When Can we Call a System Self-organizing?

Carlos Gershenson and Francis Heylighen

Centrum Leo Apostel, Vrije Universiteit Brussel
Krijgskundestraat 33, Brussels, 1160, Belgium
{cgershen, fheyligh}@vub.ac.be
http://www.vub.ac.be/CLEA

Abstract. We do not attempt to provide yet another definition of self-

organization, but explore the conditions under which we can model a system as
self-organizing. These involve the dynamics of entropy, and the purpose,
aspects, and description level chosen by an observer. We show how, changing
the level or “graining” of description, the same system can appear self-
organizing or self-disorganizing. We discuss ontological issues we face when
studying self-organizing systems, and analyse when designing and controlling
artificial self-organizing systems is useful. We conclude that self-organization
is a way of observing systems, not an absolute class of systems.

1 Introduction

There have been many notions and d efinitions of self-organization, useful in different
contexts (for a non-technical overview, see [1]). They have come from cybernetics [2,
3, 4, 5], thermodyna mics [6], mathematics [7], information theo ry [8], synergetics [9],
and others. Many people use the term “self-organizing”, but it has no generally
accepted meaning, as the abunda nce o f definitions suggests. Also, proposing such a
definition faces the philosophical problem of defining “self”, the cybernetic problem of
defining “system”, and the universal problem of defining “organization”. We will not
attempt to propose yet ano ther de finition of self-organizing system s. Nevertheless, in
order to try to understand these systems better, we will explore the following question:
which are the necessary conditions in order to call a system “self-organizing”? We do
so by combining insights from different contexts where self-organizing systems have
been studied.
The understanding of self-organizing systems is essential for the artificial life
com munity, because life obviously exhibits self-organization, and so should the
systems simulating it. As a mo st basic e xamp le, we ca n look at swarming or herding
behaviour, in which a disordered array of scattered agents gathers together into a tight
formation. Intuitively, it seems that the system of agents has become mo re
“organized”. But precisely what does this mean?
In the following section we explore the role of dynamics in self-organizing
systems. We provide examples of systems that are self-organizing at one level but not
at another one. In Section 3 we note the relevance of the observer for perceiving self-
organization. W e discu ss some dee per conceptual p roblems for understand ing self-
organizing systems in S ection 4 . In Section 5 we discu ss app lications in artificial self-
organizing systems, and when using this approach is appropriate. We draw concluding
remarks in Section 6.
2 The representation-dependent dynamics of Entropy
A property frequently used to characterize self-organiza tion is an increase of order
which is not imp osed by an external agent (not excluding environm ental interactions)
[1]. The m ost comm on way to form alize the intuitive notion of “order” is to ide ntify it
with the negative of entropy. The second law of thermodynamics states that in an
isolated system, en tropy can only decrease, no t increase. Suc h systems evolve to their
state of maximum entropy, or thermodynamic equilibrium. Therefore, self-organizing
systems canno t be isolated: they require a constant input of matter or energy with low
entropy, getting rid of the internally generated entropy through the output of heat
(“dissipation”). This allows them to produce “dissipative structures” which maintain
far from thermodynamic eq uilibrium [6]. Life is a clear example of order far from
thermodynamic equilibrium.
However, the therm odynamical concept of entropy as the dissipation of heat is not
very useful if we want to un derstand information-base d system s, such as those created
by ALife m ode llers. For that, we need the more general co ncep t of statistical entropy
(H) which is applicable to any system for which a state space can be defined. It
expresses the degree of uncertainty we have about the state s of the system, in terms of
the probability distribution P(s).

(1)

