Adaptive and Model Based Control

[From Rick Marken (960716.1330)]

Bill Leach (960715.1231)

Model based control is anything but "minor" in engineering design.

I think what you are calling "model based control" is what I would call
"adaptive control" (or reorganization in PCT).

The first controllers that began "wearing the term 'model based' on their
shirtsleaves" were the adaptive controller...These controllers monitor their
own control error and change the parameters (and even the output functions
used) to reduce average and/or peak error based upon operating history.

I agree that this kind of adaptive control is important in engineered AND in
living control systems; it's precisely what we call reorganization. Hans'
model based controller does involve some adaptation; but the essense of model
based "control" is output generation based on the model. This kind of control
of output system will be successful (from the point of view of an observer)
only after adaptation; that is, only after the appropriate "model" has been
learned.

We know from many sources that _new_ motor function tasks are controlled
poorly if at all when first attempted.

Yes. But we get better, not because we learn a model of the environment (as
Hans' system does) but because we tune (adapt) the parameters and functions
of our control organization (perceptual functions, output funcitons, gains,
lags, etc) so that we control better. Hans' model based system learns the
differential equations (environmental laws) that relate output to intended
result. Then, when there is no perceeption of the result, the system can
produce that result by simply specifying the appropriate outputs; control of
output.

As Bill Powers noted in an earlier post, when perception is available, the
model allows slightly improved control (relative to the plain vanilla PCT
model) when the time constant of model based prediction is short; but then
there is little gained from the model in terms of the ability to control when
perception is lost. When the time constant of prediction is long, the model
can produce the intended result without perceptual input for a much longer
time than the PCT model; but in this case model based control no better
than the PCT model when perception is present.

So the only thing to be gained from a model-based system like the one Hans is
talking about is the ability to produce a particular result for a relatively
long time (assuming no disturbance variations) when the perceptual
representation of that result is lost. Then the question is, do people
produce intended results for a long period of time when perception of the
result is lost? We might but we have seen no evidence presented of it and if
it does happen it is certainly "non-nominal".. Also, it's not clear that
whatever ability there is to "control" without perception cannot be handled
by the memory (and imagination) model Bill described in B:CP.

As much as Hans seems to be argueing for "generated output" I get the
feeling that he does not really mean it the way that most of us usually take
him to mean.

I have looked at the code for and run Hans' model based controller. It is not
just an adaptive controller, changing the parameters of a perceptual control
system. It is an adaptive controller that is "adapting" a linear equation
that is the system's "model" of the presumed environmental laws relating
output to input. Once the model has learned the environmental equation you
can switch off the perceptual input to the model and the system will generate
outputs that produce the intended result. Hans' model based controller uses
closed loop adaptation to learn a model that lets it act as an open loop
"controller" (I put "controller" in quotes because the system cannot resist
disturbances as long as it is running open loop).

Adaptation and tuning of lower level control loops has been posited as a
form of learning and this process is undoubtedly also a model based control
operation.

Again, I think adaptive and model-based control are two different things. I
think you are using "model based" control as a synonym for adaptive control;
I think model-based control always must involve adaptive control (to learn
the model) but not all adaptive control is "model based". Adaptive control is
good old fashioned closed loop control where the controlled variable is a
measure of the performance or another control system -- like error -- and
the adaptive system controls this variable by manipulating paramters of the
control system. Model based control, on the other hand, is aimed at output
generation; the model may be built based on closed loop adaptive processes
but the goal is to learn a model that will allow the production of intended
results "in the blind".

Best

Rick

<[Bill Leach 9960716.1754)]

[From Rick Marken (960716.1330)]

I think what you are calling "model based control" is what I would call
"adaptive control" (or reorganization in PCT).

Well we certainly are dealing with a matter of terms that can be difficult
to "pin down". I suspect that the "line" between reoranization and model
based control would be a fine one indeed.

One could easily consider reorganization to be "the model building
mechanism" but I question the idea that all model changes are necessarily
the result of random operations. For example, is it reorganization that
"recognizes" a similarity between a new task and previous tasks? My idea
there is that "it seems" that we sometimes employ existing task control
loops to a new task by "connecting" them in some new and unique way while
tuning the individual loop parameters. While it might be totally
illusionary, it seems as though at least sometimes there is almost immediate
success with the new task. Admittedly this is only conjecture.

... but the essense of model based "control" is output generation based on

the >model. This kind of control of output system will be successful (from the

point of view of an observer) only after adaptation; that is, only after

the >appropriate "model" has been learned.

And this is a sticky part. In the sense that a model for getting to the
store generates very high level reference signals I can sort of agree. When
a task exists that involves "controlling" a perception that is a future
perception that task is accomplished by setting references for the control
current perceptions that, based upon one or more models should lead to the
"control" of the future perception. This is most emphatically NOT S-R.

Yes. But we get better, not because we learn a model of the environment (as
Hans' system does) but because we tune (adapt) the parameters and functions
of our control organization (perceptual functions, output funcitons, gains,
lags, etc) so that we control better. Hans' model based system learns the
differential equations (environmental laws) that relate output to intended
result.

The setting or selection of output functions would, in my opinion, BE an
example of a model based control function. I don't see Hans' example program
as having much to do with control from a PCT perspective. The use of math in
the creation of a model of a model based control system is no more strange
than our own use of math in the standard PCT models. It is the method of use
of digital computers for emulation of analog operations.

Then, when there is no perceeption of the result, the system can
produce that result by simply specifying the appropriate outputs; control of
output.

The idea that organisms construct models of the environment (aspects or
characteristics of the environment), can and do use those models to create
reference signals for lower level systems does not bother me a bit. I don't
even have a problem with the idea that such systems do not perceive the
perception that is to be controlled for over 95% of the time that the system
is running as long as one accepts that there _are_ current perceptions, the
current perceptions are related to the desired future perception by the
model and that there are current perceptions that are being controlled in
the usual PCT manner.

... Also, it's not clear that whatever ability there is to "control"

without >perception cannot be handled by the memory (and imagination) model
Bill >described in B:CP.

The described handling would be model based control.

... that produce the intended result. Hans' model based controller uses
closed loop adaptation to learn a model that lets it act as an open loop
"controller" (I put "controller" in quotes because the system cannot resist
disturbances as long as it is running open loop).

And I agree. I don't consider Hans' model appropriate to model based control
beyond the concepts of how the controller might create the model in the
first place. It would take a great deal more thinking and experimentation
than I have done to see how such a system might actually be able to work at
the program level or above. I suspect that the implementation would be
anything but trivial.

Again, I think adaptive and model-based control are two different things. ...

Yes, sort of anyway. The problem in my mind with making the distinction that
is always appropriate to engineered systems is that the adaptive system in
living beings is created by the system itself and that process is a model
based process.
-bill
b.leach@worldnet.att.org
ars: kb7lx