This file is at http://www.cs.bham.ac.uk/research/projects/cogaff/misc/aims-of-ai.txt Originally posted to two news groups in 1994 by Aaron Sloman: http://www.cs.bham.ac.uk/~axs/ School of Computer Science, The University of Birmingham, B15 2TT, UK PAPERS: http://www.cs.bham.ac.uk/research/cogaff/ http://www.cs.bham.ac.uk/research/projects/cosy/papers/ TALKS http://www.cs.bham.ac.uk/research/projects/cogaff/talks/ FREE TOOLS: http://www.cs.bham.ac.uk/research/poplog/freepoplog.html ============================================ Also available, better formatted, but slightly out of date at: http://www.philosophyofinformation.net/sloman.htm ============================================ NOTE: links in this document were updated on 28 Jan 2007 Many more links to further discussions of the topics below are available via my web page, and the above "papers" and "talks" links. ============================================ Date: 23 Jul 1994 12:18:10 GMT From: A.Sloman@cs.bham.ac.uk (Aaron Sloman) Newsgroups: comp.ai.philosophy, comp.ai Subject: AI and neural nets (was: Re: Free will again ?) . . . How to approach the question There are different ways of trying to answer the question: What are the aims of AI? a. Try to articulate what you yourself think you are doing and what larger set of goals it fits into. This is what many AI practitioners do. Some are unaware of what lots of others do. b. Repeat some definition of AI that you have read, especially if originally produced by one of the founders or gurus of AI. This is what many undergraduates, recent graduates, recent recruits, journalists, and outsiders do. c. Look at what actually goes on in AI conferences, AI journals, books that purport to be AI, AI research labs, and try to characterise the superset. This could be what a sociologist or historian of science might do. Many AI people now tend to go only to their specialist conferences and read only their specialist journals, so they lose the general vision. d. Like (c) but instead of simply characterising the superset, try to find some underlying theme or collection of ideas which could generate that superset. This is what I've been trying to do for the last 15 years or so, which led me to the notion of trying to understand AI as a study of "design-space", though my ideas about this have been gradually refined and extended, and I've now included requirements-space (or niche-space) as separate from design-space, and added the study of mappings between design-space and requirements-space. I'll elaborate on this, and try to explain why it makes most definitions of AI too narrow. How most characterisations of AI are too narrrow From the standpoint of (d), i.e. looking for a theme covering all the activity of respectable AI researchers, people both within and outside AI mostly offer characterisations of AI that can be criticised as too narrow, a famous example being Marvin Minsky's much quoted (22 year old?) definition of AI as: The science of making machines do things that would require intelligence if done by men. It is quoted in Margaret Boden's 1978 book Artificial Intelligence and Natural Man, and often crops up in comp.ai and elsewhere. Why it is too narrow will emerge presently. (Certainly Marvin's recent work goes beyond this.) But there are worse forms of narrowness! I have found several different types of narrowness (there's some overlap between these headings): (i) definitions that limit AI to a branch of engineering. (ii) definitions that characterise AI as making or doing as opposed to understanding. [This is not the same thing as (i), since *good* engineering (as opposed to mere craft) includes understanding what you have done!!] (iii) definitions that restrict ultimate goals to the understanding or replication of human intelligence or ``human-level'' intelligence, (iv) definitions that try to avoid treading on the toes of other disciplines (e.g. excluding what could be classed as psychology or philosophy or linguistics -- see below), (v) definitions that presuppose a restricted set of representational formalisms (e.g. logical formalisms, possibly extended to include fuzzy and probabilistic logic, etc.) (vi) definitions that emphasise the search for algorithms or AN algorithm (often found in Searley critics of AI). (vii) definitions that restrict AI to a branch of computer science. (viii) definitions which require AI to be tied to computers, where "computer" may either be defined as o whatever can in principle be simulated exactly on a Turing machine and has the power to implement a Turing machine, (possibly subject to problems of adding memory as needed), o or something more vaguely and intuitively specified! I think these are all too restrictive to match the full range of activities that can currently be found in respectable places or publications or conferences under the umbrella of AI. Why are these too narrow? (i) AI isn't just engineering. It is already clear from previous discussions that there are respected AI practitioners who claim to be trying to understand something general without necessarily wishing to apply it to any engineering problems (though they don't rule that out and some of them use engineering problems to help drive their "basic" research). Later I'll offer a diagnosis of why people think it's engineering. They have a point, but they misdescribe it. (ii) People who talk about aiming merely to create something, however ambitious, risk being open to various criticisms including that there are wide-spread well-established cheaper and easier ways of making intelligent systems than doing AI. More importantly, merely achieving something without understanding how it works, why it works, what its limitations are, and how it compares with alternatives, etc. is intrinsically of less interest and of less practical value than having the additional understanding. Much work in AI is in fact concerned