[Martin Taylor 950104 10:10]
I have been thinking a bit more about the issue of reorganization, as a
consequence of the interchanges with Bill Powers, and there seems to be
a feature of the localized structure that was not initially obvious.
When I first drew the diagram that simply relocated the elements from
Rick's diagram of the "standard" separate reorganization structure, the
"Intrinsic Variable" control units were, of course, the top-level units
in the hierarchy, just as they are the top-level units in the separate
reorganization structure. This bothered my a bit, because it seems
unnatural for the mature system to have outputs from the intrinsic variables
providing the reference signals for the "principles" level or perceptual
control--the top level of the "normal" hierarchy.
What occurs to me now is that the localized reorganization principle
allows the intrinsic variable control systems to send output that contributes
to the reference signals at ANY level of the main perceptual control hierarchy,
even very low down. The IV contril units do not need to be at the top of
the hierarchy for localized reorganization to function throughout the
hierarchy.
How would this work? To recapitulate prior discussion, in the localized
reorganization proposal, the only criterion leading to reorganization is
sustained and (especially) increasing error in some ECU (any ECU, including
those whose perceptual functions depend on the values of intrinsic variables.
The result of reorganization, in any of the proposals, is some change in
the linkage weights or the forms of the I/O functions of that ECU, or the
growth of new ECUs having some relation to the one(s?) exhibiting the error.
Sustained, and increasing error in an ECU is a symptom not necessarily of
incorrect sign in its feedback loop, but might instead be a sign of its
being involved in some conflict with another control system even though
both it and its adversary are correctly connected and each would work well
in the absence of the other. If two control systems work through the same
environmental path, if only in part, they may be in conflict.
Let's consider a much oversimplified situation. Imagine that there is
an Intrinsic Variable ECU (IVU) that provides output that forms part of the
reference signal for ECU 1, somewhere in the perceptual control hierarchy.
Also within the perceptual control hierarchy there is a higher ECU, 2,
which also contributes to the reference signal of ECU 1. If all is well,
this situation need not lead to conflict between IVU and ECU 2, because
each may be able to use other means to bring its perceptual signal to its
reference level. But if other conditions are unfortunate, it is possible
that ECU 2 cannot bring its perceptual signal to its reference while at
the same time IVU brings its intrinsic variable to its reference--i.e.
conflict can occur. And when there is conflict, the error in both
conflicted ECUs normally increases over time.
If there is conflict between IVU and ECU 2, then according to the localized
reorganization proposal either or both may change weights or internal
functions, or a new ECU may develop. Such changes will keep occurring
until the conflict goes away. This might happen because the external
circumstances (the environmental feeback paths) change, or because the
reorganization has constructed pathways whereby one or other of IVU or
ECU 2 can bring its error to zero without requiring ECU 1 to come to a
particular level, or because ECU 2 has changed WHAT it perceives so that
ECU 1 is no longer a critical component of its perception.
Reorganization in the localized structure can, in this way, involve ECUs
at a level in the main hierarchy above the level at which the IVU's output
contributes to the reference signals, and if The Bomb works as I think it
does, that effect can propagate upwards (and downwards) through the
hierarchy. Failure of an IVU to maintain its intrinsic variable
near the reference level could induce reorganization throughout the
main hierarchy.
If this notion works at all, it removes at least one of my objections to
the localized reorganization proposal--that the Intrinsic Variable ECUs
provided output to the reference signals only at the Principles level
of the hierarchy. There are other objections, but none that (for me)
applied so much more strongly to the localized structure than to
the standard separate structure.
Martin