NetNews Usenet Archive 1992 #31/NN_1992_31.iso/spool/bit/listserv/csgl/1936

Text File | 1992-12-22 | 15.0 KB | 355 lines

Comments: Gated by [e-mail]
Path: sparky!uunet!gatech!darwin.sura.net!paladin.american.edu!auvm!BEN.DCIEM.DND.CA!MMT
Message-ID: <[e-mail]>
Newsgroups: bit.listserv.csg-l
Date:         Mon, 21 Dec 1992 18:29:24 EST
Sender:       "Control Systems Group Network (CSGnet)" <[e-mail]>
From:         [e-mail]
Subject:      Re: Rick on (off) Shannon
Lines: 344

[Martin Taylor 921221 18:00]
(Rick Marken 921221 1300)

Rick ,
Welcome back from your trip.  I'm sorry it didn't go well, and left
you in a bad mood, when you saw my little twitting of you.  I didn't intend
you to take it as an insult, any more than you presumably intended
your comment to me a couple of weeks ago as an insult, when I took it
badly.  Sorry you took it that way.

I think I posted a reasonable summary of the interactions, but you
seem to think I misled the other readership.  I had asked you to post
everything, to avoid that kind of problem.  Now it seems that to preserve
honour (a rather silly concept, but we tend to have reference levels
for that kind of thing), I have to waste net bandwidth by reposting
the discussion entire.

Either I still misread what you posted, or you misread what I said.
It still seems to me that you were arguing for precisely what I claimed
in my summary, which you now claim is ridiculous.
----------------------
>If (like Martin) you don't undertand PCT, DON'T TRY TO
>TEACH IT. Those of us who do understand it  -- and
>have done the grunt work necessary to grasp the
>fundementals --  find condescending tutorials on PCT
>to be cloying.

I have a feeling it is my various attempts to look for the fundamentals
that you (Rick) find cloying (should I guess at threatening?).  I know very
well that I don't understand PCT.  I don't think anyone on the net claims
to except you, least of all Bill P.  I'm sorry that you don't want to
look beyond the behaviour of control hierarchies to see how and why
they work in addition to what they do.  I hope that "those of us who do
understand it" is not a large group who find my postings "cloying."

One of the best ways of learning is to try to teach.  But I don't think that
I have ever been trying to do that (except locally here at DCIEM).  What I
have tried to do is to develop the implications of PCT, to enquire as to
problems that seem to arise, to get at the basics of PCT.  If you don't
like it, I'll quit, and just work with our group here at DCIEM.  There's
no benefit to me in finding an automatic rejection of my ideas when they
are proposed to the net.  It is really Bill P's encouragement more than
anything else that keeps me going--that, and my belief that PCT is the
Newtonian revolution in psychology.  But the latter part is something
I can develop on my own, or in private communication with Bill P and my
local collaborators.

------------------------
Here are the postings, so far as I still have them.

Martin

===========================
===========================
(I don't seem to have the first mail, to which this is Rick's response)
=========================
From [e-mail] Thu Dec 10 14:30:23 1992
Subject: Re:  Rick's proof: error not controlled
To: @MVE.AERO.ORG:[e-mail]
Status: RO

Martin

You seem to have sent this to me personally -- so I am sending it
back to you personal (I think; it's often hard for me to tell where
mail came from when I'm using this mail system). But I'd be happy
to have the put on the net -- as usual.

>I have no complaint about your proof.  Different people see things better
>from different viewpoints.  I thought I was buttressing, not caviling.

OK. What made me think that you disagreed with the proof was the
following:

>>Where in the world (take the literal meaning of that expression) is the
>>possibility for disturbing the error in an ECS?

Now you say:

>A disturbed error signal is not an error signal as defined.

If you cannot disturb error (and again I point to Bill's diagram to
show that you can) then my proof is invalid because it is based on
the idea that o = k(r-p+de) -- were de is the disturbance to error.

As I mentioned in my post with the equations, if the perceptual signal
were disturbed in the same way (by having a disturbance added to
the neural signal itself) the perceptual signal would still be controlled
(made to equal the reference -- ie. the disturbance would have no effect.
So again, while I agree with your conclusions, I disagree with the
arguments you use to come to them.

>I suppose I would agree that error could be controlled if it were to be
>provided as input to the perceptual input function of a control system that
>maintained the error at some reference level.