with understanding WHEN and WHY things work or fail to work. (Even some work on neural nets!) Alan Bundy's work on "rational reconstruction" emphasises this aspect of AI. E.g. a rational reconstruction of an AI program distinguishes the key ideas that are relevant to explaining its performance from arbitrary implementation details. I think that about 22 years ago I heard John McCarthy characterise the merely doing as the "look ma, no hands" approach to AI. Similar points were, I think, made in Drew McDermott's SIGART 1976 article "Artificial Intelligence Meets Natural Stupidity" (reprinted in Haugeland's collection: Mind Design). (iii) Definitions that restrict ultimate goals to the understanding or replication of human intelligence or human-level intelligence are too narrow because they: (a) ignore the possibility of understanding or creating new kinds of intelligence, and, on the other hand (b) ignore the fact that if AI is a way, or the best way, to understand or replicate human intelligence then it is very likely to be equally relevant to the scientific and engineering goals of understanding or replicating the intelligence of other animals (including for example insects, spiders, birds, etc.) Much good biological science is comparative, but lacks the conceptual tools for understanding behavioural capabilities, as opposed to anatomical and physiological structures and processes. A comparative study of the relationships between different designs and different behavioural capabilities of organisms is a natural extension of existing biology and if it is possible at all it must essentially share concepts, tools, methods and aims with the study of human behavioural capabilities. (I include internal behaviour -- like solving problems, acquiring a taste for art, or enjoying a joke, as well as external behaviour. I.e. I am not taking a behaviourist approach to behaviour.) (iv) The argument that something isn't AI because it is psychology or whatever, ignores the possibility that AI may be a superset of (parts of) other disciplines. E.g. Christopher Longuet-Higgins originally named his AI department in Edinburgh as the Theoretical Psychology Unit. It is quite possible (I would say actually the case) that several other disciplines have problems that cannot be solved except in the framework of AI (as defined below), in something like the way that there are problems in the theory of natural numbers that cannot be solved except in the context of a broader theory incorporating complex numbers, or whatever. Incidentally, Cognitive Science seems to me to be the subset of AI that is concerned primarily with understanding human beings and how they work. (And possibly other animals) We are doing something more general in AI, which subsumes thao. (v) People who define AI in terms of the use of what I call applicative or Fregean formalisms (logics of all kinds, fuzzy or otherwise, and most standard mathematical formalisms) simply ignore the fact that there are people exploring other formalisms, including spatial, pictorial, notations which work in totally different ways, or more abstract formalisms such as patterns of weights in a neural net. Although some textbooks and courses on AI ignore the fact, discussion of the role of multiple forms of representation in intelligence has been visible in AI work since at least as far back as IJCAI71 (where John McCarthy and I had a sort of debate) and is the focus of much current research and debate (e.g. see the special issue of Computational Intelligence Nov 1993). In fact I think it goes back to AI work done in the 1960s. Much of the history of science, mathematics and culture has involved the invention of new formalisms or notations: who knows what new kinds will have to be invented for AI purposes? (E.g. I suspect vision systems need something utterly different from anything now in use in AI labs.) Defining AI in terms of current formalisms would be as silly as defining physics in terms of known mathematics would have been in the time of Galileo. (vi) Definitions that emphasise the search for algorithms or AN algorithm ignore the fact that much actual AI design work (most obviously in robotics) is concerned with larger issues, such as what kinds of architectures are appropriate for various kinds of capabilities. Of course, algorithms can be and often need to be included in an architecture. Some of those who attack AI, including Searle and Penrose, seem to think that AI is the search for an algorithm, or THE algorithm, for intelligence. I suspect that if there are people who think it is all NOTHING BUT algorithms that's because they think that anything done on computers ultimately involves principled production of a stream of instructions (e.g. machine code or microcode instructions) and underlying that stream there ``must'' be an algorithm. Even if this is true (which I challenged in my review of Penrose in the AI journal 1992), arguing in that way is like arguing that AI is about atoms and molecules because all implementations of AI systems involve the behaviour of atoms and molecules. Such arguments fail to take account of the importance of different levels of description. Thus for a robot designer the analysis of architecture, i.e. the functional decomposition into coexisting capabilities and their interactions and interdependencies, is not the same thing as analysis of algorithms. The study of high level architectures to support intelligence, of which Minsky's Society of Minds is an example, though not the only example, will, I suspect, play an increasingly important role in AI. (Maybe I am just prejudiced because it is one of my current concerns! But important or not, it's there in AI, and it's broader than the study of algorithms.) (vi) Whether definitions that restrict AI to a branch of computer science are too narrow is a slightly tricky question as it depends on how one