A Way to Reify Emergent Phenomena in Multiagent Simulations?
Daniel David and R
emy Courdier
EA2525 - LIM - IREMIA, University of La R
2 rue Joseph Wetzel, Sainte-Clotilde, R
eunion, France
Emergence, Simulation, Multiagent Systems, Metaknowledge.
Emergence is a fascinating concept for most scientists, and multiagent simulations are known to allow and
facilitate its representation. Research in this area yield to several definitions and classifications of emergent
phenomena, but only a few of them offers a solution for a concrete reification of emergence in simulation.
This paper deals with this important notion of emergence reification that, as we know, does not have yet formal
mathematic definition, if any could be expressed. We need to progress on the conceptual meaning, leading to
more global definitions but allowing to give a general conceptual framework that makes possible the reification
of emergent phenomena in multiagent simulations. We define emergence as being a metaknowledge and we
present a conceptual framework in which emergent phenomena can be detected and injected into simulation
systems and be handled like other entities.
Simulation of natural and social systems is a cross-
disciplinary activity in which computer scientists
work with many other researchers who are the experts
of their fields (geography, biology, economics, etc.).
These experts, that we call thematicians, define the
models that lie at the core of any simulation: they in-
put the properties and functional descriptions of the
system entities. Their goal is to use computer simu-
lations to imitate operations of real-world processes
or systems over time (Banks, 2000) and then improve
the collective knowledge about the system. They are
the owners of the knowledge on the models, and for
them computer science is simply seen as a tool.
However, most of them have to face the question
of emergence during the design of simulation mod-
els. Early defined by greek philosophers with the
now-famous phrase “the Whole is more than its Parts”
(Palmer, 2000), the emergence concept is present in
most scientific fields. Many are the real-world ex-
amples that thematicians try to understand and ex-
plain: collective behavior among ants colonies, reg-
ulation of stock markets, flocking of birds, etc. In this
context, the scientists try to find answers to a series
of recurring questions (Di Marzo Serugendo et al.,
2006; Fromm, 2005), including: What does emerge?
What are the properties of the phenomenon that has
emerged? or Can we understand the emergence of
this phenomenon? that are all linked with the most
general one: What is emergence?
This question, which arises for real-world phe-
nomena, also arises for the virtual simulations we cre-
ate to imitate them. That is why working on the con-
cept of emergence, its representation and its integra-
tion in simulations is therefore essential to improve
our understanding (i) of the simulation itself and (ii)
of the real phenomena that are represented.
Emergence is so an extensively studied concept
in the complex systems field. Many work has been
done in software engineering (Hu et al., 2007; Abbott,
2007) and many definitions and classifications have
been proposed by the multiagent community (Deguet
et al., 2007). Multiagent Simulation (MAS) constitute
for emergence a very good expressing place when-
ever emergent phenomena are known to be unilat-
eral or bilateral (Castelfranchi, 1998), weak or strong
(Dessalles et al., 2007), synchronic or diachronic
(Stephan, 2002), intrinsic or causal (Boschetti and
Gray, 2007). Thus, MAS are particularly suited for
the emergence concept. They allow us to discover and
highlight the phenomena that emerge in the system.
A good and simple example is the one of ant-
agents who try to link their nest to a food source (Dro-
goul and Ferber, 1994). The ants have a very simple
behavior but after a while all ant-agents follow the
David D. and Courdier R. (2009).
SEE EMERGENCE AS A METAKNOWLEDGE - A Way to Reify Emergent Phenomena in Multiagent Simulations?.
In Proceedings of the International Conference on Agents and Artificial Intelligence, pages 564-569
DOI: 10.5220/0001803305640569
shortest path between their nest and the food source.
This shortest path has “emerged” from the collective
behavior of agents and its existence is due to the sys-
tem consideration as a whole.
But emergence should not be considered as a
simple observation result and emergent phenomena
should have a real place in simulation. For exam-
ple, if we want to simulate a lagoon and its fishes,
it would be better to detect potential shoals of hun-
dred fishes that can be formed, and then being able
to inject them into the system instead of any of the
hundred fishes belonging to the shoals. That is called
emergence reification. There would be there a sig-
nificant interest for the system comprehension and in
terms of complexity drop during the simulation. But
for computer scientists emergence is often manipu-
lated thanks to their knowing of simulation platforms
and specific concepts while for thematicians it is often
seen only as a result.
