A diagram of this architecture
Environment
Capabilities
Properties
Issues
The design
of Icarus was motivated by the goal of developing a
psychologically plausable
information processing system capable of functioning
in a
complex, physical environment.
Therefore many design choices were made on
the basis of greater psychological validity. This is especially true of the
choice to explicitly
organize memory
as a heirarcy of probabilistic concepts,
which allows for
knowledge inconsistency
and produces phenomena such as
forgetting ,
learning,
basic-level effects
(Gluck & Corter 1985),
typicality effects (Rosch & Mervis, 1975) and fan
effects (Anderson, 1975).
Icarus' architecture
is designed around a specific representation of
long term memory.
Icarus represents all knowledge in a
heirarcy of probabilistic concepts.
This heirarchy is encapsulated into an independant
, asynchronous module called
Labrynth. The architecture also includes three
other independent, asynchronous modules responsible for
perception,
planning
and
effecting -- known as
Argus,
Daedalus, and
Meander respectively.
Labrynth plays the central role in the architecture. It is not only
responsible for long term knowledge, but it acts as the
central communication
facility for Icarus' other modules. Labrynth contains an area, known as the
active memory, through which each module can place requests for information
from long term memory, and read and operate on the results. The flow of
information is best described by following some given perception by Argus
through to the resulting reaction by Meander, relating how each component is
involved in creating the overall behaviour.
Argus, the perceptual component, parces the environment into
qualitative states
which it posts in the active memory area of Labrynth.
Labrynth is continually monitoring the active area and attemps to classify
all new perceptions. If the resulting classificaion is associated with a
problem, then Labrynth poses the problem as a pair of qualitative states and
places it, with an
associated priority,
back into the active memory. Daedalus
, meanwhile, may be working on solving a previously posed problem, but is also
continually monitoring the active memory. If this new problem is associated
with a higher priority than the one it is currently working on, Daedalus will
interrupt its planning and begin work on the new problem, placing all results
in active memory. Once a solution is complete, Daedalus can resume work on
the previous problem. Meander, all along, has been monitoring active memory
and has the
ability to begin executing partial plans if it believes that
the probability that Daedalus will not have to backtrack is low (as learned by
Labrynth. Argus, then
has the ability to
complete the loop by montitoring those aspects of the
environment that Meander is effecting, and, if the perceptions do not
sufficiently match the expected results, posing a new observation which
Labrynth can attempt to classify and create a new problem for Daedalus to work
on.
In Icarus
, knowledge is not represented
uniformly. For example,
Daedalus can not explictly create a plan to change the way Argus is performing the event
parsing of the environment. The knowledge on how that is done is seperate. Similaraly,
the control knowledge about how to keep statistics of utility (Labrynth)
of concepts is unaccessable to the rest of the system. Furthermore, the choice of
a
heirarchy of probablistic concepts as an explicit
memory structure
limits accessto
the
global memory
by any individual component to
only that knowledge which is requested. Threrefore rules can only examine other rules
if they already know about, which limits the inferental ability of the system.
Icarus, can, on the other hand, manage
inconsistant knowledge.
For example, if icarus were to learn that , first, tables had legs, and then, secondly
that tables do not have legs, then (assuming these were the only two facts it was
exposed to about tables and legs), Icraus would believe that there was a 50% chance of
a table having legs. Furthermore, the more 'tables with legs' it was exposed to, it would
learn
that there was a better chance of a table having legs than not. Therefore, it can
learn
contradictory knowledge and has the
intersting property
of not replacing the old knowledge but empirically determing which is best and
eventually forgetting the other. This shows that the
learning mechanismsin Icarus are
reflexive because learning
is always occurring, but
deliberative
because Icarus reasons (via probabilities) about
which knowledge to forget.
Icarus does have the ability to be programmed with
new knowledge,by simply specifiying a background heirarchy to Labrynth. This
heirarcy is also beleived to be very general, in that it does not presuppose
a specific type of knowledge. Taskability is also possible by simply specifying
a perception, as Argus would have, to Labrynth for it to classify and create
a problem for.
The Icarus
architecture is a knowldge based architecture capable
of learning
(Labrynth), planning
(Daedalus) along with sensing and acting
(Argus and Meander).
The combination of these components are able to produce
a varitey of behaviours for dealing with a
complex environment. It can perform
both
single minded and distractable planning
by adjusting the
priorities
placed on certain problems depending on their
expected outcome.
(As learned by
Labrynth)
This same property, along with Argus's ability to
focus attentionallows Icarus to
react to emergencies noticed in the environment.
It has the ability to remember
past plans
and use them again, and the ability to
forget them
if it turns out they aren't of much utility.It also is able to dynamically
increase effeciency by
interleaving planning and action when plans
are expected to be
sufficienlty stable.
Icarus was
designed with the goals of creating a rational agent capable
of
navigating through and
manipulatingobjets in a
complex
physical environment .
The architecture was designed to be
interruptable, to a degree,
through its
priortizing of problems and can therefore deal with unpredictable
emergency situations well. The processing
speed required
combinatorial natureof such an environment
is accounted for in two ways. First, the perception and storgae of actions and
events as qualitative states effectively and concisely represents large
quantities of knowledge, and allows for efficient processing of such knowledge.
And, secondly, efficient retrieval of such knowledge is achieved through the
explicit use of a probabilistic heirarchial structure. The authors also claim
that some degree of physological validity is furter proof of Icarus' ability
to cope in a complex, dynamic environment.
Icarus,
at this time, is still in modular form, the pieces have not
been integrated as described above. This exhibits one of the possible
drawbacks to a multi component
design. Many bottlenecks
occurred in the integration of the components.
Other Architectures
Back to the Title Page