construes computer science. If CS is restricted to the study of a certain limited class of structures and processes then it could turn out that there are brain processes essential for certain sorts of capabilities that are not within the field of that sort of computer science, any more than the behaviour of stretched soap films would now be seen as part of computer science, even if some architects use them as analogue computers. But computer science as actually practiced (and I include many forms of software engineering as well as theoretical computer science, language development, etc.) has been steadily broadening its scope over the last forty years or so, and if it evolves into the most general study of processes and how they can be generated and controlled then that's broad enough to subsume AI! It's certainly broader than early conceptions of control theory or cybernetics, which were once thought to be sufficiently general for the purpose, e.g. by Wiener (and are still being plugged by some people, e.g. William T Powers). There's a concept of computation that is concerned with abstract structures, e.g. sequences of ``abstract machine states'' whose properties have nothing to do with occurring in time or having causal influences. Much theoretical computer science is about such abstractions. That mathematical notion of computation has nothing specific to do with causation or control, though it can be applied to patterns and structures embedded in causal mechanisms. It can equally well be applied to ordered sets of Go"del numbers and other entities in realms from which time and causation are entirely absent. That sort of study is related to AI in something the way that mathematics is related to physics. It provides part of the conceptual toolset. But AI goes beyond that, as it is in part about what is physically realisable, as discussed below. (vii) Whether constraining AI to be based on computers is too narrow is also a tricky question partly because the notion of a computer is pretty ill defined and in some sense everything can be seen as a computer. (E.g. anything can be used to compute the behaviour of something else that's similar: as wind-tunnels used for design work illustrate.) On the other hand if we restrict the notion of computer in some way, e.g. to exclude real randomness, or to exclude continuous change, or whatever, then we may find that some of the tools required for AI are not computers. Similarly, if it turns out that quantum computers are significantly different from Turing machines e.g. because they change the complexity classes of various problems, then restricting AI to what can be implemented on a Turing machine may be too restrictive. In particular, if the goals of AI include explaining behavioural capabilities that involve TIME CONSTRAINTS (e.g. the ability to do vision and motor control in real time in THIS physical universe) then that might, in this universe, actually rule out physically implemented Turing machines for certain purposes. (E.g. a potentially indefinitely extendable memory may be physically incompatible with bounded access times.) If you regard AI as not being concerned with such trivial details as physical realisability, but only with the study of what is in some sense possible in the abstract in some sort of universe, then perhaps Turing machines (or collections of asynchronously linked Turing machines) will suffice. That's close to treating AI as a branch of mathematics rather than engineering. (And that's a perfectly respectable sub-discipline of AI. But it's certainly not all of AI.) In many ways the time (and space) constraints are what make AI interesting and hard. There are many problems that have trivial solutions using exhaustive search, where these solutions have explosive time complexity. (E.g. exhaustive search in chess, or theorem proving.) Much work in AI is concerned with how to replace such solutions with others that are usable in a real working agent. I believe that many features of how the human mind works are direct consequences of the need to trade space for time, often leading to messy and complex mechanisms that would not be required if speed were not an issue. E.g. if you cache a lot of partial results in case they are needed again quickly later, and then you later change some of the general assumptions from which they were derived, it may be impossible to maintain consistency. The attempt to use "truth-maintenance" systems to do this can itself explode combinatorially in some contexts. Returning to the suggestion that AI is essentially concerned with computers: it is stupid to DEFINE a discipline in such a way as to exclude the possibility of it extending some of its basic concepts and tools. (Compare physics now with what Newton might have defined as physics.) It seems to me that the important thing about computers from the standpoint of AI is that they are a particularly powerful and flexible type of control system, i.e. a mechanism for producing behaviour in a principled way in the context of some environment: no other known and well understood mechanism is capable of such diverse behaviour and such rapid and varied change in its capabilities, largely because a computer provides a substratum for the implementation of a huge variety of different virtual machines, with very different and even rapidly changing architectures. Brains are, of course, equally apt vehicles for intelligence: only we don't know how they work (e.g. how much of the chemical-level processing is significant?). Maybe they too provide a vehicle for implementing rapidly changing virtual structures, but in a totally different way. [Digression: Incidentally I suspect the concern with processes and realisability subject to constraints, mentioned above, is what people who say that AI is an engineering discipline are really getting at. But this doesn't make AI engineering: it is a