Issues Deriving from Soar as a Unified Theory of Cognition
There are many issues related to the Soar
architecture's proposal as a candidate
unified theory of
cognition. Since a discussion of the intricacies of the proposal
is beyond the scope of this document, specific issues related to these
details are also avoided here. However, two simple points are made
to demonstrate that
Soar as a unified theory of cognition is
incomplete.
-
Unbounded Short-Term Memory
In Soar, working memory is analogous
to short-term memory in humans. The size of this short-term memory has
been estimated to be between five and seven symbols (or chunks). In Soar,
the size of working memory is unbounded;
elaboration adds
working memory elements which remain until the
context slot for which they
were generated is removed. Thus these elements also reside in working
memory for an indeterminate amount of time, which is also psychologically
implausible. A related discussion regards the lack of a
goal reconstruction mechanism in the Soar
architecture.
-
Deductive Learning Mechanism
Chunking is directly analogous to
explanation-based learning, a
method which is
deductive in
nature. Inductive learning has been demonstrated in Soar (see, for
example, Simple Concept Acquisition)
but it is not the bottom-level mechanism and requires mechanisms on top
of the system to realize induction. Although not proven, such a mechanism
seems counter to human ease in induction (so easy that erroneous
inductions are often made). Since Soar proposes a
single, uniform mechanism for
learning, a decision between an inductive and deductive approach had to
be taken and the latter, as chunking, was chosen.
Press
UP to go to the list of issues.
To return, press HOME.