🚀 go-pugleaf

This article arises out of recent discussion in sci.cognitive on how to
distinguish AI and Cognitive Science.

-----------------------------------------------------------------------

Nearly all attempts at precise definitions of fields of enquiry are
pointless, because knowledge has no intrinsic boundaries, and the
divisions that are found between journals, or between conferences, or
between departments in a university or college, are normally accidental
products of particular historical episodes, fashions, academic politics,
etc.

Thus there are no sharp boundaries between:

 o the activities of those who call themselves linguists and those who
   call themselves psychologists
 o the activities of those who call themselves physicists and those
   who call themselves mathematicians
 o those who call themselves engineers and those who call themselves
   scientists
 o those who call themselves philosophers and those who call themselves
   intellectual historians
 o those who call themselves philosophers and those who call themselves
   scientists.
and so on.

Academic disciplines are a bit like species of birds: if you try to
define species in terms of ability to mate you can find groups of birds
of types A, B, C, D, E spread across a continent such that A and B can
mate, and B and C can mate and C and D can mate and D and E can mate,
but A and E cannot mate.

Similarly there are different clusters of people who say they are doing
AI, doing cognitive science, doing cognitive psychology, etc. and if you
look very closely you'll just find lots of patterns of overlap and
interaction, with no clear subdivisions.

If you survey all the research reports from AI labs, all the papers that
have been published in AI journals, all the contents of major AI
conferences (e.g. International Joint Conference on AI, every two years
from 1969 onwards, AISB in the UK, AAAI in the USA, ECAI in Europe,
etc.) the breadth of coverage is huge, including activities that could
be classed as engineering or as science, as philosophy, as logic, as
mathematics, as psychology, as linguistics, as theory of music, and so
on.

Of course, people can define goals for their own AI labs or their own AI
research, or the journals they edit, as being restricted in some way,
e.g. just a branch of engineering. But nobody can legislate for a
community that has in fact been "defined" by a wide spread of activities
under the name of AI over many years and in many countries.

For example, here in the University of Birmingham we are starting a new
undergraduate degree in AI from October 1993 which will be taught in
combination with a variety of other disciplines, because we think AI has
potential links with *all* disciplines either because they contribute to
the study of principles for designing or explaining intelligent systems,
or because they provide subject matter to which AI can be applied (what
knowledge is involved in that subject, how it is represented, how is it
learnt, how is it used, how is it extended, etc.)

Taking a broad perspective, then, we can say that as a matter of
historical, social, fact (i.e. not a matter of definition), AI
activities encompass two main kinds of goals ENGINEERING and SCIENTIFIC,
both concerned with the study of *designs* for intelligent systems,
where "intelligent" is deliberately left undefined because we don't yet
know what options there are for drawing boundaries within the whole
space of behaving systems. (It is therefore premature to try to define a
sub-set of behaving systems as intelligent.)

So AI thus conceived includes:

1. Engineering
    Attempting to design useful machines that do things that require
    intelligence (especially human-like intelligence), for example:
        robots in factories,
        teaching systems,
        diagnosis systems,
        planning systems,
        "intelligent front ends",
        visual inspection systems,
        decision support systems,
        mechanisms for inducing regularities and patterns,
        etc.

2. Science
    Attempting to understand the human mind and other kinds of
    intelligent systems: existing types and theoretically possible
    types, biological and artificial types; using a variety of
    methodologies including searching for general principles, exploring
    design options and trade-offs (between architectures, mechanisms,
    formalisms, etc.), building working models, testing them, analysing
    them in order to understand why they succeed or why they fail,
    comparing them with performance of humans and other animals, and
    so on.

What is common to all these aspects of AI is the adoption of a
"design-based" approach (roughly what Dan Dennett called the "design
stance" in his 1978 book Brainstorms, but with more emphasis on
distinguishing requirements, constraints, designs, mechanisms, layers of
implementation, etc.). For instance, where a psychologist who calls
himself a cognitive scientist would be happy just to measure aspects of
human cognitive performance in various situations and publish graphs
showing relationships between context and performance, someone in AI
will typically be concerned to consider how that relationship might
arise out of design principles embodied in underlying mechanisms.

But within this broad framework of AI there are many differences, e.g.
o according to the topic of interest (chess, vs logical theorem proving,
  vs vision, vs language, vs planning, vs emotions, etc.),
o according to whether an attempt is made to design a complete agent or
  just model some fragment of intelligence,
o according to whether mere abstract competence is to be explained or
  details of performance also (e.g. errors, timing),
o according to whether researchers work top down from requirements for
  high levels of intelligence or bottom up from modelling simple
  organisms,
o according to whether they restrict themselves to one kind of mechanism
  (e.g. symbol manipulation, neural networks, etc.) or keep an open mind
  and use whatever sorts of mechanisms are appropriate for different
  sub-functions; or even whether they assume that it can all be done on
  a single serial computer or wether some other kinds of mechanisms
  (e.g. chemical mechanisms) will play an essential role.
o according to whether they expect explanatory designs to be explicitly
  constructed by AI researchers or whether they think only automated
  mechanisms simulating evolution (e.g. genetic algorithms) have any
  hope of searching the space of designs in a reasonably short time.

There are also differences between those who appear to focus only the
*content* of the information required for certain capabilities without
being concerned with *processes* involved in acquiring or using that
information, or *mechanisms* enabling those processes, and differences
between those who think human-like intelligence can be modelled or
explained, and those who think it can also be *replicated*.

All that is just a sketchy and to some extent inadequate summary of the
breadth and variety of activities encompassed within AI since its
inception somewhere in the last 40 or so years (e.g. depending whether
you think it began with Alan Turing or somewhere else).

Given all that complexity, I find the search for a one or two line
definition of AI or cognitive science is doomed to failure or vacuity:
most attempts to draw sharp boundaries between AI and cognitive science
ignore the facts about what's going on in the name of AI or in the name
of Cognitive Science.

If forced to characterise their relationship I would be inclined to say
that they share a *huge* common core, except that some people call
themselves cognitive scientists who don't care about explanatory
mechanisms (they merely collect observations), and there are some people
who claim to be doing AI who don't care about trying to explain existing
intelligent systems (they merely want to  solve some practical
problems).

-----------------------------------------------------------------------
Advert: my chairperson's introduction to AISB93 conference proceedings
(Conference theme: "Prospects for AI as the general science of
intelligence") will enlarge on these points, a little!

Aaron
---
-- 
Aaron Sloman,
School of Computer Science, The University of Birmingham, B15 2TT, England
EMAIL   A.Sloman@cs.bham.ac.uk  OR A.Sloman@bham.ac.uk
Phone: +44-(0)21-414-3711       Fax:   +44-(0)21-414-4281