Compromised actions are enabled in the behavior system via the use of
motor preferences. This design is considerably more efficient than
that of a free-flow hierarchy [Tyrrell1993b]. In a free-flow
hierarchy, every behavior must compute a recommendation for the next
action. There will be no winner until all of the recommendations
arrive at the motor command level, where a winning action is decided
by fusing all the recommendations (see Appendix 
for
more detail about free-flow hierarchy). While a free-flow hierarchy
pays no special attention to the most desirable behavior given the
current situation, our design of the intention generator takes the
opposite approach. It picks a winning behavior at an early stage such
that irrelevant sensory information need not be processed. This
reflects the advantages of a winner-takes-all selection
process. However, unlike a conventional winner-takes-all process, the
use of the motor preferences allows unselected behaviors to influence
how the winning behavior is performed. We believe that our
implementation is a good compromise between efficiency and
functionality.
Moreover, the qualitative nature of the motor preferences simplifies
both the process of generating them and the way of using them (see
Section ). This is in direct contrast to the
quantitative action recommendations used in Tyrrell's implementation.
Merging the recommendations into a single motor command was shown by
Tyrrell to be a complicated and difficult problem even when all the
actions are specified in a 2D discrete world. It will thus be even
more problematic if the actions are defined in a 3D continuous world.
In addition, it is doubtful that most animals make their decisions to
act by carefully evaluating and optimizing all possible actions. For
this reason, using qualitative motor preferences seems to be more
biologically plausible.
| Xiaoyuan Tu | January 1996 |