In this formulation, the second law of thermodynamics can be expressed as “every

system tends to its most probable state” [4]. This is in as sense a tautological law of
nature, since the probab ilities of the states are determined by us according to the
tendencies of systems. At a molecular level, the most proba ble state of an isolated
system is that of maximum entropy or thermodynamic equilibrium, where the
molecules are distributed homogeneo usly, erasing any structure or differentiation. But
does this apply as well to a real or artificial living organism?
W e have to be aware that probabilities are relative to a level of observation, and
that what is most proba ble at o ne level is not necessarily so at another. M oreo ver, a
state is defined by an observer, being the conjunction of the values for all the variables
or attributes that the observer considers relevant for the phenomenon being modelled.
Therefore, we can have different degrees of order or “entropies” for different models or
levels of observation of the sam e entity.
Let us illustrate this with the following, very simple exam ple. Co nsider a system
with four possible “microstates”, a1, a2, b1, and b2, at the lowest, most detailed level
of description. At the higher, more ab stract level of description, we aggregate the
microstates two by two, defining two macrostates: A = {a1, a2} and B = {b1, b2}. This
means that the system is in macrostate A if it is either in microstate a1 or in microstate
a2. The probab ilities of the macrostates are simply the sum of the probabilities of their
sub-states. Let us suppose that we start from an initial probability distribution so that
P(a1) = P (b1) = 0.1 and P(a2) = P (b2) = 0.4. T his implies P(A) = P(B) = 0.5 . W e can
calculate the statistical entrop y H using (1), getting H.1.72 at the lower level, and H=1
at the higher level.
Now consider a second distribution P(a1) = P (a2) = 0.2 while P(b1) = P (b2) = 0.3.
Therefore, P(A) = 0.4 and P(B) =0.6. Now we have H.1.97 at the lower and H.0.97 at
the higher levels. S ubtracting the initial H from the second, we have MH/Mt.0.24 at a
lower level and MH/Mt.!0.029 at the higher level. We have a change of distribution
where entropy is decreased at the higher level (“self-organization”), and increased at
the lower level (“self-disorganization”). To get the inverse change, we can just assume
the final states to be the initial ones and vice versa. W e would then have self-
organization at the lower level and self-disorganization at the higher level.
This can be represented graphically in Figure 1, where tones of gray represent the
probabilities of the states (darker colour = lower value). The macrostates provide a
coarse-grained [10] representation of the system, while the microstates provide a fine-
grained one. We can visualize entropy as homogeneity of colours/probabilities. At the
lower level, the d istribution becomes more homogeneous, and entropy increases. At
the higher level, the distribution becomes more differentiated.

Figure 1. Entropy, seen as homogeneity, increases at lower level, while it might

increase or decrease at the higher level, depending on how we divide the states

Entropy not only depends on higher or lower levels of abstraction, but also on how
we set the boundaries between states. Let us de fine alternative macrostates A’, which
has a1 and b1 as sub-states, and B’, with a2 and b2 as sub-states. Using the same
values for the probabilities, we see that for the alternative macrostates the initial
H.0.72 and the final H=1. So we have that MH/Mt.0.27, which means that the
statistical entropy increases in this macrorepresentation, while it decreases in the
previous one. The system appears to be self-organizing or disorganizing, depending not
only on the level at which we observe it, but also on how we do the “coarse-graining”
of the system, that is to say, which variables we select to define the states.
The variables defined by the v alues (A, B), respe ctively (A’, B’), represent two
aspects of the sam e system, where the observer has focussed on different, independent
prop erties. For example, a particle’s state includes both its position in space and its
momentum or velocity. A subsystem is defined as a physical part of a system, limited
to some of its components. Similarly, an aspect system can be defined as a functional
part of a system, limited to some of its pro perties o r aspects [11 ].
Let us illustrate this with a typical ALife model: swarming behaviour. Groups of
agents can be seen as subsystems of the swarm. The positions of all agents define one
aspect system, while their ve locities d efine ano ther asp ect system . Assum e we start
with non-moving agents scattered all over the simulated spa ce. T he po sition asp ect is
characterized by maximum entropy (age nts can be anywhere in space), while the
veloc ity aspect has m inimum entrop y (all have the same zero velocity). Acco rding to
typical swarm ing rules, the agents will start to move with varying speeds towards the
centre of the swarm while mutually adjusting their velocities so as not to bum p into
each other. T his means that their states become more concentrated in position space,
but more d iffuse in velocity space. In other word s, entropy decreases for the po sitions,
while increasing for the velocities. Depending on the aspect we consider, the swarm
self-organizes o r self-disorganizes!
T his example may appear too specific to support our argument. Let us therefore
show that dynamical processes in general exhibit this kind of aspect-dependen t or
level-dependent behaviour, either increasing or decreasing entropy. A dynamical
system, where every state is mapped deterministically onto a single next state, can be
seen as special case of a Markov process, where a state s i is mapped stocha stically onto
any number of other states s j with a fixed transition p robability P(s i 6 s j ). To turn a
deterministic dynam ics into a stocha stic one, it suffices to ap ply coarse-graining,
aggregating a numb er of microstates into a single macrostate. Transitions from this
macrostate can now go to any other macrostate that includes one of the microstates that
were the initial destinations of its microstates.
It can be proven that the statistical entropy of a distribution cannot decrease if and
only if the Marko v process that describes the mapping from initial to final distribution
is dou bly stocha stic [12]. This means that for the matrix of transition probabilities
between states, the sum over a row and the sum over a column must be one. The sum
over a row (probability of a given state end ing in any state of the state spac e) is
necessarily one, by definition of probability. However, the sum over a column is no t a
probab ility but can be seen as a measure o f the “attractiveness” or “fitness” F of a state
s:

