VMs: Would it be possible to summarize it for the list?

["Ben Preece" <ben.preece@xxxxxxxxxx>]

The app is pretty straightforward.  It runs a text through a
probablistic FSM and measures the average information rate of the text.
Presumably, the lower the information rate, the better the FSM fits the
text.  It then makes random changes and compares results.  If the
modified FSM fits better, it throws out the old FSM and uses the new
one.  Lather, rinse, repeat.

It's been a while since I've had time to work with it, so I'll need to
go back and find my papers.  I think I remember that on a Latin text, a
20-state FSM generated by the app produces text which is recognizably
trying to imitate latin, while a 50-state FSM actually starts to be a
decent imitation.  On English text, I think 20 states weren't enough to
be very recognizable.  On Hawaiian text, two states did a pretty good
superficial imitation.  I've tried it on a Romanized Chinese, but don't
have the skills to know whether the result is any good or not.

The problems with the app are the ones you expect with evolutionary
programming, like getting stuck in local minimums, or knowing how fast
to converge and how long to run.

I'll try to find my stuff on the VMs and write something up about it.
One interesting thing I remember is that some letters, like 'a' and 'o',
tend to end up with their own states;  no matter what came before the
'a', the 'a' will always take you to the same next state.  This means
that what comes next after the 'a' depends only on the fact that it
comes after the 'a', and not on anything that came before the 'a'.  The
letters 'a' and 'o' have different states, and these are not the same
state that indicates the start of a word.  So how would one interpret
all that linguistically?

-Ben
 

-----Original Message-----
From: Koontz John E [mailto:John.Koontz@xxxxxxxxxxxx] 
Sent: Wednesday, March 02, 2005 5:24 PM
To: Ben Preece
Subject: Re: VMs: RE: VMs word models --> state machines...?

Would it be possible to summarize it for the list?

John E. Koontz
http://spot.colorado.edu/~koontz


______________________________________________________________________
To unsubscribe, send mail to majordomo@xxxxxxxxxxx with a body saying:
unsubscribe vms-list