The Economy as a Distributed, Learning Control System

Abstract

The perceptual control theory of Powers is used to analyse the market mechanism as a negative feedback loop which controls the deviation between demand (goal) and supply (perception) by adjusting the amount of effort invested in the production process (action), through the the setting of the price. The interconnection of distributed control loops for the different products and services facilitates the allocation of production factors over the different products. The resulting global control system becomes more efficient by learning how to be more sensitive to deviations from the goal, and less dependent on the availability of resources. In that way, it resembles the nervous system of a supra-individual organism.

Full text

Page 1

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
The Economy as a Distributed,
Learning Control System
Francis HEYLIGHEN
PO, Free University of Brussels, Pleinlaan 2, B-1050 Brussels, Belgium
E-mail: fheyligh@vnet3.vub.ac.be
ABSTRACT. The perceptual control theory of Powers is used to analyse the
market mechanism as a negative feedback loop which controls the
deviation between demand (goal) and supply (perception) by adjusting the
amount of effort invested in the production process (action), through the
the setting of the price. The interconnection of distributed control loops for
the different products and services facilitates the allocation of production
factors over the different products. The resulting global control system
becomes more efficient by learning how to be more sensitive to deviations
from the goal, and less dependent on the availability of resources. In that
way, it resembles the nervous system of a supra-individual organism.
Introduction
When discussing social systems, it is fashionable to oppose "controlled"
systems, like the plan-based economies in communist countries, to "self-
organizing" systems, like the free market economies, which are governed by
an "invisible hand". The net effect of the invisible hand, though, is a
constraint on the economic actors, compelling them to produce specific
goods for specific prices. The present paper will argue that this constraining
action is functionally equivalent with control, as it is defined in cybernetics.
The main difference between market control and government control will
then be that the former is distributed over a multitude of interacting actors,
whereas the latter is centralized in a single institution.
To analyse control in market systems, we will use the control theory of
William T. Powers (1973, 1989), which was developed to model
autonomous, living systems, rather than mechanical or engineering
systems, which are controlled from the outside. As such, our control
theoretic model will fit into the more general view of socio-economic
systems as analogous to organisms. Several authors have argued that
human society can be viewed as a superorganism, i.e. an organism of
organisms (e.g. de Rosnay, 1975, 1996; Stock, 1993; Turchin, 1977). Powers
(1992) himself, on the other hand, would restrict the use of control theory to
individual organisms, arguing that social systems are essentially different.
Arguments for and against the interpretation of social systems as control
systems have been discussed by Heylighen and Campbell (1995). The present
paper will add another argument to the case for the control interpretation.

Page 2

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
As has been shown by Conant and Ashby (1970), the existence of control
implies the presence in the controller of a model of the system being
controlled. Therefore, if we can show that society is a distributed control
system we will also have established the presence of distributed cognition,
i.e. knowledge which is not inherent in any particular individual or
institution but in the network of control loops that makes up the economy.
In this generalized sense, all organisms, including bacteria and plants,
possess knowledge. However, in general we will only speak about cognition
in higher organisms, which have a central nervous system capable of
learning, i.e. developing new knowledge. The final sections will suggest that
the "economic organism" has such a nervous system, similar to a "global
brain" (Russell, 1996; Mayer-Kress & Barczys, 1995).
Before discussing these issues, I wish to note that I am not an economist,
but a cybernetician working on complex systems and cognition. As such
some of the remarks I make specifically about market mechanisms may
appear inaccurate or old hat to economists. Yet, I believe that a cybernetic
point of view may put some well-known phenomena in a different light,
and as such clarify old problems. I hope that my less-than-expert analysis of
economic ideas will not detract the readers from appreciating the promises
of a novel viewpoint.
Powers's control scheme
In order to analyse markets as control mechanisms, we need to agree about a
definition of control. Such a definition can be derived by abstraction from
the example of the thermostat, which in cybernetics is viewed as the
prototype of a control system. The thermostat controls the temperature in a
room, keeping it within tight bounds around its reference temperature,
which plays the role of a goal. This is achieved by counteracting all
deviations or disturbances from that goal state, which are caused by
fluctuations in outside temperature. When the temperature becomes lower
than the reference temperature, the thermostat senses the change
(perception) and switches on the heating (action). This increases the
temperature until the reference level is reached. At this moment, the
heating is switched off again. This elementary control system is also an
elementary cognitive system, since in order to function it needs to "know"
how to evaluate its perceptions, and how to choose the right action for each
of its perceptions (Heylighen, 1990; De Mey, 1982). The two states of the
thermostat, "on" and "off", in fact model the two states of the world,
"temperature too low", and "temperature high enough".
William Powers, the founder of Perceptual Control Theory (1973, 1989),
has proposed the following more general scheme to represent such negative
feedback control mechanisms (Fig. 1).

