What is an evolutionary system?

A system is a set of entities – for instance, people or machines – that interact with a common objective (Law and Kelton 1991). This objective can be implicit or explicit. Therefore, the system presents change resistance.

In practice, the definition of what is and is not a system depends on the object of study. To help in the study of the system, a system state is defined. This set of variables describes the system at a specific instant. This description tries to answer the question that motivates the construction of the model. By analyzing the system state, the first division can be established, depending on the evolution of the variables. The modification can be continuous or discrete[1].

Systems can be classified based on whether they modify their behavior during the simulation depending on the system state. Systems that modify their behavior are called evolutionary systems, the systems that learn. Those that do not learn are called interactive systems[2] (Henning 2001), where the results are obtained by the effect of the interaction between the different elements of the model.

The use of evolutionary methods rather than classical interactive methods is determined by the elements quoted below (Henning 2001):

 “If the characteristics of entities do not change, or stay in a variability range which is controlled by the factors internal to the entities (control structures, information carriers like genes in biological entities) but not a selecting influence, there is no evolution in the Darwinian sense of the term and ’interactionist’ models should suffice to describe the history of the system.” Examples of this kind of system include industrial systems, in which the variability of the different variables of the model elements (machines) is always within a specific range. Biological models are another example: the simulation time is shorter than the time required to change the behavior of the elements (study of mammals’ population dynamics).
“If the characteristics of the entities change, the characteristics are under some kind of selection and generations of characteristics are connected by tradition, i.e. there exists some form of heredity, we can meaningfully speak of and borrow from evolutionary models. These models will sometimes be confined to the biological, Neo-Darwinian type though.” Tradition appears in these models, which allows new characteristics, properties and behaviors to appear. Tradition allows the change, and selection motivates the change. Systems of this kind are able to learn.
A third case can exist. If the characteristics change but tradition is unable to create new values, the inheritance mechanism does not exist. This case is open to discussion (Henning 2001).

Systems can be classified based on time, whether the system is dynamic or static, and whether the behavior of the model elements changes. Table 1 (Henning 2001) shows how systems can be classified based on the variable pairs static/dynamic systems and interactive/evolutionary systems.

Static systems
Dynamic systems
Static interactions. The only changes are in the composition. For instance, systems that are not modified over time. Through Monte Carlo simulations, an approximate value can be obtained.[3]
System interaction. Changes in the interactions between the different model components. For instance, an industrial plant.
Evolutionary selection. Random acquisition of variations that change the composition of types.
Evolutionary system feedback that influences the supply of variation and the speed of evolution. Changes in type depend on the history of the system. For instance, the evolution of a society or wildfire with the interaction of an extinction model.

Table 1: System classification.

The classification of a system as one type or another depends on the temporal scope of the problem being studied. For instance, if the lifetime of the simulation model is short and the feedback from the modifications to the behavior of the entities has no effect on the model, then an interactive model should suffice. Other more complex classifications can be done, for instance Marín and Mehandjiev (2006) proposes a classification for AMAS[4] based on its definition and the relation with the environment: “MAS situated in an open environment and capable to self-modify its structure and internal organization by varying its elements’ interactions according to environmental changes”.

with the environment
Control system
Semi-Isolated Evolution
Complex Interactions

Table 2: AMAS System classification (Marín and Mehandjiev 2006)

 Table 2 shows that depending on the relation with the environment of the systems, these can be classified in five different categories: Automaton, Control System, Semi-Isolated Evolution, Complex Interactions, and Ecosystem”. The emergent behavior that appears[5], caused by the selection rule, can be classified in two main classes, depending on the interaction of the model individual elements:

  1. Collaboration
  2. Competition

With these two types of interaction and the possibility of a system being either interactive or evolutionary, entities can present entirely different behaviors.

 Table 3 shows the possible evolution of a model (Henning 2001).

Interactive systems
Evolutionary systems
Interactive emergence
Combinatory emergence

Table 3: Evolution due to individual behaviors.

In an interactive system, where the various elements do not evolve, differentiation appears if competition between the elements causes them to separate and become differentiated. Destabilization appears if the elements evolve or if they can develop their own strategies or behaviors in order to beat their adversaries. However, if the elements collaborate, a global behavior emerges.

The next section discusses evolutionary systems and studies the tradition mechanism that makes evolution possible. There are two main families of systems, depending on whether tradition exists: neo-Darwinist systems and evolutionary systems.

Neo-Darwinist systems are based on the assumption that mutations, or changes in the system rules, are random and not directed at any goal. Evolutionary systems are based on the assumption that changes are not random – even if they are not directed at a specific goal. In evolutionary systems, changes are directed by the existence of a structure – the structure of the organisms – and the limits it places on changes. For a more complete view of these concepts, see Riedl (1975) and Maynard and Szathmary (1995).

To avoid this dilemma, the concept of mutation can be distinguished from the changes that result from mutation. Mutation can be random and need not favor one specific direction. Hence, only a small proportion of these mutations can be applied to an organism and modify its behavior in the long term.

[1] A model can represent a system at a specific time (a static representation). The introduction describes a classification based on the existence of time evolution in a system: static systems and dynamic systems. Systems can also be classified based on whether random variables are used in the simulation model. Models that use random variables are called random models. Models without random variables are considered deterministic models.

[2] “The interest of the interactionist perspective is usually focusing on the emergence of aggregated patterns of behavior and its effect in larger systems.” (Henning 2001).

[3] One example of an interactive static system is pi-calculus using Monte Carlo simulations. In this example, there is no time but the composition is modified during the simulation because the procedure produces an increasing number of points (with more points, the value is closer to the value of pi).

[4] Adaptative multiagent-systems. adaptive agent is that who has knowledge about its own structure and evolutionary capacities (i.e. meta-knowledge), so that it can dynamically modify its behavior by changing its own structure (Guessoum 2004). These systems are in the Evolutionary section of the Henning classification.

[5] According (Marín and Mehandjiev 2006) the two kind of systems that allows the emergence of a behavior are Ecosystems and Complex-Interactions.


Guessoum, Z. 2004 Adaptive agents and multiagent systems. IEEE Distributed Systems Online 5(7) http://dsonline.computer.org/ (accessed 15 January 2007).

Henning Reschke, Carl. 2001 Evolutionary perspectives on simulations of social systems, Journal of Artificial Societies and Social Simulation, Vol. 4, No. 4. [e-journal] http://jasss.soc.surrey.ac.uk/4/4/8.html (accessed 12 January 2006).

Law, A. M.; Kelton, W. D. 1991. Simulation modeling and analysis. McGraw-Hill.

Marín, César A.; Mehandjiev, Nikolay. 2006, A Classification Framework of Adaptation in Multi-Agent Systems, Lecture Notes in Computer Science, Cooperative Information Agents X, pp 198-212,   Springer Berlin / Heidelberg, ISBN:   978-3-540-38569-1, http://www.springerlink.com/content/aq84716t957ujt11 (accessed 16 January 2007)

Riedl, R. 1975. Die Ordnung des Lebendigen. Systembedingungen der Evolution. Paperback ed. Munich: Piper.

Werner, Roland. 2000. Structure, flow, change: Towards a social systems simulation methodology. Social Systems Simulation Group. San Diego State University.

Pau Fonseca i Casas
Department of Statistics and Operations Research

Universitat Politècnica de Catalunya - BarcelonaTECH
North Campus - C5218 Room
Barcelona, 08034, SPAIN

Tel. (+34) 93 4017035
Fax. (+34) 93 4015855