Mystery Number Five - Computational Linguistics and Talking Robots - page 96

the LA-rec1 grammar will also handle any extensions of the variable definition
(requiring concomitant extensions of (i) the agent's hardware for recognition
and/or action and (ii) the memory, but not of the LA-rec1 grammar).
LA-act and LA-rec grammars fit into the component structure of diagram
4.5.3. Both receive input from the I/O component, and both are part of the rule
component.
An LA-act grammar receives a stimulus as input and selectively activates
proplets in the Word Bank, copies of which are passed as blueprints to the I/O
component for realization. 9 An LA-rec grammar receives guided patterns from
the I/O component, uses them for lexical lookup in the Word Bank, and con-
nects the sequence of lexical proplets into stimulus-response pairs 10 - ready to
be activated by the co-designed LA-act grammar.
Intuitively, the meaning of the concepts red, green, blue, strgt, left, and
right equals that of the English words, used as placeholders. Procedurally,
the concepts have straightforward implementations in terms of artificial vision
(measuring electromagnetic frequencies) and locomotion. The intuitive and
the procedural characterization of concepts are equally necessary to ensure
that the procedures are attached to the correct placeholders used as handles. 11
The agent-internal representation of the external world by means of these
concepts is extremely sparse - in concord with the tenets of subsumption ar-
chitecture in robotics (Brooks 1991). For example, a fixed behavior robot try-
ing to execute the sequence 6.2.2 in rough terrain could not possibly model this
terrain, given the limits of its recognition. Instead the realization of the motion
steps is left to a loosely coupled, massively parallel, analog walking machine
as described by Tilden and Hasslacher (1994). This machine is subsymbolic
because it uses “digital pulse trains ... for motor drive and control” (op. cit.).
For higher-level reasoning, such subsymbolic procedures must be related to
symbolic ones. Thereby, nouvelle AI proceeds from the subsymbolic to the
symbolic, while DBS proceeds from the symbolic to the subsymbolic. For
example, the fixed behavior agent outlined above is symbolic, but assumes a
subsymbolic, procedural realization of the core values in terms of the agent's
elementary recognition and action procedures.
For learning, however, the relevant transition is not from the symbolic to
the subsymbolic or vice versa, but from fixed behavior to adaptive behavior.
In contradistinction to expanding the repertoire of a fixed behavior agent by
means of guided patterns (provided by a scientist), adaptation and learning
must be autonomous, driven by automatic appraisal and schema derivation.
9 This is analogous to 4.6.2.
10 This is analogous to 4.6.1.