This week, I continued experimentation on my classification routines to
generate more comprehensive results.  Things are still running the background
(it takes a while, but more on this later).  In short, I am running my
classifier at 55%, 60%, and 65% match threshold on the various grids
(noise levels 0-45%, and real robot results).  I will complete a more
comprehensive result chart once all the data is back (the one for last
week is only a short summary).

The next level of experiments that I want to get into is to attempt to
combine the various grids (classified by the classifier) into a "mega-grid"
for the "place" for the follwoing reasons
1 - Combining the grids may yield a grid with more information that the
    grids in part (i.e., each grid may only share a part of a region,
    a region that the robot may currently is in and has passed in a local
    occupancy grid for comparison).
2 - May ease registration/computational issues (don't know on this one yet,
    may just have to try and see)
3 - (The biggest reason...)  It will allow the subsequent level of experiments
    when we have to deal with a dynamic environment.  Once things start
    moving around, it would be much easier to update the one "mega-grid"
    rather than worrying about the diverse temporal issues of multiple
    grids.

But combining the grids also have its share of problems:
1 - Knowing when to stop - since some (maybe all) the grids are connected
    one could conceivably continually "combine" grids to form one entire
    grid for the entire environment, which is clearly not what we want.
    I am assuming that there exists enough discontinuities with the narrow
    doorway that this won't be a problem, but one never knows. If that is not
    so, we may need a method to determine a "place."
2 - What is the grid-combination scheme - A simple scheme exists in my mind
    right now, (its more intuitive than mathematical)
	(1)  We combine the grids serially; i.e. we take the first two grids,
   	     combine them, then pick up the third grid, etc.
	(2)  If one grid has info on a certain spot whereas the other grid
 	     does not, we use the info from the grid with info.
	(3)  If both grids have no info on a certain spot, we leave it 0.5.
	(4)  If the two grids disagree about the info on a certain spot we
	     could either (a) average the values (b) we leave it 0.5 (i.e.
 	     to be settled later... but this would favor later grids)
    I like the average idea better in that we slowly squeeze out noise with
    increasing evidence.  A lot of the literature talks about a Bayesian
    Estimation scheme to combine grids, but I have trouble finding the
    exact methods that they are using.  Maybe we could resolve this issue
    when we meet this week.
3 - How does errors play into this - a miscalculation in the previous
    experiment and the prevous stage can be noted in slowing the learning
    curve or increasing the error rate.  An error here, in attempting to
    combine an erronously categorized grid may be fairly detrimentral.
    While over the long run, errors might be "averaged" out, it also might
    take a very long while to recover.  Another issue to leave to experiment.
    (This may also mean that we require a much stricter error tolerance
    for phase one, in that we would rather have different classes for the
    same place rather than have erronous places within the same class.)

A side issue that has come up is that the classification scheme could
potentially take a long time to run.  (And it gets worst as the class library
gets bigger an bigger.)  This might not be a problem once we start combining
grids.  I will see how Mehran is doing in some of his speedier techniques
at registration (courser registrations and neural networks).

That's all for right now.  I guess we will more clearly flesh out issues
when we meet later this week.