SOWA@IBM.COM (John Sowa) (04/28/88)
In response to some recent questions, I thought that it might be useful
to cite a few historical references:
1. The first use of the term demon in AI was for the system Pandemonium
by Oliver Selfridge (1958). He developed it as a system for
learning to recognize human-keyed Morse code. It consisted of
low-level demons that looked for patterns. When a demon found
its pattern, it would "shout". Higher-level demons listened for
shouts from lower-level demons. They, in turn, would shout when
they heard a characteristic pattern of shouts.
2. The term "demon" was introduced into physics by Maxwell, who
used it in thought experiments in thermodynamics; e.g. imagine
a demon who watched molecules bouncing around and opened a trap
door to allow only the fast ones to pass through. In principle,
it could reduce entropy by separating hot gas from cool gas.
However, the entropy of the demon itself would increase. For
a discussion of demons in physics, see von Neumann (1951), who
contributed to physics as well as logic, set theory, and even
computers.
3. While we're mentioning von Neumann, I have heard some people
distinguish highly parallel computers from "von Neumann machines."
However, von Neumann (1958) wrote one of the first books about
parallel computation and the possibility of simulating the brain.
So the term "von Neumann machine" could refer either to conventional,
single-CPU machines or to highly parallel connectionist machines.
4. A previous note mentioned Carl Hewitt's PLANNER as a source for the
three-way distinction between if-needed, if-added, and if-deleted
demons. The MIT reports may not be easy to find, but there is a
paper by Hewitt (1969) in the first IJCAI. That paper is confusing
and hard to read, but you can find the three-way distinction in it.
Although Hewitt did not invent if-needed or if-added demons, I do
not know of any earlier version of an if-deleted demon.
5. Goal-directed or if-needed patterns were well developed in the
General Problem Solver. The most definitive reference to GPS is
the book by Ernst & Newell (1969), but there are papers on early
versions dating back to 1959.
6. The 1969 version of GPS also had a well developed use of "schemas,"
which were frame-like structures that predated frames by at least
6 or 7 years. A schema always had unbound variables. When all
its variables were instantiated, it was called a "model."
7. The term schema was introduced to Newell & Simon by Adriaan de Groot,
who visited Carnegie in the 1960s. De Groot (1965) wrote a highly
influential book on thinking processes in chess, in which he applied
the theories of the psychologist Otto Selz (1913, 1922). Selz had
a theory of "schematic anticipation" in which a schema served as a
goal towards which the thinking processes were directed. Selz even
described backtracking search procedures as a way of satisfying
the goal and used a network notation for his schemas. Quillian,
who studied with Newell & Simon, cited Selz in his thesis (1966),
but the abridged version reprinted in Minsky (1968) doesn't
mention Selz.
John Sowa
References:
O. G. Selfridge (1968) "Pandemonium: a paradigm for learning," in
Mechanisation of Thought Processes, Proceedings of a symposium held
at the National Physical Laboratories, Nov. 1958, Her Majesty's
Stationery Office, London, pp. 511-531.
J. von Neumann (1951) Mathematical Foundations of Quantum Mechanics,
Princeton University Press, Princeton, NJ.
J. von Neumann (1958) The Computer and the Brain, Yale University Press,
New Haven.
C. Hewitt (1969) "PLANNER: a language for proving theorems in robots,"
Proceedings of IJCAI, pp. 295-301.
G. W. Ernst & A. Newell (1969) GPS: A Case Study in Generality and
Problem Solving, Academic Press, New York.
A. de Groot (1965) Thought and Choice in Chess, Mouton, The Hague.
O. Selz (1913) Ueber die Gesetze des geordneten Denkverlaufs,
Spemann, Stuttgart.
O. Selz (1922) Zur Psychologie des produktiven Denkens und des Irrtums,
Friedrich Cohen, Bonn.
M. R. Quillian (1966) Semantic Memory, Report AD-641671, Clearinghouse
for Federal Scientific and Technical Information.
M. Minsky (1968) Semantic Information Processing, MIT Press, Cambridge, MA.