According to this, we propose in this paper a first
step toward a formalism which is useful but for which
we still need progress. In the next section, we will
make a key proposition by defining emergence as a
metaknowledge. Then, we will take advantage of
this definition to propose a conceptual framework in
which emergent phenomena can be reified through
emergent structures. Within this framework, emer-
gent phenomena can be detected and injected in the
simulation system through mechanisms of introspec-
tion and intercession to then be manipulated like any
other entity.
Attempting to reach a possible reification of emergent
phenomena, we focused on a concept that appears in
many scientific fields related to the design of systems
and to artificial intelligence: metaknowledge (Pitrat,
1990; Paguette, 1998; Kalfoglou et al., 2000).
We found in literature that metaknowledge is
linked with many philosophical concepts or theories
which are very close to our point of view about emer-
gence, like meronomy or mereology, research fields
that deal with relations between a Whole and its Parts
(Keet, 2006). And when we talk about Parts and
Whole, emergence is never far away. In the exam-
ple “the rum of this punch”, the punch is more than
the sum of its parts (rum, sugar, fruits, etc.), because
these ones can not be separated since the punch was
created (and irreducibility is a characteristics of some
emergent phenomena (Stephan, 2002)). This encour-
aged us in order to find a link between emergence and
Metaknowledge is a tool to work on knowledge
and it has been defined as “knowledge on knowledge
rather than knowledge about a particular area such
as mathematics, medicine or geology” (Pitrat, 1990).
The concept is very broad but it can be refined so we
can consider among others:
the metaknowledge describing knowledge;
the metaknowledge on the use of knowledge;
the metaknowledge to discover knowledge;
the metaknowledge to manipulate knowledge.
Thus, seeing metaknowledge as a tool to discover, de-
scribe, utilize and manipulate knowledge is in full fit
with the fundamental concept of emergence, includ-
ing the ones of radical novelty (knowledge discovery)
and of interdependence levels (use and manipulation
of knowledge) (Stephan, 2002). According to this, in
our approach we consider the following definition.
Definition. Emergence is a metaknowledge.
This is a key proposition of this paper. Metaknowl-
edge is at the heart of the process of transforming
information into knowledge. It offers a greater vari-
ety of attitudes and a better way to adapt changes oc-
curring in environments (Luzeaux, 1997) and in that
sense this concept is particularly suited to the study
of emergence in MAS. So this definition of emer-
gence as a metaknowledge enables us to have new
approaches for simulation modeling in order to take
into account the emergence of phenomena as best as
3.1 Modeling Approach
To reify emergent phenomena that do occur, we need
to built knowledge and metaknowledge on the simu-
lation, even if we do not forget that metaknowledge is
only a supplementary knowledge: the difference lies
in the levels of abstraction that can be obtained for
example through a subjective external observer.
We know that MAS provides knowledge on studied
systems, and it is the study of this knowledge that will
provide us metaknowledge on the initial systems. For
the sake of genericity, this proposal takes place in a
MAS model that is most general as possible. We con-
ceive that agents are entities acting through mecha-
nisms of perception, influence and interaction within
SEE EMERGENCE AS A METAKNOWLEDGE - A Way to Reify Emergent Phenomena in Multiagent Simulations?
one or more environments. The proposal is at a con-
ceptual level that allows us not to take into account
the specific concepts that are integrated and handled
the different multiagent simulation platforms.
As shown on Figure 1, the idea is to construct the
knowledge about the MAS by using mechanisms of
observation on the simulation. This will provide us
the useful elements about the agents, their environ-
ment(s), and their evolutions. The metaknowledge
about the MAS, and so the emergent phenomena that
may occur in simulation, is then defined using the el-
ements of the constructed knowledge. That means
that all the emergent phenomena will be described us-
ing combinations of whatever happened (or may hap-
pen, or... may have happen) in the simulation sys-
tem. Moreover, and by definition of metaknowledge,
when some phenomena emerge, their belonging to the
metaknowledge about the MAS means that they be-
long to the knowledge about it too, and of course to
make sense they have to be add to the MAS itself.
Figure 1: From the MAS to the metaknowledge.
This modeling approach makes possible to themati-
cians to think about the simulated systems at more
abstract levels. Indeed, emergent properties cannot be
expressed with the same level used to define the sys-
tem entities. For example, it is not possible to express
the notion of shortest path in the simulation of ants
colonies moving to a source of food with the same
parameters used to define ants behaviors.
