Click here to login if you're an NAE Member
Recover Your Account Information
Author: Zoltán Toroczkai and Stephen Eubank
Control over agent-based systems can be achieved via modeling tools.
Researchers have made considerable advances in the quantitative characterization, understanding, and control of nonliving systems. We are rather familiar with physical and chemical systems, ranging from elementary particles, atoms, and molecules to proteins, polymers, fluids, and solids. These systems have interacting particles and well defined physical interactions, and their properties can be described by the known laws of physics and chemistry. Most important, given the same initial conditions, their behavior is reproducible (at least statistically).
However, other types of ubiquitous systems are all around us, namely systems that involve living entities (i.e., agents) about which we have hardly any quantitative understanding, either on an individual or collective level. In this paper, we refer to collectives of living entities as “agent-based systems” or “agent systems” to distinguish them from classical particle systems of inanimate objects. Although intense efforts have been made to study these systems, no generally accepted unifying framework has been found. Nevertheless, understanding, and ultimately controlling the behavior of agent systems, which have applications from biology to the social and political sciences to economics, is extremely important. Ultimately, a quantitative understanding can be a basis for designing agent systems, like robots or rovers that can perform tasks collectively that would be prohibitive for humans. Examples include deep-water rescue missions, minefield mapping, distributed sensor networks (for civil and military uses), and rovers for extraterrestrial exploration.
Even though there is no unifying understanding of agent systems, some control over their behavior can be achieved via agent-based modeling tools. The idea behind agent-based modeling is rather simple—build a computer model of the agent system under observation using a bottom-up approach by trying to mimic as much detail as possible. This can be rather expensive, however, because it requires (1) data collection, (2) model building, (3) exploitation of the model and the collection of statistics, and (4) validation, which normally means comparing the output of the model with additional observations of the real system.
The agent models described in this paper took about nine years to develop at Los Alamos National Laboratory. However, the framework for these models can be used to simulate many similar circumstances and to make predictions.
Properties of Agent Systems
Agent-based systems are hard to describe and understand within a unified approach because they differ from classical particle systems in at least two ways. First, an agent is a complex entity that cannot be represented by a simple function, such as a Hamiltonian function of a classical system (e.g., a spin system). Second, the interaction topology, namely the rules by which particles interact with each other, is generally represented by a complex, dynamic graph (network), unlike the regular lattices of crystalline solids or the continuous spaces of fluid dynamics. In many cases, the notion of “locality” itself is elusive in agent-based networks; in social networks, for example, the physical or spatial locality of agents may have little to do with social “distances” and interactions among them. To illustrate the complex structure of a “particle,” or agent, and its consequences we can use traffic, namely people (agents) driving on a highway, as an example. Keep in mind, however, that the statements in this description are generally applicable to other agent systems.
Agents have the following qualities: