[Hans Blom, 930309]
I'm really jealous of all you people who can spend so much time on the net.
Thank you all for your replies, especially Bill Powers, Rick Marken and
Martin Taylor. I just don't have the time to reply to everything that y'all
said, so just a few remarks.
Rick Marken (930217.1100):
I give. I thought you might have some EVIDENCE for your point of view.
Sorry to be so irritating, but 'evidence' is, in my opinion, a word very
much like 'fact'. What is evidence to one person may not convince another.
As Popper noted, evidence is an ephemeral thing. There is never enough of
it to prove your point. Like Popper, I think that counter-evidence (debunk-
ing) is much more important. One counterexample may prove a theory wrong.
But then again, one needs a common understanding of what would be a
counter-example. In other words, there might be an infinite regression in a
search for common ground. If the languages -- or the world views -- are too
different, no agreement whatsoever will be possible ('three men and an
elephant'). My personal solution to this predicament is to try to
temporarily adopt the other's world view, however foreign it might appear,
and try to make sense of what the other means. Contradictions often dis-
appear and prove to be just different perspectives. This process has its
own difficulties, of course. But sometimes this process appears to succeed
and it suddenly seems as if you have another perspective on the same old
'reality' which you suddenly realize you saw only partly before.
You know all this. My point is, that in engineering, as in life, we soon
discover that a point of view or a solution is good iff it works for you.
But here the same regression threatens: when does it work? What goals does
it satisfy? It helps to have a common (scientific, engineering, natural)
language. It may not be the best language we have, but without SOME common
language we cannot cooperate.
···
--------
Bill Powers (930217.1030):
When we study human
behavior, we aren't comparing it with some "optimal" or "best"
way of controlling. We're just trying to understand what people
are actually controlling under various circumstances. In some
regards, people control things very well indeed, by clever means
that surpass what any engineer knows how to build. In other ways,
people control stupidly and poorly, and suffer the consequences.
That is not my impression. In my opinion, in the billions of years of
experimentation through evolution people (and organisms in general) have
found superb ways to realize their goals. If we think that they are stupid,
then we are in error: we just have not properly identified their (many!)
goals. This is in line with your remark that
Much of the apparently chaotic nature of behavior
becomes more understandable when we ask about higher-level goals.
In my world view, an organism's behavior is perfectly in line with its top-
level goals. Reaching idiosyncratic goals may, of course, be hindered by
the laws of nature and of society. Every organism is always at its own
local optimum. Of course, we may not agree with its definition of optimum
and think that it is just plain stupid. We may even have convinced the
organism of that 'fact'.
I realize that this is a personal world view that can in no way be proven.
Nevertheless, it is one of my basic life rules, until a better-working one
appears. By the way, your use of 'suffer the consequences' applies in any
case. Behavior has unforeseeable short and long range side-effects, always.
Our perception is limited, although training may improve things slightly.
--------
The highest-level goal is to win the contest, not to jump as high
as possible.
How come you know? The rules of the game are usually considered to be: when
I invent a hypothetical situation, I know what goes on in that situation,
because I invented it. You go against the rules here. I say, in effect,
'assume that X', and you reply 'no, I cannot assume X, I assume Y'. You do
not play according to what I think the rules are.
When I think of a reason, I can only come up with the suggestion that high-
jumping looks different to you than to me. Your high-jumper wants to win
the contest. My high-jumper really wants to jump as high as possible; he is
not interested in winning the contest since he already knows that he is by
far the best of those he meets today. No, he is setting his sights much
higher: he is training for the next Olympics. He has to compete not with
his direct competitors this day, he has to compete with the figures in the
World Records book that he studies every day. But not even that is enough.
He knows that a world record holds only for six years on average. He wants
to do better than that and hold the record for many years to come. He will
just give this jump his very best effort.
Are these extra perceptions helpful in seeing the situation differently?
You could have been right. Your understanding might have explained somebody
else's behavior. But in different persons identically looking actions may
result from completely different motives. A few lines later you do seem to
take that position:
You can't tell what a person is doing just by looking at what the
person is doing.
And later again:
My point is that pure reason isn't going to identify the actual
variable under control by a given person in a given circumstance.
A guess about what someone is controlling for could be quite
right, or quite wrong.
Yes, this is the whole discussion of 'facts' versus hypotheses.
--------
I suspect that this is another of those myths about control, this
time about spinal control systems. For a long time, it was
thought that the tendon reflex had the purpose of "limiting"
muscle tension to prevent damage.
