This week, I began exploration of the second major extention of the
Langley/Pfleger paper: that of feature selection. I don't forsee that
usage of this and other techniques to replace the current 
grid-matching paradigm.  Rather, it seems more likely to be used
complementary with grid-matching.  This allow us to address the main
concern of the current architecture, that of the rather computational-
intensive nature of all-translations grid-matching, while preserving
the fundmental probabilistic and cased-based nature of the architecture.
The following techniques can aid in either just in the registration phase
or can help to select only the best places for full consideration using
grid-matching. 

One quick idea to speed up matching with the case library is to somehow
merge the various instances together to form one "represenative" for each
place that can be quickly checked to get a rough estimate of recognition
within that location.  I am currently trying simple, non-probabilistic
techniques right now (such as averaging) just to see if the idea is
viable.  The Bayesian updating scheme of Elfes can probably be used
here (where instead of getting new sensor readings, we are now getting
more evidence grids).  My idea for such a system is one that would match
the incoming grid with each of the represenative of the places, then do
rigorous grid matching from among the most likely candidate for final
selection.

(Merging grids also have the added benefits of often having more information
than the grids by themselves.  Combining grids often eliminate gaps
(caused by angle of observations, obstacles, and by specular reflection)
that yields better description of the location.  This is analogous to the
improved performance observed by Langley/Pfleger with multiple sample
point than a stationary sonar sweep.)

The next idea hopes to simplify the all-translation matching scheme.  Current
work this week involved taking derivatives of grids (thus, highlighting
areas of large differentials of probabilities).  A variety of feature/
landmark recognition can then be used on the resulting grid.  A line-segment
approach (as suggested by Schiele and Crowley) might proved reasonable
here (though unlike S and C, we are not just using lines of just high
probabilites, we are using lines of high differentials of probabilities).

Whatever the ultimate algorithm, it should suggests several optimum
translations to try with the resulting grids.  This would greatly improve
on the computations currently done (a few translations versus all
possible translations).

The above two ideas can be combined.  We would first build our case-library
(supervised or unsupervised).  We would then build a represenative case
sample.  We then run our "derivitivize" processor on our represenatives.
With an incoming new grid, we "derivitivize" the sample and use
feature-recognizer to suggest good translations.  We then do quick matches
between the sample and the represenatives.  Then, from among the best
candidate, we can do more rigorous matching to pick the best place.  Hopefully,
all of the above will be faster than doing all-translations best matching
between the incoming grid and each of the instances of the case library.

The above architecture is also responsive to a dynamic environment.  The
combined grid represenative can slowly outvote outdated information.  The
feature selector can also slowly focuses on the true, unchanging features
of the room (such as walls and fixed furnitures).

Experiments and coding awaits...