Page 3

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
disturbances
perception
action
reference signal
p
a
r
d
Fig. 1: Powers's control scheme.
The scheme consist of a feedback loop with two inputs, the reference signal r
and the disturbance d. The arm of the loop going from r to d is called the
action, a. The one going from d to r is called the perception, p. The relation
between these elements should be such that the perception is brought as
closely as possible to the reference signal, by the action compensating for the
effect of the disturbance. In the simplest case, each component can be
represented by a one dimensional variable, and the relation between them
by two linear expressions:
a = K (r – p)
p = E (d – a)
(1)
(2)
The action is proportional to the difference between the reference level and
the perceived level. The larger this deviation ("error") between perception
and goal, the larger the correcting action needed. Similarly, the perceived
state is proportional to the difference between the disturbance and the
compensating action. K and E are two constants expressing characteristics of
the control system. To better understand their meaning, the system of
equations can be solved for a and p. This produces the following
expressions:
a = KE(d– r/E) /(KE – 1)
p = KE(r – d/K)/(KE – 1)
(3)
(4)
Until now, the whole scheme is symmetric with respect to permutations of
(r, a) with (d, p). In Fig. 1, we can turn the loop upside down, and then the
reference becomes the disturbance and the action becomes the perception.
However, this does not accord with our intuitive understanding of control
as an asymmetric relation, in which the controller (r and a) controls the
perturbations (d and p), but not the other way around. The asymmetry can

Page 4

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
be introduced by assuming that K is very large, and much larger than E. In
that case, through equation (4), p becomes almost equal to r:
K » E then p ≅ r (5)
This means that control is achieved: the deviation between goal and
perception is practically reduced to zero. Whatever the size of the
disturbance d, the control system manages to keep the state very close to its
preferred state r. K expresses the factor of "amplification": the smallest
deviation (r – p) is immediately sensed and compensated by a large action
a = K(r – p). Although the action is large relative to the deviation, it is in
general it is not large in absolute terms (as expressed by eq. (3)). The reason is
that because of the negative feedback relation the deviation is immediately
compensated before it can grow larger. Thus, since deviations remain small
(by equation (5)), really large actions will never be needed. The amplification
factor measures both the "power" of the control system, in the sense of the
amount of energy it has available to take action, and its "sensitivity", in the
sense of the precision with which it reacts to the tiniest deviations.
Since the scheme is symmetric in E and K, we can make a similar
interpretation of the implications of a large factor E. This would mean that
the perception is very sensitive to disturbances, in other words that small
perturbations would make the system deviate strongly from the reference
level. This describes a situation with poor or no control. In conclusion, the
larger K with respect to E, the less the perception will deviate from the goal,
and the better the control.
Let us go back to the example of the thermostat. This does not perfectly
fit in the above scheme, since the scheme is continuous: actions will
continuously increase or decrease depending on the sensed deviations.
Traditional thermostats, on the other hand, have discrete actions: either the
heating is switched on or it is switched off. Let us instead assume that a
corresponds to a continuously varying intensity of heating. In that case, r
corresponds to the reference temperature of the thermostat, d to the heat
loss caused by the lower outside temperature, and p to the actual
temperature in the room. K will then basically measure the power of the
heating installation: the more heat is produced per time interval, the
quicker the deviation (r – p) will be corrected, and the smaller it will remain.
E will then measure the sensitivity of the room to heat flow: the larger the
room or the better the insulation, the smaller E, the less the outside
temperature will influence the inside temperature p, and the less corrective
action a the thermostat will need to take. We noted that in a traditional
thermostat the state of the environment is mapped onto a discrete variable
with two possible values, "on" and "off". In the continuous thermostat, or
in Powers's scheme, on the other hand, the control system has a model of
the world consisting of a one dimensional variable with an infinite number
of possible values.
The supply and demand cycle
Let us try to apply this control scheme to the market mechanism. We need
to find a cycle or feedback loop with two or more inputs, one of which can be
identified as a goal or reference level, while the others can be interpreted as
disturbances. As shown by Richardson (1991, p. 62), Adam Smith's original

