Agent-Based Model

Agent-based modelling is the modelling of phenomena as dynamical systems of interacting agents. An agent-based model (ABM) is a specific individual based computational model for the computer simulation of complex systems.

Definition

In agent-based modeling the word 'model' often means to specify 'the rules of the game': to specify exactly what kind of participants exist (the agents and their states) and how they interact (the rules). As Bonabeau says in his paper about agent-based modeling "at the simplest level, an agent-based model consists of a system of agents and the relationships between them". ABMs are the equivalent to games in game theory. A strategic game in game theory consists of a set of players, a set of moves (or strategies) available to those players, and a specification of payoffs for each combination of strategies. An ABM consists of a set of (simplified) agents, a set of possible basic behaviors or elementary interactions, and a set of rules which specifies the evolution or development of the system.

Contrary to general multi agent systems, an ABM describes the interactions among individual agents and their environment in a specific situation which leads to particular organizational patterns and emergent properties. Model in 'agent-based model' is used in the sense of an abstract representation of a concrete system from a particular viewpoint. ABMs can be used to explain collective human behavior in the social sciences, to understand complex systems, and they enable the user to run large scale virtual experiments without altering the corresponding real system.

Ingredients

Agent-based models involve three basic ingredients: agents, an environment and rules. The environment is the medium or space which separates the agents, it can consist of the agent themselves, too. Rules define the behavior of the agents.

The art of agent-based modeling

John H. Holland says about the art of creating models (here)

"The real art of model-building is in selecting the subset of the system that you want to model. Do not attempt to model the entire system. Throw away as much detail about the system as possible. If you build a model that includes all of the information in the system, then the model is no more helpful to you than the system itself".

In agent-based modeling less is more. George Johnson says in his book "Fire in the Mind" (Vintage, 1996): "The mark of a good simulation is that it separates the essential from the incidental, cutting through what is deemed irrelevant detail to get at the heart of the problem. This involves making instinctual judgements about which details are crucial and which can be ignored. [...] Like a theory, a simulation must be more general and abstract than what it is intended to explain. It must serve as a compression" (page 244).

ABMs can be used "to capture the principal laws behind the exciting variety of new phenomena that become apparent when the many units of a complex system interact", as Tamas Vicsek says (see [1]). Yet it is not easy to design models that are complex enough but not too complex, especially if the system we want to model is very large or very complex (if we have neither suitable experimental data nor a solid theory)

• If the simulation is too simple and matches your own theoretical ideas, then no matter how good these ideas are it is always easy to criticize that the simulation is either not realistic enough or only constructed to illustrate your own ideas and assumptions. If the model is too abstract to validate it with real data, it isn't science, although the modeling may be useful as an thought experiment, as a metaphor or as an illustration of a principal law in a complex system

• If the simulation is too complex and matches every official experimental data, everything takes a lot amount of time (creation, setup and execution of the experiment and finally the cumbersome analysis of the complex outcomes), and it becomes increasingly difficult to identify the principal laws, because it is easy to get lost in the data or bogged down in details. The simulated results about collective behavior can be as difficult to understand as reality itself if the model is too complex. It is clearly not helpful if we cannot understand the model any better than the real system.

It is possible to obtain similar results with different agent structures, rules, achitectures and models. Which of all these models is the minimal model ? It is difficult to find simple models with complex results. At best, the field of agent-based models can be considered currently as an art - the "art of agent-based modeling".

Examples

The boids from Craig Reynolds are a famous example. Reynolds has a long list of other individual-based models at his site. A larger number of models can also be found at the NetLogo site.

Literature

• Nigel Gilbert and Pietro Terna, How to build and use agent-based models in social science, Mind & Society, no. 1 (2000) 57-72

• Robert L. Goldstone and Marco A. Janssen, Computational models of collective behavior, TRENDS in Cognitive Sciences, Vol.9 No.9 September (2005) 424-430

• Michael W. Macy, Robert Willer, From Factors to Actors: Computational Sociology and Agent-Based Modeling, Annual Review of Sociology, Vol. 28 (2002) 143-166

• Grimm et al., Pattern-oriented modeling of agent-based complex systems, Science Vol. 310. no. 5750 (2005) 987-991,

• Steven C. Bankes, Agent-based modeling: A revolution ?, PNAS May 14 Vol. 99 Suppl. 3 (2002) 7199-7200,

• Eric Bonabeau, Agent-based modeling: Methods and techniques for simulating human systems, PNAS May 14 Vol. 99 Suppl. 3 (2002) 7280-7287

• Laszlo Gulyas, On the Transition to Agent-Based Modeling, Social Science Computer Review, Vol. 20 No. 4 (2002) 389-399

see also the special PNAS issue in 2002 about Capturing Complexity through Agent-Based Modeling

Links

• Wikipedia Entry for ABM
• Scholarpedia Entry for ABM
• The Brookings Institution, The Road to Agent-based Models

• Bonabeau's paper about agent-based modeling
• Reynolds' list of Individual-Based Models