[Hans Blom, 920831]
Although I have been listening in on this list for some years now, I
have never actively contributed. I enjoy reading the list. You are a
very creative and inspiring bunch of people. Being a control engineer,
PCT is a 'so what' thing for me, nothing new. Some of the fields of
application, psychology and (although implicit) philosophy, fascinate
me. Allow me one contribution and I promise to keep silent again for a
long time. Besides, it already takes me far too much time to monitor
the list; I wonder how many hours there are in a day for some of the
people who are the most active on this list. I admire you, but I
cannot follow you: too many other things to do.
Some general remarks first. One: control is not everything. There is
also a lot of non-control in the world. Two: what do we mean by
'control'? Do PCT and Skinner talk about different things or are they
just different perspectives? Three: where does control come from? How
does it originate? Four: when you are explicit and build models, the
type of control that you use seems to be just the old-fashioned type
PID-control. There is a lot in favor of PID-control, but a great
variety of other types of control have been explored since: adaptive
control, dual control, robust control, to name a few.
Also, when I read the things that you discuss in this list, I often
notice that I see some things very differently and/or that I see
different things. Take one of your popular examples: the control of
movements by e. coli. I remodelled and reprogrammed this 'control
system' from the descriptions that I found in the list discussions of
the last few months. My model looks as follows:
- The environment is a point source of nourishment (let's take sugar)
at x = 0, y = 0. The concentration profile in coli's (two-
dimensional) environment is an inverse square law (different laws
do not make much difference). Coli can sense the sugar
concentration at the point where it is. The concentration at the
point source is too high for coli (poisoning), far away it is too
low (hunger). A radius of 10 is optimal for coli, i.e. coli will
'control for' a radius of 10. At a radius of 100 or more, coli
cannot sense the sugar concentration anymore. Coli's initial
position is at a radius of 50.
- Per iteration coli does the following:
1. Coli tumbles, i.e. selects a new random direction.
2. Coli swims, i.e. takes a step in the new random direction. The
step size is inversely related to its sensed sugar
concentration minus its optimum sugar concentration. This makes
for large steps in case of hunger or poisoning, and small steps
if coli feels fine. The step size is, however, limited to a
maximum value (when coli is at a radius of 100 or more) and
also to a (small) minimum value. This minimum value ensures
that coli will always have to reestablish its position and
cannot relax after reaching the goal. Alternatively, a small
random displacement of its position (modeling physical effects
of water flow and such) provides a similar kind of disturbance.
Pick your parameters as you like (their values do not matter much),
but make sure that coli needs at least some ten steps from a radius
of 50 to a radius of 10 but not more than a few hundred.
Looking at coli's behavior in a number of simulation runs, I notice
the following. In about half of the simulations, coli takes off into
the blue beyond, where its sensors do not work anymore and where it is
essentially lost, despite the fact that its initial position and at
least its initial ten to twenty steps are within the area where its
sensors do a perfect job. In the other half of the simulations, coli
goes towards the radius of 10 and finds it or almost finds it. The
path often looks very crooked with lots of moves towards, and then
again away from, the optimal radius. Sometimes coli reaches a radius
of 11 or 12 but subsequently moves away again and gets lost. In other
cases, coli finds the radius of 10, lingers there for some time,
escapes, returns, repeats this sequence a number of times, but
eventually it escapes and gets lost in what for it is infinity.
Looking at the simulation as it unfolds is an esthetically very
satisfying experience, like art: you _almost_ believe that you
'understand'.
Now what I see is very reasonable behavior given the real environment
in which a real coli lives. But it is not control. You might call it
an attempt to control. Coli has more control than an inert protein
molecule, but not much. Coli is not fully dependent on the Brownian
movement of the molecules of the water it lives in: it looks as if it
has some small say in the matter. I look at coli's behavior as a
demonstration of _emergence_, of how control, in an as yet very
primitive form, comes into being in the tiniest organisms.
It is not difficult to improve upon coli's control, but that requires
additional equipment, either extra sensors or memory. Either provides
a higher-dimensional view of the world and, given the right actuators,
therefore also actions in more dimensions than before.
Better goal directed behavior results if I give my simulated coli a
memory (just a single bit) that remembers whether the new
concentration is 'better' than the old. If so, I make coli continue in
the same old direction. If not, I make coli select a new random
direction. The result is that the new coli (e. coli seems too small to
have such a memory, but this could be the model of a larger sized
bacterium) takes only very _short_ moves away from but _much longer_
moves towards its optimum environment. Eventually, this new organism
reached the radius of 10 in all simulations, and each one stayed close
to it forever.