As you will see in a reply I plan to make to Greg, there is no
supposing necessary in this case. The error signal (which is now
a perceptual input to a control system) would DEFINITELY be
controlled.

> But what on earth would that
> do other than enforce a situation in the ECS whose error was being
controlled
>that would ordinarily arise only through unresolvable conflict.  I can't
>see why any hierarchy would include such a mechanism--but I suppose there
>might be a reason somewhere.

But this is precisely the way reorganization works. It controls errors by
changing properties of the control system itslef (like its perceptual
function) -- properties which are presumably responsible for the conflict
that is creating the chronic error. Reorganization is control of error
(where error is the input perceptual variable to the reorganization control
system) -- and control is effected by acting on properties of the control
systems themselves. Since there is no way to know HOW to change the
control systems in order to reduce error, there must be a random
componenet to this kind of control -- Bill has modelled as a control
system where the RATE at which random changes occur is inversely
proportional to the perceived value of error.

Best

Rick
============================
From mmt Thu Dec 10 16:45:11 1992
To:  [e-mail]
Subject: Re:  Rick's proof: error not controlled
Cc:  ./marken
Status: RO

Rick,

As usual, we were talking a little at cross purposes. There's no real
argument.  You were assuming that there might be some way that a disturbance
could be applied to the error within an ECS, and when you mention electrodes,
that can be true.  I was working within the normal hierarchy, in which the
only inputs to an ECS are through its perceptual input function and its
reference inputs, and within that frame, there's no way to apply a "de".
That doesn't argue against your proof at all.  If(x) then (y) is the proof.
Not(x) does not say not(y).

On a different topic, you say:

>> But what on earth would that
>> do other than enforce a situation in the ECS whose error was being
>>controlled
>>that would ordinarily arise only through unresolvable conflict.  I can't
>>see why any hierarchy would include such a mechanism--but I suppose there
>>might be a reason somewhere.
>
>But this is precisely the way reorganization works. It controls errors by
>changing properties of the control system itslef (like its perceptual
>function) -- properties which are presumably responsible for the conflict
>that is creating the chronic error.

Reorganization may well work this way. (All we know experimentally is that
Bill says that my approach to reorganization is the only one he knows to
work by simulation, and then only in respect to gain functions; even if we
eventually discover various ways in which it COULD work, we won't know
that it DOES work this way).  But the thing I was saying was that however
reorganization works, it seems unlikely to develop stable systems that do
useless or obstructionist things.

> Reorganization is control of error
>(where error is the input perceptual variable to the reorganization control
>system) -- and control is effected by acting on properties of the control
>systems themselves. Since there is no way to know HOW to change the
>control systems in order to reduce error, there must be a random
>componenet to this kind of control -- Bill has modelled as a control
>system where the RATE at which random changes occur is inversely
>proportional to the perceived value of error.

Actually, what Bill has done seems to be a little more complex. He has
found that the derivative of the squared error is a better criterion.
In our experiments, we have been planning to use K1(e^^2)+K2(e*de/dt),
where de/dt is the derivative of error (2e*de/dt is the derivative of
e^^2).  We think that both error and the rate of increase of error are
important.  It shouldn't matter if the error is momentarily large, provided
it is decreasing well.  When I say "criterion" I mean the Poisson rate of
reorganization events.

Does this help to reduce the cross of our purposes?

Martin
====================================
From [e-mail] Tue Dec 15 18:54:38 1992
Subject: Re:  Rick's proof: error not controlled
To: @MVE.AERO.ORG:[e-mail]
Status: RO

Martin

>Actually, what Bill has done seems to be a little more complex. He has
>found that the derivative of the squared error is a better criterion.

You mean, as the variable controlled by the reorganization system, right?

>In our experiments, we have been planning to use K1(e^^2)+K2(e*de/dt),
>where de/dt is the derivative of error (2e*de/dt is the derivative of
>e^^2).  We think that both error and the rate of increase of error are
>important.  It shouldn't matter if the error is momentarily large, provided
>it is decreasing well.  When I say "criterion" I mean the Poisson rate of
>reorganization events.

Now I don't understand. The "Poisson rate.." is a criterion? It sounds like a
variable. Why would your model be concerned about the "Poisson rate of
reorganization events"? It's not controlling that rate, is it? Bill's model
controls  the derivative of the squared error(you are correct about this,
I believe); it controls this variable by changing some parameter of the
systems
that are experiencing the error; the time between such changes increases as
the
difference between the derivative of the squared error and zero decreases. As
a
side effect the Poisson rate of reorganization events changes during the
course
of reorganization.

