Agent-based modeling

from Wikipedia, the free encyclopedia

Agent-based modeling is a special, individual-based method of computer-aided modeling and simulation, closely linked to complex systems , multi- agent systems , evolutionary programming and cellular automata .

history

Agent-based modeling has its roots in the modeling of cellular automata as well as in the various areas of artificial intelligence . When compared, the agent-based modeling can also be interpreted as an extension of cellular automata. It is a special case of microsimulation . Agent-based models are based on the theory of multi-agent systems.

theory

In contrast to other types of modeling (e.g. system dynamics ), agent-based modeling has many small units (agents) that can make decisions or act. In this sense, this type of modeling allows the connections between the micro and macro levels to be explicitly modeled or examined. The system behavior results from the behavior of the individual agents and is not specified at the system level. If there are effects on the system level that cannot be directly derived from the decision-making algorithms of the individuals, this is called emergence . In addition, system behavior that is separate from the individual decisions can be implemented.

Two crucial aspects of agent-based modeling are the ability to explicitly depict heterogeneous behavior and dependencies on other individuals.

This type of modeling is mainly used when the focus of a question is not the stability of an equilibrium or the assumption that a process will return to equilibrium, but rather the question of how a system can adapt to changed framework conditions ( robustness ). This takes into account the knowledge that complex problems require the micro-level, i.e. the decisions of individuals, their heterogeneity and their interactions, to be examined directly.

Application examples

Very different applications fall into the area of ​​agent-based modeling. They differ, for example, in the level of the agents' modeled intelligence and in the modeling of physical or social space. What all these approaches have in common is that decision-making behavior is implemented at the level of the individual.

A few examples illustrate this range.

Simple simulation of the formation of traffic jams

Vehicles (the agents) move in one dimension. The drivers or vehicles have a certain acceleration and braking behavior and maintain a minimum distance from the car in front of them. The complexity of the simulated environment is so low and the necessary artificial intelligence of the agents so limited that in this case one can also speak of a microsimulation . Nevertheless, interesting statements can be made with these models.

A traffic model with discretely modeled space (the vehicles move on grid cells) is the Nagel-Schreckenberg model , an example of a vehicle following model with continuous space is the Wiedemann model .

Formation of ant trails

Simulated ants, which secrete scents while searching for food and follow the scents of other ants, suffice with a similarly simple intelligence. The fragrances are lost over time. The two-dimensional environment can be much more complex here, for example it can contain food sources and obstacles. Even if the behavior of the individuals is simple, a complex swarm intelligence can develop here . See also the simulation of the formation of an ant road implemented in NetLogo . So-called ant algorithms were also derived from this behavior to solve combinatorial optimization problems.

segregation

The agents in Schelling's segregation model show somewhat more complex decision-making behavior. There agents make a choice based on different preferences as to which district to move to. In addition to the spatial environment, there is the social environment. The behavior of the agents depends on the behavior and preferences of other agents (social embeddedness). See also the simulation implemented in NetLogo. The model goes back to Thomas Schelling .

Social networks

Space can take a back seat if the agents' decision-making behavior no longer depends on where they are, but on the other agents with whom they have contact, such as consumer behavior or the spread of cultural norms. Social networks are simulated for this purpose . Exchanges only take place with agents with whom there is a network relationship. Here the decision-making behavior of the individual agents can become more complicated and multi-layered and, for example, as with consumats, contain repetition, imitation, social comparison and reflection.

Artificial economies

The scientific discipline Agent-based Computational Economics deals with the simulation of economic decision-making behavior on the level of individuals. Investigated questions range from auction behavior to individual work ( moral hazard ) to behavior in social dilemmas .

Social simulation

The area of ​​social simulation includes the modeling of concrete, observable situations that are examined in case studies. The resulting agent-based models depict the behavior of people in the study areas, for example farmers in a river catchment area. At the same time, they can be coupled with more or less complex models of the physical environment and contain corresponding feedback.

Agent-based modeling and economics

The application in Artificial Economics is particularly noteworthy, because the assumption of rationally acting individuals ( Homo oeconomicus ) was always a description on the aggregate level. The aggregated behavior of economically acting individuals can be described as if the individuals were acting rationally. This may be true for markets with a lot of information, lots of learning opportunities, enough time and motivation. But there are enough examples of situations in which assumptions of rational behavior do not make good predictions about actual human behavior. The interesting scientific questions, especially in relation to public goods and social dilemmas , belong to these situations. But since there is no other theory of human behavior that lends itself to aggregation in the same way as that of rationality, in such questions it is necessary to examine the heterogeneous, actually observed behavior of people. Agent-based modeling is a method to simulate this behavior and to set up and investigate hypotheses about the connections between the micro-behavior of the individuals and the macro-behavior of the system.

See also

literature

Individual evidence

  1. ↑ Štefan Emrich. "Comparison of mathematical models and development of a hybrid approach for the simulation and forecast of influenza epidemics within heterogeneous populations" (PDF; 1.8 MB), Vienna University of Technology, 2007 , Vienna, Austria
  2. ↑ Dr. Johannes Kottonau: Teacher Online - Simulation of an ant road with NetLogo . Teacher-online.de. November 3, 2004. Retrieved August 20, 2010.
  3. ↑ Uri Wilensky: NetLogo Models Library: Segregation. Retrieved November 27, 2018 .