But its 'towards' is not a 'steepest descent towards'. That is
possible only with another sensor, with which a gradient can be
established. When I give my simulated coli the capability to sense and
use the sugar concentration gradient (again, e. coli cannot do that,
but something the size of an amoeba can), it immediately takes the
shortest path to its optimum and stays there ever after.
These simulations give a lot to think about. First, I see an
_emergence_ of control, from no control at all (type 1; in a protein
molecule, virus or small bacterium with no own modes of movement) to
the primitive partial control that I see in my simulated coli (type 2)
to the gradient descent (type 3) to the steepest descent (type 4). In
reality, of course, there are no discrete types, and even an oxygen
atom can be said to 'control for' the kind of chemical reactions that
it will allow. This makes 'control' a rather fuzzy issue, very unlike
a set of linear differential equations with fixed coefficients from
which you can calculate P, I and D-terms.
Contrast penni sibun (920818.2000)'s general description of type 1-4
behavior
i don't think behavior is ``what *needs* to be done.'' i think it's
what *is* done.
with Rick Marken (920819.1000)'s more restricted description of type 4
behavior
I have spent the last ten years trying to develop demonstrations that
would show that precisely that assumption is wrong.
Is 'leaning on the world' a more appropriate term than 'control' for
type 1 behavior? For which types of behavior is Rick Marken
(920819.1000)'s
The computer, using "the test for the controlled variable", can tell
which of these behaviors is being done intentionally--so it is reading
the subject's intention (mind)--hence, the program does "mind reading".
appropriate? Are often posed questions like "Why is there such
reluctance on the part of those working on the hot approaches to
behavior to even consider the possibility that behavior is the control
of perception?" (this one from Rick Marken) showing a confusion about
type 1 versus type 4 behavior?
Second, these four types of control have little to do with an
'increase of loop gain'. A larger loop gain for coli would mean larger
strokes and the real danger of moving past a food source so fast that
it cannot be located. Actually, coli's loop gain is very robust; a
large range results in almost the same almost optimal behavior. Give
it a too small loop gain and it cannot move; give it a too large loop
gain and it gets lost immediately.
Third, the kind of control that is being discussed in this list is
mainly the type 4 control. But even in humans I see all different
types. Sometimes there is nothing you can do; you are just swept along
with the winds of what is for you just random change. Sometimes you
have a small say in the matter; you only have a vague 'holistic' sense
the too complex situation and hardly know what to do (you cannot
convert what you sense into meaningful actions). Sometimes you can
tell when a new situation is 'better' but not the direction of 'best'.
Sometimes you do know what is 'best', i.e. in which direction to go.
All this has to do with the possibilities and limitations of your
equipment, sensors, memory cells, actuators and the connections
between them. In a sense, sensors and memory cells have the same
function in that the latter can be viewed as sensors that provide
access to (an encoding of) past experiences. Both increase the
dimensionality of the 'impression' that an organism can have at any
moment of time. If the dimensionality of the impression becomes too
low relative to the complexity of the problem, when processing
capabilities are too limited, when actions are futile, or when no
clear goal exists, penni sibun (920825)'s
why conduct elaborate deductions about yr surrounding when
you can look and see? in particular, why maintain elaborate control
structures when you can look and see what needs to be done? why make
highly detailed Plans when you can improvise? why required instant
expertise when you can improve by just keeping on doing it? why try
figuring it out yourself when you can collaborate w/ others who have
been there? why insist on figureing out every situation afresh when
you can trust yor accumulated experience?
strategy, sometimes called 'intuition', may work best. But note that
the overwhelming majority of our experience has been accumulated
through our evolutionary path through the eons, and that therefore our
problem solving methods will be those that are best for _humans_ (in
my opinion, even that is frequently too broad a generalization).
Fourth, we experience the behavior of simulated coli as extremely
complex, despite the fact that the 'laws' on which that behavior is
based are extremely simple. That is because coli's behavior is
unpredictable, 'chaotic'. We cannot nicely formulate the link between
the randomness of its steps and the emerged (limited) order except in
subjective terms such as 'sometimes some colis seem to like being
around the radius of 10'. 'Going down a level' for an explanation is,
because it is so non-obvious, a never-ending investigation into the
randomness of nature. Often, what remains are statistics, a subject
that most of you seem not to like. But in physics, statistics is quite
acceptable as a tool to obtain a sufficiently accurate picture of the
world, such as in statistical mechanics which does not need quantum
theory to arrive at 'emergent' quantities like temperature and
pressure. Here, 'going up a level' is just as impossible, or simply
too difficult for our limited human resources. The 'three body
problem' from celestial mechanics is a classic example, and chaos
researchers will tell you that 'almost all' problems are of this
nature. It seems that the proportion of analyzable problems in nature
is vanishingly small. We may have to 'kludge' forever.
