Re: Total recall undo and redo!

[http://dummy.us.eu.org/robert (robert)]

 > From: Scott Schwartz <http://www.bio.cse.psu.edu/~schwartz>
 > Date: Wed  Apr  8,  1:45pm
 >
 > | Hi.  I was wondering: have you ever done anything with rk?  I use it every
 > | day and love it.  But I would like to improve its performance and accuracy in
 > | prediction.  (I have the book and have been going through it.)
 > 
 > Sure I've used it.  But I stopped around the time that I gave up using
 > xterm in favor of 9term, which does no cursor addressing.  The idea is
 > good, but to really improve it you'd have to rewrite from scratch, and
 > I never got around to that.  Part of the accuracy problem is that the
 > book talks about ways to handle trie overflow, but the sample code
 > doesn't do it---it just stops learning once the thing gets full.

Yeah.  That's just one problem.  (I was thinking of fixing that code.)

Another problem I don't like is that sequences are based purely upon frequency.
I'd also like to have sequences partially based upon recency of use (most-
recently-used).

Another problem is that (this may be a consequence of old nodes not
getting deleted) memory gets used up somewhat quickly.  I wish that rk
used a better, more compact representation for its model, or even another
model all together.  (I must say, the model in rk is very clever, but
there's still a great deal of redundancy in the resulting trees that might
be got rid of; I just don't know how at this point.)

John Darragh never seems to answer his e-mail, so I guess I'll have to do it
alone.

Are there any other papers you know about that talk about or improve upon
the model that Reactive Keyboard uses?