Re: Central Control

["David A. Fisher" <dfisher@cert.org>]

>David,
>
>I'd like to pose a counter-example to your statement, or perhaps it is an
>opportunity to refine my understanding of global knowledge:
>
>Never having studied 'central control',  I'll offer some reflections
>nevertheless.  It seems that central control can also depend upon layered
>knowledge of the system - take for example a CEO in charge of an
>institution, who holds little _global_ knowledge of the state of the
>system, but instead relies upon the layers of control below to have
>knowledge of activity within their purview.   The layers of control in this
>case are the different levels of approval for taking certain actions
>(mostly expenditure related).  The CEO's knowledge does not have to include
>information about what is happening everywhere, all the time.

Now you are getting at the heart of the matter of why one might prefer
emergent algorithms without central control and global visibility.  In real
world situations we seldom have true central control or global visibility,
but instead attempt to synthesize these capabilities through hierarchical
mechanisms.

>The CEO's
>does have knowledge, however, that control at a lower level can be trusted
>or controlled per se.  Perhaps this is one place where centralized systems
>break down.  So for example, at TERC project directors have approval
>authority for x dollars, cluster leaders have approval authority for y
>dollars, where y>x, etc...  Project directors also coordinate the
>day-to-day decisions of project teams to align them with project goals.

As you point out, control must pass down through the hierarchy and access
to global information must pass up. This is fine under the assumptions that
all of the intermediaries can be trusted to correctly interpret and pass
along the critical control or information.  Each superior position in the
hierarchy represents a single point for potential failure, attack,
misinterpretation, etc. relative to control and information within their
subdivisions with, in your example, the CEO representing the single point
of greatest risk.  Virtual corporations and team empowerment approaches
tend to overcome the risks within certain levels.

Emergent algorithms (without the centralization) avoid these
vulnerabilities.  Species depend on neither global visibility nor central
control for their survivable.  Ditto for birds maintaining safe distances
when flying in flock formations.

>In star-logo this might be interpreted as turtles who have the ability to
>oversee and override the behavior of a finite number of other turtles -
>quite possibly examining the other turtles code and modifying it.

Here we may want to get more technically precise.  We would generally
assume that we are dealing with finite collections of nodes, turtles, or
whatever.  We would refer to the system or network as "unbounded" if the
total number of participants can grow without anyone of the participants
ever knowing (i.e. having global visibility) as to total number of
participants or the totality of their relationships.  For example, no one
ever knows the complete topology or number on nodes on the Internet, not
even the most recently updated router within the Internet.  Nevertheless,
we still assume the system is finite.

The line we tend to draw is between information that in proportional in
size to, depends on, is bounded by a constant and information that is
proportional to or depends of a portion of the system whose size is
proportional to the size of the  system.  Under this view, it is legitimate
for local actions to depend on nonlocal information (i.e. neighbor nodes or
neighbor turtles) providing that the number of neighbors is bounded by a
constant (and not for example a value proportional to the number of nodes).


>The
>controlling turtles do not know the details of the 'decisions' each
>individual turtle makes, but oversee the general behavior of the turtles.

Maybe.  We have no formal difficulty if the number overseen is independent
of the total population of turtles, but to achieve this it eventually will
be necessary to increase the number of levels as the number of turtles
increases.  We do not recommend this form of algorithm because the number
of points of risk (i.e. faulty, compromised, or rogue turtles) would then
also increase with the population size.  We would prefer
algorithms/situations in which the vulnerability of the system as a whole
declines as the population increases.

>This example extends, I think, to Mark Guzdials students who are learning
>C++.  At a higher level, the students per se can be considered the central
>control - although this level might be removed by students programming in
>teams.

Yes.