Subsumption summary

Subsumption summary

To summarize subsumption:
 Subsumption has a loose definition of behavior as a tight coupling of sensing
and acting. Although it is not a schema-theoretic architecture, it can

be described in those terms. It groups schema-like modules into layers of
competence, or abstract behaviors.
 Higher layersmay subsume and inhibit behaviors in lower layers, but behaviors
in lower layers are never rewritten or replaced. From a programming
standpoint, this may seem strange. However, it mimics biological
evolution. Recall that the fleeing behavior in frogs (Ch. 3)was actually the
result of two behaviors, one which always moved toward moving objects
and the other which actually suppressed that behavior when the object
was large.
 The design of layers and component behaviors for a subsumption implementation,
as with all behavioral design, is hard; it is more of an art than
a science. This is also true for all reactive architectures.
 There is nothing resembling a STRIPS-like plan in subsumption. Instead,
behaviors are released by the presence of stimulus in the environment.
 Subsumption solves the frame problem by eliminating the need to model
the world. It also doesn’t have to worry about the open world being
non-monotonic and having some sort of truth maintenance mechanism,
because the behaviors do not remember the past. There may be some
perceptual persistence leading to a fixed-action pattern type of behavior
(e.g., corridor following), but there is no mechanism which monitors for
changes in the environment. The behaviors simply respond to whatever
stimulus is in the environment.
 Perception is largely direct, using affordances. The releaser for a behavior
is almost always the percept for guiding the motor schema.
 Perception is ego-centric and distributed. In the wander (layer 2) example,
the sonar polar plot was relative to the robot. A new polar plot was
created with each update of the sensors. The polar plot was also available
to any process which needed it (shared global memory), allowing
user modules to be distributed. Output from perceptual schemas can be
shared with other layers.