At least, that was my impression this summer in Durango.

Regards

Rick
===============================
From mmt Tue Dec 15 20:34:04 1992
To:  [e-mail]
Subject: Re:  Rick's proof: error not controlled
Cc:  ./marken
Status: RO

Rick,

It's words again getting in the way...

Bill's model doesn't control the derivative of the squared error, in any
normal sense.  That derivative has more the function of "error" in a
standard ECS, in that it determines the output of the reorganizing function.
That output is a reorganizing event that occurs from time to time in the
main hierarchy.  The higher the value of the derivative of the squared
error, the more likely a reorganization even is to occur.  I don't know
whether in Bill's simulations the rate was determinate or Poisson.  Ours
is Poisson, which means that the probability of a reorganization event
happenning in a small delta-T is proportional to the value of the criterion
(derivative of squared error in Bill's case, unless he was using a
determinate time-interval between events).

Now what is controlling what?  I think what is being controlled is the value
of some intrinsic variable.  It has some error, and all this stuff about
derivatives and Poisson rates are attributes of the output function of
the control system for that intrinsic variable.  As with any control system,
whether in the main hierarchy or in the reorganizing system, the actual
outputs are not controlled in themselves.  Outputs are, shall we say, blind.
So the "criterion" is not controlling the Poisson rate.  It is just a
component of the output function of a controller that controls an intrinsic
variable.

(Bill and I have an unresolved, and perhaps unresolvable, disagreement here
as to what might constitute an intrinsic variable, but that disagreement is
minor and not germane to this discussion.)

Martin

PS.  It feels great to be back on the air.  I hope the link stays in action
for a while.
===============================
From [e-mail] Wed Dec 16 11:59:39 1992
Subject: Re:  Rick's proof: error not controlled
To: @MVE.AERO.ORG:[e-mail]
Status: RO

Martin

Let's put this on the Net.

>Bill's model doesn't control the derivative of the squared error, in any
>normal sense.

It does. I'll explain on the net.

>That derivative has more the function of "error" in a
>standard ECS

Because the  intrinsic reference (for de^2) is 0.

>Now what is controlling what?  I think what is being controlled is the value
>of some intrinsic variable.

de^2 is the intrinsic variable being controlled relative to an intrinsic
reference which happens to be 0; it could be something else (maybe
evolution wants the system to operate with de^2 at just a tad above
zero -- keeping you on your toes).

de^2 is controlled because it is both a cause and a result of reorganization;
and the sense of the feedback is (or should be) negative.

Regards

Rick
===============================
From mmt Wed Dec 16 14:10:23 1992
To:  [e-mail]
Subject: Re:  Rick's proof: error not controlled
Cc:  ./marken
Status: RO

Rick,
I don't mind putting the discussion on the net, but I don't think it worthwhile.
There are too many long wavelength fish involved.

>Because the  intrinsic reference (for de^2) is 0.
>
>>Now what is controlling what?  I think what is being controlled is the value
>>of some intrinsic variable.
>
>de^2 is the intrinsic variable being controlled relative to an intrinsic
>reference which happens to be 0;

This makes no sense to me.  In words, you are saying that whatever the
error is in the intrinsic variable (say blood CO2 level, for example),
what is controlled is that this error should not change.  To me, the
only thing that makes sense is that there is some reference level for
blood CO2, and if this is too high or too low, and particularly if it
is moving in the wrong direction, then the main hierarchy needs reorganizing.
The reason for using the derivative of the squared error in the intrinsic
variable as part of the output function is that it is a short way of
ensuring that the error is both large and increasing, whichever sign it
has.  Other functions having the same properties would also be useful.
But e*de/dt is not a controlled variable, any more than (in fact less
than) is the error in a normal ECS.

Another way of seeing that d(e^^2)/dt is not controlled is to note that
zero is not a reference level for it. Negative values are even better,
becasue they show that the intrinsic variable is really being controlled.
A zero value is neutral in this respect, in that the error in the intrinsic
variable may be large but unvarying, or the error may be small while changing
wildly.  Neither indicates good control, but both are compatible with good
control (momentarily).

If you want to put it on the net, collect all the postings and send them
out as one.  But I don't think it worthwhile.  I haven't seen anyone else
that seems to be interested in the matter (except presumably Bill, by
inference)

Martin