[Hans Blom, 931217]
My parents recently had a new home heating system installed. To my sur-
prise, it came with a new twist: it has a predictive controller.
In this country, most heating systems employ a furnace -- traditionally in
the cellar or on the attic -- that heats the circulating water, which is
subsequently pumped to radiators in all rooms. In such a system, getting
the living room from night to day temperature in the morning typically
takes an hour or more when it is really cold outside, but much less when
it is not. People usually don't bother about this delay time and adjust
the thermostat's clock in such a way that the temperature is comfortable
when they get up. This new system, as far as I understand it (the descrip-
tion was minimal), reads the night-to-day switching time and switches the
system to the higher setpoint so far in advance that the temperature is
_at_ the day level at the arranged switching time. This is in contrast
with the usual systems that switch the temperature setpoint up at that
time, and where the temperature is thus at the day level an unpredictable
time interval later.
How does the prediction work? It seems that the controller stores the
times that were needed to get the living room to its day temperature on
the ten previous days, takes the average of those times, and switches to
the new setpoint that much in advance. No more details were given, but if
I were to design a system like this, I would perform some additional
checks on those times and maybe discard some outliers.
No big advertisement campaigns for this gadget, no fuss, just a silent
small technological advance that results in extra comfort and increased
fuel economy, made possible by the low price of microcontrollers.
This system reminded me again of the 'feedforward' discussion of some time
ago, where I proposed a "reverse blink" cursor tracking experiment, which,
in my experience, can succeed only when the controller has a similar type
of memory of the immediate past. Rick promised to solve this problem using
the standard PCT model, however. How are you doing, Rick?
Greetings,
Hans
[From Rick Marken (931217.0930)]
Hans Blom (931217)--
This system reminded me again of the 'feedforward' discussion of some time
ago, where I proposed a "reverse blink" cursor tracking experiment, which,
in my experience, can succeed only when the controller has a similar type
of memory of the immediate past. Rick promised to solve this problem using
the standard PCT model, however. How are you doing, Rick?
I spent some time on it but I didn't really know how to proceed in
a way that might best inform this discussion of feedforward. I
set up a pursuit tracking task where the target moved back and forth
horizontally with a slow, triangular wave motion (about .2 Hz).
Either the target or the cursor could be made invisible for
periods of time during a tracking run; that is, the experimenter
could vary the "duty cycle" of the invisible periods during
the run. So the subject was tracking "open loop" (at least with
respect to the visual perception of curson-target difference)
for fixed periods of time during a tracking run. The results
were not very surprising -- the accuracy of tracking during the
invisible (open loop) periods remained high when these periods
were brief. As the invisible periods got longer, accuracy
declined toward the end of the invisible period.
The tracking was done with a mouse that often sliped or got stuck
on the pad -- so, although no disturbances were added intentionally,
these mechanical disturbances really screwed things up. If such a
disturbance occurred during an invisible period, tracking
performance went south and remained poor during the remainder
of the invisible interval (since the cursor position is an
integration of mouse output). The existence of these disturbances
made it difficult for me to compare the human performance to that
of a PCT model. The PCT model that I used behaved just like the
subject during the short invisible periods, by the way. The
model just continued to send the "appropriate" reference to
the lower level system (that controls the "mouse") based on
the perceived position of the target (or, when the target
was invisible, based on a "perfectly imagined" representation
of target position). Of course, when there were mechanical
disturbances to the mouse during the invisible periods (and toward
the end of long invisible periods, disturbances or not) the model
behavior did not mimic subject behavior -- subject behavior was
much worse than the model's becuase the effects of the model's
output's encountered no disturbance during the invisible or
"open loop" phases).
In order to do this research properly, I would have to eliminate
(as best as I could) disturbances to the effects of the output
device (like by getting a better mouse or pad?). Then I could
have the computer add disturbances, the SAME disturbances for
subject and model. Then one would see that the simple control
model acts (during the short invisible periods) just like the
subject. When the model has no visual perception to control (during
the invisible periods) it can still produce lower level outputs
but these output will not resist disturbances to the visual variable
(for subject or model). There is also a "drift" in the subject's
data from the target position during the invisible phases. This
drift seems to be driven by integration of "error" variance in the
output signal rather than by disturbance. This variance would be
tough to match to the subject's behavior -- though it is easily
modelled.
Anyway, I think this "feedforward" research should be done; we need
to show how perceptual control systems operate when they go "open
loop" with respect to a subset of the variables they are controlling.
Does this seem like the right direction for the research, Hans? I
would appreciate any suggestions, especially from you since you are
the "advocate" for feedforward control. Perhaps I could refine
the experiment or the model so that it reveals the feedforward
control that you imagine is so prevalent; so far, all I've been
able to see in these experiments is good old-fashioned feedback
control. But I'm willing to look.
Best
Rick