(2)

High fitness (F > 1) of a state s means that on average more transitions enter s than
leave s (the sum of all transition probabilities leaving s can never be >1). Thus, while
the process runs, high fitness states become mo re prob able, and low fitness states less
probable, increasing their differentiation, as would be expected from a pro cess
undergoing self-organization.
A doubly stochastic process is defined by the requirement that F (s i ) = 1 for all
states s i . This corresponds to the entropy increasing processes studied in traditional
thermod ynamics. In a more general process, probability eventually concentrates in the
high fitness states, d ecrea sing overall entro py if the initial distribution is more
homogeneous, increasing it if it is more “peaked”. Therefore, a necessary and
sufficient cond ition for a Markov pro cess to allow self-organization is that it has a non-
trivial fitness functio n, i.e. there exist states such that F(s) … 1. The sum of all
probabilities must be 1. Therefore, for every F(s)>1, there must be at least one F(s’)<1.
This condition is equivalent to saying that the dynamics has a differential “preference”
for certain states s over other states s’. Any dynamics that allows attractors has such an
inbuilt preference for attractor states over basin states [1]. This is the most general
case, and the o ne typically found in ALife simulations.
For any initial and final distribution of pro bab ilities, such as the one s in the
examples we disc ussed, it is possible to determine a ma trix of transition proba bilities
that maps the one onto the other. This matrix will in general not be doubly stochastic,
and therefore allow self-organization as well as disorganization. Therefore, the same
system, described in different aspects or leve ls of abstraction can b e mo delled as self-
organizing in the one case, as self-disorganizing in another.
3 The Role of the Observer
W e have to be aware that even in mathematical and physical models of self-organizing
systems, it is the observer who ascribes properties, aspects, states, and prob abilities;
and therefore entropy or order to the system. But organization is more than low
entrop y: it is structure that has a function or purpose [13]. Stafford Beer [4] noted a
very important issue: what under some circumstances can be seen as organization,
under others can be seen as disorder, depending on the purpo se of the system. He
illustrates this idea with the following example: When ice cream is taken from a
freezer, and put at room temperature, we can say that the ice cream disorganizes, since
it loses its purpose of having an icy consistency. B ut from a physical point of view, it
becomes more ordered by achieving equilibrium with the room, as it had done with the
freezer1 . Again, the purpose of the system is not an objective property of the system,
but something set by an observer.
W . Ross Ashb y noted decades ago the importance of the role of the observer in
relation to self-organizing systems: “A substantial part of the theory of organization
will be concerned with properties that are not intrinsic to the thing but are relational
between observer and thing” ([3], p. 258, emphasis in original).
Of course there should be a correlate in the world to the observations. The question
now is: How frequent is this correlate, so that we can observe self-organization? W hat
we need is a collection of elements that interact. By generalizing the second law of
thermod ynamics, we can see that the system through time will reach a more “probable”
or “stable” configuration. We can say that it will reach an equilibrium or attractor 2 . The
observer then needs to focus his/her viewpoint, in order to set the purpose of the
system so that we can see the attractor as an “organized” state and to see it at the right
level and in the right aspect, and the n self-orga nization w ill be observed. W e can see
that this is much more comm on than what intuition tells us. Not on ly lasers, magnets,
Bénard rolls, ant colonies, or economies can be said to be self-organizing. Even an
ideal gas can be said to be self-organizing, if we say (contrary to thermodynamics) that
the equilibrium state where the gas homogeneously fills its container, is “ordered” [4].
Any dynamical system can be said to be self-organizing [3]. Self-organization is a way
of modelling systems, not a class of systems. This does not mean that the re is no self-
organization independently of the observer, but rather that self-organization is
everywhere.
Of course, not all systems are usefully described as self-organizing. Most natural
systems can be easily fit into the class “self-organizing”, unlike the simple mechanisms
we find in physics textbooks. Most artificial systems are hard to see as self-organizing.
Many are not dynamic, others involve only one element (actually no system), and mo st
of the rest follow sequence s of rules that ca n be easily understoo d. T herefo re there is
no need to explain their functioning with the rather esoteric conc ept of “self-
organization”.
W e have said that any dyna mical system, if observed “prop erly”, can be seen as
self-organizing. But if we set a different purpose or description level, then any

1
Thermodynamic entropy can be seen as order or disorder in different situations (e.g. [6,
4]. This also occurs with information entropy, as the debate between Wiener and Shannon showed.
2
In some chaotic systems, this can take practically infinite time. But as systems approach
an attractor, we can say that they follow this law.
dynamical system can be self-disorganizing. An eco nom y will not be seen as self-
organizing if we look only at a short timescale, or if we look at the scale of only one
small business. An ant colony will not be self-organizing if we describe only the global
behaviour of the co lony (e.g. as an element of an ecosystem), or if we only list the
behaviours of indiv idual ants. W e have to rememb er that the description of self-
organization is partially, but strongly, dependent on the observer.