When you only have a hammer, you see nails everywhere. I do not want
to detract from the value of control theory (a hammer is, after all,
sometimes a very useful instrument), but when you have control theory
_only_, you see control systems everywhere. 'Control' is an elusive
thing, just like 'temperature'. Control 'exists' because we want it to
exist, maybe even because we _need_ it to exist to bring order into
our chaos. In nature, the mechanisms that compose what we call a
control system often seem to be thrown together from the lower-level
stuff that happened to be available in an almost random, still badly
understood process that we call evolution. Most emphatically,
evolution has no goal; it just looks that way to the naive observer.
If there is no goal, 'control' may be as naive a concept.
Avery.Andrews says
This discussion is gotten completely out of hand, and, like Martin,
I'm rather baffled by it. I suspect that people haven't done enough
homework on the other guys' stuff to justify the things they're
saying about it.
I agree. A lot of confusion and misunderstanding arises when people
speak a different language. Learning each other's language is a
prerequisite for an exchange of thoughts. Regrettably, this is a life-
long process.
Hans Blom
Who am I? I studied electrical engineering, majoring in measurement
and control theory. I work at a university and have done so for more
than 20 years. My tasks are both teaching and doing research in
'medical engineering'. I have done work in a number of subjects:
modelling, parameter estimation, control, adaptive control, dual
control, man-machine interfacing (now called 'ergonomics'), usually
applied to anesthesia and intensive care monitoring and control
systems. My long-term project is called 'servo-anesthesia', but that
has proven to be a very elusive goal. I have looked around in
artificial intelligence, mainly expert systems, neural networks and
genetic algorithms. Also psychology, physiology, and some philosophy.
My Ph.D. work was the design of a real time expert systems toolbox
(special purpose programming language, its compiler and its inference
engine). An application was the design of an expert system based blood
pressure control system, where the computer controls the flow of an
infusion pump that delivers a drug in such a way as to stabilize the
patient's blood pressure at a lower than normal level. The core of the
system is a PID-controller, but that core is surrounded by a 'safety
shell' that contains expert knowledge, both medical (what to do on
special occasions such as shock) and about how to tune a PID-
controller. Tuning must keep the controller both fast and stable in
the face of large variations in the patient parameters. Difficult to
handle is the variability of the patient's sensitivity to the drug,
which can vary unpredictably by a factor of about 80. Most difficult
is, however, that we cannot explicitly specify what is best for the
patient.
What impressed me most in my research was the influence of the
_unknown_. Some examples: how to learn the initially unknown
characteristics of the patient as fast as possible (time, especially
medical time, is money) when the controller is started, while
guaranteeing safety (predictable, 'almost' optimal behavior); how to
keep track of changes of the patient's characteristics while
controlling stably (no test signals); and how to handle missing
feedback measurements (a temporarily broken feedback loop) while still
ensuring safety for the patient. That cannot always be done. Sometimes
the system has to alarm and transfer control back to the physician.
You have to design in what to do when more and more sensors fail. That
is not easy.
My expert system's logic has three truth values: true, false and
unknown. Ensuring that the system reaches a conclusion of either true
or false ('is the patient OK, yes or no?') but not the uninformative
'unknown' when given the largest number of input 'unknown's generally
does not seem a mathematically tractable problem. Still, we want to
design solutions. Creativity seems the only method, but one that
cannot be formalized, regrettably.
Again, control is not everything, and PCT is even more limited. Just
one example. 'Dual control' theory considers optimization of control
over some time in an uncertain world. It recognizes that, in general,
control will be better the more accurate a model of the world is
available. Thus, in many cases _active_ learning (system
identification) is called for. Active learning, however, means
introducing 'test signals' and thus a disturbance of the observation,
which of course _degrades_ the quality of control. But only in the
short run. Over a long time, the observations of the response of the
test signals yield an improved model which provides a better control.
PCT does not consider this active learning. Its 'reorganization' is,
if I understand correctly, a random process that only occurs when
errors remain large for some time. There is no provision to
temporarily _create_ small errors in order to prevent later larger
errors. This type of active learning ('curiosity', 'exploration',
'play') is pervasive in (higher) animals and humans.
Wittgenstein once remarked that people cannot think unlogically. For
me this means that every remark, however stupid it initially seems,
has its point and can be learned from (maybe that's why I read so
slowly). Sometimes others have just a different perspective on the
same 'truth'. Sometimes someone has an impression that we can show to
be inconsistent with what science shows us. But even then I think that
we have to take that impression seriously and try to find out where it
comes from and what are the sources of misunderstanding.