Page 5

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
represented by such a feedback cycle. I have slightly rephrased the
components of this loop so that the parallelism with Powers's control
scheme becomes most clear (see Fig. 2).
scarcity of
resources
effort
demand
p
a
r
d
supply
Fig. 2: the market mechanism as a control loop.
The demand for goods and services plays the role of a goal or reference
level, a desire that the market tries to satisfy. We here understand demand
as the amount of commodities the consumers are capable and willing to
purchase, independently of whether these commodities are available or not.
(Note that this is different from the way demand was understood in an
earlier paper (Heylighen & Campbell, 1995), where it reflected what
consumers want on top of what is already available. In the present paper
this would seen as the deviation or error signal. This different
interpretation explains why the earlier paper did not treat demand as an
input but as a dependent variable, and therefore did not recognize it as the
goal of a control loop).
The supply is what is really available to satisfy the demand. The
difference between demand and supply determines the error signal. This is
reflected by the price of the commodity: the larger the discrepancy, the more
expensive the good. If there is no disparity, i.e. if any demand whatsoever is
completely and immediately fulfilled, the price is nil. An example of such a
commodity would be air, which is everywhere plentiful. (The fact that air is
normally free does not mean that people would not be willing to pay for it
in a situation where it would be rare, e.g. while diving or on a space ship.
Therefore, in the present usage of "demand", there is always a demand for
air.) In some countries, tap water is free as well. This is an exceptional
situation, however: for most commodities, there will always be moments
when consumers would like to have more than is presently available. In
economics, this deviation from the ideal or goal state is called "scarcity".
The larger the discrepancy between what the consumers want and what

Page 6

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
increases the incentive for producers to produce more of the scarce good.
Therefore, they will invest more capital, labor and research (in short,
"effort") in production. However, in order to produce goods, they depend on
the availability of resources (e.g. raw materials, energy, components) over
which they have no direct control. The more scarce these resources, the
more difficult it will be to produce a sufficient amount of goods. Therefore,
the scarcity of resources plays the role of an outside disturbance. Effort and
resources together (the "factors of production") determine the total
production and therefore the supply.
Although this scheme is of course highly simplified, ignoring many
other factors and interactions, the negative feedback loop with two
independent inputs, consumer demand and availability of resources seems
well established. In order to determine whether this is really a control
system, we need to investigate whether there is amplification. Let us
therefore try to map the different components of the supply-demand loop
onto the variables in equations (1)-(4), keeping in mind that we are working
with a highly idealized model. Equation (1) then merely expresses the fact
that the amount of effort invested in production will be proportional to the
difference between demand and supply, as signalled by the price. Equation
(2) seems at first sight counterintuitive, until we note that the signs of the
different variables are purely conventional. By replacing E by –E in (2), we
get a formula that tells us that the supply will be proportional to the effort
invested minus the scarcity, or, in other words, to the effort plus the
availability of resources.
Now we need to estimate the relative size of the two constants E and K.
Let us consider an example. Imagine a poor Ethiopian farming community
dependent on water for growing their crops. The amount of water they need
on the land determines the demand. The amount of rain determines the
availability of resources. The supply will depend on the rain and on the
effort invested in irrigation, pumping, combating evaporation, etc.
Obviously, in this case, whether supply meets demand will depend much
more on the natural phenomenon of rain than on human effort. There is
little control, and K is small with respect to E.
Now imagine a similar scenario in a rich, technologically developed
region, e.g. California. The methods for irrigation and pumping are much
more sophisticated, getting water from much farther away with much less
loss or evaporation. In this case too, a drought will increase the price of
water, but the capability of the producers to respond to the demand will be
much less dependent on the natural availability. K will be much larger than
E, and the system will have much better control over the amount of water
needed for agriculture. The slightest increase in the need for water will
immediately lead to a larger volume of water being pumped to the farms.
Let us analyse in more detail which factors determine amplification and
therefore control. High K means high sensitivity to changes in the
difference between supply and demand. These changes will be expressed in
the fluctuations of the price. The price is determined in the market, the
place where producers and consumers, buyers and sellers, meet. By coming
in contact with many consumers, producers will get an idea of how much
demand there really is. By looking at the proposals of competing producers
they will get an idea of how much supply there really is. By comparing the
average demand with the average supply they will get an idea of what price
they can realistically ask for their good or service. Sensitivity means that