4 Ontological Issues
One of the most common problems when discussing self-organizing systems is the
meaning of emergence. Self-organizing systems typically have higher level properties
that cannot be observed at the level of the eleme nts, and that can be seen as a product
of their interactions (more than the sum of the parts). Some people call these properties
emergent. The problem we face is o ntological. Accord ing to A ristotelea n logic, a
system cannot be more than one thing at a time. In this case, a system cann ot be at the
same time a set of elements and a whole with emergent properties. But by introducing
an ontological distinction [14], we can clarify the issue.
W e can distinguish two types of being: relative and absolute. The relative (rel-
being) is experienced by an observer with a finite cognitive capacity. It therefore
depends on her/his context, and is limited. Strictly speaking, every co gnizer has a
different rel-being of anything, since every co gnizer has a different co ntext.
Theoretically, we can assume that there exists an absolute being (abs-being), which
wou ld be “the real thing” (K ant’s Ding-an-sich), indep endent of the obse rver, which
observers correlate to their rel-beings. W e can o bserv e any abs-being from an infinity
of perspectives and describe an infinity of potential properties or aspects. Nevertheless,
most rel-beings and contexts are similar, since they are inspired by the same abs-being
seen by similar obse rvers fro m a similar point of view. This enables us to share
knowledge, but it is because of the different nuances in the different rel-beings and
contexts that we fail to agree in every situation.
W e can then say that the observation of a system at different abstraction leve ls or in
different aspects is merely a difference in the perspective, and therefore the system rel-
is different (only for the observers). But the system abs-is the sam e thing itself,
indep endently of how we describe it. W e can ob serve a cell as rel-being a bunch of
molecules or as rel-being a living struc ture. But it abs-is both and even mo re. Rel-
beings can be seen as different models or metaphors for describing the same thing. A
change in the metaphor d oes not cha nge the thing. If we define emergence as a p rocess
that requires a change of the model [15] in order to better understand and predict the
system [8], then it becomes clear that there is no magic. Any dynamical system abs-is
self-organizing and self-disorganizing at the same time, in the sense that it can
potentially rel-be both.
Another confusion may arise when people describe systems as the lower level
causing change in the emergent properties. Vice-versa, downward causation is the idea
that higher level prop erties constrain or control components at the lower level [16].
Speaking about causality between abstraction levels is not accurate [14], because
actually they abs-are the same thing. W hat we could say is that when we observe
certain conditions in the lower level, we can expect to observe certain properties at a
higher level, and vice versa. There is correlation, but not actual causation.
This leads us to wha t is probably the most fundamental problem. If we can describe
a system using different levels, aspects, or represe ntations, which is the one w e should
choose? As P rem [17] suggests, the level should be the one where the prediction of the
behaviour of the system is easiest; in other words, where we need least information to
make predictions3 [8]. A possible way to formalize this requirement is by choosing the
representation that minim izes the conditional entropy, i.e. the average uncertainty of
the next state given the present state [5]. We can speculate that this is possible because
of regularities in the system at that particular level, and that this is what leads pe ople to
try to describe how the simp le properties cause the not so simple ones, either upward or
dow nward .

5 Artificial Self-organizing Systems

Independently of the definition of self-organizing systems, if we see them as a
perspective for studying system s, we can use this perspe ctive for designing, build ing,
and controlling systems. A key cha racteristic of an artificial self-organizing system is
that structure and function of the system “emerge” from interactions between the
elements. The purpose should not be explicitly designed, programmed, or controlled.
The components should interact freely with each other and with the environment,
mutua lly adapting so as to reach an intrinsically “preferable” or “fit” configuration
(attractor), thus defining the purpose of the system in an “emergent” way [13].
Certainly this is not the only approach for designing and controlling systems, and in
many cases it is not appropriate. But it can be very useful in complex systems where
the observer cannot a priori conceive of all possible configurations, purposes, or
problems that the system may be confronted with. Examples of these are organizations
(corpo rations, governments, communities), traffic control, proteomics, distributed
robo tics, allocation of ecologic resources, self-assembling nanotubes, and complex
software systems [13], such as the semantic web.
For artificial life, we believe that the perspec tive of self-organizing system s is
essential, because we cannot explain the emergence, evolution, and development of
life, whether in vivo or in silico, by restricting our models to a single level of
abstraction. We need to und erstand how properties such as life, autocatalysis, or
autop oiesis can em erge fro m intera ctions o f elements without these properties, or how
species or social properties can arise from individual interactions. T his can only be
done from a perspective of self-organizing systems. Therefore, it is important that the
issues we have d iscussed are taken into account by artificial life researchers.

