[Martin Taylor 930318 16:00]
(Rick Marken 930316.2000)
I really didn't understand your description of the study. But after
reading this far I am getting the distinct impression that the
control systems you plan to study are simulations.
Not really. The end (far distant) goal is an "intelligent" computer
interface. If it works, a byproduct may be better understanding of
human use of language.
Why not do a
study with real control systems -- like people? If you are intending
to study people then I need some clarification of the vector
representation of the letters and all that.
Yes, so would I. In my mind (and certainly not in any way as a simulation)
the vectors represent the degree to which the symbols have attributes.
If we were simulating a recognizer of printed words, the features might
be letters contained, plus such things as length, porportional position
of ascenders and descenders, location of desnity maxima of marks on the
page, and so forth. If it were spoken words, we might have spectral
vectors and their time derivatives (or any of a host of other features).
If it were concepts, the attributes might be featheriness, weight,
linearity, patchiness of colour, ... In the experiment we propose, our
intention is to assign letters arbitrarily to locations in the vector
space (we are going to cheat, though, and mentally group the letters
into sets such as vowels, liquids, stops, voiced ... and colocate them
according to our mental groupings. The control system won't know anything
about that, though it may learn).
I sure hope you are planning
to do this with living control systems. What we really need in PCT
is good data; who cares how a particular implementation of an algorithm
works? Unless your working in AI or Artificial Life, of course, in which
case you only care about the behavior of the model itself.
I agree with the need, but that's not what we are doing. As I said,
the ultimate goal is a human factors one, but the immediate goal is to
study the properties of different forms of reorganization, to see whether
simple control system can learn to control under the artificial world
conditions to which they will be exposed.
You may not have noticed, but the task world is that of the Little Man,
and the letter symbols simply drive the cursor to be tracked by the finger.
Can a Little Man that reorganizes (a.k.a. the Little Baby) learn where
the finger is likely to have to go if the motions of the tracking cursor
are directed by a synthetic grammar? Will it learn to predict the letter
group before it learns to predict the actual letter (or over)? The latter
question is tantamount to asking whether a control hierarchy will (as I
think it will) learn to paraphrase.
···
------------
In respect of your proposed study, I am reminded of a peculiar result
obtained by somebody here when I was a summer student around 1958. So far
as I know, it was never published because no-one knew what to do with it.
The experiment was very simple. There were 25 slides, 23 of which showed
in big letters ABDCE (or some such). Two showed ADBCE. These were placed
18th and 23rd (roughly) in the sequence of 25. The subjects had to report
for each slide the sequence of letters. Easy? No subject reported the
deviant sequence at slide 18. All subjects reported the deviant sequence
at slide 23. When asked afterward whether they had noticed any other
deviant slide than 23, no-one acknowledged having done so. But I'll bet
they would have noticed if slide 18 had GVLYK. I'm not sure if this
relates to your result of faster control when the letters change than
when the sequence changes, but it has lain in the back of my mind for
many years as something needing explanation.
Martin