 The Frame Problem
  A radical split
The Frame Problem
a radical split
Peter Doomen - Masters AI Program - Cognitive Science Option
Two kinds of FP's
If (1) it is possible to understand "the" frame problem and (2) I at least understood what
some others say about it, then (3) there are two frame problems. The
first one is not solvable by any physical system—be it a computer or a human being. The second one is
currently only solved by biological beings, but it is not
principally impossible for computers to do so.
FP as a problem of complete knowledge
As Toth (1995) remarks, omniscience is needed to deal with the first version of the frame problem, i.e.
to know what is going to happen assumes that one has a
complete description of the current world state, plus complete knowledge of all laws of nature. Only
Laplace's ghost can do that, of course. It is therefore
unrealistic to expect AI systems to be able to find a solution for it. The FP stated this way belongs
to the class of "prove that the world exists" problems.
FP as a problem of flat knowledge
A second version (Toth, 1995 p. ) handles the difference between human and computer on the level of
metaknowledge. People are usually very good at knowing
what they can possibly know, computers don't. When asked "who is currently the prime minister of
foreign affairs of Zimbabwe?" people immediately say "I don't
know". A computer consults its database, and depending on its size, returns "X not found".
This is what I call "flat knowledge": knowledge without knowing what
you know.
Whatever computer it may be, when it claims to be intelligent, it should be able to solve this FP version.
Implementing metaknowledge, however, is not only a
technical issue: a better understanding of human understanding may be needed.
Reference
Toth, J.A. (1995). Book review of Reasoning agents in a Dynamic World: the Frame Problem (K.M.
Ford and P.J. Hayes, eds.), Artificial Intelligence,
73: 323-369.
Part I
The Frame Problem
Part I: The Papert Principle
Peter Doomen - Master's Programme in AI
"Some crucial steps in the development of mental capabilities does not consist in achieving
new skills, but in achieving new ways of handling achieved
knowledge" (Minsky, 1988, p. 105)
Cartoon-movie makers are happy with computer aided graphics, for it solves their problem of the moving
duckling and the fixed background. It would be nice to
have a computer that solves the AI Frame Problem. This is certainly not impossible; in fact, Piaget
has shown (Minsky, 1988, p.102) that human FP solving
capabilities (Hayes, 1987, p. 6) are learned—not innate. Up to a certain age, children are unaware of
strange changes in the environment (comparable to the
unloading of the YSP gun), whereas older children seem surprised.
The only explanation is that certain mental restructurations take place. From flat knowledge,
children develop metaknowledge. Metaknowledge helps when
omniscience (Toth, 1995, p. 340) is not possible. Knowing what you know and what not is
necessary. When asked "Who is the minister of foreign affairs of
Zimbabwe", we immediately know that we do not know the answer. A computer, however, will consult
its database and conclude that it cannot find the name
(presumed that it isn't).
Furthermore, people have extensive hierarchies of knowledge that are extremely context-dependent. This
is how we solve the Frame Problem: we quickly learn
that most things don't change after an action. We also learn what does change—depending on the action.
Our mental skills are extremely versatile. The way these
structures develop is connected with the notion of "aboutness" (Toth, 1995, p. 333):
our mental model of a cat depends on our experience with hairy, meowing
mammals that beg for food when you're eating and the like. Defining the cat this way is almost impossible:
therefore, the only way of telling a computer what a cat
is in human terms, is to let it experience it that way.
But current AI systems claim at best to be as intelligent as a two year old child. Probably, a mental
restructuration would help them, too, when faced with the
Frame Problem and its descendants. Therefore, a purely technical solution is not probable: concise study
of human mental capabilities and their development will
prove necessary.
References
Hayes, P.J. (1987). What the frame problem is and isn't , in Z.W. Pylyshyn (ed.), The Robot's
Dilemma, Norwood NJ: Ablex, pp. 123-137.
Minsky, M. (1988). Het denken: De menselijke geest als maatschappij. Amsterdam: Bert Bakker.
Toth, J.A. (1995). Book review of Reasoning agents in a Dynamic World: the Frame Problem (K.M.
Ford and P.J. Hayes, eds.), Artificial Intelligence,
73: 323-369.
Part III
The Frame Problem
Part III: Groundations
Peter Doomen - Master's Programme in AI
"…on pain of infinite regress, a symbol system cannot be the right model for what is going on
in our heads." (Harnad, p. 15)
"In the symbolic paradigm, the context of a symbol is manifest around it and consists of other
symbols; in the subsymbolic paradigm, the context of a
symbol is manifest inside it and consists of subsymbols." (Smolensky, 1988, p. 17)
Harnad rightly states that the symbolic approach in AI fails, not because of practical reasons (clearly,
some symbolic systems deliver very good results), but
because AI is more than problem solving. An important aspect of AI is that we be able to understand
natural cognitive systems with aid of artificial systems. The
frame problem and the related problem of symbol grounding are a manifest demonstration of the symbolic
failure. Therefore, mimicking aspects of human
cognition will prove to be important for theoretical AI.
In his excellent 1988 article, Paul Smolensky advocates what he calls the "subsymbolic approach"
to AI. In the subsymbolic paradigm, fine-grained constituents
of symbols are used as building blocks of cognition. As a good example serves the "rooms-experiment"
of Rumelhart and Smolensky. Several persons are asked
to describe rooms with aid of 40 features (e.g. contains coffee pot). This description is fed
into a neural network, which then behaves as if it had concepts for
different room-types. This behaviour is emergent: no one programmed it with room concepts. The
conclusion disarms the critique that frame-like things cannot
be done with neural networks (Harnad, p.14).
More than 80 percent of our brains serve processing of input or output information. Harnad calls these
"transducers"; for him, they play a major role in cognition.
Indeed, building up a model of our environment is what we most of the time do; in fact, this helps us
solve the frame problem. But still, no causal powers are
needed: transduction is done by our senses and primary brain parts. This constitutes no attack on the
strong AI view that our brain is a Universal Turing
Computer (which may be wrong—but not the way some opponents want it to). It simply shows that AI will
pass via the study of human knowledge and its
development.
References
Harnad, S. and others (1993). Grounding symbols in the analog world with neural nets: a hybrid
model. The symposium: connectionism versus symbolism,
special issue Think: ITK, 9-78.
Smolensky, P. (1988). On the proper treatment of connectionism. BBS, 11, 1-74.
Tip: the left pane works like windows explorer!
Please send your comments to Peter Doomen.
This document was updated 27/05/00. 
|