Re: VMs: Another method different from Cardano Grilles

[Jacques Guy <jguy@xxxxxxxxxxxxxxxx>]

26/01/2005 2:21:43 PM, "Marcelo Dos Santos" <mdossantos@xxxxxxxxx> wrote:

>Instead of Cardano Grilles, Violat describes there the use of three
>concentric cardboard circles, inspired by the three circles surrounding the
>Moon´s face as seen on f67r. This cardoard circles carry the prefixes,
>midfixes and suffixes (one kind on each circle). Nailed by the center and
>allowed to spin freely, you can generate large amounts of voynichese words,
>far easier and faster than Cardano Grilles would allow.

Obviously. And it is a far likelier theory. But it is only
that: a theory. I know that the concept is not easy to grasp 
to anyone but seasoned statisticians or gamblers, but I will try.

>He claims he could reproduce, with this simple method, the whole text on
>f67r.

I can reproduce the whole text on f67r with a much simpler wheel.
A single wheel bearing the "letters" of EVA (or those of Frogguy).
But consider: how many OTHER texts will spinning that wheel produce?

Now, imagine that I have refined my machine by using three wheels 
instead of one, each wheel with, not single letters, but clusters 
of letters drawn from the analysis of the VMS, just like Jorge Stolfi 
did, three wheels like Violat's.

This new machine allows me (with a great deal of luck) to reproduce
f65r, with far less garbage than my earlier one-wheel method.

I can refine it to produce only the words of f65r, not only
"only those words" but even only those words in their
frequency of occurrence on f65r. There are several ways of
doing this. A trivial one is to have one big wheel with
the "words" instead of individual letters, each word 
occurring as many times as it is seen to occur in f65r.
 
And that is called "overfitting the data".

"Overfitting the data" is what you do when your predictive
method fails and you keep refining it until it succeeds.

Overfitting the data happens when the quantity of 
information (in the mathematical sense of Shannon's 
information theory) in your predictive machine (your
spinning wheels for instance) converges with the
amount of information in the data set which the
machine tries to reproduce. 

>and, adding more circles (up to five) he says he could reconstruct the
>whole 40,000 words in the VMs.

Voilà! Overfitting the data. But not good enough yet, by far. 
For how many words which DO NOT occur will those five wheels 
reconstruct? 



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