Received: from spot.Colorado.EDU by karazm.math.UH.EDU with SMTP id AA09295
  (5.65c/IDA-1.4.4 for <glove-list@karazm.math.uh.edu>); Tue, 15 Oct 1991 12:30:01 -0500
Received: by spot.Colorado.EDU id AA24142
  (5.65b+/IDA-1.4.3/CNS-2.0 for glove-list@karazm.math.uh.edu); Tue, 15 Oct 91 11:26:04 -0600
Date: Tue, 15 Oct 91 11:26:04 -0600
From: SCHILDBACH WOLFGANG <schildba@spot.Colorado.EDU>
Message-Id: <9110151726.AA24142@spot.Colorado.EDU>
To: glove-list@karazm.math.uh.edu
Subject: LPC vs. polynomial predicting
Is it really a good idea to use Linear Predictive Coding for the glove?
From what I understand LPC works with fourier coefficients of the data
given, or, in other words, it tries to model the data as a sum of sines
and cosines. Now I would guess that most movements you make are not sinu-
soidal but rather polynomial (i.e. the velocity should go like a parabola).
(Of course, some research has to be done at that, but it shouldn't be to
hard for those of you who have gloves to write a little hack that records
your glove movements on time and display this data.)
	Now if the velocity is mainly polynomial, wouldn't it be better
to model it as a polynomial?

	For references on Kalman filters, LPC etc. try

Teukolsky, Vetterling, Press, Flannery: Numerical Recipes in (C|FORTRAN|PASCAL)

which is technical (the matter is, after all) but easy to understand. It
contains a lot of sources that might help.

	For filtering the glove output, I would suggest what I would call a
"plausibility filter": If the (pos,vel) measurement at time 0 is (x0,v0) and
time t is x1, solve x1=x0+v0 t+1/2 a t^2 for a (the acceleration) and check if
it gives a reasonable value. If not, reject the measurement and predict it in-
stead.

Ciao
 Wolfgang