Imagination and model-based control

[From Bill Powers (940202.1400 MST)]

Mark Olson (940202.1306) --(Direct, with copy to CSG-L)

I'm not sure what you mean by imagined perceptions since in my
post I was suggesting that all perceptions are imagined unless
sense data constrain it.

In model-based control, the output of a control system goes two
places: to the inputs of a local computational model and to the
reference inputs of lower-level systems. If the model has the
same characteristics as the lower systems + environment, its
behavior will be the same as that of the perceptual signals in
the next lower level. Thus perceiving and controlling the outputs
of the computational model is equivalent to perceiving and
controlling via the lower-level control systems.

The advantage of model-based control is that control can
continue, and appropriate outputs can be sent to lower systems,
even if there are interruptions in the perceptual systems
(walking through a house in the dark). The disadvantage is that
this scheme can work only in environments that change very slowly
or not at all, and unexpected disturbances of the world can't be
corrected (because they're not represented in the model).

A couple of years ago I tentatively proposed on CSGnet a scheme
whereby the perceptual functions of a higher system received the
sum of the model's output and the appropriate error signals from
lower systems. Those error signals represent the difference
between the reference signals being sent to the lower systems and
the perceptual signals that actually result. If the error signals
are properly added to the output of the model, the result is just
as though the lower perceptual signals were being received, and
transient disturbances can be handled in real time.

It's a lot easier to think in general about schemes like this
than it is to figure out how such a scheme could actually be
implemented. One great problem is the fact that perceptual
signals are in general a function of multiple simultaneously-
changing lower-level perceptual signals. A model-based system
would have to reproduce all of these lower perceptual signals in
order to simulate the world realistically. Not only that, but
different systems at the higher level generally receive
information from the same subset of lower perceptual signals,
giving different interpretations to them but still controlling by
means of contributing to the same set of lower-level reference
signals. My BYTE articles show a simple example of how this
works. Model-based control by multiple systems at a given level
would therefore require coordinating the various models so they
were consistent with each other in terms of lower-level system
behavior. That would require a rather extensive system for
perceiving the coordinations and making sure there were no

When I realized these problems, I decided that model-based
control is less attractive than it seemed at first. The
mechanisms just get too elaborate; real-time control is vastly
simpler, and therefore to my mind more likely. It's easiest to
let the environment be its own model.

We will have to deal with cases in which perceptual inputs are
interrupted while at least some kind of crude control continues.
I think the key word is "crude." When you lose feedback you also
lose the ability to carry out skillful control. What control is
left, as several of us decided some months ago while talking
about this subject, probably is simply control of those
perceptual inputs that are still available, such as tactile and
kinesthetic inputs.

At one time I was thinking that model-based control might prove
to be a viable idea for the bulk of perceptual control. This no
longer looks likely. Too many problems show up when you try to
imagine how such a system could actually be organized. Of course
there are some valid examples, mostly at the cognitive levels.
When you write a check and subtract the amount from a running
balance in your checkbook tally, you're running a model of what
you imagine your bank balance to be. But models are fallible, and
the imagined reality can easily depart signficantly from the one
you will discover later, when the bank statement comes in.

maybe these seemingly unimagined imagined perceptions
exist at and above one level and not below.

Yes, that's a better bet. By "imagined," incidentally, I mean
internally generated; the imagination connection is discussed in
BCP, pp. 205 ff, as part of the discussion of memory.



Bill P.