[Martin Taylor 941007 13:45]
Hans Blom, 941005
Forgive me my background in control engineering, but what matters _there_ is
what is sometimes called "multi-sensor/multi-actuator integration".
The problem is then how to design a system in
such a way that an "optimal compromise" is reached among the goals, not all of
which can often be realized at the same time under all circumstances. Bill
Powers proposes a hierarchy as the mechanism to realize such a cooperation.
What is more common in complex control systems is a "society of mind"-type of
organization, a number of parallel control loops that do not only monitor
their own process but are also, somehow (through cross-connections), aware of
what other control loops are doing.
I may be wildly wrong here, but it seems to me that there are two separate
issues, not one. One is based on the background thought in PCT that the
elementary signal is always the series of impulses on a single neural axon,
and is thus necessarily scalar, whether each impulse is treated singly or
impulses are integrated to define a scalar "neural current". The other is
that we subjectively feel that we have perceptions of different kinds, one
kind being built on the basis of perceptions of other kinds, as, for example,
perception of a sequence is built on successive perceptions of events. These
"kinds of perception" form the Powers series of levels in the hierarchy.
The "kinds of perception" argument sometimes strikes me as unwarranted,
inasmuch as one can obtain, say, transition functions without separately
having intensity functions, and transitions and intensities can be combined
into the same function. Nevertheless, the argument does lead to at least
a surface validity for the idea that there is a perceptual hierarchy (an
idea that is at least 150 years old in psychology, anyway, going back to
Donders--a Netherlander). If there is a perceptual hierarchy, and if
2a'. for all Y (perceptions), Y can be used by some (controller) X
then there is a hierarchy of perceptual control.
The second (or should I say first) point is the identification of perceptual
and other signals with the signals on neural axons. This is not a necessary
component of perceptual control theory, but again it is a tempting mapping
If it is true, then ALL controlled perceptions are scalar, not vector, even
though an outside analyst might well observe that the control hierarchy
acts in the same way as an adaptive vector controller.
The question of whether individual control systems "know what each other
is doing" is independent of whether the controlled signals are scalar.
One ECU could "know what the other is doing" if, say, the output of one
affected the gain of another, or if the perceptual signal on one contributed
to the sensory input of the other. Both connections are envisaged as
possibilities in PCT, the second being normal in the standard simple
hierarchy. Only the connections of this kind are always between levels
and are never reciprocal. Such interactions are not usually considered
as happening within a level (but see below).
There is a tendency to think that parallel non-interacting control systems
will have problems in a non-linear environment with local minima (traps),
but, as Bill Powers' Crowd demo shows, that thought is not always correct.
The entity with 4 or 6 parallel (single level) control systems can find its
way quite effectively and often nearly optimally through a complex set of
obstacles that include pockets that might have been expected to trap it.
An open question in my mind, which perhaps you could address, is whether
there exists any theoretical basis for saying that the following statement
is either true or false:
"Any multisensor, multi-actuator control system of one level can be emulated
by some hierarchy of scalar control systems that do not interact within
a level of the hierarchy."
There is one place in which I think that interactions must exist within a
level of the hierarchy, and that is what is called the "category level" in
the Powers hierarchy. I have had strong criticism from the PCT old-timers
on this issue, but I have not found any way around it, and maybe it is there
that the above statement can be shown to be false. If so, then a weaker
statement might hold:
"Any multisensor, multi-actuator control system of one level that does not
involve any use of categories or logic can be emulated by some hierarchy
of scalar control systems that do not use categories and that have no
interactions within a level."
In most experimental tests of PCT, the variable subjected to the Test is
unidimensional, though Tom Bourbon has done many 2-D tests, and I suppose
there are others I don't know of. In such tests there is little opportunity
to investigate whether scalar hierarchies are adequate to describe more
complex real-world control in which there might be static or dynamic conflicts
between scalar control systems (juggling payment and purchase schedules
when the total income might be adequate but the momentary requirements
fluctuate, for example). Simulations or analysis are required for such
situations.
You say that a higher-level system adjusts the lower-level goals. Now assume,
just for a minute, that there is no higher-level system and that the highest-
level reference level is fixed. Are the lower-level reference levels now fixed
as well? If not, might we not say that it is the _perceptions_ that adjust the
lower goals?
In fact, one of the presumptions often made is that the highest level references
are fixed. The consequence is that when any perception changes, there may
be error signals induced at the highest level, causing output at that level,
which fans out to become reference signal changes at the next (and all lower)
levels, which cause error signals to change, some reducing, some increasing,
and so forth. So you _could_ say that the perceptions "adjust the lower
goals" but not in any predictable or consistent way. It may be true, but
it isn't very helpful.
Let's not quibble about words. You know what I mean: the system is designed in
such a way, that simple-minded perusal of one goal cannot destruct the whole
system.
Can't it? What about drug addicts, teen-aged drivers playing "chicken" ...?
as a control engineer, it strikes me as funny to say that
all that matters to an ECU is what it itself does.
I think you are using two senses of the word "matters." In one sense, it
matters to the ECU if its sensors are blinded, or its energy supply vanishes.
In the other sense, it simply "is," and as such, it acts only according to
the waveform of its error signal. Nothing else "matters" to it, inasmuch
as nothing else affects its output. If there is a cross-linked signal that
affects the gain of the ECU, then that signal "matters" to the ECU. But
since we are usually considering hierarchies without such cross-linking,
it makes eminent sense to say that the only thing that matters to the ECU
is the difference between its reference level and the level of its perceptual
signal.
Martin