Conditional stability

[Martin Taylor 920718 12:00]

I am hesitant about starting a new and possibly long thread before dealing
with outstanding issues, but in writing up my extended abstract for the
Paris talk, I came across a phenomenon I think could be quite important
in various areas. I think specifically of the mass murderer who is always
the kindest, quietest kid on the block. Since it hasn't been talked about
while I have been on this group, I'm wondering if it is a new discovery or
one that has been long dismissed. (By the way, the Paris paper is the one
for which I asked assistance some months ago, on HCI, where I have to give
a keynote presentation that I intend to base around PCT. It is the reason
I have been (relatively) quiet for the last couple of weeks.)

Consider two ECSs at the same level, labelled Y and Z. Their perceptual
input functions are, say, y(a-b+c) and z(b-c+d) where a,b,c,d are sensations
coming from A,B,C,D which may be lower level ECS perceptions or objects in
the real world. Notice that Y and Z both control for functions that contain
b and c, but also have another variable.

The outputs of Y and Z affect A, B, C, and D. Y affects A,B,C, and Z affects
B,C,D. The signs of the output connections determine whether Y and Z have
proper negative feedback control (getting them that way is what reorganization
is all about). Let us assume that they do, so that all of them move in such
a way as to oppose the error in both Y and Z.

Y and Z have their reference signal levels set from above. If their reference
levels change oppositely, then Y tries to (say) increase a and c, but decrease
b, whereas Z tries to decrease b and d while increasing c. B and C are affected
the same way by both, and probably b and c change dramatically while a and d
change only a little.

If the reference levels of Y and Z change in the same direction, then their
outputs affect B and C oppositely. If they have roughly equal gains and
equal shifts of reference level, b and c will not change at all, and Y and
Z will reduce their errors by causing large changed in a and d respectively.
Everything is still under control, but the outside observer will not see
that B and C are in play. It will seem as though Y is concerned only with A
and Z is concerned only with D.

Now comes the interesting part. If you haven't drawn the situation on paper
yet, I suggest you do so, to make the argument easier to follow.

Let us change the sign of one of the outputs of Z, specifically the one
affecting D. Now if Z acts alone, it will cause b and c to change so as to
correct any error, but d will change so as to increse the error. Assuming
they all have equal responsiveness, the changes in b and c will dominate
the change in d, so that Z maintains control and all looks normal. Neither
Z nor an outside observer can detect that the loop through D has positive
feedback.

With this changed Z, reconsider what happens when Y comes into play. If Y
collaborates with Z in its effect on b and c, the situation is normal. All
is well controlled. But what happens when Y opposes Z in respect of b and c?
The levels of b and c do not change. A moves in such a way that the error in
Y is reduced, but D moves to increase the error in Z. There is positive
feedback. Z goes out of control. And supposing that Z contributes z to the
perceptual input function of higher level ECSs that contribute to its
reference level, there is the possibility that they, too, will go out of
control.

There are mechanisms that would reduce the probability of propagation of
control loss up the hierarchy. If the connections from above, say from ECSs
labelled P and Q were such that Z contributed only a small amount to the
perceptual input function of either, they might be able to retain control
through Y and other ECSs at the level of Y and Z. Also, saturating
non-linearities in the perceptual input and/or the gain function of Z would
reduce the likelihood of a runaway loss of control, since the contribution
of D to the propagated error would be limited to a fixed maximum that could
be countered by control exercised through other variables.

But sometimes, the error induced by a positive feedback loop would propagate.
The hierarchy could work flawlessly, with near zero error, for a long time.
Then a particular pattern of reference signals might occur at some level
of the hierarchy, preventing some ECS from exercising control through
loops having negative feedback and making it act through a loop that
had unsuspected positive feedback. And then, the hierarchy might go crazy.

I haven't gone too much further with this analysis to see whether the
propagation of errors induced by hidden positive feedback loops will normally
proceed, be suppressed, or whether it is a matter of chance. But it does
seem to show that a hierarchy that has reorganized and has near zero errors
in all ECSs (and in its intrinsic variables), still has the potential for
violent and maladaptive (non-controlling) action. It can lose its temper.

A system in this state would be hard to detect, either from outside or from
inside. It exhibits no errors most of the time, but goes mad under some
ill-defined kind of stress. I cannot see how any reorganizing principle
we have yet discussed (including my own) can avoid producing such a brittle
system. But it is possible that it could be avoided by an annealing process
in which the developing system is exposed to many different kinds of minor
stress, in conjunction with either Bill's or my reorganizing principle.
The minor stresses might sample the space of reference level changes, and
could trigger these conditional instabilities, inducing reorganization
that should remove the dangerous loops. But that kind of sampling might
not always work, leaving around a potential mass murderer who does not
know that he is one. And he would be the quiet, well-behaved type who
has always been in control.

I don't know if the above is correct, but it seems so to me at the moment.
I am still trying hard to write this paper, which had a deadline of last
Wednesday, so don't expect too much follow-up from me for a while.

Martin