Previous: Alignment Up: The GLAIR Architecture

Consciousness

As we pointed out above, we identify the Knowledge level with consciously accessible data and processing; the Perceptuo-Motor level with ``hard-wired'', 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 distinction of conscious (Knowledge) levels vs. unconscious (Perceptuo-Motor and Sensori-Actuator) levels is convenient as an anthropomorphic metaphor, as it allows us to separate explicitly represented and reasoned about knowledge from implicitly represented and processed knowledge. This corresponds grosso modo to consciously accessible and not consciously accessible knowledge for people. Although we are aware of the pitfalls of introspection, this provides us with a rule of thumb for assigning knowledge (and skills, behaviors, etc.) to the various levels of the architecture. We believe that our organization is to some extent psychologically relevant, although we have not yet undertaken any experimental investigations in this respect. The real test for our architecture is its usefulness in applications to physical (robotic) autonomous agents (section ).

Knowledge in GLAIR can migrate from conscious to unconscious levels. In [Hexmoor et al. 1993a] we show how a video-game playing agent learns how to dynamically ``compile'' a game playing strategy that is initially formulated as explicit reasoning rules at the Knowledge level into an implicit form of knowledge at the Perceptuo-Motor level, a Perceptuo-Motor Automaton (PMA).

There are also clear computational advantages to our architectural organization. A Knowledge Representation and Reasoning system as used for the conscious Knowledge level is by its very nature slow and requires lots of computational resources. The implementation mechanisms we use for the unconscious levels, such as PMAs, are much faster and require much less resources. Since the three levels of our architecture are semi-independent, they can be implemented in a (coarse-grained) parallel distributed fashion; at least each level may be implemented on distinct hardware, and even separate mechanisms within the levels (such as individual reflex behaviors) may be. Our Robot Waiter agent, for instance, uses distinct hardware for the three levels (section ).