dr. Nils Nilsson (*) of the Stanford Research Institute's Artificial Intelligence Centre and author of the graduate textbook The problem-solving methods in Artificial Intelligence.
* An outsider's view of the Experimental Programming Unit at Edinburgh University, obtainable from School of Artificial Intelligence, Edinburgh.
What is to be done
We are in the embarrassing situation in Britain that in order to carry out significant work over the next few years in the context of international competition, it will be essential to import American machines - specifically the DEC System 10 (formerly known as the PDP 10). I think that everybody would be happier about the case for allowing an American importation now if steps were at the same time taken to see that British AI research never found itself in such a predicament again. What would have to be done if this desirable state of affairs were to be brought about?
Why does the need arise? It is not only, or primarily, because of the superiority of the architecture of the DEC System 10 for AI-type uses. The over-riding consideration is access to the rapidly accumulating fund of AI-oriented software and applications programs in the big American laboratories. The key to the situation is the absence of software for British machines, either present or new range, suitable for AI work, which has its own very peculiar needs. These needs are peculiar. It is hardly more sensible to speak of making do with, say, general scientific software developed without reference to AI than to suggest that, say, plasma physicists short of experimental fusion equipment should make do by borrowing linear accelerators from the particle physicists!
The kind of software development needed if AI workers are ultimately to be put in business as users of the new range of British computers (I do not necessarily intend this phrase to be exclusively confined to ICL) comes under two headings:
- Development of experimental operating systems, compilers and packages, as has been done in a small way on the ICL 4130 at Edinburgh. But the new effort should aim to embrace the entire standard range of facilities which every AI worker should be entitled to take for granted LISP, POP-2, SNOBOL, QA4, PLANNER, CONNIVER, etc. etc., and, ultimately far more important, leap-frogging into the future both by adapting the latest advances of AI research work where appropriate, and by innovation within the R & D effort itself. Also to be considered are operating system features for handling funny peripherals (experimental robots, speech input devices, etc.) and basic packages for front end functions such as, say, video and speech input, robot control functions, language pre-processing. In addition the design and development of advanced peripherals (eg for robotics) should be regarded as an integral part of the job although (as with software) the more standard aspects of instrumentation should be contracted out to industry wherever possible.
- Communality aids whereby new research programs and software developed in overseas laboratories can be made immediately available on demand for British research workers to test out and either accept or reject as tools for their own needs. Communality can be achieved by various means and these means will vary according to the nature of the case, but they include software/hardware interfaces to the ARPA net, and emulation (for example by microprogramming) of the 'donor' machine from which the program is to be adapted.
If a fully-fledged development project is to be got up to full speed by around 1977 then forward studies could usefully be started now. It is already obvious that early installations of a PDP-10 in an active centre of British AI research is a precondition if these studies are to develop fruitfully, since immediate access to the latest AI research materials (and intimate contact with advanced AI research) will be as essential to the specification and development of new research facilities as it is for those who will later be using them. Until the new facilities exist, the only point of access and contact will be through British groups equipped compatibly with their American counterparts, - ie with PDP-10s.
In this field successive workers in a given area should be able to stand on the shoulders of their predecessors through the medium of successive contributions to a common stock of new language aids and library packages. This will not happen unless someone makes it his business continually to scoop in what is new and useful and build it into a properly documented and integrated system. The level at which the British AI community will be able to contribute in the late 1970s, as judged by competitive international standards, will be crucially affected by the sophistication of the available software.