6 Conclusions
W e proposed that self-organizing systems, rather than a type of systems, are a
perspective for studying, understand ing, designing, controlling, and b uilding systems.
This perspective has advantages and disadvantages, and there are systems that benefit
from this approach, and others for which it is redundant. But eve n in the general case
when the systems dynamics allows self-organization in the sense of entropy decrease,
the crucial factor is the observer, who has to describe the process at an ap propriate
level(s) and aspects, and to define the purpo se of the system. All these “make” the
system to be self-organizing. In that sense, self-organization can be everywhere: it just
needs to be observed.

3
This argument could be also followed to decide which “graining” to choose.
 W e believe that this discussion on when and how to best mod el a system as self-
organizing should be carried further in the artificial life community, since we all study
and build systems from a self-organizing perspective. This would benefit not only the
com munity, but every dom ain whe re the no tion of self-organization is useful.

References
[1] Heylighen, F. (2003). The Science of Self-organization and Adaptivity, in: Knowledge
Management, Organizational Intelligence and Learning, and Complexity, in: The
Encyclopedia of Life Support Systems, EOLSS Publishers Co. Ltd.
[2] Von Foerster, H (1960). On Self-Organizing Systems and their Environments. In Yovitts,
M. C. and S. Cameron (Eds.), Self-Organizing Systems, Pergamon, pp. 31-50.
[3] Ashby, W. R. (1962). Principles of the Self-organizing System. In von Foerster, H. and G.
W. Zopf, Jr. (Eds.), Principles of Self-organization. Pergamon Press, pp. 255-278.
[4] Beer, S. (1966). Decision and Control: The Meaning of Operational Research and
Management Cybernetics. John Wiley & Sons, Inc.
[5] Heylighen, F. and C. Joslyn (2001). Cybernetics and Second Order Cybernetics, in: R. A.
Meyers (ed.), Encyclopaedia of Physical Science & Technology, Vol. 4 (3rd ed.), Academic
Press, pp. 155-170.
[6] Nicolis, G. and I. Prigogine (1977). Self-Organization in Non-Equilibrium Systems: From
Dissipative Structures to Order Through Fluctuations, Wiley.
[7] Lendaris, G. G. (1964). On the Definition of Self-Organizing Systems. Proceedings of the
IEEE, March.
[8] Shalizi, C. R. (2001). Causal Architecture, Complexity and Self-Organization in Time
Series and Cellular Automata. Doctoral Dissertation. University of Wisconsin at Madison.
[9] Haken, H. (1981). Synergetics and the Problem of Selforganization. In Roth, G. and H.
Schwegler (eds.). Self-organizing Systems: An Interdisciplinary Approach, Campus Verlag,
pp. 9-13.
[10] Hobbs J. (1985). 'Granularity', in : Proceedings of the 9th International Joint Conference
on Artificial Intelligence(vol. 1), p. 423.
[11] ten Haaf W., H. Bikker, D.J. Adriaanse (2002) Fundamentals of Business Engineering and
Management, A Systems Approach to People and Organisations, Delft University Press
[12] Koopman, B. (1978). Entropy Increase and Group Symmetry. In Tribus, M. and I. Levine
(Eds.) Maximum Entropy Formalism. MIT Press.
[13] Heylighen, F. and C. Gershenson (2003). The Meaning of Self-organization in Computing.
IEEE Intelligent Systems, section Trends & Controversies - Self-organization and
Information Systems, May/June 2003. [in press]
[14] Gershenson, C. (2002). Complex Philosophy. Proceedings of the 1st Biennial Seminar on
Philosophical, Methodological & Epistemological Implications of Complexity Theory. La
Habana, Cuba. Also in InterJournal of Complex Systems, 544.
[15] Rosen, R. (1985). Anticipatory Systems. Pergamon.
[16] Campbell, D. T. (1974). ‘Downward causation’ in Hierarchically Organized Biological
Systems, In Ayala, F. J. and T. Dobzhansky (Eds.), Studies in the Philosophy of Biology,
Macmillan, pp. 179-186.
[17] Prem, E. (1993). Understanding Self-Organization: What can the speaking lion tell us?
Oesterreichisches Forschungsinstitut fuer Artificial Intelligence, Wien, TR-93-14.