I was not talking about the tendon reflex but about pain as a protective
mechanism. I cannot agree with the following statement:
"Pain" is not an either-or sensation; it begins at zero and rises
from there, with some level being considered "too much" and
calling for action to reduce it. Most "pain", I suspect, is
really just an ordinary sensation, like the sensation of having a
fold of skin squeezed.
You seem to have a weird conception of what pain is. Your "pain" sensors
seem to be my pressure/deformation sensors. That does not correspond with
my experience. When my dentist gives me a shot of painkiller, extraction of
a tooth does not cause pain, but it does cause a massive sensation of
"pull". For me, these are different things. In my opinion, pain is a
stimulus generated by our body when attention needs to be drawn to more or
less massive ongoing or threatening distruction of bodily tissues. Pain, in
this view, is not normally present. Pain, moreover, is a strong motivator
to get away from the situation that brings it forth if still possible and,
probably even more important, a very strong motivator to avoid similar
situations in the future. This is slightly paradoxical: we want to keep
away from something that we do not feel. Isn't such a construct, control
relative to an 'imagined' reference, possible in PCT?
--------
About my formula
x (t + T) = a * x (t) + b * u (t) + e (t)
you say:
Why not the position of the car relative to a point 1 foot to the
right of the middle of the road? You're sneaking a reference
condition into this argument without mentioning it. ...
In a steady crosswind, e certainly does not have an average value
of zero. Assuming disturbances with an average value of zero
conceals the real control problem -- such as standing up in a
gravitational field.
You are right, of course. However, in control systems a CONSTANT never
provides a problem. Control of CONSTANT disturbances is trivial; one might
not even call a constant disturbance a disturbance. Add a constant to the
formula:
x (t + T) = a * x (t) + b * u (t) + c + e (t)
The constant c can be given three different equivalent interpretations:
1. it is an offset for x, say 1 foot to the right of the middle of the
road;
2. it is a constant disturbance, say a steady crosswind or a gravita-
tional field;
3. it is an offset for u, say a base rate of its metabolism to keep the
engine running.
The third interpretation shows why we have a non-problem.
So this control law will leave the car weaving back and forth
from one side of the road to the other under each gust of wind,
with the driver attempting to steer only when a limit gets too
close. If this is how you drive, I'm not sure I would like to be
a passenger! Nor do I think that this behavior would look much
like the way a real human driver steers a car.
When you work out the control law, it shows up as 'behavior' that does not
attach much weight to BEING AT at the middle of the road, but a great deal
to GOING TOWARDS the middle of the road. Mathematically, those are very
different things.
--------
The simplest kind is a reference setting of zero.
If you set your reference level for the perception of a loose
tiger to zero, then any perception of a loose tiger constitutes
an error, and you will act to reduce the perception of the tiger
to zero by moving it away or yourself away from it.
My problem: if you "set your reference level for the perception ... to
zero", then, if you succeed, you see nothing. How can seeing nothing tell
you where you are? "No", you say, "you still use any perception ... by
moving away". You seem to have to balance on the edge between light and
darkness. My problem with your solution is that you are stuck there.
I have encountered the same problem in engineering. We are currently
designing a muscle relaxation control system for use in the operating room.
Its goal is, of course, to abolish motion reflexes that might cause the
surgeon's knife to cut where it should not. Muscle relaxation is estimated
from the EMG (obtained from a muscle in the hand) that is evoked by apply-
ing a supramaximal stimulus (actually a train of four short pulses) to the
nerve going to the muscle. Measurements have shown that the sensitivity of
patients to the drug can vary a great deal. Some patients require little of
it, some a lot. Giving a massive dose of the drug is certain to provide
relaxation in all cases but this is regarded a primitive method: overdosing
is not nice practice. Moreover, it may be harmful to the patient if he is
extremely sensitive or allergic to the drug. One would therefore like to
just barely abolish the response. This approach is unacceptable, however:
going from 0% to 100% relaxation by infusing the drug _in a controlled way_
takes upto 15 minutes. Current manual practice is to give an initial bolus
dose that is barely enough for the least sensitive patients but overdoses
the more sensitive ones. This takes about 3 minutes in the worst case. The
difference of 12 minutes at the start of surgery cost thousands of dollars;
surgeons do not come cheap. A control system must therefore adopt the cur-
rent practice method, even if it has to temporarily relinquish control. A
consequence is, that for tens of minutes or even several hours the control-
ling system does not know "where the patient is" and therefore cannot
answer questions like "when we stop the infusion now, how long will it take
before the patient is able to breathe again", which are of great practical
importance.