Page 7

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
they will immediately raise prices if demand increases, or diminish prices if
supply increases, thus affecting the investment in further production.
The mechanism is not so simple or automatic as it is usually portrayed.
Imagine a small market in an Ethiopian village where there are just a few
buyers and a few sellers of different agricultural products. Prices will
fluctuate strongly depending on the bargaining power of the consumer the
producer happens to be discussing with. If suddenly a rich consumer with a
high need appears, prices will jump up. Although this may seem like high
sensitivity, what is sensed is in general not a reliable measure of the overall
need of the Ethiopian population. There is simply too little information
about the demands and supplies of the people who happen not to be present
on that particular day on that particular market. The fact that a producer
managed to sell his produce for a high price today, does not in the least
imply that he will be as lucky the next day. Therefore, there is little
incentive for him to immediately increase his production.
Now, by way of contrast, imagine the world-wide, computer controlled
stock exchange. Any order to buy a particular stock, coming from any part of
the world, will immediately, in whatever small amount, raise the price of
that stock. Any order to sell will similarly lower the price. The effective price
is a continuously adjusted average resulting from hundreds of thousands of
buy/ sell decisions. If we ignore the effects of speculation at this point, the
price of a stock is an extremely sensitive and reliable measure of overall
demand and supply. As soon as the price starts moving up, producers will
start raising their production. The difference with the Ethiopian market is
the immediate transmission and integration of information from all
different parts of the market.
In addition to sensitivity, the amplification factor K, as we said, depends
on the "power" of the control system to act. In this case, power corresponds
to the amount of production factors (capital and labor) that is available to
increase production. The richer the economy, the more capital will be
available for investment, and the more quickly new production lines can be
started up. (This again distinguishes the Ethiopian economy from the
Western one.) This capital should not only exist, it should also be able to
quickly move to the place where it is needed. This is usually called
"liquidity" in economics.
Let us now look at the factor E, which measures the effect of
disturbances on production. A good control system will be characterized by
low E, i.e. low sensitivity to fluctuations in the amount of resources. This
can be achieved in two ways. First, resources or products can be stockpiled.
This creates a reserve, buffering against changes in availability. For example,
if water is collected in a large reservoir, the distribution of rain fall over the
year will have little effect on the daily distribution of water. Second,
productivity can be increased. This means producing more output (goods,
services) with less input (resources, labor). Buckminster Fuller (1969) noted
that technological advances lead to a continuing decrease of the input
needed to produce a given output. He called this evolution towards doing
more with less "ephemeralization". Low buffering can be compensated by
high productivity. When productivity is high, a relatively small number of
goods need to be kept in store, since it is easy to produce more when
demand increases. This principle underlies the trend towards "just in time"
production.
In conclusion, it is clear that technological progress and increasing
wealth lead to an ever tighter control over supply by the consumer's