Moreover, it offers the possibility for thematicians to
inject into a simulation the properties, behaviors and
patterns that emerged. In order to make this concep-
tual framework intent to detect and learn emergent
properties of the simulated systems, we propose in
the followings a first step toward a useful formalism
of these two aspects of emergence reification. Draw-
ing on works found in philosophy and about reflection
and metaobjects (Kiczales, 1991), we called them in-
trospection and intercession processes.
3.2 Introspection Process
In philosophy, introspection is a method of observa-
tion and analysis to study one’s own person and to
acquire self-knowledge. To yield to the detection of
emergent phenomena, the MAS has to do its own in-
trospection. Obviously, this process is based on the
analysis of the facts that are happening in the simu-
lation. So in the context of our study, we define the
introspection process as follows.
Definition. The introspection process consists in
creating and expressing knowledge and metaknowl-
edge on the simulation in order to detect phenomena
that do emerge.
To detect the emergent phenomena that appear in a
simulation, we need to know the set (or a set) of all
the facts that are occurring (or have occured in the
case of a post-simulation analysis) in the MAS. As
we said before, these facts are related to the agents
behavior and to their environment. They are function
of space, time, communication, and, more generally,
they depend on the different entities and the different
metrics present in the MAS.
Using the observation mechanisms offered by
simulation platforms (probes, publish/subscribe
mechanism) to observe the interactions of the agents
and the evolutions of the environment(s) in which
they evolve, we can build K , a set defined as follows:
K = { f acts}
This set represents the useful knowledge (for emer-
gent phenomena analysis) that can be studied on the
MAS. Moreover, we can refine K as follows :
K = K
where K
, which gathers the facts produced by
agents, is defined by:
= {in f luences, perceptions,interactions}
and where K
is the set of all the facts produced
within the environment.
The emergent phenomena are defined through the
study of the set of facts K that represents the knowl-
edge on the system. Such an analysis of K consists to
establish relations between facts. This helps to high-
light the changes that occur in the simulation and that
ICAART 2009 - International Conference on Agents and Artificial Intelligence
represent emergent phenomena. To detect these po-
tential changes, we propose to define an extensible set
of functions R
defined as follows:
= { f : K
boolean}, n N
Each function of this set detects if combinations of
parameter facts define an emergent phenomenon. We
define these functions as emergence revelators.
Using these emergence revelators, we can now de-
fine the P
set of emergent phenomena:
= { f R
/ f = true}
It is important to notice that these mechanisms of
knowledge construction and analysis do not limit
themselves to simple conjunctions of facts: time,
space, or communication events can be considered to-
gether, as any other types of facts, to reveal complex
emergence situations. For example, back to the in-
trinsic emergence (as defined in (Boschetti and Gray,
2007)) example of the lagoon and its fishes, we can
consider that the emergent phenomenon “shoal of
fishes” only occurs when hundred fishes are in the
same neighborhood, during a minimum period, and
when the fish agents interact to know in which di-
rection they should evolve. Thus, our approach is
very powerful because we give the possibility to de-
tect emergent phenomena that occurs in complex sit-
uations mixing informations on space, time, commu-
nication, or whatever.
3.3 Intercession Process
Once emergent phenomena have been detected in
simulations, we need to “give them life”. In our con-
ceptual framework, this is done during the interces-
sion process that we define as follows.
Definition. The intercession process consists in the
injection into the MAS of the emergent phenomena
that have been detected during the introspection
Most of the time, emergent phenomena that occur in
the real world are at least materialized or character-
ized. Their existence impact directly on other enti-
ties of the real world. Some of them may take part
of the emergent phenomena, others are influenced by
these phenomena, and others have their perception of
world modified from the presence of these phenom-
ena. We decided to set up such mechanisms in the
virtual world of simulations. Thus, the detection of
emergent phenomena sometimes yields to the defi-
nition of entities that will directly influence agents
behaviors in the MAS. These entities manifest them-
selves through different kind of elements that we de-
fine hereafter : emergence agents and interposition
elements managed by emergent metastructures with
which they constitute emergent structures.