--------
human beings
roam free through an undisciplined environment that is far more
complex than any of them can understand. That environment is also
full of disturbances that can't be predicted (weather, for
example) or even be sensed before they occur. Most of our
"predictions" are statistical in nature; sometimes they work and
sometimes they don't. So there's no way that living systems could
evolve to anticipate every circumstance or act correctly every
time.
That is not my point. My point is that the human perceptual + conceptual
systems are so beautifully designed that they even extract information from
very 'noisy' perceptions. It is extremely common to 'filter' stochastic
observations in such a way that the information contained in them is pre-
served whereas the noise is discarded. Kalman filtering is one of the
engineering methods to do this. Originally designed for satellite tracking
and position control, it is a technique to extract the maximum possible
information from few and imprecise measurements. Mathematically, it is
similar to 'real-time' ANOVA. What is does is to 'pack' a great number of
observations into just a few numbers through a least squares averaging
process.
--------
My job is actually easier than yours. I'm not trying to optimize
anything -- just to match the behavior of a model with that of a
real human subject.
Have to be precise here: our jobs are very similar. You ARE trying to opti-
mize something: you are trying to find an optimal match between a model and
a real human subject.
Of course real control engineers know a lot more than I do about
the design of complex control systems ...
Maybe, maybe not. Anyway, that extra knowledge may not account for much
when it comes down to designing good control systems. After all, there is
not much good theory around to travel by. 'Feeling' and 'intuition' are
required as substitutes for knowledge. I don't think you lack those. I have
to agree with Avery.Andrews (930220.1130):
On the topic of `real' control engineers...
There may simply not be much in the way of theorems that
help with understanding how complex living control systems work.
--------
Martin Taylor (930218 10:40):
I always thought
psychology was essentially a problem in engineering, which is why I seemed
to switch fields (according to society--I never thought I switched).
Great! Agreed! But how about this: once in a while I sit back and take the
opposite perspective and consider engineering a problem in psychology. Why,
for instance, do we think that our current scientific approaches are so
great? Why do we believe in grand unifying theories? I frequently see
chance piled upon chance when I ponder questions like: why are there
humans? Where do I come from, biologically and mentally? Where do our
theories come from? Are other approaches feasible? Attribution theory is
one of the psychological tools that 'explains' why similar perceptions lead
to different conceptions (or higher level perceptions) in different persons
or in the same person under different conditions. What is pure sensory
input and what are the personal 'illusions' that I mix in. Infinite regres-
sion again, yet a process akin to vacuum-cleaning my mind.
Where there is no feedback, CSG-L tends to use terms such as "affect,"
"influence," "linkage," and the like.
The question of 'control' versus 'affect' seems to have to do with either
intended versus unintended or full versus partial correlation. In either
case, it has to do with our limited predictive powers. The first raises the
question what it means to 'intend' or to have 'goals'. The second raises
the problem that actions will always have effects in addition to those
'intended'. Control must always be limited; the world is just too complex
for our three pounds of brains to model it and our fifty pounds or so of
muscles to subdue it.
--------
I would call this a selection between different possibilities of action.
That's exactly what I would call "decision." Do we have another source of
confusion based solely on a different dictionary? What do you mean by
"decision?"
For me, decision has the connotation of 'willed' or 'conscious'. I wanted
to avoid that. Why? Well, from modelling theory we know that, whatever we
want to model any part of the world, an infinite variety of models is
possible. We also know that the simpler the model, i.e. the fewer the
number of degrees of freedom that it contains, the faster it converges. But
simple models of complex realities are of course less accurate. In every
model built up from real world information there is a compromise between
detailedness and accuracy. In particular, in a changing world a very
detailed model is impossible; convergence will not be reached, i.e. the
variance of the parameter estimates will remain large. In information
theoretic terms you might say that the (limited) available information is
(often about equally) distributed amongst the degrees of freedom of the
model.
That is why frequently different models are employed. A very detailed one
can be obtained and used if the world that it is about proves to be statio-
nary (a posteriori; if you know this a priori than you need only one
model). When the world changes rapidly, only a coarse model can be obtain-
ed. How selection amongst the models takes place I do not know; somehow the
most appropriate one must be selected. Child psychologists do indeed find a
developmental sequence of models which become more and more detailed the
more experience the child gathers. Sometimes, when the current model does
not work, there is a 'regression' to an earlier model. This regression is
no tragedy, at least not for the individual that shows it; it provides the
opportunity to backtrack and start a different, hopefully better, new
model.
