Previous: Architecture levels
Up: Architecture levels
Next: The Knowledge Level
The three levels of our architecture are of organizational as well as
theoretical importance. Organizationally, the layered architecture allows
us to work on individual levels in a relatively independent manner,
although all levels are constrained by the nature of their interactions
with the adjoining level(s). The architecture is hierarchical, in that
level can only communicate with levels
and
, if any.
The levels of our architecture are semi-independent. While control flows mainly top-down and data mainly bottom-up, local control mechanisms at any level can preempt higher-level control, and these local mechanisms filter the data stream for their own purpose, in parallel with higher-level ones. Representations become coarser-grained from bottom to top, while control data becomes more fine-grained from top to bottom. The terms in the Knowledge Level's KRR system model conscious awareness of the world (and the body), and the perception and motor capabilities in the other levels provide the grounding for an embodied semantics of the former. Routine, reflex-like activities are controlled by close coupling of perception with motor actions at the (unconscious) Perceptuo-Motor and Sensori-Actuator levels. This close coupling avoids having to exert control over these activities from the conscious level, as in purely top-down structured architectures with a symbol level at the top of the hierarchy. In the latter kind of system, signals must first be transformed to symbols and vice versa. The low-level coupling provides for better real-time performance capabilities, and relieves the Knowledge level of unnecessary work.
In general, we have multi-level layered representations of objects,
properties, events, states of affairs, and motor capabilities, and the
various levels are aligned. By alignment we mean a correspondence
between representations of an entity at different levels. This organization
contributes to the robustness and computational efficiency of
implementations. The semi-autonomous nature of the levels allows for
graceful degradation of system performance in case of component failure or
situation-dependent incapacitatedness. Lower levels can function to some
extent without higher-level control, and higher levels can function to some
extent without lower-level input.
Our architecture allows us to elegantly model a wide range of
behaviors: from mindless, spontaneous, reflex-like, and automatic
behavior, e.g., ``stop if you hit an obstacle'', to plan-following,
rational, incremental, and monitored behavior, e.g., ``Get in the car
now, if you want to go to LA on Friday''.
In anthropomorphic terms, we identify the Knowledge level with consciously accessible data and processing; the Perceptuo-Motor level with ``hardwired'', not consciously accessible processing and data involved with motor control and perceptual processing; and the Sensori-Actuator level with the lowest-level muscular and sensor control, also not consciously accessible. The substrate of grounding and embodiment [Suchman 1988][Lakoff 1987][Harnad 1990] of actions, concepts, and reasoning is mainly the Perceptuo-Motor level and to some extent the Sensori-Actuator level.
We will now explore representation and computation at the individual levels in more detail.