Page 8

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
demand, making supply virtually independent of fluctuations in the
availability of resources. For example, in developed economies fruits and
vegetables remain available throughout the year, while the seasonal
fluctuations in price tend to become smaller and smaller. Although it could
be debated whether a primitive economy is really a control system, there is
little room for such argument when discussing developed economies.
Distributed vs. centralized control
In our simple model, we have represented supply and demand by one-
dimensional variables. In reality, there are different supply and demand
variables for each of the millions of goods and services provided in society.
The supply of each of these goods will be governed by a control loop.
However, these loops are interdependent. One thing they have in common
is the amount of effort and resources invested in the production of any one
good. The total amount of effort and resources available in a society is in
first approximation constant. What is used for the production of one
commodity cannot be used simultaneously for the production of another
commodity. The strength of the market is its capacity to efficiently allocate
these "factors of production", i.e. to distribute them over the different
control loops.
The principle is simple: product control loops where the deviation
between supply and demand is large will be characterized by high prices and
will therefore attract more factors of production. Those where the deviation
is relatively small will attract only few factors of production. However, if
production would decrease too much because of that, the deviation and
therefore the price will go up, thus again attracting more production factors.
This allocation of production factors can be seen as a higher order, collective
control system, which coordinates the control loops for the individual
products. Its goal is not to let any deviation become too large with respect to
other deviations. Its action consists in the movement of production factors
from low deviation loops to high deviation ones.
This global coordination of control loops is complemented by local
interactions between control loops. For example, if the demand for cars
increases, this will spur car manufacturers to assemble more cars. However,
to do that they will need more components. This will increase the demand
for goods such as tyres and steel. The control loop for tyres will thus be
subordinated through its reference level to the control loop for cars (see Fig.
3). The control loop for steel, on the other hand, will only be partially
dependent on the demand for cars, since steel is also needed for other
products, like cans, buildings, bridges, etc. All control loops are thus linked
via different interactions, where the reference level or resources for one
control loop are partially or completely determined by the functioning of
other control loops.

Page 9

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
demand for cars
demand for tyres
supply of
rubber
supply of other
car components
supply of tyres
supply of
cars
Fig. 3: example of an interaction between two control loops, where the right loop, controlling
for tyres, is subordinated to the left one, controlling for cars.
Such a fluid regulation of investment and exchange is difficult to
achieve in a more traditional, centralized control system. Centralized
economies need to plan how much they want to invest in each of the
different types of production. (Planning control is based on feedforward
rather than feedback). This leads to the well-known "calculation problem":
an enormous amount of information needs to be collected and processed in
order to find the optimal allocation of investment over all the different
production lines. In the distributed, market system, decisions to shift
investment from one product to another are made locally, without
knowledge of the global distribution of investment. It is the control loop
which will reward (and thus reinforce) good investments, and punish (and
thus inhibit) bad investments. Thus, investments are continuously flowing
from one place to another while staying only in those places where there is a
large enough demand with respect to the supply. This guarantees a
relatively balanced distribution of labor and capital, with a minimal need for
information processing.
Does the efficiency of distributed market control mean that there is no
need for control by the government, as the laissez-faire economists argue (cf.
Heylighen & Campbell, 1995)? Not at all. The efficiency of the market is in
making sure that demand is met. However, demand is a given variable, a
"reference signal" which enters the control loop from outside, but which is
in general not itself under control. This would not be a problem if demand
would represent the objective needs of society. In practice, however,
demand represents what consumers are ready to pay for a certain good or
service. This obviously depends on how much money the consumers
possess. Poor consumers may have high needs, but they contribute little to
the demand for the products they need. In the limit, a person with no
money would effect no demand whatsoever, and would therefore be
completely ignored by the control loop.
The fact that money is limited causes other paradoxical effects. In the
Giffen paradox, increasing the prize of a particular good, say bread, may
actually increase its sales, in contradiction with the traditional law of supply
and demand. Imagine a situation where poor people divide their fixed food

