[From Rick Marken (941218.1430)]
Peter Burke (941216.1000) --
The question is, how are those outputs selected. Its one thing to say
they are selected because they control the inputs, but it is another
thing entirely to show exactly how the error signal accomplishes this
proper selection of behavior. It may not be reinforcement that
selects the behavior, but the behavior is nevertheless selected
by some mechanism. Let's see some disucssion of this issue!
I think things are going to stay pretty confused if we insist on talking
about learning before we are clear about what it is that is learned.
"Behavior" is what is learned. So what is "behavior"? Peter talks about
selecting (learning, I presume) "outputs" and "behavior" as though
they were the same thing. In PCT they are NOT the same thing.
"Outputs" are effects that a controller has on a controlled variable;
"behavior" is the _process_ of varying those outputs as necessary in
order to maintain a perceptual variable in the state specified by the
reference signal. Behavior IS control. Outputs are the means used to
achieve control.
Outputs are the controller's contribution to all the influences on a
perception that is under control. In PCT, the word "behavior" refers to
what control systems do; they control. Recall that the title of the book
that started this whole thing off is "Behavior: The control of
perception". I think Bill Powers picked this title very carefully and that
it means what it says: behavior is the process of controlling our own
perceptual experience. The book is about how we are able to do this.
That is, it is about how we are able to vary our _outputs_ (effects on
our own perceptual experience) appropriately in order to achieve _control_.
In PCT, learning involves changing or building a control system or a
set of control systems; what is learned is a control process. Variations
in output that produce control (behavior) must be distinguished from
variations in output that result from building or changing control
systems. It is impossible to make this distinction (and start to study
learning) until we are able to monitor the behavior of at least one
control system. We have to know what perceptual variable is being
controlled and we have to be able to track changes in the reference state
of that variable. If we can't do this, then we have no idea whether
random looking variations in output are 1) a response to disturbance
variations 2) used to keep a perception matching a variable reference
specification 3) the result of reorganization of an existing control
system or 4) all of the above.
Before we start studying reorganization (learning) we have to have a
real good handle on how to identify and study skilled behavior (high
gain control). For example, we have to have a good handle on the
controlling done by a rat in a Skinner box (what variable(s) it's
controlling, the gain of the control system(s), etc) before we can start
trying to understand the process by which the rat acquired this skill.
There is precious little research on learning in PCT -- probably because
the first order of business must be the study of skilled control, and
there's hardly any of THAT kind of research around either. But there is
at least one excellent study of reorganization. It was done by Robertson
and Glines (Perceptual & Motor Skills, 61, 86-102, 1983). Those interested
in the PCT view of learning might want to check it out.
Another way to get a sense of the PCT view of learning is to play
around a bit with my spreadsheet model of a hierarchy of control
systems (the Lotus version is available on Dag's "PCT software demos"
disk). The spreadsheet shows the operation of a "skilled" hierarchy of
control systems; it represents a living control system that has learned
whatever control skills it needs to control all the variables it is
controlling, despite changes (disturbances) in the environment. In
PCT, "learning" is a change in some aspect of this hierarchy; a change
in one of the perceptual functions or in the connection between one
system and another, for example. Most of these changes result in _loss of
control_ somewhere in the hierarchy. The person who is operating on the
hierarchy is playing the role of the "reorganizing system". Research on
learning is needed to tell us something about how the reorganizing
system operates -- whether it is random or systematic in some way, for
example. But such studies cannot proceed coherently (in my opinion)
unless we (like the user of the spreadsheet) have a pretty good idea of
what the "skilled" control hierarchy is up to; ie. what variables it's
controlling, how it's controlling them and why.
Best
Rick