Mystery Number Four - Computational Linguistics and Talking Robots - page 86

Subjunctive transfer is also essential for constructing the agent's discourse
model in dialogue (cf. Dominey and Warneken, in press). Consider, for exam-
ple, two agents discussing going on a vacation together. Their respective dis-
course models include not only knowledge of the subject matter, e.g., where
to go, how to get there, where to stay, what to do there, how to pay for it, etc.,
but also knowledge of the other agent. While knowledge of the subject mat-
ter is subactivated by means of intersection, knowledge of the other agent's
cognitive state is based also on subjunctive transfer.
Even though subjunctive transfer will provide maximum information about
another agent's viewpoints, feelings, tolerances, preferences, dislikes, de-
mands, etc., it may not be enough. For example, agent A may inadvertently
hit a sensitive spot of B, hitherto unknown and with explosive consequences.
Thus, goal-driven behavior is much less predictable than data-driven behav-
ior - though both use the same inference chain, e.g., 5.2.1.
In summary, instinctive, habitual, or rote behavior provided by data-driven
inferencing is the same over and over again and is therefore predictable in
another agent. Another agent's goal-driven behavior, 20 in contrast, is unpre-
dictable because it depends on data stored in the other agent's memory and
therefore not directly accessible to the partner in discourse.
The DBS inferencing described is continuously triggered by subactivation
and intersection, which in turn is activated by current language and nonlan-
guage recognition. The result is a stream of new inference data written to the
now front - in addition to current language and nonlanguage recognition. This
raises the more general question of how to avoid overflow of the agents' large,
but finite memory.
Given that a content-addressable memory cannot be changed, incoming data
should be selected and cleaned up prior to storage. Following the natural pro-
totype, this may be done by providing a short-term and a long-term memory,
and by doing selection and cleanup in short-term memory prior to long-term
storage. For example, subjunctive and imperative action sequences may be
derived in short-term memory, but only the imperative sequences need to be
remembered long-term.
The other method to stave off memory overflow, also following the natural
prototype, is forgetting. Though a limiting case of data change (and therefore
20 Goal-driven behavior is important to nouvelle AI in general (cf. Braitenberg 1984, Brooks 1991) and
BEAM robotics (for B iology, E lectronics, A esthetics, M echanics) in particular (cf. Tilden and Has-
slacher 1996). However, while BEAM has the goal to model goal driven behavior with the simplest
means possible, based on analogical sensors without microprocessors, DBS uses representations of
content in a database. This is because DBS starts out from natural language, while BEAM robotics
proceeds from insect behavior.