Page 10

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
means that they will have less money left to buy meat. But since they need a
minimum amount of calories to survive, they will compensate for less
meat consumption by consuming more bread, thus increasing the demand
for bread.
Wealth and poverty tend to be self-reinforcing. Poor people do not have
the means to get a good education, to start a business, to invest money, or do
any of the other things which typically lead to high income. This vicious
circle or positive feedback loop can only be broken by outside intervention,
e.g. by the government providing free education, or subsidizing business
start-ups.
Positive feedback is a more general problem affecting the demand
variable. A traditional example is speculation, where an increasing demand
makes people anticipate even higher prices, and thus gains for goods or
assets bought now and sold later. Another example is technological
development, where increased adoption of an innovative technological
product (say a particular type of video cassette or computer operating
system) diminishes the costs needed to produce further units, and
encourages the public to use that technology, rather than less well-spread
competing designs. Thus, increase of supply leads to increase of demand
("increasing returns", see Arthur, 1989). On the other hand, demand for the
competing design will diminish, even though that design may be
objectively superior (e.g. the Betamax standard for video cassettes).
Such positive feedback loops can severely distort or destabilize the
economic system. The only way not to let them get out of hand is to control
the demand. This is typically done by governments through taxes or
restrictions when they want to reduce demand (e.g. for alcohol, tobacco or
polluting products, or during currency speculation), or through subsidies
when they want to strengthen demand (e.g. for new technologies, public
transport or education). Controlling the reference variable in a control loop
defines a second order control system, i.e. a metasystem (Heylighen, 1995).
Such a system is potentially much more flexible and intelligent than a first
order system. It seems that government control is more effective when it is
applied at the metalevel (the demand, i.e. the goal of the market control
loop), than at the object level (the production, i.e. the means to achieve the
goal). The reason is that control at the metalevel is much simpler, since all
object level fluctuations in supply are already minimized by the highly
effective market.
Controlling the negative effects of market mechanisms seems more
effective by feedback control of its reference levels than by completely
substituting it with a planning economy. This seems to be the lesson that
can be drawn from the collapse of communist regimes and the success of
socially adjusted market economies. The question remains open which is
the most efficient way to implement such second order control of the
economy. Perhaps a more distributed system may here too be worth
exploring. For example, it has been proposed to control pollution by
providing firms with a limited number of tokens, each entitling the owner
to generate one unit of pollution, which can be bought or sold. In such a
system, highly polluting factories would need to buy a lot of tokens from
less polluting factories. This would make them less competitive and thus
spur them to become less polluting themselves.

Page 11

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
The learning control system
The picture of the socio-economic system we see emerging is that of a
complex control system, consisting of an intricate network of connected
control loops for the different products and services. The overall goal of the
control system is to satisfy the needs of the society, both directly through the
demands consumers make in the marketplace, and indirectly through the
government, which tries to regulate the individual demands to better reflect
long-term, collective needs. The perturbations which the system tries to
minimize are the internal and external fluctuations of different resources
which the system needs for its survival and growth. In this view, the socio-
economic system is similar to a living organism.
Like all control systems, the socio-economic system needs specialized
knowledge to control the perturbations. Indeed, every combination of
values for the availability of the different resources requires a specific
combination of counteractions, and the system must know which action
corresponds to which perturbation (Heylighen, 1994). The mapping from the
set of all possible perturbations to the set of all possible counteractions
defines a model of the perturbations (Conant and Ashby, 1970), and
therefore a model of the environment in which the control system tries to
survive. In artifical control systems, such as thermostats and automatic
pilots, the model was created by the human designer. In primitive
organisms, such as plants and bacteria, this model is carried by the genes,
which are the result of blind variation and natural selection.
The most sophisticated control systems, such as higher animals and
people, however, are moreover able to learn, i.e. to improve their control
models through experience. (Powers, 1974, 1989 calls this process
"reorganization"). In animals, control is implemented by the nervous
system, which connects sensors (sensitive cells in the eye, ear, skin, etc.),
responsible for perception, to effectors (muscles), responsible for action. In
the simplest organisms, the connection of sensor to effector is direct, like in
a simple control loop (Turchin calls this the level of simple reflexes, 1977).
In more complex organisms, though, signals coming from different external
and internal sensors need to be integrated and processed by a central
nervous system in order to decide about the best combination of signals that
need to be sent to the effectors (Turchin calls this the level of complex
reflexes, 1977). This seems similar to a market with many different products
and services, where the demands for different products are processed
together in order to efficiently allocate the production factors.
At the next level of complexity, the network of nerves processing the
information will change its pattern of connections depending on its
experience. The more complex and changing the environment, the greater
the need to continuously adapt and fine-tune the different connection
strengths. Some of the mechanisms to change connection strengths, such as
back propagation and Hebbian learning, are well known through computer
simulations of neural networks (Rumelhart & McClelland, 1986). We might
speculate that similar mechanisms underlie the development of an
economy. As a first step towards elaborating and testing this hypothesis, we
may look for analogies between individual learning and socio-economic
learning, which may provide inspiration
sophisticated theory.
The simplest level at which learning seems to occur is in the evolution
of the control constants E and K. Like we showed in our example of the
for developing a more

