Capabilities of Subsumption Architectures
Capabilities
of Subsumption Architectures
The basic subsumption architecture was designed with elementary capabilities
in mind. There is a
simple navigation scheme
for moving around in the
world either randomly or in a specific, exploratory direction while
avoiding objects.
Brooks describes three capabilities which were important in the original
conception and design of the subsumption architecture which were addressed
in the
hormonal activation modification he
proposed:
Mataric describes a modification to the subsumption architecture that
allows an agent a rudimentary
learning
skill. The agent learns the spatial
structure of its environment by noting landmarks. Additional information
on this topic is discussed under
ability to learn.
An elementary
planning
capability exists within the
modified subsumption agent as well. A goal location is selected by
a user and the agent makes its way to the location based on a activation
gradient which decreases according to an inverse square law. By following
paths along increasing activation level, the agent takes the shortest
path to the goal location.
The subsumption architectures are not (currently) capable of any of the
following activities:
high-level learning,
natural language understanding,
explaining the reasons
for behavior,
coordination and
cooperation with other agents and many others.
A subsumption agent should exhibit a
coherence
of actions and goals when
observed. This means that is should not constantly switch between
alternate goals nor should two competing or interfering behaviors be active
at once. The damping effect of the hormonal system (mentioned in
connection with reactivity addresses this
capability.
Salience is a capability deriving from the architecture's
rationality, implying that the agent
undertakes appropriate behaviors. For example, the agent should not seek
to recharge its power supply when the system is fully charged.
Adequacy is achieved when all the behaviors appropriate for some task are
completed in task specific order. For example, to clean a floor, an agent
should take the mop and bucket out of the janitor's supply closet before
attempting to mop; however, the order in which the mop and bucket are taken is
inconsequential. Adequacy is generally pre-programmed (e.g., see
taskability).
To return, press HOME.
To go to the next document, press NEXT.