An emergence agent is an agent that runs on a
MAS platform. It thus evolves in the same environ-
ment as all other agents of the system and interacts
with them through the mechanisms of influence and
perception offered by the platform. Several emer-
gence agents can be created to reify the same phe-
An interposition element is a modification of one
or several environments. It changes (as appropriate by
altering, improving, restricting, etc.) the perception
or influence mechanisms associated with one or more
These two elements are controlled by emergent
metastructures that we call ms
, which are them-
selves governed by laws of emergence. These emer-
gence laws are all the elements of the set L
as follows:
= { f : P
}, n N
where S
represents the set of all the emergent struc-
An emergent structure is defined by a tuple
<emergent metastructure, emergence agents, interpo-
sition elements>. Each emergent structure is man-
aged (created, modified, deleted) by its own emergent
An emergent phenomenon p
is materialized
in the MAS through an emergent structure s
in function of the laws defined in L
. Depending of
its properties, this emergent phenomenon p
is con-
cretely injected inside the simulation as a one or more
emergence agents, as one or more interposition ele-
ments, or as a combination of both.
The independent or complementary use of emer-
gence agents and interposition elements allow us to
take into account different types of emergence that
could occur in MAS. Thus, in the shoal of fish ex-
ample, the shoal will be represented directly in the
system by an emergence agent. The fishes that con-
stitute the shoal of fishes may continue to evolve in
their environment, but will have their influences and
perceptions changed by elements of interposition con-
trolled by the ms
corresponding to the shoal of fish.
Here, as soon as the thematicians are able to observe
and then characterize the behavior of the shoal of fish,
it is possible to keep it into the system and to remove
each one of the hundred fish-agents that may consti-
tute it. At least we can imagine that the fishes that
evolve in the shoal will not need to exchange mes-
sages to find collectively the best direction for mov-
SEE EMERGENCE AS A METAKNOWLEDGE - A Way to Reify Emergent Phenomena in Multiagent Simulations?
Figure 2: Complete framework for emergence reification.
ing, but that the shoal will take this decision for all of
them. No more interactions to deal with for hundred
agents : the complexity drop is important.
Notice that this approach supports the dynamism
of emergent phenomena, especially their volatility:
when an emergent phenomenon is no longer ob-
served, its corresponding emergent structures are
deleted from the MAS. Indeed, the functions in P
(the ones that activated the corresponding laws in L
are no longer defined.
3.4 Emergence Reification
As said in introduction, emergence should have a bet-
ter place in the design cycle of models and simula-
tions. Figure 2 shows a summary of our proposal,
with both the “inside simulation” aspect and the con-
ceptual vision for the complete cycle of emergence
reification that we define as follows.
Definition. The emergence reification consists in
doing the complete cycle of the introspection and
intercession processes
We can also see on Figure 2 that because emergent
structures can have their own behavior, they also con-
tribute (via their own influences and perceptions) to
create new knowledge that can be added to the set of
facts. Once more, this contributes to refine the knowl-
edge on the MAS.
At the junction of the conceptual works and the
MAS representation, we identified three services that
simulation platforms should have to reach such a reifi-
cation of emergent phenomena:
observation services (to set up the observation
mechanisms useful to build knowledge on the
an agent manipulation service (to create, delete
and modify agents cycle of life);
an environment manipulation service (to manipu-
late the interposition elements).
Notice that the MAS platform that we develop in our
lab fits with the conceptual framework we defined in
this paper. This argues for the feasibility of our ap-
proach. Because of the genericity of the proposal, it
is a reuse advantage that would benefit to other simu-
lation platforms that would like to set up such emer-
gence reification mechanisms.
Before concluding, it is important to remind that
with this conceptual framework, thematicians have a
new role to play during the modeling cycle of simu-
lations: thinking about the R
and L
sets that ap-
pears in the processes of introspection and interces-
sion. This is consistent with what happen in the real
world: thematicians do not have any innate knowl-
edge on emergent phenomena and it is only from ob-
servation that they have learn to recognize, character-
ize or name them. The first phenomena that emerge
and that may be detected by mechanisms of observa-
tion are the result of the descriptions initially provided
by the thematicians. So forth, the detected emergent
phenomena will constitute a kind of new knowledge,
that it is important to add to the global knowledge of
the system. This will yield to a better understanding
ICAART 2009 - International Conference on Agents and Artificial Intelligence
of the simulation system itself and to the discovery
of new emergent phenomena that could not have been
identified with the thematicians initial knowledge.