fascinating topic for scientific study independently of any particular applications. I.e. it is primarily a scientific research area, which can be informed by and applied to engineering activities. End-Digression.] So, what then is AI? After all that preparation, I'll now repeat what I wrote in a previous message, but which experience suggests fails to communicate with the majority of people in AI, for reasons hinted at above. AI is: The general study of sophisticated self modifying information-driven control systems, both natural (biological) and artificial, both actual (evolved or manufactured) and possible (including what might have evolved or might be made). By the ``general study'', I mean to include not just the creation of any particular such system, but an understanding of what the options are, and how they differ and why. I include not only individual agents, but also societies and the like: social systems add new constraints and design possibilities. What this means is understanding the design principles underlying different sorts of designs, and knowing how those designs account for different sets of capabilities (= satisfy different sets of requirements = fit different possible environmental niches) and accounting for various tradeoffs when it is impossible simultaneously to satisfy all requirements, or to optimise them all at once. Both design-space and niche-space have very rich and intricate structures with many kinds of discontinuities, and complex mappings between designs and niches. E.g. there need not be uniquely optimal designs corresponding to particular niches: usually there are only partial matches and tradeoffs. (Perhaps a new branch of topology will have to be invented before we can think clearly and deeply about these mappings??) I don't think it is helpful to ask for a precise delimitation of the set of control systems that AI is concerned with. Phrases such as "intelligent" or "human like" reflect particular vague sub-regions but we should keep an open mind as to how far we'll be exploring in the future. Most AI people are not particularly interested in such simple things as thermostats and simple homeostatic devices, but you can, in principle, gradually add increasing sophistication to a home temperature control until you have something that holds conversations about the temperature, adjusts the temperature to suit the preferences and current activities of occupants of different rooms etc. Where should we draw the line and say intelligence, or the relevance of AI, starts? Why bother to draw lines? Thermostats are kind of limiting case and if we wished to include them then AI could subsume all of what is normally done under the rubric of control theory. In fact at present AI people tend to require a certain minimum level of sophistication, whether in representational capability, flexibility of response, or whatever, which is why words like "intelligent" tend to be used in describing their aims. But I don't think there's any point at this stage trying to define the boundaries. Perhaps later when we know more about the significant discontinuities in design-space we may be able to agree that one of them should define the limits of AI (e.g. something to do with capabilities for internal monitoring and self modification?). Insisting that there must be a sharp boundary would be like trying to insist on defining the boundary between physics and chemistry, or between chemistry and biology. Why bother? The role of ``toy'' problems Finally, John McCarthy raised an important issue concerning the role of so-called toy problems, in an email message about grand challenges for AI: Basic research in AI should not be identified with theory. Experimental basic research in AI exists and needs to be supported. Unfortunately, its practitioners are criticized for working on "toy problems" rather than applications. This is like criticizing geneticists for working with fruit flies. "Who needs a better fruit fly?, they could be asked. ... [quoted with permission] I think much valuable work in AI, as in other sciences, can be done by thinking about "toy" problems which are toy in the sense that they identify a key issue that is currently unsolved and leave out a lot of messy detail that (it is hoped) can be dealt with separately. Of course, you have to make sure that what you leave out isn't part of the solution! It is fashionable (and very easy) to criticise work on such abstract and simplified problems, but in fact they are often the very things that yield key insights. The invention of computers by Turing and others came from abstracting away from a great deal of the detail of the process of mathematical thinking, for example. Of course people who simplify while unaware of what they are ignoring and what its implications are can be criticised. Equally people who address supposed non-toy problems e.g. by building ``real'' robots with ``real'' visual systems and ``real'' motors etc. may spend a lot of time trying to get the machine to do trivial things that ignore most of what human and animal vision or motor control are about. IT'S POSSIBLE TO DISGUISE A TOY PROBLEM BY BURYING IT IN A LOT OF EQUIPMENT. That's all for now. Some of these points are elaborated in a paper available here: http://www.cs.bham.ac.uk/research/projects/cogaff/81-95.html#27 Aaron Sloman: Explorations in Design Space in Proc ECAI94, 11th European Conference on Artificial Intelligence Edited by A.G.Cohn, John Wiley, pp 578-582, 1994. [[previously available here ftp://ftp.cs.bham.ac.uk/pub/dist/papers/cog_affect/Aaron.Sloman_explorations.ps.Z But no longer]] In that paper (to appear in ECAI94 proceedings) I offer a definition of AI as exploration of design-space and its relationships with niche-space. Aaron ------------------ Aaron Sloman, http://www.cs.bham.ac.uk/~axs/ School of Computer Science, The University of Birmingham, B15 2TT, UK