Conclusion - Computational Linguistics and Talking Robots - page 245

or regaining balance. 28 Which blueprint for action is selected for realization
is determined by continuously recalculating the weights associated with the
agent's current needs and the consequences of available options. The calcula-
tion is based on a cost-benefit analysis, computed in real-time.
Fine-tuning the DBS system to simulate the balance of a robot in a terrain,
i.e., in a co-designed changing environment, requires an actual robot. After all,
without the robot's external and internal interfaces we would have to recreate
every nook and cranny of the changing environment by hand (as in Model
Theory, cf. 4.3.1). This would violate a basic principle of nouvelle A.I., namely
that The world is its own best model (Brooks 1989).
To manage the massive, multiple, parallel search required for operating a
DBS robot in real time, retrieval must be based on efficient database operations
and provide highly differentiated recall. 29 This task may and must be solved
theoretically, i.e., without any need for actual robot hardware. The following
aspects may be distinguished: (i) the organization of competing retrieval tasks
at any given moment and (ii) the quality of the search mechanism itself.
For the synchronization of competing retrieval tasks, DBS uses the time-
linear derivation order (cf. Herlihy and Shavit, in press, for related issues).
All parallel derivation strands apply their current step, e.g., a rule application,
simultaneously. An example of time-linear derivation strands running in par-
allel is the simultaneous operation of (i) the hear mode, (ii) subactivation, (iii)
intersection, and (iv) inferencing.
In a step, each strand produces a set of retrieval tasks, specified by the DBS
rules and database operations to be applied. These jobs may be executed in
parallel or sequentially in some suitable order, depending on the hardware
and the operating system. This approach to organizing parallel operations is
not only efficient but also transparent - which is essential for debugging and
optimization of the DBS robot.
The other aspect of optimal retrieval is the quality of the individual search
operations. In DBS, it (i) depends on the speed of the retrieval mechanism
and (ii) must have the expressive power needed for the kind of queries to be
expected. For speed, DBS uses the schema of a content-addressable database
(Word Bank) and the use of pointers. For expressive power, DBS utilizes the
28 As shown in Chaps. 5 and 6, this is based in part on rule-governed and goal-governed inferencing,
fixed behavior, and trial and error.
29 In the hear mode, retrieval is used to determine the correct token line for storing proplets at the now
front . In the think mode, retrieval must activate a successor proplet somewhere in the Word Bank,
using an address. In the speak mode, the content to be realized must be mapped into a sequence
of language surfaces, written to the now front , which again requires finding the proper token line.
Activated content must find inferences with a matching antecedent.