Icarus - Labrynth. (Langly, P)
Architecture
Labrynth, the heirarchy of probabilistic concepts, stores
events,
plans,
actions, and physical objects by creating a 'concept' and
classifying it according to its similarity to pre existing, abstract concept
classes. If there are no sufficiently similar classes, the new concept is
abstracted and a new class created for it. Concepts are classified,
within a class, according to an abstraction heirarcy, where abstract concepts
are the interior nodes and specific concepts are the exterior nodes. This
process of classification is also Icarus's principle
learningmethod.
The definition of a concept depends upon what
type of knowledge is
being represented. If the knowledge is about a physical object, then the
object is parsed into component parts (i.e. a table may be parsed into
four legs and a flat top). There also then exists composite concepts, where
roles within that concepts represent conposite parts of an object and have
associated with them the probabilitiy that that part is included in the
overall object. So, for example, a composite concepts of the table would
include a role for a leg plus the probability that something was a table also
had a leg. Actions, events and plans, however, are classified as qualitative
states. Qualitative states consist of a set of objects (which are classified
as above) and an interval of time during which the qualitative aspects of the
state changes in a fixed direction. Labrynth classifies problems as paris of
qualitative states, represendting the goal and final states, and it classifies
plans as a problem, an operator, an initial subplan, and a final subplan or
problem solving trace.
Capabilities
Labryinth, along with its responsibilities of
storing knowledge
and providing for
intermodule communication, is also responsible for all the
learning
that occurrs in Icarus. Labrynth learns as it classifies new concepts
- adjusting the probabilities of the entire class. Once a concept is
classified, it can
learn by generalizing more abstract concepts from a
set of concepts. Labrynth can also learn by keeping statistics on the
probability that a given subplan or operator will, first, prove useful in
solving analogous subplans, second be abandoned during the planning phase (the
chance of having to backtrack) and thirdly, actually have the result, once
executed by Meander, that was intended. Labrynth is able to accomplish this
through feedback from
Argus,
Daedalus, and
Meander. The second and third
statistics allow Icarus to operate in
open loop mode, because Meander can
begin actions before a plan is complete if there is little chance
of backtracking,and if there is little chance of the expected action being
violated, Argus need not pay attention to the results of Meanders actions.