Goal Reconstruction in the Teton Architecture
Goal Reconstruction in the Teton Architecture
Teton's
architecture
includes features that enable the blending of situated and planned
action. The architecture is based on a
production system.
Declarative knowledge is
stored in the
working memory.
Procedural knowledge in the form of operator
specifications and control rules is stored in the
knowledge base. Two modifications to
this basic design facilitate goal reconstruction:
operator shortcuts and the
situation in working memory.
The authors show how the combination of 2 things, a
truth maintenance
system (TMS) and reconstruction,
complement each other in maintaining the goal structure. A TMS can be used to
maintain a goal structure. A goal structure consists of a top-level goal,
and the
subgoals (not necessarily in a stack) that can be inferred from top-level
goal, and the
current state of the environment.
If the state changes, the derived goals change. So the
2 ways of updating the derived goals are:
- Reconstruction: remove all subgoals and rederive them. This is relatively
slow, but does not require that any of the derived subgoals be remembered.
- TMS: keep track of the derivations of all of the subgoals,
and only update
what needs to be updated. This is faster, but requires
that the derived subgoals
be remembered.
The idea is to use the fastest method, which depends on whether or not all of the
subgoals are remembered. Since
short-term memory is supposedly limited, it is
entirely conceivable that many of the subgoals will be forgotten.
There is nothing
inherently LIFO about the derived subgoals.
The authors did not adequately explain
how situated and planned action are blended.
Return to the top of this architecture.
Go to a discussion of this capability
for multiple architectures.
Current Location: Teton - Capability - Goal Reconstruction