I am not a diehard PCT guy (have not set to really exploring the
possibilities, yet) but I can see the imediate advantages in control theory.
Recently I looked over the ??? notes and saw a possible connection to other
ideas circulating arround.
from Avery Andrews 920818
The driving story illustrates nicely what I at the moment consider to be
the biggest potential problem for PCT (ignorance of how perception works
is of course another big problem, but it afflicts everybody) - it often
seems to happen that one can
think of certain abstract&high-level variables as being maintained
(car moving down road = 1; car lying on it or beside it as a junk heap =
0; reference level = 1) by means of various relatively concrete and
straightforward low level variables as being maintained, but in between
there's a vast zone where all sorts of stuff might be going on, which,
at least initially, seems to have about as much structure as a plate of
spaghetti. e.g. sometimes one is controlling for `car near middle of
road', other times `car in snow rut', other times `car to right (or is
it left??)' of reflectors, & on top of that there's anticipatory
compensation (knowledge of the road), however that works, etc.
The reference signal does not have to be "digital". I personally would
set it to an analog function. Instead of the car being on or off the road I
prefer to say where on the road "I" want the car. Ie: -1 for the ditch on the
left and 1 for the ditch on the right ( -2 ... offroading?) with values
inbetween.
The problem now exists of how to determine what the "best" refrence
signal or even a good enough refrence signal (position on the road/ditch) is.
This, I think, could fall into pattern recognition which can be handled by
neural networks (tedious and difficult ... yes! but possible). The networks
can be taught to generalize as to the conditions/customs of driving. If the
"car" was transplanted into a new enviornment it may crash or it may "learn"
(just like a tourist), the only way to tell is to try it out (in simulation or
at worst, not in my neighbourhood).
Getting away from networks, if more sensors and input devices are put
in place to deal with narrow roads, high speeds, memory input, etc then these
can be incorporated into the feedback loop by changing the reference signal, or
adding to the disturbance signal. This will mean that instead of trying to
drive down the right side of the road when in england memory input changes the
reference signal (*[-1]) so that the car goes down the left side, or the
middle, or arround some oil or other substance (would you really avoid oil if
you saw some liquid on the road?). If one control loop is responsible for the
control of the car, can`t another be in control of the first? I am getting
too far ahead of myself, I hope that I have made some points that may have been
unclear.
I would like to thank B. Powers for that vapourous explanation kindly
given and wish Tom good luck (that is what I get to look forward to in 8
months). It was very helpful, now it is time to pick up some more lititure and
try to digest all that I have swallowed here (not water) in the last couple
weeks. This term I will be building an autonomous vehicle. I will try to
organize it to accept PCT concepts and hopefully it will serve as a viable
testing ground for different theorys. Should anyone have suggestions
pertaining to traps to avoid or areas that need to be addressed My ears are
always open and any critisim (or that elusive praise) is more than welcome.
Incidently the platform base will probably be one of those battery operated
cars from Toys Are Us (not my idea or first choice but it will save time,
maybe). A great way to test out some of these "Driving Miss Daisy" ideas.
-John
-a tenderfoot in the way of electron behavior
···
--------------------------------------------------------------
John van Loon: John_Van_Loon.XRCC%xerox.com@uunet.ca