Most of the symbols in PopBeast have names which look like English words:
opens, assertion, in. When looking at programs that
use such symbols, it's very easy to fall into the trap of thinking that
they mean a lot more to the program than they do. When we see the
symbol opens, it induces a rich network of concepts and
associations. None of these, however, are available to PopBeast: to it,
opens only gains meaning by virtue of the way it is manipulated
by various parts of the program.
Another way to make this point is that PopBeast would behave exactly the same
if I were to systematically replace every name by its equivalent in some
other language. I could, for example, replace square by
vierkant, key by sleutel, me by mij
...As long as I have done this so that names are different in this
new language wherever they were different in the original, the program
won't suffer. I could of course also use made-up names: square by
z, key by zz, me by zzz ...This is an important point, well worth bearing in mind when reading AI
programs and books. Most of the examples of symbolic representations you
will see use English names, for the simple reason that authors and
programmers find it easier to work with names with familiar
connotations. These names mean so much to us that we can easily forget
they have no such meaning to their program. Beware of this. I recommend
reading Artificial intelligence and natural stupidity by Drew
McDermott from Mind Design by Haugeland (MIT 1981, PSY KH:H029)
as a preventative. It assumes some experience with semantic nets, so you
may want to come back to it when you've read about those.
If the internal symbols don't gain meaning from their names, how do they
gain it? PopBeast's brain does not operate in isolation, but perceives and
acts in what, to it, is an external reality. How does the internal symbol
door come to be connected with what PopBeast perceives when it sees
a # in Eden? This is the symbol-grounding problem: how,
in an artificial or natural symbol-manipulating system, the internal
symbols come to be connected to external referents and actions. In PopBeast,
the answer is that although most of the symbols' names are completely
arbitrary, this is not true of the images on PopBeast's retina, nor of the
motor actions that it obeys. These are determined by Eden's ``laws of
physics'', and somewhere inside PopBeast there is a
consistent mapping between them and symbols like door - a
mapping set up by me. The symbol-grounding problem becomes more acute in
learning systems which must identify for themselves novel
features and properties for which their programmer has not provided
pre-defined symbols.