Dick Robertson (930219) points to the selection mechanism:
... Daniel Dennett ... proposes that our thinking consists of "multiple
drafts" in which different control systems (my terminology, not his) in
the brain compete for consciousness...and therefore we often don't know
what we think until we hear ourselves say something. I thought of it in
terms of different versions of a program competing to satisfy some error
in a principle-level system.
--------
Luckily, humans are wired in such a way that they can sense their
outputs; this is called the "body image".
No. They sense inputs from many sensors, some of which are detecting the
conditions inside the body. But it is quite possible (and discussed in
BCP) that the anticipated (imagined) effects of outputs can be used as
inputs through what you call models.
A model basically consolidates observed correlations. If I want to know
what the world is like, I'd better correlate my actions on that world with
the way that world reacts to my actions. It is therefore important to me to
have an accurate sensory picture of what I do to the world. And I do seem
to have the sensors for that, in particular muscle spindles, Golgi tendon
organs, skin pressure and temperature sensors. These provide a sensory
image as far from the body and as close to the world as is possible. That
is what I meant.
Bill Powers again:
When I think of the "output" of a system, I mean the physical
effect on the environment that is due to the actions of the
behaving system ALONE. In the human system, this would mean
muscle tensions, because that's that last place in the chain of
outgoing effects where environmental disturbances can't get into
the process and alter the consequences.
Very much in line with what I said above.
So this is more a matter of labeling than ideology. I'm sure you
would agree that a servomechanism doesn't control the torque
applied to the armature of its motor, but only some consequence
of that torque measured farther downstream in the causal chain.
and
I think that my way of defining output and control is the least
ambiguous.
This IS a matter of labelling. In engineering, we take great liberty in
defining inputs, outputs and systems. I can take for an input anything that
I can manipulate and for an output anything that I can measure. A system is
anything in between. Ambiguity does not appear as long as we clearly state
which is which. One person's choice may differ from another one's, but in
your example that appears hardly relevant as long as the relations between
different output choices are rigid (mathematically: can be one-to-one
transformed into each other).
--------
The only aspect of a control loop that is under reliable control,
therefore, is the sensor signal.
In practice, this is true only if the sensor signal is ever-present and
noise-free. If the sensor signal contains (much) noise, as is often true in
engineering applications, some means is required to separate the signal
from the noise. You often call such means an input function. There are two
general approaches to getting rid of the noise: averaging over multiple
sensors (redundancy) or averaging over time (filtering).
If the sensor signal may be absent for shorter or longer periods, a model
(in the sense of a built-in or acquired approximation of the object to be
controlled) is required that temporarily provides an alternative means of
pseudo-feedback. Such models are usually not highly accurate and therefore
drift will occur away from the optimal operating point.
An easy-to-do experiment to demonstrate both phenomena:
The next time you go for a walk, try the following in a well-know, unsur-
prising environment. First, close your eyes while you walk. Your visual
perception, though not completely absent, will be so disturbed that you
have to fall back on the feedback that other senses provide. You will feel
a strong urge to slow down and you will undoubtedly do so: you need more
time to do the same thing (averaging over time). [Try the test of averaging
over multiple sensors when you can find a number of companions: link arms,
eyes closed. Your common walking speed will tend to go up.]
Second, force yourself to walk in your natural rhythm. Open your eyes
briefly whenever you feel like it, but try to do so as infrequently as
possible without feeling uncomfortable. You will probably find that you
will have to open your eyes only very briefly (like an inverse 'blink')
every three or four steps or less. An internal model (memory) provides the
required additional pseudo-feedback. A similar notion of yours is the
'imagination mode'.
Martin Taylor (930223 14:20) hints at the same things in his answer to Bill
Powers (930223.0800):
When a control system is trying to keep a percept near a reference, it
can do so only to the extent that it can imagine or analyze the incoming
sensory data.
--------
The idea is that reference signals ARE played-back recordings of
perceptual signals, in organisms. This will remain only an idea
until someone does the implied experiments, to see if reference
signals are ever set to values that have never been experienced.
I make no predictions one way or the other.
In adaptive control systems, we can discriminate between two types of
reference signals: those that are hardwired (analogues to hunger, thirst
and bodily integrity) and those that are acquired (analogues to social
conventions). In A.I., the former are sometimes called primary goals, the
latter can be called secondary, derived or contingent goals. The former
must always be met, regardless of outside conditions, the latter depend on
external circumstances to a much greater degree. Hints of nature and
nurture...
Best to all,
Hans Blom