Reading Wiener in Rio

Felipe Pait
Universidade de S ao Paulo
Email: [email protected]
AbstractThis selfconsciously somewhat rambling paper re-
visits the connections between 3 dening themes of Wieners
cybernetics: communications, control, and cognition. Although
Wiener emphasizes the unity among these concepts leading
to a science of information, in contrast with preceding scientic
concerns with matter and energy the strands have developed
along largely separate paths. This is understandable in terms of
the internal logic and difculties of scientic development and the
search for applicable results, but is also to some extent caused by
centripetal forces in professional academic investigation leading
to compartmentalization of knowledge inside university depart-
ments.
We consider connections between cybernetics and the social
sciences, which seem to be more fruitful in the cases of com-
munications and cognition. In the rst case they include the
socalled Toronto school of communication theory, which studies
how societies and political institutions are shaped by the prevalent
communications technology. The latter are too numerous to
mention and beyond the scope of this paper.
The connections are more indirect in regards to feedback
control, despite the economic applications of game theory, which
although not unrelated to cybernetics is of a somewhat different
character. We regard the phenomena of bubbles and bursts
in adaptive control, which suggest that the bubble and bursts
in nancial economies are related to secondorder or adaptive
feedback loops. This example serves a dual purpose. First is
to advance the proposition that the time is ripe for a better
application of feedback control, as understood by Wiener, to the
social sciences. Second is to suggest that a crucial concept that
might serve to further rebridge the gaps formed over decades of
independent research is adaptation.
Tuning algorithms as employed in adaptive control and lter-
ing are strongly modelbased and very dependent of the nature
of the models used. More restrictively, they are often dependent
on the linear dependence of parameters to some tuning signal
such as an identication or control error. This is to some extent
unavoidable because as soon as we move away from linear errors
and quadratic objetives the analysis becomes too complex to
permit hard bounds on the errors. However this requirement has
long being an obstacle to the more widespread use of adaptive
algorithms, and a barrier between the more rigorous aspects of
cybernetics and those that defy our attempts at formalization.
Of course the optimization literature is full of algorithms that
can be used in a nonlinear setting. The question is to what
extent they can be employed in a real time problem, where
parameter convergence must be effected while the parameters are
employed, say, to close a possibly unstable feedback loop. Also
important is to note that an a priori mathematical formulation
of the objective function to be minimized is not available in a
realtime adaptation problem. For example, in adaptive control
the objective function is only computed a posteriori as a result
of measuring the actual performance of the closedloop control
system when a certain control design is employed. This strongly
suggests consideration of direct, or derivativefree, optimization.
Without making claim of any sort of grand synthesis, we
advocate that adaptation can employ a direct optimization
method based on the computation of the barycenter of a sequence
of evaluations or measurements of a given function. The weight
given to each point xi is e
f(x
i
,t
i
)
, where f() is the function
to be optimized, xi is a point probed at time ti, and is
a suitable constant. The method is conceptually transparent,
straightforward to code, and the free design parameters have
an intuitive interpretation.
This being a paper about Norbert Wieners work, it cannot
fail to close with a proper theorem. We obtain some useful
results concerning the barycenter method in a continuous-time
framework, which is perhaps more complex mathematically but
affords a clearer view of the connections with more traditional
model-based optimization techniques. Consider a continuous-
time sequence of test points x(t) R
n
and the corresponding
cost time-function f(x(t)). The barycenter of the exponential
cost is given by the formula
x(t) =