In this paper, we focused on the emergence issue and
on its representation in MAS. We consider that we
need to improve the way this concept is taken into ac-
count in simulations and we proposed a conceptual
framework that enables the reification of emergence
in simulations.
Actually, this is allowed by the analysis of the
knowledge on the simulation. That is why the main
issue of our approach is the definition of emergence
as a metaknowledge on the MAS: it is the key con-
cept that we used to propose and describe a concep-
tual framework for the detection and injection of the
different kinds of emergent phenomena. Thanks to
this, we identified the services that should be avail-
able in simulation platforms to take emergence into
Our experience on large-scale simulation projects
and our long-time wondering about the representation
of emergence in MAS (Marcenac et al., 1998) has led
us to make this conceptual framework the most sim-
ple and generic as possible. In that sense, we consider
it as a first step (i) toward a formalism which is use-
ful but for which we still need progress and (ii) in
the way of designing models and programming with
In future works, we will use the conceptual frame-
work we described in this paper to improve a multia-
gent application of energy simulation under develop-
ment on our simulation platform GEAMAS-NG, in
the context of a research program financed by the Re-
union Island. Drawing on this conceptual framework
we will also extend our platform in order to take emer-
gence into account as soon as we start the conception
of the simulation agents, while keeping a clear sepa-
ration between initial behaviors and emergent ones.
This would improve agents capacities by giving to
them the possibility of reasoning on themselves and
so on emergent phenomena that will get back some
piece of magic.
Abbott, R. (2007). Putting complex systems to work. Com-
plex., 13(2):30–49.
Banks, J. (2000). Simulation in the future. Proceedings of
the 32nd conference on Winter simulation, pages 7–
Boschetti, F. and Gray, R. (2007). A turing test for emer-
gence. Advances in Applied Self-organizing Systems,
pages 349–364.
Castelfranchi, C. (1998). Modelling social action for ai
agents. Artif. Intell., 103(1-2):157–182.
Deguet, J., Demazeau, Y., and Magnin, L. (2007). Emer-
gence and software development based on a survey of
emergence definition. Studies in Computationnal In-
telligence, 56:13–21.
Dessalles, J., M
uller, J., and Phan, D. (2007). Emergence in
multi-agent systems: conceptual and methodological
issues. Agent-based modelling and simulation in the
social and human sciences, pages 327–355.
Di Marzo Serugendo, G., Gleizes, M.-P., and Karageorgos,
A. (2006). Self-Organisation and Emergence in Multi-
Agent Systems: An Overview. Informatica, 30(1):45–
Drogoul, A. and Ferber, J. (1994). Multi-agent simulation
as a tool for modeling societies: Application to social
differentiation in ant colonies. Artificial Social Sys-
tems, (830):3–23.
Fromm, J. (2005). Ten questions about emergence.
arXiv:nlin/0509049v1 [nlin.AO].
Hu, J., Liu, Z., Reed, G., and Sanders, J. (2007). Posi-
tion paper: Ensemble engineering and emergence (and
ethics?). UNU-IIST Technical Report, 390.
Kalfoglou, Y., Menzies, T., Althoff, K.-D., and Motta,
E. (2000). Meta-knowledge in systems design:
panacea...or undelivered promise? The Knowledge
Engineering Review, 15(4):381–404.
Keet, C. (2006). Part-whole relations in object-role mod-
els. Object-Role Modelling (ORM 2006), pages 1118–
Kiczales, G. (1991). The art of the metaobject protocol. The
MIT Press.
Luzeaux, D. (1997). Learning knowledge-based systems
and control of complex systems. 15th IMACS World
Marcenac, P., Courdier, R., Calderoni, S., and Soulie, J.-C.
(1998). Towards an emergence machine for complex
systems simulation. Lecture Notes in Computer Sci-
ence, 1416:785–794.
Paguette, G. (1998). Metaknowledge representation: appli-
cation to learning systems engineering. TLNCE Tech-
nical Reports.
Palmer, J. A. (2000). Aristotle on the ancient theologians.
Apeiron, 33(3):181–205.
Pitrat, J. (1990). M
eta-connaissances, futur de
l’intelligence artificielle. Hermes.
Stephan, A. (2002). Emergence. Encyclopedia of Cognitive
Science, 1:1108–1115.
SEE EMERGENCE AS A METAKNOWLEDGE - A Way to Reify Emergent Phenomena in Multiagent Simulations?