Page 12

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
with social and technological development. Since K measures sensitivity of
actions to deviations from the reference level, the process seems similar to
sensitization, the process by which an organism becomes more sensitive to
important stimuli. Similarly, the decrease of E, which measures the
sensitivity to outside perturbations, seems analogous to habituation, the
process by which an organism learns to ignore irrelevant stimuli. This
analogy is quite superficial, though, since the technological development
because of which E and K change is much more complicated than the
mechanism underlying habituation and sensitization in neurons.
Habituation and sensitization take place on the level of the simple
control loops. Learning on the level of complex networks is based on the
process of association, the strengthening of connections between neural
pathways. A classical learning mechanism is reinforcement, which
underlies the more complex mechanism of back propagation in neural
networks. Reinforcement strengthens neural connections that contribute to
achieving the organism's goals, while weakening those that move it away
from those goals. An analogue of such a connection in the economy would
be the dependence of one control loop on the products of another loop, like
in the example of the car manufacturers needing steel.
Suppose that car producers can either use steel or plastic to build their
cars. This determines two connections, from plastic to cars and from steel to
cars. Suppose that thanks to new technologies the producers using plastic
manage to produce more cars than those using steel, for the same amount of
investment and for the same quality of the final product. In that case the
connection from plastic to cars will be reinforced, while the one from steel
to cars will be weakened. One way to represent such learning might be to
assume that the constant Ecar-plastic measuring the sensitivity of car
production to plastic supply decreases, relative to the constant Ecar-steel
measuring the sensitivity to steel supply. This would increase the demand
for plastic coming from car manufacturers, while decreasing its demand for
steel to a larger extent. The resulting overall lower sensitivity to plastic or
steel supply means amplification of car production, and therefore better
satisfaction of the demand for cars. This is a reinforcement of the shift in the
connection strength.
A second, more indirect mechanism of association is Hebbian learning,
where the connection between neural pathways is strengthened when the
two pathways are activated simultaneously. For example, assume that each
time the demand for cars increases the demand for petroleum also increases.
There is no direct dependency of petroleum production on car production,
or vice versa. Yet, the petroleum producers would be wise to watch the
demand for cars, and to increase their production as soon as they see the
demand for cars increase. This would allow them to anticipate the increase
in demand for petroleum which is likely to follow the increase in demand
for cars.
Discussion
The perceptual control model seems to offer many promising insights into
the functioning of economic systems. It in particular may provide a more
solid, possibly even mathematical, basis for the approaches which view
societies as similar to organisms. However, this will require a much more
detailed analysis of the functioning of market mechanisms in terms of
control loops. In particular need of study are the precise definition of the