t
0
e
f(x(s))
x(s)ds

t
0
e
f(x(s))
ds
.
It is reasonable to pick at each instant of time the test point x(t) =
x(t)+z(t), where z(t) may be regarded as the curiosity of the
optimization algorithm. We show that, under some simplifying
assumptions, if the statistics z are those of a white noise, then
the barycenter update rule is related to gradient-type methods.
Although the algorithms discussed are completely derivativefree,
these relationships serve as a useful design guide in the choice
of the relevant free parameters.
I. CYBERNETICS AND THE SOCIAL SCIENCES.
Communication theory along the lines studied by the
Toronto school developed in dialog, although not invariably
in friendly cooperation, with Wieners cybernetics. The core
idea is to study the role of communications in the evolution
of societies, emphasizing how political institutions are shaped
not only by material considerations but also by the prevalent
communications technology.
The connections between, on the one hand, mathematical
cognitive sciences, including but not limited to articial intel-
ligence, and on the other hand social sciences centered around
psychology are manifold and beyond the scope of the present
paper.
But we must admit that we are not aware of a similar
interaction between the remaining theme of cybernetics
control and feedback theory and the social sciences. The
closest such connection may be through game theory. While
neither exactly a particular case nor a generalization of optimal
control, the mathematical theory of games and the inevitably
resulting minimax problems are certainly a more structured
approach to questions similar to those considered in cybernet-
ics. At the same time game theory brings about, in a natural
if perhaps not inevitable way, a worldview which stands in
opposition to that of Wieners. This point is very salient in the
intriguing double biography of Wiener and von Neumann[1].
Wieners worldview is softer, more humanistic, compared to
Johnnys more hawkish and uncompromising ways.
Perhaps the similarities are more poetic that scientic; and
certainly one should not read too much into the moods sug-
gested by different mathematical terminology and approaches.
Yet it is impossible not to see a parallel with the differences
between the phase margins and stability and oscillation com-
promises of classical feedback control, on the one hand; and
the harsher minimize against your opponents maximizing
strategy gametheoretic paradigms.
How can the controls aspect of cybernetics, in a more
Wienerian avor, nd wider applications to the social sci-
ences? To give one possible answer, it is instructive to con-
template the present condition of the international economy
from a cybernetic perspective. In the digression that follows
we shall discuss the state of the economies of the richer part
of the world in the last decade or so, as amply discussed in the
leading newspapers, without making use of academic citation.
In this regard we may notice en passant that the saltwater
school of economics, to which Wieners MIT belongs, has
comported itself admirably in the proper mix of rigorous
mathematical reasoning and skepticism towards excessive for-
malism. Beforehand let us talk about a wellknown but still
poorly understood phenomenon in adaptive system.
II. BUBBLES AND BURSTS IN ADAPTIVE CONTROL.
Rich economies have been operating below capacity since
the burst of the most recent bubble. Bubbles and bursts are
a form of instability typical of adaptive processes, in fact
they are familiar in adaptive control since the famous Rohrs
counterexamples. Let us describe in words the phenomenon
that was identied in[2] and further studied the subsequent
literature.
Bursts tend to appear when adaptive systems are designed
to control a simplied, reducedorder model of a process.
While the controls are exploring, trying to learn, the closed
loop performance tends to be poor. Adaptive controllers are
designed so that large mismatches and errors force adapta-
tion to proceed. Once a properly stabilized regime settles
in, adaptation continues unless explicitly turned off
although at a much reduced pace. However the welldesigned
controller is able to keep the modeled, dominant modes of the
process quiescent. Errors will therefore be attributable mainly
to process noise, or to dynamic modes which are not present
in the process models employed for identication. Adaptation
thus tends to be driven mostly by the higherorder, unmodeled
process dynamics, and control parameters often drift away
from the regime values where the process is stabilized.
Once a situation is reached where the frozentime controller
is no longer stabilizing, and with this regards their signals turn
awry. The process turns unstable, and adaptation wakes up and
proceeds at its desired, fast pace. The controller learns from its
mistakes, focus on the dominant modes, and converges again
to some stabilizing value. Process states converge back to their
quiescent situation, and the cycle is repeated. The adaptive
controllers success contains the seeds of its own future tsuris.
III. BUBBLES AND BURSTS IN FINANCIAL ECONOMIES.
It can be argued that bubbles are a somewhat permanent
feature of our international economy, characterized by large
volatile nancial capitals the controltheoretic concept of
fast feedback and high gains comes to mind. The long periods
of nancial stability such as the postwar prosperity that
lasted until the oil crisis, and the Great Moderation starting
in the 1980s are the long quiescent phases where the agents
have adapted to focus on what matters for the real economy.
And the bursts occur when the agents become greedy and
try to optimize their earnings on the basis of models which,
forgotten during the moderation, are at best imperfect.
In this light the policies of the rich countries Europe
in particular can be interpreted as an attempt to avoid
bubbles and bursts by keeping the economy running well
below its potential. The practice of detuning a controller to
avoid instability is well know among control practitioners and
well justied by theory. This is a charitable view of policies
whose motivations include such human foibles as haughtiness,
callousness, bureaucratic inertia, and biased selfinterest. It
may in a sense be considered cautious but fails to take into
account the suffering of the almost half of young Europeans
left unemployed by the caution. A cybernetic view also
reveals that the precaution is misguided: the source of bubbles
is not efciency in the human use of human beings the fact
that each person can nd meaningful employment in the eld
where they are best able to contribute. The source of nancial
bubbles, damaging as they are for the proper operation of a
modern economy, must be found, if our analysis is correct, in
the secondorder, or adaptive, feedback loop.
IV. ADAPTATION.
V. DIRECT OPTIMIZATION.
Without making claim of any sort of grand synthesis, we
advocate a direct optimization method based on the com-
putation of the barycenter of a sequence of evaluations or
measurements of a given function. The weight given to each
point x
i
is e
f(xi,ti)
, where f() is the function to be
optimized, x
i
is a point probed at time t
i
, and is a suitable
constant. We develop formulas that relate the derivative-free
barycenter update rule to gradient-type methods. Extensions of
note include optimization on Riemannian manifolds, and the
use of complex weightings, deriving intuition from Richard
Feynmans interpretation of quantum electrodynamics.
The direct optimization method is quite promising as a basis
for adaptation of various sorts, be that in adaptive control, in
estimation for multichannel communications, or.....
This being a paper about Norbert Wieners work, it cannot
fail to include a proper theorem. We prove some useful
results concerning the barycenter method in a continuous-time
framework, which is perhaps more complex mathematically
but affords a clearer view of the connections with more
traditional model-based optimization techniques.
REFERENCES
[1] S. Heims, John Von Neumann and Norbert Wiener: from
mathematics to the technologies of life and death, ser.
Science biography. MIT Press, 1982. [Online]. Available:
http://books.google.com/books?id=qr0dAQAAMAAJ
[2] C. E. Rohrs, L. Valavani, M. Athans, and G. Stein, Robustness of
continuous-time adaptive control in the presence of unmodeled dynam-
ics, IEEE Trans. Automatic Control, vol. 30, no. 9, pp. 881888, Sep.
1985.