Page 13

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
different variables of the control loop, such as supply, demand, scarcity,
effort, the quantitative dependencies of these variables (as expressed in a
simplified way by the constants E and K), the connections between the
control loops for different products and services and the role of metalevel
control by the government.
An even more challenging issue for further research is the cognitive
development of the economic control system. The analogies with neural
mechanisms of learning we proposed barely scratch the surface of this
problem domain. Further inspiration may be found in some more detailed
models for supra-individual learning via computer networks which we
proposed earlier (Heylighen & Bollen, 1996; Bollen & Heylighen, 1996).
Computer and communication technologies provide particularly effective
ways to implement an efficient "nervous system" for the socio-economic
organism. We already discussed the example of the computer-controlled
stock exchange to illustrate the enormous gains in sensitivity that can be
achieved in that way.
In addition to the software agents and learning webs which are discussed
in our earlier papers, the software technology based on "workflow" holds
particular promise. The main idea behind workflow is that tasks, belonging
to a larger project, are automatically sent by the computer system to the
person or group most competent to handle them. Once a part of the job has
been finished, the provisional results together with the subsequent
questions are forwarded to the next one in line, and so on. For example, the
manufacture of cars requires the completion of a number of subtasks, such
as buying tyres, steel, plastic, etc. An automated computer system, when
receiving an order for a certain amount of cars would automatically order
the required amounts of tyres and other components to the suppliers that
make the best offer, and continuously monitor in how far the orders are
fulfilled. By comparing orders fulfilled by different producers, it might even
learn which suppliers are most reliable and efficient and take this into
account when making further orders. Such technologies could make the
socio-economic system much more efficient and intelligent, providing it
with a real "global brain" (Heylighen & Bollen, 1996; Russell, 1996; de
Rosnay, 1996; Mayer-Kress & Barczys, 1995). Further development of the
control system model of the economy may help us to support that
development.
Acknowledgments:
The author is supported as a Senior Research Assistant by the FWO (Fund
for Scientific Research-Flanders)
References
Arthur W. B. (1989): "Competing Technologies, Increasing Returns, and Lock-in by Historical
Events", The Economic Journal 99, pp. 106-131.
Bollen J. & Heylighen F. (1996): "Algorithms for the self-organisation of distributed, multi-
user networks. Possible application to the future World Wide Web", in: Cybernetics and
Systems '96 R. Trappl (ed.), (Austrian Society for Cybernetics), p. 911-916.
Buckminster Fuller R. (1969): Utopia or Oblivion (Bantam, New York).

Page 14

Heylighen F. (1997): "The Economy as a Distributed, Learning Control System", Communication & Cognition-
AI 13, nos. 2-3, p. 207-224.
Conant R. C. & Ashby W.R. (1970): "Every Good Regulator of a System Must Be Model of
that System", Int. J. Systems Science 1:2, p. 89-97.
De Mey M. (1982): The Cognitive Paradigm (Reidel, Dordrecht)
de Rosnay J. (1975): Le Macroscope (Seuil, Paris)
de Rosnay J. (1996): L'Homme Symbiotique (Seuil, Paris),
Heylighen F. (1990): Representation and Change. A Metarepresentational Framework for
the Foundations of Physical and Cognitive Science, (Communication & Cognition, Gent).
Heylighen F. (1994) “Fitness as Default: the evolutionary basis for cognitive complexity
reduction”, in: Cybernetics and Systems '94, R. Trappl (ed.), (World Science, Singapore),
p.1595-1602.
Heylighen F. (1995): "(Meta)systems as constraints on variation", World Futures: the Journal
of General Evolution .45, p. 59-85.
Heylighen F. & D. T. Campbell (1995): "Selection of Organization at the Social Level:
obstacles and facilitators of Metasystem Transitions", World Futures: the journal of
general evolution 45:1-4 , p. 181.
Heylighen F. & J. Bollen (1996): "The World-Wide Web as a Super-Brain: from metaphor to
model" in: R. Trappl (ed.): Cybernetics and Systems '96 (Austrian Society for Cybernetic
Studies), p. 917.
Mayer-Kress G. & C. Barczys (1995): "The Global Brain as an Emergent Structure from the
Worldwide Computing Network", The Information Society 11 (1).
Powers W. T. (1973): Behavior: the Control of Perception, (Aldine, Chicago).
Powers W. T. (1989): Living control systems, (Control Systems Group, Gravel Switch: KY).
Powers W. T. (1992): "Control Theory Psychology and Social Organizations", in: Living
Control Systems II , Powers, W. (ed.), (New View Publications, Chapel Hill, NC), p. 91-
127.
Richardson G.P. (1991): Feedback thought in social science and systems theory (Univ.
Pennsylvania Press, ).
Rumelhart D.E. & McClelland J.L. (eds.) (1986): Parallel Distributed Processing:
Explorations in the Microstructure of Cognition, (Bradford Books/ MIT Press, Cambridge,
MA).
Russell P. (1996): The Global Brain Awakens: Our next evolutionary leap (Global Brain).
Stock G. (1993): Metaman: the merging of humans and machines into a global superorganism,
(Simon & Schuster, New York, ).
Turchin V. (1977): The Phenomenon of Science. A cybernetic approach to human evolution,
(Columbia University Press, New York).